Transportation Network Analysis

Volume I: Static and Dynamic Traffic Assignment

Stephen D. Boyles
Civil, Architectural, and Environmental Engineering
The University of Texas at Austin

Nicholas E. Lownes
Civil and Environmental Engineering
University of Connecticut

Avinash Unnikrishnan
Civil and Environmental Engineering
Portland State University

Version 0.90 (Public beta)

Updated September 15, 2022
ii

Preface
This book is the product of more than ten years of teaching transportation
network analysis, at the The University of Texas at Austin, the University of
Wyoming, the University of Connecticut, Portland State University, and West
Virginia University. The project began during a sabbatical visit by the second
author to The University of Texas at Austin, and has continued since. This
version is being released as a public beta, with a final first edition hopefully
to be released within a year. In particular, we aim to improve the quality of
the figures, add some additional examples and exercises, and add historical and
bibliographic notes to each chapter. We are also developing a companion set of
lecture slides and assignments. A second volume, covering transit, freight, and
logistics, is also under preparation.
Any help you can offer to improve this text would be greatly appreciated,
whether spotting typos, math or logic errors, inconsistent terminology, or any
other suggestions about how the content can be better explained, better orga-
nized, or better presented.
We gratefully acknowledge the support of the National Science Foundation
under Grants 1069141/1157294, 1254921, 1562109/1562291, 1636154, 1739964,
and 1826320. Travis Waller (University of New South Wales), Chris Tampère
(Katholieke Universiteit Leuven), and Xuegang Ban (University of Washington)
hosted visits by the first author to their respective institutions, and provided
wonderful work environments where much of this writing was done.

Target Audience
This book is primarily intended for first-year graduate students, but is also
written with other potential audiences in mind. The content should be fully
accessible to highly-prepared undergraduate students, and certain specialized
topics would be appropriate for advanced graduate students as well. The text
covers a large number of topics, likely more than would be covered in one or
two semesters, and would also be useful for self-paced learners, or practitioners
who may want in-depth learning on specific topics. We have included some
supplementary material in optional sections, marked with an asterisk, which we
believe are interesting, but which can be skipped without loss of continuity.
The most important prerequisites for this book are an understanding of
multivariate calculus, and the intellectual maturity to understand the tradeoffs
involved in mathematical modeling. Modeling does not involve one-size-fits-all
approaches, and dogma about the absolute superiority of one model or algorithm
over another is scarce. Instead, the primary intent of this book is to present a
survey of important approaches to modeling transportation network problems,
as well as the context to determine when particular models or algorithms are
appropriate for real-world problems. Readers who can adopt this perspective
will gain the most from the book.
Appendix A covers the mathematics needed for the network models in this
 iii

book. Readers of this book will have significantly different mathematical back-
grounds, and this appendix is meant to collect the necessary results and concepts
in one place. Depending on your background, some parts of it may need only a
brief review, while other parts may be completely new.
While this book does not explicitly cover how to program these models and
algorithms into a computer, if you have some facility in a programming language,
it is highly instructive to try to implement them as you read. Many network
algorithms are tedious to apply by hand or with other manual tools (calculator,
spreadsheet). Computer programming will open the door to applying these
models to interesting problems of large, realistic scale.

Difficulty Scale for Exercises

Inspired by Donald Knuth’s The Art of Computer Programming, the exercises
are marked with estimates of their difficulty. The key reference points are:
0 : A nearly trivial problem that you should be able to answer without any
pencil-and-paper work.
20 : A straightforward problem that may require a few minutes of effort, but
nothing too difficult if you have given the chapter a good read.
40 : A typical problem requiring some thought, or a few attempts at solution
in different ways, but the answer should yield after some dedicated effort.

60 : A problem of above-average difficulty, where the correct approach is not

obvious. You may require a bit of scratch paper or computer work as you
try out different approaches before settling on a solution.
80 : A highly challenging or involved problem, which may be appropriate as a
course project or other long-term study.
100 : An open problem in the research literature, whose solution would be a
substantial advance in the understanding of transportation networks.
The tens digit indicates the intellectual difficulty of the exercise, while the ones
digit indicates the amount of calculation required. An exercise rated at 50 may
require more cleverness and insight than one ranked at 49, but the ultimate
solution is shorter. Of course, each student will find different problems more
challenging than others.
iv
Contents

I Preliminaries 1
1 Introduction to Transportation Networks 3
1.1 Transportation Networks . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Examples of Networks . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Notion of Equilibrium . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Traffic Assignment . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.5 Static Traffic Assignment . . . . . . . . . . . . . . . . . . . . . . 13
1.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5.2 Critique . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 Dynamic Traffic Assignment . . . . . . . . . . . . . . . . . . . . . 16
1.6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.6.2 Critique . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.7 Historical Notes and Further Reading . . . . . . . . . . . . . . . 20
1.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 Network Representations and Algorithms 23

2.1 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Acyclic Networks and Trees . . . . . . . . . . . . . . . . . . . . . 26
2.3 Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 Shortest Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.1 Shortest paths on acyclic networks . . . . . . . . . . . . . 34
2.4.2 Shortest paths on cyclic networks: label correcting . . . . 36
2.4.3 Shortest paths on general networks: label setting . . . . . 38
2.4.4 Shortest paths from one origin to one destination . . . . . 40
2.5 Historical Notes and Further Reading . . . . . . . . . . . . . . . 42
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3 Mathematical Techniques for Equilibrium 51

3.1 Fixed-Point Problems . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2 Variational Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3 Convex Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.1 One-dimensional problems . . . . . . . . . . . . . . . . . . 61
3.3.2 Bisection method . . . . . . . . . . . . . . . . . . . . . . . 64
3.3.3 Multidimensional problems with nonnegativity constraints 66

v
vi CONTENTS

3.3.4 Constrained problems . . . . . . . . . . . . . . . . . . . . 67

3.4 Historical Notes and Further Reading . . . . . . . . . . . . . . . 69
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

II Static Traffic Assignment 73

4 Introduction to Static Assignment 75
4.1 Notation and Concepts . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1.1 Commentary . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2 Principle of User Equilibrium . . . . . . . . . . . . . . . . . . . . 80
4.2.1 A trial-and-error solution method . . . . . . . . . . . . . . 83
4.3 Three Motivating Examples . . . . . . . . . . . . . . . . . . . . . 86
4.4 Historical Notes and Further Reading . . . . . . . . . . . . . . . 93
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

5 The Traffic Assignment Problem 97

5.1 Mathematical Formulations . . . . . . . . . . . . . . . . . . . . . 97
5.1.1 Multifunctions and application to network equilibrium (*) 101
5.2 Properties of User Equilibrium . . . . . . . . . . . . . . . . . . . 104
5.2.1 Existence and link flow uniqueness . . . . . . . . . . . . . 105
5.2.2 Path flow nonuniqueness, entropy, and proportionality . . 107
5.2.3 Aggregation by origins or destinations . . . . . . . . . . . 114
5.3 Alternative Assignment Rules . . . . . . . . . . . . . . . . . . . . 117
5.3.1 System optimal assignment . . . . . . . . . . . . . . . . . 118
5.3.2 Perception errors . . . . . . . . . . . . . . . . . . . . . . . 119
5.3.3 Bounded rationality (*) . . . . . . . . . . . . . . . . . . . 122
5.4 User Equilibrium and System Optimal . . . . . . . . . . . . . . . 125
5.4.1 Externalities . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.4.2 Mathematical equivalence . . . . . . . . . . . . . . . . . . 128
5.4.3 Price of anarchy . . . . . . . . . . . . . . . . . . . . . . . 130
5.5 Historical Notes and Further Reading . . . . . . . . . . . . . . . 132
5.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

6 Algorithms for Traffic Assignment 137

6.1 Introduction to Assignment Algorithms . . . . . . . . . . . . . . 138
6.1.1 Why do different algorithms matter? . . . . . . . . . . . . 138
6.1.2 Framework . . . . . . . . . . . . . . . . . . . . . . . . . . 139
6.1.3 Convergence criteria . . . . . . . . . . . . . . . . . . . . . 141
6.2 Link-Based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 143
6.2.1 Method of successive averages . . . . . . . . . . . . . . . . 143
6.2.2 Frank-Wolfe . . . . . . . . . . . . . . . . . . . . . . . . . . 149
6.2.3 Small network example . . . . . . . . . . . . . . . . . . . 151
6.2.4 Large network example . . . . . . . . . . . . . . . . . . . 153
6.2.5 Conjugate Frank-Wolfe . . . . . . . . . . . . . . . . . . . 155
6.3 Path-Based Algorithms . . . . . . . . . . . . . . . . . . . . . . . 162
CONTENTS vii

6.3.1 Gradient projection . . . . . . . . . . . . . . . . . . . . . 165

6.3.2 Manifold suboptimization . . . . . . . . . . . . . . . . . . 169
6.4 Bush-Based Algorithms . . . . . . . . . . . . . . . . . . . . . . . 171
6.4.1 Bush labels . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.4.2 Shifting flows on a bush . . . . . . . . . . . . . . . . . . . 177
6.4.3 Improving bushes . . . . . . . . . . . . . . . . . . . . . . . 185
6.5 Likely Path Flow Algorithms . . . . . . . . . . . . . . . . . . . . 186
6.5.1 Primal method . . . . . . . . . . . . . . . . . . . . . . . . 192
6.5.2 Dual method . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.5.3 Traffic assignment by paired alternative segments . . . . . 198
6.6 Historical Notes and Further Reading . . . . . . . . . . . . . . . 215
6.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

7 Sensitivity Analysis and Applications 225

7.1 Sensitivity Analysis Preliminaries . . . . . . . . . . . . . . . . . . 225
7.2 Calculating Sensitivities . . . . . . . . . . . . . . . . . . . . . . . 229
7.2.1 Changes to the OD matrix . . . . . . . . . . . . . . . . . 230
7.2.2 Changes to a link performance function . . . . . . . . . . 233
7.3 Network Design Problem . . . . . . . . . . . . . . . . . . . . . . . 235
7.4 OD Matrix Estimation . . . . . . . . . . . . . . . . . . . . . . . . 242
7.5 Historical Notes and Further Reading . . . . . . . . . . . . . . . 247
7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247

8 Extensions of Static Assignment 251

8.1 Elastic Demand . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.1.1 Demand functions . . . . . . . . . . . . . . . . . . . . . . 252
8.1.2 Gartner’s transformation . . . . . . . . . . . . . . . . . . 254
8.1.3 Variational inequality formulation . . . . . . . . . . . . . 255
8.1.4 Optimization formulation . . . . . . . . . . . . . . . . . . 255
8.1.5 Solution method . . . . . . . . . . . . . . . . . . . . . . . 257
8.1.6 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
8.2 Link Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
8.2.1 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 261
8.2.2 Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
8.2.3 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 267
8.3 Stochastic User Equilibrium . . . . . . . . . . . . . . . . . . . . . 272
8.3.1 Discrete choice modeling . . . . . . . . . . . . . . . . . . . 273
8.3.2 Logit route choice . . . . . . . . . . . . . . . . . . . . . . 275
8.3.3 (*) The Markov property and the logit formula . . . . . . 280
8.3.4 Stochastic network loading . . . . . . . . . . . . . . . . . 286
8.3.5 Stochastic user equilibrium . . . . . . . . . . . . . . . . . 293
8.3.6 Relationships between logit loading and most likely path
flows (*) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
8.3.7 Alternatives to logit loading . . . . . . . . . . . . . . . . . 301
8.4 Historical Notes and Further Reading . . . . . . . . . . . . . . . 302
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
viii CONTENTS

III Dynamic Traffic Assignment 311

9 Network Loading 313

9.1 Link Model Concepts . . . . . . . . . . . . . . . . . . . . . . . . . 314
9.1.1 Sending and receiving flow . . . . . . . . . . . . . . . . . 315
9.1.2 Point queue . . . . . . . . . . . . . . . . . . . . . . . . . . 317
9.1.3 Spatial queue . . . . . . . . . . . . . . . . . . . . . . . . . 319
9.2 Node Model Concepts . . . . . . . . . . . . . . . . . . . . . . . . 321
9.2.1 Links in series . . . . . . . . . . . . . . . . . . . . . . . . 324
9.2.2 Merges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326
9.2.3 Diverges . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
9.3 Combining Node and Link Models . . . . . . . . . . . . . . . . . 329
9.4 Elementary Traffic Flow Theory . . . . . . . . . . . . . . . . . . 331
9.4.1 Traffic state variables . . . . . . . . . . . . . . . . . . . . 333
9.4.2 Cumulative counts and conservation . . . . . . . . . . . . 335
9.4.3 The Lighthill-Whitham-Richards model . . . . . . . . . . 339
9.4.4 Characteristics and the Newell-Daganzo method . . . . . 344
9.5 LWR-based Link Models . . . . . . . . . . . . . . . . . . . . . . . 350
9.5.1 Cell transmission model . . . . . . . . . . . . . . . . . . . 350
9.5.2 Link transmission model . . . . . . . . . . . . . . . . . . . 355
9.5.3 Point and spatial queues and the LWR model (*) . . . . . 360
9.5.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
9.6 Fancier Node Models . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.6.1 Basic signals . . . . . . . . . . . . . . . . . . . . . . . . . 365
9.6.2 Equal priority movements . . . . . . . . . . . . . . . . . . 366
9.6.3 Intersections with priority . . . . . . . . . . . . . . . . . . 372
9.7 Historical Notes and Further Reading . . . . . . . . . . . . . . . 376
9.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377

10 Time-Dependent Shortest Paths 387

10.1 Time-Dependent Shortest Path Concepts . . . . . . . . . . . . . 387
10.1.1 Time-expanded networks . . . . . . . . . . . . . . . . . . 389
10.1.2 Bellman’s principle in time-dependent networks . . . . . . 391
10.2 Time-Dependent Shortest Path Algorithms . . . . . . . . . . . . 392
10.2.1 FIFO networks . . . . . . . . . . . . . . . . . . . . . . . . 393
10.2.2 Discrete-time networks, one departure time . . . . . . . . 394
10.3 Departure Time Choice . . . . . . . . . . . . . . . . . . . . . . . 398
10.3.1 Artificial origins and destinations . . . . . . . . . . . . . . 398
10.3.2 Arrival and departure time preferences . . . . . . . . . . . 400
10.4 Dynamic A∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404
10.5 Historical Notes and Further Reading . . . . . . . . . . . . . . . 404
10.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
CONTENTS ix

11 Dynamic User Equilibrium 409

11.1 Towards Dynamic User Equilibrium . . . . . . . . . . . . . . . . 409
11.1.1 Flow Representation . . . . . . . . . . . . . . . . . . . . . 410
11.1.2 Travel time calculation . . . . . . . . . . . . . . . . . . . . 411
11.1.3 Determining splitting proportions . . . . . . . . . . . . . . 413
11.1.4 Principle of dynamic user equilibrium . . . . . . . . . . . 415
11.2 Solving for Dynamic Equilibrium . . . . . . . . . . . . . . . . . . 417
11.2.1 General framework . . . . . . . . . . . . . . . . . . . . . . 417
11.2.2 Convex combinations method . . . . . . . . . . . . . . . . 419
11.2.3 Simplicial decomposition . . . . . . . . . . . . . . . . . . . 420
11.2.4 Gradient projection . . . . . . . . . . . . . . . . . . . . . 422
11.3 Properties of Dynamic Equilibria . . . . . . . . . . . . . . . . . . 425
11.3.1 Existence: competition at merges . . . . . . . . . . . . . . 425
11.3.2 Uniqueness: Nie’s merge . . . . . . . . . . . . . . . . . . . 427
11.3.3 Nie’s merge redux . . . . . . . . . . . . . . . . . . . . . . 431
11.3.4 Efficiency: Daganzo’s paradox . . . . . . . . . . . . . . . . 432
11.3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 435
11.4 Historical Notes and Further Reading . . . . . . . . . . . . . . . 435
11.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436

A Mathematical Concepts 441

A.1 Indices and Summation . . . . . . . . . . . . . . . . . . . . . . . 441
A.2 Vectors and Matrices . . . . . . . . . . . . . . . . . . . . . . . . . 444
A.3 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449
A.4 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454
A.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 462

B Optimization Concepts 465

B.1 Components of an Optimization Problem . . . . . . . . . . . . . 465
B.2 Index Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
B.2.1 Single index notation . . . . . . . . . . . . . . . . . . . . . 472
B.2.2 Multiple index notation . . . . . . . . . . . . . . . . . . . 476
B.2.3 Additional information on index notation . . . . . . . . . 478
B.3 Vector and Matrix Notations . . . . . . . . . . . . . . . . . . . . 479
B.4 Examples of Basic Optimization Problems . . . . . . . . . . . . . 482
B.5 More Examples of Optimization Formulations . . . . . . . . . . . 489
B.6 General Properties of Optimization Problems . . . . . . . . . . . 495
B.6.1 Local and global solutions . . . . . . . . . . . . . . . . . . 496
B.6.2 Objective function transformations . . . . . . . . . . . . . 497
B.6.3 Existence of solutions . . . . . . . . . . . . . . . . . . . . 498
B.6.4 Necessary and sufficient conditions . . . . . . . . . . . . . 500
B.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500
x CONTENTS

C Optimization Techniques 505

C.1 More Line Search Algorithms for One-Dimensional Optimization 505
C.1.1 Golden section method . . . . . . . . . . . . . . . . . . . . 506
C.1.2 Newton’s method . . . . . . . . . . . . . . . . . . . . . . . 508
C.2 Linear Programming . . . . . . . . . . . . . . . . . . . . . . . . . 511
C.2.1 Basic properties and graphical solution method . . . . . . 511
C.2.2 Standard form of a linear program . . . . . . . . . . . . . 519
C.2.3 Simplex algorithm . . . . . . . . . . . . . . . . . . . . . . 521
C.3 Unconstrained Nonlinear Optimization . . . . . . . . . . . . . . . 528
C.3.1 Optimality conditions . . . . . . . . . . . . . . . . . . . . 528
C.3.2 Solution framework . . . . . . . . . . . . . . . . . . . . . . 532
C.3.3 Convergence criteria . . . . . . . . . . . . . . . . . . . . . 533
C.3.4 Descent direction . . . . . . . . . . . . . . . . . . . . . . . 533
C.3.5 Step size . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
C.3.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 535
C.4 Constrained Nonlinear Optimization . . . . . . . . . . . . . . . . 540
C.4.1 Optimality conditions . . . . . . . . . . . . . . . . . . . . 540
C.4.2 Examples of applying KKT conditions . . . . . . . . . . . 543
C.4.3 Algorithms for linear constraints . . . . . . . . . . . . . . 549
C.4.4 Convex combination methods . . . . . . . . . . . . . . . . 550
C.4.5 Gradient projection . . . . . . . . . . . . . . . . . . . . . 553
C.4.6 Manifold suboptimization . . . . . . . . . . . . . . . . . . 556
C.5 Integer Programming . . . . . . . . . . . . . . . . . . . . . . . . . 560
C.5.1 Branch and bound algorithm . . . . . . . . . . . . . . . . 565
C.5.2 Total unimodularity . . . . . . . . . . . . . . . . . . . . . 568
C.6 Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 569
C.6.1 Simulated annealing . . . . . . . . . . . . . . . . . . . . . 571
C.6.2 Genetic algorithms . . . . . . . . . . . . . . . . . . . . . . 577

D Algorithms and Complexity 585

D.1 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
D.2 Pseudocode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
D.3 Algorithmic Efficiency . . . . . . . . . . . . . . . . . . . . . . . . 590
D.4 Complexity Classes . . . . . . . . . . . . . . . . . . . . . . . . . . 596
D.4.1 Decision problems vs. optimization . . . . . . . . . . . . . 597
D.4.2 The classes P , N P , N P -hard, and N P -complete . . . . . 598
D.4.3 Examples demonstrating N P -completeness . . . . . . . . 602

Bibliography 607
 Part I

Preliminaries

1
Chapter 1

Introduction to
Transportation Networks

This introductory chapter lays the groundwork for traffic assignment, providing
some overall context for transportation planning in Section 1.1. Some examples
of networks in transportation are given in Section 1.2. The key idea in traf-
fic assignment is the notion of equilibrium, which is presented in Section 1.3.
The goals of traffic assignment are described in Section 1.4. Traffic assignment
models can be broadly classified as static or dynamic. Both types of models are
described in this book, and Sections 1.5 and 1.6 provide general perspective on
these types of models.

1.1 Transportation Networks

Planning helps ensure that transportation spending and policies are as effective
as possible. As transportation engineers and researchers, we support this process
by developing and running models which predict the impact of potential projects
or policies — for instance, what would be the impact on city traffic and emissions
if an extra lane was added on a major freeway? If the toll on a bridge was
reduced? If streetcar lines are installed downtown? In this way, the benefits
of projects can be compared with their costs, and funding and implementation
priorities established. Depending on the models used, a variety of measures
of effectiveness can be considered, and one may want to know the impacts of a
project on mobility, congestion, emissions, equity, toll revenue, transit ridership,
infrastructure maintenance needs, or countless other metrics.
This is rather difficult. To predict ridership on a new transit line with
complete accuracy would require knowing how many trips every single possible
rider makes, and the decision process each one of these potential riders uses when
deciding whether or not to use transit. Unlike trusses or beams, human beings
can behave in ways that are impossible to predict, maddeningly inconsistent,
and motivated by a variety of factors difficult to observe (many of which occur

3
4 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

at a subconscious level). Further, most transportation infrastructure lasts for

decades, meaning that effective planning must also predict the impact of projects
and policies decades into the future. And so far, we’ve only considered the pure
engineering dimension: introduce local, state, and federal politics into the mix,
other stakeholders such as neighborhood associations and transit agencies, and
a public with a variety of priorities (is it more important to reduce congestion,
increase safety, or have livable communities?), and the picture only grows more
complicated. What to do?
Enter the mathematical model. The purpose of a mathematical model is to
translate a complicated, but important, real-world problem into precise, quan-
titative language that can be clearly and unambiguously analyzed. By their
nature, models cannot account for all of the possible factors influencing plan-
ning. As the famous statistician George Box once quipped, “All models are
wrong, but some models are useful.” A useful model is one which provides
enough insight that good decisions can be made. To do this, a model must
capture the most important characteristics of the underlying system; must not
require more input data than what is available for calibration; and must not
require more time and memory than what available hardware permits.
Further, just because a model is useful does not mean it cannot be improved.
Indeed, this is the goal of transportation researchers around the world. The
usual pattern is to start with a model which is simple, transparent, insightful...
and also wrong. This simple model can then be improved in ways to make it
more correct and useful, and this is the general pattern which will be seen in
this book. The first network models you will see are such gross simplifications
of reality that you may question whether they can truly be of value in practice.
Perhaps they can, perhaps they can’t; but in any case, they form a foundation
for more advanced and realistic models which relax the assumptions made earlier
on.
For this reason, as a student of transportation planning, you should
always be looking for the assumptions involved in everything you
see. All models make assumptions which are not entirely correct. The relevant
questions are, how much does this assumption limit the applicability of the
model, and how easy would it be to relax this assumption? If you’re looking for
a research topic, finding an existing model and relaxing an assumption is often a
good approach. With this book, if you clearly understand all of the assumptions
underlying each model, and how they differ from those made in other models
introduced, you’re 90% of the way there.
Networks are a type of mathematical model which are very frequently used
in the study of transportation planning. This introductory chapter gives a
very brief overview of transportation networks, and provides a sketch for the
remainder of the book.
This book covers both static and dynamic network models. Static models
assume that network conditions are at steady-state, while dynamic models rep-
resent changes in congestion and demand patterns over the course of several
hours or a day. Static models were the first to be developed historically, and
remain the most commonly-used in current transportation planning practice.
1.2. EXAMPLES OF NETWORKS 5

Dynamic models are more realistic in portraying congestion, but require more
data for calibration and validation, and more computational resources to run.
Solving and interpreting the output of dynamic models is also more difficult.
As research progresses, however, more planners are using dynamic models, par-
ticularly for applications when travel conditions are changing rapidly during
the analysis period. This chapter will present a balanced perspective of the
advantages and disadvantages of dynamic traffic assignment vis-à-vis static as-
signment, but one of them is worth mentioning now: dynamic traffic assignment
models are inherently mode-specific.
That is, unlike in static assignment (where it is relatively easy to build “mul-
timodal” networks mixing roadway, transit, air, and waterway infrastructure),
the vast majority of dynamic traffic assignment models have been specifically
tailored to modeling vehicle congestion on roadways. In recent years, researchers
have started integrating other modes into dynamic traffic assignment, and this
area is likely to receive more attention in years to come. However, the congestion
model for each mode must be custom-built. This is at once an advantage (in that
congestion in different modes arises from fundamentally different sources, and
perhaps ought to be modeled quite differently) and a disadvantage (a “generic”
dynamic traffic assignment model can only be specified at a very high level). For
this reason, this book will focus specifically on automobile traffic on roadway
networks. This is not meant to suggest that dynamic traffic assignment cannot
or should not be applied to other modes, but simply an admission that covering
other modes would essentially require re-learning a new theory for each mode.
Developing such theories would make excellent research topics.

1.2 Examples of Networks

Networks are fundamental to the study of large-scale transportation models rep-
resenting an entire metropolitan area, a state, or multistate regions. They can
be applied in many contexts, including alternatives analysis, developing conges-
tion pricing plans, identifying bottlenecks and critical infrastructure, shipping
and freight logistics, multimodal planning, and disaster evacuation planning,
to name only a few. The reason network models are so useful, and so broadly
applicable, is because a mathematical network is a simple, compact, and flexible
way to represent a large, complicated system.
A network consists of links and nodes. In transportation applications, a link
usually represents a means of travel from one point to another: a road segment
between two intersections, a bus route between two stops, and so on, as seen
in Figure 1.1. The nodes, in turn, are the endpoints of the links. Quite often,
nodes are adjacent to multiple links, so a node representing an intersection may
adjoin multiple links representing road segments. Nodes and links may also be
more abstract; for instance, links in a multimodal network might represent a
transfer from one transport mode to another. The level of detail in a network
varies from application to application. For multistate freight models, major
highways may be the only links, and major cities the only nodes. For a city’s
6 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

Park-and-ride Park-and-ride

Bus stop Bus stop

Roadway link
Bus route Bus route link
Transfer link

Physical infrastructure Network representation

Figure 1.1: Nodes and links in transportation networks.

Table 1.1: Nodes and links in different kinds of transportation networks.

Network type Nodes Links
Roadway Intersections Street segments
Public transit Bus or train stops Route segments
Freight Factories, warehouses, retailers Shipping options
Air Airports Flights
Maritime Ports Shipping channels

planning model, all major and minor arterials may be included as well. For a
more detailed model, individual intersections may be “exploded” so that dif-
ferent links represent each turning movement (Figure 1.2). Other examples of
transportation networks are shown in Table 1.1.

1.3 The Notion of Equilibrium

The nature of transportation systems is that of multiple interacting systems.
Congestion is determined by the choices travelers make: where, when, how of-
ten to travel, and by what mode. At the same time, these choices depend on
congestion: travelers may choose routes or departure times to avoid congestion.
These two “systems” (travel choices and system congestion) are thus interde-
pendent and interrelated, with a circular or chicken-and-egg quality to their
relationship. This interdependency lies at the root of transportation analysis.
It is at once interesting, because of the complexity of transportation systems
involving both humans and physical systems; challenging, because we must find
a way to resolve this circular dependency; and frustrating, because obvious-
looking policy interventions can actually be counterproductive. Some examples
1.3. THE NOTION OF EQUILIBRIUM 7

Figure 1.2: Two representations of the same intersection.

Traffic flow model

Path travel times
Path flows

Route choice model

Figure 1.3: Traffic assignment as interacting systems.

of “paradoxical” effects will be seen in Chapters 4 and 11.

The schematic in Figure 1.3 shows the dependency between the choices made
by travelers (sometimes called the demand side), and the congestion and delay
in the system (sometimes called the supply side) in the basic traffic assignment
problem. Each of these systems requires a distinct set of models. Demand-side
models should be behavioral in nature, identifying what factors influence travel
choices, and how. Supply-side models are often based in traffic flow theory,
queueing theory, computer simulation, or empirical formulas describing how
congestion will form.
It is not difficult to imagine other types of mutually-dependent transporta-
tion systems. Figure 1.4 shows how one might model traffic assignment in a
region with a privately-operated toll road. Now, there are three systems. In ad-
dition to the demand side and supply side from before, the private toll operator
can also influence the state of the network by choosing the toll in some way, such
as maximizing toll revenue. But this choice is not made in isolation: as the toll
is increased, drivers will choose alternate routes, suggesting that driver choices
8 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

Traffic flow model

Path travel times
Path flows

Route choice model

Revenue
Selected toll

Toll-setting model

Figure 1.4: A more complicated traffic assignment problem, with tolls.

are affected by tolls just as the toll revenue is determined by driver choices.
It is fruitful to think of other ways this type of system can be expanded. For
instance, a government agency might set regulations on the maximum and min-
imum toll values, but travelers can influence these policy decisions through the
voting process.
The task of transportation planners is to somehow make useful predictions
to assist with policy decision and alternatives analysis, despite the complexities
which arise when mutually-dependent systems interact. The key idea is that
a good prediction is mutually consistent in the sense that all of the systems
should “agree” with the prediction. As an example, in the basic traffic assign-
ment problem (Figure 1.3), a planning model will provide both a forecast of
travel choices, and a forecast of system congestion. These should be consistent
in the sense that inputting the forecasted travel choices into the supply-side
model should give the forecast of system congestion, and inputting the fore-
casted system congestion into the demand-side model should give the forecast
of travel choices. Such a consistent solution is termed an equilibrium.
The word equilibrium is meant to allude to the concept of economic equilib-
rium, as it is used in game theory. In game theory, several agents each choose
a particular action, and depending on the choices of all of the agents, each
receives a payoff (perhaps monetary, or simply in terms of happiness or satis-
faction). Each agent wants to maximize their payoff. The objective is to find a
“consistent” or equilibrium solution, in which all of the agents are choosing ac-
tions which maximize their payoff (keeping in mind that an agent cannot control
another agent’s decision). A few examples are in order.
1.3. THE NOTION OF EQUILIBRIUM 9

Table 1.2: Alice and Bob’s game; Alice chooses the row and Bob the column.
Bob
Cactus Café Desert Drafthouse
Cactus Café (−1, −1) (1, 1)
Alice
Desert Drafthouse (1, 1) (−1, −1)

Consider first a game with two players (call them Alice and Bob), who
happen to live in a small town with only two bars (the Cactus Café and the
Desert Drafthouse). Alice and Bob have recently broken off their relationship,
so they each want to go out to a bar. If they attend different bars, both of
them will be happy (signified by a payoff of +1), but if they attend the same
bar an awkward situation will arise and they will regret having gone out at all
(signified by a payoff of −1). Table 1.2 shows the four possible situations which
can arise — each cell in the table lists Alice’s payoff first, followed by Bob’s.
Two of these are boldfaced, indicating that they are equilibrium solutions: if
Alice is at the Cactus Café and Bob at the Desert Drafthouse (or vice versa),
they each receive a payoff of +1, which is the best they could hope to receive
given what the other is doing. The states where they attend the same bar are
not equilibria; either of them would be better off switching to the other bar.
This is a game with two equilibria in which Alice and Bob always attend the
same bar each week. 1
A second game involves the tale of Erica and Fred, two criminals who have
engaged in a decade-long spree of major art thefts. They are finally apprehended
by the police, but for a minor crime of shoplifting a candy bar from the grocery
store. The police suspect the pair of the more serious crimes, but have no hard
evidence. So, they place Erica and Fred in separate jail cells. They approach
Erica, offering her a reduced sentence in exchange for testifying against Fred for
the art thefts, and separately approach Fred, offering him a reduced sentence
if he would testify against Erica. If they remain loyal to each other, they will
be convicted only of shoplifting and will each spend a year in jail. If Erica
testifies against Fred, but Fred stays silent, then Fred goes to jail for 15 years
while Erica gets off free. (The same is true in reverse if Fred testifies against
Erica.) If they both testify against each other, they will both be convicted of
the major art theft, but will have a slightly reduced jail term of 14 years for
being cooperative. This game is diagrammed in Table 1.3, where the “payoff”
is the negative of the number of years spent in jail, negative because more years
in jail represents a worse outcome. Surprisingly, the only equilibrium solution
is for both of them to testify against each other. From Erica’s perspective,
she is better off testifying against Fred no matter what Fred will do. If he is
1 There is also a third equilibrium in which they each randomly choose a bar each weekend,

but equilibria involving randomization are outside the scope of this book.
10 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

Table 1.3: Erica and Fred’s game.

Fred
Testify Remain silent
Testify (−14, −14) (0, −15)
Erica
Remain silent (−15, 0) (−1, −1)

Table 1.4: Ginger and Harold’s game.

Harold
Heads Tails
Heads (+1, −1) (−1, +1)
Ginger
Tails (−1, +1) (+1, −1)

going to testify against her, she can reduce her sentence from 15 years to 14 by
testifying against Fred. If he is going to stay silent, she can reduce her sentence
from one year to zero by testifying against him. Fred’s logic is exactly the same.
This seemingly-paradoxical result, known as the prisoner’s dilemma, shows that
agents maximizing their own payoff can actually end up in a very bad situation
when you look at their combined payoffs!
A third game, far less dramatic than the first two, involves Ginger and
Harold, who are retirees passing the time by playing a simple game. Each of
them has a penny, and on a count of three each of them chooses to reveal either
the head or the tail of their penny. If the pennies show the same (both heads
or both tails), Ginger keeps them both. If one penny shows heads and the
other shows tails, Harold keeps them both. (Table 1.4). In this case, there is
no equilibrium solution: if Ginger always shows heads, Harold will learn and
always show tails; once Ginger realizes this, she will start showing tails, and so
on ad infinitum.
You may be wondering how these games are relevant to transportation prob-
lems. In fact, the route choice decision can be seen as a game with a very large
number of players. Some drivers may choose to avoid the freeway, anticipating
a certain level of congestion and trying to second-guess what others are doing —
but surely other drivers are engaging in the exact same process. 2 Each of these
three games has bearing on the traffic assignment problem. The game with
Alice and Bob shows that some games have more than one equilibrium solution
(an issue of equilibrium uniqueness), although the two equilibrium solutions are
2 To borrow from Yogi Berra, nobody takes the freeway during rush hour anymore — it’s

too congested.
1.4. TRAFFIC ASSIGNMENT 11

the same if you just look at the total number of people at each bar and not
which individuals go at each. What does it mean for transportation planning if
a model can give several seemingly valid predictions? The game with Erica and
Fred shows that agents individually doing what is best for themselves may lead
to an outcome which is quite bad overall, an issue of equilibrium efficiency. As
we will see later on, in transportation systems this opens the door for seemingly
helpful projects (like capacity expansion on congested roads) to actually make
things worse. The game with Ginger and Harold is a case where there is no
equilibrium at all (an issue of equilibrium existence). If this could happen in a
transportation planning model, then perhaps equilibrium is the wrong concept
to use altogether. These questions of uniqueness, efficiency, and existence are
important, and will appear throughout the book.
The three example games described above can be analyzed directly, by enu-
merating all the possible outcomes. However, transportation systems involve
thousands or even millions of different “players” and an analysis by enumera-
tion is hopeless. The good news is that the number of players is so great that
little is lost in assuming that the players can be treated as a continuum. 3
This allows us to work with smooth functions, greatly simplifying the process
of finding equilibria. The remainder of this chapter introduces the basic traffic
assignment problem in terms of the equilibrium concept and with a few moti-
vating examples, but still in generally qualitative terms and restricted to small
networks. The following two chapters provide us with the mathematical vocab-
ulary and network tools needed to formulate and solve equilibrium on realistic,
large-scale systems.

1.4 Traffic Assignment

There are many possible measures of effectiveness for evaluating the impacts of a
roadway transportation project or policy. However, many of these can be calcu-
lated if one can predict the number of drivers on each roadway segment. These
are called link flows. Predicting link flows allows a city or state government to
evaluate different options.
If link flows are the output of a planning model, the main input is demo-
graphic data. That is, given certain information about a population (number
of people, income, amount of employment, etc.), we want to predict how many
trips they will make, and how they will choose to travel. Census records form an
invaluable resource for this, often supplemented with travel surveys. Commonly,
a medium-to-large random sample of the population is offered some money in
exchange for keeping detailed diaries indicating all of the trips made within the
next several weeks, including the time of day, reason for traveling, and other
3 This is analogous to solving structural design problems by assuming the usual stress-strain

relationships, which assume a continuous material. In reality, a beam or column is composed

of many distinct atoms, not a homogeneous material — but surely it is unnecessary to model
each atom separately. The continuum assumption works almost as well and is much, much
easier to work with.
12 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

Demographic data

Total number of trips

1. Trip generation 2. Trip distribution

Trip start and end points

4. Route choice 3. Mode choice

Start points, end points,
and distribution by mode
Link flows

Figure 1.5: Schematic of the four-step process.

details.
To get link flows from demographic data, most regions use the so-called four-
step model (Figure 1.5). The first step is trip generation: based on demographic
data, how many trips will people make? The second is trip distribution: once
we know the total number of trips people make, what are the specific locations
people will travel to? The third is mode choice: once we know the trip locations,
will people choose to drive, take the bus, or use another mode? The fourth and
final step is route choice, also known as traffic assignment: once we know the
modes people will take to their trip destinations, what routes will they choose?
Thus, at the end of the four steps, the transition from demographic data to link
flows has been accomplished. 4
Demographics are not uniform in a city; some areas are wealthier than others,
some areas are residential while others are commercial, some parts are more
crowded while other parts have a lower population density. For this reason,
planners divide a city into multiple zones, and assume certain parameters within
each zone are uniform. Clearly this is only an approximation to reality, and the
larger the number of zones, the more accurate the approximation. (At the
extreme, each household would be its own zone and the uniformity assumption
becomes irrelevant.) On the other hand, the more zones, the longer it takes to
run each model, and at some point computational resources become limiting.
Typical networks used for large metropolitan areas have a few thousand zones.
Zones are often related to census tracts, to make it easy to get demographic
information from census results.
The focus of this book is the last of the four steps, traffic assignment. In the
beginning, we assume that the first three steps have been completed, and we
know the number of drivers traveling between each origin zone and destination
zone. From this, we want to know how many drivers are going to use each road-
4 In more sophisticated models, the four steps may be repeated again, to ensure that the

end results are consistent with the input data. There are also newer and arguably better
alternatives to the four-step model.
1.5. STATIC TRAFFIC ASSIGNMENT 13

Figure 1.6: Centroids (shaded) coinciding with existing infrastructure, and ar-
tificial centroids. Dashed links on the right represent centroid connectors.

way segment, from which we can estimate congestion, emissions, toll revenue,
or other measures of interest.
We’ve already discussed several of the pieces of information we need in or-
der to describe traffic assignment models precisely, including zones and travel
demand. The final piece of the puzzle is a representation of the transportation
infrastructure itself: the transportation network described more in the next
chapter.
It is usually convenient to use a node to represent each zone; such nodes
are called centroids, and all trips are assumed to begin and end at centroids.
The set of centroids is thus a subset of the set of nodes, defined in the next
chapter. Centroids may coincide with physical nodes in the network. Centroids
may also represent artificial nodes which do not correspond to any one physical
point, and are connected to the physical infrastructure with links called centroid
connectors (dashed lines in Figure 1.6).

1.5 Static Traffic Assignment

Figure 1.3 is the template for all traffic assignment models, be they static or
dynamic: the choices of travelers lead to congestion patterns in the network
(as predicted by a traffic flow model), and these patterns in turn influence
the choices travelers make. The difference between static and dynamic traffic
assignment lies in the traffic flow models used. Historically, static assignment
models were the first to be developed, and research into dynamic models arose
from the need to improve earlier, static models. Dynamic traffic assignment
thus has a large number of parallels with static assignment; but where they
differ, this difference is often intentional and important. Understanding these
distinctions is key to knowing when dynamic models are appropriate to use.

1.5.1 Overview
In static assignment, the traffic flow model is based on link performance func-
tions, which map the flow on each link to the travel time on that link. Mathe-
14 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

matically, if the notation (i, j) is used to refer to a roadway link connecting two
nodes i and j, then xij is the flow on link (i, j) and the function tij (xij ) gives
the travel time on link (i, j) as a function of the flow on (i, j). These functions
tij (·) are typically assumed to be nonnegative, nondecreasing, and convex, re-
flecting the idea that as more vehicles attempt to drive on a link, the greater
the congestion and the higher the travel times will be. A variety of link per-
formance functions exist, but the most popular is the Bureau of Public Roads
(BPR) function, which takes the form

 β !
xij
tij (xij ) = t0ij 1+α (1.1)
uij

where t0ij and uij are the free-flow time and capacity of link (i, j), respectively,
and α and β are shape parameters which can be calibrated to data. It is common
to use α = 0.15 and β = 4.
With such functions, the more travelers choose a path, the higher its travel
time will be. Since travelers seek to minimize their travel time, travelers will
not choose a path with high travel time unless there is no other option avail-
able. Indeed, if travelers only choose paths to minimize travel time, and if they
have perfect knowledge of network conditions, then the network state can be
described by the principle of user equilibrium: all used paths between the same
origin and destination have equal and minimal travel times, for if this were not
the case travelers would switch from slower routes to faster ones, which would
tend to equalize their travel times.
It is not difficult to show that this user equilibrium state is not socially op-
timal, and that other configurations of traffic flow can reduce the average travel
time (or even the travel time for all drivers) compared to the user equilibrium
state. In other words, individual drivers seeking to minimize their own travel
times will not always minimize travel times throughout the network, and this
latter system optimal state can be contrasted with the user equilibrium one.
The prime advantage of using link performance functions like that in equa-
tion (1.1) is that the user equilibrium and system optimum states can be found
with relative ease, even in realistic networks involving tens of thousands of
links. Part II of the book discusses this in detail, showing how the static assign-
ment problem can be formulated using the mathematical tools of optimization,
fixed point, and variational inequality problems. These three representations of
the equilibrium problem can be linked to powerful mathematical results which
assure the existence and uniqueness of user equilibrium solutions under mild
conditions. Efficient algorithms allow these states to be identified in a matter
of minutes on large-scale networks.
For these reasons, static traffic assignment has been widely used in trans-
portation planning practice for decades, and remains a powerful tool that can
be used for performing alternatives analysis and generating forecasts of network
conditions.
1.5. STATIC TRAFFIC ASSIGNMENT 15

1.5.2 Critique
There are also a number of serious critiques of static assignment models, focused
primarily on the link performance functions. By definition, static models do
not monitor how network conditions (either demand or congestion) change over
time, and implicitly assume a steady-state condition. This is clearly not the
case. There are additional, subtler and more fundamental problems with link
performance functions. This section describes a few of these problems.
First, not all vehicles on a link experience the same travel time. Even if
the demand for travel on a link exceeds the capacity, the first vehicles to travel
on that link will not experience much delay at all, while vehicles which arrive
later may experience a very high delay. Together with the principle of user
equilibrium, this means that the paths chosen by travelers will also depend on
when they are departing. Route choices during periods of high congestion will be
different from route choices made while this congestion is forming or dissipating.
Furthermore, the travel time faced by a driver on a link depends crucially on the
vehicles in front of them, and very little on the vehicles behind them. (A driver
must adjust their speed to avoid colliding with vehicles downstream; a driver
has no obligation to change their speed based on vehicles behind them.) This
asymmetry is known as the anisotropic property of traffic flow, and it is violated
by link performance functions — an increase in flow on the link is assumed to
increase the travel time of all vehicles, and directly using link performance
functions in a dynamic model would lead to undesirable phenomena, such as
vehicles entering a link immediately raising the travel time for all other vehicles
on the link, even those at the downstream end.
Second, the use of the word “flow” in static assignment is problematic. In
traffic engineering, flow is defined as the (time) rate at which vehicles pass a
fixed point on the roadway, and capacity is the greatest possible value for flow.
By definition, flow cannot exceed capacity. However, the BPR function (1.1)
imposes no such restriction, and it is common to see “flow-to-capacity” ratios
much greater than one in static assignment. 5 Instead, the xij values in static
assignment are better thought of as demand rather than actual flow, since there
is no harm in assuming that the demand for service exceeds the capacity, but it is
impossible for the flow itself to exceed capacity. And for purposes of calibration,
demand is much harder to observe than actual flow. These issues do not have
clean resolutions.
Third, and related to the previous issue, link performance functions suggest
that lower-capacity links have higher travel times under the same demand. But
consider what happens at a freeway bottleneck, such as the lane drop shown in
Figure 1.7. Congestion actually forms upstream of a bottleneck, and downstream
of the lane drop there is no reason for vehicles to flow at a reduced speed. In
5 There are several ways to add such a restriction, but these are less than satisfactory. A

link performance function which tends to +∞ as capacity is reached introduces numerical

issues in solving for equilibrium. Explicitly adding link capacity constraints to the traffic
assignment problem may make the problem infeasible, during peak periods there may be no
way to assign all vehicles to the network without (temporarily) exceeding capacity.
16 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

Traffic flow

Figure 1.7: Congestion arises upstream of a bottleneck, not on the link with
reduced capacity.

reality, it is upstream links that suffer when the demand for traveling on a link
exceeds the capacity, not the bottleneck link itself.
Fourth, in congested urban systems it is very common for queues to fill the
entire length of a link, causing congestion to spread to upstream links. This is
observed on freeways (congested offramp queues) and in arterials (gridlock in
central business districts) and is a major contributor to severe delay. In addition
to the capacity, which is a maximum flow rate, real roadways also have a jam
density, a maximum spatial concentration of vehicles. If a link is at jam density,
no more vehicles can enter, which will create queues on upstream links. If these
queues continue to grow, they will spread even further upstream to other links.
Capturing this phenomenon can greatly enhance the realism of traffic models.
For all of these reasons, dynamic traffic assignment models shift to an en-
tirely different traffic flow model. Rather than assuming simple, well-behaved
link performance functions for each link we turn to traffic flow theory, to find
more realistic ways to link traffic flow to congestion. Some early research in
dynamic traffic assignment attempted to retain the use of link performance
functions — for instance, modeling changes in demand by running a sequence
of static assignment models over shorter time periods, with methods for linking
together trips spanning multiple time periods. While this addresses the obvi-
ous shortcoming of static models, that they cannot accommodate changes in
demand or model changes in congestion over time, it does nothing to address
the more serious and fundamental problems with link performance functions
described above. For this reason, this approach has largely been abandoned in
favor of entirely different traffic flow models.

1.6 Dynamic Traffic Assignment

Dynamic traffic assignment arose from efforts to resolve the problems with static
assignment noted in the previous section. While it is being used more and more
in practice, it has not completely supplanted static traffic assignment. This
is partially due to institutional inertia, but is also due to a few drawbacks
associated with more realistic traffic flow models. Both these advantages and
1.6. DYNAMIC TRAFFIC ASSIGNMENT 17

drawbacks are discussed in this section.

1.6.1 Overview
The biggest difference between static and dynamic traffic assignment is in the
traffic flow models used. A number of different traffic flow models are available,
and a number of them are discussed in the following chapter. At a minimum, a
traffic flow model for dynamic traffic assignment must be able to track changes
in congestion at a fairly fine resolution, on the order of a few seconds. To do
this, the locations of vehicles must be known at the same resolution. Most of
them also address some or all of the shortcomings of link performance functions
identified above.
The behavior of drivers is similar to that in static assignment in that drivers
choose paths with minimum travel time. However, since congestion (and there-
fore travel time) changes over time, the minimum-time paths also change with
time. Therefore, the principle of user equilibrium is replaced with a principle
of dynamic user equilibrium: All paths used by travelers departing the same
origin, for the same destination, at the same time, have equal and minimal
travel times. Travelers departing at different times may experience different
travel times; all the travelers departing at the same time will experience the
same travel time at equilibrium, regardless of the path they choose. By virtue
of representing demand changes over time, dynamic traffic assignment can also
incorporate departure time choice of drivers, as well as route choice. Some dy-
namic traffic assignment models include both of these choices, while others focus
only on route or departure time choice. Which choices are appropriate to model
depends on which is more significant for a particular application, as well as the
data and computational resources available. Most chapters of this book focus
only on route choice, and as a general rule we will assume that departure times
are fixed. In a few places we show how departure time choice can be added in.
It is important to specify that this equilibrium is based on the travel times
the drivers actually experience, not the “instantaneous” travel times at the
moment they depart. That is, we do not simply assume that drivers will pick
the fastest route based on current conditions (as would be provided by most
advanced traveler information services), but that they will anticipate changes in
travel times which will occur during their journey. This suggests that drivers are
familiar enough with the network that they know how congestion changes with
time. This distinction is important — dynamic traffic assignment equilibrates
on experienced travel times, not instantaneous travel times.
This means that it is impossible to find the dynamic user equilibrium in one
step. Experienced travel times cannot be calculated at the moment of departure,
but only after the vehicle has arrived at the destination. Therefore, dynamic
traffic assignment is an iterative process, where route choices are updated at
each iteration until an (approximate) dynamic user equilibrium solution has
been found. This iterative process virtually always involves three steps, shown
in Figure 1.8:
18 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

1. Calculate route
travel times 2. Find shortest paths

3. Adjust route choices

toward equilibrium

Figure 1.8: Three-step iterative process for dynamic traffic assignment.

Network loading: This is the process of using a traffic flow model to calcu-
late the (time-dependent) travel times on each link, taking the routes and
departure times of all drivers as inputs. In static assignment, this step
was quite simple, involving nothing more than evaluating the link perfor-
mance functions for each network link. In dynamic traffic assignment, this
involves the use of a more sophisticated traffic flow model, or even the use
of a traffic simulator. Network loading is discussed in Chapter 9.

Path finding: Once network loading is complete, the travel time of each link,
at each point in time, is obtained. From this, we find the shortest path
from each origin to each destination, at each departure time. Since we
need experienced travel times, our shortest path finding must take into
account that the travel time on each link depends upon the time a vehicle
enters. This requires time-dependent shortest path algorithms, which are
discussed in Chapter 10.

Route updating: Once time-dependent shortest paths have been found for
all origins, destinations, and departure times, vehicles can be shifted from
their current paths onto these new, shortest paths. As in static assignment,
this step requires care, because shifting vehicles will change the path travel
times as well. A few options for this are discussed in Chapter 11, along
with other issues characterizing dynamic equilibrium. Unfortunately, and
in contrast with static assignment, dynamic user equilibrium need not
always exist, and when it exists it need not be unique.

1.6.2 Critique
The primary advantage of dynamic traffic assignment, and a significant one,
is that the underlying traffic flow models are much more realistic. Link per-
formance functions used in static assignment are fundamentally incapable of
representing many important traffic phenomena. However, dynamic traffic as-
signment is not without its drawbacks as well.
1.6. DYNAMIC TRAFFIC ASSIGNMENT 19

Dynamic assignment requires considerably more computational time and

memory than static assignment. As computers advance, this drawback is be-
coming less severe; but regardless of the computational resources available, one
can run a large number of static traffic assignments in the time required for a
single dynamic traffic assignment run. It may be advantageous to examine a
large number of scenarios with static assignment, rather than a single run with
dynamic traffic assignment, particularly if there is a lot of uncertainty in the
input data.
Dynamic assignment models also require more input data for calibration.
In addition to the usual link parameters such as capacity and free-flow time,
dynamic traffic assignment models require a time-dependent origin-destination
matrix, often known as a demand profile. Estimating even a static origin-
destination matrix is difficult; estimating a dynamic demand profile can be
even harder.
Furthermore, in addition to simply requiring more input data, dynamic traf-
fic assignment also requires more accurate input data. Dynamic traffic assign-
ment tends to be much more sensitive to the input data. Unfortunately, these
models are more sensitive precisely because they are more realistic — features
such as queue spillback mean that even a single erroneous capacity value can
have ramifications throughout the entire network, not just on the link with the
wrong value. If there is great uncertainty in the inputs (for instance, when
attempting to predict demand decades into the future), then using a dynamic
traffic assignment model may actually be further away from the truth than a
static model, despite its more “realistic” congestion model.
Separately, dynamic traffic assignment is a relatively young field relative to
static assignment, and there is no consensus on a single formulation. There are
a large number of software packages (and an even larger number of academic
models) which make differing assumptions and can lead to distinct results. By
contrast, the optimization, variational inequality, and fixed point formulations in
static assignment are completely standard, and therefore these models are more
transparent. This book attempts to provide a high-level perspective, along with
detailed discussions of a few of the most common modeling choices.
Finally, dynamic traffic assignment generally lacks neat, exact mathematical
properties. In many dynamic traffic assignment models, one can create exam-
ples where no dynamic user equilibrium exists, or where this equilibrium is not
unique. Proving convergence of iterative schemes is also more difficult.
All of these drawbacks must be traded off against the increased realism of
dynamic traffic assignment models. Both static and dynamic traffic assignment
models have their place as distinct tools for transportation engineering, and you
should learn to identify circumstances where one or the other is more appropri-
ate. As a general rule of thumb, dynamic models are most appropriate when the
input data are known with high certainty (as in present-day traffic studies), and
when queue lengths or other detailed congestion measures are required. Static
models, by contrast, perform best when there is considerable uncertainty in the
input data, or when it is more important to run a large number of scenarios
than to obtain detailed congestion information.
20 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS

1.7 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition. )
There is a long history of using equilibrium models in the field of economics;
an early example in the transportation field is Pigou (1920), who gave an exam-
ple showing that the equilibrium, “user optimal” solution need not be efficient,
in terms of maximizing total utility across travelers. (His example is described
later in the book, in Section 4.3.) Among traffic engineers, Wardrop (1952) is
credited with introducing two principles of route choice that correspond to the
user optimal and system optimal states. Readers interested in learning more
about game theory and economic equilibrium can consult the textbook by Fu-
denberg and Tirole (1991).
Static traffic assignment, as described in this book, was first formulated
by Beckmann et al. (1956), who also introduced mathematical representations
which will be described further in Chapter 5. A great deal of subsequent research
further developed and extended the basic static traffic assignment model. The
books by Sheffi (1985), Patriksson (1994), and Bell and Iida (1997) discuss many
of these further developments.
Dynamic traffic assignment is relatively newer. The first dynamic traffic as-
signment models were developed by Merchant and Nemhauser (1978a,b). These
models received increasing attention starting in the late 1990s, as more powerful
computers became available, and Peeta and Ziliaskopoulos (2001) give a review
of these advances. Another perspective, tailored towards practicing engineers
and planners, was provided in a primer (Chiu et al., 2010).

1.8 Exercises
1. [10] If all models are wrong, how can some of them be useful?
2. [10] All of the links in Figure 1.1 have a “mirror” connecting the same
nodes, but in the opposite direction. When might mirror links not exist?
3. [23] A network is called planar if it can be drawn in two dimensions with-
out links crossing each other. (The left network in Figure 1.2 is planar,
but not the network on the right.) Do we expect to see planar graphs in
transportation network modeling? Does it depend on the mode of trans-
portation? What other factors might influence whether a transportation
network is planar?
4. [35] Table 1.1 shows how networks can represent five types of transporta-
tion infrastructure. List at least five more systems (not necessarily in
transportation) that can be modeled by networks, along with what nodes
and links would represent.
5. [45] What factors might determine how large of a geographic area is mod-
eled in a transportation network (e.g., corridor, city, region, state, na-
1.8. EXERCISES 21

tional)? Illustrate your answer by showing how they would apply to the
hypothetical projects or policies at the start of Section 1.1.
6. [45] What factors might determine the level of detail in a transportation
network (e.g., freeways, major arterials, minor arterials, neighborhood
streets)? Illustrate your answer by showing how they would apply to the
hypothetical projects or policies at the start of Section 1.1.
7. [21] Provide an intuitive explanation of the prisoner’s dilemma, as de-
scribed in the Erica-Fred game of Table 1.3. Why does it happen? Name
other real-world situations which exhibit a similar phenomenon.
8. [20] For each of the following games, list all equilibria (or state that none
exist). In which of these games do equilibria exist; in which are the equi-
libria unique; and in which are there some equilibria which are inefficient?
These games are all played by two players A and B: A chooses the row
and B chooses the column, and each cell lists the payoffs to A and B, in
that order.

(a) L R (b) L R
U (8, 13) (5, 14) U (12, 12) (2, 10)
D (10, 10) (7, 12) D (10, 2) (4, 5)
(c) L R (d) L R
U (8, 6) (10, 8) U (4, 5) (5, 6)
D (2, 3) (8, 4) D (3, 4) (6, 3)

9. [1] Explain the difference between the demand for travel on a link, and
the flow on a link.
10. [42] Explain why a less realistic model less sensitive to correct input data
may be preferred to a more realistic model more sensitive to correct inputs,
and in what circumstances. Give specific examples.
22 CHAPTER 1. INTRODUCTION TO TRANSPORTATION NETWORKS
Chapter 2

Network Representations
and Algorithms

This chapter introduces networks as they are used in the transportation field.
Section 2.1 introduces the mathematical terminology and notation used to de-
scribe network elements. Section 2.2 discusses two special kinds of networks
which are used frequently in network analysis, acyclic networks and trees. Sec-
tion 2.3 then describes several ways of representing a network in a way computers
can use when solving network problems. This third section can be skipped if
you do not plan to do any computer programming. Section 2.4 describes the
shortest path problem, a classic network algorithm which plays a central role in
traffic assignment.

2.1 Terminology
Because of its relative youth, there are a variety of notational conventions in the
transportation networks community. A common notation is adopted in the book
to present the methods and topics in a consistent manner, but you should be
aware that other authors may use slightly different conventions and definitions.
Table 2.1 gives an example of some terms which are often used synonymously
(or nearly synonymously) with ours. These differences are mainly a matter of
style, not substance, but when reading other books or papers you should pay
careful attention to the exact wording of their definitions.
The fundamental construct we will utilize is the network. A network is
a collection of nodes, and a collection of links which connect the nodes. A
network is denoted G = (N, A), where N is the set of nodes and A is the
set of links. Figure 2.1(a) shows a simple network, with four nodes in the set
N = {1, 2, 3, 4} and five links in the set A = {(1, 2), (1, 3), (2, 3), (2, 4), (3, 4)}.
Notice that the notation for each link contains the two nodes connected by the
link: the upstream node is called the tail of the link, and the downstream node
the head. We will often refer to the total number of nodes in a network as n

23
24 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

Table 2.1: A thesaurus of network terminology

Terminology in this book Alternative terms
Network Graph
Node Vertex
Link Arc, edge, line
Tree Arborescence

2
(1, 2) (2, 4)

1 (2, 3) 4

(1, 3) (3, 4)

3
Figure 2.1: Example network notation

and the total number of links as m. This book is solely concerned with the
case where n and m are finite — networks with infinitely many nodes and links
are sometimes studied in theoretical mathematics, but rarely in transportation
applications.
All of the networks in this book are directed. In a directed network, a link
can only be traversed in one direction, specified by the ordering of the nodes.
Therefore, (1,2) and (2,1) represent different links: they connect the same nodes,
but represent travel in opposite directions. Unless stated otherwise, we assume
that there are no “self-loops” (i, i) from a node to itself, and no parallel links, so
the notation (i, j) is unambiguous. This is not usually a restrictive assumption,
since we can introduce artificial nodes to split up self-loops or parallel links,
as shown in Figure 2.2. The upper left panel in this figure shows a network
with a self-loop (2,2). By introducing a fourth node in the bottom left, we have
divided the self-loop into two links (2,4) and (4,2). In the upper right panel of
the figure, there are two networks connecting nodes 2 and 3, so the notation
(2,3) does not tell us which of these two links we are referring to. By introducing
a fourth node in the bottom right, we now have three links: (2,3), (2,4), and
(4,3), which collectively represent both of the ways to travel between nodes 2
and 3 in the original network, but without parallel links. These new nodes
are artificial in the sense that they do not represent physical transportation
infrastructure, but are inserted for modeling convenience. Artificial nodes (and
2.1. TERMINOLOGY 25

1 2 3 1 2 3

1 2 3 1 2 3

4
4

Figure 2.2: Artificial links can be introduced to simulate self-loops and parallel
links.

(i, j)
i j

Figure 2.3: Generic two-node network

artificial links) play an important role in simplifying certain network problems,

as discussed throughout the book.
To introduce some more terminology, Figure 2.3 shows a two-node network
with nodes i and j connected by the link (i, j). In this figure, we say that link
(i, j) is incident to both i and j because it is connected to both. We would
further say that (i, j) is an outgoing link of i and an incoming link of j. If
(i, j) ∈ A, then we say that node j is adjacent to node i. In this example, j is
adjacent to i, but i is not adjacent to j. Adjacency can also be applied to links;
two links are adjacent if the head of the first link is the tail of the second link.
In Figure 2.1, both links (2,3) and (2,4) are adjacent to (1,2).
The forward star of a node i is the set of all outgoing links, denoted Γ(i);
the reverse star is the set of all incoming links, denoted Γ−1 (i). In Figure 2.1,
Γ(2) = {(2, 3), (2, 4)} and Γ(3) = {(3, 4)}, while Γ−1 (2) = {(1, 2)} and Γ−1 (3) =
{(1, 3), (2, 3)}.
The degree of a node is the total number of links incident to that node. The
node degree can be separated into the indegree and outdegree of a node. The
indegree is the total number of incoming links at a node, while the outdegree
is the number of outgoing links at a node. The degree is then the sum of the
indegree and outdegree. Referring back to Figure 2.1, node 2 has an indegree
of 1, an outdegree of 2 and a degree of 3.
A path π is a sequence of adjacent links connecting two nodes i0 and ik . We
can either write π as an ordered set of links

{(i0 , i1 ), (i1 , i2 ), (i2 , i3 ), . . . , (ik−1 , ik )} ,

26 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

(a) (b) (c)

Figure 2.4: Networks which are (a) strongly connected; (b) connected, but not
strongly connected; (c) disconnected.

or more compactly, by the nodes passed on the way with the notation

[i0 , i1 , i2 , i3 , . . . , ik−1 , ik ] .

A path is a cycle if the starting and ending nodes are the same (i0 = ik ). Paths
which contain a cycle are called cyclic; paths which do not have a cycle as a
component are called acyclic. Cyclic paths are uncommon in transportation
problems, so unless stated otherwise, we only consider acyclic paths. Let Πrs
denote the set of all acyclic paths connecting nodes r and s, and let Π denote
the set of all acyclic paths in the entire network, that is, Π = ∪(r,s)∈N 2 Πrs . A
network is connected if there is at least one path connecting any two nodes in
the network, assuming that the links can be traversed in either direction (that
is, ignoring the direction of the link); otherwise it is disconnected. A network
is strongly connected if there is at least one path connecting any two nodes
in the network, obeying the given directions of the links. Figure 2.4 shows
networks which are strongly connected; connected, but not strongly connected;
and disconnected.

2.2 Acyclic Networks and Trees

Networks which do not have any cycles are called acyclic networks. Acyclic
networks are extremely important, because the lack of cycles can greatly simplify
analysis. Even when networks actually have cycles in reality, many efficient
transportation network algorithms temporarily divide the network into a set of
acyclic subnetworks.
A defining characteristic of acyclic networks is that a topological order can
be established. A topological order is a labeling of each node with a number
between 1 and n (n is the number of nodes), such that every link connects a
node of lower topological order to a node of higher topological order. Every
path, therefore, consists of a sequence of nodes in increasing topological order.
As an example, in Figure 2.1, the nodes are labeled in a topological order: each
path ([1, 3, 4], [1, 2, 4], and [1, 2, 3, 4]) traverses nodes in increasing numerical
order. (The phrase “a” topological order is used because it may not be unique,
and there may be several ways to label the nodes so that links always connect
2.2. ACYCLIC NETWORKS AND TREES 27

Figure 2.5: Cyclic network with no topological order.

lower-numbered nodes to higher-numbered ones.) In general networks, a topo-

logical order may not exist (see Figure 2.5, where a cycle makes it impossible
to label nodes so that the numbers increase when traversing any link). The
following theorem shows that the existence of a topological order is a defining
characteristic of acyclic networks.
Theorem 2.1. A topological order exists on a network if and only if it is acyclic.
Proof. (⇒) Assume a topological order exists on G. Then every path contains
a sequence of nodes whose topological order is strictly increasing (if this were
not so, then there is a link connecting a higher topological order node to a lower
one, contradicting the definition of topological order). Therefore, no path can
repeat the same node more than once, and the network is acyclic.
(⇐) Assume the network G is acyclic. Then we can prove existence of a
topological order by construction. Let o(i) be the topological order of node
i. We will describe a procedure to build such a topological order one step at a
time. At any point in time, a “marked” node is one which has a topological order
assigned, and an “unmarked” node is one which does not yet have its topological
order. Because G is acyclic, there is at least one node r with no incoming links.
(You will be asked to prove this statement as an exercise.) Let o(r) = 1. Again,
because G is acyclic, there is at least one unmarked node with no incoming
links from an unmarked node; mark this node with the topological order 2.
This process can be repeated until all nodes are marked; if at any point every
unmarked node has an incoming link from another unmarked node, then a cycle
exists, contradicting the assumption of acyclicity. Otherwise, the topological
order will have been constructed. The topological order so constructed is valid
because when a node is assigned an order, the only incoming links are from nodes
which have already been marked, and thus have a lower topological order.
A tree is a special type of acyclic network, which also shows up frequently
in transportation network algorithms. A tree is defined as a network in which
there is a unique node r (called the root) which has the property that exactly
one path exists between r and every other node in the network.1 Figure 2.6
shows a tree. Other useful properties of trees are:
1 In graph theory, this is usually called an arborescence, and trees are defined for undirected

graphs. However, the transportation community generally uses the term “tree” in both cases,
a convention followed in this book.
28 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

ˆ A tree has at least two nodes with degree one.

ˆ In a tree, there is at most one path between any two nodes.
ˆ A tree would become disconnected if any of its links were deleted.
ˆ Every node i in a tree (except the root) has a unique incoming link; the
tail node of that unique link is called the parent of node i. Similarly, the
head nodes of the links in the forward star of i are its children.

Figure 2.6: A tree

2.3 Data Structures

In practice, the methods and techniques in this book are carried out by com-
puter programs. Thus, it is important not only to have a convenient way to
theoretically represent a network (the network structure) but a way to store
data so that it can be easily accessed and utilized by computer programs. To
demonstrate, Figure 2.7 presents a network similar to those earlier, with addi-
tional information provided. In this case, the extra information is the travel
time of the link from i to j, denoted tij .
The first data structure is the node-link incidence matrix, shown in Fig-
ure 2.8. The columns are indexed by the network links, and the rows by the
network nodes. A “1” in the matrix indicates an outgoing link from that node,
while a “−1” indicates an incoming link. A “0” indicates that the link is not
incident to that particular node. This is a fairly inefficient data structure, es-
pecially for large networks, as in nearly any case there are going to be a large
number of zeros in every row. Note that the forward and reverse stars of node
i can respectively be identified with the columns in the matrix which have a 1
or −1 in the i-th row.
The second data structure, a node-node adjacency matrix is a simpler repre-
sentation of a network but also suffers from inefficiencies. Each row represents a
node i and each column a node j. A “1” indicates the existence of a link (i, j).
2.3. DATA STRUCTURES 29

tij
i j

2
2 4

1 3 4

1 5

3
Figure 2.7: Graph used for network representation examples

(1, 2) (1, 3) (2, 3) (2, 4) (3, 4)

 
1 1 1 0 0 0
2 −1 0 1 1 0  
3 0 −1 −1 0 1 
4 0 0 0 −1 −1

Figure 2.8: Node-link incidence matrix for the network in Figure 2.7.
30 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

1 2 3 4
 
1 0 1 1 0
2
0 0 1 1

3 0 0 0 1 
4 0 0 0 0

Figure 2.9: Node-node adjacency matrix for the network in Figure 2.7.

In this structure, a non-zero entry provides both existence and direction infor-
mation. An example of a node-node adjacency matrix is given in Figure 2.9.
In very dense networks, with many more links than nodes (m  n), node-node
adjacency matrices can be efficient.
A shortcoming of the previous data structures is that they do not contain
information beyond the existence and direction of the links in a network. In
transportation applications we are typically working with networks for which
we want to store and use more information about each element. For example,
we may want to store information about the travel time, number of lanes (ca-
pacity), speed limit, signalization, etc. for each link. Adjacency lists give us an
opportunity to store this information efficiently in a manner that is compatible
with programming languages and their built-in data structures.
Storing a network as an array is more efficient in terms of space, and is
easier to work with in some programming languages. Storage is not wasted
on a large number of zeros, as with the adjacency matrices. A “forward star”
representation is presented below. In the forward star representation, links
are sorted in order of their tail node, as shown in Table 2.2. A second array,
called point, is used to indicate where the forward star of each node begins. If
the network has n nodes and m links, the point array has n + 1 entries, and
point(n + 1) is always equal to m + 1; the reason why is discussed below. As
shown in Table 2.2, point(2) = 3 because the link 3 is the first link adjacent
to node 2; point(3) = 5 because link 5 is the first link adjacent to node 3, and
so forth. Three conventions associated with the star representation are:
1. The number of entries in the point array is one more than the number of
nodes.
2. We automatically set point(n+1) = m + 1.
3. If there are no outgoing links, point(i) = point (i+1)
With these conventions, we can say that the forward star for any node i con-
sists of all links whose IDs are greater than or equal to point(i), and strictly
less than point(i+1). This representation is most useful when we frequently
need to loop over all of the links leaving a particular node, which can be accom-
plished by programming a “for” loop between point(i) and point(i+1) - 1,
and referring to the link arrays with these indices. We can make this statement
universally, because we defined the point array to have one more entry than
the number of nodes. If point only had n entries, then referring to point(n+1)
2.4. SHORTEST PATHS 31

Table 2.2: Forward star representation of a network.

Node Point Link tail head cost capacity
1 1 1 1 3 1 ···
2 3 2 1 2 2 ···
3 5 3 2 3 3 ···
4 6 4 2 4 4 ···
5 6 5 3 4 6 ···

would be meaningless in the above statements. It is also possible to store a net-

work in a “reverse star” representation along similar lines, which is most useful
when we frequently need to loop over all the links entering a particular node.
The exercises explore this further, along with an array-based method for easily
iterating over both the forward and reverse stars of a node.
The purpose of this brief discussion of data structures is not to provide a
comprehensive treatment of the topic. Rather, it is intended to present some
basic ideas related to network storage and data representation. It should also
prompt you to think about data structures carefully when working with net-
works. Transportation network problems tend to be large in size and compli-
cated in nature. The difference between computer code that takes minutes to
run, as opposed to hours, is often the way data is stored and passed.

2.4 Shortest Paths

As a first network algorithm, we’ll discuss how to find the shortest (least cost)
path between any two nodes in a network in an efficient and easily-automated
way. In shortest path algorithms, we are given a network G = (N, A); each link
(i, j) ∈ A has a fixed cost cij . The word “cost” does not necessarily mean a
monetary cost, and refers to whatever total quantity we are trying to minimize.
In traffic applications, we often use the link travel times tij as the cost, to find the
least travel-time path. If there are tolls in a network, the cost of each link may
be the sum of the toll on that link, and the travel time on that link multiplied by
a “value of time” factor converting time into monetary units. Costs may reflect
still other factors, and some of these are explored in the exercises. Because we
can solve all of these problems in the same way, we may as well use a single name
for all of these quantities we are trying to minimize, and “cost” has become the
standard term.
The word “fixed” in the last paragraph is emphasized because in many trans-
portation problems the cost is dependent on the total flow and vehicle route
choices, and it is not reasonable to assume fixed costs. However, even in such
network problems the most practical solution methods involve solving several
shortest path problems as part of an iterative framework, updating paths as
travel times change. The context for a shortest path algorithm is to find the
least travel-time path between two nodes, at the current travel times — with ev-
32 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

3
2
4

1
1 4
5
2
2

Figure 2.10: Example network for shortest paths (link costs indicated).

erybody’s route choices held fixed. By separating the process of identifying the
shortest path from the process of shifting path flows toward the shortest path,
we simplify the problem and end up with something which is relatively easy to
solve. This problem can also be phrased in the language of optimization, as
shown in Appendix B.4, by identifying an objective function, decision variables,
and constraints. Here, we develop specialized algorithms for the shortest path
problem which are more efficient and more direct than general optimization
techniques.
Although the shortest path problem most obviously fits into transportation
networks, many other applications also exist in construction management, ge-
ometric design, operations research, and many other areas. For instance, the
fastest way to solve a Rubik’s Cube from a given position can be solved using
a shortest path algorithm, as can the fewest number of links needed to connect
an actor to Kevin Bacon when playing Six Degrees of Separation.
A curious fact of shortest path algorithms is that finding the shortest path
from a single origin to every other node is only slightly harder than finding the
shortest path from that origin to a single destination. This also happens to be
the reason why we can find shortest paths without having to list all of the zillions
of possible paths from an origin to a destination, and adding up their paths.
This common reason is Bellman’s Principle, which states that any segment of a
shortest path must itself be a shortest path between its endpoints. For instance,
consider the network in Figure 2.10, where the costs cij are printed next to each
link. The shortest path from node 1 to node 4 is [1, 2, 3, 4]. Bellman’s Principle
requires that [1, 2, 3] also be a shortest path from nodes 1 to 3, and that [2, 3, 4]
be a shortest path from nodes 2 to 4. It further requires that [1, 2] be a shortest
path from node 1 to node 2, [2, 3] be a shortest path from node 2 to node 3,
and [3, 4] be a shortest path from node 3 to node 4. You should verify that this
is true with the given link costs.
To see why this must be the case, assume that Bellman’s Principle was
violated. If the cost on link (1, 3) was reduced to 2, then [1, 2, 3] is no longer a
shortest path from node 1 to node 3 (that path has a cost of 3, while the single-
2.4. SHORTEST PATHS 33

r i
g

Figure 2.11: Illustration of Bellman’s Principle (dashed lines indicated shortest

paths between two nodes).

link path [1, 3] has cost 2). Bellman’s Principle then implies that [1, 2, 3, 4] is
no longer the shortest path between nodes 1 and 4. Why? The first part of the
path can be replaced by [1, 3] (the new shortest path between 1 and 3), reducing
the cost of the path from 1 to 4: [1, 3, 4] now has a cost of 4. In general, if a
segment of a path does not form the shortest path between two nodes, we can
replace it with the shortest path, and thus reduce the cost of the entire path.
Thus, the shortest path must satisfy Bellman’s Principle for all of its segments.
The implication of this is that we can construct shortest paths one node at a
time, proceeding inductively. Let’s say we want to find the shortest path from
node r to a node i, and furthermore let’s assume that we’ve already found the
shortest paths from r to every node which is directly upstream of i (nodes f , g,
and h in Figure 2.11). The shortest path from r to i must pass through either f ,
g, or h; and according to Bellman’s Principle, the shortest path from r to i must
be either (a) the shortest path from r to f , plus link (f, i); the shortest path
from r to g, plus link (g, i); or the shortest path from r to h, plus link (h, i). This
is efficient because, rather than considering all of the possible paths from r to i,
we only have to consider three, which can be easily compared. Furthermore, we
can re-use the information we found when finding shortest paths to f , g, and h,
and don’t have to duplicate the same work when finding the shortest path to i.
This idea doesn’t give a complete algorithm yet — how did we find the shortest
paths to f , g, and h, for instance? — but gives the flavor of the shortest path
algorithms presented next.
Bellman’s Principle also gives us a compact way of expressing all of the
shortest paths from an origin to every other node in the network: for each node,
simply indicate the last node in the shortest path from that origin. This is
called the backnode vector qr , where each component qir is the node immediately
preceding i in the shortest path from r to i. If i = r, then qir is not well-defined
(i is the origin itself; what is the shortest path from the origin to itself, and
if we can define it, what node immediately precedes the origin?) so we say
qrr ≡ −1 by definition. For the network in Figure 2.10 (with the original costs),
34 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

1 1 1 1
 have q1 = −1, q2 = 1, q3 = 2, and q4 = 3, or, in vector notation,
we thus
1
q = −1 1 2 3 .
The backnode vector can be used as follows: say we want to look up the
shortest path from node 1 to node 4. Starting at the destination, the backnode
of 4 is 3, which means “the shortest path to node 4 is the shortest path to
node 3, plus the link (3, 4).” To find the shortest path to node 3, consult its
backnode: “the shortest path to node 3 is the shortest path to node 2, plus the
link (2, 3).” For the shortest path to node 2, its backnode says: “the shortest
path to node 2 is from node 1.” This is the origin, so we’ve found the start
of the path, and can reconstruct the original path to node 4: [1, 2, 3, 4]. More
briefly, we can use the backnodes to trace the shortest path back to an origin,
by starting from the destination, and reading back one node at a time.
We will also define Lri to be the total cost on the shortest path from origin r
to node i (the letter L is used because these values are often referred to as node
labels), with Lrr ≡ 0, so in this example we would have L11 = 0, L12 = 2, L13 = 3,
and L14 = 5.
This chapter presents four shortest path algorithms. The first, in Sec-
tion 2.4.1, only applies when the network is acyclic but is extremely fast and
simple. Sections 2.4.2 and 2.4.3 next present the two most common approaches
for solving shortest paths in networks with cycles: label setting and label correct-
ing. Dijkstra’s algorithm and the Bellman-Ford algorithm are the quintessential
label setting and label correcting algorithms, respectively, and they are used to
show the difference between these approaches. These three algorithms actually
find the shortest path from the origin r to every other node, not just the des-
tination s; this exploits Bellman’s Principle, because the shortest path from r
to s must also contain the shortest path from r to every node in that path.
Further, both of them contain a set of labels Lri associated with each node,
representing the shortest path from r to node i. With a label setting approach,
each node’s label is determined once and only once. On the other hand, with a
label correcting approach, each node’s label can be updated multiple times.
The fourth algorithm, A∗ is presented in Section 2.4.4 . This algorithm can
be faster than Dijkstra’s algorithm if we are only interested in the shortest path
from a single origin to a single destination.

2.4.1 Shortest paths on acyclic networks

Recall from Section 2.2 that an acyclic network is one in which no cyclic paths
exist; it is impossible to visit any node more than once on any path. Conversely,
a cyclic network is one where a cycle does exist, and where (in theory) one could
repeatedly drive around in a circle forever. Transportation networks are usually
cyclic, and in fact any network where you can reach every node from every
other node must be cyclic. (Why?) However, we’ll take a short diversion into
acyclic networks for the purposes of finding shortest paths, for two reasons: (1)
finding the shortest path on an acyclic network is much simpler and faster, and
makes for a good first illustration; and (2) more advanced solution methods take
advantage of the fact that people do not use cyclic paths (because of the shortest
2.4. SHORTEST PATHS 35

path assumption), so we can only look at an acyclic portion of the network and
thereby use the much faster shortest path algorithm for acyclic networks.
Once we have a topological order on a network, it becomes very easy to
find the shortest path from any node r to any other node s (clearly r has lower
topological order than s), using the following algorithm which is based directly
on Bellman’s Principle:

1. Initialize by setting Lri = ∞ and qir = −1 for all nodes i ∈ N to indicate

that we have not found any shortest paths yet. Then set Lrr = 0, because
the distance from the origin to itself is zero.

2. Let i be the node topologically following r.

3. We can find the shortest path from r to i using Bellman’s Principle, using
these formulas:
Lri = min {Lrh + chi } (2.1)
(h,i)∈Γ−1 (i)

qir ∈ arg min {Lrh + chi } (2.2)

(h,i)∈Γ−1 (i)

(Essentially, these formulas have us search all of the links arriving at node
i for the approach with minimum cost.)

4. Does i = s? If so, stop, and we have found the shortest path from r to
s, which has cost Lrs . Otherwise, let i be the next node topologically, and
return to step 3.

Bellman’s Principle lets us find all of the shortest paths in one pass over the
nodes (in topological order), because there are no cycles which could make us
loop back.
Here’s how to apply this algorithm to the network in Figure 2.10, with node
1 as origin and node 4 as destination.
   
Step 1: Initialize: L1 ← 0 ∞ ∞ ∞ and q1 ← −1 −1 −1 −1 .

Step 2: We set j ← 2.

Step 3: The only approach to node 2 is (1, 2), so Γ−1 (2) is just {(1, 2)} and
the minimization is trivial: L12 ← L11 + c12 = 0 + 2 = 2 and q21 ← 1.

Step 4: 2 6= 4, so we advance j to the next node topologically: j ← 3 and

return to step 3.

Step 3: There are two approaches to node 3, (1, 3) and (2, 3). So L13 ←
min{L11 + c13 , L12 + c23 } = min{0 + 4, 2 + 1} = 3, which corresponds to
approach (2, 3), so q31 ← 2.

Step 4: 3 6= 4, so we advance j to the next node topologically: j ← 4 and

return to step 3.
36 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

Step 3: There are two approaches to node 4, (2, 4) and (3, 4). So L14 ←
min{L12 + c24 , L13 + c34 } = min{2 + 5, 3 + 2} = 5, which corresponds to
approach (3, 4), so q41 ← 3.
Step 4: We have reached the destination (node 4) and terminate.
A simple induction proof shows that this algorithm must give the correct
shortest paths to each node (except for those topologically before r, since such
paths do not exist). By “give the correct shortest paths,” we mean that the
labels L give the lowest cost possible to each node when traveling from r, and
that the backnodes q yield shortest paths when traced back to r. Assume
that the nodes are numbered in topological order. Clearly Lrr and qrr are set
to the correct values in the first step (the shortest path from r to itself is
trivial), and are never changed again because the network proceeds in increasing
topological order. Now assume that the algorithm has found correct L and q
values for all nodes between r and k. In the next step, it updates labels for
node k + 1. Let (i, k + 1) be the last link in a shortest path from r to k + 1. By
Bellman’s principle, the first part of this path must be a shortest path from r to
i. Since i is topologically between r and k, by the induction hypothesis the Lr
and q r labels are correct for all nodes immediately upstream of k. Therefore,
Lri +ci,k+1 ≤ Lrj +cj,k+1 for all immediately upstream nodes j, and the algorithm
makes the correct choices in step 3 for node k + 1.

2.4.2 Shortest paths on cyclic networks: label correcting

When there are cycles in the network, this previous approach can’t be ap-
plied, because there is no clear sequence in which to examine nodes and apply
Bellman’s Principle. However, we can generalize the approach. Rather than
scanning nodes in a rigid order, we can fan out from the origin, keeping in mind
that we may have to scan a node multiple times in case of a cycle. To keep
track of the nodes we need to examine, we define a scan eligible list SEL, a set
of nodes that we still need to examine before we have found all of the short-
est paths. This is a label correcting approach, because nodes may be scanned
multiple times, and the labels updated.

1. Initialize every label Lri to ∞, except for the origin, where Lrr ← 0, and
the backnode vector qr ← −1.
2. Initialize the scan eligible list to contain all nodes immediately downstream
of the origin, that is, SEL ← {i : (r, i) ∈ Γ(r)}.
3. Choose a node i ∈ SEL and remove it from that list.
4. Scan node i by applying the same equations as in the previous algorithm:

Lri = min {Lrh + chi } (2.3)

(h,i)∈Γ−1 (i)

qir = arg min {Lrh + chi } (2.4)

(h,i)∈Γ−1 (i)
2.4. SHORTEST PATHS 37

5. If the previous step changed the value of Lri , then add all nodes immedi-
ately downstream of i to the scan eligible list, that is, SEL ← SEL ∪ {j :
(i, j) ∈ Γ(i)}.

6. If SEL is empty, then terminate. Otherwise, return to step 3.

Repeating the same example, this algorithm works as follows:

   
Step 1: We set L1 ← 0 ∞ ∞ ∞ and q1 ← −1 −1 −1 −1 .

Step 2: We set SEL ← {2, 3}.

Step 3: Choose node 3 and remove it from SEL, so i ← 3, and SEL = {2}.

Step 4: The two possible approaches are (1, 3) and (2, 3), so L13 ← min{L11 +
c13 , L12 + c23 } = min{0 + 4, ∞ + 1} = 4, which corresponds to approach
(1, 3), so q31 ← 1. Note that, unlike in the case of the acyclic network, we
are not claiming that this is the shortest path to node 3. We are simply
claiming that this is the shortest path we have found thus far.

Step 5: The previous step reduced L13 from ∞ to 4, so we add the downstream
node 4 to SEL, so SEL = {2, 4}.

Step 6: SEL is nonempty, so we return to step 3.

Step 3: Choose node 4 and remove it from SEL, so i ← 4, and SEL = {2}.

Step 4: There are two approaches to node 4, (2, 4) and (3, 4). So L14 ←
min{L12 + c24 , L13 + c34 } = min{∞ + 5, 4 + 2} = 6, which corresponds to
approach (3, 4), so q41 ← 3.

Step 5: The previous step reduced L14 from ∞ to 6, but there are no down-
stream nodes to add, so SEL = {2}.

Step 6: SEL is nonempty, so we return to step 3.

Step 3: Choose node 2 and remove it from SEL, so i ← 2, and SEL = ∅.

Step 4: The only approach to node 2 is (1, 2), so {(h, i) ∈ Γ−1 (i)} is just
{(1, 2)} and the minimization is trivial: L12 ← L11 + c12 = 0 + 2 = 2 and
q21 ← 1.

Step 5: The previous step reduced L12 from ∞ to 2, so we add the downstream
nodes 3 and 4 to SEL, so SEL = {3, 4}.

Step 6: SEL is nonempty, so we return to step 3.

Step 3: Choose node 3 and remove it from SEL, so i ← 3, and SEL = {4}.
Note that this is the second time we are scanning node 3. We have to do
this, because we found a path to node 2, and there is a possibility that
this could lead to a shorter path to node 3 as well.
38 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

Step 4: The two possible approaches are (1, 3) and (2, 3), so L13 ← min{L11 +
c13 , L12 + c23 } = min{0 + 4, 2 + 1} = 3, which corresponds to approach
(2,3), so q31 ← 2. Note that we have changed the labels again, showing
that we have indeed found a better path through node 2.

Step 5: The previous step reduced L13 from 4 to 3, so we add the downstream
node 4 to SEL, so SEL = {4}.

Step 6: SEL is nonempty, so we return to step 3.

Step 3: Choose node 4 and remove it from SEL, so i ← 4, and SEL = ∅.

Step 4: There are two approaches to node 4, (2, 4) and (3, 4). So L14 ←
min{L12 + c24 , L13 + c34 } = min{2 + 5, 3 + 2} = 5, which corresponds to
approach (3, 4), so q41 ← 3.

Step 5: The previous step reduced L14 from 6 to 5, but there are no down-
stream nodes to add, so SEL remains empty.

Step 6: SEL is empty, so we terminate.

As you can see, this method required more steps than the algorithm for
acyclic networks (because there is a possibility of revisiting nodes), but it does
not rely on having a topological order and can work in any network. One can
show that the algorithm converges no matter how you choose the node from
SEL, but it is easier to prove if you choose a systematic rule. If you operate
SEL in a “first-in, first-out” manner, always choosing a node which has been
in the queue for the greatest number of iterations, it is possible to show that
after m iterations, you have certainly found the shortest paths from r which
consist of only a single link. (You’ve probably found quite a few more shortest
paths, but even in the worst case you’ll have found at least these.) After 2m
iterations, you will have certainly found the shortest paths from r which consist
of one or two links only, and so on. So, after mn iterations, you will have found
all of the shortest paths, since a shortest path cannot use more than n links.
The exercises ask you to fill in the details of this proof sketch.

2.4.3 Shortest paths on general networks: label setting

In the label-correcting method, we accounted for cycles by allowing nodes to
be scanned more than once, relieving us of having to scan nodes in the “right”
order when there is no topological order to guide us. A label-setting method
works in cyclic networks, yet only has to scan each node once. This requires a
bit more care to scan nodes in the proper order, since there is no opportunity
to revisit nodes to correct their labels later on. So, while label-setting methods
only require one scan per node, determining which node to scan requires more
effort at each iteration.
The quintessential label setting algorithm was developed by Edsger Dijkstra.
Dijkstra’s algorithm finds the shortest path from the origin r to every other node,
2.4. SHORTEST PATHS 39

when every link has a nonnegative cost. Furthermore, at each step it finds the
shortest path to at least one additional node. It uses the concept of finalized
nodes, that is, nodes to which the shortest path has already been found. It
uses the same L and q labels as before. The primary distinction between label
setting and label correcting is that label setting requires fewer iterations than
label correcting, but each iteration takes a greater amount of effort. Which
one is superior depends on the network topology, the specific implementation of
each algorithm, and the skill of the programmer.
Dijkstra’s algorithm can be stated as follows:
1. Initialize every label Lri to ∞, except for the origin, where Lrr ← 0.
2. Initialize the set of finalized nodes F = ∅ and the backnode vector qr ←
−1.
3. Find an unfinalized node i (i.e., not in F ) for which Lri is minimal.
4. Finalize node i by adding it to set F ; if all nodes are finalized (F = N )
then terminate.
5. Update labels for links leaving node i: for each (i, j) ∈ A, update Lj =
min{Lrj , Lri + cij }.
6. Update backnodes: for each (i, j) ∈ A, if Lrj was reduced in the previous
step set qjr ← i.
7. If all nodes are finalized, stop. Otherwise, return to step 3
This algorithm is considered label setting, because once a node’s label is
updated, the node is finalized and never revisited again. We consider how this
algorithm can be applied to the same example used for the other algorithm
demonstrations.
   
Initialization. We initialize L1 = 0 ∞ ∞ ∞ , F = ∅, and q1 = −1 −1 −1 −1 .
Iteration 1. The unfinalized node with least L value is the origin, so we
set i = 1 and finalize it: F = {1}. We update the downstream labels:
L12 = min{∞, 0 + 2} = 2 and L13 = 4. Both labels were reduced, so we
update the backnodes: q21 = q31 = 1.
Iteration 2. Of the unfinalized nodes, i = 2 has the lowest label, so we finalize
the node: F = {1, 2}. We update the downstream labels: L13 = min{4, 2 +
1} = 3 and L14 = 7. Both labels were reduced, so we update the backnodes:
q31 = q41 = 2.
Iteration 3. Of the unfinalized nodes, i = 3 has the lowest label, so we finalize
it: F = {1, 2, 3}. We update the downstream label L14 = min{7, 3 + 2} = 5
and backnode q41 = 3.
Iteration 4. Node 4 is the only unfinalized node, so i = 4. We finalize it, and
since F = N we are done.
40 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

The correctness of Dijkstra’s algorithm can be proved by an induction argu-

ment, on the size of the set of finalized nodes |F |. Specifically, we show that at
every step of the algorithm the backnodes to every finalized node represent a
shortest path. The algorithm always finalizes the origin r first, so when |F | = 1
the only finalized node is the origin, and the shortest path to this node is trivial.
Now assume that when |F | = k, the algorithm has correctly found shortest
paths to every finalized node, and consider what happens when the k + 1-th
node is added to this set. This occurs in step 4 of the algorithm. Call this
newly finalized node j, and say that at this point the backnode is qjr = i. By
contradiction, assume that there is a shorter path from r to j than the one
through node i. Notice that the entries of the backnode vector which are not
−1 all point to finalized nodes, because step 6 is the only step which changes
this vector, and only changes entries of this vector to a newly finalized node.
Furthermore, from the induction hypothesis, the backnode vectors from all of
these previously-finalized nodes describe shortest paths.
So, if there is a shorter path to j than the one through i, it must differ
in the last link, that is, this shorter path must end with a link (h, j) where
h 6= i. Either h is a finalized node, or it is not. If h were finalized, then by the
induction hypothesis Lrh is the correct shortest path cost to r, and since it leads
to a shorter path to j we must have Lrh +chj < Lri +cij . But if this were the case,
then steps 5 and 6 would have set qjr to h when node h was finalized, so qjr could
not equal i, a contradiction. Or, if h were not finalized, then Lrh > Lrj , since in
step 3 we always finalize a node with the smallest L value. Since all links have
nonnegative costs, we cannot have Lrh + chj < Lrj , which is what we need if h
led to a shorter path to j. Since both cases lead to contradictions, we conclude
that a correct shortest path to node j has been found, and by induction, this
argument extends to every node.

2.4.4 Shortest paths from one origin to one destination

The previous sections gave algorithms to find the shortest path from one origin
to all destinations. Slight modifications to these algorithms can find the shortest
path from all origins to one destination (see Exercise 47). As discussed above,
Bellman’s principle lets us re-use information from one shortest path when find-
ing another, and as a result even finding a shortest path from one origin to one
destination can provide many other shortest paths “for free.” In traffic assign-
ment, where we are modeling the flows of many drivers, it often makes sense to
use an algorithm that finds many shortest paths simultaneously.
However, there are times when we are only concerned with a single origin and
destination, and do not care about the “free” shortest paths to other destinations
we get. Examples are in-vehicle routing systems, or if the number of origin-
destination pairs with positive demand is small compared to the overall size of
the matrix (i.e., the OD matrix is sparse) so there are relatively few paths we
need to find. In such cases we can use a more focused algorithm to find a single
shortest path in less time than it takes to find shortest paths to all destinations.
The A∗ algorithm is a simple modification to label-setting, which in some
2.4. SHORTEST PATHS 41

cases can dramatically reduce the running time, in exchange for limiting the
scope to one origin r and one destination s. This algorithm requires an addi-
tional value gis for each node, representing an estimate of the cost of the shortest
path from i to s. This estimate (often called the “heuristic”) should be a lower
bound on the actual cost of the shortest path from i to s. Some examples of
how these estimates are chosen are discussed below.
Once the estimates gis are chosen, the label-setting algorithm from Sec-
tion 2.4.3 proceeds as before with a small modification in Step 3: rather than
choosing a node i in E which minimizes Lri , we choose a node which minimizes
Lri + gis . Everything else proceeds exactly as before. The intuition is that we
are taking into account our estimates of which nodes are closer to the destina-
tion. The algorithm in Section 2.4.3 fans out in all directions from the origin
(by simply looking at Lri ), rather than directing the search towards a particular
destination.
It can be shown that this algorithm will always yield the correct shortest
path from r to s as long as the gis are lower bounds on actual shortest path costs
from i to s. If this is not the case, A∗ is not guaranteed to find the shortest
path from r to s. Some care must be taken in how these estimates are found.
Two extreme examples are:
ˆ Choose gis = 0 for all i. This is certainly a valid lower bound on the
shortest path costs (recall that label-setting methods assume nonnegative
link costs), so A∗ will find the shortest path from r to s. However, zero is
a very poor estimate of the actual shortest path costs. With this choice
of gis , A∗ will run exactly the same as Dijkstra’s algorithm, and there is
no time savings.
ˆ Choose gis to be the actual shortest path cost from i to s. This is the
tightest possible “lower bound,” and will make A∗ run extremely quickly
— in fact, it will only scan nodes along the shortest path, the best pos-
sible performance that can be achieved. However, coming up with these
estimates is just as hard as solving the original problem! So in the end we
aren’t saving any effort; what we gain from A∗ is more than lost by the
extra effort we need to compute gis in the first place.
So, there is a tradeoff between choosing tight bounds (the closer gis to the true
costs, the faster A∗ will be) while not spending too long in computing the esti-
mates (which might swamp the savings in A∗ itself). Luckily, in transportation
networks, there are several good bounds available which can be computed fairly
quickly. For instance:
ˆ The Euclidean (“as the crow flies”) distance between i and s, divided by
the fastest travel speed in the network, is a lower bound on the travel time
between i and s.
ˆ Replace every link with a lower bound on its cost (say, free-flow travel
time) and find shortest paths between all nodes and all destinations (re-
peatedly using one of the previous algorithms from this chapter). This
42 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

takes more time, but only needs to be done once and can be done as a
preprocessing step. As we will see in Chapter 6, solving traffic assignment
requires many shortest path computations. The extra time spent finding
these costs once might result in total time savings over many iterations.

You may find it instructive to think about other ways we can estimate gis values,
and how they might be used in transportation settings.

2.5 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
Networks arise frequently in transportation and optimization, particularly
network flow problems such as shortest path, maximum flow, and minimum
cost flow. The text by (Ahuja et al., 1993) discusses a number of ways to
formulate and solve such problems, along with a breadth of applications. This
text also provides more detail on data structures used for working with networks
in computer programs. Bellman’s principle was stated in Bellman (1957).
The first label-correcting shortest path algorithm was developed by Ford
(1956). Subsequent experience has shown that the practical performance of this
method depends significantly on the order in which nodes are retrieved from
the scan list; good choices are “first-in, first-out,” (FIFO) in which the node
selected from SEL is always one of the oldest there, and a double-ended queue
implementation which provides an exception to FIFO in that when nodes are
return to the scan list after being scanned earlier they are moved to the front
of the queue. These choices were first described in Bellman (1958) and Pape
(1974).
The quintessential label-setting shortest path method is that of Dijkstra
(1959). The A∗ method is essentially an extension to Dijkstra’s algorithm, and
was developed by Hart et al. (1968). For more thorough reviews of shortest
path algorithms, the reader is referred to the survey paper by Deo and Pang
(1984), and to Ahuja et al. (1993).
There are many variants of the shortest path program, for instance, adding
side constraints on the amount of some other resource consumed (such as money
spent on tolls or battery charge from an electric vehicle). Ziegelmann (2001)
provides a review of approaches for this problem.
Another common variant is aimed at addressing reliability and uncertainty in
link costs, by having the link costs be drawn from some probability distribution
rather than known exactly in advance. If one aims to minimize the expected
travel time, it is enough to replace each link cost with its expected value. How-
ever, travelers may also care about reliability specifically. This can be modeled
by adding variance or standard deviation to the objective function (Xing and
Zhou, 2011; Shahabi et al., 2013; Khani and Boyles, 2015; Zhang et al., 2016),
adding a constraint on variance (Sivakumar and Batta, 1994), adopting a robust
optimization perspective (Yu and Yang, 1998; Montemanni and Gambardella,
2.6. EXERCISES 43

1 2 3

4 5 6

7 8

Figure 2.12: Network for Exercises 1–3

2004; Shahabi et al., 2015), applying multiobjective optimization (Sen et al.,

2001), introducing a nonlinear utility function to represent arrival time prefer-
ences (Loui, 1983; Eiger et al., 1985; Murthy and Sarkar, 1996; Boyles, 2006;
Gao, 2005), or changing the objective entirely to maximize the probability of
on-time arrival (Fan et al., 2005; Nie and Wu, 2009).
Yet another way to address reliability and real-time information provision
is to allow the traveler’s path to change en route based on information learned
while traveling; this leads to a class of online shortest path formulations. For
examples of such formulations, see Andreatta and Romeo (1988), Psaraftis and
Tsitsiklis (1993), Polychronopoulos and Tsitsiklis (1996), Miller-Hooks (2001),
Waller and Ziliaskopoulos (2002), Provan (2003), Gao (2005), Boyles (2006),
Boyles and Rambha (2016), and Boyles (2009).

2.6 Exercises
1. [9] In the network in Figure 2.12, list the indegree, outdegree, degree,
forward star, and reverse star of each node.
2. [3] In the network in Figure 2.12, list all of the paths between nodes 1
and 5.
3. [4] State whether the network in Figure 2.12 is or is not (a) cyclic; (b) a
tree; (c) connected; (d) strongly connected.
4. [15] For each of the following, either draw a network with the stated
properties, or explain why no such network can exist: (a) connected, but
not strongly connected; (b) strongly connected, but not connected; (c)
cyclic, but not strongly connected.
5. [25] If m and n are the number of links and nodes in a network, show that
m < n2 .
6. [25] If a network is connected, show that m ≥ n − 1.
7. [25] If a network is strongly connected, show that m ≥ n.
44 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

Figure 2.13: Network for Exercise 11

|Γ−1 (i)| = m.
P P
8. [10] Show that i∈N |Γ(i)| = i∈N

9. [1] Why must transportation infrastructure networks be cyclic?

10. [15] Although we do not usually expect drivers to use cyclic paths, there
are some exceptions. Name one.
11. [15] Find a topological order for the network in Figure 2.13.
12. [11] Is the topological order for an acyclic network unique? Either explain
why it is, or provide a counterexample showing it is not.
13. [42] Show that any acyclic network has at least one node with no incoming
links. This was the “missing step” in the proof of Theorem 2.1. (Since
this is part of the proof, you can’t use the result of this theorem in your
answer.)
14. [36] Let i and j be any two nodes in an acyclic network. Give a procedure
for calculating the number of paths between i and j which involves at
most one calculation per node in the network.
15. [57] Give a procedure for determining whether a given network is strongly
connected. Try to make your method require as few steps as possible.
16. [58] Give a procedure for determining whether a given network is con-
nected. Try to make your method require as few steps as possible.
17. [30] Show that there is at most one path between any two nodes in a tree.
18. [30] Show that any node in a tree (excluding its root) has exactly one link
in its reverse star (and thus one parent).
19. [30] Show that any tree is an acyclic network.
20. [30] Show that removing any link from a tree produces a disconnected
network.
21. [50] Show that any tree has at least two nodes of degree one.
2.6. EXERCISES 45

22. [55] Consider a rectangular grid network of one-way streets, with r rows
of nodes and c columns of nodes. All links are directed northbound and
eastbound. How many paths exist between the lower-left node (southwest)
and the upper-right (northeast) node?

23. [14] Write down the node-node adjacency matrix of the network in Fig-
ure 2.12.

24. [16] Consider the network defined by this node-node adjacency matrix:
 
0 0 0 1 1 0 0
1 0 1 0 0 0 0
 
0 0 0 1 0 0 0
 
0 0 0 0 1 0 0
 
0 0 0 0 0 0 0
 
1 1 0 0 0 0 0
0 1 0 0 0 1 0

(a) Draw the network.

(b) How many links does the network have?
(c) How many links enter node 1? How many links leave node 1?

25. [17] Is the network represented by the following node-node adjacency ma-
trix strongly connected?
 
0 0 0 1 1 0 0
1 0 1 0 0 0 0
 
0 0 0 1 0 0 0
 
0 0 0 0 1 0 0
 
0 0 0 0 0 1 0
 
1 1 0 0 0 0 0
0 1 0 0 0 1 0

26. [10] If the nodes in an acyclic network are numbered in a topological order,
show that the node-node adjacency matrix is upper triangular.

27. [37] Let A be the node-node adjacency matrix for a network. What is the
interpretation of the matrix product A2 ?

28. [42] Let A be the

P∞node-node adjacency matrix for an acyclic network.
First show that n=1 An exists, and give an interpretation of this sum.

29. [65] A unimodular matrix is a square matrix whose elements are integers
and whose determinant is either +1 or −1. A matrix is totally unimodular
if every nonsingular square submatrix is unimodular. (Note that a to-
tally unimodular matrix need not be square). Show that every node-link
incidence matrix is totally unimodular.
46 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

30. [10] One disadvantage of the forward star representation is that it is time-
consuming to identify the reverse star of a node — one must search through
the entire array to find every link with a particular head node. Describe a
“reverse star” data structure using arrays, where the reverse star can be
easily identified.

31. [52] By combining the forward star representation from the text and the
reverse star representation from the previous exercise, we can quickly iden-
tify both the forward and reverse stars of every node. However, a naive
implementation will have two different sets of arrays, one sorted accord-
ing to the forward star representation, and the other sorted according to
the reverse star representation. This duplication wastes space, especially
if there are many attributes associated with each link (travel time, cost,
capacity, etc.) Identify a way to easily identify the forward and reverse
stars of every node, with only one set of arrays of link data, by adding an
appropriate attribute to each link.

32. [58] In the language of your choice, write computer code to do the follow-
ing:

(a) Produce the node-node adjacency matrix of a network when given

the node-link incidence matrix.
(b) Produce the node-link incidence matrix of a network when given the
node-node adjacency matrix.
(c) Produce the node-node adjacency matrix of a network when given
the forward star representation of the network.
(d) Produce the forward star representation of a network when given the
node-node adjacency matrix.

33. [13] After solving a shortest path problem from node 3 to every other
node, I obtain the backnode vector shown in Table 2.3. Write the shortest
paths (a) from node 3 to node 5; (b) from node 3 to node 7; (c) from node
4 to node 8.

34. [26] Find the shortest path from node 1 to every other node in the network
shown in Figure 2.14. Report the final labels and backnodes (L and q
values) for all nodes.

35. [25] The network in Figure 2.15 has a link with a negative cost. Show that
the label-correcting algorithm still produces the correct shortest paths in
this network, while the label-setting algorithm does not.

36. [57] Prove or disprove the following statement: “Any network with neg-
ative costs can be transformed into a network with nonnegative costs by
adding a large enough constant to every link’s cost. We can then use
the label-setting algorithm on this new network. Therefore, the label-
setting algorithm can find the shortest paths on any network, even if it
2.6. EXERCISES 47

Table 2.3: Backnode vector for Exercise 33.

Node Backnode
1 4
2 6
3 −1
4 7
5 4
6 5
7 3
8 10
9 5
10 6

3
2
4

3 1
1 4
5
2
2

Figure 2.14: Network for Exercises 34 and 41.

3
1
2

1 -2 4
4
3
2

Figure 2.15: Network for Exercise 35.

48 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

7 7
11
5
3 10 1 8
4
8 4 5
6
6
2
1 3 9

Figure 2.16: Network for Exercise 37, link labels are costs.

has negative-cost links.” (Proving this statement means showing that it

is true in any network; to disprove it, it is sufficient to find a single coun-
terexample.)
37. [37]. Find the shortest paths on the network in Figure 2.16, using both the
label-correcting and label-setting algorithms. Sketch the shortest path tree
(that is, the network with only the links implied by the backnode vector)
produced by each algorithm. Which algorithm required fewer iterations?
38. [33]. You and your friends are camping at Yellowstone Park, when you
suddenly realize you have to be at the Denver airport in ten hours to
catch a flight. You don’t have Internet access, but you do have an atlas
showing the distance and travel time between selected cities (Figure 2.17).
Assuming these travel times are accurate, can you make it to Denver in
time for your flight?
39. [43]. Consider the following variation of the shortest path problem: in-
stead of the costs cij being fixed, insteadassume that they are random,
and the cost of link (i, j) takes the values c1ij , c2ij , . . . , csij with probabil-

ities p1ij , p2ij , . . . , psij independent of the cost of any other link. How can
you adapt the shortest path algorithms in this chapter to find the path
with the least expected cost?
40. [53]. Assume that each link (i, j) in the network fails with probability pij ,
and that link failures are independent. (Failure can represent some kind
of damage or disruption, the probability of being detected in a military
routing problem, etc.) A path fails if any link in that path fails. Explain
how you can find the most reliable path (that is, the one with the least
failure probability) with a shortest path algorithm.
41. [23] Instead of trying to find the shortest path between two nodes, let’s
try to find the longest path between two nodes. (As we will see later,
2.6. EXERCISES 49

Figure 2.17: Atlas page for Exercise 38.

there are actually cases when this is useful.)

(a) Modify the algorithm presented in Section 2.4.1 to find the longest
path in an acyclic network.
(b) If you modify the label-correcting or label-setting algorithms for gen-
eral networks in a similar way, will they find the longest paths suc-
cessfully? Try them on the network in Figure 2.14.

42. [25]. As a modification to the node-node adjacency matrix you might

use to represent the network in a computer program (cf. Section 2.3), you
could store the cost of the links in this matrix, with ∞ where no link exists
(as opposed to ‘1’ where links exist and ‘0’ where they do not). Find the
shortest path between nodes 7 and 4 using the modified adjacency matrix
below, where the entry in row i and column j is the cost cij if this link
exists, and ∞ if it does not.
 
∞ ∞ ∞ 5 3 ∞ ∞
 4 ∞ 7 ∞ ∞ ∞ ∞
 
∞ ∞ ∞ 6 ∞ ∞ ∞
 
∞ ∞ ∞ ∞ 2 ∞ ∞
 
∞ ∞ ∞ ∞ ∞ 4 ∞
 
3 2 ∞ ∞ ∞ ∞ ∞
∞ 1 ∞ ∞ ∞ 4 ∞

43. [22]. In the game “Six Degrees of Kevin Bacon,” players are given the
name of an actor or actress, and try to connect them to Kevin Bacon in
as few steps as possible, winning if they can make the connection in six
50 CHAPTER 2. NETWORK REPRESENTATIONS AND ALGORITHMS

steps or less. Two actors or actresses are “connected” if they were in the
same film together. For example, Alfred Hitchcock is connected to Kevin
Bacon in three steps: Hitchcock was in Show Business at War with Orson
Welles, who was in A Safe Place with Jack Nicholson, who was in A Few
Good Men with Kevin Bacon. Beyoncé Knowles is connected to Kevin
Bacon in two steps, since she was in Austin Powers: Goldmember, where
Tom Cruise had a cameo, and Cruise was in A Few Good Men with Bacon.
Assuming that you have total, encyclopedic knowledge of celebrities and
films, show how you can solve “Six Degrees of Kevin Bacon” as a shortest
path problem. Specify what nodes, links, costs, origins, and destinations
represent in the network you construct.

44. [14] In a network with no negative-cost cycles, show that −nC is a lower
bound on the shortest path cost between any two nodes in a network.
45. [57] Show that if the label-correcting algorithm is performed, and that at
each iteration you choose a node in SEL which has been in the list the
longest, at the end of kn iterations the cost and backnode labels correctly
reflect all shortest paths from r which are no more than k links long.
(Hint: try an induction proof.)
46. [44] Assume that the label-correcting algorithm is terminated once the
label for some node i falls below −mC. Show that the following the
current backnode labels from i will lead to a negative-cost cycle.

47. [32] Modify three of the shortest path algorithms in this chapter so that
they find shortest paths from all origins to one destination, rather than
one origin to all destinations:
(a) The acyclic shortest path algorithm from Section 2.4.1.
(b) The label-correcting algorithm from Section 2.4.2.
(c) The label-setting algorithm from Section 2.4.3.
48. [73] It is known that the label correcting algorithm will find the correct
shortest paths as long as the initial labels L correspond to the distance of
some path from the origin (they do not necessarily need to be initialized
to +∞). Assume that we are given a vector of backnode labels q which
represents some tree (not necessarily the shortest paths) rooted at the
origin r. Develop a one-to-all shortest path algorithm that uses this vector
to run more efficiently. In particular, if the given backnode labels q do
correspond to a shortest path tree, your algorithm should recognize this
fact and terminate in a number of steps linear in the number of network
links.
Chapter 3

Mathematical Techniques
for Equilibrium

This chapter provides a survey of mathematical techniques used in network

analysis. There is no attempt to be comprehensive; many indeed, books have
been written about each of the sections in this chapter. Rather, the intent is to
cover topics which are used frequently in transportation network problems.
Several of the appendices may be useful at this point. Appendix A reviews
definitions and facts related to vectors, matrices, sets, and functions which are
needed, including the topics of convex functions, convex sets, and multivariable
calculus (the gradient vector, and the Jacobian and Hessian matrices will play
a particularly important role). If these topics are new to you, you may wish to
consult more extended treatments of these topics in other books or references.
Appendix B introduces basic concepts of optimization, focusing on what will
ultimately be relevant for static and dynamic traffic assignment. Optimization
is a much richer and deeper field than what is used in traffic assignment, and
Appendices C and D go into further detail here — the material from these latter
two appendices is not strictly necessary for this book, but we believe it to be of
interest to many readers nevertheless.
This chapter focuses on mathematical material which is both specialized to
the traffic assignment problem, and which is likely to be new to anticipated
readers of the book. This will involve discussion of three main techniques: the
fixed point problem, in Section 3.1; the variational inequality, in Section 3.2, and
convex optimization, in Section 3.3. Chapter 1 characterized equilibrium as a
“consistent” state in which no driver can improve his or her satisfaction by uni-
laterally changing routes. Under the assumption of continuous flow variables,
we can use calculus to greatly simplify the problem. All three of the mathemat-
ical techniques discussed in this chapter formalizes this equilibrium principle in
different ways.
Each approach has its own advantages from the standpoint of this book. The
fixed point formulation is perhaps the most intuitive and generally-applicable

51
52 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

definition, but does not give much indication as to how one might actually find
this equilibrium. The variational inequality formulation lends itself to physical
intuition and can also accommodate a number of variations on the equilibrium
problem. The convex optimization approach provides an intuitive interpretation
of solution methods, provides an elegant proof of equilibrium uniqueness in
link flows, and powers the best-known solution algorithms, but the connection
between the equilibrium concept and optimization requires more mathematical
explanation and is less obvious at first glance.

3.1 Fixed-Point Problems

A fixed point of a function f is a value x such that f (x) = x, that is, the
value x is unchanged by f . As an example, the function f1 (x) = 2x − 1 has
only one fixed point at x = 1, because f1 (1) = 2 × 1 − 1 = 1, the function
f2 (x) = x2 has two fixed points at 0 and 1 (02 = 0 and 12 = 1), while the
function f3 (x) = x2 + 1 has no fixed points at all. A helpful visual illustration
is that the fixed points of a function are the points of intersection between the
function’s graph and the 45-degree line y = x (Figure 3.1).
Equilibrium solutions can often be formulated as fixed points of a suitable
function. For instance, in the traffic assignment problem, a route choice model
and a congestion model are mutually interdependent: drivers choose routes to
avoid congestion, but congestion is determined by the routes drivers choose.
(Figure 1.3). An equilibrium solution is consistent in the sense that drivers are
satisfied with the travel times calculated by the paths they chose. Feeding the
route choices into the congestion model, then feeding the resulting travel times
into the route choice model, one can obtain the original route choices back.
This is reminiscent of fixed point problems: when you evaluate a function at
its fixed-point, after performing whatever calculations the function requires you
obtain the fixed-point again. Fixed points thus arise naturally when dealing
with these kinds of “circular” dependencies.
As a first example, consider the problem of trying to estimate the number of
bus riders in a dense urban area. The bus system is subject to congestion; when
there are x riders, the average delay to customers is given by the function t =
T (x) which is assumed continuous and increasing. (Assuming that the fleet of
buses and timetables are fixed, more riders mean longer boarding and offloading
times, more frequent stops, and the possibility of full buses skipping waiting
passengers.) However, the number of bus riders depends on the congestion in
the system — as the buses become more crowded, some riders may switch to
alternate modes of transportation or combine trips, so we can write x = X(t) for
some function X which is continuous and decreasing. For a concrete example,
assume that the system is designed such that T (x) = 10 + x, and that the
ridership function is given by X(t) = [20 − t]+ , when x and t are measured in
appropriate units (say, thousands of passengers and minutes).1
1 The notation [·]+ is used to mean the positive part of the term in brackets, that is,

[y]+ = max{y, 0}. If the term in brackets is negative, it is replaced by zero; otherwise it is
3.1. FIXED-POINT PROBLEMS 53

The goal is to find the ridership x; but x = X(t) and t = T (x), which means
that we need to find some value of x such that x = X(T (x)). This is a fixed
point problem! Here the function f is the composition of X and T :

x = X(T (x))
= [20 − T (x)]+
= [20 − (10 + x)]+
= [10 − x]+

Assuming that 10 − x is nonnegative, we can replace the right-hand side of the

last equation by 10 − x. Solving the resulting equation x = 10 − x, we obtain
x = 5 as the fixed point. Checking, when there are 5 riders the average travel
time will be 15 minutes (based on T ); and when the travel time is 15 minutes,
there will indeed be 5 riders (based on X). Finally, if x = 5, our assumption
that 10 − x was nonnegative was true so this solution is valid.
For this example, it is relatively easy to find the fixed point by substituting
the definitions of the functions and performing some algebra. However, in more
complex problems it will be difficult or impossible to calculate the fixed point
directly. Despite this, fixed points are important because of so-called “fixed-
point theorems” which guarantee the existence of a fixed point under certain
conditions of the function f . A fixed point theorem by Brouwer is provided
below, and another by Kakutani is provided in Section 5.1.1. These fixed point
theorems are non-constructive because they give us no indication of how to find
a fixed point, they simply guarantee that at least one exists. Brouwer’s theorem
can be stated as:

Theorem 3.1. (Brouwer). Let f be a continuous function from the set K to

itself, where K is convex and compact. Then there is at least one point x ∈ K
such that f (x) = x.

The exercises ask you to show that each of the conditions in Brouwer’s
theorem is necessary; you might find it helpful to visualize these conditions
geometrically similar to Figure 3.1.
Notice the stipulation that f be a function “from the set K to itself;” this
means that the range of its function must be contained in its domain. Intuitively,
this means that any “output” of the function f must also be a valid “input”
to that same function. In other words, iteration is possible: starting from any
value x in its domain, you can apply the function over and over again to produce
a sequence of values x, f (x), f (f (x)), . . .. If this condition does not hold, then
Brouwer’s theorem does not guarantee anything about a fixed point.
Let us apply Brouwer’s theorem to the transit ridership example. Both
X and T are continuous functions, so their composition X ◦ T is continuous
as well. (Alternately, by substituting one function into the other we obtain
X(T (x)) = [10 − x]+ , which is evidently continuous.) What are the domain
unchanged.
54 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

f(x) x = f(x)

Figure 3.1: Visualizing fixed-point problems; the intersection of f (x) and the
45-degree line is a point where x = f (x).

and range of X(T (x))? Since X(T (x)) is the positive part of 20 − T (x), then
x = X(T (x)) ≥ 0. Further note that because x ≥ 0, T (x) ≥ 10, so x =
X(T (x)) ≤ 10. That is, we have shown that x must lie between 0 and 10, so
the function X(T (x)) = [10 − x]+ can be defined from the set [0, 10] to itself.
This set is convex and compact, so Brouwer’s theorem would have told us that
at least one fixed point must exist, even if we didn’t know how to find it.

3.2 Variational Inequalities

Fixed point problems often lend themselves to elegant theorems like Brouwer’s,
which prove that a fixed point must exist. However, such problems often lack
easy solution methods. The variational inequality can be more convenient to
work with in this respect. Variational inequalities can be motivated with a
physical analogy. Imagine an object, initially stationary, which is confined to
move within some frictionless container (Figure 3.2) and cannot leave. This
object is acted on by a force whose magnitude and direction can be different at
each point. If the object is in the interior of the container, the object will begin
to move in the same direction as the force at that point. If the object starts
at the edge of the container, it may not be able to move in the same direction
as the force, but it might slide along the side of the object. The problem is to
determine where in the container (if anywhere) the object can be placed so that
it will not move under the action of the force field.
Figure 3.3 shows some examples. In this figure, the direction of the force
is drawn with black arrows, shown only at the points under consideration for
clarity. At three of the points (A, B, and C) the object will move: at A in the
direction of the force, and at B and C sliding along the edge of the container
in the general direction of the force. At the other two points (D and E), the
3.2. VARIATIONAL INEQUALITIES 55

Figure 3.2: A container with a force field.

A
B
D

C
E

Figure 3.3: Two solutions, and three non-solutions, to a variational inequality.

object will not move under the action of the force, being effectively resisted by
the container wall.
How can we think about such problems in general? A little thought should
convince you that (1) if the object is on the boundary of the container, but not
at a corner point, it will be unmoved if and only if the force is perpendicular
to the boundary (point D in Figure 3.3), and (2) if the object is at a corner
of the container, it will be unmoved if and only if the force makes a right or
obtuse angle with all of the boundary directions (point E). These two cases can
be combined together: a point on the boundary is an equilibrium if and only if
the force makes a right or obtuse angle with all boundary directions. In fact, if
the force makes such an angle with all boundary directions, it will do so with
any other direction pointing into the feasible set (Figure 3.4). So, we see that
a point is unmoved by the force if and only if the direction of the force at that
point makes a right or obtuse angle with any possible direction the object could
move in.
The mathematical definition of a variational inequality is little more than
translating the above physical problem into algebraic terminology. The “con-
tainer” is replaced by a set K of n-dimensional vectors, which for our purposes
56 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

Figure 3.4: Stable points (in green) make an obtuse or right angle with all
feasible directions; unstable points (in red) make an acute angle with some
feasible directions.

can be assumed compact and convex (as in all of the figures so far). The “force
field” is replaced by a vector-valued function F : Rn → Rn which depends on
n variables and produces an n-dimensional vector as a result. Recalling vector
operations, specifically equation (A.16), saying that two vectors make a right
or obtuse angle is equivalent to saying that their dot product is negative. A
“solution” to the variational inequality is a point which is unmoved by the force
field. (In the above example, we want to say that D and E are solutions to the
variational inequality problem created by the container shape and force field,
while A, B, and C are not.) Therefore, rewriting the condition in the previous
paragraph with this mathematical notation, we have the following definition:
Definition 3.1. Given a convex set K ⊆ Rn and a function F : K → Rn , we
say that the vector x̂ ∈ K solves the VI(K, F) if, for all x ∈ K, we have

F(x̂) · (x − x̂) ≤ 0 . (3.1)

In other words, as x ranges over all possible points in K, x − x̂ represents

all possible directions the object can move from x̂. The point x̂ solves the
variational inequality precisely when the dot product of the force and all possible
directions is negative.
There is a relationship between fixed point problems and variational inequal-
ities, which can be motivated again by the physical analogy of a force acting
within a container. The solutions to the variational inequality are quite literally
“fixed points” in the sense that an object placed there will not move. Con-
sider some point x under the action of the force F. Assume furthermore that
this force is constant and does not change magnitude or direction as this point
moves. Then, since K is convex, the trajectory of the object can be identified
with the curve projK (x + αF(x)) where α ranges over all positive numbers and
projK means projection onto the set K as defined in Section A.3. See Figure 3.5
for a few examples. If a point is a solution to VI(K, F), then the corresponding
“trajectory” will simply be the same point no matter what α is. So, for the sake
3.3. CONVEX OPTIMIZATION 57

Figure 3.5: Trajectories of different points (red) assuming the force (black) is
constant.

of convenience we arbitrarily choose α = 1 and look at the location of the point

projK (x + F(x)). If this is the same as the initial point x, then x is a solution
to the variational inequality. So, finally, if we let

f (x) = projK (x + F(x)) , (3.2)

then the fixed points of f coincide exactly with solutions to VI(K, F).
In many cases of practical interest, F will be a continuous function. Further-
more, it can be shown that the projection mapping onto a convex set (such as
K) is a well-defined (i.e., single-valued), continuous function. Then by Proposi-
tion A.6, the function f defined by equation (3.2) is a continuous function. So,
if the set K is compact in addition to being convex, then Brouwer’s theorem
shows that the variational inequality must have at least one solution:

Theorem 3.2. If K ⊆ Rn is a compact, convex set and F : Rn → Rn is a

continuous function, then the variational inequality VI(K, F) has at least one
solution.

You should convince yourself that all of these conditions are necessary: if the
container K is not bounded or not closed, or if the force field F is not continuous,
then it is possible that an object placed at any point in the container will move
under the action of the force field.

3.3 Convex Optimization

As we will see in the coming chapters, fixed points and variational inequalities
are relatively intuitive ways to represent the idea of an equilibrium in transporta-
tion systems. A fixed point captures the idea that “at equilibrium, nobody can
make a better choice than the one they are currently making; therefore the state
of the system is the same from one day to the next.” The transit example in
Section 3.1 illustrates how such a fixed point problem might arise. The physical
58 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

analogy with variational inequalities suggests how travelers might change their
behavior from one day to the next; the “container” represents the set of pos-
sible travel choices, and the “force field” represents travelers’ desires to move
to lower-cost options. An equilibrium solution is one where there is no way to
move in such an improving direction.
Both of these methods have disadvantages. Fixed point theorems are “non-
constructive,” which means they often lack methods guaranteed to find a fixed
point, even if one exists. Brouwer’s theorem gives us conditions under which a
fixed point exists, but tells us nothing about how to find it. Sometimes applying
f repeatedly from a starting point will converge to a fixed point, but not always.
The force field analogy in variational inequalities suggests a natural algorithm
(pick a starting point, and see where the force carries you), but again this is
not always guaranteed to work. It is also possible to have multiple solutions
to a fixed point or variational inequality problem. From the standpoint of
transportation planning this is inconvenient — how can you consistently rank
alternatives if you have several different predictions for what might happen
under each alternative?
Convex optimization is a more powerful tool in that we have uniqueness
guarantees on solutions, and efficient algorithms that provably converge to an
optimal point. The downside is that it is not immediately obvious how a user
equilibrium problem can be formulated in terms of optimization. Chapter 5
takes up this task; this section presents what you need to know about convex
optimization for this to make sense. If you have never encountered optimiza-
tion problems before, please read Appendix B before proceeding further, to
get familiar with the terminology and notation used in presenting optimization
problems. This subsection will focus on convex optimization, a specific kind of
optimization problem which is both easier to solve, and well-suited for solving
transportation network problems. If you are interested in other applications,
Appendix C discusses methods and properties of other kinds of optimization
problems.
The reason to restrict attention to convex optimization is that some opti-
mization problems are much easier to solve than others. For instance, the func-
tion in Figure 3.6 has many local minima and is unbounded below as x → −∞,
both of which can cause serious problems if we’re trying to minimize this func-
tion. Usually, the best that a software program can do is find a local minimum.
If it finds one of the local minima for this function, it may not know if there
is a better one somewhere else (or if there is, how to find it). Or if it starts
seeking x values which are negative, we could run into the unbounded part of
this function.
On the other hand, some functions are very easy to minimize. The function
in Figure 3.7 only has one minimum point, is not unbounded below, and there
are many algorithms which can find that minimum point efficiently.
What distinguishes these is a property called convexity, which is defined in
Appendix A. If the feasible region is a convex set, and if the objective function
is a convex function, then it is much easier to find the optimal solution. Check-
ing convexity of the objective function is not usually too difficult. To check
3.3. CONVEX OPTIMIZATION 59

Figure 3.6: A function which is not convex.

Figure 3.7: A convex function.

60 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

convexity of the feasible region, the following result is often useful.

Theorem 3.3. Consider an optimization program whose constraints all take
the form gi (x) ≤ 0 or hj (x) = 0, where i = 1, 2, . . . , k indexes the inequality
constraints and j = 1, 2, . . . , ` indexes the equality constraints. If each function
g1 (x), . . . , gk (x) is convex, and if each function h1 (x), . . . , h` (x) is linear, then
the feasible region X = {x ∈ Rn : gi (x) ≤ 0, hj (x) = 0, i ∈ {1, . . . , k}, j ∈
{1, . . . , `}} is a convex set.
Proof. Let Yi = {x ∈ Rn : gi (x) ≤ 0} represent the values of x which satisfy
the i-th inequality constraint, and let Zj = {x ∈ Rn : hj (x) = 0} be the values
of x which satisfy the j-th equality constraint. From Proposition A.14, all of
the sets Yi are convex. From Example A.2, all of the sets Zj are convex. The
feasible region X is the set of vectors x which satisfy all of the inequality and
equality constraints, that is, the intersection of all of the sets Yi and Zj . By
Proposition A.5, therefore, X is convex.
This is a very common situation, where the functions representing the in-
equality constraints are convex, and the functions representing equality con-
straints are linear. From this theorem, this means that the feasible region must
be convex.
This subsection collects a few useful results justifying why we’re spending
some time on convexity: minimizing a convex function over a feasible region
that is a convex set is extremely advantageous. Every local minimum is a global
minimum, every stationary point is a local minimum, and the set of global
minima is a convex set. Furthermore, if the function is strictly convex, the
global minimum is unique. In other words, unlike in elementary calculus, you
don’t have to perform any “second derivative tests” on solutions to ensure they
are truly minima, or distinguish between local and global minima. The set of
global minima being a convex set is useful because it means that all solutions
are “connected” or “adjacent” in some sense — there are no far-flung optimal
solutions.
First, a few definitions; in everything that follows, we are trying to minimize
a function f (x) over a feasible region X. (That is, X is the set of all x which
satisfy all of the constraints.)
Proposition 3.1. If f is a convex function and X is a convex set, then every
local minimum of f is also a global minimum.
Proof. By contradiction, assume that x1 is a local minimum of f , but not a
global minimum. Then there is some x2 ∈ X such that f (x2 ) < f (x1 ). Because
f is a convex function, for all λ ∈ (0, 1] we have

f ((1 − λ)x1 + λx2 ) ≤ (1 − λ)f (x1 ) + λf (x2 ) = f (x1 ) + λ(f (x2 ) − f (x1 ))

and furthermore all points (1 − λ)x1 + λx2 are feasible since X is a convex set
and x1 and x2 are feasible. Since f (x2 ) < f (x1 ), this means that

f ((1 − λ)x1 + λx2 ) < f (x1 )

3.3. CONVEX OPTIMIZATION 61

even as λ → 0, contradicting the assumption that x1 is a local minimum.

Proposition 3.2. If f is a convex function and X is a convex set, then the set
of global minima is convex.

Proof. Let X̂ be the set of global minima of f over the feasible region X. Choose
any two global optima x̂1 ∈ X̂ and x̂2 ∈ X̂, and any λ ∈ [0, 1].
Since x̂1 and x̂2 are global minima, f (x̂1 ) = f (x̂2 ); let fˆ denote this common
value. Because X is a convex set, the point (1 − λ)x̂1 + λx̂2 is also feasible.
Because f is a convex function,

f ((1 − λ)x̂1 + λx̂2 ) ≤ (1 − λ)f (x̂1 ) + λf (x̂2 )

= (1 − λ)fˆ + λfˆ
= fˆ

Therefore f ((1 − λ)x̂1 + λx̂2 ) ≤ fˆ But at the same time, f ((1 − λ)x̂1 + λx̂2 ) ≥ fˆ
because fˆ is the global minimum value of f . So we must have f ((1 − λ)x̂1 +
λx̂2 ) = fˆ which means that this point is also a global minimum and (1 − λ)x̂1 +
λx̂2 ∈ X̂ as well, proving its convexity.

Proposition 3.3. (Uniqueness.) If f is a strictly convex function and X is a

convex set, the set X̂ contains at most one element.

Proof. By contradiction, assume that X̂ contains two distinct elements x̂1 and
x̂2 . Repeating the proof of the previous proposition, because f is strictly convex,
the first inequality becomes strict and we must have f ((1 − λ)x̂1 + λx̂2 ) < fˆ.
This contradicts the assumption that x̂1 and x̂2 are global minima.

We next discuss techniques which are common in transportation network

analysis. We begin with simple problems: optimization problems with a single
decision variable, and optimization problems with no constraints, before mov-
ing to more general problem classes. In this section, we’ll assume that both
the objective function and feasible region are convex and differentiable, which
greatly simplifies matters.

3.3.1 One-dimensional problems

To start off, consider a simple minimization problem in one variable with no
constraints:

min f (x)
x

Because f is convex, any local minimum is also a global minimum. So, all
we need is to know when we’ve reached a local minimum. Thus, we’re looking
for a set of optimality conditions, that is, a set of equations and/or inequalities
62 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

in x which are true if and only if x is optimal. For the unconstrained case, this
is easy: we know from basic calculus that x̂ is a local minimum if

f 0 (x̂) = 0

(We don’t have to check whether x̂ is a local minimum or a local maximum

because f is convex.)
Example 3.1. Find the value of x which minimizes f (x) = x2 − 3x + 5.
Solution. f 0 (x) = 2x − 3, which vanishes if x = 3/2. Therefore x = 3/2
minimizes f (x). 
It’s a little bit more complicated if we add constraints to the picture. For
instance, consider the function in Figure 3.7 (which could very well be the
function from Example 3.1), but with the added constraint x ≥ 0. In this case,
nothing is different, and the optimum still occurs where f 0 (x) vanishes, that
is, at x = 3/2. But what if the constraint was x ≥ 2? In this case, x = 3/2
is infeasible, and f 0 (x) is always strictly positive in the entire feasible region.
This means that f is strictly increasing over the entire feasible region, so the
minimum value is obtained at the smallest possible value of x, that is, x = 2. So
we see that sometimes the local minimum of a constrained optimization problem
can be at a point where f 0 (x) is nonzero.
To simplify things a little bit, assume that the constraint is of the form
x ≥ 0, that is, we are trying to solve

min f (x)
x
s.t. x ≥0

As Figure 3.8 shows, there are only two possibilities. In the first case, the
minimum occurs when x is strictly positive. We can call this an interior min-
imum, or we can say that the constraint x ≥ 0 is nonbinding at this point. In
this case, clearly f 0 (x) must equal zero: otherwise, we could move slightly in
one direction or the other, and reduce f further. The other alternative is that
the minimum occurs for x = 0, as in Figure 3.8(b). For this to be a minimum,
we need f 0 (0) ≥ 0 — if f 0 (0) < 0, f is decreasing at x = 0, so we could move to
a slightly positive x, and thereby reduce f .
Let’s try to draw some general conclusions. For the interior case of Fig-
ure 3.8(a), we needed x ≥ 0 for feasibility, and f 0 (x) = 0 for optimality. For the
boundary case of Figure 3.8(b), we had x = 0 exactly, and f 0 (x) ≥ 0. So we see
that in both cases, x ≥ 0 and f 0 (x) ≥ 0, and furthermore that at least one of
these has to be exactly equal to zero. To express the fact that either x or f 0 (x)
must be zero, we can write xf 0 (x) = 0. So a solution x̂ solves the minimization
problem if and only if

x̂ ≥ 0
0
f (x̂) ≥ 0
x̂f 0 (x̂) = 0.
3.3. CONVEX OPTIMIZATION 63

(a) Convex function where the constraint is not binding at the minimum.

(b) Convex function where the constraint is binding at the minimum.

Figure 3.8: Two possibilities for minimizing a convex function with a constraint.
64 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

Whenever we can find x̂ that satisfies these three conditions, we know it is op-
timal. These are often called first-order conditions because they are related to
the first derivative of f . The condition x̂f (x̂) = 0 is an example of a comple-
mentarity constraint because it forces either x̂ or f (x̂) to be zero.

3.3.2 Bisection method

The bisection method allows us to solve one-dimensional problems over bounded
feasible regions. Consider the one-dimensional optimization problem

min f (x)
x
s.t. a ≤ x ≤ b

where f is continuously differentiable and convex. The bisection method works

by constantly narrowing down the region where the optimal solution lies. After
the k-th iteration, the bisection method will tell you that the optimum solution
lies in the interval [ak , bk ], with this interval shrinking over time (that is, bk −
ak < bk−1 − ak−1 ). A natural termination criterion is to stop when the interval
is sufficiently small, that is, when bk − ak < , where  is the precision you want
for the final solution.
The idea is that the sign of the derivative of the midpoint tells you where the
optimum is. If the derivative is negative at ak , but positive at bk , the optimum
occurs somewhere in-between, at a point where it is zero. So, if the derivative is
positive at the midpoint, we know that the zero point has to happen somewhere
to its left; if negative, somewhere to its right. If we happen to get lucky, the
derivative at the midpoint will be exactly zero, and you can stop — but this is
really rare. Figure 3.9 illustrates how bisection works. (The exercises ask you
to show that bisection converges to the optimum solution even if the derivative
initially has the same sign at both endpoints; the case where f 0 is positive at
ak but negative at bk is impossible for a convex function.)
Here’s how the algorithm works.
Step 0: Initialize. Set the iteration counter k = 0, a0 = a, b0 = b.
Step 1: Evaluate midpoint. Identify the midpoint ck = (ak + bk )/2 and cal-
culate the derivative of f there, dk = f 0 (ck )
Step 2: Bisect. If dk > 0, set ak+1 = ak , bk+1 = ck . Otherwise, set ak+1 = ck ,
bk+1 = bk .
Step 3: Iterate. Increase the counter k by 1 and check the termination crite-
rion. If bk − ak < , then terminate; otherwise, return to step 1.
Example 3.2. Use the bisection algorithm to find the minimum of f (x) =
(x − 1)2 + ex on the interval x ∈ [0, 2], to within a tolerance of  = 0.01.
Solution. You may find it useful to follow along in Table 3.1. We start off
with k = 0, a0 = 0, and b0 = 2. We calculate the derivative at the midpoint:
3.3. CONVEX OPTIMIZATION 65

Figure 3.9: Bisection method.

Table 3.1: Demonstration of the bisection algorithm with f (x) = (x − 1)2 + ex ,

x ∈ [0, 2].
k ak bk ck = (ak + bk )/2 dk
0 0 2 1 2.71 > 0
1 0 1 1/2 0.649 > 0
2 0 1/2 1/4 −0.216 < 0
3 1/4 1/2 3/8 0.205 > 0
4 1/4 3/8 5/16 −0.00816 < 0
5 5/16 3/8 11/32 0.0978 > 0
6 5/16 11/32 21/64 0.0446 > 0
7 5/16 21/64 41/128 0.0182 > 0
8 5/16 41/128 81/256

f 0 (x) = 2(x − 1) + ex , so f 0 (1) = 2.71, which is positive. Since f 0 is positive at

x = 1, the minimum must occur to the left of this point, that is, somewhere in
the interval [0, 1]. We set a1 and b1 equal to these new values, and repeat. The
new midpoint is 1/2, and f 0 (1/2) = 0.649 is again positive, so the minimum
must occur to the left of this point, in the interval [0, 1/2]. The midpoint of 0
and 1/2 is 1/4, and f 0 (1/4) = −0.216 is negative, so the minimum is to the right
of the midpoint. Thus, the new interval is [1/4, 1/2], and we repeat as shown
in Table 3.1.
During the eighth iteration, the width of the interval is 41/128 − 5/16 =
0.00781, which is less than the tolerance of 0.01; therefore we stop, and return
our best guess of the optimum as the midpoint of this interval: x̂ ≈ 81/256 =
0.31640625. The true minimum point occurs at x̂ = 0.3149230578...; if we had
chosen a smaller tolerance , the algorithm would have narrowed the interval
further, with both ends converging towards this point. 
66 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

There are additional methods for solving one-dimensional convex optimiza-

tion problems like these. Appendix C also describes the “golden section” method,
which is useful when the objective function is not differentiable, and a method
based on Newton’s method, which is useful when the objective is twice differ-
entiable.

3.3.3 Multidimensional problems with nonnegativity con-

straints
Most interesting optimization problems have more than one decision variable.
However, we will keep the assumption that the only constraint on the decision
variables is nonnegativity. Using the vector x to refer to all of the decision
variables, we solve the problem

min f (x)
x
s.t. x ≥0

Using the same logic as before, we can show that x̂ solves this problem if
and only if the following conditions are satisfied for every decision variable xi :

x̂i ≥ 0
∂f (x̂)
≥ 0
∂xi
∂f (x̂)
x̂i = 0
∂xi

You should convince yourself that if these conditions are not met for each de-
cision variable, then x̂ cannot be optimal: if the first condition is violated, the
solution is infeasible; if the second is violated, the objective function can be re-
duced by increasing xi ; if the first two are satisfied but the third is violated, then
x̂i > 0 and ∂f∂x(x̂)
i
> 0, and the objective function can be reduced by decreasing
xi .
This can be compactly written in vector form as

0 ≤ x̂ ⊥ ∇f (x̂) ≥ 0 (3.3)

where the ⊥ symbol indicates orthogonality, i.e. that the dot product of x̂ and
∇f (x̂) is zero.
Unfortunately, the bisection algorithm does not work nearly as well in higher
dimensions. It is difficult to formulate an extension that always works, and those
that do are inefficient. We’ll approach solution methods for higher-dimensional
problems somewhat indirectly, tackling a few other topics first: addressing con-
straints other than nonnegativity, and a few highlights of linear optimization.
3.3. CONVEX OPTIMIZATION 67

3.3.4 Constrained problems

Constraints can take a very large number of forms. For the purposes of this
book, we can restrict attention to only two types of constraints: linear equality
constraints and nonnegativity constraints. The previous two sections showed
you how to deal with nonnegativity constraints; this section discusses linear
equality constraints. A linear equality constraint is of the form
X
ai xi = b , (3.4)
i

where the xi are decision variables, and the ai and b are constants.
We can handle these using the technique of Lagrange multipliers. This tech-
nique is demonstrated in the following example for the case of a single linear
equality constraint.
min x21 + x22
x1 ,x2
s.t. x1 + x2 =5
(It is a useful exercise to verify that x21 + x22 is a strictly convex function, and
that {x1 , x2 : x1 + x2 = 5} is a convex set.)
The main idea behind Lagrange multipliers is that unconstrained problems
are easier than constrained problems. The technique is an ingenious way of
nominally removing a constraint while still ensuring that it holds at optimality.
The equality constraint is “brought into the objective function” by multiplying
the difference between the right- and left-hand sides by a new decision variable
κ (called the Lagrange multiplier), adding the original objective function. This
creates the Lagrangian function

L(x1 , x2 , κ) = x21 + x22 + κ(5 − x1 − x2 ) (3.5)

It is possible to show that the optimal solutions of the original optimization

problem correspond to stationary points of the Lagrangian function, that is, to
values of x1 , x2 , and κ such that ∇L(x1 , x2 , κ) = 0. To find this stationary
point, take partial derivatives with respect to each variable and set them all
equal to zero:

∂L
= 2x1 − κ = 0 (3.6)
∂x1
∂L
= 2x2 − κ = 0 (3.7)
∂x2
∂L
= 5 − x1 − x2 = 0 (3.8)
∂κ
Notice that the third optimality condition (3.8) is simply the original constraint,
so this stationary point must be feasible. Equations (3.6) and (3.7) respectively
tell us that x1 = κ/2 and x2 = κ/2; substituting these expressions into (3.8)
gives κ = 5, and therefore the optimal solution occurs for x1 = x2 = 5/2.
68 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

This technique generalizes perfectly well to the case of multiple linear equal-
ity constraints. Consider the general optimization problem

min f (x1 , . . . , xn )
x1 ,...,xn Pn
s.t. Pi=1 a1i xi = b1
n
i=1 a2i xi = b2
..
Pn .
i=1 a mi x i = bm

where f is convex. The corresponding Lagrangian is

 
n
X
L(x1 , . . . , xn , κ1 , . . . , κm ) = f (x1 , . . . , xn ) + κ1 b1 − a1j xj  +
j=1
   
n
X n
X
κ2 b2 − a2j xj  + · · · + κm bm − amj xj 
j=1 j=1

For an optimization problem that has both linear equality constraints and
nonnegativity constraints, we form the optimality conditions by combining the
Lagrange multiplier technique with the complementarity technique from the
previous section. Thinking back to Section 3.3.3, in the same way that we
replaced the condition f 0 (x̂) = 0 for the unconstrained case with the three
conditions x̂ ≥ 0, f 0 (x̂) ≥ 0, and xf 0 (x̂) = 0 when the nonnegativity constraint
was added, we’ll adapt the Lagrangian optimality conditions. If the optimization
problem has the form

min f (x1 , . . . , xn )
x1 ,...,xn Pn
s.t. Pi=1 a1i xi = b1
n
i=1 a2i xi = b2
..
Pn .
i=1 ami xi = bm
x1 , . . . , x n ≥0

where m ≤ n, then the Lagrangian is

 
n
X
L(x1 , . . . , xn , κ1 , . . . , κm ) = f (x1 , . . . , xn ) + κ1 b1 − a1j xj  +
j=1
   
n
X n
X
κ2 b2 − a2j xj  + · · · + κm bm − amj xj 
j=1 j=1
3.4. HISTORICAL NOTES AND FURTHER READING 69

and the optimality conditions are

∂L
≥0 ∀i ∈ {1, . . . , n}
∂xi
∂L
=0 ∀j ∈ {1, . . . , m}
∂κj
xi ≥ 0 ∀i ∈ {1, . . . , n}
∂L
xi =0 ∀i ∈ {1, . . . , n}
∂xi
Be sure you understand what each of these formulas implies. Each decision
variable must be nonnegative, the partial derivative of L with respect to this
variable must be nonnegative, and their product must equal zero (for the same
reasons as discussed in Section 3.3.3). For the Lagrange multipliers (κ1 , . . . , κm ),
the corresponding partial derivative of L must be zero. Notice how this is a
combination of the two techniques.
For small optimization problems, we can write down each of these conditions
and solve for the optimal solution, as above. However, for large-scale problems
this process can be very inconvenient. Later chapters in the book explain meth-
ods which work better for large problems. As a final note, the full theory of
Lagrange multipliers is more involved than what is discussed here. However,
it suffices for the case of a convex objective function and linear equality con-
straints. Optimality conditions for some other cases are given in Appendix C.

3.4 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
There are a great number of resources on single and multivariable calcu-
lus, and on linear algebra, that readers can consult if they require additional
background than what is provided in this chapter on these subjects.
For the mathematical concepts introduced in this chapter, the main result
on fixed point problems (Brouwer’s theorem) was presented in Brouwer (1910).
Variational inequalities are described at length in the book by Facchinei and
Pang (2003). For extended treatments of convex optimization, see textbooks of
Rockafellar (1997) and Bertsekas (2003).
Mathematical optimization has a long history, including important contri-
butions by Fermat, Newton, the Bernoullis, Lagrange, and Gauss. These earlier
methods are largely based in calculus and analytical in nature (with Newton’s
method being a notable exception). With the advent of computers in the early
20th century, and the logistics demands imposed by World War II, the field of
optimization took on additional foci centered on computation and solution of
large-scale problems. The seminal work of Dantzig (1963) in linear programming
substantially expanded both optimization theory and the range of applications
where optimization was used.
70 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM

Optimization problems are often classified based on the structure of the

objective function, decision variables, and constraints. Contemporary treat-
ments can be found in linear programming (Bertsimas and Tsitsiklis, 1997),
nonlinear programming (Bazaraa et al., 2006; Bertsekas, 2016), integer pro-
gramming (Wolsey, 1998), stochastic optimization (Birge and Louveaux, 1997),
and network optimization (Ahuja et al., 1993).
Simulated annealing was proposed by Kirkpatrick et al. (1983), building on
an earlier algorithm of Metropolis et al. (1953). Genetic algorithms for opti-
mization were first proposed by Rechenberg (1973) and Schwefel (1977). Other
metaheuristics not described in the text include GRASP (Feo and Resende,
1995), tabu search (Glover, 1989, 1990), ant swarms (Dorigo and Stützle, 2004),
and bee swarms (Pham et al., 2005).

3.5 Exercises
1. [31] For each of the following functions, find all of its fixed points or state
that none exist.
(a) f (x) = x2 , where X = R
(b) f (x) = 1 − x2 , where X = [0, 1]
(c) f (x) = ex , where X = R.
 
−x1
(d) f (x1 , x2 ) = where X = {(x1 , x2 ) : −1 ≤ x1 ≤ 1, 0 ≤ x2 ≤ 1}.
x2
 
−y
(e) f (x, y) = where X = {(x, y) : x2 + y 2 ≤ 1} is the unit disc.
x
2. [54] Brouwer’s theorem guarantees the existence of a fixed point for the
function f : X → X if f is continuous and X is closed, bounded, and
convex. Show that each of these four conditions is necessary by creating
examples of a function f and set X which satisfy only three of those
conditions but do not have a fixed point. Come up with such examples
with each of the four conditions missing. (The notation f : X → X means
that the range of the function must be contained in its domain; every
“output” from f is also a valid “input.”) Hint: you will probably find
it easiest to work with simple functions and sets whenever possible, e.g.
something like X = [0, 1]. Visualizing fixed points as intersections with
the diagonal line through the origin may help you as well.
3. [45] Find all of the solutions of each of the following variational inequalities
VI(K, F).
(a) F (x) = x + 1, K = [0, 1]
(b) F (x) = x2 , K = [−1, 1]
 
0
(c) F (x, y) = , K = {(x, y) : x2 + y 2 ≤ 1}
−y
3.5. EXERCISES 71

4. [58] Theorem 3.2 guarantees the existence of a solution to the variational

inequality VI(K, F) if K is closed, bounded, and convex, and if F is con-
tinuous. Show that each of these three conditions is necessary by creating
counterexamples of functions F and sets K which satisfy only three of
these conditions, but for which VI(K, F) has no solution. It may be help-
ful to include sketches.
72 CHAPTER 3. MATHEMATICAL TECHNIQUES FOR EQUILIBRIUM
 Part II

Static Traffic Assignment

73
Chapter 4

Introduction to Static
Assignment

4.1 Notation and Concepts

As discussed in Chapter 1, there are many possible measures of effectiveness
for evaluating the impacts of a transportation project or policy. Recall that
the traffic assignment problem assumes that the number of drivers traveling
between each origin zone and destination zone is given, and we want to find
the number of drivers using each roadway segment. This provides information
on congestion, emissions, toll revenue, or other measures of interest. We are
also given the underlying network, which has a set of links A representing the
roadway infrastructure, and a set of nodes N representing junctions, and a set
of centroids Z where trips are allowed to begin and end. Every centroid is
represented by a node, so Z is a subset of N .
Let drs denote the number of drivers whose trips start at centroid r and
end at centroid s. It is often convenient to write these values in matrix form,
the origin-destination matrix (OD matrix, for short) D where the number of
trips between r and s is given by the value in the r-th row and s-th column.
Together, an origin r and destination s form an OD pair (r, s).
Associated with each link is its flow, denoted xij , representing the total
number of vehicles wanting to use link (i, j) during the analysis period. The
flow is also known as the volume or demand. For reasons described below,
“demand” is actually the most accurate term, but “flow” and “volume” are the
most common for reasons of tradition, and “flow” is used in this book. The
travel time on link (i, j) is expressed as tij , and to represent congestion, we let
this travel time be a function of xij and write tij (xij ). Because of congestion
effects, tij is typically increasing and convex, that is, its first two derivatives
are typically positive. The function used to relate flow to travel time is called a
link performance function. The most common used in practice is the Bureau of

75
76 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

3
10x
50+x

10+x 10+x
1 4
50+x
10x
2

Figure 4.1: Example network for demonstration, with link performance func-
tions shown.

Public Roads (BPR) function, named after the agency which developed it:
 β !
0 xij
tij (xij ) = tij 1 + α , (4.1)
uij

where t0ij is the “free-flow” travel time (the travel time with no congestion), uij
is the practical capacity (typically the value of flow which results in a level of
service of C or D), and α and β are shape parameters which can be calibrated
to data. The values α = 0.15 and β = 4 are commonly used if no calibration is
done.
Notice that this function is well-defined for any value of xij , even when flows
exceed the stated “capacity” uij . In the basic traffic assignment problem, there
are no explicit upper bounds enforced on link flows. The interpretation of a
“flow” greater than the capacity is actually that the demand for travel on the
link exceeds the capacity, and queues will form. The delay induced by these
queues is then reflected in the link performance function. Alternately, one can
choose a link performance function which asymptotically grows to infinity as
xij → uij , implicitly enforcing the capacity constraint. However, this approach
can introduce numerical issues in solution methods. More discussion on this
issue follows in Section 4.1.1, but the short answer is that properly addressing
capacity constraints in traffic assignment requires a dynamic traffic assignment
model, which is the subject of Part III.
From the modeler’s perspective, the goal of traffic assignment is to determine
the link flows in a network. But from the standpoint of the travelers themselves,
it is easier to think of them each choosing a path connecting their origin to their
destination. Let drs denote the number of vehicles which will be departing origin
r to destination s, where r and s are both centroids. Using hπ to represent the
4.1. NOTATION AND CONCEPTS 77

number of vehicles who choose path π, a feasible assignment is defined as a

vector of path flows h satisfying the following conditions:
1. hπ ≥ 0 for all paths π ∈ Π. That is, path flows must be nonnegative.
π rs
P
2. π∈Πrs h = d for all OD pairs (r, s). Together with the nonnegativity
condition, this requires that every vehicle traveling from r to s is assigned
to exactly one of the paths connecting these zones.
Let H denote the set of feasible assignments.
Link flows xij and path flows hπ are closely linked, and we can obtain the
π
former from the latter. Let δij denote the number of times link (i, j) is used by
π π
path π, so δij = 0 if path π does not use link (i, j), and δij = 1 if it does. With
this notation, we have
XX X
π π
xij = δij h . (4.2)
r∈Z s∈Z π∈Πrs

This can be more compactly written using matrix notation as

x = ∆h , (4.3)

where x and h are the vectors of link and path flows, respectively, and ∆ is the
link-path adjacency matrix. The number of rows in this matrix is equal to the
number of links, and the number of columns is equal to the number of paths
in the network, and the value in the row corresponding to link (i, j) and the
π
column corresponding to path π is δij .
Given a feasible assignment, the corresponding link flows can be obtained
by using equation (4.2) or (4.3). The set of feasible link assignments is the set
X of vectors x which satisfy (4.3) for some feasible assignment h ∈ H.
Similarly, the path travel times cπ are directly related to the link travel times
tij : the travel time of a path is simply the sum of the travel times of the links
comprising that path. By the same logic, we have
X
cπ = π
δij tij (4.4)
(i,j)∈A

or, in matrix notation,

c = ∆T t . (4.5)
A small example illustrates these ideas. Consider the network in Figure 4.1
with four nodes and six links. The centroid nodes are shaded, and the link
performance functions are as indicated. For travelers from node 2 to node 4,
there are two paths: {(2, 3), (3, 4)} and {(2, 4)}. Using the compact notation
for paths, we could also write these as [2, 3, 4] and [2, 4]. Therefore, the set of
acyclic paths between these nodes is Π24 = {[2, 3, 4], [2, 4]}. You should verify
for yourself that

Π14 = {[1, 2, 4], [1, 2, 3, 4], [1, 3, 4], [1, 3, 2, 4]} .

78 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

It will be useful to have specific indices for each path, i.e. Π24 = {π1 , π2 } so
π1 = [2, 3, 4] and π2 = [2, 4], and similarly Π14 = {π3 , π4 , π5 , π6 } with π3 =
[1, 2, 4], π4 = [1, 2, 3, 4], π5 = [1, 3, 4], and π6 = [1, 3, 2, 4].
Let’s say that the demand for travel from centroid 1 to centroid 4 is 40
vehicles, and that the demand from 2 to 4 is 60 vehicles. Then d14 = 40 and
d24 = 60, and we must have
X
hπ = h3 + h4 + h5 + h6 = d14 = 40 (4.6)
π∈Π14

and X
hπ = h1 + h2 = d24 = 60 . (4.7)
π∈Π24

Let’s assume that the vehicles from each OD pair are divided evenly among
all of the available paths, so hπ = 10 for each π ∈ Π14 and hπ = 30 for each
π ∈ Π24 . We can now use equation (4.2) to calculate the link flows. For instance
the flow on link (1,2) is
( ( ))
X X X
π
δ1,2 hπ =
r∈Z s∈Z π∈Πrs
nn oo nn oo
3
= δ(1,2) h3 + δ(1,2)
4
h4 + δ(1,2)
5
h5 + δ(1,2)
6
h6 + 1
δ(1,2) h1 + δ(1,2)
2
h2 =
= {{1 × 10 + 1 × 10 + 0 × 10 + 0 × 10}} + {{0 × 30 + 0 × 30}} = 20 . (4.8)

where the braces show how the summations “nest.” Remember, this is just a
fancy way of picking the paths which use link (1,2), and adding their flows. The
equation is for use in computer implementations or for large networks; when
solving by hand, it’s perfectly fine to just identify the paths using a particular
link by inspection — in this case, only paths π11,4 and π21,4 use link (1,2). Repeat-
ing similar calculations, you should verify that x13 = 20, x23 = 40, x32 = 10,
x24 = 50, and x34 = 50.
From these link flows we can get the link travel times by substituting the
flows into the link performance functions, that is, t12 = 10x12 = 200, t13 =
50 + x13 = 70, t23 = 10 + x23 = 50, t32 = 10 + x32 = 20, t24 = 50 + x24 = 100,
and t34 = 10x34 = 500. Finally, the path travel times can be obtained by either
adding the travel times of their constituent links, or by applying equation (4.4).
You should verify that c1 = 550, c2 = 100, c3 = 300, c4 = 750, c5 = 570, and
c6 = 190.
The role of traffic assignment is to choose one path flow vector ĥ for purposes
of forecasting and ranking alternatives, out of all of the feasible assignments in
the network. An assignment rule is a principle used to determine this path flow
vector. The most common assignment rule in practice is that the path flow
vector should place all vehicles on a path with minimum travel time between
their origins and destinations, although other rules are possible as well and will
be discussed later in the book.
4.1. NOTATION AND CONCEPTS 79

x = Dh
Path flows h Link flows x

Assignment Link performance

rule functions

Path travel times c Link travel times t

c = DTt

Figure 4.2: Traffic assignment as an iterative process.

4.1.1 Commentary
The equations and concepts mentioned in the previous subsection can be related
as follows. Given a vector of path flows h, we can obtain the vector of link flows
x from equation (4.3); from these, we can obtain the vector of link travel times t
by substituting each link’s flow into its link performance function; from this, we
can obtain the vector of path travel times c from equation (4.5). This process
is shown schematically in Figure 4.2.
The one component which does not have a simple representation is how to
obtain path flows from path travel times using an assignment rule, “completing
the loop” with the dashed line in the figure. This is actually the most compli-
cated step, and answering it will require most of the remainder of the chapter.
The main difficulty is that introducing some rule for relating path travel times
to path flows creates a circular dependency: the path flows would depend on
the path travel times, which depend on the link travel times, which depend on
the link flows, which depend on the path flows, which depend on the path travel
times and so on ad infinitum. We need to find a consistent solution to this
process, which in this case means a set of path flows which remain unchanged
when we go around the circuit: the path flows must be consistent with the travel
times we obtain from those same path flows. Furthermore, the assignment rule
must reflect the gaming behavior described in Section 1.3.
At this point, it is worthwhile to discuss the assumptions that go into the
traffic assignment problem as stated here. The first assumption concerns the
time scale over which the network modeling is occurring. This part of the book
is focused entirely on what is called static network modeling, in which we assume
that the network is close to a “steady state” during whatever length of time we
choose to model (whether a peak hour, a multi-hour peak period, or a 24-hour
model), and the link flows and capacities are measured with respect to the entire
time period. That is, the capacity is the capacity over the entire analysis period,
so if a facility has a capacity of 2200 veh/hr and we are modeling a three-hour
peak period, we would use a capacity of 6600. Likewise, the link flows are the
total flow over the three hours.
80 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

How long the analysis period should be is a matter of some balancing. Ob-
viously, the longer the analysis period, the less accurate the steady state as-
sumption is likely to be. In the extreme case of a 24-hour model, it is usually
a (very big) stretch to assume that the congestion level will be the same over
all 24 hours. On the other hand, choosing too short an analysis period can be
problematic as well. In particular, if we are modeling a congested city or a large
metropolitan area, trips from one end of the network to the other can easily
take an hour or two, and it is good practice to have the analysis period be at
least as long as most of the trips people are taking.
Properly resolving the issue of the “steady state” assumption requires mov-
ing to a dynamic traffic assignment (DTA) model. DTA models have the po-
tential to more accurately model traffic, but are harder to calibrate, are more
sensitive to having correct input data, and require more computer time. Fur-
ther, DTA models end up requiring very different formulations and approaches.
In short, while useful in some circumstances, they are not universally better
than static models, and in any case they are surprisingly dissimilar. As a result,
a full discussion of DTA models is deferred to the final part of this volume.
Another assumption we make is that link and path flows can take any non-
negative real value. In particular, they are not required to be whole numbers,
and there is no issue with saying that the flow on √a link is 12.5 vehicles or that
the number of vehicles choosing a path is, say, 2. This is often called the
continuum assumption, because it treats vehicles as an infinitely-divisible fluid,
rather than as a discrete number of “packets” which cannot be split. The reason
for this assumption is largely computational — without the continuum assump-
tion, traffic assignment problems become extremely difficult to solve, even on
small networks. Further, from a practical perspective most links of concern have
volumes in the hundreds or thousands of vehicles per hour, where the difference
between fractional and integer values is negligible. Some also justify the con-
tinuum assumption by interpreting link and path flows to describe a long-term
average of the flows, which may fluctuate to some degree from day to day.

4.2 Principle of User Equilibrium

The previous section introduced the main ideas and components of static traffic
assignment, but did not provide much explanation of assignment rules other
than claiming that a reasonable assignment rule results in all used routes con-
necting an origin and destination to have equal and minimal travel time, and
that the “gaming” equilibrium idea described in Section 1.3 is relevant to route
choice. This section explains this assignment rule in more detail.
Assignment rules are more complex than the other steps in traffic assignment
shown in Figure 4.2, for two reasons. The first is that, being behavioral, it does
not follow as mechanically as the other three steps, and has a more complicated
answer. The other reason, alluding to the discussion in the previous section, is
that adding this link creates a cycle of dependency. Linking path flows to path
travel times, we now have four quantities which all depend on each other, and
4.2. PRINCIPLE OF USER EQUILIBRIUM 81

Table 4.1: Potential criteria in choosing a route.

Low travel time
Reliable travel time
Low out-of-pocket cost (tolls, fuel, etc.)
Short distance
Bias toward (or away from) freeways
Low accident risk
Few potholes
Scenic view

untangling this cycle requires a little bit of thought.

To handle this, let’s consider a simpler situation first. Rather than trying
to find path flows (which require stating every driver’s route choice), let’s stick
to a single driver. Why do they pick the route that they do? The number
of potential paths in a network between two even slightly distant points in a
network is enormous; in a typical city there are literally millions of possible
routes one could theoretically take between an origin and destination. The vast
majority of these are ridiculous, say, involving extraneous trips to the outskirts
of the city and then back to the destination even though the origin is nearby.
Why are such paths ridiculous on their face? Table 4.1 gives a list of potential
criteria which a desirable route would have, given in roughly decreasing order (at
least according to my tastes for a trip to work or school). The most important
times of day to properly model are the morning and evening peak periods,
when most of the trips made are work-related commutes. For these types of
trips, most people will choose the most direct route to their destination, where
“direct” means some combination of travel time, cost, and distance (which are
usually correlated). For simplicity, we’ll use “low travel time” as our starting
point, specified as Assumption 4.1.

Assumption 4.1. (Shortest path assumption.) Each driver wants to choose

the path between their origin and their destination with the least travel time.

Notice that this principle does not include the impact a driver has on other
drivers in the system, and a pithy characterization of Assumption 4.1 is that
“people are greedy.” If this seems unnecessarily pejorative, Section 4.3 shows
a few ways that this principle can lead to suboptimal flow distributions on
networks. Again, I emphasize that we adopt this assumption because we are
modeling urban, peak period travel which is predominantly composed of work
trips. If we were modeling, say, traffic flows around a national park during
summer weekends, quality of scenery may be considerably more important than
being able to drive at free-flow speed, and a different assumption would be
needed.
Further, the basic model developed in the first few weeks could really func-
tion just as well with cost or some other criterion, as long as it is separable and
additive (that is, you can get the total value of the criterion by adding up its
82 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

value on each link — the travel time of a route is the sum of the travel times
on its component links, the total monetary cost of a path is the sum of the
monetary costs of each link, but the total scenic quality of a path may not be
the sum of the scenic quality of each link), and can be expressed as a function of
the link flows x. So even though we will be speaking primarily of travel times,
it is quite possible, and sometimes appropriate, to use other measures as well.
A second assumption follows from modeling commute trips:
Assumption 4.2. Drivers have perfect knowledge of link travel times.
In reality, drivers’ knowledge is not perfect — but commutes are typically
habitual trips, so it is not unreasonable to assume that drivers are experienced
and well-informed about congestion levels at different places in the network,
at least along routes they might plausibly choose. We will later relax this
assumption, but for now we’ll take it as it greatly simplifies matters and is not
too far from the truth for commutes. (Again, in a national park or other place
with a lot of tourist traffic, this would be a poor assumption.)
Now, with a handle on individual behavior, we can try to scale up this
assumption to large groups of travelers. What will be the resulting state if
there are a large number of travelers who all want to take the fastest route
to their destinations? For example, if you have to choose between two routes,
one of which takes ten minutes and the other fifteen, you would always opt for
the first one. If you are the only one traveling, this is all well and good. The
situation becomes more complicated if there are others traveling. If there are
ten thousand people making the same choice, and all ten thousand pick the first
route, congestion will form and the travel time will rapidly increase. According
to Assumption 4.2, drivers would become aware of this, and some people would
switch from the first route to the second route, because the first would no longer
be faster. This process would continue: as long as the first route is slower, people
would switch away to the second route. If too many people switch to the second
route, so the first becomes faster again, people would switch back.
With a little thought, it becomes clear that if there is any difference in the
travel times between the two routes, people will switch from the slower route
to the faster one. Note that Assumption 4.1 does not make any allowance for
a driver being satisfied with a path which is a minute slower than the fastest
path, or indeed a second slower, or even a nanosecond slower. Relaxing Assump-
tion 4.1 to say that people may be indifferent as long as the travel time is “close
enough” to the fastest path leads to an interesting, but more complicated line
of research based on the concept of “bounded rationality,” which is discussed
later, in Section 5.3.3. It is much simpler to assume that Assumption 4.1 holds
strictly, in which case there are only three possible stable states:
1. Route 1 is faster, even when all of the travelers are using it.
2. Route 2 is faster, even when all of the travelers are using it.
3. Most commonly, neither route dominates the others. In this case, people
use both Routes 1 and 2, and their travel times are exactly equal.
4.2. PRINCIPLE OF USER EQUILIBRIUM 83

Because the third case is most common, this basic route choice model is
called user equilibrium: the two routes are in equilibrium with each other. Why
must the travel times be equal? If the first route was faster than the second,
people would switch from the second to the first. This would decrease the travel
time on the second route, and increase the travel time on the first, and people
would continue switching until they were equal. The reverse is true as well: if
the second route were faster, people would switch from the first route to the
second, decreasing the travel time on the first route and increasing the travel
time on the second. The only outcome where nobody has any reason to change
their decision, is if the travel times are equal on both routes.
This is important enough to state again formally:

Corollary 4.1. (Principle of user equilibrium.) Every used route connecting

an origin and destination has equal and minimal travel time.

Unused routes may of course have a higher travel time, and used routes
connecting different origins and destinations may have different travel times,
but any two used routes connecting the same origin and destination must have
exactly the same travel times. Notice that we call this principle a corollary
rather than an assumption: the real assumptions are the shortest path and full
information assumptions. If you believe these are true, the principle of user equi-
librium follows immediately and does not require you to assume anything more
than you already have. The next section describes how to solve for equilibrium,
along with a small example.

4.2.1 A trial-and-error solution method

We can develop a simple method for solving for path flows using the principle of
user equilibrium, using nothing but the definition itself. For now, assume there
is a single OD pair (r, s). The method is as follows:

1. Select a set of paths Π̂rs which you think will be used by travelers from
this OD pair.

2. Write equations for the travel times of each path in Π̂rs as a function of
the path flows.

3. Solve the system of equations enforcing equal travel times on all of these
paths, together with the requirement that the total path flows must equal
the total demand drs .

4. Verify that this set of paths is correct; if not, refine Π̂rs and return to step
2.

The first step serves to reduce a large potential set of paths to a smaller set of
reasonable paths Π̂rs , which you believe to be the set of paths which will be
used by travelers from r to s. Feel free to use engineering judgment here; if this
is not the right set of paths, we’ll discover this in step 4 and can adjust the set
84 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

accordingly. The second step involves applying equations (4.4) and (4.2). Write
the equations for link flows as a function of the flow on paths in Π̂rs (assuming
all other paths have zero flow). Substitute these expressions for link flows into
the link performance functions to get an expression for link travel times; then
write the equations for path travel times as a function of link travel times.
If Π̂rs truly is the set of used paths, all the travel times on its component
paths will be equal. So, we solve a system of equations requiring just that. If
there are n paths, there are only n − 1 independent equations specifying equal
travel times. (If there are three paths, we have one equation stating paths one
and two have equal travel time, and a second one stating paths two and three
have equal travel time. An equation stating that paths one and three have
equal travel time is redundant, because it is implied by the other two, and so
it doesn’t help us solve anything.) To solve for the n unknowns (the flow on
each path), we need one more equation: the requirement that the sum of the
path flows must equal the total flow from r to s (the “no vehicle left behind”
equation requiring every vehicle to be assigned to a path). Solving this system
of equations gives us path flows which provide equal travel times.
The last step is to verify that the set of paths is correct. What does this
mean? There are two ways that Π̂rs could be “incorrect”: either it contains
paths that it shouldn’t, or it omits a path that should be included. In the
former case, you will end up with an infeasible solution (e.g., a negative or
imaginary flow on one path), and should eliminate the paths with infeasible
flows, and go back to the second step with a new set Π̂rs . In the latter case,
you have a feasible solution and the travel times of paths in Π̂rs , but they are
not minimal: you have missed a path which has a faster travel time than any
of the ones in Π̂rs , so you need to include this path in Π̂rs and again return to
the second step.
Let’s take a concrete example: Figure 4.3 shows 7000 travelers traveling
from zone 1 to zone 2 during one hour, and choosing between the two routes
mentioned above: route 1, with free-flow time 20 minutes and capacity 4400
veh/hr, and route 2, with free-flow time 10 minutes and capacity 2200 veh/hr.
That means we have
  x 4 
1
t1 (x1 ) = 20 1 + 0.15 , (4.9)
4400
  x 4 
2
t2 (x2 ) = 10 1 + 0.15 . (4.10)
2200

This example is small enough that there is no real distinction between paths
and links (because each path consists of a single link), so link flows are simply
path flows (x1 = h1 and x2 = h2 ), and path travel times are simply link travel
times (c1 = t1 and c2 = t2 ). As a starting assumption, we assume that both
paths are used, so Π12 = {π1 , π2 }. We then need to choose the path flows h1
and h2 so that t1 (h1 ) = t2 (h2 ) (equilibrium) and h1 + h2 = 7000 (no vehicle
left behind). Substituting h2 = 7000 − h1 into the second delay function, the
4.2. PRINCIPLE OF USER EQUILIBRIUM 85

7000 7000
1 2

Figure 4.3: Small example using the two links.

equilibrium equation becomes

 4 !  4 !
h1 7000 − h1
20 1 + 0.15 = 10 1 + 0.15 . (4.11)
4400 2200

Using a numerical equation solver, we find that equilibrium occurs for h1 = 3376,
so h2 = 7000 − 3376 = 3624, and t1 (x1 ) = t2 (x2 ) = 21.0 minutes. Alternately,
we can use a graphical approach. Figure 4.4 plots the travel time on both routes
as a function of the flow on route 1 (because if we know the flow on route
1, we also know the flow on route 2). The point where they intersect is the
equilibrium: h1 = 3376, t1 = t2 = 21.0.
In the last step, we verify that the solution is reasonable (no paths have
negative flow) and complete (there are no paths we missed which have a shorter
travel time). Both conditions are satisfied, so we have found the equilibrium
flow and stop.
Let’s modify the problem slightly, so the travel time on link 1 is now
  x 4 
1
t1 (x1 ) = 50 1 + 0.15 .
4400
In this case, solving
 4 !  4 !
h1 7000 − h1
50 1 + 0.15 = 10 1 + 0.15
4400 2200

leads to a nonsensical solution: none of the answers involve real numbers without
an imaginary part. The physical interpretation of this is that there is no way to
assign 7000 vehicles to these two paths so they have equal travel times. Looking
at a plot (Figure 4.5), we see that this happens because path 2 dominates path
1: even with all 7000 vehicles on path 2, it has a smaller travel time.
86 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

Route 1

Route 2

Figure 4.4: The equilibrium point lies at the intersection of the delay functions.

So, the solution to the modified problem is: h1 = 0, h2 = 7000, t1 = 50,

and t2 = 39.2. This still satisfies the principle of user equilibrium: path 1 is not
used, so it is fine for its travel time to not equal that of path 2.
To find this solution in the trial-and-error method, we would start by re-
moving both paths (since imaginary path flows are not meaningful), and then
adding in the shortest path at free-flow (since there must be at least one used
path). When there is only one path in Π̂rs , the equilibrium state is easy to find,
because the “system of equations” reduces trivially to placing all demand onto
the single used path.
In more complicated networks, writing all of the equilibrium equations and
solving them simultaneously is much too difficult. Instead, a more systematic
approach is taken, and will be described in Chapter 6.

4.3 Three Motivating Examples

It is worth asking whether user equilibrium is the best possible condition. “Best”
is an ambiguous term, but can be related towards our general goals as trans-
portation engineers. User equilibrium probably does not lead to the flow pattern
with, say, the best emissions profile, simply because there’s no reason to believe
that the principle of user equilibrium has any connection whatsoever to emis-
sions — it is based on people trying to choose fastest paths. But maybe it is
related to congestion in some way, and you might find it plausible that each in-
dividual driver attempting to choose the best path for himself or herself would
minimize congestion system-wide. This section presents three examples which
should shatter this innocent-seeming idea. In this section, we will not concern
4.3. THREE MOTIVATING EXAMPLES 87

Route 1

Route 2

Figure 4.5: There is no intersection point: path 1 is dominated by path 2.

ourselves with how a feasible assignment satisfying this assignment rule is found
(the trial-and-error method from the previous section would suffice), but will
focus instead on how equilibrium can be used to evaluate the performance of
potential alternatives.
In the first example, consider the network shown in Figure 4.6 where the
demand between nodes 1 and 2 is d12 = 30 vehicles. Using h↑ and h↓ to
represent the flows on the top and bottom paths in this  network, the set of
feasible assignments are the two-dimensional vectors h = h↑ h↓ which satisfy


the conditions h↑ ≥ 0, h↓ ≥ 0, and h↑ + h↓ = 30. The figure shows the link

performance functions for each of the two links in the network. Notice that the
travel time on the top link is constant irrespective of the flow, while the travel
time on the bottom link increases with its flow. However, at low values of flow,
the bottom link is faster than the top one. This corresponds to a scenario where
one link is shorter, but more subject to congestion, while the other link is longer
but free of congestion.
With our stated assignment rule, the solution to the traffic assignment prob-
lem is h↑ = 25, h↓ = 5, because this will result in link flows of 25 and 5 on the
top and bottom links, respectively, giving travel times of 50 on both paths. In
this state, all vehicles in the network experience a travel time of 50.
Now, suppose that it is possible to improve one of the links in the network.
The intuitive choice is to improve the link which is subject to congestion, say,
changing its link performance function from 45 + x↓ to 40 + 21 x↓ , as shown in
Figure 4.7. (Can you see why this is called an improvement?) However, in
this case the path flow solution which corresponds to the assignment rule is
h↑ = 10, h↓ = 20, because this results in equal travel times on both the top and
bottom paths. These travel times are still 50! Even though the bottom link
88 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

50
30
1 2

45+x

Figure 4.6: A two-link network.

50
30
1 2

40+x/2

Figure 4.7: The two-link network with an improvement on the bottom link.

was improved, the effect was completely offset by vehicles switching paths away
from the top link and onto the bottom link.
In other words, improving a link (even the only congested link in a network)
does not necessarily reduce travel times, because drivers can change their behav-
ior in response to changes on the network. (If we could somehow force travelers
to stay on the same paths they were using before, then certainly some vehicles
would experience a lower travel time after the improvement is made.) This is
called the Knight-Pigou-Downs paradox.
The second example was developed by Dietrich Braess, and shows a case
where building a new roadway link can actually worsen travel times for all
travelers in the network, after the new equilibrium is established. Assume we
have the network shown in Figure 4.8a, with the link performance functions
next to each link. The user equilibrium solution can be found by symmetry:
since the top and bottom paths are exactly identical, the demand of six vehicles
will evenly split between them. A flow of three vehicles on the two paths leads
to flow of three vehicles on each of the four links; substituting into the link
performance functions gives travel times of 53 on (1, 3) and (2, 4) and 30 on
4.3. THREE MOTIVATING EXAMPLES 89

3
10x
50+x

1 4
50+x
10x
2

(a)

3
10x
50+x

10+x
1 4
50+x
10x
2

(b)

Figure 4.8: Braess network, before and after construction of a new link.

(1, 2) and (3, 4), so the travel time of both paths is 83, and the principle of user
equilibrium is satisfied.
Now, let’s modify the network by adding a new link from node 2 to node 3,
with link performance function 10 + x23 (Figure 4.8b). We have added a new
path to the network; let’s label these as follows. Path 1 is the top route [1, 3, 4];
path 2 is the middle route [1, 2, 3, 4], and path 3 is the bottom route [1, 2, 4].
Paths 1 and 3 each have a demand of three vehicles and a travel time of 83.
Path 2, on the other hand, has a flow of zero vehicles and a travel time of 70, so
the principle of user equilibrium is violated: the travel time on the used paths
is equal, but not minimal.
Assumption 4.1 suggests that someone will switch their path to take advan-
tage of this lower travel time; let’s say someone from path 1 switches to path 2,
so we have h1 = 2, h2 = 1, and h3 = 3. From this we can predict new link flows:
x12 = 4, x13 = 2, x23 = 1, x24 = 3, x34 = 3. Substituting into link performance
functions gives new link travel times: t12 = 40, t13 = 52, t23 = 11, t24 = 53,
t34 = 30, and finally we can recover the new path travel times: c1 = 82, c2 = 81,
and c3 = 93.
This is still not an equilibrium; perhaps someone from path 3 will switch
90 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

Saturation flow 30

35
1 2

Saturation flow 60

Figure 4.9: Network for Smith’s paradox.

to path 2 in an effort to save 12 minutes of time. So now h1 = h2 = h3 = 2,

the link flows are x12 = x34 = 4 and x13 = x23 = x24 = 2, the travel times
are t12 = t34 = 40, t13 = t24 = 52 and t23 = 12, and the path travel times are
c1 = c2 = c3 = 92. We have found the new equilibrium, and it is unconditionally
worse than the old one. Before adding the new link, each driver had a travel
time of 83; now every driver has a travel time of 92. Nobody is better off.
Everyone is worse off. This is the famous Braess paradox.
The third example, adapted from M. J. Smith, shows how the traffic assign-
ment problem can provide more insight into other traffic engineering problems
such as signal timing. Consider the network shown in Figure 4.9. Drivers choose
one of two links, which join at a signalized intersection at the destination. As-
sume that each of these links has the same free-flow time (t↑0 = t↓0 = 1 min),
but that the saturation flow of the bottom link is twice that of the top link
(s↑ = 30 veh/min and s↓ = 60 veh/min). Vehicles enter the network at a de-
mand level of d = 35 veh/min. The signal has a cycle length of C = 1 min and,
for simplicity’s sake, assume that there is no lost time so that the green times
allocated to the top and bottom links equal the cycle length: G↑ + G↓ = 1.
In traditional traffic signal analysis, the capacity u of an approach is the
saturation flow scaled by the proportion of green time given to that approach,
so ui = si (Gi /C) where i can refer to either the top or bottom link. The degree
of saturation X for an approach is the ratio of the link flow to the capacity,
so Xi = xi /ui = (xi C)/(si Gi ). With these quantities, the total travel time on
each link (free-flow time plus signal delay) can be written as
9 C(1 − Gi /C)2 Xi2
 
ti = 1 + + . (4.12)
20 1 − xi /si xi (1 − Xi )
Assume that initially, the signal is timed such that G↑ = 40 sec and G↓ =
20 sec. Then the formulas for delay on the top and bottom links are functions of
their flows alone, obtained from (4.12) by substituting the corresponding green
times into the expressions for Xi , and the equilibrium solution can be obtained
by solving the equations t↑ = t↓ and x↑ + x↓ = 35 simultaneously. With
these green times, the equilibrium solution occurs when x↑ = 23.6 veh/min and
x↓ = 11.4 veh/min, and the reader can confirm that the travel times on the two
links are t↑ = t↓ = 2.11 min.
4.3. THREE MOTIVATING EXAMPLES 91

So far, so good. Now assume that it has been a long time since the signal was
last retimed, and a traffic engineer decides to check on the signal and potentially
change the timing. A traditional rule in traffic signal timing is that the green
time given to an approach should be proportional to the degree of saturation.
However, with the given solution, the degrees of saturation are X ↑ = 0.982
and X ↓ = 0.953, which are unequal — the top link is given slightly less green
time than the equisaturation rule suggests, and the bottom link slightly more.
Therefore, the engineer changes the green times to equalize these degrees of sat-
uration, which occurs if G↑ = 48.3 sec and G↓ = 11.7 sec, a smallish adjustment.
If drivers could be counted on to remain on their current routes, all would be
well. However, changing the signal timing changes the delay formulas (4.12) on
the two links. Under the assumption that drivers always seek the shortest path,
the equilibrium solution will change as drivers swap from the (now longer) path
to the (now shorter) path. Re-equating the travel time formulas, the flow rates
on the top and bottom links are now x↑ = 23.8 veh/min and x↓ = 11.2 veh/min,
with equal delays t↑ = t↓ = 2.26 min on each link. Delay has actually increased,
because the signal re-timing (aimed at reducing delay) did not account for the
changes in driver behavior after the fact.
A bit surprised by this result, our diligent traffic engineer notes that the
degrees of saturation are still unequal, with X ↑ = 0.984 and X ↓ = 0.959,
actually further apart than before the first adjustment. Undeterred, the engineer
changes the signal timings again to G↑ = 48.5 s and G↓ = 11.5 s, which results
in equal degrees of saturation under the new flows. But by changing the green
times, the delay equations have changed, and so drivers re-adjust to move toward
the shorter path, leading to x↑ = 23.9 veh/min and x↓ = 11.1 veh/min, and new
delays of 2.43 minutes on each approach, even higher than before!
You can probably guess what happens from here, but Table 4.2 tells the
rest of the story. As our valiant engineer stubbornly retimes the signals in a
vain attempt to maintain equisaturation, flows always shift in response. Fur-
thermore, the delays grow faster and faster, asymptotically growing to infinity
as more and more adjustments are made. The moral of the story? Changing
the network will change the paths that drivers take, and “optimizing” the net-
work without accounting for how these paths will change is naive at best, and
counterproductive at worst.
 92
CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT
Table 4.2: Evolution of the network as signals are iteratively retimed.
Iteration G↑ (s) G↓ x↑ (v/min) x↓ t↑ (min) t↓ X↑ X↓
0 48 12 23.6 11.4 2.11 2.11 0.982 0.953
1 48.3 11.7 23.8 11.2 2.26 2.26 0.984 0.959
2 48.5 11.5 23.9 11.1 2.43 2.43 0.986 0.965
3 48.7 11.3 24.1 10.9 2.63 2.63 0.988 0.969
4 48.9 11.1 24.2 10.8 2.86 2.86 0.99 0.974
5 49.1 10.9 24.3 10.7 3.12 3.12 0.991 0.977
10 49.5 10.5 24.7 10.3 5.11 5.11 0.996 0.989
20 49.88 10.12 24.91 10.09 16.58 16.58 0.9988 0.9971
50 49.998 10.002 24.998 10.002 855.92 855.93 0.99998 0.99995
∞ 50 10 25 10 ∞ ∞ 1 1
4.4. HISTORICAL NOTES AND FURTHER READING 93

4.4 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
Beckmann et al. (1956) were the first to pose the static traffic assignment as-
signment problem in its full generality, although similar concepts were expressed
earlier in Pigou (1920) and Wardrop (1952).
The first motivating example is actually that of Pigou’s in the previous
reference; it is referred to as the “Knight-Pigou-Downs” paradox as the same
effect was discovered by Knight (1924) and Downs (1962). The second ex-
ample was first described by Braess (1969), and an up-to-date list of research
and popular works further exploring this paradox is maintained at https://
homepage.ruhr-uni-bochum.de/Dietrich.Braess/#paradox. The third ex-
ample is adapted from Smith (1979b).

4.5 Exercises
1. [15] Expand the diagram of Figure 1.4 to include a “government model”
which reflects public policy and regulation regarding tolls. What would
this “government” agent influence, and how would it be influenced?

2. [23] How realistic do you think link performance functions are? Name at
least two assumptions they make about traffic congestion, and comment
on how reasonable you think they are.

3. [23] How realistic do you think the principle of user equilibrium is? Name
at least two assumptions it makes (either about travelers or congestion),
and comment on how reasonable you think they are.

4. [12] Develop analogies to the equilibrium principle to represent the fol-

lowing phenomena: (a) better restaurants tend to be more crowded; (b)
homes in better school districts tend to be more expensive; (c) technolo-
gies to improve road safety (such as antilock brakes) can increase reckless
driving. Specify what assumptions you are making in these analogies.

5. [34] (All-or-nothing solutions are corner points.) The point x is a corner

point of the convex set X if it is impossible to write x = λy + (1 − λ)z
for y ∈ X, z ∈ X, and λ ∈ (0, 1). Show that if H is the set of feasible
path flows, h is a corner point if and only if it is an all-or-nothing solution
(a solution where all of the flow from each OD pair is on a single path).
Repeat for the set of feasible link flows.

6. [14] There are three routes available to travelers between a single OD

pair. For each case below, you are given the travel time t on each route,
as well as the number of travelers x using each route. Indicate whether
each case satisfies the principle of user equilibrium. If this principle is
violated, explain why.
94 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

1 5 6 3

2 7 8 4

Figure 4.10: Network for Exercise 7.

(a) t1 = 45, t2 = 60, and t3 = 30 while x1 = 0, x2 = 50, and x3 = 0.

(b) t1 = 45, t2 = 30, and t3 = 30 while x1 = 0, x2 = 50, and x3 = 25.
(c) t1 = 45, t2 = 30, and t3 = 30 while x1 = 0, x2 = 50, and x3 = 0.
(d) t1 = 15, t2 = 30, and t3 = 45 while x1 = 30, x2 = 0, and x3 = 0.
(e) t1 = 15, t2 = 30, and t3 = 45 while x1 = 30, x2 = 20, and x3 = 10.
(f) t1 = 15, t2 = 30, and t3 = 30 while x1 = 0, x2 = 50, and x3 = 30.

7. [26] The network in Figure 4.10 has 8 nodes, 12 links, and 4 zones. The
travel demand is d13 = d14 = 100, d23 = 150, and d24 = 50. The dashed
links have a constant travel time of 10 minutes regardless of the flow on
those links; the solid links have the link performance function 15 + x/20
where x is the flow on that link.
(a) For each of the four OD pairs with positive demand, list all acyclic
paths connecting that OD pair. In total, how many such paths are in
the network?
(b) Assume that the demand for each OD pair is divided evenly among all
of the acyclic paths you found in part (a) for that OD pair. What are
the resulting link flow vector x, travel time vector t, and path travel
time vector C?
(c) Does that solution satisfy the principle of user equilibrium?
(d) What is the total system travel time?

8. [32] Consider the network and OD matrix shown in Figure 4.11. The
travel time on every link is 10 + x/100, where x is the flow on that link.
Find the link flows and link travel times which satisfy the principle of user
equilibrium.
9. [36] Find the equilibrium path flows, path travel times, link flows, and link
travel times on the Braess network (Figure 4.8b) when the travel demand
from node 1 to node 4 is (a) 2; (b) 7; and (c) 10.
10. [52] In the Knight-Pigou-Downs network, the simple and seemingly rea-
sonable heuristic to improve the congested link (i.e., changing its link
4.5. EXERCISES 95

1 3

3 4
5 6
1 5,000 0
2 0 10,000
2 4

Figure 4.11: Network and OD matrix for Exercise 8.

performance function to lower the free-flow time) did not help. Can you
identify an equally simple heuristic that would lead you to improve the
other link? (Ideally, such a heuristic would be applicable in many kinds
of networks, not just one with an uncongestible link.)
11. [20] Give nontechnical explanations of why uniqueness, efficiency, and ex-
istence of equilibrium solutions have practical implications, not just theo-
retical ones. Concrete examples may be helpful.
12. [77] In the trial-and-error method, identify several different strategies for
choosing the initial set of paths. Test these strategies on networks of
various size and complexity. What conclusions can you draw about the
effectiveness of these different strategies, and the amount of effort they
involve?
13. [55] The trial-and-error method generally involves the solution of a system
of nonlinear equations. Newton’s method for solving a system of n nonlin-
ear equations is to move all quantities to one side of the equation, express-
ing the resulting equations in the form F(x) = 0 where F is a vector-valued
function mapping the n-dimensional vector of unknowns x to another n-
dimensional vector giving the value of each equation. An initial guess is
made for x, which is then updated using the rule x ← x − (JF(x))−1 F(x),
where (JF(x))−1 is the inverse of the Jacobian matrix of F, evaluated at
x. This process continues until (hopefully) x converges to a solution. A
quasi-Newton method approximates the Jacobian with a diagonal matrix,
which is equal to the Jacobian along the diagonal and zero elsewhere,
which is much faster to calculate and convenient to work with. Extend
your experiments from Exercise 12 to see whether Newton or quasi-Newton
methods work better.
14. [84] What conditions on the link performance functions are needed for
Newton’s method (defined in the previous exercise) to converge to a solu-
tion of the system of equations? What about the quasi-Newton method?
Assume that the network consists of n parallel links connecting a single
origin to a single destination. Can you guarantee that the solution to the
system of equations only involves real numbers, and not complex numbers?
96 CHAPTER 4. INTRODUCTION TO STATIC ASSIGNMENT

Hint: it may be useful to redefine the link performance functions for nega-
tive x values. This should not have any effect on the ultimate equilibrium
solution, since negative link flows are infeasible, but cleverly redefining the
link performance functions in this region may help show convergence.
15. [96] Repeat the previous exercise, but for a general network with any
number of links and nodes, and where paths may overlap.
Chapter 5

The Traffic Assignment

Problem

This chapter formalizes the user equilibrium traffic assignment problem defined
in Chapter 4. Using the mathematical language from Chapter 3, we are pre-
pared to model and solve equilibrium problems even on networks of realistic
size, with tens of thousands of links and nodes. Section 5.1 begins with fixed
point, variational inequality, and optimization formulations of the user equilib-
rium problem. Section 5.2 then introduces important existence and uniqueness
properties of the user equilibrium assignment, as was first introduced with the
small two-player games of Section 1.3. Section 5.3 names several alternatives
to the user equilibrium rule for assigning traffic to a network, including the
system optimal principle, and the idea of bounded rationality. This chapter
concludes with Section 5.4, exploring the inefficiency of the user equilibrium
rule and unraveling the mystery of the Braess paradox.

5.1 Mathematical Formulations

The previous section introduced a trial-and-error method for solving the user
equilibrium problem. To solve equilibrium on larger networks, we need some-
thing more sophisticated than trial-and-error, but doing so requires a more
convenient representation of the principle of user equilibrium in mathematical
notation, using the tools defined in the preceding chapters. This section does
so by first formulating the equilibrium problem as the solution to a variational
inequality; as the solution to a fixed point problem based on the variational
inequality; and as the solution to a convex optimization problem. The reason
for presenting all three of these formulations (rather than just one) is that each
is useful in different ways. Just as true fluency in a language requires being
able to explain the same concept in different ways, fluency in traffic assignment
problems requires familiarity with all of these ways of understanding the prin-
ciple of user equilibrium. (It is also possible to formulate the principle of user

97
98 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

f(x)
30

H
H
x

30

Figure 5.1: Feasible path-flow sets H for the simple two-link network (left) and
a schematic representing H for more complex problems (right).

f(x)
30

H
H
x

30

Figure 5.2: “Force” vectors −c for the simple two-link network (left) and a
schematic representing H for more complex problems (right).

equilibrium as a fixed point problem without referring to a variational inequality

through the use of multifunctions, as described in the optional subsection 5.1.1.)
The variational inequality formulation proceeds as follows. For each feasi-
ble path flow vector, we can associate a direction representing travelers’ desire
for lower travel time paths. Figure 5.1 shows the feasible set for the two-link
equilibrium problem on the left, and a conceptual schematic of the feasible set
for the general equilibrium problem on the right. Notice that in all cases the
feasible set is convex and has no “gaps.” For each feasible vector h, we can cal-
culate the path travel time vector c associated with these path flows. Associate
with each point a vector pointing in the opposite direction as c, that is, each
point h is associated with the direction −c(h). You can think about this as a
force acting on the point h. The interpretation is that the path flow vector is
“pulled” in the direction of decreasing travel times, and we will shortly show
that an equilibrium is a point which is unmoved by this force. Of course, this
intuitive interpretation must be established mathematically as well, and we will
show that a vector h satisfies the principle of user equilibrium if, and only if,
the force criterion is satisfied.
A few points are shown in Figure 5.2 as examples. Recall that a non-corner
point is unmoved by such a force only if the force is perpendicular to the bound-
5.1. MATHEMATICAL FORMULATIONS 99

ary of the feasible set at that point. At a corner point, the force must make a
right or obtuse angle with all boundary directions. So we can characterize stable
path flow vectors ĥ as points where the force −c(ĥ) makes a right or obtuse
angle with any feasible direction h − ĥ, where h is any other feasible path flow
vector. Recalling vector operations, saying that two vectors make a right or
obtuse angle with each other is equivalent to saying that their dot product is
nonpositive. Thus, ĥ is a stable point if it satisfies

− c(ĥ) · (h − ĥ) ≤ 0 ∀h ∈ H (5.1)

or, equivalently,
c(ĥ) · (ĥ − h) ≤ 0 ∀h ∈ H . (5.2)
This is a variational inequality (VI) in the form shown in Section 3.2. We
now prove that solutions of this VI correspond to equilibria, as shown by the
next result:

Theorem 5.1. A path flow vector ĥ solves the variational inequality (5.2) if
and only if it satisfies the principle of user equilibrium.

Proof. The theorem can equivalently be written “a path flow vector ĥ does not
solve (5.2) if and only if it does not satisfy the principle of user equilibrium,”
which is easier to prove. Assume ĥ does not solve (5.2). Then there exists some
h ∈ H such that c(ĥ) · (ĥ − h) > 0, or equivalently c(ĥ) · ĥ > c(ĥ) · h. Now,
c(ĥ) · ĥ is the total system travel time when the path flows are ĥ, and c(ĥ) · h
is the total system travel time if the travel times were held constant at c(ĥ)
even when the flows changed to h. For the latter to be strictly less than the
former, switching from ĥ to h must have reduced at least one vehicle’s travel
time even though the path travel times did not change. This can only happen
if that vehicle was not on a minimum travel time path to begin with, meaning
that ĥ does not satisfy user equilibrium.
Conversely, assume that ĥ is not a user equilibrium. Then there is some OD
pair (r, s) and path π such that ĥπ > 0 even though cπ (ĥ) > minπ0 ∈Πrs cπ0 (ĥ).
Let π̂ be a minimum travel time path for this OD pair. Create a new path
flow vector h which is the same as ĥ except that ĥπ is reduced by some small
positive amount  and ĥπ̂ is increased by . As long as 0 <  < ĥπ the new
point h remains feasible. By definition all the components of ĥ − h are equal
to zero, except the component for π is  and the component for π̂ is −. So
c(ĥ) · (ĥ − h) = (cπ (ĥ) − cπ̂ (ĥ)) > 0, so ĥ does not solve (5.2).

This variational inequality formulation leads directly to a fixed point formu-

lation, as was discussed in Section 3.2. Theorem 5.1 shows that the equilibrium
path flow vectors are exactly the solutions of the VI (5.2), that is, the stable
points with respect to the fictitious force −c. Therefore, an equilibrium vector
ĥ is a fixed point of the function projH (h − c(h)), that is, for all equilibrium
solutions ĥ we have
ĥ = projH (ĥ − c(ĥ)) . (5.3)
100 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

Finally, user equilibrium can also be shown as the solution to a convex

optimization problem. The main decision variables are the path flows h, and
we need to choose one from the set of feasible assignments H; the trick is
identifying an appropriate function f (h) in terms of the path flows, which is
minimized when the path flows satisfy the principle of user equilibrium. So, the
constraint set of our optimization problem consists of the equations defining H:
X
hπ = drs ∀(r, s) ∈ Z 2 (5.4)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (5.5)

Different assignment rules will lead to different objective functions. It turns

out that the objective function corresponding to the principle of user equilibrium
has a somewhat unintuitive form. Therefore, we will derive this function in
the same way that it was originally derived, by working backwards. Rather
than writing down the optimization problem, and then deriving the optimality
conditions, we start by writing down the optimality conditions we want, then
determining what kind of optimization problem has that form.
Writing the (unknown!) objective function as f (h), using the procedures
described in Chapter 3 we can Lagrangianize the constraints (5.4) introducing
multipliers κrs for each OD pair, providing the following optimality conditions:

∂f
π
− κrs ≥ 0 ∀(r, s) ∈ Z 2 , π ∈ Πrs (5.6)
 ∂h 
∂f
hπ − κ rs
=0 ∀(r, s) ∈ Z 2 , π ∈ Πrs (5.7)
∂hπ
X
hπ = drs ∀(r, s) ∈ Z 2 (5.8)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (5.9)

The last two of these are simply (5.4) and (5.5), requiring that the solution be
feasible. The condition (5.7) is the most interesting of these, requiring that the
product of each path’s flow and another term involving f must be zero. For
∂f
this product to be zero, either hπ must be zero, or ∂h π = κ
rs
must be true;
∂f rs
and by (5.6), for all paths ∂hπ ≥ κ . That is, in a solution to this problem,
∂f ∂f
whenever hπ is positive we have ∂h π = κ
rs
; if a path is unused, then ∂h π ≥ κ
rs

— in other words, at optimality all used paths connecting OD pair (r, s) must
∂f rs
have equal and minimal ∂h π , which is equal to κ . According to the principle
of user equilibrium, all used paths have equal and minimal travel time... so if
∂f π
we can choose a function f (h) such that ∂h π = c , we are done!

So, the objective function must involve some integral of the travel times. A
first guess might look something like

XZ ?
f (h) = cπ dhπ , (5.10)
π∈Π ?
5.1. MATHEMATICAL FORMULATIONS 101

where the bounds of integration and other details are yet to be determined.
The trouble is that cπ is not a function of hπ alone: the travel time on a path
depends on the travel times on other paths as well, so the partial derivative of
this function with respect to hπ will not simply be cπ , but contain other terms
as well. However, these interactions are not arbitrary, but instead occur where
paths overlap, that is, where they share common links. In fact, if we try writing
a similar function to our guess but in terms of link flows, instead of path flows,
we will be done.
To be more precise, let x(h) be the link flows as a function of the path flows
h, as determined by equation (4.2). Then the function
X Z xij (h)
f (h) = tij (x) dx (5.11)
(i,j)∈A 0

satisfies our purposes. To show this, calculate the partial derivative of f with
respect to the flow on an arbitrary path π, using the fundamental theorem of
calculus and the chain rule:
∂f X ∂xij X
π
π
= tij (xij (h)) π = δij tij (xij (h)) = cπ (5.12)
∂h ∂h
(i,j)∈A (i,j)∈A

where the last two equalities respectively follow from differentiating (4.2) and
from (4.4).
Finally, we can clean up the notation a bit by simply introducing the link
flows x as a new set of decision variables, adding equations (4.2) as constraints
to ensure they are consistent with the path flows. This gives the following
optimization problem, first formulated by Martin Beckmann:
X Z xij
min tij (x)dx (5.13)
x,h 0
(i,j)∈A
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (5.14)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (5.15)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (5.16)

Section 5.2 shows that the objective function is convex, and the feasible
region is a convex set, so this problem will be relatively easy to solve.

5.1.1 Multifunctions and application to network equilib-

rium (*)
(This section is optional and can be skipped. However if you are curious about
the “other road” to using fixed points to show equilibrium existence, read this
section to learn about multifunctions and Kakutani’s theorem.)
102 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

50
30
1 2

45+x

Figure 5.3: A simple two-link network for demonstration.

In the transit ridership example of Section 3.1, we could formulate the solu-
tion to the problem as a fixed point problem directly, without first going through
a variational inequality. The traffic assignment problem is slightly more com-
plex, because at equilibrium all used routes will have equal travel time. So, if
we try to apply the same technique as in the transit ridership problem, we will
run into a difficulty with “breaking ties.” To see this concretely, consider the
two-route network in Figure 5.3, and let h and c be the vectors of route flows
and route travel times. As Figure 1.3 suggests, we can try to define two func-
tions: H(c) gives the vector of path flows representing route choices if the travel
times were fixed at c, and C(h) gives the vector of path travel times when the
path flows are h. The function C is clearly defined: simply calculate the cost
on each route using the link performance functions. However, if both paths are
used at equilibrium, then c1 = c2 , which means that any path flow vector would
be a valid choice for H(c). Be sure you understand this difficulty: because H
assumes that the travel times are fixed, if they are equal any path flow vector is
consistent with our route choice assumptions. If we relax the assumption that
the link performance functions are fixed, then H is no simpler than solving the
equilibrium problem in the first place.
To resolve this, we introduce the concept of a multifunction.1 Recall that
a regular function from X to Y associates each value in its domain X with
exactly one value in the set Y . A multifunction, on the other hand, associates
each value in the domain X with some subset of Y . If F is such a multifunction
the notation F : X ⇒ Y can be used. For example, consider the multifunction
R(c). If c1 < c2 (and these travel times were fixed), then the only consistent
path flows are to have everybody on path 1. Likewise, if c1 > c2 , then everyone
would have to be on path 2. Finally, if c1 = c2 then people could split in any
proportion while satisfying the rule that drivers are using least-time paths. That

1 Also known by many other names, including correspondence, point-to-set map, set-valued

map, or set-valued function.

5.1. MATHEMATICAL FORMULATIONS 103

h1

30

c2 - c1

Figure 5.4: Multifunction corresponding to the two-link network; notice multiple

values when c1 = c2 .

is, nh io


 30 0 c1 < c2
nh io
R(c1 , c2 ) = 0 30 c1 > c2 (5.17)

 nh i o
 h 30 − h : h ∈ [0, 30]

c1 = c2

The notation in equation (5.17) is chosen very carefully to reflect the fact that
H is a multifunction, which means that its values are sets, not a specific number.
In the first two cases, the set only consists of a single element, so this distinction
may seem a bit pedantic; but in the latter case, the set contains a whole range of
possible values. In this way, multifunctions generalize the concept of a function
by allowing R to take multiple “output” values for a single input. This is
graphically represented in Figure 5.4, using the fact that the function can be
parameterized in terms of a single variable c2 − c1 .
Fixed-point problems can be formulated for multifunctions as well as for
functions. If F is an arbitrary multifunction defined on the set K, then a fixed
point of F is a point x such that x ∈ F (x). Note the use of set inclusion ∈
rather than equality = because F can associate multiple values with x. Just as
Brouwer’s theorem guarantees existence of fixed points for functions, Kakutani’s
theorem guarantees existence of fixed points for multifunctions under general
conditions:
Theorem 5.2. (Kakutani). Let F : K ⇒ K be a multifunction (from the set
K to itself ), where K is convex and compact. If F has a closed graph and F (x)
is nonempty and convex for all x ∈ K, then there is at least one point x ∈ K
such that x ∈ F (x).
The new terminology here is a closed graph; we say that the multifunction
F has a closed graph if the set {(x, y) : x ∈ X, y ∈ F (x)} is closed. Again,
each of these conditions is necessary; you might find it helpful to visualize these
conditions geometrically similar to Figure 3.1.
As a result of Kakutani’s theorem, we know that an equilibrium solution
must always exist in the traffic assignment problem. Let H denote the set of
104 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

all feasible path flow vectors, that is, vectors h such that π∈Πrs hπ = drs for
P
all OD pairs (r, s) and hπ ≥ 0 for all π ∈ Π. Let the function C(h) represent
the path travel times when the path flows are h.2 Let the multifunction R(c)
represent the set of path flows which could possibly occur if all travelers chose
least cost paths given travel times c. Mathematically
 
π π π0 2 rs
R(c) = h ∈ H : h > 0 only if c = min 0 rs
c for all (r, s) ∈ Z , π ∈ Π
π ∈Π
(5.18)
This is the generalization of equation (5.17) when there are multiple OD pairs
and multiple paths connecting each OD pair — be sure that you can see how
(5.17) is a special case of (5.18) when there is just one OD pair connected by
two paths.
Theorem 5.3. If the link performance functions tij are continuous, then at
least one equilibrium solution exists to the traffic assignment problem.
Proof. Let the multifunction F be the composition of the multifunction R and
the function C defined above, so F (h) = R(C(h)) is the set of path flow choices
which are consistent with drivers choosing fastest paths when the travel times
correspond to path flows h. An equilibrium path flow vector is a fixed point of
F : if h ∈ F (h) then the path flow choices and travel times are consistent with
each other. The set H of feasible h is convex and compact by Lemma 5.1 in
the next section. Examining the definition of R in (5.18), we see that for any
vector c, R(c) is nonempty (since there is at least one path of minimum cost),
and compact by the same argument used in the proof of Lemma 5.1. Finally,
the graph of F is closed, as you are asked to show in the exercises.

5.2 Properties of User Equilibrium

This section uses the mathematical formulations from the previous section to
explore basic properties of user equilibrium solutions. In Section 5.2.1 we give
fairly general conditions under which there is one (and only one) user equilibrium
link flow solution in a network. Although this section is mostly mathematical,
these properties are actually of great practical importance. We ultimately want
to use traffic assignment as a tool to help evaluate and rank transportation
projects, using the principle of user equilibrium to predict network conditions.
If it is possible that there is no feasible assignment which satisfies the principle
of user equilibrium, then we would be at a loss as to how projects should be
ranked, and we would need to find another assignment rule. Or, if there could
be multiple feasible assignments which satisfy the principle of user equilibrium,
again it is unclear how projects should be ranked.
Interestingly, while the conditions for a unique link flow user equilibrium
are fairly mild, there will almost certainly be multiple path flow solutions which
satisfy the principle of equilibrium — infinitely many, in fact. As explained in
2 If you look back to Chapter 4, you will see that C(h) = ∆T t(∆h) in matrix notation.
5.2. PROPERTIES OF USER EQUILIBRIUM 105

Section 5.2.2, this suggests that the principle of user equilibrium is not strong
enough to identify path flows, simply link flows. If path flows are needed to
evaluate a project, an alternative approach is needed, and this section explains
the concepts of maximum entropy and proportionality which provide this alter-
native.
Lastly, when solving equilibrium problems on large networks both the link
flow and path flow representations have significant limitations — algorithms
only using link flows tend to be slow, while algorithms only using path flows
can require a large amount of memory and tend to return low-entropy path
flow solutions. A compromise is to use the link flows, but distinguish the flow
on each link by its origin or destination. While explored more in Chapter 6,
Section 5.2.3 lays the groundwork by showing how flow can be decomposed this
way, and that at equilibrium the links with positive flow from each origin or
destination form an acyclic network.

5.2.1 Existence and link flow uniqueness

We would like to use the mathematical formulations from the previous section
— variational inequality, fixed point, and convex optimization — to identify
conditions under which one (and only one) user equilibrium solution exists.
This section relies heavily on mathematical formulations and results proved
previously; interested readers may refer to those sections for more detail, and
others can still read the proofs here to obtain the general idea. Recall the
following results from Chapter 3:
ˆ Brouwer’s Theorem 3.1: Let X be a compact and convex set, and
let f : X → X be a continuous function. Then there is at least one fixed
point x ∈ X.
ˆ Proposition 3.3: Let X be a convex set, and let f : X → R be a strictly
convex function. Then there is a unique global minimum x̂ of f on X.
Since all of these results rely on some properties of H and X, we establish
them first:
Lemma 5.1. The set of feasible path assignments H and the set of feasible link
assignments X are both compact and convex.
Proof. To show that H is compact, we must show that it is closed and bounded.
The set H is defined
P by a combination of linear weak inequalities (hπ ≥ 0) and
linear equalities ( h = drs ), so together Propositions A.3b and A.4a–b show
π

that it is closed. For boundedness, consider any h ∈ H, and OD pair (r, s), and
rs π rs π
P path ππ ∈ Πrs . We must have h ≤ d , since each h is nonnegative and
any
π∈Πrs h = d . Therefore, if D is the largest p
entry in the OD matrix, we
π
have h ≤ D for any path π. Therefore |h| ≤ |Π|D, so H is contained in the
ball B√|Π|D (0) and H is bounded.
Finally, for convexity, consider any h1 ∈ H, any h2 ∈ H, any λ ∈ [0, 1], and
the resulting vector h = λh1 + (1 − λ)h2 . For any path π, hπ1 ≥ 0 and hπ2 ≥ 0,
106 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

so hπ = λhπ1 + (1 − λ)hπ2 ≥ 0 as well. Furthermore, for any OD pair (r, s) we

have
X X
hπ = [λhπ1 + (1 − λ)hπ2 ]
(r,s)∈Πrs (r,s)∈Πrs
X X
=λ hπ1 + (1 − λ) hπ2
(r,s)∈Πrs (r,s)∈Πrs

= λdrs + (1 − λ)drs
= drs

so h ∈ H as well.
Every feasible link assignment x ∈ X is obtained from a linear transforma-
tion of some h ∈ H by (5.14) so X is also closed, bounded, and convex.

To show existence of solutions, we use Brouwer’s Theorem based on the

formulation of user equilibrium as a fixed point of the function f (h) = projH (h−
c(h)).

Proposition 5.1. If the link performance functions tij are continuous for each
link (i, j) ∈ A, then there is a feasible assignment satisfying the principle of user
equilibrium.

Proof. Taking H as the set of feasible assignments, the range of the func-
tion f (h) = projH (h − c(h)) clearly lies in H because of the projection. By
Lemma 5.1, H is compact and convex, so if we show that f is continuous, then
Brouwer’s Theorem guarantees existence of a fixed point. The discussion in
Section 5.1 showed that each fixed point is a user equilibrium solution, which
will be enough to prove the result.
We use the result that the composition of continuous functions is continuous
(Proposition A.6). The function f is the composition of two other functions
(call them f1 and f2 ), where f1 (h) = projH (h) and f2 (h) = h − c(h). By
Proposition A.7, f1 is continuous because H is a convex set. Furthermore, f2
is continuous if c is a continuous function of h, which is true by hypothesis.
Therefore, the conditions of Brouwer’s Theorem are satisfied, and f has at least
one fixed point (which satisfies the principle of user equilibrium).

The easiest way to show uniqueness of solutions is to makeP

use Rof the convex
xij
optimization formulation, minimizing the function f (x) = ij 0 tij (x) dx
over the set X.

Proposition 5.2. If the link performance functions tij are differentiable, and
t0ij (x) > 0 for all x and links (i, j), there is exactly one feasible link assignment
satisfying the principle of user equilibrium.

Proof. The previous section showed that user equilibrium link flow solutions
correspond to minimum points of f on X. If we can show that f is strictly
convex, then there can only be one minimum point. (Differentiability implies
5.2. PROPERTIES OF USER EQUILIBRIUM 107

continuity, so we can assume that at least one minimum point exists by Propo-
sition 5.1). Since f is a function of multiple variables (each link’s flow), to show
that f is convex we can write its Hessian matrix of second partial derivatives.
∂f
The first partial derivatives take the form ∂x ij
= tij (xij ). So, the diagonal en-
∂2f
tries of the Hessian take the form ∂x2ij
= t0ij (xij ), while the off-diagonal entries
2
∂ f
take the form ∂xij ∂xk` = 0. Since the Hessian is a diagonal matrix, and its diag-
onal entries are strictly positive by assumption, f is a strictly convex function.
Therefore there is only one feasible link assignment satisfying the principle of
user equilibrium.

It is also possible to prove the same result under the slightly weaker condition
that the link performance functions are continuous and strictly increasing; this
is undertaken in the exercises.
To summarize the above discussion, if the link performance functions are
continuous, then at least one user equilibrium solution exists; if in addition
the link performance functions are strictly increasing, then exactly one user
equilibrium solution exists. For many problems representing automobile traffic,
these conditions seem fairly reasonable, if not universally so: adding one more
vehicle to the road is likely to increase the delay slightly, but not dramatically.

5.2.2 Path flow nonuniqueness, entropy, and proportion-

ality
The previous subsection gave a relatively mild condition for the equilibrium
solution to be unique in terms of link flows. However, a more subtle point
is that the equilibrium solution need not be unique in terms of path flows.
Consider the example shown in Figure 5.5. With the paths as numbered  in
the
 figure, the reader
 can verify that both h1 = 20 10 20 10 and h2 =
30 0 10 20 produce the equilibrium link flows, and therefore satisfy the
principle of user equilibrium (all paths having equal travel time). Furthermore,
any weighted average of h1 and h2 also produces the same link flows, so there
are an infinite number of feasible path flow solutions which satisfy the principle
of user equilibrium.
At first glance this may appear to contradict the proof of Proposition 5.2,
which showed that the function f was strictly convex. Although f is strictly
convex as a function of x, it is not strictly convex as a function of h. This
is easy to see mathematically — if there are two distinct path flow solutions
h1 and h2 which satisfy the principle of user equilibrium, then λh1 + (1 −
λ)h2 is an equilibrium as well for λ ∈ (0, 1). Since all three are equilibria,
f (h1 ) = f (h2 ) = f (λh1 + (1 − λ)h2 ) even though strict convexity would require
f (λh1 + (1 − λ)h2 ) < λf (h1 ) + (1 − λ)f (h2 ) = f (h1 ). From a more intuitive
standpoint, in typical networks there are many more paths than links; therefore,
it is very likely that multiple path flow solutions correspond to the same link
flow solution. Since path travel times only depend on the values of the link
flows (because link flows determine link travel times, which are added up to get
108 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

30 40
A B C
30 20

Path h1 h2
20 30
10 0
20 10
10 20

Figure 5.5: Nonuniqueness of equilibrium in terms of path flows

path travel times), and since the principle of user equilibrium is defined in terms
of path travel times, it is reasonable to expect multiple path flow solutions to
satisfy the equilibrium principle.
The existence of multiple path flow equilibria (despite a unique link flow
equilibrium) is not simply a technical curiosity. Rather, it plays an important
role in using network models to evaluate and rank alternatives. Thus far, we’ve
been content to ask ourselves what the equilibrium link flows are, but in practice
these numbers are usually used to generate other, more interesting measures of
effectiveness such as the total system travel time and total vehicle-miles traveled.
If a neighborhood group is worried about increased traffic, link flows can support
this type of analysis. If we are concerned about safety, we can use link flows as
one input in estimating crash frequency, and so forth.
Other important metrics, however, require more information than just the
total number of vehicles on a link. Instead, we must know the entire paths used
by drivers. Examples include

Select link analysis: In transportation planning parlance, a “select link anal-

ysis” goes beyond asking “how many vehicles are on a link,” to “exactly
which vehicles are using this link?” This is used to produce nice visual-
izations, as well as to identify which origins and destinations are using a
particular link. This is needed to identify which neighborhoods and re-
gions are affected by transportation improvements, both positively and
negatively.

Equity and environmental justice: As a special case of the above, plan-

ners are focusing more and more on the notion of equity. Most trans-
portation projects involve both “winners” and “losers.” For instance, con-
structing a new freeway through an existing neighborhood may improve
regional mobility, but cause significant harm to those whose neighborhood
is disrupted in the process. Historically, disadvantaged populations tend
5.2. PROPERTIES OF USER EQUILIBRIUM 109

to bear the brunt of such projects. To know who is benefiting or suffer-

ing from a project, we need to know the origins and destinations of the
travelers on different links.
Emissions: Vehicles emit more pollutants toward the start of trips, when the
engine is cold, than later on once the catalytic converter has warmed up.
Therefore, to use the output of a network model to predict emissions from
vehicles on a link, ideally we need to distinguish between vehicles which
have just started their trips from vehicles which have already traveled
some distance. This requires knowing path flows.
Pavement loading: Heavy vehicles cause much more damage to pavement
than lighter vehicles; a commonly-used relation in pavement engineering
estimates that pavement damage is proportional to the fourth power of
vehicle weight. If a network model is being used to forecast pavement
deterioration, it is therefore important to know which links are being used
by heavy vehicles. Again, this requires more than just the total flow on a
link; we must know what specific type of vehicle it is (which often depends
heavily on the origin-destination pair).
In fact, software vendors often discuss such abilities as key features of their
applications. In this light, the fact that equilibrium path flows are nonunique
should trouble you somewhat. If there are multiple path flow equilibria (and
therefore multiple values for emissions forecasts, equity analyses, etc.), which
one is the “right” one? Simply picking one possible path flow solution and
hoping that it’s the right one is not particularly rigorous, and inappropriate for
serious planning applications.
So, the principle of user equilibrium is only strong enough to determine
link flows, not path flows, even though having a path flow solution would
be very helpful in analyzing engineering alternatives. This has motivated the
search for the “most likely path flow” solution which satisfies equilibrium. Even
though user equilibrium is not a strong enough concept to determine path flows
uniquely, we can still ask “of all the path flow vectors which satisfy the principle
of user equilibrium, which do we think is most likely to occur?” This section de-
scribes the concepts most commonly used to this end. To my knowledge, there
has been little to no research directly validating this principle, but nobody has
come up with a better, generally agreed-upon solution to this issue.
A physical analogy is presented to motivate the idea. Figure 5.6 shows the
locations of n gas molecules within a box. In the left panel, all n molecules
happen to be located in the top half of the box, while in the right panel n/2
molecules are located in the top half and n/2 in the bottom half. Both of these
situations are physically possible (that is, they satisfy the laws of mechanics,
etc.) although you would certainly think that the scenario on the right is more
likely or plausible than the scenario on the left. In exactly the same way,
although there are many path flow solutions which satisfy the “law” of user
equilibrium, we are not forced to conclude that all such solutions are equally
likely to occur.
110 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

Figure 5.6: Molecules of an ideal gas in a box. Which situation is more likely?

Why does the scenario on the left seem so unlikely? If the molecules form an
ideal gas, the location of each molecule is independent of the location of every
other molecule. Therefore, the probability that any given molecule is in the top
half of the box is 1/2, and the probability that all n molecules are as in the left
figure is
 n
L 1
p = . (5.19)
2
To find the probability that the distribution of molecules is as in the right figure
(pR ), we use the binomial distribution:
   n/2  n/2  n
n 1 1 n! 1
pR = = (5.20)
n/2 2 2 (n/2)!(n/2)! 2

Since n!  [(n/2)!]2 , we have pR  pL , that is, the situation on the right

is much, much likelier than that on the left even though both are physically
possible.
We want to use the same principle for path flows. The principle of user equi-
librium assumes that each user is identical (in the sense that they choose routes
only according to minimal travel time) and chooses routes independent of other
drivers (except insofar as other drivers’ choices affect travel times, drivers do not
coordinate and strategize about their route choices with each other). Therefore,
we can apply the same logic as with the gas molecules. For now, assume there
is a single OD pair, with integer demand d, and let Π̂ = {π1 , · · · , πk } denote
the set of minimum travel-time paths at the equilibrium solution. For simplicity
assume that these paths have constant travel times independent of flow. Since
drivers only care about travel time, each path is essentially identical, and the
probability any particular driver chooses any particular path  is 1/k. The  proba-
bility that the path flow vector h takes a particular value h1 · · · hk is then
given by the multinomial distribution:
 d
d! 1
p= (5.21)
h1 !h2 ! · · · hk ! k
5.2. PROPERTIES OF USER EQUILIBRIUM 111

and the most likely path flow vector is the one which maximizes this product.
Since (1/k)d is a constant, the most likely path flow vector simply maximizes
d!
p= (5.22)
h1 !h2 ! · · · hk !
Now we introduce a common trick: since the logarithm function is strictly
increasing, the path flows which maximize p also maximize log p, which is
k
X
log p = log d! − log hπ ! (5.23)
π=1

To simplify
√ further, we use Stirling’s approximation, which states that n! ≈
nn e−n 2πn, or equivalently log n! ≈ n log n − n + (1/2) log(2πn). This ap-
proximation
√ is asymptotically exact in the sense that the ratio between n! and
nn e−n 2πn approaches 1 as n grows large. Further, when n is large, n is much
larger than log n, so the last term can be safely ignored and log n! ≈ n log n − n.
Substituting into (5.23) we obtain
X
log p ≈ (d log d − d) − (hπ log hπ − hπ ) (5.24)
π
P
Since d = π hπ , we can manipulate (5.24) to obtain
X
log p ≈ − hπ log(hπ /d) (5.25)
π

and the
P path flows h1 , . . . , hk maximizing this quantity (subject to the constraint
d = π hπ and nonnegativity) approximately maximize the probability that
this particular path flow vector will occur in the field if travelers are choosing
routes independent of each other.
To move towards the traffic assignment problem we’re used to, we need to
make the following changes:
1. In the traffic assignment problem, the demand d and path flows hπ do not
need to be integers, but can instead take on real values. The corresponding
interpretation is that the demand is “divided” into smaller and smaller
units, each of which is assumed to act independently of every other unit.
(This is how user equilibrium works with continuous flows.) This doesn’t
cause any problems, and in fact helps us out: as we take the limit towards
infinite divisibility, Stirling’s approximation becomes exact — so we can
replace the ≈ in our formulas with exact equality.
2. There are multiple origins and destinations. This doesn’t change the basic
idea, the formulas just become incrementally more complex as we sum
equations of the form (5.25) for each OD pair.
3. Travel times are not constant, but are instead flow dependent. Again,
this doesn’t change any basic ideas; we just have to add a constraint
112 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

stating that the path flows correspond to an equilibrium solution. Since

the equilibrium link flows are unique, we simply have to state that the
path flows correspond to the equilibrium link flows x̂.

Putting all this together, we seek the path flows h which solve the optimization
problem
X X
max − hπ log(hπ /drs ) (5.26)
h
(r,s)∈Z 2 π∈Π̂rs
X
π
s.t. δij hπ = x̂ij ∀(i, j) ∈ A (5.27)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (5.28)
π∈Π̂rs
hπ ≥ 0 ∀π ∈ Π (5.29)

The objective function (5.26) is called the entropy of a particular path flow
solution, and the constraints (5.28), (5.27), and (5.29) respectively require that
the path flows are consistent with the OD matrix, equilibrium link flows, and
nonnegativity. The set Π̂rs is the set of paths which are used by OD pair (r, s).
The word entropy is meant to be suggestive. The connection with the ther-
modynamic concept of entropy may be apparent from the physical analogy at
the start of this section.3 In physics, entropy can be interpreted as a measure of
disorder in a system. Both the left and right panels in Figure 5.6 are “allowable”
in the sense that they obey the laws of physics. However, the scenario in the
right has much less structure and much higher entropy, and is therefore more
likely to occur.
It is not trivial to solve the optimization problem (5.26)–(5.29). It turns out
that entropy maximization implies a much simpler condition, proportionality
among pairs of alternate segments. Any two paths with a common origin and
destination imply one or more pairs of alternate segments where the paths di-
verge, separated by links common to both paths. In Figure 5.5, there are two
pairs of alternate segments: the top and bottom links between nodes A and B,
and the top and bottom links between B and C. It turns out that path flows h1
are the solution to the entropy maximization problem.
You might notice some regularity or patterns in this solution. For instance,
looking only at the top and bottom links between A and B, at equilibrium these
links have equal flow (30 vehicles each). Paths 1 and 3 are the same except for
between nodes A and B, and these paths also have equal flow (20 vehicles each).
The same is true for paths 2 and 4 (10 vehicles each). Or, more subtly, between
B and C, the ratio of flows between the top and bottom link is 2:1, the same as
the ratio of flows on paths 1 and 2 (which only differ between B and C), and
the ratio of flows on paths 3 and 4. This is no coincidence; in fact, we can show
that entropy maximization implies proportionality.
3 Actually, the term here is more directly drawn from the fascinating field of information

theory, which is unfortunately beyond the scope of this text.

5.2. PROPERTIES OF USER EQUILIBRIUM 113

A E
40 120

40
C D
120

120 40

B F

Figure 5.7: Multiple OD pair example of proportionality.

Before proving this fact, we show that it holds even for different OD pairs.
The network in Figure 5.7 has 40 vehicles traveling from origin A to destination
F, and 120 vehicles from B to E, and the equilibrium link flows are shown in
the figure. Each OD pair has two paths available to it, one using the top link
between C and D, and the other using the bottom link between C and D. Using
an upward-pointing arrow to denote the first type of path, and a downward-
pointing arrow to describe the second, the four paths are h↑AF , h↓AF , h↑BE , and
h↓BE . Solving the optimization problem (5.26)–(5.29), we find the most likely
path flows are h↑AF = 10, h↓AF = 30, h↑BE = 30, and h↓BE = 90. The equilibrium
flows for the top and bottom links between C and D have a ratio of 1:3; you
can see that this ratio also holds between h↑AF and h↓AF , as well as between h↑BE
and h↓BE .
In particular, the obvious-looking solution h↑AF = 40, h↓AF = 0, h↑BE = 0,
h↓BE = 120 has extremely low entropy, because it implies that for some reason all
travelers from one OD pair are taking one path, and all travelers from the other
are taking the other path, even though both paths are perceived identically by
all travelers and even though travelers are making choices independent of each
other. This is exceptionally unlikely.
We now derive the proportionality condition by defining it a slightly more
precise way:

Theorem 5.4. Let π1 and π2 be any two paths connecting the same OD pair.
If the path flows h solve the entropy-maximizing problem (5.26)–(5.29), then
the ratio of flows h1 /h2 for paths π1 and π2 is identical regardless of the OD
pair these paths connect, and only depends on the pairs of alternate segments
distinguishing these two paths.

The proof is a bit lengthy, and is deferred to the end of the section. But
even though the derivation is somewhat involved, the proportionality condition
itself is fairly intuitive: when confronted with the same set of choices, travelers
from different OD pairs should behave in the same way. The proportionality
114 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

condition implies nothing more than that the share of travelers choosing one
alternative over another is the same across OD pairs. Proportionality is also a
relatively easy condition to track and enforce.
It would be especially nice if proportionality can be shown fully equivalent
to entropy maximization. Theorem 5.4 shows that entropy maximization im-
plies proportionality, but is the converse true? Unfortunately, the result is no,
and examples can be created which satisfy proportionality without maximizing
entropy. Therefore, proportionality is a weaker condition than entropy max-
imization. The good news is that proportionality “gets us most of the way
there.” In the Chicago regional network, there are over 93 million equal travel-
time paths at equilibrium; after accounting for the equilibrium and “no vehicle
left behind” constraints (5.27) and (5.28), there are still over 90 million degrees
of freedom in how the path flows are chosen. Accounting for proportionality
reduces the number of degrees of freedom to 91 (a reduction of 99.9999%!)
So, in practical terms, enforcing proportionality seems essentially equivalent to
maximizing entropy.

5.2.3 Aggregation by origins or destinations

Thus far, we have looked at flow representations in one of two ways: path flows h
or link flows x. In a way, these can be thought of as two extremes. The path flow
solution contains the most information, showing the exact path chosen by every
traveler in the network. The price paid for this level of detail is the amount
of memory needed. The number of paths in a typical network is very, very
large and storing the entire vector of path flows is impractical. The link flows,
on the other hand, are much more compact and provide enough information
for many common measures of effectiveness (such as vehicle-hours or vehicle-
miles traveled, and volumes or delays on individual links). The link flow vector
can be thought of as an “aggregated” form of the path flow vector, where we
lump together all the travelers using the same link (regardless of which specific
path, they are using, or which specific OD pair they are from). Naturally, some
information is lost in this process of aggregation. So, there is a tradeoff: path
flow solutions contain a great deal of information, at the expense of requiring a
large amount of memory; link flow solutions are more compact but contain less
information.
So, it is natural to ask if there is an intermediate “Goldilocks” flow represen-
tation which contains more detail than link flow solutions, without invoking the
combinatorial explosion in storage space associated with a path flow solution.
One such representation involves aggregating all travelers departing the same
origin together.4 On a link (i, j), let xrij be the flow on this link who left origin
r, that is,
X X
xrij = π π
δij h (5.30)
s∈Z π∈Πrs

4 Everything in this section would apply equally well to an aggregation by destination; the

presentation from here on will be in terms of origins only to avoid repetition.

5.2. PROPERTIES OF USER EQUILIBRIUM 115

and let xr be the vector of link flows associated with origin r.

Many of the most efficient algorithms for solving the traffic assignment prob-
lem use such a representation, often called an origin-based or bush-based repre-
sentation. The term “origin-based” describes how the aggregation occurs. The
term “bush-based” will make more sense once we get to Chapter 6. Let ĥ be
a feasible assignment satisfying the principle of user equilibrium, and let xr be
the link flows associated with origin r at this solution.
It turns out that, at equilibrium, the links with positive xrij values form an
acyclic subnetwork. As a result, they have a topological order, which allows
network calculations to be done much more rapidly than in general networks.
This property is exploited heavily by bush-based traffic assignment algorithms,
which are discussed in Section 6.4.
The following result shows that this must be true when each link’s travel
time is strictly positive. With slightly more effort, a similar result can be shown
even when there are zero-time links.

Proposition 5.3. Let xr be the link flows associated with origin r at some
feasible assignment satisfying the principle of user equilibrium. If tij > 0 for all
links (i, j), then the subset of arcs Ar = {(i, j) ∈ A : xrij > 0} with positive flow
from origin r contains no cycle.

Proof. By contradiction, assume that there is a cycle [i1 , i2 , . . . , ik , i1 ], where

(i1 , i2 ), (i2 , i3 ), . . . , (ik , i1 ) are all in Ar . Let Lri represent the travel time on the
shortest path from origin r to node i. Consider any link (i, j) ∈ Ar ; this implies
that xrij > 0 and that some used path starting from origin r includes (i, j).
Because the principle of user equilibrium holds, this path must be a shortest
path to some destination. Clearly Lri + tij ≥ Lrj by the definition of the labels
L; but if Lri + tij > Lrj , then there is a shorter path to node j than any of
them passing through node i, and such a path cannot be used at equilibrium.
Therefore we must have Lri + tij = Lrj for all links in Ar ; since all travel times
are strictly positive this implies Lri < Lrj for all (i, j) ∈ Ar . So, for the cycle
under consideration, we have L1 < L2 < · · · < Lk < L1 , which is impossible.
Therefore the links in Ar cannot contain a cycle.

Proof of Theorem 5.4:

Begin by forming the Lagrangian for the entropy-maximizing problem:

  !
X X hπ X X
π
L(h, β, γ) = − hπ log + βij x̂ij − δij hπ
drs
(r,s)∈Z 2 π∈Π̂rs (i,j)∈A π∈Π
 
X X
+ γrs drs − hπ  (5.31)
(r,s)∈Z 2 π∈Π̂rs

using β and γ to denote the Lagrange multipliers. Note that the nonnegativity
constraint can be effectively disregarded since the objective function is only
116 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

defined for strictly positive h. Therefore at the optimum solution the partial
derivative of L with respect to any path flow must vanish:
 
∂L hπ X
π
= −1 − log − δij βij − γrs = 0 (5.32)
∂hπ drs
(i,j)∈A

where (r, s) is the OD pair connected by path π. Solving for hπ we obtain

 
X
hπ = drs exp −1 − π
δij βij − γrs  (5.33)
(i,j)∈A

Likewise we have
∂L X
= drs − hπ0 = 0 (5.34)
∂γrs 0 rs π ∈Π̂

so
 
X X X
drs = hπ0 = drs exp (−1 − γrs ) exp − π
δij βij  (5.35)
π 0 ∈Π̂rs π∈Π̂rs (i,j)∈A

substituting the result from (5.33). Therefore we can solve for γrs :
  
X X 0
π
γrs = −1 + log  exp − δij βij  (5.36)
π 0 ∈Π̂rs (i,j)∈A

Finally, substituting (5.36) into (5.33) and simplifying, we obtain

 
rs
d X
hπ = P  P  exp − π
δij βij  (5.37)
exp − δ π0 β
π 0 ∈Π̂rs (i,j)∈A ij ij (i,j)∈A

Noting that the fraction in (5.37) only depends on the OD pair (r, s), we can
simply write
 
X
hπ = Krs exp − π
δij βij  (5.38)
(i,j)∈A

where Krs is a constant associated with OD pair (r, s).

Let A1 be the set of links not in either π1 or π2 , A2 the set of links common
to both paths, A3 the links in π1 but not π2 , and A4 the links in π2 but not π1 .
5.3. ALTERNATIVE ASSIGNMENT RULES 117

Then the ratio h1 /h2 can be written

 P 
π1
h1 Krs exp − (i,j)∈A δij βij
=  P  (5.39)
h2 π2
Krs exp − (i,j)∈A δij βij
 P π1 P π1
− (i,j)∈A1 δij βij − (i,j)∈A2 δij

βij
Krs exp P π1 P π1
− (i,j)∈A3 δij βij − (i,j)∈A4 δij βij
=  P π2 P π2 (5.40)
− (i,j)∈A1 δij βij − (i,j)∈A2 δij βij

Krs exp P π2 P π2
− (i,j)∈A3 δij βij − (i,j)∈A4 δij βij
 P 
π1 P π1
Krs exp − (i,j)∈A2 δij βij − (i,j)∈A3 δij βij
=  P  (5.41)
π2 P π2
Krs exp − (i,j)∈A2 δij βij − (i,j)∈A4 δij βij
 P   P 
π1 π1
Krs exp − (i,j)∈A2 δij βij exp − (i,j)∈A3 δij βij
=  P   P  (5.42)
π2 π2
Krs exp − (i,j)∈A2 δij βij exp − (i,j)∈A4 δij βij
 P 
π1
exp − (i,j)∈A3 δij βij
=  P  (5.43)
π1
exp − (i,j)∈A4 δij βij

where the steps of the derivation respectively involve substituting (5.37), ex-
panding A into A1 ∪ A2 ∪ A3 ∪ A4 , using the definitions of the Ai sets to identify
π
sums where δij = 0, splitting exponential terms, and canceling common factors.
Thus in the end we have
 P 
π1
h1 exp − (i,j)∈A3 δij βij
=  P  (5.44)
h2 exp − (i,j)∈A4 δij π1
βij

regardless of the OD pair h1 and h2 connect, and this ratio only depends on A3
and A4 , that is, the pairs of alternate segments distinguishing π1 and π2 .

5.3 Alternative Assignment Rules

The previous section described the most important assignment rule, the prin-
ciple of user equilibrium. However, it is not the only possible assignment rule,
and it is not hard to see that other factors may enter route choice beyond travel
time, or that other assumptions of the user equilibrium model are not perfectly
realistic. Over time researchers have introduced a number of alternative assign-
ment rules that can be used. This text by necessity can only cover a few of
these alternative assignment rules, and three representative rules were chosen:
assignment with perception errors, system optimal assignment, and bounded
rationality.
118 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

Assignment with perception errors relaxes the assumption that drivers have
perfect knowledge of travel times. The interpretation is that every driver per-
ceives a particular travel time on each path (which may or may not equal its
actual travel time), and chooses a path that they believe to be shortest based on
these perceptions. The net effect is that travelers are distributed across paths
with different travel times, but are concentrated more on paths with lower travel
times. (If a path is very much longer than the shortest one, it is unlikely that
someone’s perception would be so far off as to mistakenly think it is shortest.)
The mathematical tool most commonly used to model perception errors is
the theory of discrete choice from economics. The resulting model is termed
stochastic user equilibrium, and this important model is the focus of Section 8.3.
The other two alternative assignment rules are discussed next.

5.3.1 System optimal assignment

Imagine for a moment that route choice was taken out of the hands of individual
travelers, and instead placed in the hands of a dictator who could assign each
traveler a path that they would be required to follow. Suppose also that this
dictator was benevolent, and wanted to act in a way to minimize average travel
delay in the network. The resulting assignment rule results in the system optimal
state.
While this scenario is a bit fanciful, the system optimal state is important
for several reasons. First, it provides a theoretical lower bound on the delay
in a network, a benchmark for the best performance that could conceivably be
achieved that more realistic assignment rules can be compared to. Second, there
are ways to actually achieve the system optimal state even without taking route
choice out of the hands of travelers, if one has the ability to freely charge tolls
or provide incentives on network links. Third, there are some network problems
where a single agent can exert control over the routes chosen by travelers, as
in certain logistics problems where a dispatcher can assign specific routes to
vehicles.
At first, it may not be obvious that this assignment rule is meaningfully
different from the user equilibrium assignment rule. After all, in user equilibrium
each traveler is individually choosing routes, while in system optimum a single
agent is choosing routes for everyone, but both are doing so to minimize travel
times. To see why they might be different, consider the network used for the
Knight-Pigou-Downs paradox in Section 1.5: a two-link network (Figure 4.6)
with a constant travel time of 50 minutes on the top link, a link performance
function 45 + x↓ on the bottom link, and a total demand of 30 vehicles. The
user equilibrium solution was x↑ = 25, x↓ = 5, with equal travel times of 50
minutes on both top and bottom routes.
So, in the user equilibrium state the average travel time is 50 minutes for all
5.3. ALTERNATIVE ASSIGNMENT RULES 119

vehicles. Is it possible to do better? We can write the average travel time as

1 1
50x↑ + (45 + x↓ )x↓ = 50(30 − x↓ ) + (45 + x↓ )x↓



50 50
1
(x↓ )2 − 5x↓ + 1500 . (5.45)


=
50
This is a quadratic function which obtains its minimum at x↓ = 2.5. At the
solution x↑ = 27.5, x↓ = 2.5, the travel times are unequal (t↑ = 50, t↓ = 47.5),
but the average travel time of 49.8 minutes is slightly less than the average
travel time of 50 minutes at the user equilibrium solution.
Therefore, travelers individually choosing routes to minimize their own travel
times may create more delay than would be obtained with central control. This
reveals an important, but subtle point about user equilibrium assignment, a
point important enough that Section 5.4 is devoted entirely to explaining this
issue. So, we will not belabor the point here, but it will be instructive to start
thinking about why the user equilibrium and system optimal states need not
coincide.
To be more precise mathematically, the system optimal assignment rule
chooses the feasible assignment minimizing the average travel time. Since the
total number of travelers is a constant, we can just as well minimize the total
system travel time, defined as
X X
T ST T = hπ cπ = xij tij (5.46)
π∈Π (i,j)∈A

where it is not hard to show that calculating T ST T through path flows and link
flows produces the same value.

5.3.2 Perception errors

An alternative assignment rule tries to relax the assumption that travelers
choose the path with exactly the shortest travel time. After all, this assumption
implicitly requires drivers to have perfect knowledge of the travel times on all
routes in the network. In reality, we know this is not true: do you know the
travel times on literally all routes between an origin and destination? And can
you accurately distinguish between a route with a travel time of 16 minutes, and
one with a travel time of 15 minutes and 59 seconds? And perhaps travelers
take other factors into account anyway, travel time may be just one of several
criteria used to choose routes.
Both of these factors can be modeled using discrete choice concepts. In
discrete choice models, an individual chooses one option from a set of alter-
natives to maximize his or her utility, consisting of both an observed portion
and an unobserved portion. Rather than requiring that each driver follow the
true shortest path between each origin and destination, we assume that drivers
follow the path they believe to be shortest, but allow for some perception error
between their belief and the actual travel times. An alternative, mathematically
120 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

equivalent, interpretation (explained below) is that drivers do in fact perceive

travel times accurately, but care about factors other than travel time. This
leads to another important traffic assignment model commonly called stochastic
user equilibrium (SUE). We discuss this model briefly here, and at length in
Section 8.3.
Consider a traveler leaving origin r for destination s. They must choose
one of the paths π connecting r to s, that is, they must make a choice from
the set Πrs . The most straightforward way to generalize the principle of user
equilibrium to account for perception errors is to set the observed utility equal
to the negative of path travel time, so

U π = −cπ + π (5.47)

with the negative sign indicating that maximizing utility for drivers means min-
imizing travel time. Assuming that the π are independent, identically dis-
tributed Gumbel random variables, we can use the logit formula (8.37) to ex-
press the probability that path π is chosen:
exp(−θcπ )
pπ = P (5.48)
π 0 ∈Πrs exp(θCπ )
0

The comments in the previous section apply to the interpretation of this formula.
As θ approaches 0, drivers’ perception errors are large relative to the path travel
times, and each path is chosen with nearly equal probability. (The errors are
so large, the choice is essentially random.) As θ grows large, perception errors
are small relative to path travel times, and the path with lowest travel time is
chosen with higher and higher probability. At any level of θ, there is a strictly
positive probability that each path will be taken.
For concreteness, the route choice discussion so far corresponds to the inter-
pretation where the unobserved utility represents perception errors in the utility.
The other interpretation would mean that π represents factors other than travel
time which affect route choice (such as comfort, quality of scenery, etc.). Either
of these interpretations is mathematically consistent with the discussion here.
The fact that the denominator of (8.39) includes a summation over all paths
connecting r to s is problematic, both theoretically and practically. From a
theoretical standpoint, it implies that drivers are considering literally every
path between the origin and destination, even when this path is very circuitous
or illogical (driving all around town when going to a store a half mile away).
Presumably no driver would ever choose such a path, no matter how ill-informed
they are about travel conditions or how poorly they can estimate travel times —
yet the logit formula (8.39) suggests that some travelers will indeed choose such
paths. From a practical standpoint, evaluating the formula (8.39) first requires
enumerating all of these paths. Since the number of paths grows exponentially
with the network size, any approach which requires an explicit listing of all the
paths in a network will not scale to realistic-sized problems.
To address this fact, we can restrict the choice set somewhat. Rather than
using all paths Πrs connecting r to s, we can restrict path choices to a subset
5.3. ALTERNATIVE ASSIGNMENT RULES 121

of reasonable paths (denoted Π̂rs ). There are different ways to identify sets of
reasonable paths, but they should address both the theoretical and practical
difficulties in the previous paragraph. That is, the paths in Π̂rs correspond
to a plausible-sounding behavioral principle (why might travelers only consider
paths in this set) and lead to a formula which is efficiently computable even in
large networks. There may be some tension between these ideas, in that there
may be very efficient formulas which do not correspond to realistic choices, or
that the most realistic models of reasonable paths may not lead to an efficient
formula. Section 8.3 discusses these in more detail.
So, equation (8.39) leads us an expression for path flows in terms of travel
times, which is the assignment rule filling the place of the question mark in
Figure 4.2. To be explicit, the formula for path flows as a function of path
travel times is
exp(−θcπ )
hπ = drs P (5.49)
π 0 ∈Πrs exp(θCπ )
0

where (r, s) is the OD pair corresponding to path π. The complete traffic as-
signment problem with this assignment rule can then be expressed as follows:
find a feasible path flow vector h∗ such that h∗ = H(C(h∗ )). This is a stan-
dard fixed-point problem. Clearly H and c are continuous functions if the link
performance functions are continuous, and the feasible path set is compact and
convex, so Brouwer’s theorem immediately gives existence of a solution to the
SUE problem.
Notice that this was much easier than showing existence of an equilibrium
solution to the original traffic assignment problem! For that problem, there was
no equivalent of (8.73). Travelers were all using shortest paths, but if there were
two or more shortest paths there was no rule for how those ties should be broken.
As a result, we had to reformulate the problem as a variational inequality and
introduce an auxiliary function based on movement of a point under a force.
With this assignment rule, there is no need for such machinations, and we can
write down the fixed point problem immediately.
That said, it is also possible to formulate the assignment problem with this
rule as the solution to a variational inequality, and as the solution to a convex
minimization problem. However, the objective function is rather complicated
and computationally expensive to evaluate, and thus it is less useful for identi-
fying an equilibrium point. The objective function is
X   X X Z xij
rs π
z(x) = − d E minrs U + xij tij − tij (x) dx (5.50)
π∈Π 0
(r,s)∈Z 2 (i,j)∈A (i,j)∈A

where the expectation is taken with respect to the “unobserved” random vari-
ables , and U π and tij in the first two terms are understood to be functions of
xij . It can be shown that this function is convex, and therefore that the solu-
tion is unique. Unfortunately, this function is less useful for actually providing
a solution. The first term in (8.112) involves an expectation over all paths con-
necting an OD pair, and as a result requires enumerating all paths, which is
impractical in large networks.
122 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

30 Flow (top)
Flow (bottom)
25

20

Flow
15

10

0
0 0.5 1 1.5 2 2.5
Theta

Figure 5.8: Flows with different perception errors.

60
Time (top)
58 Time (bottom)
Time (average)
56

54
Flow

52

50

48

46

0 0.5 1 1.5 2 2.5

Theta

Figure 5.9: Travel times with different perception errors.

To demonstrate what this assignment rule looks like, refer again to the net-
work in Figure 5.3. Figures 5.8 and 5.9 show how the flows and travel times
in the network vary with θ. A few observations worth noting: first, the travel
times on the two links are not equal, due to the presence of perception er-
rors. However, as drivers’ perceptions become more accurate (θ increasing), the
travel times on the two paths become closer, and the solution asymptotically
approaches the user equilibrium solution.

5.3.3 Bounded rationality (*)

(This optional section discusses an assignment rule which relaxes the assumption
that all travelers are on the shortest path.)
An alternative assignment rule is based on another variation of shortest path-
seeking behavior. Under bounded rationality, drivers will choose paths that are
within some threshold of the true shortest path travel time. The behavioral
rationale is that a driver will switch paths from a longer path to a shorter one,
but only if the new path is sufficiently faster than their current choice. Perhaps
5.3. ALTERNATIVE ASSIGNMENT RULES 123

a driver is unlikely to switch paths to save, say, a tenth of a second in travel

time: the burden of learning a new path outweighs any travel time savings which
are possible.
There is some evidence from behavioral economics that humans often make
choices in a boundedly rational way. Specifically in network assignment, bounded
rationality can explain observed reactions to changes in a network in a way that
the principle of user equilibrium cannot. Of course, these advantages must be
balanced against other concerns: there may be many distinct bounded rational
assignments, and efficient methods for finding bounded rational assignments in
large networks have not yet been developed.
The boundedly rational user equilibrium (BRUE) problem can be stated as
follows. Let  represent some tolerance value in terms of travel time, and let κrs
be the travel time on the shortest path connecting origin r to destination s.
Definition 5.1. (Principle of boundedly rational user equilibrium [BRUE].)
Every used path has travel time no more than  in excess of the travel time on
the shortest path connecting these nodes. That is, the feasible assignment h ∈ H
is a BRUE if, for each (r, s) ∈ Z 2 and π ∈ Πrs , hπ > 0 implies cπ ≤ κrs + .
As an example, consider once again the network in Figure 5.3, but now
assume we are seeking a BRUE where the tolerance is  = 2 minutes. That is,
travelers are willing to use any path as long as it is within one minute of being
the shortest path. The user equilibrium solution x↑ = 25, x↓ = 5 is clearly a
BRUE as well. However, there are others: if x↑ = 26, x↓ = 4, then the travel
times on the top and bottom paths are 50 and 49 minutes, respectively. Even
though the majority of travelers (those on the top link) are not on the shortest
path, the travel time on their path is close enough to the shortest path that the
situation is acceptable, and this solution is BRUE as well. However, the flows
x↑ = 28, x↓ = 2 are not BRUE: the travel times are 50 and 47 minutes, and the
difference between the (used) top path and the shortest path (3 minutes) exceeds
the tolerance value (2 minutes). You should convince yourself that x↑ ∈ [23, 27]
and x↓ = 30 − x↑ characterize all of the BRUE solutions in this simple network.
To come up with a mathematical formulation of BRUE assignment, it is
helpful to introduce an auxiliary variable ρπ indicating the amount by which
the travel time on path π falls short of the maximum acceptable travel time on
a path. Specifically, define
+
ρπ = [κrs +  − cπ ] ∀(r, s) ∈ Z 2 , π ∈ Πrs (5.51)

So, in the example of the previous paragraph, if  = 1, x↑ = 26 and x↓ = 4, then

t↑ = 50, t↓ = 49, and thus ρ↑ = [49 + 2 − 50]+ = 1 and ρ↓ = [49 + 2 − 49]+ = 2.
If the path π is unacceptable (cπ > κrs + ), then ρπ = 0 by (5.51). We can
then define the effective travel time ĉπ as

ĉπ = cπ + ρπ . (5.52)

These new quantities are helpful because the principle of BRUE can now be
formulated in a more familiar way:
124 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

Proposition 5.4. A feasible assignment h ∈ H is a BRUE if and only if every

used path has equal and minimal effective travel time, that is, for each (r, s) ∈ Z 2
0
and π ∈ Πrs , hπ > 0 implies ĉπ = minπ0 ∈Πrs ĉπ .
Proof. Let h ∈ H satisfy the principle of BRUE, and consider any OD pair
(r, s) and path π ∈ Πrs with hπ > 0. Since h is BRUE, cπ ≤ κrs + , so
ρπ = κrs +  − cπ and ĉπ = cπ + κrs +  − cπ = κrs + . But for any π 0 ∈ Πrs ,
0 0 0 0 0
ĉπ = cπ + ρπ 0 ≥ cπ + κrs +  − cπ = κrs + , so ĉπ = minπ0 ∈Πrs ĉpi .
0
Contrariwise, let h ∈ H be such that hπ > 0 implies ĉπ = minπ0 ∈Πrs ĉπ ,
and consider any OD pair (r, s). Let π̂ be a shortest path connecting r to s,
so ĉπ̂ = κrs + [κrs +  − κrs ]+ = κrs +  and this must be minimal among all
effective travel times on paths connecting r to s. For any path π ∈ Πrs with
hπ > 0, we thus have ĉπ = κrs + , so cπ = ĉπ − ρπ = κrs +  − ρπ ≤ κrs + 
since ρπ ≥ 0. Thus h satisfies the principle of BRUE as well.

Proposition 5.4 may give you hope that we can write a convex optimization
problem whose solutions correspond to BRUE. Unfortunately, this is not the
case, for reasons that will be shown shortly. However, we can write a variational
inequality where both the path flows h and the auxiliary variables ρ are decision
variables. In what follows, let κrs (h) = minπ∈Πrs cπ (h) and, with slight abuse
of notation, let κπ refer to the κrs value corresponding to the path π, and κ be
the vector of κπ values.
Proposition 5.5. A vector of path flows ĥ ∈ H is a BRUE if there exists a
|Π|
vector of nonnegative auxiliary costs ρ̂ ∈ R+ such that
   
ĉ(ĥ, ρ̂) ĥ − h
· ≤0 (5.53)
c(ĥ) + ρ̂ − κ(ĥ) −  ρ̂ − ρ
|Π|
for all h ∈ H, ρ ∈ R+ .

Proof. Assume that (ĥ, ρ̂) solve the variational inequality (5.53). The first
component of the VI ĉ(ĥ, ρ̂) · (ĥ − h) ≤ 0 shows that all used paths have equal
and minimal effective cost, and by Proposition 5.4 satisfy the principle of BRUE,
assuming that ρ is consistent with (5.51). To show this, consider the second
component of the VI: (κ(ĥ) +  − ρ̂) · ρ̂ − ρ ≤ 0, or equivalently
X X
(cπ (ĥ) + ρ̂π − κπ (ĥ) − )ρ̂π ≤ (cπ (ĥ) + ρ̂π − κπ (ĥ) − )ρπ (5.54)
π∈Π π∈Π

|Π|
for all ρ ∈ R+ . If ρ̂π > 0 for any path, inequality (5.54) can be true only if
ρπ = κπ (ĥ) +  − cπ (ĥ); or, if ρ̂π = 0, then inequality (5.54) can be true only if
ρπ = 0 ≤ κπ (ĥ) +  − cπ (ĥ). In either case (5.51) is satisfied.
|Π|
In the reverse direction, assume that ĥ is BRUE. Choose ρ̂ ∈ R+ according
to (5.51). By Proposition 5.4 surely ĉ(ĥ, ρ̂) · (ĥ −Ph) ≤ 0 for all h ∈ H. For
any path where ρπ = κπ (ĥ) +  − cπ (ĥ), clearly π∈Π (cπ (ĥ) + ρ̂π − κπ (ĥ) −
5.4. USER EQUILIBRIUM AND SYSTEM OPTIMAL 125

π π π π
P π − πρ ) = 0; andπ for any path πwhere ρ = 0, κ (ĥ) +  − c (ĥ) ≥ 0 and
)(ρ̂
π∈Π (c (ĥ)+ ρ̂π −κ (ĥ)−)(ρ̂π −ρ ) ≥ 0. In either case the second component
of the VI is satisfied as well.

This BRUE formulation has been helpful in explaining observed changes in

network flows after a disruption. For instance, assume that a link is removed
from the network due to a disaster of some sort, and that flows adjust towards
a new equilibrium in the network without the affected link. When the link is
restored, flows will adjust again. If the principle of user equilibrium is true, the
flows will move back to exactly the same values as before. However, in practice
there has been some “stickiness” observed, and not all drivers will return to the
same routes they were initially on. The BRUE framework provides a logical
explanation for this: when the network is disrupted, certain drivers were forced
to choose new paths. When the network is restored, they will only switch back
to their original paths if the travel time savings are sufficiently large. Otherwise,
they will remain on their new paths.
However, an implication of this finding is that the BRUE solution need not
be unique, even in link flows. (This was also seen in the small example at the
start of this subsection.) In general, the set of BRUE solutions is not convex
either — which immediately shows that there is no convex optimization problem
whose solutions correspond to BRUE assignments, since the sets of minima to
convex optimization problems are always convex. However, there is always at
least one BRUE solution by Proposition 5.1, since any path flows satisfying the
principle of user equilibrium also satisfy BRUE.
Still, the lack of uniqueness means that BRUE should be applied carefully.
While it is certainly more realistic than the principle of user equilibrium (since
user equilibrium can always be obtained as a special case with  = 0), it is
considerably more difficult to use BRUE to rank projects (which BRUE as-
signment should be used?) and at present it is not clear what value of  best
represents travel choices. Whether the added realism offsets these disadvantages
is application and network dependent.

5.4 User Equilibrium and System Optimal

The relationship between the user equilibrium and system optimal assignment
rules is worth studying further, for a few reasons. First, the fact that these
rules lead to distinct solutions means that travelers cannot be relied upon to
independently behave in a way that minimizes congestion5 , even though each
driver is minimizing his or her own travel time. This means that there is a
role for transportation professionals to reduce congestion through planning and
operational control. Exploring why this happens leads to a discussion of eco-
nomic externalities, which are illustrated in the Braess paradox discussed in
Section 4.3.
5 Perhaps not a very surprising observation to those used to driving in crowded cities.
126 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

Second, the mathematical structure of the system optimal assignment prob-

lem is actually very similar to that of the user equilibrium problem. In fact, if
one has a “black box” procedure for solving user equilibrium assignment, one
can just as easily solve system optimal assignment using that same procedure,
making some easy modifications to the network. Likewise, if one has a proce-
dure for solving system optimal assignment, one can just as easily solve user
equilibrium by modifying the network. So, when we discuss algorithms for solv-
ing the user equilibrium traffic assignment problem in Chapter 6, they all apply
equally well to the system optimal problem. This relationship is explored in
Section 5.4.2.
Finally, there is a more recent and interesting set of results which attempts
to quantify how far apart the user equilibrium and system optimal assignments
could possibly be, in terms of total system travel time. At first glance it seems
that these assignment rules would tend to give relatively similar total travel
times, since both are based on travel time minimization (one individual and the
other collective). However, the Braess paradox shows that these can diverge in
surprising ways. Nevertheless, in many cases one can in fact bound how much
inefficiency is caused by allowing individual drivers to pick their own routes, a
set of results called the price of anarchy and discussed in Section 5.4.3.

5.4.1 Externalities
Recall the three examples in Section 4.3, in which equilibrium traffic assignment
exhibited counterintuitive results. Perhaps the most striking of these was the
Braess paradox, in which adding a new roadway link actually worsened travel
times for all travelers. This is counterintuitive, because if the travelers simply
agreed to stay off of the newly-built road, their travel times would have been
no worse than before. This section describes exactly why this happens. But
before delving into the roots of the Braess paradox, let’s discuss some of the
implications:

User equilibrium does not minimize congestion. In principle, everyone could

have stayed on their original paths and retained the lower travel time of
83; the only catch was that this was no longer an equilibrium solution
with the new link, and the equilibrium solution with the link is worse.
Therefore, there may be traffic assignments with lower travel times (for
everybody!) than the user equilibrium one.

The “invisible hand” may not work in traffic networks. In economics, Adam
Smith’s celebrated “invisible hand” suggests that each individual acting
in his or her own self-interest also tends to maximize the interests of so-
ciety as well. The Braess paradox shows that this is not necessarily true
in transportation networks. The two drivers switched paths out of self-
interest, to reduce their own travel times; but the end effect made things
worse for everyone, including the drivers who switched paths.
5.4. USER EQUILIBRIUM AND SYSTEM OPTIMAL 127

There may be room to “improve” route choices. A corollary of the pre-

vious point, when the invisible hand does not function well, a case can be
made for regulation or another coordination mechanism to make everyone
better off. In this case, this mechanism might take the form of certain
transportation policies. What might they be?

If you have studied economics, you might already have some idea of why the
Braess paradox occurs. The “invisible hand” requires certain things to be true
if individual self-interest is to align with societal interest, and fails otherwise.
A common case where the invisible hand fails is in the case of externalities: an
externality is a cost or benefit resulting from a decision made by one person, and
felt by another person who had no say in the first person’s decision. Examples
of externalities include industrial pollution (we cannot rely on industries to self-
regulate pollution, because all citizens near a factory suffer the effects of poor air
quality, even though they had no say in the matter), driving without a seatbelt
(if you get into an accident without a seatbelt, it is likely to be more serious,
costing others who help pay your health care costs and delaying traffic longer
while the ambulance arrives; those who help pay your costs and sit in traffic had
no control over your decision to not wear a seatbelt), and education (educated
people tend to commit less crime, which benefits all of society, even those who
do not receive an education). The first two examples involve costs, and so are
called negative externalities; the latter involves a benefit, and is called a positive
externality.
Another example of an externality is seen in the prisoner’s dilemma (the
Erica-Fred game of Section 1.3). When Erica chooses to testify against Fred,
she does so because it reduces her jail time by one year. The fact that her choice
also increases Fred’s jail time by fourteen years was irrelevant (such is the nature
of greedy decision-making). When both Erica and Fred behaved in this way, the
net effect was to dramatically increase the jail time they experienced, even
though each of them made the choices they did in an attempt to minimize their
jail time.
The relevance of this economics tangent is that congestion can be thought
of as an externality. Let’s say I’m driving during the peak period, and I can
choose an uncongested back road which is out of my way, or a congested freeway
which is a direct route (and thus faster, even with the congestion). If I choose
the freeway, I end up delaying everyone with the bad luck of being behind me.
Perhaps not by much; maybe my presence increases each person’s commute by a
few seconds. However, multiply these few seconds by a large number of drivers
sitting in traffic, and we find that my presence has cost society as a whole a
substantial amount of wasted time. This is an externality, because those other
drivers had no say over my decision to save a minute or two to take the freeway.
(This is why the user equilibrium assumption is said to be “greedy.” It assumes
a driver will happily switch routes to save a minute even if it increases the total
time people spend waiting in traffic by an hour.) These concepts are made more
precise in the next subsection.
As a society, we tend to adopt regulation or other mechanisms for minimizing
128 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

the impact of negative externalities, such as emissions regulation or seatbelt

laws (although you can probably think of many negative externalities which
still exist). Transfer payments are another option: if a factory had to directly
compensate each citizen for the harm they suffer as a result of pollution, the
externality no longer exists because the factory now has to account for the costs
it imposed on society, and will only continue to operate if its profits exceed the
cost of the pollution it causes. What types of regulation or transfer payments
might exist in the transportation world?

5.4.2 Mathematical equivalence

Recall that the user equilibrium traffic assignment was the solution to the fol-
lowing convex optimization problem:
X Z xij
min tij (x)dx (5.55)
x,h 0
(i,j)∈A
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (5.56)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (5.57)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (5.58)
while the system optimal traffic assignment solves this convex optimization prob-
lem:
X
min tij (xij )xij (5.59)
x,h
(i,j)∈A
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (5.60)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (5.61)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (5.62)
Inspecting these, the only difference is in their objective functions, and even
these are quite similar: both are sums of terms involving each link in the net-
work. The problems share essentially the same structure, the only difference
is the term corresponding
R x to each link. For the user equilibrium problem, the
term for link (i, j) is 0 ij tij (x) dx, while for the system optimum problem it is
tij (xij )xij .
Suppose for a moment that we had some algorithm which can solve the user
equilibrium traffic assignment problem for any input network and OD matrix.
(A number of these algorithms will be discussed in Chapter 6.) Now suppose
that we had to solve the system optimal problem. If we were to replace the link
performance functions tij (xij ) with modified functions t̂ij (xij ) defined by
t̂ij (xij ) = tij (xij ) + t0ij (xij )xij (5.63)
5.4. USER EQUILIBRIUM AND SYSTEM OPTIMAL 129

where t0ij is the derivative of the link performance functions, then

Z xij Z xij
tˆij (x) dx = (tij (x) + t0ij (x)x) dx (5.64)
0 0

Integrating the second term by parts, this simplifies to

Z xij
t̂ij (x) dx = tij (xij )xij (5.65)
0

which is the same term for link (i, j) used in the system optimal function.
That is, solving user equilibrium with the modified functions t̂ij produces the
same link and path flow solution as solving system optimum with the original
performance functions tij . The interpretation of the formula (5.63) is closely
linked with the concept of externalities introduced in the previous subsection.
Equation (5.63) consists of two parts: the actual travel time tij (xij ), and an
additional term t0ij (xij )xij . If an additional vehicle were to travel on link (i, j),
it would experience a travel time of tij (xij ). At the margin, the travel time on
link (i, j) would increase by approximately t0ij (xij ), and this marginal increase
in the travel time would be felt by the other xij vehicles on the link. So, the
second term in (5.63) expresses the additional, external increase in travel time
caused by a vehicle using link (i, j). For this reason, the modified functions t̂ij
can be said to represent the marginal cost on a link.
Since the system optimal solution is a “user equilibrium” with respect to the
modified costs t̂ij , we can formulate a “principle of system optimum” similar to
the principle of user equilibrium:
Definition 5.2. (Principle of system optimum.) Every used route connecting
an origin and destination has equal and minimal marginal cost.
So, system optimum can be seen as a special case of user equilibrium, with
modified link performance functions. The converse is true as well. Suppose we
had an algorithm which would solve the system optimal problem for any input
network and OD matrix. If we replace the link performance functions tij (xij )
with modified functions t̃ij (xij ) defined by
R xij
tij (x) dx
t̃ij (xij ) = 0 (5.66)
xij

when xij > 0 and t̃ij (xij ) = 0 otherwise, a similar argument shows that the ob-
jective function for the system optimal problem with link performance functions
t̃ij is identical to that for the user equilibrium problem with link performance
functions tij .
As a result, despite very different physical interpretations, the user equilib-
rium and system optimal problems are essentially the same mathematically. If
you can solve one, you can solve the other by making some simple changes to the
link performance functions. So, there is no need to develop separate algorithms
for the user equilibrium and system optimal problems.
130 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

5.4.3 Price of anarchy

The discussion of the user equilibrium and system optimal problems so far has
gone in several directions. On the one hand, it seems like the problems are fairly
similar: both involve routes chosen according to some travel time minimization
rule, and they are mathematically equivalent. On the other hand, the real-world
interpretations are very different (individual control vs. centralized control),
and the Braess paradox shows that user equilibrium can behave in a far more
counterintuitive manner than system optimal. This subsection concludes the
discussion by showing that, while the user equilibrium solution may be different
than the system optimal solution, and worse in terms of higher total system
travel time, in many cases there is a limit to how much worse it can be.
In particular, if the link performance functions are linear (that is, of the
form tij (xij ) = aij + bij xij ), the ratio between the total system travel time at
user equilibrium, and the total system travel time at system optimum, can be
no greater than 43 . This is often called the price of anarchy, since it represents
the amount of additional delay which can potentially be caused by allowing
individual drivers to choose their own routes, compared to a centrally-controlled
solution. This bound holds no matter what the network structure or the level of
demand, as long as the link performance functions are affine. This is an elegant
and perhaps surprising result, which is actually not difficult to show using the
variational inequality formulation of equilibrium.
Theorem 5.5. Given a network with link performance functions of the form
tij (xij ) = aij + bij xij with aij ≥ 0 and bij ≥ 0 for all links (i, j), let feasible
link flows x̂ ∈ X satisfy the principle of user equilibrium and ŷ ∈ X the system
optimum principle. Denoting the total system travel times corresponding to these
solutions as T ST T (x̂) and T ST T (ŷ), we have

T ST T (x̂) 4
≤ (5.67)
T ST T (ŷ) 3

Proof. Since x̂ satisfies the principle of user equilibrium, we have

t(x̂) · (x̂ − ŷ) ≤ 0 (5.68)

or, equivalently,

t(x̂) · x̂ ≤ t(x̂) · ŷ (5.69)

= (t(x̂) − t(ŷ)) · ŷ + t(ŷ) · ŷ (5.70)

which is the same as

T ST T (x̂) ≤ (t(x̂) − t(ŷ)) · ŷ + T ST T (ŷ) . (5.71)

If we can show that

1
(t(x̂) − t(ŷ)) · ŷ ≤ T ST T (x̂) (5.72)
4
5.4. USER EQUILIBRIUM AND SYSTEM OPTIMAL 131

1
1
1 2

Figure 5.10: The price of anarchy bound is tight.

then we have established the result, by substitution into (5.71). We do this by

showing an even stronger result, namely
1
(tij (x̂ij ) − tij (ŷij ))ŷij ≤ tij (x̂ij )x̂ij (5.73)
4
for all links (i, j) which is clearly sufficient.
If x̂ij ≤ ŷij , then the left-hand side of (5.73) is nonpositive (the link per-
formance functions are nondecreasing since bij ≥ 0), while the right-hand side
is nonnegative (since aij ≥ 0), so the inequality is certainly true. On the other
hand, if x̂ij > ŷij , then

(tij (x̂ij ) − tij (ŷij ))ŷij = bij (x̂ij − ŷij )ŷij . (5.74)

Now, the function bij (x̂ij − z)z is a concave quadratic function in z, which
obtains its maximum when z = x̂ij /2. Therefore

bij (x̂ij )2 1
(tij (x̂ij ) − tij (ŷij ))ŷij ≤ ≤ (aij + bij x̂ij )x̂ij = tij (x̂ij )x̂ij ) , (5.75)
4 4
proving the result.
Furthermore, this bound is tight. Consider the more extreme version of the
Knight-Pigou-Downs network shown in Figure 5.10 where the demand is 1 unit,
the travel time on the upper link is 1 minute, and the travel time on the bottom
link is t↓ = x↓ . The user equilibrium solution is x↑ = 0, x↓ = 1, when both
links have equal travel times of 1 minute, and the total system travel time is
1. You can verify that the system optimal solution is x↑ = x↓ = 21 , when the
total system travel time is 34 . Thus the ratio between the user equilibrium and
system optimal total system travel times is 34 .
You may be wondering if a price of anarchy can be found when we relax the
assumption that the link performance functions are affine. In many cases, yes;
for instance, if√ the link performance functions are quadratic, √
then the price of
3 3 434
anarchy is 3√ 3−2
, if cubic, then the price of anarchy is √
3
4 4−3
, and so on. In
all of these cases the modified Knight-Pigou-Downs network can show that this
bound is tight. On the other hand, if the link performance functions have a
132 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

vertical asymptote (e.g., tij = (uij − xij )−1 ), then the ratio between the total
system travel times at user equilibrium and system optimum may be made
arbitrarily large.

5.5 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
The variational inequality formulation of the traffic assignment problem was
presented by Smith (1979a), Dafermos (1980), and Smith (1983). The convex
optimization formulation of the traffic assignment problem was first presented in
Beckmann et al. (1956). The fixed-point formulation described in the optional
section relies on a more general fixed point theorem due to Kakutani (1941).
Although not described in this book, the traffic assignment problem can also
be described as a nonlinear complementarity problem (Aashtiani and Magnanti,
1981).
Rossi et al. (1989) proposed the use of entropy to distinguish a “most likely”
path flow solution among all of those satisfying the principle of user equilibrium.
Bar-Gera and Boyce (1999) and Bar-Gera (2006) discuss how the more intuitive
(but slightly weaker) proportionality condition can be derived from entropy
maximization.
The bounded rationality formulations and results presented in this chapter
are drawn from Mahmassani and Chen (1987), Lou et al. (2010), and Di et al.
(2013). The first results on the price of anarchy in traffic assignment are due to
Roughgarden (2002), and the specific proof given in this book is from Correa
et al. (2004)
non-additive, RECOURSE, and so on. Or maybe this should go in Chapter
9?

5.6 Exercises
1. [34] Consider a network with two parallel links connecting a single origin
to a single destination; the link performance function on each link is 2 + x
and the total demand is d = 10.
(a) Write the equations and inequalities defining the set of feasible path
flows H, and draw a sketch.
(b) What are the vectors −C(h) for the following path flow vectors? (1)
h = [0, 10] (2) h = [5, 5] (3) h = [10, 0] Draw these vectors at these
three points on your sketch.
(c) For each of the vectors from part (b), identify the point projH (h −
C(h)) and include these on your sketch.
2. [45] Consider a network with two parallel links connecting a single origin
to a single destination; the link performance function on the first link is
5.6. EXERCISES 133

10 + x, and the link performance function on the second link is x2 /20.

The total demand is d = 30.
 
(a) Demonstrate that the solution ĥ = 20 10 is not an equilibrium,
and then provide a vector h showing that the variational inequal-
ity (5.2) is not satisfied.
 
(b) Demonstrate that the solution ĥ = 10 20 is an equilibrium, and
prove that the variational inequality (5.2) is satisfied no matter what
h ∈ H is.
3. [43] For the Braess network of Figure 4.8(b), write down the optimality
conditions corresponding to Beckmann’s formulation (5.13)–(5.16).
4. [27] Which of the following multifunctions F are closed? Each of these
multifunctions maps [0, 1] to subsets of [0, 1]. Draw sketches of each of
these multifunctions.
(a) F (x) = {y : 0 ≤ y ≤ x}
(b) F (x) = {y : 0 < y < 1}
(c) F (x) = {0} ∪ {1 − x} ∪ {1}

(d) F (x) = x2
5. [62] Show that if f (x) is a continuous function with a compact domain,
the single-valued “multifunction” F (x) = {f (x)} is closed.
6. [23] For each of the multifunctions in Exercise 4, identify all of the fixed
points.
7. [62] Specify the multifunction R(c) for the Braess network (Figure 4.8b),
and identify all of its fixed points.
8. [42] Complete the proof of Theorem 5.3 by showing that the graph of F
is indeed closed.
9. [51] Create a simple network where one or more link performance functions
are not continuous, and where no user equilibrium solution exists. (Don’t
worry about creating “realistic” functions for this problem.)
10. [51] Create a simple network where one or more link performance functions
are not strictly increasing, and where the user equilibrium link flows are
not unique.
11. [11] Show that the BPR link delay function (4.1) does not satisfy the
condition t0ij (xij ) > 0 for all xij if β > 1.
12. [46] In spite of Exercise 11, the BPR link delay functions are strictly
increasing, and the resulting link flow equilibrium solution is still unique.
Generalize the proof of Proposition 5.2 to handle the case when the link
delay functions are differentiable and strictly increasing, by showing that
134 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

300 225 3
1 2
100 175 100 25
75 0
4 5 6
0 75
25 100 175 100
7 225 8 300 9

Figure 5.11: Network for Exercises 14 and 15. Each link has link performance
function 10 + xij

the Beckmann function is still strictly convex even if t0ij (xij ) is not always
strictly positive.
13. [51] Show that the user-equilibrium and system-optimal link flows are the
same if there is no congestion, that is, if ta (xa ) = τa for some constant τa .
14. [10] Consider the network of Figure 5.11, ignoring the labels next to the
link in the figure. The demand in this network is given by d19 = 400.
Does the following path-flow solution satisfy the principle of user equilib-
rium? 100 vehicles on path [1,2,3,6,9], 100 vehicles on path [1,2,3,6,9], 100
vehicles on path [1,4,5,6,9], and 100 vehicles on path [1,4,5,2,3,6,9]. You
can answer this without any calculation.
15. [45] In the network in Figure 5.11, the demand is given by d13 = d19 =
d17 = d91 = d93 = d97 = 100. The flow on each link is shown in the figure.

(a) Find a path flow vector h which corresponds to these link flows.
(b) Show the resulting origin-based link flows x1 and x9 from the two
origins in the network.
(c) Show that the links used by these two origins form an acyclic sub-
network by finding topological orders for each subnetwork.
(d) Determine whether these link flows satisfy the principle of user equi-
librium.

16. [35] Consider the network in Figure 5.12, along with the given (equilib-
rium) link flows. There is only one OD pair, from node 1 to node 3.
Identify three values of path flows which are consistent with these link
flows, in addition to the most likely (entropy-maximizing) path flows.
17. [37] In the network shown in Figure 5.13, 320 vehicles travel from A to
C, 640 vehicles travel from A to D, 160 vehicles travel from B to C, and
320 vehicles travel from B to D. The equilibrium link flows are shown.
5.6. EXERCISES 135

90 120
1 30 2 3
60
60

Figure 5.12: Network for Exercise 16.

C D
480 960
720 720 240

1 2 3 4
240 720 720
960 480
A B

Figure 5.13: Network for Exercise 17.

(a) Give a path flow solution which satisfies proportionality and produces
the equilibrium link flows.
(b) At the proportional solution, what fraction of flow on the top link
connecting 3 and 4 is from origin A?
(c) Is there a path flow solution which produces the equilibrium link flows,
yet has no vehicles from origin A on the top link connecting 3 and 4?
If so, list it. If not, explain why.

18. [38] Consider the network in Figure 5.14, where 2 vehicles travel from 1
to 4 (with a value of time of $20/hr), and 4 vehicles travel from 2 to 4
(with a value of time of $8/hr). The equilibrium volumes are 3 vehicles
on Link 1, and 3 vehicles on Link 2.

(a) Assuming that the vehicles in this network are discrete and cannot be
split into fractions, identify every combination of path flows which
give the equilibrium link volumes (there should be 20). Assuming
each combination is equally likely, show that the proportional division
of flows has the highest probability of being realized.
(b) What is the average value of travel time on Link 1 at the most likely
path flows? What are the upper and lower limits on the average value
of travel time on this link?

19. [65] Derive the optimality conditions for the system-optimal assignment,
and provide an interpretation of these conditions which intuitively relates
them to the concept of system optimality.
136 CHAPTER 5. THE TRAFFIC ASSIGNMENT PROBLEM

1 Link 1

3 4

2 Link 2

Figure 5.14: Network for Exercise 18.

20. [33] Find the system optimal assignment in the Braess network (Fig-
ure 4.8b), assuming a demand of 6 vehicles from node 1 to node 4.
21. [51] Consider a network where every link performance function is linear,
of the form tij = aij xij . Show that the user equilibrium and system
optimum solutions are the same.
22. [34] Calculate the ratio of the total system travel time between the user
equilibrium and system optimal solutions in the Braess network (Fig-
ure 4.8b).

23. [46] Find the set of boundedly rational assignments in the Braess network
(Figure 4.8b).
Chapter 6

Algorithms for Traffic

Assignment

This chapter presents algorithms for solving the basic traffic assignment problem
(TAP), which was defined in Chapter 5 as the solution x̂ to the variational
inequality
t(x̂) · (x̂ − x) ≤ 0 ∀x ∈ X , (6.1)
which can also be expressed in terms of path flows ĥ as

c(ĥ) · (ĥ − h) ≤ 0 ∀h ∈ H , (6.2)

or equivalently as the solution x̂ to the optimization problem

X Z xij
min tij (x)dx (6.3)
x,h 0
(i,j)∈A
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (6.4)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (6.5)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (6.6)

While there are general algorithms for variational inequalities or nonlinear

optimization problems, TAP involves tens of thousands or even millions of vari-
ables and constraints for practical problems. So, these general algorithms are
outperformed by specialized algorithms which are designed to exploit some spe-
cific features of TAP. The most significant features to exploit are the network
structure embedded in the constraints, and the fact that the objective function
is separable by link.
This chapter presents four types of algorithms for TAP. The first three are
aimed at finding an equilibrium solution (either link flows x̂ or path flows ĥ).
Broadly speaking, these algorithms can be divided into link-based, path-based,

137
138 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

and bush-based algorithms, according to the way the solution is represented.

As discussed in Section 5.2.2, this is not trivial even though it plays a major
role in certain types of applications, such as select link analysis.
The chapter focuses entirely on the basic TAP, and not the alternative as-
signment rules from Section 5.3. System optimal assignment can be transformed
mathematically into a user equilibrium problem with modified cost functions,
so all of these algorithms can be easily adapted for the system optimal problem,
but most of these algorithms cannot be directly applied to the variations with
bounded rationality or perception errors.
Section 6.1 provides an introduction, justifying the need for efficient algo-
rithms for TAP and discussing issues such as convergence criteria which are rel-
evant to all algorithms. The next three sections respectively present link-based,
path-based, and bush-based algorithms for finding user equilibrium solutions.

6.1 Introduction to Assignment Algorithms

This introductory section touches on three topics: first, the advantages and
disadvantages of link-based, path-based, and bush-based algorithms; second,
the general framework for solving equilibrium problems; and third, the question
of convergence criteria, which is pertinent to all of these.

6.1.1 Why do different algorithms matter?

It may not be clear why we need to present so many different algorithms for the
same problem. Why not simply present the one “best” algorithm for solving
TAP? Link-based, path-based, and bush-based algorithms all exhibit advan-
tages and disadvantages relative to each other. In this introductory section,
these algorithms are compared qualitatively. As you read through the following
sections, which include details and specifications of each algorithm, keep these
concepts in mind.
Link-based algorithms only keep track of the aggregate link flows x, not
the path flows h which lead to these link flows. Since the number of links in a
large network is much less than the number of paths, link-based algorithms are
very economical in terms of computer memory. For this reason, link-based algo-
rithms were the first to be used in practice decades ago, when computer memory
was very expensive. Link-based algorithms also tend to be easier to parallelize.
With modern desktop machines containing multiple cores, this parallelization
can reduce computation time significantly. Finally, link-based algorithms tend
to be easy to code and implement (even in a spreadsheet). A significant draw-
back is that they tend to be slow, particularly when high precision is demanded.
The first few iterations of link-based algorithms make good progress, but they
stall quickly, and final convergence to the equilibrium solution can be extremely
slow.
Path-based algorithms, by contrast, keep track of the path flow vector h,
which can be used to generate the link flows x whenever needed. Although this
6.1. INTRODUCTION TO ASSIGNMENT ALGORITHMS 139

representation requires more memory than a link-based solution, it also retains

a great deal of information which is lost when one aggregates path flows to
link flows. This information can be exploited to converge much faster than link-
based algorithms, especially when a very precise solution is needed. Because the
number of paths in a network is so large, much of the effort involved in coding
path-based algorithms is associated with clever data structures and algorithm
schemes which are aimed at minimizing the number of paths which need to be
stored. This increases the complexity of the coding of these algorithms, even if
the algorithmic concepts are not difficult.
Bush-based algorithms are the most recent developed, and aim to of-
fer speed comparable to path-based algorithms, while requiring less memory
(although still significantly more than link-based algorithms). Bush-based al-
gorithms do this by selectively aggregating the path flows: as shown in Sec-
tion 5.2.3, at equilibrium the set of paths used by all travelers from the same
origin (or traveling to the same destination) forms an acyclic subgraph (a bush).
Bush-based algorithms aim to identify these bushes for each origin and desti-
nation, exploiting the fact that calculations in acyclic networks are very fast.
Bush-based algorithms also tend to return higher-entropy solutions than path-
based algorithms, which is important when interpreting the path flow solution.
The downside to these algorithms is that their design involves more complex-
ity, and the success of these algorithms depends highly on implementation and
programming skill.

6.1.2 Framework
It turns out that none of these solution methods get to the right answer im-
mediately, or even after a finite number of steps. There is no “step one, step
two, step three, and then we’re done” recipe for solving large-scale equilibrium
problems. Instead, an iterative approach is used where we start with some fea-
sible assignment (link or path), and move closer and closer to the equilibrium
solution as you repeat a certain set of steps over and over, until you’re “close
enough” to quit and call it good. One iterative algorithm you probably saw in
calculus was Newton’s method for finding zeros of a function. In this method,
one repeats the same step over and over until the function is sufficiently close
to zero.
Broadly speaking, all equilibrium solution algorithms repeat the following
three steps:
1. Find the shortest (least travel time) path between each origin and each
destination.
2. Shift travelers from slower paths to faster ones.
3. Recalculate link flows and travel times after the shift, and return to step
one unless we’re close enough to equilibrium.
The shortest path computation can be done quickly and efficiently even in
large networks, as was described in Section 2.4. The third step is even more
140 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

10 + x1
50 vehicles
1 2 traveling
20 + x2 from 1 to 2

Figure 6.1: Two-link example for demonstration.

straightforward, and is nothing more than re-evaluating the link performance

functions on each link with the new volumes. The second step requires the most
care; the danger here is shifting either too few travelers onto faster paths, or
shifting too many. If we shift too few, then it will take a long time to get to the
equilibrium solution. On the other hand, systematically shifting too many can
be even more dangerous, because it creates the possibility of “infinite cycling”
and never finding the true equilibrium.
In the simple example in Figure 6.1, by inspection the equilibrium is for
thirty travelers to choose the top route, and twenty to choose the bottom route,
with an equal travel time of 40 minutes on both paths. Solving this example
using the above process, initially (i.e., with nobody on the network) the fastest
path is the top one (step one), so let’s assign all 50 travelers onto the top path
(step two). Performing the third step, we recalculate the travel times as 60
minutes on the top link, and 20 on the bottom. This is not at all an equilibrium,
so we go back to the first step, and see that the bottom path is now faster, so we
have to shift some people from the top to the bottom. If we wanted, we could
shift travelers one at a time, that is, assigning 49 to the top route and 1 to the
bottom, seeing that we still haven’t found equilibrium, so trying 48 and 2, then
47 and 3, and so forth, until finally reaching the equilibrium with 30 and 20.
Clearly this is not efficient, and is an example of shifting too few travelers at a
time.
At the other extreme, let’s say we shift everybody onto the fastest path in
the second step. That is, we go from assigning 50 to the top route and 0 to
the bottom, to assigning 0 to the top and 50 to the bottom. Recalculating link
travel times, the top route now has a travel time of 10 minutes, and the bottom
a travel time of 70. This is even worse!1 Repeating the process, we try to fix
this by shifting everybody back (50 on top, 0 on bottom), but now we’re just
back in the original situation. If we kept up this process, we’d keep bouncing
back and forth between these solutions. This is clearly worse than shifting too
few, because we never reach the equilibrium no matter how long we work! You
might think it’s obvious to detect if something like this is happening. With
this small example, it might be. Trying to train a computer to detect this, or
trying to detect cycles with over millions OD pairs (as is common in practice),
is much harder. The key step in all of the algorithms for finding equilibria is
1 By “worse” we mean farther from equilibrium.
6.1. INTRODUCTION TO ASSIGNMENT ALGORITHMS 141

determining how much flow to shift.

6.1.3 Convergence criteria

A general issue is how one chooses to stop the iterative process, that is, how
one knows when a solution is “good enough” or close enough to equilibrium.
This is called a convergence criterion. Many convergence criteria have been
proposed over the years; perhaps the most common in practice is the relative
gap, which is defined first. Unfortunately, the relative gap has been defined in
several different ways, so it is important to be familiar with all of the definitions.
Remembering that the multiplier κrs represents the time spent on the fastest
path between origin r and destination s, one definition of the relative gap γ1 is
P
(i,j)∈A tij xij t·x
γ1 = P rs rs
−1= − 1. (6.7)
(r,s)∈Z 2 κ d κ·d

The numerator of the fraction is the total system travel time (TSTT). The
relative gap is always nonnegative, and it is equal to zero if and only if the flows
x satisfy the principle of user equilibrium. It is these properties which make
the relative gap a useful convergence criterion: once it is close enough to zero,
our solution is “close enough” to equilibrium. For most practical purposes, a
relative gap of 10−4 –10−6 is small enough.
A second definition of the relative gap γ2 is based on the Beckmann function
itself. Let f denote the Beckmann function, and fˆ its value at equilibrium
(which is a global minimum). In many algorithms, given a current solution x,
it is not difficult to generate upper and lower bounds on fˆ based on the current
solution, respectively denoted f¯(x) and f (x). A trivial upper bound is its value
¯
at the current solution: f¯ = f (x), since clearly fˆ ≤ f (x) for any feasible link
assignment x. Sometimes, a corresponding lower bound can be identified as
well. Assuming that these bounds can become tighter over time, and that in
the limit both f¯ → fˆ and f → fˆ, the difference or gap f¯(x) − f (x) can be used
as a convergence criterion.¯ These values are typically normalized,
¯ leading to
one definition of the relative gap:

f¯ − f
γ2 = ¯ , (6.8)
f
¯
or a slightly modified version

f¯ − max f
γ3 = ¯ , (6.9)
max f
¯
where max f is the greatest lower bound found to date, in case the sequence of f
values is not¯ monotone over iterations. A disadvantage of these definitions of the
¯
relative gap is that different algorithms calculate these upper and lower bounds
differently. While they are suitable as termination criteria in an algorithm, it is
142 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

not possible (or at least not easy) to directly compare the relative gap calculated
by one algorithm to that produced by another to assess which of two solutions
is closer to equilibrium.
One drawback of the relative gap (in all of its forms) is that it is unitless and
does not have an intuitive meaning. Furthermore, it can be somewhat confusing
to have several slightly different definitions of the relative gap, even though they
all have the same flavor. A more recently proposed metric is the average excess
cost, defined as

κrs drs
P P
(i,j)∈A tij xij − (r,s)∈Z 2 t·x−κ·d
AEC = P = . (6.10)
(r,s)∈Z 2 drs d·1

This quantity represents the average difference between the travel time on each
traveler’s actual path, and the travel time on the shortest path available to him
or her. Unlike the relative gap, AEC has units of time, and is thus easier to
interpret.
Another convergence measure with time units is the maximum excess cost,
which relates directly to the principle of user equilibrium. The maximum excess
cost is defined as the largest amount by which a used path’s travel time exceeds
the shortest path travel time available to that traveler:
 
π rs
M EC = max 2 max {c − κ } . (6.11)
(r,s)∈Z π∈Πrs :hπ >0

This is often a few orders of magnitude higher than the average excess cost.
One disadvantage of the maximum excess cost is that it is only applicable when
the path flow solution is known. This is easy in path-based or bush-based
algorithms. However, since many path-flow solutions correspond to the same
link-flow solution (cf. Section 5.2.2), M EC is not well suited for link-based
algorithms.
Finally, this section concludes with two convergence criteria which are infe-
rior to those discussed thus far. The first is to simply use the Beckmann function
itself; when it is sufficiently close to the global optimal value fˆ, terminate. A
moment’s thought should convince you that this criterion is not practical: there
is no way to know the value of fˆ until the problem has already been solved.
(Upper and lower bounds are possible to calculate, though, as with γ2 .) A more
subtle version is to terminate when the Beckmann function stops decreasing,
or (in a more common form) to terminate the algorithm when the link or path
flows stabilize from one iteration to the next. The trouble with these conver-
gence criteria is that they cannot distinguish between a situation when the flows
stabilize because they are close to the equilibrium solution, and when they sta-
bilize because the algorithm “gets stuck” and cannot improve further due to a
flaw in its design or a bug in the programming. For this reason, it is always
preferable to base the termination criteria on the equilibrium principle itself.
6.2. LINK-BASED ALGORITHMS 143

6.2 Link-Based Algorithms

Link-based algorithms for traffic assignment are the simplest to understand and
implement, and require the least amount of computer memory. Given a current
set of link flows x, a link-based algorithm attempts to move closer to equilibrium
by performing two steps. First, a target point x∗ is identified; moving in the
direction of x∗ from x should lead towards an equilibrium solution. Then, a
step of size λ ∈ [0, 1] is taken in this direction, updating the link flows to
λx∗ + (1 − λ)x. This is a specific way of implementing the first two steps of the
framework specified in the previous section: λ represents the fraction of flow
which is shifted from the paths at the current solution x to the paths at the
target solution x∗ . If λ = 1, all of the flow has shifted to the target solution; if
λ = 0, no flow has shifted at all. Since the set of feasible link flow solutions X
is convex, as long as x and x∗ are feasible, we can be assured that the convex
combination λx∗ + (1 − λ)x is feasible as well.
Link-based algorithms differ in two primary ways: first, how the target x∗ is
chosen; and second, how the step size λ is chosen. If λ is too small, convergence
to equilibrium will be very slow, but if λ is too large, the solution may never
converge at all — the example from Figure 6.1 in the previous section is an
example of what can happen if λ = 1 for all iterations. This section presents
three link-based algorithms, in increasing order of sophistication (but in increas-
ing order of convergence speed.) The first is the method of successive averages
(MSA), which is perhaps the simplest equilibrium algorithm. The second is
the Frank-Wolfe algorithm, which can be thought of as a version of MSA with
a more intelligent choice of step size λ. (Frank-Wolfe was the most common
used in practice for several decades.) The third is the conjugate Frank-Wolfe
algorithm, which can be thought of as a version of Frank-Wolfe with a more
intelligent choice of target x∗ .
All of these algorithms use the following framework; the only difference is
how x∗ and λ are calculated.

1. Generate an initial solution x ∈ X.

2. Generate a target solution x∗ ∈ X.
3. Update the current solution: x ← λx∗ + (1 − λ)x for some λ ∈ [0, 1].
4. Calculate the new link travel times.
5. If the convergence criterion is satisfied, stop; otherwise return to step 2.

6.2.1 Method of successive averages

Although the method of successive averages (MSA) is not competitive with
other equilibrium solution algorithms, its simplicity and clarity in applying the
three-step iterative process make it an ideal starting place. To specify MSA, we
need to specify how the target solution x∗ is chosen, and how the step size λ is
chosen.
144 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

The target x∗ is an all-or-nothing assignment. That is, assuming that the

current travel times are fixed, identify the shortest path between each origin
and destination, and load all of the demand for that OD pair onto that path.
Thus, x∗ is the state which would occur if literally every driver was to switch
paths onto what is currently the shortest path. Of course, if we were to switch
everybody onto these paths, which would occur if we choose λ = 1, those paths
would almost certainly not be “shortest” anymore. But x∗ can still be thought
of as a target, or a direction in which travelers would feel pressure to move.
So, what should λ be? As discussed above, there are problems if you shift
too few travelers, and potentially even bigger problems if you shift too many.
MSA adopts a reasonable middle ground: initially, we shift a lot of travelers,
but as the algorithm progresses, we shift fewer and fewer until we settle down
on the average. The hope is that this avoids both the problems of shifting too
few (at first, we’re taking big steps, so hopefully we get somewhere close to
equilibrium quickly) and of shifting too many (eventually, we’ll only be moving
small amounts of flow so there is no worry of infinite cycling).
Specifically, on the i-th iteration, MSA uses λ = 1/(i + 1). So, the first time
through, half of the travelers are shifted to the current shortest paths. The
second time through, a third of the people shift to the current shortest paths
(and two thirds stay on their current path). On the third iteration, a fourth
of the people shift to new paths, and so on. (MSA can also be applied with
different step size rules; see Exercise 11.)
At this point, it’s worth using the Beckmann formulation to show that the
choice of an all-or-nothing assignment for x∗ has mathematical justification, in
addition to the intuitive interpretation of shifting towards shortest path. Let
x be the current set of link flows, and x∗ an all-or-nothing assignment. As a
result of a shift of size λ, the Beckmann function will be changed from f (x)
to f (x0 ), and we want to show that it’s possible to choose λ in some way to
guarantee f (x0 ) ≤ f (x). That is, we want to show that we can reduce the
Beckmann function (and thus move closer to the equilibrium solution) by taking
a (correctly-sized) step in the direction x∗ − x. Define f (x(λ)) = f ((1 − λ)x +
λx∗ ) to be the Beckmann function after taking a step of size λ. Using the
multivariate chain rule, the derivative of f (λ) is
df X ∂f dxij X
= = tij ((1 − λ)xij + λx∗ij )(x∗ij − xij ) .
dλ ∂xij dλ
(i,j)∈A (i,j)∈A

Evaluating this derivative at λ = 0 gives

df X
(0) = tij (xij )(x∗ij − xij ) .
dλ
(i,j)∈A

Now, x∗ was specifically P

chosen to put all vehiclesP on the shortest paths at
travel times t(x), and so (i,j)∈A x∗ij tij (xij ) ≤ (i,j)∈A xij tij (xij ), and there-
df
fore dλ (0) ≤ 0. Furthermore, if we are notPat the equilibrium solution al-
ready, (i,j)∈A x∗ij tij (xij ) is strictly less than (i,j)∈A xij tij (xij ). This implies
P
6.2. LINK-BASED ALGORITHMS 145

df
dλ (0)< 0 or, equivalently, we can decrease the Beckmann function if we take a
small enough step in the direction x∗ − x, by shifting people from longer paths
onto shorter ones.
Two examples of MSA are shown below. A proof of convergence is sketched
in Exercise 12.

Small network example Here we solve the small example of Figure 6.1 by
MSA, using the relative gap γ1 to measure how close we are to equilibrium.

Initialization. Find the shortest paths: with no travelers on the network, the
top link has a travel time of 10, and the bottom link has a travel
 time of
20. Therefore the top link is the shortest path,so x∗ = 50 0 . We take
this to be the initial solution x ← x∗ = 50 0 . Recalculating the travel
times, we have t1 = 10 + x1 = 60 and t2 = 20 + x2 = 20 (or, in vector
form, t = 60 20 ).

Iteration 1. With the new travel times, the shortest path is now the bottom
link, so κ = 20 and the relative gap is

t·x 50 × 60 + 0 × 20
γ1 = −1= − 1 = 2.
κ·d 20 × 50

This is far too big, so we continue with the second iteration. If everyone
were to take the new shortest path, the flows would be x∗ = 0 50 .
 

Because this is the first iteration, we shift

 1/2 of the travelers
 onto this
path, so x ← (1/2)x∗ + (1/2)x
 =  0 25 + 25 0 = 25 25 . The new
travel times are thus t = 35 45 .

Iteration 2. With the new travel times, the shortest path is now the top link,
so κ = 35 and the relative gap is

t·x 25 × 35 + 25 × 45
γ1 = −1= − 1 = 0.143 .
κ·d 35 × 50

If
 everyone
 were to take the new shortest path, the flows would be x∗ =
50 0 . Because this is the second iteration, we shift1/3of the travelers
∗

 so x ← (1/3)x + (2/3)x =
onto this path, 50/3 0 + 50/3 50/3 =
  
100/3 50/3 . The new travel times are thus t = 43.33 36.67 .

Iteration 3. With the new travel times, the shortest path is now the bottom
link, so κ = 36.67 and the relative gap is γ1 = 0.121. A bit better,
∗
0 50 , x ← (1/4)x∗ +

but still too
 big, so we
  carry on.
 Here x =

 = 0 50/4 + 25 50/4 = 25 25 . The new travel times are
(3/4)x
t = 35 45 . Note that we have returned to the same solution found in
Iteration 1. Don’t despair; this just means the last shift was too big. Next
time we’ll shift fewer vehicles (because λ is smaller with each iteration).
146 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

1 3 3 4
1
4 1 5,000 0
2
3 2 0 10,000
5 6
5 6
7
2 4

Figure 6.2: Larger example with two OD pairs. (Link numbers shown.)

Iteration 4. With the new travel times, the shortest path is now the top link,
so
 κ = 35 and the relative gap is γ1 = 0.143. The new target is x∗ =
∗

50 0 , x ←  (1/5)x + (4/5)x = 30 20 . The new travel times are
t = 40 40 . With the new travel times, the shortest path is the top
link, so κ = 40 and the relative gap is γ1 = 0, so we stop. In fact, either
path could have been chosen for the shortest path. Whenever there is a tie
between shortest paths, you are free to choose among them.

Larger network example Here we apply MSA to a slightly larger network

with two OD pairs, shown in Figure 6.2, where each link has the link performance
function t(x) = 10 + x/100.
There are four paths in this network; for OD pair (1,3) these are denoted
[1, 3] and [1, 5, 6, 3] according to their link numbers, and for OD pair (2,4) these
are [2, 5, 6, 4] and [2, 4]. In this example, we’ll calculate the average excess cost,
rather than the relative gap.

Initialization. Find the shortest paths: with no travelers on the network, paths
[1, 3], [1, 5, 6, 3], [2, 5, 6, 4], and [2, 4] respectively have travel times of 10,
30, 30, and 10. Therefore [1, 3] is shortest for OD pair (1,3), and [2,4]
is shortest for OD pair (2,4), so x∗ = 5000 0 0 0 0 0 10000 .2


Since is the first iteration, we simply set

x ← x∗ = 5000
 
0 0 0 0 0 10000 .

Recalculating the travel times, we have

 
t = 60 10 10 10 10 10 110 .

Iteration 1. With the new travel times, the shortest path for (1,3) is now
[1, 5, 6, 3], with a travel time of 30, so κ13 = 30. Likewise, the new shortest
2 For each OD pair, we add the total demand from the OD matrix onto each link in the

shortest path.
6.2. LINK-BASED ALGORITHMS 147

path for (2,4) is [2, 5, 6, 4], so κ24 = 30 and the average excess cost is

t·x−κ·d
AEC =
d·1
5000 × 60 + 10000 × 110 − 30 × 5000 − 30 × 10000
= = 63.33 .
5000 + 10000
This is far too big and suggests that the average trip is 63 minutes slower
than the shortest paths available! If everyone were to take the new shortest
paths, the flows would be

x∗ = 0 5000 15000 5000 10000 10000 0 .

 

(Be sure you understand how we calculated this.) Because this is iteration
1, we shift 1/2 of the travelers onto this path, so

x ← (1/2)x∗ +(1/2)x = 2500 2500 7500 2500 5000 5000 5000 .

 

The new travel times are thus

 
t = 35 35 85 35 60 60 60 .

Iteration 2. With the new travel times, the shortest path for (1,3) is now [1, 3],
with κ13 = 35. The new shortest path for (2,4) is [2, 4], so κ24 = 60 and
the average excess cost is

AEC =
 
2500 × 35 + 2500 × 35 + 7500 × 85 + 2500 × 35 + 5000 × 60
+ 5000 × 60 + 5000 × 60 − 35 × 5000 − 60 × 10000
15000
= 68.33 .

This is still big (and in fact worse), but we persistently continue with the
second iteration. If everyone were to take the new shortest paths, the
flows would be

x∗ = 5000 0 0 0 0 0 10000 ,
 

so

x ← (1/3)x∗ +(2/3)x = 3333

 
1667 5000 1667 3333 3333 6667 .

The new travel times are

 
t = 43.3 26.7 60 26.7 43.3 43.3 76.7 .

Iteration 3. With the new travel times, the shortest path for (1,3) is still
[1, 3], with κ13 = 43.3, and the shortest path for (2,4) is still [2, 4] with
148 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

κ24 = 76.7 and the average excess cost is AEC = 23.6. Continuing the
fourth iteration, as before

x∗ = 5000 0 0 0 0 0 10000 ,
 

so

x ← (1/4)x∗ +(3/4)x = 3750

 
1250 3750 1250 2500 2500 7500 .

The new travel times are

 
t = 47.5 22.5 47.5 22.5 35 35 85 .

Iteration 4. With the new travel times, the shortest path for (1,3) is still [1, 3],
with κ13 = 47.5, and the shortest path for (2,4) is still [2, 4] with κ24 = 85
and the average excess cost is AEC = 9.42. Continuing the fifth iteration,
as before
x∗ = 5000 0 0 0 0 0 10000 ,
 

so

x ← (1/5)x∗ +(4/5)x = 4000

 
1000 3000 1000 2000 2000 8000 .

The new travel times are

 
t = 50 20 40 20 30 30 90 .

Iteration 5. With the new travel times, the shortest path for (1,3) is still [1, 3],
with κ13 = 50, and the shortest path for (2,4) is still [2, 4] with κ24 = 90
and the average excess cost is AEC = 3.07. Note that the shortest paths
have stayed the same over the last three iterations. This means that we
really could have shifted more flow than we actually did. The Frank-Wolfe
algorithm, described in the next section, fixes this problem. We have

x∗ = 5000 0 0 0 0 0 10000 ,
 

so

x ← (1/6)x∗ + (5/6)x = 4167

 
833 2500 833 1667 1667 8333 .

The new travel times are

 
t = 51.7 18.3 35 18.3 26.7 26.7 93.3 .

Iteration 6. With the new travel times, the shortest path for (1,3) is still
[1, 3], with κ13 = 51.7, but the shortest path for (2,4) is now [2, 5, 6, 4]
with κ24 = 88.3. The average excess cost is AEC = 3.80. Note that the
OD pairs are no longer behaving “symmetrically,” the shortest path for
6.2. LINK-BASED ALGORITHMS 149

(1,3) stayed the same, but the shortest path for (2,4) has changed. We
have
x∗ = 5000 0 10000 0 10000 10000 0 .
 

so

x ← (1/7)x∗ + (6/7)x = 4286

 
714 3571 714 2857 2857 7142 .

The new travel times are

 
t = 52.9 17.1 45.7 17.1 38.6 38.6 81.4 .

This process continues over and over until the average excess cost is suffi-
ciently small. Even with such a small network, MSA requires a very long time
to converge. An average excess cost of 1 is obtained after eleven iterations, 0.1
after sixty-three iterations, 0.01 after three hundred thirty-two, and the rate of
convergence only slows down from there.

6.2.2 Frank-Wolfe
One of the biggest drawbacks with MSA is that it has a fixed step size. Iteration
i moves exactly 1/(i + 1) of the travelers onto the new shortest paths, no matter
how close or far away we are from the equilibrium. Essentially, MSA decides its
course of action before it even gets started, then sticks stubbornly to the plan of
moving 1/(i+1) travelers each iteration. The Frank-Wolfe (FW) algorithm fixes
this problem by using an adaptive step size. At each iteration, FW calculates
exactly the right amount of flow to shift to get as close to equilibrium as possible.
We might try to do this by picking λ to minimize the relative gap or average
excess cost, but this turns out to be harder to compute. Instead, we pick λ
to solve a “restricted” VI where the feasible set is the line segment connecting
x and x∗ . It turns out that this is the same as choosing λ to minimize the
Beckmann function (6.3) along this line segment. Both approaches for deriving
the step size λ are discussed below.
Define X 0 to be the link flows lying on the line segment between x and x∗ .
That is, X 0 = {x0 : x = λx∗ + (1 − λ)x for some λ ∈ [0, 1]}. The restricted VI
is: find x̂0 ∈ X 0 such that t(x̂0 ) · (x̂0 − x0 ) ≤ 0 for all x0 ∈ X 0 .
This VI is simple enough to be solved as a single equation. The set X has
two endpoints (x and x∗ , corresponding to λ = 0 and λ = 1, respectively). For
now, assume that the solution x̂0 to the VI is not at one of these endpoints.3
In this case, the force vector −t(x̂0 ) is perpendicular to the direction x∗ − x.
(Figure 6.3), so −t(x̂0 ) · (x∗ − x) = 0. Writing this equation out in terms of
individual components, we need to solve
X
tij (x̂0ij ) x∗ij − xij = 0


(6.12)
ij

3 Exercise13 asks you to show that the solution methods provided below will still give the
right answer even in these cases.
150 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

x*

x’

t(x’)
x
X

Figure 6.3: A solution to the restricted VI in the Frank-Wolfe method.

or equivalently
X X
tij (x̂0ij )x∗ij = tij (x̂0ij )xij (6.13)
ij ij

The same equation can be derived based on the Beckmann function. Recall
the discussion above, where we wrote the function f (x(λ)) = f ((1 − λ)x + λx∗ )
to be the value of the Beckmann function after taking a step of size λ, and
furthermore found the derivative of f (x(λ)) to be

df X
= tij ((1 − λ)xij + λx∗ij )(x∗ij − xij ) . (6.14)
dλ
(i,j)∈A

It is not difficult to show that f (x(λ)) is a convex function of λ, so we can

find its minimum by setting the derivative equal to zero, which occurs if the
condition
X X
x∗ij tij ((1 − λ)xij + λx∗ij ) = xij tij ((1 − λ)xij + λx∗ij ) . (6.15)
(i,j)∈A (i,j)∈A

is satisfied, which is the same as (6.13).

Study these equations carefully: the coefficients xij and x∗ij are constants
and do not change with λ; the only part of this condition which is affected
by λ are the travel times. You can interpret this equation as trying to find a
balance between x and x∗ in the following sense: different values of λ correspond
to shifting a different number of travelers from their current paths to shortest
paths, which will result in different travel times on all the links. You want to
pick λ so that, after you make the switch, both the old paths x and the old
shortest paths x∗ are equally attractive in terms of their travel times.
6.2. LINK-BASED ALGORITHMS 151

It is convenient to write equation (6.12) or (6.15) as a function in terms of

λ as X
ζ 0 (λ) = tij (λx∗ij + (1 − λ)xij ) x∗ij − xij = 0 ,


(6.16)
ij

which we need to solve for λ ∈ [0, 1] and ζ(λ) is used as a shorthand for f (x(λ)).
Since the link performance functions are typically nonlinear, we cannot expect
to be able to solve this equation analytically to get an explicit formula for λ.
General techniques such as Newton’s method or an equation solver can be used;
but it’s not too difficult to use an enlightened trial-and-error method such as a
binary search or bisection. Details for all of these line search techniques were
presented in Section 3.3.1, and will not be repeated here.
To summarize: in FW, x∗ is an all-or-nothing assignment (just as with
MSA). The difference is that λ is chosen to solve (6.16). So, there is a little bit
more work at each iteration (we have to solve an equation for λ instead of using
a precomputed formula as in MSA), but the reward is much faster convergence
to the equilibrium solution. Two examples of FW now follow; you are asked to
provide a proof of correctness in Exercise 14 and 15.

6.2.3 Small network example

Here we solve the small example of Figure 6.1 by FW. Some steps are similar
to MSA, and therefore omitted. Here, when we do the bisection method, we do
five interval reductions, so we are within 1/25 = 1/32 of the correct λ∗ value.
When solving by computer, you would usually perform more steps than this,
because the bisection calculations are very fast.

Initialization. As before, we load

 everybody on the initial shortest path, so
x = x∗ = 50 0 and t = 60 20 .
 

Iteration 1. As before, the relative gap is γ1 = 2. With the new shortest

paths, x∗ = 0 50 . Begin the bisection method.


Bisection Iteration 1. Initially λ∗ ∈ [0, 1]. Calculate ζ 0 (1/2) = (0 −

50) × (10 + 25) + (50 − 0) × (20 + 25) = 500 > 0 so we discard the
upper half.
Bisection Iteration 2. Now we know λ∗ ∈ [0, 1/2]. Calculate

ζ 0 (1/4) = (0 − 50) × (10 + 37.5) + (50 − 0) × (20 + 12.5) = −750 < 0

so we discard the lower half.

Bisection Iteration 3. Now we know λ∗ ∈ [1/4, 1/2]. Calculate

ζ 0 (3/8) = (0 − 50) × (10 + 31.25) + (50 − 0) × (20 + 18.75) = −125 < 0

so we discard the lower half.

152 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Bisection Iteration 4. Now we know λ∗ ∈ [3/8, 1/2]. Calculate

ζ 0 (7/16) = (0 − 50) × (10 + 28.125) + (50 − 0) × (20 + 21.875) = 187.5
so we discard the upper half.
Bisection Iteration 5. Now we know λ∗ ∈ [3/8, 7/16]. Calculate
ζ 0 (13/32) = (0−50)×(10+29.6875)+(50−0)×(20+20.3125) = 31.25
so we discard the upper half.
From here we take the midpoint of the last interval [3/8, 13/32]
 to estimate
λ∗ ≈ 25/64
 = 0.390625,
 so x = (25/64)x ∗
+ (39/64)x = 30.47 19.53
and t = 40.47 39.53 .
Iteration 2. The relative gap is calculated as γ1 = 0.014. (This is an order of
magnitude smaller than the relative gap MSA found by this point.) The
shortest paths are still x∗ = 0 50 , and we begin bisection.
Bisection Iteration 1. Initially λ∗ ∈ [0, 1]. Calculate
ζ 0 (1/2) = 900 > 0
so we discard the upper half.
Bisection Iteration 2. Now we know λ∗ ∈ [0, 1/2]. Calculate
ζ 0 (1/4) = 435 > 0
so we discard the upper half.
Bisection Iteration 3. Now we know λ∗ ∈ [0, 1/4]. Calculate
ζ 0 (1/8) = 203 > 0
so we discard the upper half.
Bisection Iteration 4. Now we know λ∗ ∈ [0, 1/8]. Calculate
ζ 0 (1/16) = 87 > 0
so we discard the upper half.
Bisection Iteration 5. Now we know λ∗ ∈ [0, 1/16]. Calculate
ζ 0 (1/32) = 29 > 0
so we discard the upper half.
∗ ∗
The midpoint  interval is λ ≈ 1/64,
 of the final  so x = (1/64)x +
(63/64)x = 29.99 20.01 and t = 39.99 40.01 .
Iteration 3. The relative gap is now γ1 = 0.00014, so we quit and claim we
have found flows that are “good enough” (the difference in travel times
between the routes is less than a second).
Alternately, using calculus, we could have identified λ∗ during the second
iteration as exactly 0.40, which would have found the exact equilibrium after
only one step.
6.2. LINK-BASED ALGORITHMS 153

6.2.4 Large network example

Here we apply FW to the network shown in Figure 6.2, using the same notation
as in the MSA example.

Initialization. Path [1, 3] is shortest for OD pair (1,3), and path [2, 4] is short-
est for OD pair (2,4), so

x∗ = 5000 0 0 0 0 0 10000
 

and
x = x∗ = 5000
 
0 0 0 0 0 10000 .
Recalculating the travel times, we have
 
t = 60 10 10 10 10 10 110 .

Iteration 1. With the new travel times, the shortest path for (1,3) is now
[1, 5, 6, 3], and the new shortest path for (2,4) is [2, 5, 6, 4], so AEC = 63.33
If everyone were to take the new shortest paths, the flows would be

x∗ = 0 5000 15000 5000 10000 10000 0 .

 

Begin the bisection method to find the right combination of x∗ and x.

Bisection Iteration 1. Initially λ∗ ∈ [0, 1]. Calculate

ζ 0 (1/2) = (0 − 5000) × (10 + 2500/100) + . . .

+ (0 − 10000) × (10 + 5000/100) = 2050000 > 0

so we discard the upper half.

Bisection Iteration 2. Now we know λ∗ ∈ [0, 1/2]. Calculate

ζ 0 (1/4) = (0 − 5000) × (10 + 3750/100) + . . .

+ (0 − 10000) × (10 + 7500/100) = 550000 > 0

so we discard the upper half.

Bisection Iteration 3. Now we know λ∗ ∈ [0, 1/4]. Calculate

ζ 0 (1/8) = (0 − 5000) × (10 + 4375/100) + . . .

+ (0 − 10000) × (10 + 8750/100) = −200000 < 0

so we discard the lower half.

Bisection Iteration 4. Now we know λ∗ ∈ [1/8, 1/4]. Calculate

ζ 0 (3/16) = (0 − 5000) × (10 + 4062/100) + . . .

+ (0 − 10000) × (10 + 8125/100) = 175000 > 0

so we discard the upper half.

154 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Bisection Iteration 5. Now we know λ∗ ∈ [1/8, 3/16]. Calculate

ζ 0 (5/32) = (0 − 5000) × (10 + 4219/100) + . . .

+ (0 − 10000) × (10 + 8437/100) = −12500 < 0

so we discard the lower half.

The final interval is λ∗ ∈ [5/32, 3/16], so the estimate is λ∗ = 11/64 and
11 ∗ 53  
x= x + x = 4141 859 2578 859 1719 1719 8281 .
64 64
The new travel times are thus
 
t = 51.4 18.6 35.8 18.6 27.2 27.2 92.8 .

Iteration 2. With the new travel times, the shortest path for (1,3) is now
[1, 3], but the shortest path for (2,4) is still [2, 5, 6, 4]. The relative gap is
AEC = 2.67 (roughly 30 times smaller than the corresponding point in
the MSA algorithm!) We have

x∗ = 5000 0 10000 0 10000 10000 0 .

 

We begin the bisection method to find the right combination of x∗ and x.

Bisection Iteration 1. Initially λ∗ ∈ [0, 1]. Calculate

ζ 0 (1/2) = 637329 > 0

so we discard the upper half.

Bisection Iteration 2. Now we know λ∗ ∈ [0, 1/2]. Calculate

ζ 0 (1/4) = 154266 > 0

so we discard the upper half.

Bisection Iteration 3. Now we know λ∗ ∈ [0, 1/4]. Calculate

ζ 0 (1/8) = 36063 > 0

so we discard the upper half.

Bisection Iteration 4. Now we know λ∗ ∈ [0, 1/8]. Calculate

ζ 0 (1/16) = 7741 > 0

so we discard the upper half.

Bisection Iteration 5. Now we know λ∗ ∈ [0, 1/16]. Calculate

ζ 0 (1/32) = 1302 > 0

so we discard the upper half.

6.2. LINK-BASED ALGORITHMS 155

The final interval is λ∗ ∈ [0, 1/32], so the estimate is λ∗ = 1/64 and

1 ∗ 63  
x= x + x = 4154 845 2694 845 1848 1848 8152 .
64 64
The new travel times are thus
 
t = 51.5 18.5 36.9 18.5 28.5 28.5 91.5 .

At this point, the average excess cost is around 1.56 min; note that FW is
able to decrease the relative gap much faster than MSA. However, we’re still
quite far from equilibrium if you compute the actual path travel times. In this
case, even though we’re allowing the step size to vary for each iteration, we are
forcing travelers from all OD pairs to shift in the same proportion. In reality,
OD pairs farther from equilibrium should see bigger flow shifts, and OD pairs
closer to equilibrium should see smaller ones. This can be remedied by more
advanced algorithms.

6.2.5 Conjugate Frank-Wolfe

If the distinction between FW and MSA is that FW chooses the step size λ
in a more clever way, the distinction between conjugate Frank-Wolfe (CFW)
and plain FW is that CFW chooses the target x∗ in a more clever way. To
understand why CFW is more clever, we first need to understand why using the
all-or-nothing assignment as target can be problematic.
Viewed in terms of the set of feasible link assignments X, the all-or-nothing
assignments correspond to corner points of X. That is, Frank-Wolfe must limit
itself to the corner points of the feasible region when determining where to move.
In Figure 6.4, FW is constrained to follow the trajectory shown by the thin lines,
and is unable to take a direct step like that indicated by the thick arrow. While
these directions are effective in the early iterations, as the algorithm approaches
the equilibrium point its converge slows down dramatically, and “zigzagging”
behavior is observed.
So, how can we choose the target solution in a smarter way, so that steps in
the direction of the target still move toward equilibrium, while granting more
flexibility than the use of an all-or-nothing assignment? Conjugate Frank-Wolfe
provides one approach towards doing so.
Understanding CFW requires introducing the concept of conjugacy, which is
done here. Temporarily ignoring the context of the traffic assignment problem,
assume that we are trying to find the minimum point of a convex quadratic
function f (x1 , x2 ) of two variables, when there are no constraints. Figure 6.5
shows a few examples of these functions. Any quadratic function of two variables
can be written in the form

f (x) = xT Qx + bT x , (6.17)
 
where x = x1 x2 and b are two-dimensional vectors, and Q is a 2 × 2 matrix.
Figure 6.5 shows the Q matrix and b vector corresponding to each example.
156 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

H5*

H7*
H6*

H4*

H
H1*

H2*

H3*

Figure 6.4: Frank-Wolfe can only move towards extreme points of the feasible
region.

How would you go about finding the minimum of such a function? Given
some initial solution (x1 , x2 ), one idea is to fix x1 as a constant, and find the
value of x2 which minimizes f . Then, we can fix x2 , and find the value of x1
which minimizes f , and so on. This process will converge to the minimum, as
shown in Figure 6.6, but in general this convergence is only asymptotic, and the
process will never actually reach the minimum. The exception is when Q is the
identity matrix, as in Figure 6.6(a). In this case, the exact optimum is reached
in only two steps.
In fact, it is possible to reach the exact optimum in only two steps even
when Q is not the identity matrix, by changing the search directions. The
process described above (alternately fixing x1 , and  then x2 ) can be thought of
as alternating between searching in the direction 0 1 , then searching in the
 
direction 1 0 . As shown in Figure 6.7, by making a different choice for the
two search directions, the minimum can always be obtained in exactly two steps.
This happens if the two directions d1 and d2 are conjugate, that is, if

d1 T Qd2 = 0 (6.18)

Conjugacy generalizes the concept of orthogonality (or perpendicularity). If Q

is the identity matrix, equation (6.18) reduces to d1 · d2 = 0, the definition of
perpendicular vectors.
Now, returning to the traffic assignment problem, we want to use the concept
of conjugacy to choose a more intelligent search direction. In particular, we want
the target x∗ to be chosen so that the search direction x∗ − x is conjugate to
6.2. LINK-BASED ALGORITHMS 157

   
1 0 0
Q= b=
0 1 0

(a) f (x1 , x2 ) = x21 + x22

   
1 1/2 0
Q= b=
1/2 1 0

(b) f (x1 , x2 ) = x21 + x22 + x1 x2

   
1 1 0
Q= b=
1 1 0

(c) f (x1 , x2 ) = x21 + x22 + 2x1 x2

   
1 1/2 1
Q= b=
1/2 1 −1

(d) f (x1 , x2 ) = x21 + x22 + x1 x2 + x1 − x2

Figure 6.5: Four examples of convex quadratic functions of the form f (x) =
xT Qx + bT x
158 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

   
1 0 0
Q= b=
0 1 0

(a) f (x1 , x2 ) = x21 + x22

   
1 1/2 1
Q= b=
1/2 1 −1

(b) f (x1 , x2 ) = x21 + x22 + x1 x2 + x1 − x2

Figure 6.6: Searching in orthogonal directions finds the optimum in two steps
only if Q = I.

Figure 6.7: Searching in conjugate directions always leads to the optimum in

two steps.
6.2. LINK-BASED ALGORITHMS 159

the previous search direction. Before the derivation, there are a few differences
between traffic assignment and the unconstrained quadratic program used to
introduce conjugacy which should be addressed.

ˆ The Beckmann function is not a function of two variables. This is not a

huge problem. Finding the unconstrained minimum of a quadratic func-
tion of n variables requires n conjugate steps (so, in the above examples
with two variables, two steps sufficed), due to a result known as the ex-
panding subspace theorem. Of course...

ˆ The Beckmann function is in general not a quadratic function. (Can you

think of a case when it is?) Instead of the matrix Q, we will instead use
the Hessian of the Beckmann function Hf . Therefore, we cannot hope
for exact convergence in a finite number of iterations. However, when
the solution gets closer and closer to equilibrium, the Beckmann function
can be better and better approximated by a quadratic by taking the first
two terms of its Taylor series. This is good news, because zigzagging in
plain Frank-Wolfe becomes worse and worse as the equilibrium solution is
approached.

ˆ The optimization problem (6.3)–(6.6) has constraints. Again, this is not

a major problem, since the feasible region is convex. As long as we ensure
that the target point x∗ is feasible, any choice of λ ∈ [0, 1] will retain
feasibility.

So, how can we make sure that the new target vector x∗ is chosen so that the
search direction is conjugate to the previous direction, and that x∗ is feasible?
Since X is a convex set, feasibility can be assured by choosing x∗ to be a convex
combination of the old target vector (x∗old ) and the all-or-nothing assignment
xAON :
x∗ = αx∗old + (1 − α)xAON , (6.19)

for some α ∈ [0, 1]. Choosing α = 0 would make the all-or-nothing assignment
the target (as in plain FW), while choosing α = 1 would make the target in this
iteration the same as in the last. In fact, α should be chosen so that the new
direction is conjugate to the last, that is,

(x∗old − x)T H(x∗ − x) = 0 , (6.20)

where H is the Hessian of the Beckmann function evaluated at the current

solution x. Substituting (6.19) into (6.20), we can solve for α as follows:

(x∗old − x)T H(x∗ − x) = 0 (6.21)

⇐⇒ (x∗old − x) T
H(αx∗old + (1 − α)x AON
− αx − (1 − α)x) = 0 (6.22)
⇐⇒ (x∗old − x) H(α(x∗old − x) + (1 − α)(xAON − x)) = 0
T
(6.23)
160 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

or, after rearrangement,

α[(x∗old − x)T H(x∗old − xAON )] = −(x∗old − x)T H(xAON − x) (6.24)

(x∗old − x)T H(xAON − x)
⇐⇒ α = (6.25)
(x∗old − x)T H(x∗AON − xold )

Now, for TAP, the Hessian takes a specific form. Since the Beckmann func-
tion is
X Z xij
f (x) = tij (x) dx , (6.26)
(i,j)∈A 0

its gradient is simply the vector of travel times at the current flows

∇f (x) = vect{tij (xij )} , (6.27)

and its Hessian is the diagonal matrix of travel time derivatives at the current
flows
Hf (x) = diag{t0ij (xij )} . (6.28)
So, the matrix products in equation (6.25) can be written out explicitly, giving
∗ AON
− xij )t0ij
P
(i,j)∈A ((xold )ij − xij )(xij
α= P ∗ AON − (x∗ ) )t0
, (6.29)
(i,j)∈A ((xold )ij − xij )(xij old ij ij

where the derivatives t0ij are evaluated at the current link flows xij .
Almost there! A careful reader may have some doubts about the formula
in (6.29). First, it is possible that the denominator can be zero, and division
by zero is undefined. Second, to ensure feasibility of x∗ , we need α ∈ [0, 1],
even though it is not obvious that this formula always lies in this range (and in
fact, it need not do so). Furthermore, α = 1 is undesirable, because then the
current target point is the same as the target point in the last iteration. If the
previous line search was exact, there will be no further improvement and the
algorithm will be stuck in an infinite loop. Finally, what should you do for the
first iteration, when there is no “old” target x∗ ?
To address the first issue, the easiest approach is to simply set α = 0 if
the denominator of (6.29) is zero (i.e., if the formula is undefined, simply take
a plain FW step by using the all-or-nothing solution as the target). As for
the second and third issues, if the denominator is nonzero we can project the
right-hand side of (6.29) onto the interval [0, 1 − ] where  > 0 is some small
tolerance value. That is, if equation (6.29) would give a value greater than 1 − ,
set α = 1 − ; if it would give a negative value, use zero. Finally, for the first
iteration, simply use the all-or-nothing solution as the target: x∗ = xAON .
So, to summarize the discussion, choose α in the following way. If it is the
first iteration or the denominator of (6.29) is zero, set α = 0. Otherwise set
0 ∗
P AON
!
(i,j)∈A tij ((xold )ij − xij )(xij − xij )
α = proj[0,1−] P 0 ∗ AON − (x∗ ) )
(6.30)
(i,j)∈A tij ((xold )ij − xij )(xij old ij
6.2. LINK-BASED ALGORITHMS 161

Then the target solution x∗ is calculated using (6.19). The value of the step
size λ is chosen in the same way as in Frank-Wolfe, by performing a line search
(e.g., using bisection or Newton’s method) to solve (6.16).

Large network example Here we apply CFW to the network shown in Fig-
ure 6.2, using the same notation as in the FW and MSA examples. The tolerance
 is chosen to be a small positive constant, 0.01 in the following example.
Initialization. Generate the initial solution by solving an all-or-nothing assign-
ment. Path [1, 3] is shortest for OD pair (1,3), and path [2, 4] is shortest
for OD pair (2,4), so
xAON = 5000 0 0 0 0 0 10000
 

and
x = xAON = 5000
 
0 0 0 0 0 10000 .
Recalculating the travel times, we have
 
t = 60 10 10 10 10 10 110 .

Iteration 1. Proceeding in the same way as in the large network example for
Frank-Wolfe, the all-or-nothing assignment in this case is
xAON = 0 5000 0 10000 15000 5000 10000 .
 

which is used as the target x∗ since this is the first iteration of CFW.
Repeating the same line search process, the optimal value of λ is 19/120,
producing the new solution
 
x = 4208 792 8417 1583 2375 792 1583 .

Iteration 2. Again as with regular Frank-Wolfe, based on the updated travel

times the all-or-nothing assignment is
xAON = 5000 0 0 10000 10000 0 10000 .
 

However, here FW and CFW take different paths. Rather than using
xAON as the target vector, CFW generates a conjugate search direction.
First calculate the right-hand side of (6.29). Since for this problem tij =
1/100 for all links (regardless of the flow), the formula is especially easy to
compute using x and x∗ from the previous iteration and xAON as just now
computed. The denominator of (6.29) is nonzero, and the formula gives
−2.37; projecting onto the set [0, 1 − ] thus gives α = 0. So, calculating
the target x∗ from equation (6.19) with α = 0 we have
x∗ = 5000 0 0 10000 10000 0 10000
 

and, using a line search between x and x∗ , find that λ = 0.0321 is best,
resulting in
 
x = 4233 766 8146 1853 2620 766 1854 .
162 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Iteration 3. With the new flows x, the travel times are now
 
t = 52.3 17.7 91.5 28.5 36.2 17.7 28.5
and the all-or-nothing assignment is
xAON = 5000 0 10000
 
0 0 0 0 .
Calculating the right-hand side of (6.29), we see that the denominator
is nonzero, and the formula gives 0.198, which can be used as is since it
lies in [0, 1 − ]. So, calculating the target x∗ from equation (6.19) with
α = 0.198 we have
x∗ = 5000 0 8024 1976 1976 0 1976 .
 

Note that unlike any of the other link-based algorithms in this section,
the target flows are not an all-or-nothing assignment (i.e., not an extreme
point of X). Performing a line search between x and x∗ , find that λ =
0.652 is best, resulting in
 
x = 4733 267 8067 1933 2200 267 1933 .
which solves the equilibrium problem exactly, so we terminate.
In this example, CFW found the exact equilibrium solution in three itera-
tions. This type of performance is not typical (even though it is generally faster
than regular FW or MSA). In this example, the link performance functions are
linear, so the Beckmann function is quadratic. Iterative line searches with con-
jugate directions lead to the exact solution of quadratic programs in a finite
number of iterations, as suggested by the above discussion. This performance
cannot be assured with other types of link performance functions.
An even faster algorithm known as biconjugate Frank-Wolfe chooses its tar-
get so that the search direction is conjugate to both of the previous two search
directions. This method converges better than CFW, but is not explained here
because the details are a little more complicated even though the idea is the
same.

6.3 Path-Based Algorithms

This section introduces equilibrium algorithms which work in the space of path
flows H, rather than the space of link flows X. These tend to be faster, espe-
cially when high precision solutions are needed, but they require more computer
memory. Furthermore, achieving the full potential of these algorithms for rapid
convergence requires considerably more programming skill than for link-based
algorithms.
Before explaining path-based algorithms, it is worth explaining why link-
based algorithms are slow to converge. The following criticisms are specifically
aimed at the Frank-Wolfe algorithm, but apply to other link-based algorithms
as well.
6.3. PATH-BASED ALGORITHMS 163

3
7
3

3 7
1 4
3
7
2

Figure 6.8: A persistent cycle in Frank-Wolfe. (Link flows shown.)

It treats all OD pairs equally. If an OD pair is close to equilibrium, only a

small flow shift among its paths is needed, while if an OD pair is far from
equilibrium, a larger flow shift is needed. The Frank-Wolfe method uses
the same λ shift for all OD pairs regardless of how close or far away each
one is from equilibrium.

It uses a restricted set of search directions. If one imagines the space of

feasible traffic assignments, the “all-or-nothing” x∗ solutions generated
by the Frank-Wolfe algorithm represent extreme points or corners of this
region. In other words, the Frank-Wolfe algorithm is only capable of
moving towards a corner. Initially, this is fine, but as one nears the optimal
solution, this results in extensive zig-zagging when a much more direct
path exists (Figure 6.4). Conjugate Frank-Wolfe is a bit better in this
regard, but still uses a fairly restrictive set of target points.

It is unable to erase cyclic flows. Consider the network in Figure 6.8, with
the flows as shown. Such a flow might easily arise if [1, 2, 3, 4] is the
shortest path during the first iteration of Frank-Wolfe, [1, 3, 2, 4] is the
shortest path during the second, and λ = 0.3. With only one OD pair, it
is impossible for both links (2, 3) and (3, 2) to be used at equilibrium, as
discussed in Section 5.2.3. However, the Frank-Wolfe method will always
leave some flow on both links unless λ = 1 at any iteration (which is
exceedingly rare, especially in later iterations when λ is typically very
close to zero).

These difficulties can all be avoided by tracking the path flows h, rather than
the link flows x. The path flows contain much more information, tracking flow by
origin and destination, as opposed to link flows which are aggregated together.
On balance, the number of elements in the path flow vector is many orders of
magnitude larger than that of the link flows, easily numbering in the millions
for realistic networks. Algorithms which require us to first list off all paths in
the network are not tractable. Instead, path-based algorithms only track the
164 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

paths which an OD pair actually uses, that is the set Π̂rs = {π ∈ π rs : hπ > 0}.4
This is often referred to as the set of working paths for each OD pair.5 A rough
description of path-based algorithms can then be described as

1. Initialize Π̂rs ← ∅ for all OD pairs.

2. Repeat the following steps for each OD pair (r, s):

(a) Find the shortest path π̂rs . Add it to Π̂rs if it’s not already used.
(b) Shift travelers among paths to get closer to equilibrium.
(c) Update travel times.

3. Drop paths from Π̂rs if they are no longer used; return to step 2 unless a
convergence criterion is satisfied.

On the surface, this scheme looks quite similar to the link-based methods
presented earlier. Why might it converge faster? Recall the three factors de-
scribed above. First, each OD pair is now being treated independently. With
MSA and Frank-Wolfe, the same step-size λ was applied across all links (and
therefore, across all OD pairs). If one OD pair is very close to equilibrium, while
another is far away, we should probably make a finer adjustment to the first OD
pair, and a larger adjustment to the second one. Link-based methods allow no
such finesse, and instead bluntly apply the same λ to all origins. In practice,
this means that λ becomes very small after a few iterations: we can’t move
very far without disturbing an OD pair which is already close to equilibrium.
As a result of this, it takes a really long time to solve OD pairs which are far
from equilibrium. Equivalently, the set of search directions is broader in that
we can vary the step size by OD pairs; and lastly, it is quite possible to erase
cyclic flows in a path-based context, because the extra precision allows us to
take larger steps.
Step 3 is where different path-based algorithms differ. This section describes
two path-based algorithms: gradient projection and manifold suboptimization.
The latter is sometimes called projected gradient, because both algorithms use
the same two ingredients: exploiting the fact that the gradient is the direction
of steepest ascent (and therefore, in a minimization problem, we should move in
the opposite direction to descend as quickly as possible); and having to consider
the constraints in the problem by using a projection operation to stay in the
feasible set. (Recall from Chapter 3 that projection involves finding the point
within a set which lies closest to another point.) Where they differ is in the
order these steps are applied.
In gradient projection, we first take a step in the opposite direction of the
gradient, which will typically result in an infeasible point. The projection is
done after the flow shift, not before: we do the projection after we make use of
4 Those familiar with other types of optimization problems might recognize this as a column

generation scheme.
5 You may recognize similarities with the “trial-and-error” method from Chapter 4. Path-

based algorithms are essentially a more clever form of this method.

6.3. PATH-BASED ALGORITHMS 165

the gradient, so to speak. We then apply a projection operation to return to the

feasible set. In manifold suboptimization, we first calculate the gradient, then
project the gradient onto the feasible set so that our search direction respects
the demand constraint.

6.3.1 Gradient projection

The gradient projection method identifies the direction of steepest descent for
the Beckmann function, and calculates a new point in this direction. In case
this point is infeasible, a projection operation is applied to return to the closest
feasible point. Writing the Beckmann function in terms of path flows h, rather
than link flows as is customary, we have
P π π
X Z π∈Π δij h
f (h) = tij (x) dx , (6.31)
(i,j)∈A 0

and its partial derivative with respect to any path flow variable is
!
∂f X
π
X 0 0
π π
X
π
π
= δij tij δij h = δij tij (xij ) = cπ , (6.32)
∂h 0
(i,j)∈A π ∈Π (i,j)∈A

so the direction of steepest descent s is the negative gradient:

− ∇h f = −vect(cπ ) . (6.33)

So, a first attempt at gradient projection would be to update h ← projH (h−

∇h f ). Unfortunately, projecting onto the set H is not particularly easy. If we
apply a suitable change of variables, though, the projection can be made much
easier.
Define the basic path for OD pair (r, s) to be a path π̂rs with minimum travel
time. All of the other paths are called nonbasic paths. We can eliminate the
basic path flow variable by expressing it in terms of the nonbasic path flows:
X
hπ̂rs = drs − hπ . (6.34)
π∈Π̂rs :π6=π̂rs

Substituting this into the path-based Beckmann function (6.31), the partial
derivative with respect to one of the nonbasic path flows is now

∂ fˆ
= cπ − cπ̂rs ∀(r, s) ∈ Z 2 , π ∈ Π̂rs − {π̂rs } , (6.35)
∂hπ

denoting the Beckmann function as fˆ instead of f because the function has been
modified by using (6.34) to eliminate some of the path flow variables.
So, the change in path flows will be the negative of the gradient. Since the
gradient is given by (6.35) and cπ ≥ cπ̂rs (because the basic path is by definition
166 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

the shortest one), moving in this direction means that every nonbasic path flow
will decrease.
Since the transformation (6.34) eliminated the demand satisfaction con-
straint, the only remaining constraint is that the nonbasic path flow variables
be nonnegative. Projecting onto this set is trivial: if any of the nonbasic path
flow variables is negative after taking a step, simply set it to zero. At this point,
the basic path flow can be calculated through equation (6.34).
Furthermore, the larger the difference in travel times, the larger the cor-
responding element of the gradient will be. This suggests that more flow be
shifted away from paths with higher travel times. We can go a step further, and
estimate directly how much flow should be shifted from a nonbasic path to a
basic path to equalize the travel times, using Newton’s method.
Let ∆h denote the amount of flow we shift away from non-basic path π and
onto the basic path π̂, and let cπ (∆h) and cπ̂ (∆h) denote the travel times on
path π and π̂ after we make such a shift. We want to choose ∆h so these costs
are equal, that is, so

g(∆h) = cπ (∆h) − cπ̂ (∆h) = 0 ; (6.36)

that is, g is simply the difference in travel times between the two paths. To
apply Newton’s method, we need to find the derivative of g with respect to ∆h.
Using the relationships between link travel times and path travel times, we
have X
π π̂
g(∆h) = (δij − δij )tij (xij (∆h)) ,
(i,j)∈A

so
X dtij dxij
g 0 (∆h) = π
(δij π̂
− δij )
dxij d∆h
(i,j)∈A

by the chain rule. For each arc, there are four possible cases:
π π̂
Case I : δij = δij = 0, that is, neither path π nor path π̂ uses link (i, j). Then
π π̂ dtij dxij
(δij − δij ) dxij d∆h = 0 and this link does not contribute to the derivative.

π π̂
Case II : δij = δij = 1, that is, both paths π and path π̂ use link (i, j).
π π̂ dtij dxij
Then (δij − δij ) dxij d∆h = 0 and this link again does not contribute to the
derivative. (Another way to think of it: since both paths use this arc, its
total flow will not change if we shift travelers from one path to another.)
π π̂
Case III : δij = 1 and δij = 0, that is, path π uses link (i, j), but path π̂ does
π π̂ dtij dxij dt dxij dt dxij
not. Then (δij − δij ) dxij ∆h = dxijij d∆h = − dxijij since d∆h = −1.

π π̂
Case IV : δij = 0 and δij = 1, that is, path π̂ uses link (i, j), but path π does
π π̂ dtij dxij dt dxij dt dxij
not. Then (δij − δij ) dxij ∆h = − dxijij d∆h = − dxijij since d∆h = 1.
6.3. PATH-BASED ALGORITHMS 167

Putting it all together, the only terms which contribute to the derivative are
the links which are used by either π or π̂, but not both. Let A1 , A2 , A3 , and
A4 denote the sets of links falling into the four cases listed above. Then
X dtij
g 0 (∆h) = − .
dxij
(i,j)∈A3 ∪A4

which is simply the negative sum of the derivatives of these links, evaluated at
the current link flows.
Then, starting with an initial guess of ∆h = 0, one step of Newton’s method
gives an improved guess of
cπ − cπ̂
∆h = 0 − g(0)/g 0 (0) = P dtij
.
a∈A3 ∪A4 dxij

That is, the recommended Newton shift is given by the difference in path costs,
divided by the sum of the derivatives of the link performance functions for links
used by one path or the other, but not both. Therefore, the updated nonbasic
and basic path flows are given by
cπ − cπ̂
hπ̂ ← hπ̂ + P dtij
a∈A3 ∪A4 dxij

and
cπ − cπ̂
hπ ← hπ − P dtij
.
a∈A3 ∪A4 dxij

This process is repeated for every nonbasic path.

This is demonstrated on the example in Figure 6.2 as follows.

Iteration 1, Step 1. Initially Π̂13 = Π̂24 = ∅.

Iteration 1, Step 2a, OD pair (1,3). Find the shortest path for (1,3): with
no travelers on the network, the top link has a travel time of 10. This is
not in the set of used paths, so include it: Π̂13 = {[1, 3]}.

Iteration 1, Step 2b, OD pair (1,3). Since there is only one used path, we
simply have h13
[1,3] = 5000.

Iteration 1, Step 2c, OD pair (1,3). Update travel times:

 
t = 60 10 10 10 10 10 10

Iteration 1, Step 2a, OD pair (2,4). Find the shortest path for (2,4): with
no travelers on the network, link 7 has a travel time of 10. This is not in
the set of used paths, so include it: Π̂24 = {[2, 4]}.

Iteration 1, Step 2b. Since there is only one used path, h24
[2,4] = 10000.
168 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Iteration 1, Step 2c. Update travel times:

 
t = 60 10 10 10 10 10 110

Iteration 1, Step 3. All paths are used, so return to step 2. The relative gap
is γ1 = 2.11.

Iteration 2, Step 2a, OD pair (1,3). The shortest path is now [1,5,6,3]. This
is not part of the set of used paths, so we add it: Π̂13 = {[1, 3], [1, 5, 6, 3]}.

Iteration 2, Step 2b, OD pair (1,3). The difference in travel times between
the paths is 30 minutes; and the sum of the derivatives of links 1, 2, 3,
and 4 is 0.04. So we shift 30/0.04 = 750 vehicles from [1,3] to [1,5,6,3],
h13 13
[1,3] = 4250 and h[1,5,6,3] = 750.

Iteration 2, Step 2c, OD pair (1,3). Update travel times:

 
t = 52.5 17.5 17.5 17.5 10 10 110

Note that the two paths have exactly the same cost after only one step!
This is because Newton’s method is exact for linear functions.

Iteration 2, Step 2a, OD pair (2,4). The shortest path is now [2,5,6,4]. This
is not part of the set of used paths, so we add it: Π̂24 = {[2, 4], [2, 5, 6, 4]}.

Iteration 2, Step 2b, OD pair (2,4). The difference in travel times between
the paths is 72.5 minutes; and the sum of the derivatives of links 5, 3, 6,
and 7 is 0.04. So we shift 72.5/0.04 = 1812.5 vehicles from [2,4] to [2,5,6,4],
h24 24
[2,4] = 8187.5 and h[2,5,6,4] = 1812.5.

Iteration 2, Step 2c, OD pair (2,4). Update travel times:

 
t = 52.5 17.5 35.625 17.5 28.125 28.125 91.875

Note that the two paths again have exactly the same cost. However, the
equilibrium for the first OD pair has been disturbed.

Iteration 2, Step 3. All paths are used, so return to step 2. The relative gap
is γ1 = 0.0115.

Iteration 3, Step 2a, OD pair (1,3). The shortest path is now [1,3], which
is already in the set of used paths, so nothing to do here.

Iteration 3, Step 2b, OD pair (1,3). The difference in travel times between
the paths is 18.125 minutes; and the sum of the derivatives of links 1, 2,
3, and 4 is 0.04. So we shift 18.125/0.04 = 453 vehicles from [1,5,6,3] to
[1,3], h13 13
[1,3] = 4703 and h[1,5,6,3] = 297.
6.3. PATH-BASED ALGORITHMS 169

Iteration 3, Step 2c, OD pair (1,3). Update travel times:

 
t = 57.0 13.0 31.1 13.0 28.1 28.1 91.9

Again the first OD pair is at equilibrium, up to rounding error.

Iteration 3, Step 2a, OD pair (2,4). The shortest path is again [2,5,6,4].
This is already in the set of used paths, so nothing to do here.

Iteration 3, Step 2b, OD pair (2,4). The difference in travel times between
the paths is 4.6 minutes; and the sum of the derivatives of links 5, 3, 6,
and 7 is 0.04. So we shift 4.6/0.04 = 114 vehicles from [2,4] to [2,5,6,4],
h24 24
[2,4] = 8073 and h[2,5,6,4] = 1927.

Iteration 3, Step 2c, OD pair (2,4). Update travel times:

 
t = 57.0 13.0 32.2 13.0 29.3 29.3 90.7

Iteration 3, Step 3. All paths are used, so return to step 2. The relative gap
is γ1 = 0.00028.

Note that after three iterations of gradient projection, the relative gap is two or-
ders of magnitude smaller than that from the Frank-Wolfe algorithm. Although
not demonstrated here for reasons of space, the performance of gradient projec-
tion relative to Frank-Wolfe actually improves from here on out. Frank-Wolfe
usually does most of its work in the first few iterations, and then converges very
slowly after that.6 On the other hand, gradient projection maintains a steady
rate of progress throughout, with a nearly constant proportionate decrease in
gap from one iteration to the next.

6.3.2 Manifold suboptimization

The manifold suboptimization algorithm is also based on the Beckmann formu-
lation, applying an algorithm of Rosen from nonlinear optimization. As with
gradient projection, moving in the direction of the negative gradient will travel
along the direction of steepest descent. Such a direction will tend towards the
equilibrium solution, which minimizes the Beckmann function. However, we
need to be careful not to leave the region of feasible path flows H, making sure
that our search direction still satisfies the demand constraint and retains non-
negative path flows. Unlike gradient projection, which fixes an infeasible move
by projecting back to the feasible region, manifold suboptimization avoids ever
leaving the feasible region in the first place by modifying the steepest descent
direction. This algorithm goes by several names in the literature; sometimes
called projected gradient and at other times even called gradient projection
(even though it is different than the gradient projection method of Section 6.3.1).
6 It’s been said that Frank-Wolfe converges, but just barely.
170 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

To avoid confusion, we refer to it by manifold suboptimization, a convention also

followed by Bertsekas (2016).
Recall from equation (6.33) that the gradient of the Beckmann function is

∇h f = vect(cπ ) . (6.37)

Assuming that the current path flow solution h is feasible, we must move in a
direction that does not violate any of the constraints,
P only using the working
paths Π̂rs . For instance, if the demand constraint π∈Π̂rs hπ = drs is satisfied
for all OD pairs (r, s), it must remain so after taking a step in the direction ∆h:
X
(hπ + ∆hπ ) = drs . (6.38)
π∈Π̂rs

This in turn implies

X X
∆hπ = drs − hπ (6.39)
π∈Π̂rs π∈Π̂rs
X
∆hπ = 0 (6.40)
π∈Π̂rs

So, we need to find the

P projection of the steepest descent direction s onto
the space ∆H ≡ {∆h : π∈Π̂irs ∆hπ = 0}. It turns out that this projection
has a remarkably simple closed form expression.
Proposition 6.1. Using c̄rs = (1/|Π̂rs |) π∈Π̂rs cπ to denote the average travel
P
time of the working paths for OD pair (r, s), the direction s0 with components
s0π = cπ − c̄rs is the projection of the steepest descent direction s onto the set
∆H.
Proof. To show that s0 is the projection of s onto ∆H, we must show that
s0 ∈ ∆H, and that s − s0 is orthogonal to ∆H. Regarding the first part, we have
X X X
s0π = (cπ − c̄rs ) = cπ − |Π̂rs |c̄rs = 0 , (6.41)
π∈Π̂rs π∈Π̂rs π∈Π̂rs

so s0 ∈ ∆H.
Regarding the second part, let ∆h be any vector in ∆H. We now show that
(s − s0 ) · ∆h = 0. We have
X
(s − s0 ) · ∆h = [cπ − (cπ − c̄rs )]∆hπ (6.42)
π∈Π̂rs
X
= −c̄rs ∆hπ (6.43)
π∈Π̂P rs
=0 (6.44)

since ∆h ∈ ∆H.
6.4. BUSH-BASED ALGORITHMS 171

So, given current path flows hrs for OD pair (r, s), we use ∆hrs = vect(cπ −
c̄ ) as the search direction. To update the path flows hrs ← hrs + µ∆hrs , we
rs

need an expression for the step size. For any path π ∈ Π̂rs for which ∆hπ < 0,
the new path flows would be infeasible if µ > hπ /∆hπ . Therefore, the largest
possible step size is
hπ
µ̄ = min π
. (6.45)
π∈Π̂rs :∆hπ <0 ∆h

The actual step size µ ∈ [0, µ̄] should be chosen to minimize the Beckmann
function. This can be done either through bisection or one or more iterations
of Newton’s method.

6.4 Bush-Based Algorithms

As seen in the previous section, path-based algorithms converge faster than link-
based algorithms, and allow us find much more accurate equilibrium solutions in
a fraction of the time. The prime disadvantage is a huge memory requirement,
with potentially millions of paths available for use in large networks. Further-
more, many of these paths are “redundant” in some way, as they overlap: the
same sequence of links might be used by many different paths.
Bush-based algorithms (also known as origin-based algorithms) try to ad-
dress these limitations. Rather than treating each OD pair separately (as link-
based algorithms do), bush-based algorithms simultaneously consider every des-
tination associated with a single origin. Instead of considering every possible
used path (as path-based algorithms do, and there are potentially very many
of these), they maintain a set of links called a bush, which are the only links
which travelers from that origin are permitted to use. One can think of a bush
as the set of links obtained by superimposing all of the paths used by travelers
starting from an origin. In particular, a bush must be:

ˆ Connected ; that is, using only links in the bush, it is possible to reach
every node which was reachable in the original network.

ˆ Acyclic; that is, no path using only bush links can pass the same node more
than once. This is not restrictive, because travelers trying to minimize
their own travel time would never cycle back to the same node, and greatly
speeds up the algorithm, because acyclic networks are much simpler and
admit much faster methods for finding shortest paths and other quantities
of interest.

There is no particular reason why bushes have to be “origin-based” rather than

“destination-based,” and all of the results in this section can be derived in
a parallel way for bushes terminating at a common destination, rather than
starting at a common origin.
An example of a bush is shown in panels (a) and (b) of Figure 6.9, where
the thickly-shaded links are part of the bush, and the lightly-shaded links are
not. The bush in panel (a) is a special type of bush known as a tree, which
172 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

(a) (b)

(c) (d)

Figure 6.9: Examples of bushes (panels (a) and (b)) and non-bushes (panels (c)
and (d)).

has exactly one path from the origin to every node. The thickly-shaded links
in panel (c) do not form a bush, because it is not connected; there is no way to
reach the nodes at the bottom of the network only using bush links. Likewise,
the thick links in panel (d) do not form a bush either, because a cycle exists and
it would be possible to revisit some nodes multiple times using the bush links
(find them!).
Notice that because a bush is acyclic, it can never include both directions
of a two-way link. This implies that at equilibrium, on every street travelers
from the same origin must all be traveling in the same direction. (This follows
from Proposition 5.3 in Section 5.2.3). Interestingly, link-based and path-based
algorithms cannot enforce this requirement easily; and this is yet another reason
that bush-based algorithms are a good option for solving equilibrium. There is
one bush for every origin; this means that if there are z origins and m links, we
need to keep track of at most zm values. By contrast, a link-based approach
(such as Frank-Wolfe) requires storage of only m values to represent a solution,
while a path-based approach could conceivably require z 2 2m values.7
The first well-known origin-based algorithm was developed by Hillel Bar-
Gera in his dissertation (circa 2000), and was simply called origin-based assign-
ment (OBA). Bob Dial developed another, simpler method that was published
7 These values are very approximate, but give you an idea of the scale.
6.4. BUSH-BASED ALGORITHMS 173

in 2006 as “Algorithm B,” which also seems to work faster than OBA. Yu
(Marco) Nie compared both algorithms and developed additional variations by
combining features of both, and contributing some ideas of his own. Most re-
cently, Guido Gentile has developed the LUCE algorithm, and Hillel Bar-Gera
has provided a new algorithm called TAPAS which simultaneously solves for
equilibrium and proportional link flows (approximating entropy maximization).
This section focuses on Algorithm B, since it is simpler to explain on its own.
All bush-based algorithms operate according to the same general scheme:

1. Start with initial bushes for each origin (the shortest path tree with free-
flow times is often used as a starting point).

2. Shift flows within each bush to bring each origin closer to equilibrium.

3. Improve the bushes by adding links which can reduce travel times, and by
removing unused links. Return to step 2.

Step 1 is fairly self-explanatory: collecting the shortest paths from an origin

to every destination into one bush results in a connected and acyclic set of links,
and we can simply load the travel demand to each destination on the unique
path in the bush.
Step 2 is where most bush-based algorithms differ.8 This section explains
three different methods for shifting flows within a bush: Algorithm B, origin-
based assignment (OBA), and linear user cost equilibrium (LUCE). Each of
these algorithms is based on labels calculated for bush nodes and links, and
Section 6.4.1 defines each of these. The three algorithms themselves are pre-
sented in Section 6.4.2.
Step 3 requires a little bit of thought to determine how to adjust the bush
links themselves to allow movement towards equilibrium when Step 2 is per-
formed on the new bushes. Section 6.4.3 shows how this can be done.

6.4.1 Bush labels

Algorithm B, OBA, and LUCE all make use of “bush labels” associated with
each node and link in a bush. All of these labels can be calculated in a straight-
forward, efficient manner using the topological order within each bush. As you
read this section, pay attention to which of these labels are calculated in ascend-
ing topological order (i.e., the formulas only involve nodes with lower topological
order) and which are calculated in descending order (based only on nodes with
higher topological order.) In the discussion below, assume that a bush is given
and fixed, with origin node r. The set of bush links is denoted B and, for
the purposes of this section, assume that the travel times tij and travel time
derivatives t0ij of all bush links are given and constant. All of these labels are
only defined for bush links, and the formulas only involve bush links. From the
standpoint of Step 2 of bush-based algorithms, non-bush links are completely
8 If this is starting to sound familiar, good! How to shift flows to move closer to equilibrium

is also where link-based and path-based algorithms differ most.

174 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

4 +2x2 2 + x2 (1,1/7) (7,7/10)

1 2 3 1 2 3
4 +2x2 (1,1) (5,5/7)
40 +2x2 2+ x2 2+ x2 (1,1/7) (3,3/10)
4 5 6 4 5 6
2 + x2 (5,1)
2+x 2 2 + x2 (7,1) (3,1)
7 8 9 7 8 9
2 + x2 2 + x2 (3,1) (3,1)
(a) Link performance functions (b) Bush flows: (x, ) labels

1 6 2 51 3 1 4 2 14 3
6 4 6
42 27 11 4 10
4 5 6 4 5 6
27 10
51 11 14 6
7 8 9 7 8 9
11 11 6 6
(c) Bush travel times t (d) Bush travel time derivatives t'

Figure 6.10: Example bush used to demonstrate label calculation.

irrelevant. Table 6.1 shows all of the labels defined in this section, and which
labels are used in which algorithms, and in the bush-updating steps described
in Section 6.4.3.
We start with two different ways to represent the travel patterns on each
bush, starting with x labels. The label xB ij associated with each link indicates
the number of travelers starting at node r (the root of bush B) and traveling
on bush link (i, j). The superscript B indicates that we are only referring to
the flow on this link associated with the bush B, and the total link flow xij is
the sum of xB ij across all bushes B. However, using these superscripts tends to
clutter formulas, and often times it is clear that we are only referring to flows
within the context of a specific bush. In this case, we can simply write xij with
it being understood that this label refers to the flow on a bush link. Within
this section, we are only concerned with a single bush and the superscript will
be omitted for brevity.
The network in Figure 6.10(a) will be used as to demonstrate the labels
introduced in this section. The thick links comprise the bush, and the link
performance functions for all links in the network are shown. The origin in this
case is node 7, and the demand is 10 vehicles from node 7 to node 3. You can
verify that a topological ordering of the nodes on this bush is 7, 8, 9, 6, 4, 1, 5,
2, 3.
The corresponding label xi associated with each node indicates the total
number of vehicles using node i on the bush B, including flow which is termi-
nating at node i. The flow conservation equations relating xij and xi labels are
as follows: X X
xi = xhi = dri + xij ∀i ∈ N ,
(h,i)∈B (i,j)∈B
6.4. BUSH-BASED ALGORITHMS 175

where the first expression defines the node flow in terms of incoming link flows,
and the second in terms of outgoing link flows. The two definitions are equiva-
lent.
It is sometimes convenient to refer to the fraction of the flow at a node
coming from a particular link. For any node i with positive node flow (xi > 0),
define αhi to be the proportion of the node flow contributed by the incoming
link (h, i), that is

αhi = xhi /xi . (6.46)

Clearly each αhi is nonnegative, and, by flow conservation, the sum of the αhi
values entering each node i is one. The definition of αhi is slightly trickier
when xi = 0, because the formula (6.46) then involves a division by zero. To
accommodate this case, we adopt this rule: when xi = 0, the proportions αhi
may take any values whatsoever, as long as they are nonnegative and sum to
one. It is important to be able to define αhi values even in this case, because
the flow-shifting algorithms may cause xi to become positive, and in this case
we need to know how to distribute this new flow among the incoming links.
If we are given α labels for each bush link, it is possible to calculate the
resulting node and link flows xi and xij , using this recursion:

X
xi = dri + xij ∀i ∈ N (6.47)
(i,j)∈B

xij = αij xj ∀(i, j) ∈ B (6.48)

The sum in (6.47) is empty for the node with the highest topological order.
So we can start there, and then proceed with the calculations in backward
topological order.
Figure 6.10(b) shows the x and α labels for the example bush. You should
verify that the formulas (6.47) and (6.48) are consistent with these labels. The
link travel times and link travel time derivatives are shown in panels (c) and (d)
of this figure.
As has been used earlier in the text for shortest path algorithms, L is used
to denote travel times on shortest paths. The superscript B can be appended
to these labels when it is necessary to indicate that these labels are for shortest
paths specifically on the bush B, although it will usually be clear from context
which bush is meant. This section will omit such a superscript to avoid cluttering
the formulas, and it should be understood that L means the shortest path only
on the bush under consideration. The same convention will apply to the other
labels in this section. There are L labels associated with each node i, and with
each link (i, j) ∈ B: Li denotes the distance on the shortest path from r to i
using only bush links, and Lij is the travel time which would result if you follow
the shortest path to node i, then take link (i, j). These labels are calculated
176 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

using these equations:

Lr = 0 (6.49)
Lij = Li + tij ∀(i, j) ∈ B (6.50)
Li = min {Lhi } ∀i 6= r (6.51)
(h,i)∈B

The U labels are used to denote travel times on longest paths within the bush,
and are calculated in a similar way. Like the L labels, U labels are calculated
for both nodes and bush links, using the formulas:

Ur = 0 (6.52)
Uij = Ui + tij ∀(i, j) ∈ B (6.53)
Ui = max {Uhi } ∀i 6= r (6.54)
(h,i)∈B

The M labels are used to denote the average travel times within the bush,
recognizing that some travelers will be on longer paths and other travelers will be
on shorter ones. The node label Mi represents the average travel time between
origin r and node i across all bush paths connecting these nodes, weighted by
the number of travelers using each of these paths. The label Mij indicates the
average travel time of vehicles after they finish traveling on link (i, j), again
averaging across all of the bush paths starting at the origin r and ending with
link (i, j). These can be calculated as follows:

Mr = 0 (6.55)
Mij = Mi + tij ∀(i, j) ∈ B (6.56)
X
Mi = αhi Mhi ∀i 6= r (6.57)
(h,i)∈B

Figure 6.11 shows the L, U , and M labels corresponding to the example

bush in panels (a) and (b). Panel (a) shows the labels associated with links
(Lij , Mij , and Uij ), while panel (b) shows the labels associated with nodes (Li ,
Mi , and Ui ).
It is also useful to know how the average travel times (the M labels) will
change with a marginal increase in flow on a particular link or through a partic-
ular node. The set of D labels are used to represent this. As we will see shortly,
these play the role of the travel time derivatives9 , which are needed in flow
shifting rules based on Newton’s method (like those described for path-based
algorithms in the previous section.) Ideally, the Dij labels would represent the
derivative of Mj with respect to αij . However, there is no known method for
calculating these derivatives in an efficient way (“efficient” meaning they can be
calculated in a single topological pass). The following formulas for Dij can be
used as an approximation that can be calculated in a single topological pass.
9 For a mnemonic, you can associate the labels L, U , and M with the lower, upper, and

mean travel times to nodes, while D refers to the derivative.

6.4. BUSH-BASED ALGORITHMS 177

(63,63,63) (114,144,156)
1 2 3 1 2 3
(57,57,57)
(57,57,57) (63,93,105) (44,114,156)
(93,93,93) (105,105,105) (44,44,44)

4 5 6 4 5 6
(78,78,78) (51,51,51) (78,78,78) (33,33,33)
(51,51,51) (33,33,33)
7 8 9 7 8 9
(11,11,11) (22,22,22) (0,0,0) (11,11,11) (22,22,22)
(a) (L,M,U) labels for bush links (b) (L,M,U) labels for nodes

1 22 2 43.6 3 1 2 3
18 29.6 37.1
18 34
18 24
4 5 6 4 5 6
24 14 24 18
14 18
7 8 9 7 8 9
6 12 0 6 12
(c) D labels for bush links (d) D labels for nodes

Figure 6.11: Continuation of label calculation demonstration.

Table 6.1: Bush labels used in different algorithms.

Label B OBA LUCE Updating
x X
α X X
L X X
U X X
M X X
D X X

Dr = 0 (6.58)
Dij = Di + t0ij ∀(i, j) ∈ B (6.59)
X X X
2
p
Di = αhi Dhi + αhi αgi Dhi Dgi ∀i 6= r (6.60)
(h,i)∈B (h,i)∈B (g,i)∈B
(g,i)6=(h,i)

Figure 6.11 shows the D labels associated with the example bush in panels
(c) and (d). Panel (c) shows the Dij labels associated with links, and panel (d)
shows the Di labels associated with nodes.

6.4.2 Shifting flows on a bush

With the labels defined in the previous subsection, we are now ready to state
the Algorithm B, OBA, and LUCE flow shifting procedures. Any of these can
be used for Step 2 of the generic bush-based algorithm presented at the start
178 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

of the section. You will notice that all of these procedures follow the same
general form: calculate bush labels (different labels for different algorithms) in
forward topological order, then scan each node in turn in reverse topological
order. When scanning a node, use the labels to identify vehicles entering the
node from higher-cost approaches, and shift them to paths using lower-cost
approaches. Update the x and/or α labels accordingly, then proceed to the
previous node topologically until the origin has been reached.
All three of these algorithms also make use of divergence nodes (also called
last common nodes or pseudo-origins in the literature) as a way to limit the
scope of these updates. While the definition of divergence nodes is slightly
different in these algorithms, the key idea is to find the “closest” node which is
common to all of the paths travelers are being shifted among. The rest of this
subsection details these definitions and the role they play.

Algorithm B
Algorithm B identifies the longest and shortest paths to reach a node, and
shifts flows between them to equalize their travel times. That is, when scanning
a node i, only two paths (the longest and shortest) are considered. It is easy
to determine these paths using the L and U labels, tracing back the shortest
and longest paths by identifying the links used for the minimum or maximum
in equations (6.51) and (6.54). Once these paths are identified, the divergence
node a is the last node common to both of these paths.
The shortest and longest path segments between nodes a and i form a pair
of alternate segments; let σL and σU denote these path segments. Within Al-
gorithm B, flow is shifted from the longest path to the shortest path, using
Newton’s method to determine the amount of flow to shift:
(Ui − Ua ) − (Li − La )
∆h = P 0 (6.61)
(g,h)∈σL ∪σU tgh

There is also the constraint ∆h ≤ min(i,j)∈σU xij which must be imposed to

ensure that all links retain nonnegative flow after the shift. This flow is sub-
tracted from each of the links in the longest path segment σU , and added to
each of the links in the shortest path segment σL . The derivation of this formula
parallels that of gradient projection in Section 6.3.1. By shifting flow from the
longer path to the shorter path, we will either (i) equalize the travel times on
the two paths or (ii) shift all the flow onto the shorter path, and have it still
be faster than the longer path. In either case, we move closer to satisfying the
equilibrium condition.
The steps of Algorithm B are as follows:

1. Calculate the L and U labels in forward topological order.

2. Let i be the topologically last node in the bush.

3. Scan i by performing the following steps:

6.4. BUSH-BASED ALGORITHMS 179

(a) Use the L and U labels to determine the divergence node a and the
pair of alternate segments σL and σU .
(b) Calculate ∆h using equation (6.61) (capping ∆h at min(i,j)∈σU xij if
needed).
(c) Subtract ∆h from the x label on each link in σU , and add ∆h to the
x label on each link in σL .
4. If i = r, go to the next step. Otherwise, let i be the previous node
topologically and return to step 3.
5. Update all travel times tij and derivatives t0ij using the new flows x (re-
membering to add flows from other bushes.)
Demonstrating on the example in Figures 6.10 and 6.11, we start with the L
and U labels as shown in Figure 6.11(a) and (b), and start by letting i = 3, the
last node topologically. The longest path in the bush from the origin (r = 7)
to node i = 3 is [7,4,5,2,3], and the shortest path is [7,8,9,6,3], as can be easily
found from the L and U labels. The divergence node is the last node common
to both of these paths, which in this case is the origin, so a = 7. Equation (6.61)
gives ∆h = 1.56, so we shift this many vehicles away from the longest path and
onto the shortest path, giving the flows in Figure 6.12(a).
The second-to-last node topologically is node 2, and we repeat this process.
The longest and shortest paths from the origin to node 2 in the bush are [7,4,5,2]
and [7,4,1,2], respectively.10 The last node common to both of these paths is 4,
so the divergence node is a = 4 and we shift flow between the pair of alternate
segments [4,5,2] and [4,1,2]. Using equation (6.61) gives ∆h = 1.50. Shifting
this many vehicles from the longer segment to the shorter one gives the flows in
Figure 6.12(b).
The previous node topologically is node 5. Since there is only one incoming
bush link to node 5, there is nothing for Algorithm B to do. To see why, notice
that the longest and shortest bush paths are [7,4,5] and [7,4,5]. The divergence
node would be a = 5, which is the same as i, and the “pair of alternate segments”
is the empty paths [5] and [5]. Intuitively, since there is only one way to approach
node 5, there are no “alternate routes” to divert incoming flow. In fact, the same
is true for all of the previous nodes topologically (1, 4, 6, 9, 8, 7 in the reverse
of the order given above.), so there are no more flow shifts on the bush.

Origin-based assignment (OBA)

Rather than considering just two paths at a time, as Algorithm B did, OBA
shifts flow among many paths simultaneously. This can be both a blessing (in
that its moves affect more paths simultaneously) and a curse (in that its moves
are not as sharp as Algorithm B, which can focus on the two paths with the
10 Notice that we have not yet updated the travel time labels based on the shift at node 7.

You may do so if you wish, but it is not required for Algorithm B to work, and in all of the
examples in this section travel times are not updated until all flow shifts are complete for the
bush.
180 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

1 5.44 3 1 2.50 2 5.44 3

1 2
1 2.50 1.94
1 3.44 4.56 1 4.56
4 5 6 4 5 6
3.44 1.94
5.44 4.56 5.44 4.56
7 8 9 7 8 9
4.56 4.56 4.56 4.56
(a) Link flows x after scanning node 3 (b) Link flows x after scanning node 2

Figure 6.12: Algorithm B flow shifts.

greatest travel time difference). Because OBA shifts flow among many paths, it
makes use of the average cost labels M and the derivative labels D, rather than
the shortest and longest path costs (and the direct link travel times derivatives)
used by Algorithm B. OBA also makes use of a Newton-type shift, dividing a
difference in travel times by an approximation of this difference’s derivative.
When scanning a node i in OBA, first identify a least-travel time approach,
that is, a link (ĥ, i) such that Mĥi ≤ Mhi for all other approaches (h, i). This
link is called the basic approach in analogy to the basic path concept used in
path-based algorithms. Then, for each nonbasic approach (h, i), the following
amount of flow is shifted from xhi to xĥi :
Mhi − Mĥi
∆xhi = , (6.62)
Dhi + Dĥi − 2Da
with the constraint that ∆xhi ≤ xhi to prevent negative flows. It can be shown
that this formula would equalize the mean travel times on the two approaches if
they were linear functions. In reality, they are not, but the formula is still used
as an approximation.
Here a is the divergence node, defined for OBA as the node with the highest
topological order which is common to all paths in the bush from r to i, excluding
i itself. This is a reasonable place to truncate the search, since shifting flow
among path segments between a and i will not affect the flows on any earlier
links in the bush. This is a generalization of the definition used for Algorithm
B, needed since there are more than two paths subject to the flow shift.
After applying the shift ∆xhi to each of the links entering node i, we have to
update flows on the other links between a and i to maintain flow conservation.
This is done by assuming that the α values stay the same elsewhere, meaning
that any increase or decrease in the flow passing through a node propagates
backward to its incoming links in proportion to the contribution each incoming
link provides to that total flow. Thus, the x labels can be recalculated using
equations (6.47) and (6.48) to links and nodes topologically between a and i.
The steps of OBA are as follows:
1. Calculate the M and D labels in forward topological order.
2. Let i be the topologically last node in the bush.
6.4. BUSH-BASED ALGORITHMS 181

3. Scan i by performing the following steps:

(a) Determine the divergence node a corresponding to node i, and the

basic approach (ĥ, i).
(b) For each nonbasic approach (h, i) calculate ∆xhi using equation (6.62)
(with ∆xhi ≤ xhi ), subtract ∆xhi from xhi and add it to xĥi .
(c) Update αhi for every bush link terminating at i.
(d) Apply equations (6.47) and (6.48) to all nodes and links topologically
between a and i, updating their xij and xi values.

4. If i = r, go to the next step. Otherwise, let i be the previous node

topologically and return to step 3.

5. Update all travel times tij and derivatives t0ij using the new flows x (re-
membering to add flows from other bushes.)

Again demonstrating on the example in Figure 6.10, we use the M and D

labels shown in Figure 6.11. As with Algorithm B, we scan nodes in reverse
topological order. Starting with node 3, we compare the labels M23 and M63 to
determine the basic approach. Since M63 is lower, this is the least-cost approach
and named basic. For OBA, the divergence node corresponding to node 3 is node
7, since it is the only node common to all paths between nodes 7 and 3, except
for node 3 itself. Formula (6.62) is then applied to determine the amount of flow
∆x23 that should be shifted from approach (2, 3) to the basic approach (6, 3),
producing ∆x23 = 1.48. Shifting this flow updates the α values as shown in
Figure 6.13(b). To maintain flow conservation, we also need to adjust the flow
on other links in the bush. Holding the α proportions fixed at all other links,
these changes at node 3 are propagated back proportional to the original flows,
producing the link flows in Figure 6.13(a).
Next scanning node 2, we examine M12 , M42 , and M52 to determine the
basic approach: since M12 is smallest, it is deemed the basic approach. The
divergence node a corresponding to node 2 is node 4, since all paths from 7 to
2 pass through node 4 (and no other node with higher topological order). We
now apply two shifts, calculating ∆x42 and ∆x52 with equation (6.62). This
equation gives ∆x42 = 2.5, but since x42 = 0.79 it is only possible to move
0.79 units of flow from (4,2) to (1,2). For approach (5,2), ∆x52 = 1.50, so we
move 1.50 units of flow from (5,2) to (1,2). Assuming that the α values at all
other nodes remains constant, propagating this change back produces the x and
α labels in Figure 6.13(c) and (d). Notice that we only had to update flows
between the divergence node (a = 4) and the node being scanned (i = 2), since
no other links would be affected by shifting flow among approaches to node 2.
As with Algorithm B, no changes are made for the remaining nodes scanned,
since they only have one incoming link. This incoming link is trivially the
“basic” approach, and there are no nonbasic approaches to shift flow from.
182 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

1 0.79 2 5.52 3 1 1/7 2 0.552 3

0.79 1 0.448
0.79 3.94 4.48 1/7 5/7
4 5 6 4 5 6
3.94 1
5.52 4.48 1 1
7 8 9 7 8 9
4.48 4.48 1 1
(a) Link flows x after scanning node 3 (b)  labels after scanning node 3

1 3.08 2 5.52 3 1 0.558 2 0.552 3

3.08 1 0.448
0 2.44 4.48 0 0.442
4 5 6 4 5 6
2.44 1
5.52 4.48 1 1
7 8 9 7 8 9
4.48 4.48 1 1
(c) Link flows x after scanning node 2 (d)  labels after scanning node 2

Figure 6.13: OBA flow shifts.

Linear user cost equilibrium (LUCE)

The LUCE algorithm is similar in structure to OBA, using the same definition
of a divergence node and the same sets of labels (M and D). Where it differs
is in how nodes are scanned. When LUCE scans a node, it attempts to solve
a “local” user equilibrium problem, based on linear approximations to the ap-
proach travel times. To be concrete, when node i is scanned, the labels Mhi and
Dhi respectively indicate the average travel time for travelers arriving at i via
(h, i), and the derivative of this average travel time as additional vehicles arrive
via this link. Both of these are calculated based on the current bush flows, so if
the flow on approach (h, i) changes by ∆xhi , then the new average travel time
0
Mhi can be approximated by
0
Mhi ≈ Mhi + ∆xhi Dhi . (6.63)

The LUCE algorithm chooses ∆xhi values for each approach according to the
following principles:
P
(a) Flow conservation must be obeyed, that is, (h,i)∈B ∆xhi = 0.

(b) No link flow can be made negative, that is, xhi + ∆xhi ≥ 0 for all (h, i) ∈ B.
0
(c) The Mhi values should be equal and minimal for any approach with positive
0
flow after the shift, that is, if xhi + ∆xhi > 0, then Mhi must be less than
or equal to the M 0 label for any other approach to i.

The ∆xhi values satisfying this principles can be found using a “trial and error”
algorithm, like that introduced in Section 4.2.1. Choose a set of approaches (call
6.4. BUSH-BASED ALGORITHMS 183

it A), and set ∆xhi = −xhi for all (h, i) not in this set A. For the remaining
0
approaches, solve the linearPsystem of equations which set Mhi equal for all
(h, i) ∈ A and which have (h,i)∈A ∆xhi = 0. The number of equations will
equal the number of approaches in A. After obtaining such a solution, you
can verify whether the three principles are satisfied. Principle (a) will always
be satisfied. If principle (b) is violated, approaches with negative xhi + ∆xhi
should be removed from A. If principle (c) is violated, then some approach not
in A has a lower M 0 value, and that approach should be added to A. In either
of the latter two cases, the entire process should be repeated with the new A
set.
The steps of LUCE are as follows:

1. Calculate the M and D labels in forward topological order.

2. Let i be the topologically last node in the bush.

3. Scan i by performing the following steps:

(a) Determine the divergence node a corresponding to node i, and the

basic approach (ĥ, i).
(b) Use the process described above to find ∆xhi values satisfying the
local equilibrium principles (a)–(c) above, and add ∆xhi to xhi for
each approach to i.
(c) Update αhi for every bush link terminating at i.
(d) Use equations (6.47) and (6.48) to all nodes and links topologically
between a and i, updating their xij and xi values.

4. If i = r, go to the next step. Otherwise, let i be the previous node

topologically and return to step 3.

5. Update all travel times tij and derivatives t0ij using the new flows x (re-
membering to add flows from other bushes.)

Applying LUCE to the same example, we again start by scanning node 2. As-
suming that both approaches (2,3) and (6,3) will continue to be used, we solve
the following equations simultaneously for ∆x23 and ∆x63 :

M23 + D23 ∆x23 = M63 + D63 ∆x63

∆x23 + ∆x63 = 0

Substituting the M and D labels from Figure 6.11 and solving, we obtain ∆x23 =
−1.48 and ∆x63 = +1.48. Updating flows as in OBA (using the divergence node
a = 7 and assuming all α values at nodes other than 3 are fixed) gives the x and
α labels shown in Figure 6.14(a) and (b). Notice that this step is exactly the
same as the first step taken by OBA. The interpretation of LUCE and solving
a local, linearized equilibrium problem provides insight into how the OBA flow
shift formula (6.62) was derived.
184 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

1 0.79 2 5.52 3 1 1/7 2 0.552 3

0.79 1 0.448
0.79 3.94 4.48 1/7 5/7
4 5 6 4 5 6
3.94 1
5.52 4.48 1 1
7 8 9 7 8 9
4.48 4.48 1 1
(a) Link flows x after scanning node 3 (b)  labels after scanning node 3

1 2.85 2 5.52 3 1 0.516 2 0.552 3

2.85 1 0.448
0 2.67 4.48 0 0.484
4 5 6 4 5 6
2.67 1
5.52 4.48 1 1
7 8 9 7 8 9
4.48 4.48 1 1
(c) Link flows x after scanning node 2 (d)  labels after scanning node 2

Figure 6.14: LUCE flow shifts.

The shift is slightly different when there are three approaches to a node,
as happens when we proceed to scan node 2. First assuming that all three
approaches will be used, we solve these three equations simultaneously, enforcing
flow conservation and that the travel times on the three approaches should be
the same:

M12 + D12 ∆x12 = M42 + D42 ∆x42

M52 + D52 ∆x52 = M42 + D42 ∆x42
∆x12 + ∆x42 + ∆x52 = 0

Substituting values from Figure 6.11 and solving these equations simultaneously
gives ∆x12 = 2.82, ∆x42 = −1.85, and ∆x52 = −0.97. This is problematic, since
x42 = 0.79 and it is impossible to reduce its flow further by 1.85. This means
that approach (4,2) should not be used, so we fix ∆x42 = −0.79 as a constant
and re-solve the system of equations for ∆x12 and ∆x52 :

M12 + D12 ∆x12 = M42 − 0.79D42

M52 + D52 ∆x52 = M42 − 0.79D42
∆x12 − 0.79 + ∆x52 = 0

This produces ∆x12 = 2.06, and ∆x52 = −1.27, alongside the fixed value
∆x42 = −0.79. Updating flows on other links as in OBA produces the x
and α labels shown in Figure 6.14(b) and (c). No other shifts occur at lower
topologically-ordered nodes, because there is only one incoming link, and flow
conservation demands that no flow increase or decrease take place.
6.4. BUSH-BASED ALGORITHMS 185

6.4.3 Improving bushes

After shifting flow for the current bushes as described in the previous section,
the next step is to determine whether the bushes themselves need to change
(adding or dropping links) to allow us to move closer to equilibrium on the
entire network.
To begin, any link which is unused can be dropped from a bush without
disturbing the solution (unless it is needed for connectivity); and it can always
be added back later if we need to. As far as which links to add, one approach is
to add any “shortcut” links, defined based on the bush travel times. If we solve
each bush precisely to equilibrium in Step 2, then the distance from the origin r
to each node i is the same on any used path; call this Li . A “shortcut” link (i, j)
is any link for which Li + tij < Lj , where the L labels have been re-calculated
after eliminating the zero-flow bush links. Such a link can justifiably be called
a shortcut, because it provides a faster path to node j than currently exists in
the bush.
If this is our rule for adding links to the bush, it is easy to see that no cycles
are created. Reasoning by contradiction, assume that a cycle is created, say,
[i1 , i2 , . . . , ik , i1 ]. If link travel times are strictly positive, it is easy to see that
Lij+1 > Lij for any successive pair of nodes in the cycle (either (ij , ij+1 ) was
in the previous bush, in which case Lij + tij ij+1 = Lij+1 by equilibrium; or
(ij , ij+1 ) was just added, in which case Lij + tij ij+1 < Lij+1 ). Applying this
identity cyclically, we have Li1 < Li2 < . . . < Lik < Li1 , a contradiction since
we cannot have Li1 < Li1 .
The requirement that we solve each bush exactly to equilibrium in step 2 is
vital to ensuring no cycles are created. If this is not the case (and in practice, we
can only solve equilibrium approximately), another criterion is needed. To see
why, the argument used to show that no cycles would be created relied critically
on the assumption the difference in minimum cost labels for adjacent nodes in
the cycle was exactly the travel time of the connecting link (except possibly
for a link just added), which is only possible if all used paths have equal and
minimal travel time. Another one, which is almost as easy to implement, is to
add links to the bush based on the maximum travel time to a node using bush
links, as opposed to the minimum travel time. Although calculating longest
paths in general networks is difficult, in acyclic networks it can be done just as
efficiently as finding shortest paths (Section 2.4.1).
Assume that we have re-calculated the longest path labels Ui to each bush
node, after eliminating unused bush links as described above. We now define
a “modified shortcut” link as any link for which Ui + tij < Uj (this is the
same definition as a shortcut link but with maximum costs labels used instead
of minimum cost labels). Using a similar argument, we can show that adding
modified shortcut links cannot create cycles, even if the bush is not at equilib-
rium. Arguing again by contradiction, assume that the cycle [i1 , i2 , . . . , ik , i1 ] is
created by adding modified shortcut links, and consider in turn each adjacent
pair of labels Lij and Lij+1 . By the definition of maximum cost, for any link
which was in the bush before we have Uij + tij ij+1 = Uij+1 . For any link just
186 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

1 20.2 2 32.5 3 1 2 3
20.2 9.1
42 22.1
4 5 6 4 5 6
9.1 2
32.5 2 22.1
7 8 9 7 8 9
22.1 22.1
(a)Travel times t after LUCE update (b) Bush with unused links removed

1 2 3 1 2 3
(52.7,52.7) (50.7,72.9) (83.2,105.4)

4 5 6 4 5 6
(41.6,41.6) (66.3,66.3)
(32.5,32.5)

7 8 9 7 8 9
(0,0) (22.1,22.1) (44.2,44.2)
(c) Updated (L, U) labels for nodes (d) Bush with shortcut links added

Figure 6.15: Updating a bush by removing unused links and adding shortcuts.

added, we have Uij + tij ij+1 < Uij+1 ). Applying this identity cyclically, we have
Ui1 ≤ Ui2 ≤ . . . ≤ Uik ≤ Ui1 . Furthermore, since there were no cycles in the
bush during the previous iteration, at least one of these links must be new, and
for this link the inequality must be strict.
As an example, consider the bush updates which occur after performing the
LUCE example in the previous section. Figure 6.15 shows the updated link
travel times in panel (a), and the remaining bush links after zero-flow links are
removed in panel (b). Re-calculating L and U labels with the new travel times
and bush topology gives the values in Figure 6.15(c). At this point the three
unused bush links (4,2), (5,6), and (8,5), are examined to determine whether
Ui +tij < Uj for any of them. This is true for (5,6) and (8,5), since 41.6+2 < 66.3
and 22.1 + 2 < 41.6, but false for (4,2) since 32.5 + 42 ≥ 72.9. So, (5,6) and
(8,5) are added to the bush, as shown in Figure 6.15(d). From here, one can
return to the flow shifting algorithm to update flows further.

6.5 Likely Path Flow Algorithms

Recall from Section 5.2.2 that there is generally an infinite number of path
flow vectors h which satisfy the principle of user equilibrium. This contrasts
with the fact that the equilibrium link flow vector x̂ is unique as long as link
performance functions are increasing. That section introduced the principle of
entropy maximization as a way to select a equilibrium path flow believed to be
the most likely to occur in practice. The entropy-maximizing link flows solve
6.5. LIKELY PATH FLOW ALGORITHMS 187

the optimization problem

X X
max − hπ log(hπ /drs ) (6.64)
h
(r,s)∈Z 2 π∈Π̂rs
X
π
s.t. δij hπ = x̂ij ∀(i, j) ∈ A (6.65)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (6.66)
π∈Π̂rs
hπ ≥ 0 ∀π ∈ Π (6.67)

That section also introduced the proportionality condition, showing in The-

orem 5.4 that the ratio of the flows on any two paths connecting the same OD
pair depends only on the pairs of alternate segments where the paths differ,
not what that OD pair happens to be, or on any links where the paths coin-
cide. The proportionality condition is easier to verify, and entropy-maximizing
solutions must satisfy proportionality. Proportionality does not imply entropy
maximization, but in practical terms they seem nearly equivalent. Therefore,
this section focuses on finding proportional solutions, hence the term “likely
path flow algorithms” rather than “most likely.”
The proof that entropy maximization implies proportionality was based on
the Lagrangian of the entropy-maximization problem:
  !
X X hπ X X
π
L(h, β, γ) = − hπ log + βij x̂ij − δij hπ
drs
(r,s)∈Z 2 π∈Π̂rs (i,j)∈A π∈Π
 
X X
+ γrs drs − hπ  . (6.68)
(r,s)∈Z 2 π∈Π̂rs

After some algebraic manipulation, we obtained the formula

 
X
hπ = Krs exp − π
δij βij  , (6.69)
(i,j)∈A

where Krs is a proportionality constant associated with the OD pair (r, s) cor-
responding to path π. The value of KP rs can be found from the constraint that
total path flows must equal demand, π∈Π̂rs hπ = drs . The rest of the equa-
tion shows that the entropy-maximizing path flows are determined solely by the
Lagrange multipliers βij associated with each link, and therefore that the ratio
between two path flows for the same OD pair only depends on the links where
they differ.
Here we describe three ways to calculate likely path flows. The first method
is “primal,” and operates directly on the path flow vector h itself. The second
method is “dual,” operating on the Lagrange multipliers β in the entropy max-
imization problem, which can then be used to determine the path flows. Both
188 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

of these methods presume that an equilibrium link flow solution x̂ has already
been found. The third method, “traffic assignment by paired alternative seg-
ments” (TAPAS), is an algorithm which simultaneously solves for equilibrium
and likely paths.
Primal and dual methods for likely path flows have two major issues in
common:

1. The sets of equilibrium paths Π̂rs must be determined in some way. If x̂ is

exactly a user equilibrium, we can identify the shortest path for each OD
pair, and then set Π̂rs to be all the paths with that same cost between
r and s. In practice, though, we cannot solve for the equilibrium link
flows exactly, but only to a finite precision. So we cannot guarantee that
all of the used paths for an OD pair have the same travel time, and we
cannot guarantee that the unused paths are actually unused at the true
equilibrium solution.
2. The system of constraints (6.65) and (6.66) has many more variables than
equations; the number of equilibrium paths is typically much larger than
the number of OD pairs and network links. For primal methods, we need
to find and specify these “degrees of freedom” so we know how to adjust
path flows while keeping the solution feasible (matching equilibrium link
flows and the OD matrix). For dual methods, it means that many of the
link flow constraints (6.65) are redundant, and their corresponding βij
Lagrange multipliers are not needed. If these redundant constraints and
Lagrange multipliers are included, then there are infinitely many β values
that correspond to entropy maximization.
To address the first issue, we can include all paths in Π̂ that are within some
threshold  of the shortest-path travel time for each OD pair. Some care must be
taken to ensure that the resulting path sets are “consistent,” in that travelers
from different origins and destinations consider the same sets of alternatives,
when their path sets overlap. Figure 6.16 illustrates the difficulties in obtaining
a consistent solution. The top panel of the figure shows the travel times at the
true equilibrium solution, and the bottom panel shows the travel times at an
approximate equilibrium solution, such as the one obtained after stopping one
of the algorithms in this chapter after a finite number of iterations.
In this figure, assume there are two OD pairs, one from node 1 to node
2, and another from node 3 to node 4. At the true equilibrium solution,
there are three equal-cost paths between nodes 1 and 2 — [1, 2], [1, 3, 5, 6, 4, 2],
and [1, 3, 5, 7, 8, 6, 4, 2] — and two equal-cost paths between nodes 3 and 4 —
[3, 5, 6, 4] and [3, 5, 7, 8, 6, 4]. Travelers whose paths cross nodes 5 and 6 consider
two possible alternative routes between these nodes ([3, 5, 6, 4] and [3, 7, 8, 6, 4]),
regardless of which OD pair they came from. So defining Π̂12 and Π̂34 to include
the equal-cost paths listed above is consistent.
Now consider the bottom panel of the figure. There is now a unique shortest
path for each OD pair — [1, 2] and [3, 5, 6, 4] — but this does not reflect the true
equilibrium, simply the imprecision in an approximate solution. The threshold
6.5. LIKELY PATH FLOW ALGORITHMS 189

5 5 10
2 4 6 8

50 30 10

5 5 10
1 3 5 7
(a) Equilibrium link travel times
6 5 9
2 4 6 8

49 31 10

5 4 11
1 3 5 7
(b) Link travel times at an approximate equilibrium solution

Figure 6.16: Link travel times at the true equilibrium solution (top) and at an
approximate equilibrium solution (bottom).

 can be set to reflect this. Assume that  = 1, so that any path within 1 minute
of the shortest path travel time is included in Π̂. This choice gives the path sets
Π̂12 = {[1, 2], [1, 3, 5, 6, 4, 2]} and Π̂34 = {[3, 5, 6, 4], [3, 5, 7, 8, 6, 4]}. This choice
is not consistent, because it assumes travelers passing between nodes 5 and 6
consider different choices depending on their OD pair: travelers starting at node
1 only consider the segment [5, 6], while travelers starting at node 3 consider
both [5, 6] and [5, 7, 8, 6] as options. If travelers are choosing routes to minimize
cost, it should not matter what their origin or destination is. Increasing  to a
larger value would address this problem, but in a larger network would run the
risk of including paths which are not used at the true equilibrium solution.
So some care must be taken in how  is chosen. In the approximate equilib-
rium solution, we can define the acceptance gap ga to be the greatest difference
between a used path’s travel time and the shortest path travel time for its OD
pair, and the rejection gap gr to be the smallest difference between the travel
time of an unused path, and the shortest path for an OD pair. In Figure 6.17,
the OD pairs are from 1 to 6 and 4 to 9, and the links are labeled with their
travel times. The thick lines show the links used by these OD pairs, and the
thin lines show unused links. For travelers between nodes 1 and 6, the used
paths have travel times of 21 (shortest) and 22 minutes, and the unused path
has a travel time of 26 minutes. For travelers between nodes 4 and 9, the used
paths have travel times 18 (which is shortest) and 20 minutes, while the unused
path has a travel time of 21 minutes. The acceptance gap ga for this solution is
2 minutes (difference between 20 and 18), and the rejection gap is 5 (difference
between 26 and 21).
It is possible to show that if  is at least equal to the acceptance gap, but less
190 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

10 4
7 8 9

7 5 6

9 5
4 5 6

7 9 10

8 10
1 2 3

Figure 6.17: Example demonstrating rejection and acceptance gaps. Thick links
are parts of used paths at the equilibrium solution. Link labels are travel times.

than half of the rejection gap, then the resulting sets of paths Π̂rs are consistent
for the proportionality condition. That is, we need

gr
ga ≤  ≤ . (6.70)
2

In the network of Figure 6.17, any  choice between 2 and 2.5 will thus lead to
a consistent set of paths.
Exercise 21 gives a more formal definition of consistency and asks you to
prove this statement. It is not possible to choose such an  unless the equilibrium
problem is solved with enough precision for the acceptance gap to be less than
half of the rejection gap. To fully maximize entropy, rather than just satisfying
proportionality, demands a more stringent level of precision in the equilibrium
solution.
The second issue involves redundancies in the set of equations enforcing the
OD matrix and equilibrium link flow constraints. Properly resolving this issue
requires using linear algebra to analyze the structure of the set of equations (and
this in fact is the key to bridging the gap between proportionality and entropy
maximization), but for proportionality a simpler approach is possible.
A redundancy in a system of equations can be interpreted as a “degree of
freedom,” a dimension along which a solution can be adjusted. Consider the
network in Figure 6.18, which has two OD pairs (A to B, and C to D). The
equilibrium link flows are shown in the figure, along with the link IDs and
an indexing of the eight paths. The OD matrix and link flow constraints are
6.5. LIKELY PATH FLOW ALGORITHMS 191

A B

15 15
5 40 30 7
1 3
E F G
20 30
45 2 4 45
6 8

C D
OD pair (A, B) OD pair (C, D)
Path ID Links Path ID Links
1 5, 1, 3, 7 5 6, 1, 3, 8
2 5, 1, 4, 7 6 6, 1, 4, 8
3 5, 2, 3, 7 7 6, 2, 3, 8
4 5, 2, 4, 7 8 6, 2, 4, 8

Figure 6.18: Example demonstrating redundancies and “degrees of freedom”

when choosing a path flow solution. Links are labeled with equilibrium flows
(above) and link IDs (below).

reflected in the six equations

h1 + h2 + h3 + h4 = 15 (6.71)
h5 + h6 + h7 + h8 = 45 (6.72)
h1 + h2 + h5 + h6 = 40 (6.73)
h3 + h4 + h7 + h8 = 20 (6.74)
h1 + h3 + h5 + h7 = 30 (6.75)
h2 + h4 + h6 + h8 = 30 (6.76)

Equations (6.71) and (6.72) reflect the constraints that the total demand among
all paths from A to B, and from C to D, must equal the respective values in
the OD matrix. Equations (6.73)–(6.76) reflect the constraints that the flow
on links 1–4 must match their equilibrium values. Similar equations for links
5–8 are omitted, since they are identical to (6.71) and (6.72), as can easily be
verified.
This system has eight variables but only six equations, so there must be
at least two independent variables — in fact, there are four, since some of the
six equations are redundant. For example, adding equations (6.71) and (6.72)
gives you the same result as adding equations (6.73) and (6.74), so one of them
— say, (6.74) can be eliminated. Likewise, equation (6.76) can be eliminated,
since adding (6.71) and (6.72) is the same as adding (6.75) and (6.76). These
choices are not unique, and there are other equivalent ways of expressing the
192 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

same redundancies.
Each of these redundancies corresponds to an independent way to adjust the
path flows without affecting either total path flows between OD pairs, or total
link flows. A primal method uses these redundancies to adjust the path flows
directly, increasing the entropy of the solution without sacrificing feasibility
of the original path flow solution. A dual method uses these redundancies to
eliminate unnecessary link flow constraints — for instance, equations (6.74)
and (6.76) in the example above — so that the entropy-maximizing βij values
are unique.

6.5.1 Primal method

To find a proportional solution with a primal method, we need (1) the equilib-
rium path set Π̂; (2) the equilibrium link flows x̂; (3) an initial path flow solution
h0 , and (4) a list of “redundancies” in the link flow constraints, each of which
corresponds to a way to change path flows while maintaining feasibility. The
algorithm then adjusts the h values, increasing the entropy at each iteration,
until termination.11
The equilibrium path set and link flows were already described above. De-
pending on the algorithm used to solve for the equilibrium link flows, a path
flow solution may already be available. If you solved for equilibrium with a
path-based algorithm, its solution already contains the flows on each path. If
you used a bush-based algorithm, a corresponding path flow can be found using
the “within origins” adjustment technique described below. If you used a link-
based algorithm, it is not as easy to directly identify a path flow solution from
the final output — but it may be possible to track a path flow solution as the
algorithm progresses (for instance, in Frank-Wolfe, each “target” all-or-nothing
solution can be clearly identified with a path flow solution, and this can be
averaged with previous path flow solutions using the same λ values).
To achieve proportionality, it is not necessary to identify all of the redundan-
cies in the link flow constraints (it would be needed to fully maximize entropy).
It is enough to identify pairs of alternate segments between two nodes which are
used by multiple OD pairs. In Figure 6.18, links 1 and 2 are alternate segments
between nodes E and F, and links 3 and 4 are alternate segments between nodes
F and G.
The primal algorithm alternates between two steps: (1) adjusting h to
achieve proportionality for each origin r; and (2) for each pair of alternative
segments, adjusting h to achieve proportionality between those nodes. These
steps are described below; you can verify that each one of these steps preserves
feasibility of the solution, and increases entropy. If the solution remains un-
changed after performing both of these steps, then proportionality has been
achieved and we terminate. Other stopping criteria can be introduced, by mea-
suring the deviation from proportionality and terminating once this deviation
11 There are alternative solution representations that can make this algorithm much faster,

but would make the explanations more complicated. You are encouraged to think about how
to implement this algorithm efficiently, without having to list all paths explicitly.
6.5. LIKELY PATH FLOW ALGORITHMS 193

is sufficiently small.

Proportionality within origins

Section 5.2.3 described how a user equilibrium solution can be described by
the total flow from each origin r on each link (i, j), denoted xrij , and how at
equilibrium the links with positive xrij values must form an acyclic subnetwork,
a bush. It is easy to obtain such a solution if we have a path flow solution h:
X X
xrij = π π
δij h . (6.77)
s∈Z π∈Π̂rs

With these values, we can calculate the fraction of the flow from origin r ap-
proaching any node j from one specific link entering that node (i, j):
,
X
r
αij = xrij xrij (6.78)
(h,j)∈Γ−1 (j)

r
with αij defined arbitrarily if the denominator is zero. (This use of α is the
same as in the bush-based algorithms described in Section 6.4).
We can ensure that the proportionality condition holds between all paths
associated with an origin r by updating the path flows according to the formula
Y
hπ ← drs r
αij ∀s ∈ Z, π ∈ Π̂rs , (6.79)
(i,j)∈π

that is, by applying the aggregate approach proportions across all paths from
this origin to each individual path. One can show updating the path flows with
this formula will not change either the total OD flows or link flows, maintaining
feasibility, and will also increase entropy if there is any change in h.
This process is repeated for each origin r.

Proportionality between origins

To achieve proportionality between different origins, we consider pairs of alter-
nate segments between two nodes, used by multiple OD pairs. In Figure 6.18,
an example of a pair of alternate segments is links 1 and 2. In a larger network,
these alternate segments can contain multiple links. These links connect the
same two nodes (E and F), and are parts of equal-cost paths for both OD pairs
(A, B) and (C, D). If we move some flow from link 1 to link 2 from OD pair
(A, B), and move the equivalent amount of flow from link 2 to link 1 from OD
pair (C, D), we will not disturb the total link flows or OD flows, but will change
h and allow us to increase entropy.
Let σ = {σ1 , σ2 } be a pair of alternate segments, and let Z 2 (σ) be the set of
OD pairs which have paths using both of the alternate segments in σ. For each
194 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

segment σi , the segment flow g(σi ) is defined as the sum of flows on all paths
using that segment:
X X
g(σi ) = hπ , (6.80)
(r,s)∈Z 2 (σ) π∈Π̂rs :σi ⊆π

where the notation σi ⊆ π means that all of the links in the segment σi are in
the path π.
If the equilibrium path set Π̂ is consistent in the sense of (6.70), then any
path π which uses one segment of the pair has a “companion” path which is
identical, except it uses the other segment of the pair, denoted π c (σ). For
example, in Figure 6.18, if σ is the pair of alternate segments between nodes E
and F, the companion of path 1 is path 3, and the companion of path 6 is path
8.
To achieve proportionality between origins for the alternate segments in σ,
we calculate the ratios between g(σi ) values and apply the same ratios to the
path flows for each OD pair using this set of alternate segments:

c g(σi )
hπ ← (hπ + hπ (σ)
) ∀(r, s) ∈ Z 2 (σ), i ∈ {1, 2}, π ∈ Π̂rs ∧ π ⊇ σi .
g(σ1 ) + g(σ2 )
(6.81)
It is again possible to show that updating path flows with this formula leaves
total OD flows and link flows fixed, and can only increase entropy.

Example
This section shows how the primal algorithm can solve the example in Fig-
ure 6.18. Assume that the initial path flow solution is h1 = 15, h5 = 15,
h6 = 10, h8 = 20, and all other path flows zero. As shown in the first column
of Table 6.2, this solution satisfies the OD matrix and the resulting link flows
match the equilibrium link flows, so it is feasible. The entropy of this solution,
calculated using (6.64), is 47.7.12 Two pairs of alternate segments are identified:
links 1 and 2 between nodes E and F, and links 3 and 4 between nodes F and
G.
This table summarizes the progress of the algorithm in successive columns;
you may find it helpful to refer to this table when reading this section. The
bottom section of the table shows the origin-based link flows corresponding to
the path flow solution, calculated using (6.77).
The first iteration applies the within-origin formula (6.79) to origin A, and
to origin C. To apply the formula to origin A, the origin-based proportions are
first calculated with equation (6.78): α1A = 1, α2A = 0, α3A = 1, and α4A = 0,
and thus h1 ← 15, h2 ← 0, h3 ← 0, and h4 ← 0. (There is no change.) For
origin C, we have α1C = 5/9, α2C = 4/9, α3C = 1/3, and α4C = 2/3, and thus
h5 ← 8 31 , h6 ← 16 23 , h7 ← 6 23 , and h8 ← 13 13 . The entropy of this new solution
has increased to 59.6.
12 When computing this formula, 0 log 0 is taken to be zero, since limx→0+ x log x = 0.
6.5. LIKELY PATH FLOW ALGORITHMS 195

Next, we apply the between-origin formula to the pair of alternate segments

between nodes F and G. Paths 1, 3, 5, and 7 use link 3; and their companion
paths (2, 4, 6, and 8, respectively) use link 4. The segment flow for link 3 is
the sum of the path flows that use it (30), and similarly the segment flow for
link 4 is also 30. Thus formula (6.81) requires paths 1 and 2 to have equal flow,
paths 3 and 4 to have equal flow, and so on for each path and its companion.
Redistributing the flow between paths and companions in this way gives the
result in the third column of Table 6.2, and the entropy has increased to 72.5.
The between-origin formula is applied a second time to the other pair of
alternate segments, between nodes E and F. Paths 1, 2, 5, and 6 use link 1, and
the companion paths using link 2 are 3, 4, 7, and 8, respectively. The segment
flows for links 1 and 2 are 40 and 20, so for each path and its companion, the
path using link 1 should have twice the flow of the path using link 2. The fourth
column of Table 6.2 shows the results, and the entropy has increased again to
79.8.
This solution achieves proportionality (and in fact maximizes entropy). The
proportionality conditions can either be checked directly, or noticed when run-
ning the algorithm a second time does not change any path flows. For larger
networks with a more complicated structure, the algorithm generally requires
multiple iterations, and only converges to proportionality in the limit.

6.5.2 Dual method

An alternative approach involves the optimality conditions directly. Recall from
equation (6.69) that entropy-maximizing
 (and thus proportional) flow on each
π
P
path is Krs exp − (i,j)∈A δij βij , where Krs is an OD-specific constant cho-
sen so that the path flows sum to the total demand drs . We can adopt this
condition as a formula for h, and no matter what values are chosen for βij ,
the path flows we calculate satisfy proportionality. The difficulty is that they
will generally not be feasible, unless the link flows x corresponding to h happen
to equal their equilibrium values x̂. A dual algorithm tries to adjust these βij
values until x = x̂, at which point we terminate with proportional (and in fact
entropy-maximizing) path flows.
This contrasts with the primal approach in the previous section, which al-
ways maintained feasibility (the link flows in Table 6.2 never changed from the
equilibrium values) and worked toward optimality, expressed in equation (6.69).
The algorithm in this section always maintains optimality, and works toward
feasibility. Unlike the primal algorithm, the path flows in the dual algorithm
are not feasible until termination.
The idea behind the algorithm is simple enough: start with initial values
for βij on each link; calculate h from equation (6.69); calculate x from h; and
see which links have too much flow or too little flow. Adjust the βij values
accordingly, and iterate until the flow on every link is approximately equal to
its equilibrium value. There are a few details to take care of: how to find an
initial solution, when to terminate, how to adjust the βij values, and dealing
196 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Table 6.2: Demonstration of primal proportionality algorithm.

Initial solution Within-origin F/G PAS E/F PAS
Entropy 47.7 59.6 72.5 79.8
Path flows h1 15 15 7.5 5
h2 0 0 7.5 5
h3 0 0 0 2.5
h4 0 0 0 2.5
h5 15 8.33 12.5 15
h6 10 16.67 12.5 15
h7 0 6.67 10 7.5
h8 20 13.33 10 7.5
OD flows (A, B) 15 15 15 15
(C, D) 45 45 45 45
Total link flows x1 40 40 40 40
x2 20 20 20 20
x3 30 30 30 30
x4 30 30 30 30
From origin A xA
1 15 15 15 10
xA
2 0 0 0 5
xA
3 15 15 7.5 7.5
xA
4 0 0 7.5 7.5
From origin B xB
1 25 25 25 30
xB
2 20 20 20 15
xB
3 15 15 22.5 22.5
xB
4 30 30 22.5 22.5
6.5. LIKELY PATH FLOW ALGORITHMS 197

with redundancies in the system of constraints. The first two details are fairly
simple: any initial solution will do; β = 0 is simplest. Terminate when x is
“close enough” to x̂ according to some measure.
For adjusting the βij values, notice from equation (6.69) that increasing βij
will decrease xij , and vice versa. So a natural update rule is

βij ← βij + α(xij − x̂ij ) . (6.82)

where α is a step size, and xij − x̂ij is the difference between the link flows
currently implied by β, and the equilibrium values. In addition to this intuitive
interpretation, this search direction is also proportional to the gradient of the
least-squares function X
φ(x) = (xij − x̂ij )2 . (6.83)
(i,j)∈A

This function is zero only at a feasible solution, and (6.82) is a steepest descent
direction in terms of x.13
Equation (6.83) can also be used to set the step size α. One can select a trial
sequence of α values (say, 1, 1/2, 1/4, 1/8, . . .), evaluating each α value in turn
and stopping once the new x values reduce (6.83). A more sophisticated step
size rule chooses α using Newton’s method, to approximately maximize entropy.
Newton’s method also has the advantage of scaling the step size based on the
effect changes in β have on link flows. Exercise 23 develops this approach in
more detail.
A last technical detail concerns redundancies in the system of link flow equa-
tions, as discussed at the end of Section 6.5. Redundancies in the link flow
equations mean that the βij values maximizing entropy may not be unique. To
resolve this issue, redundant link flow constraints can be removed, and their βij
values left fixed at zero. Practical experience shows that this can significantly
speed convergence.

Example
The dual method is now demonstrated on the same example as the primal
algorithm; see Figure 6.18. You may find it helpful to refer to Table 6.3 when
reading this section to track the progress of the algorithm. The format is similar
to Table 6.2, except for additional rows showing the βij values.
This table summarizes the progress of the algorithm in successive columns;
you may find it helpful to refer to this table when reading this section. The
bottom section of the table shows the origin-based link flows corresponding to
the path flow solution, calculated using (6.77).
To begin, as discussed at the end of Section 6.5, two of the link flow con-
straints are redundant, and their βij values are fixed at zero. Assume that links
2 and 4 are chosen for this purpose, so β2 = β4 = 0 throughout the algorithm.
(The algorithm would perform similarly for other choices of the two redundant
13 To be precise, it is not a steepest descent direction in terms of β. An alternative derivation

of equation (6.82) is explored in Exercise 22.

198 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

links; note that we are also continuing to ignore the link flow constraints associ-
ated with links 5–8, since these are identical to the OD matrix constraints (6.71)
and (6.72).)  P 
π
Initially, β1 = β3 = 0. This means that exp − (i,j)∈A δij βij = 1 for all
paths, so h1 = h2 = h3 = h4 = KAB and h5 = h6 = h7 = h8 = KCD . To
ensure that the sum of each OD pairs’ path flows equals the total demand, we
need KAB = 3.75 and KCD = 11.25, and equation (6.69) gives the flows on each
path, as shown in the Iteration 0 column of Table 6.3.
The table also shows the link flows corresponding to this solution: links 1–4
all have 30 vehicles, whereas the equilibrium solution has x1 = 40 and x2 = 20.
The least-squares function (6.83) has the value (40 − 30)2 + (20 − 30)2 = 200.
Trying an initial step size of α = 1 would give β1 = 0 + 1 × (30 − 40) = −10. The
other βij values are unchanged: β3 remains at zero because it has the correct
link flow, while β2 and β4 are permanently fixed at zero because their link flow
constraints were redundant. Re-applying equation (6.69) with this new value
of β1 (and recalculating KAB and KCD to satisfy the OD matrix) would give
x1 = 60, x2 = 0, and x3 = x4 = 30. This has a larger least-squares function
than before (800 vs. 200), so we try again with α = 1/2. This is slightly better
(the least-squares function is 768), but still worse than the current solution.
After two more trials, we reach α = 1/8, which produces a lower mismatch
(88).
This step size is accepted, and we proceed to the next iteration. The path
and link flows are shown in the Iteration 1 column of Table 6.3. The flows on
links 1 and 2 are closer to their equilibrium values than before. Continuing as
before, we find that α = 1/8 is again the acceptable step size with the new link
flow values, so β1 ← −1.25 + 81 (36.2 − 40) = −0.42, producing the values in the
Iteration 2 column. Over additional iterations, the algorithm converges to the
final values shown in the rightmost column.
It is instructive to compare the dual algorithm in Table 6.3 with the primal
algorithm in Table 6.2. Notice how the dual algorithm always maintains propor-
tionality, and the link flows gradually converge to their equilibrium algorithms.
By contrast, the primal algorithm maintains the link flows at their equilibrium
values, and gradually converges to proportionality. The entropy also does not
change monotonically, and at times it is higher than the maximum entropy value
— this can only happen for an infeasible solution.

6.5.3 Traffic assignment by paired alternative segments

The primal and dual methods described in the preceding sections assumed that
user equilibrium link flows were already available, and then found likely path
flows as a post-processing step. Traffic assignment by paired alternative seg-
ments (TAPAS) is an algorithm which finds the equilibrium solution and pro-
portional path flows simultaneously. Interestingly, accomplishing both tasks at
once does not slow down the algorithm. TAPAS is in fact among the fastest of
the traffic assignment algorithms currently known.
6.5. LIKELY PATH FLOW ALGORITHMS 199

Table 6.3: Demonstration of dual proportionality algorithm.

Iteration 0 1 2 ··· ∞
Entropy 83.2 73.4 81.9 79.8
β1 0 −1.25 −0.42 −0.692
β3 0 0 0 0
Path flows h1 3.75 5.83 4.53 5
h2 3.75 5.83 4.53 5
h3 3.75 1.67 2.97 2.5
h4 3.75 1.67 2.97 2.5
h5 11.25 17.49 13.58 15
h6 11.25 17.49 13.58 15
h7 5.01 8.92 6.53 7.5
h8 5.01 8.92 6.53 7.5
OD flows (A, B) 15 15 15 15
(C, D) 45 45 45 45
Total link flows x1 30 46.6 36.2 40
x2 30 13.4 23.8 20
x3 30 30 30 30
x4 30 30 30 30
From origin A xA
1 7.5 11.7 9.1 10
xA
2 7.5 3.3 5.9 5
xA
3 7.5 7.5 7.5 7.5
xA
4 7.5 7.5 7.5 7.5
From origin B xB
1 22.5 35.0 27.1 30
xB
2 22.5 10.0 17.8 15
xB
3 22.5 22.5 22.5 22.5
xB
4 22.5 22.5 22.5 22.5
200 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Figure 6.19: A grid network consisting of 100 city blocks.

Recall from Section 6.4 that path- and bush-based algorithms find the equi-
librium solution by shifting flow from longer paths to shorter ones, and that
these paths often differ on a relatively small set of links (in gradient projection,
we denoted these by the set A3 ∪A4 ; in bush-based algorithms, by the concept of
a divergence node). The main insights of TAPAS are that these algorithms tend
to shift flow repeatedly between the same sets of links, and that these links are
common to paths used by between different origins and destinations. As a re-
sult, it makes sense to store these path segments from one iteration to the next,
rather than having to expend effort finding them again and again. Furthermore,
since these links are common to multiple origins, we can apply proportionality
concepts at the same time to find a high-entropy path flow solution.
Pairs of alternative segments can also form a concise representation of the
equilibrium conditions. In the grid network of Figure 6.19, the number of paths
between the origin in the lower-left and the destination in the upper-right is
rather large (in fact there are 184,756) even though the network is a relatively
modestly-sized grid of ten rows and columns. If all paths are used at equilibrium,
expressing the equilibrium condition by requiring the travel times on all paths
6.5. LIKELY PATH FLOW ALGORITHMS 201

be equal requires 184,755 equations.

The network can also be seen as a hundred “city blocks,” each of which
can either be traversed in the clockwise direction (north, then east) or in the
counterclockwise direction (east, then north). If travel times on all paths in the
network are identical, then the travel time around each block must be the same
in the clockwise and counterclockwise directions. In fact, the converse is true
as well: if the travel time is the same around every block in both orientations,
then the travel times on all paths in the network are the same as well. We can
thus express equality of all network paths with only 100 equations!
Furthermore, other nodes in the network may serve as origins and destina-
tions, not just nodes at two corners. Expressing equilibrium in terms of path
travel time equality requires additional equations for each new OD pair, but the
same 100 equations expressing equality of travel times around each block are
sufficient no matter how many nodes serve as origins or destinations.
This discussion implies that most of the 184,755 path travel-time equations
are redundant. It is not trivial to identify these linear dependencies, but meth-
ods based on paired alternative segments are a way to do so. The two ways to
travel around each block can be seen as a pair of alternative segments between
their southwestern and northeastern nodes, and the set of all such pairs of al-
ternative segments is “spanning” in the sense that one can shift flow between
any two paths with the same origin and destination by shifting flow between a
sequence of pairs of alternative segments. There are some subtleties involving
the equivalence of equilibrium conditions on pairs of alternative segments and
on paths (see Exercise 24), but this example shows how they can often simplify
the search for a user equilibrium solution.
The TAPAS algorithm represents network flows aggregated by origin, with
xrij denoting the flow on link (i, j) which started at origin r, following equa-
tion (5.30). The algorithm also involves a set of pairs of alternative segments
(PASs). Each PAS ζ is defined by two path segments σ1ζ and σ2ζ starting and end-
ing at the same nodes, and by a list of relevant origins Zζ indicating a subset of
zones which have flow on both path segments at the current solution. The steps
of the algorithm involve maintaining a set of PASs (creating new ones, updating
relevant origins, and optionally discarding inactive ones), and adjusting the link
flows xrij to move towards user equilibrium and proportionality. Notice that
TAPAS does not store the path flows h explicitly, for computational reasons.
Instead, the path flows are represented implicitly, obtained from the xrij values
using the proportional split formulas previously introduced in Section 6.5.1:
,
X
r r
αij = xij xrij (6.84)
(h,j)∈Γ−1 (j)

where X X
xrij = π π
δij h . (6.85)
s∈Z π∈Π̂rs

An example of a PAS is shown in Figure 6.20. For each link, the upper and
lower labels give the flows on that link from Origin 1 and Origin 2, respectively.
202 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

2 4 2 Flow from Origin 1

1 2 Flow from Origin 2
4
0 0 i j
1 2 1 2 5
0
2
15
d =4 1 3
PAS ζ Segment 1 Segment 2
d25 = 2 σ1=[2,4] σ2=[2,3,4]
Relevant origins: Zζ = {2}

Figure 6.20: Example for demonstrating pairs of alternative segments.

There are two path segments: [2, 4] and [2, 3, 4], and there is one relevant origin
(Origin 2). You might expect that this PAS is also relevant to Origin 1; and
indeed at the ultimate equilibrium solution this will be true. However, in large
networks it is not immediately obvious which PASs are relevant to which origins,
and the TAPAS algorithm must discover this during its steps.
There are three main components to the algorithm: PAS management, flow
shifts, and proportionality adjustments. PAS management involves identifying
new PAS, updating the lists of relevant origins, and removing inactive ones.
Flow shifts move the solution closer to user equilibrium, by shifting vehicles
from longer paths to shorter ones. Proportionality adjustments maintain the
total link flows at their current values, but adjust the origin-specific link flows
to increase the entropy of the path flow solution implied by (6.84). One possible
way to perform these steps is as follows; the rest of this subsection fills out the
details of each step.
1. Find an initial origin-disaggregated solution, and initialize the set of PASs
to be empty.
2. Update the set of PASs by determining whether new ones should be cre-
ated, or whether existing ones are relevant to more origins.
3. Perform flow shifts within existing PASs.
4. Perform proportionality adjustments within existing PASs.
5. Check for convergence, and return to step 2 unless done.
This algorithmic description may appear vague. Like many of the fastest al-
gorithms currently available, the performance of the algorithm depends on suc-
cessfully balancing these three components of the algorithm. The right amount
of time to spend on each component is network- and problem-specific, and im-
plementations that make such decisions adaptively, based on the progress of the
algorithm, can work well.
6.5. LIKELY PATH FLOW ALGORITHMS 203

Updating the set of pairs of alternative segments

In its second step, the TAPAS algorithm must update the set of PASs. Since
flow shifts mainly occur within PASs, finding the user equilibrium solution relies
on being able to identify new alternative routes which are shorter than the ones
currently being used. Given the origin-disaggregated solution xr for each origin
r, we can search for routes in the following way.
Solving a shortest path algorithm over the entire network produces a tree
rooted at an origin r, containing paths to every node. (Note that already having
an origin-disaggregated solution can greatly accelerate the process of finding
shortest paths; see Exercise 48 from Chapter 2.) At equilibrium, essentially all
of these links should be used.14 By comparing this tree to the links which are
used in the disaggregate solution xr , we can identify any links in the shortest
path tree not currently being used by an origin r even though there is flow to
their head nodes.
More specifically, let Âr denote the set of these links. A link (i, j) is in Âr if
(i, j) is in the shortest path tree rooted at r, if xrij = 0 (it is currently unused by
origin r), and if there is some other link (h, j) for which xrhj (flow from origin r
is reaching the head node j in another way). We then look for a PAS ζ whose
two segments σ1ζ and σ2ζ end with the links (h, j) and (i, j), which will allow us
to shift flow onto the shortest path segment.
Two possibilities exist: either such a PAS already exists (in which case we
add origin r to the relevant set Zζ if it is not already listed), or we create a
new one. To create a new PAS, we must choose two path segments σ1ζ and σ2ζ
which start and end at the same node. The first segment σ1ζ should consist of
links with positive flow from origin r, and the second one σ2ζ should consist of
links from the shortest path tree. We also know they must both end at node j,
but must choose an appropriate node for them to start (a divergence node). As
discussed in Section 6.4.2, there are several ways to choose divergence nodes.
For TAPAS, an ideal divergence node results in short path segments σ1ζ and
σ2ζ . Short segments both result in faster computation, and intuitively are more
likely to be relevant to more origins.
Putting these concepts together, we can search for a divergence node a for
which there is a segment σ1ζ starting at a, ending with link (h, j), and only using
links with positive flow from origin r; and a segment σ2ζ starting at a, ending
with link (i, j), and only using links in the current shortest path tree. Among
all such divergence nodes and segments, we want one for which the two path
segments are short.
As an example, consider again the network from Figure 6.20. The link
performance functions and current travel times are shown in Figure 6.21. The
bold links show the shortest path tree rooted at Origin 1, and see that there are
two links used by this origin which are not part of this tree: links (2, 4) and (3, 5).
For link (2, 4), we see that the two segments of the PAS a from the previous
14 The only exceptions would be to links leading to nodes not used by this origin, or if

multiple paths are tied for being shortest with one of them having zero flow.
204 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

2 4 2 Flow from Origin 1

1 2 Flow from Origin 2
4
0 0 i j
1 2 1 2 5
0
2
15
d =4 1 3
d25 = 2
Link performance function
50+x 4 10x Travel time
0 53 40 i j
0 10+x
1 2 11 50+x 5
52
10x
30 3
PAS a Segment 1 Segment 2
σ1=[2,4] σ2=[2,3,4]
Relevant origins: Zζ = {2}

Figure 6.21: Creating a new PAS. The top panel shows the origin-specific flows,
the bottom panel the link performance functions and current times. Bold links
are the shortest path tree for Origin 1.

example include a segment of links used by this origin [2, 4], and a segment of
links from the shortest path tree [2, 3, 4] that have a common divergence node
2. At this point, we declare Origin 1 relevant to this PAS by adding it to the
set Za .
For link (3, 5), we need to create a new PAS. There are two possibilities
for choosing segments: one choice is [3, 5] and [3, 4, 5]; and the other choice is
[2, 3, 5] and [2, 3, 4, 5] (both involve a segment of used links and a segment from
the shortest path tree, starting at a common divergence node). The first one
is preferred, because it has fewer links — and in fact the common link (2, 3)
in the second PAS is irrelevant, since shifting flow between segments will not
change flow on such a link at all). By being shorter, there are potentially more
relevant origins. If node 3 were also an origin, it could be relevant to the first
choice of segments, but not the second. Therefore we create a new PAS b, and
set σb1 = [3, 5], σb2 = [3, 4, 5], and Zb = {1}. (The choice of which segment is the
first and second one is arbitrary.)
Repeating the same process with the shortest path tree from Origin 2, we
6.5. LIKELY PATH FLOW ALGORITHMS 205

verify that it is relevant to PAS a (which it already is), and add it as relevant
to PAS b, so Za = Zb = {1, 2}. (Both origins are now relevant to both PASs).

Flow shifts
TAPAS uses flow shifts to find an equilibrium solution. There are two types of
flow shifts: the most common involves shifting flow between the two segments
on an existing PAS. The second involves identifying and eliminating cycles of
used links for particular origins.
For the first type, assume we are given a PAS ζ, and without loss of generality
assume that the current travel time on the first segment σ1ζ is greater than that
on the second σ2ζ . We wish to shift flow from the first segment to the second
one to either equalize their travel times, or to shift all the flow to the second
path if it is still shorter. The total amount of flow we need to shift to equalize
the travel times is approximately given by Newton’s method:
P P
(i,j)∈σ1ζ tij − (i,j)∈σ2ζ tij
∆h = P 0
P 0 . (6.86)
(i,j)∈σ ζ tij +
1 (i,j)∈σ ζ tij
2

We must also determine whether such a shift is feasible (would shifting this much
flow force an xrij value to become negative?) and, unlike the algorithms earlier
in this chapter, how much of this flow shift comes from each of the relevant
origins in Zζ .
To preserve feasibility, for any relevant origin r, we must subtract the same
amount from xrij for each link in the longer segment, and add the same amount
to each link in the shorter segment. Call r this amount ∆hr . The non-negativity
r r
constraints require ∆h ≤ min(i,j)∈πζ xij ; let ∆h denote the right-hand side
1
of this inequality, which must hold for every relevant origin.
If X r
∆h ≤ ∆h , (6.87)
r∈Zζ

then the desired shift is feasible, and we choose the origin-specific shifts ∆hr to
r
be proportional to their maximum values ∆h to help maintain proportionality.
r
If this shift is not feasible, then we shift as much as we can by setting ∆hr = ∆h
for each relevant origin.
An example of such a shift is shown in Figure 6.22, continuing the example
from before. The left side of the figure shows the state of the network prior
to the flow shift. The top panel shows the current origin-specific link flows;
the middle panel the current travel times and travel time derivatives; and the
bottom panel shows the structure of both PASs. Starting with PAS a, we first
calculate the desired total shift from equation (6.86):
53 − (30 + 11)
∆h = = 1. (6.88)
2 + (10 + 1)
For origin 1, we can subtract at most 2 units of flow from segment 1, and for
origin 2, we can subtract at most 1. Removing any more would result in negative
206 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

2 4 2 4/3 4 2
4 1 2 4 2/3 2
0 0 0 2/3
1 2 1 2 5 1 2 4/3 2 5
0 0
2 Flow from Origin 1 8/3
d15 = 4 1 3 Flow from Origin 2 4/3 3
d25 = 2 i j

53 4 40 52 4 40
0 1 10 0 1 10
0 0
1 2 11 5 1 2 12 5
52 52
1 1
30 1 40 1
Travel time
10 3 10 3
Derivative
i j

PAS a Segment 1 Segment 2

σ1=[2,4] σ2=[2,3,4]
Relevant origins: Za = {1,2}

PAS b Segment 1 Segment 2

σ1=[3,5] σ2=[3,4,5]
Relevant origins: Za = {1,2}

Figure 6.22: Example of flow shifts using TAPAS; left panel shows initial flows
and times, right panel shows flows and times after a shift for PAS a.
6.5. LIKELY PATH FLOW ALGORITHMS 207

origin flows on link (2, 4). We thus split ∆h in proportion to these maximum
allowable values, yielding

∆h1 = 2/3 ∆h2 = 1/3 (6.89)

and producing the solution shown in the right half of Figure 6.22. Since the link
performance functions are linear, Newton’s method is exact, and travel times
are equal on the two segments of PAS a.
Moving to the second PAS, we see that it is at equilibrium as well, and no
flow shift is done: the numerator of equation (6.86) is zero. In fact, the entire
network is now at equilibrium, but the origin-based link flows do not represent
a proportional solution. Proportionality adjustments are discussed below.
The second kind of flow shift involves removing flow from cycles. In TAPAS,
there may be occasions where cycles are found among the links with positive flow
(xrij > 0). Such cycles can be detected using the topological ordering algorithm
described in Section 2.2.
In such cases, we can subtract flow from every link in the cycle, maintaining
feasibility and reducing the value of the Beckmann function. (See Exercise 26).
Let X denote the minimum value of xrij among the links in such a cycle. After
subtracting this amount from every xrij in the cycle, the solution is closer to
equilibrium and the cycle of positive-flow links no longer exists.
Figure 6.23 shows an example of how this might happen. This network has
only a single origin, and two PASs. Applying the flow shift formula, we move
1 vehicle from segment [1, 2] to [1, 3, 2], and 1 vehicle from segment [2, 4] to
segment [2, 3, 4]. This produces the flow solution in the lower-left of the figure,
which contains a cycle of flow involving links (2, 3) and (3, 2). If we subtract 1
unit of flow from both of those links, we have the flow solution in the lower-right.
This solution is feasible and has a lower value of the Beckmann function, as you
can verify.

Proportionality adjustments

TAPAS uses proportionality adjustments to increase the entropy of the path

flow solution. Note that the path flow solution is not explicitly stored, since
the number of used paths can grow exponentially with network size. Rather,
a path flow solution is implied by the bush, using the procedure described in
Section 6.5.1, and the definitions of α and h in equations (6.84) and (6.85).
r
That is, we calculate approach proportions αij using the formula
In a proportionality adjustment, we shift flows between segments in the PAS
without changing the total flow on each link, so some origins will shift flow from
the links in σ1ζ to those in σ2ζ , while other origins will shift flow from σ2ζ to
σ1ζ . Following the previous section, we will use ∆hr to denote the amount of
flow shifted from each link in σ1ζ to each link in σ2ζ , using negative numbers
to indicate flow shifting from σ2ζ to σ1ζ . To maintain total link flows at their
208 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

3 3
5 5 6 5

1 0 0 4 1 1 0 4
5 5
5 4
2 2

(a) (b)

3 3
6 6 6 6

1 1 1 4 1 0 0 4
4 4
4 4
2 2

(c) (d)

Figure 6.23: (a) Initial flow solution; (b) After a flow shift at a PAS ending at
3; (c) After a flow shift at a PAS ending at 4; (d) After removing a cycle of flow.
6.5. LIKELY PATH FLOW ALGORITHMS 209

current levels, we will require

X
∆hr = 0 . (6.90)
r∈Zζ

To describe the problem more formally, let b denote the node at the down-
stream end of the PAS, and let xb denote the total flow through this node as
in Equation (6.4.1) . We can calculate the flow on the segments σ1ζ and σ2ζ for
each relevant origin r with the formulas
Y xrij
g r (σ1ζ ) = xrb (6.91)
xrj
(i,j)∈σ1ζ
Y xrij
g r (σ2ζ ) = xrb , (6.92)
xrj
(i,j)∈σ2ζ

(6.93)
assuming positive flow through all nodes in the segment (xrj > 0).15 If propor-
tionality were satisfied, we would have
r0 ζ
P
g r (σ1ζ ) r 0 ∈Zζ g (σ1 )
= (6.94)
g r (σ1ζ ) + g r (σ2ζ )
 
P r0 ζ r0 ζ
r 0 ∈Zζ g (σ1 ) + g (σ2 )

for all relevant origins.

After applying segment shifts of size ∆hr , the new segment flows will be
given by
Y xrij − ∆hr
g r (σ1ζ ) = xrb (6.95)
xr − ∆hr [r 6= j]
(i,j)∈π1 j
Y xrij + ∆hr
g r (σ2ζ ) = xrb , (6.96)
xrj + ∆hr [r 6= j]
(i,j)∈π2

(6.97)
using brackets for an indicator function. We aim to find ∆hr values satisfy-
ing constraint (6.90) and (6.94), where the segment flows are computed with
equations (6.95) and (6.96).
Solving this optimization problem exactly is a bit difficult because equa-
tions (6.84) and (6.85) are nonlinear. A good approximation method is de-
veloped in Exercise 27. A simpler heuristic is to adapt the “proportionality
between origins” technique from Section 6.5.1 and approximate the (nonlinear)
formulas (6.95) and (6.96) by the (linear) formulas

g r (σ1ζ ) ≈ g0r (σ1ζ ) − ∆hr (6.98)

g r
(σ2ζ ) ≈ g0r (σ2ζ ) + ∆h r
(6.99)
15 What would happen if this were not true?
210 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

where g0r (σ1ζ ) and g0r (σ2ζ ) are the current segment flows (with zero shift).
Substituting equations (6.98) and (6.99) into (6.94) and simplifying, we ob-
tain

0
g0r (σ1ζ )
P
r 0 ∈Zζ
r
∆h = g0r (σ1ζ ) − (g0r (σ1ζ ) + g0r (σ2ζ )) P   (6.100)
g0r (σ1ζ ) + g0r (σ2ζ )
0 0
r 0 ∈Zζ

as a flow shift to heuristically move toward proportionality.

This heuristic is in fact exact if the PAS is “isolated” in the sense that flow
does not enter or leave the segments in the middle, so αijr
= 1 for all links in σ1ζ
ζ
and σ2 except for the last links of each segment.
To illustrate how this procedure works, consider the example in Figure 6.24.
In this example, there are multiple destinations in addition to multiple origins.
The figure shows the path flow solution implied by the link flow solution at
the left. We emphasize that TAPAS does not maintain the path flow solution
explicitly, and the path flows are constructed from the link flows using equa-
tion (6.84). For instance, the flow on path [1, 3, 4] from Origin 1 is calculated
as 3 × 31 × 11 = 1. This solution does not satisfy proportionality. All the vehicles
from Origin 1 passing between nodes 2 and 4 use segment [2,4], while all of
those from Origin 2 passing between these nodes use segment [2,3,4]. Since the
total flow on the two segments are equal (two vehicles on each), flow from both
origins should split equally between the two segments.
Applying equation (6.100) gives the shifts shown in Figure 6.25. This figure
also shows the new origin-specific link flows and implied path flows. In this
case, the link flows on the segments of PAS a now satisfy proportionality. This
is a case where the PAS is isolated, because no vehicles entered and left the
segments in the middle, and the heuristic formula (6.100) is exact.
To show how the formula is inexact for a non-isolated PAS, consider the
modification of this example shown in Figure 6.26. The only change is that
node 3 is now a destination for Origin 1, with a demand of 1, and that as a
result the flow from Origin 1 on link (1,3) is increased by one vehicle. Repeating
the same process as above, and applying (6.100), we again one swap one vehicle
between each pair of segments. This results in the situation in Figure 6.27.
The origin-specific link flows are the same as in Figure 6.25, except for the
additional vehicle from Origin 1 on link (1,3). But the implied path flows are
quite different! This is because the additional vehicle shifted onto link (3,4) was
“split” between incoming links (1,3) and (2,3), according to equation (6.84),
rather than allocated solely to (2,3). (Again, TAPAS does not store the flows
on individual paths, and must calculate them implicitly using this formula.) As
a result, proportionality is still not satisfied: between segments [2,4] and [2,3,4],
Origin 1 splits in the ratio of 3:2, whereas Origin 2 splits in the ratio 1:1. This
is closer to proportionality from before, but not exact. Repeated applications of
the heuristic shift formula will converge to a proportional solution, in this case.
6.5. LIKELY PATH FLOW ALGORITHMS 211

Flow from Origin 1

1 3 1 Flow from Origin 2
0 2
0 i j
1 3 4

2 d14 = 3
2
0 2 d23 = 1
0
d24 = 2

Implied path flows: PAS Segment 1 Segment 2

[1,2,4] 2 [2,4] [2,3,4]
[1,2,3,4] 0 Origin 1 2 0
[1,3,4] 1 Origin 2 0 2

[2,3] 1 Total 2 2

[2,4] 0
[2,3,4] 2

Figure 6.24: Example of a proportionality adjustment in TAPAS.

212 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Flow from Origin 1

1 3 1 Flow from Origin 2
0 1
1 i j
1 2 4

2 d14 = 3
1
0 2 d23 = 1
1
d24 = 2

Implied path flows: PAS Segment 1 Segment 2

[1,2,4] 1 [2,4] [2,3,4]
[1,2,3,4] 1 Origin 1 1 1
[1,3,4] 1 Origin 2 1 1

[2,3] 1 Total 2 2

[2,4] 1
[2,3,4] 1

Figure 6.25: Updated flows after a proportionality adjustment for PAS a.

6.5. LIKELY PATH FLOW ALGORITHMS 213

Flow from Origin 1

2 3 1 Flow from Origin 2
0 2
0 i j
1 3 4
d13 = 1
2 2 d14 = 3
0 2 0 d23 = 1
d24 = 2

Implied path flows: PAS Segment 1 Segment 2

[1,2,3] 0 [2,4] [2,3,4]
[1,3] 1 Origin 1 2 0
Origin 2 0 2
[1,2,4] 2
[1,2,3,4] 0 Total 2 2
[1,3,4] 1

[2,3] 1

[2,4] 0
[2,3,4] 2

Figure 6.26: Example of a proportionality adjustment for a non-isolated PAS.

214 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Flow from Origin 1

2 3 2 Flow from Origin 2
0 1
1 i j
1 2 4
d13 = 1
2 1 d14 = 3
0 2 1 d23 = 1
d24 = 2

Implied path flows: PAS Segment 1 Segment 2

[1,2,3] 1/3 [2,4] [2,3,4]
[1,3] 2/3 Origin 1 1 2/3
Origin 2 1 1
[1,2,4] 1
[1,2,3,4] 2/3 Total 2 5/3
[1,3,4] 4/3

[2,3] 1

[2,4] 1
[2,3,4] 1

Figure 6.27: The heuristic formula is not always exact for a non-isolated PAS.
6.6. HISTORICAL NOTES AND FURTHER READING 215

6.6 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
Many gap measures have been proposed. For the relative gap variants γ1 , γ2 ,
and γ3 defined in the text, see Chiu et al. (2010), Patriksson (1994), and Boyce
et al. (2004) for more detail. The average excess cost and maximum excess cost
were proposed in Bar-Gera (2002). In studying the Philadelphia network, Boyce
et al. (2004) found that freeway link flows stabilized once a relative gap of 10−4
was reached.
Extensive reviews of algorithms for the traffic assignment problem are found
in Patriksson (1994) and Florian and Hearn (1995). For the specific methods
discussed in this text, the method of successive averages and Frank-Wolfe are
both instances of the more general “convex combinations” method, and can
in fact be applied to any convex optimization problem (Bertsekas, 2016). The
Frank-Wolfe method itself was proposed in Frank and Wolfe (1956), and the
conjugate version in Mitradjieva and Lindberg (2013). Notable link-based al-
gorithm not presented in this chapter is the simplicial decomposition methods
of Smith (1983) and Lawphongpanich and Hearn (1984); but see Chapter 8 for
discussion of this method in the setting of equilibrium with link interactions.
The gradient projection and manifold suboptimization algorithms were pre-
sented in Jayakrishnan et al. (1994) and Florian et al. (2009), respectively.
Another notable path-based algorithm not described here is the disaggregate
simplicial decomposition method of Larsson and Patriksson (1992).
Nie (2010a) described a general framework for bush-based algorithms, unit-
ing earlier work on origin-based assignment (Bar-Gera, 2002), Algorithm B (Dial,
2006), and local user cost equilibrium (Gentile, 2014). See Xie et al. (2013) for
a discussion about the close relationships between origin-based assignment and
local user cost equilibrium. Interestingly, many of the concepts in bush-based
algorithms were anticipated in the study of routing in telecommunications net-
works; see Gallager (1977) and Bertsekas et al. (1984) for examples of such
work.
The primal method for maximizing path flow entropy is described at greater
length in Bar-Gera (2006). The dual method is the conjugate gradient method
of Larsson et al. (2001); another dual method not described here is iterative
balancing; see Bell and Iida (1997). Traffic assignment by paired alternative
segments was presented in Bar-Gera (2010).

6.7 Exercises
1. [32] One critique of the BPR link performance function is that it allows
link flows to exceed capacity. An alternative link performance “function”
is tij = t0ij /(uij − xij ) if xij < uij , and ∞ otherwise, where t0ij and uij are
the free-flow time and capacity of link (i, j). First show that tij → ∞ as
xij → uij . How would using this kind of link performance function affect
216 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

3 30 6 40
G H I
2 2 2
1 10 4 20
D E F

1 3 3
A 3 B 3 C

Figure 6.28: Network for Exercise 6.

the solution algorithms discussed in this chapter?

2. [33] Show that the relative gap γ1 and average excess cost are always
nonnegative, and equal to zero if and only if the link or path flows satisfy
the principle of user equilibrium.

3. [61] Some relative gap definitions require a lower bound f on the value
of the Beckmann function at optimality. Let x denote the current solu-
tion, f (x) the value of the Beckmann function at the current solution,
and T ST T (x) and SP T T (x) the total system travel time and short-
est path travel time at the current solution, respectively. Show that
f (x) + SP T T (x) − T ST T (x) is a lower bound on the Beckmann func-
tion at user equilibrium.

4. [10] What is the value of the lower bound f = f (x)+SP T T (x)−T ST T (x)
if x satisfies the principle of user equilibrium?

5. [23] Let (h, x) and (g, y) be two feasible solutions to the Beckmann formu-
lation (6.3)–(6.6), and let λ ∈ [0, 1]. Show that (λh+(1−λ)g, λx+(1−λy)
is also feasible, directly from the constraints (without appealing to con-
vexity.)

6. [35] In the network in Figure 6.28, all trips originate at node A. The links
are labeled with the current travel times, and the nodes are labeled with
the number of trips whose destination is that node.

(a) Find the shortest paths from node A to all other nodes, and report
the cost and backnode labels upon termination.
(b) What would be the target link flow solution x̂ in the method of
successive averages or the Frank-Wolfe algorithm?

7. [73] All-or-nothing assignments x̂ play a major role in link-based algo-

rithms. A naı̈ve way to calculate these is to start with zero flows on each
link; then find the shortest path from each origin r to each destination s;
6.7. EXERCISES 217

30 10
1 2 3
10 20 10
10 10
4 5 6

Figure 6.29: Network for Exercise 8, boldface links indicate previous x̂ target.

then add drs to each link in this path. This method may require adding up
to |Z|2 terms for each link, in case every shortest path uses the same link.
Formulate a more efficient algorithm which requires solving one shortest
path problem per origin, and which requires adding no more than |Z|
terms for each link. (Hint: Do not wait until the end to calculate x̂ and
find a way to build x as you go.)
8. [42] Consider the network in Figure 6.29, with a single origin and two
destinations. Each link has the link performance function 10 + x2 , and the
boldface links indicate the links used in the previous x̂ target. Report the
new target link flows x̂, the step size λ, and the new resulting link flows,
according to (a) Frank-Wolfe and (b) conjugate Frank-Wolfe.
9. [47] Consider the network in Figure 6.30, where 8 vehicles travel from
node 1 to node 4. Each link is labeled with its delay function. For each of
the algorithms listed below, report the resulting link flows, average excess
cost, and value of the Beckmann function.
(a) Perform three iterations of the method of successive averages.
(b) Perform three iterations of the Frank-Wolfe algorithm.
(c) Perform three iterations of conjugate Frank-Wolfe.
(d) Perform three iterations of gradient projection.
(e) Perform three iterations of manifold suboptimization.
(f) Perform three iterations of Algorithm B (for each iteration, do one
flow update and one bush update)
(g) Perform three iterations of origin-based assignment (for each itera-
tion, do one flow update and one bush update)
(h) Perform three iterations of linear user cost equilibrium (for each it-
eration, do one flow update and one bush update)
(i) Compare and discuss the performance of these algorithms.

10. [48] Consider the network in Figure 6.31. The cost function on the light
links is 3 + (xa /200)2 , and the delay function on the thick links is 5 +
(xa /100)2 . 1000 vehicles are traveling from node 1 to 9, and 1000 vehicles
from node 4 to node 9. For each of the algorithms listed below, report
the resulting link flows, average excess cost, and value of the Beckmann
function.
218 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

3
10x
50+x

10+x
1 4
60+x
15 x
2

Figure 6.30: Network for Exercise 9.

7 8 9

4 5 6

1 2 3

Figure 6.31: Network for Exercise 10

(a) Perform three iterations of the method of successive averages.

(b) Perform three iterations of the Frank-Wolfe algorithm.
(c) Perform three iterations of conjugate Frank-Wolfe.
(d) Perform three iterations of projected gradient.
(e) Perform three iterations of gradient projection.
(f) Perform three iterations of Algorithm B (for each iteration, do one
flow update and one bush update)
(g) Perform three iterations of origin-based assignment (for each itera-
tion, do one flow update and one bush update)
(h) Perform three iterations of linear user cost equilibrium (for each it-
eration, do one flow update and one bush update)
(i) Compare and discuss the performance of these algorithms.

11. [43] The method of successive averages, as presented in the text, uses
the step size λi = 1/(i + 1) at iteration i. Other choices of step size can
be used, and P Exercise 12 shows
P 2 that the algorithm converges whenever
λi ∈ [0, 1], λi = ∞ and λi < ∞. Which of the following step size
choices guarantee convergence?

(a) λi = 1/(i + 2)
6.7. EXERCISES 219

(b) λi = 4/(i + 2)
(c) λi = 1/i2
(d) λi = 1/(log i)
√
(e) λi = 1/ i
(f) λi = 1/i2/3

12. [65] (Proof of convergence for the method of successive averages.) Con-
sider the method of successive averages applied to the vector of link flows.
This produces a sequence of link flow vectors x1 , x2 , x3 , . . . where xi is the
vector of link flows at iteration i. We can also write down the sequence
f 1 , f 2 , f 3 , . . . of the values taken by the Beckmann function for x1 , x2 ,
etc. To show that this algorithm converges to the optimal solution, we
have to show that either xi → x̂ or f i → fˆ as i → ∞, where x̂ is the user
equilibrium solution and fˆ the associated value of the Beckmann func-
tion. This exercise walks through one proof of this fact, for P any version
of
P 2 the method of successive averages for which λi ∈ [0, 1], λi = ∞ and
λi < ∞.

(a) Assuming that the link performance functions are differentiable, show
that for any feasible x and y there exists θ ∈ [0, 1] such that
X
f (y) = f (x) + tij (xij )(yij − xij )+
(i,j)∈A
1 X 0
tij ((1 − θ)xij + θyij )(yij − xij )2 . (6.101)
2
(i,j)∈A

(b) Setting x = xi and y = xi+1 , recast equation (6.101) into an ex-

pression for the difference in the values of the Beckmann function
between two consecutive iterations of MSA, in terms of λi and x∗i .
(c) Sum the resulting equation over an infinite number of iterations to
obtain a formula for the limiting value fˆ of the sequence f 1 , f 2 , . . ..
λi < ∞, and that fˆ has a finite
P P 2
(d) Use the facts that λi = ∞,
value to show that the limiting values of SP T T (xi ) and T ST T (xi )
must be equal, implying that the limit point is a user equilibrium.

13. [34] The derivation leading to (6.16) assumed that the solution to the
restricted VI was not at the endpoints λ = 0 or λ = 1. Show that if you
are solving (6.16) using either the bisection method from Section 3.3.2, or
Newton’s method (with a “projection” step ensuring λ ∈ [0, 1]), you will
obtain the correct solution to the restricted VI even if it is at an endpoint.

14. [55] (Linking Frank-Wolfe to optimization.) At some point in the Frank-

Wolfe algorithm, assume that the current link flows are x and the target
220 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

link flows x∗ have just be found, and we need to find new flows x0 (λ) =
λx∗ + (1 − λ)x for some λ ∈ [0, 1]. Let z(λ) be the value of the Beckmann
function at x0 (λ).

(a) Using the multi-variable chain rule, we can show that z is differen-
tiable and z 0 (λ) is the dot product of the gradient of the Beckmann
function evaluated at x0 (λ) and the direction x∗ − x. Calculate the
gradient of the Beckmann function and use this to write out a formula
for z 0 (λ).
(b) Is z a convex function of λ?
(c) Show that z 0 (0) = 0 only if x is an equilibrium, and that otherwise
z 0 (0) < 0.
(d) Assume that the solution of the restricted variational inequality in
the Frank-Wolfe algorithm is for an “interior” point λ∗ ∈ (0, 1). Show
that z 0 (λ∗ ) = 0.
(e) Combine the previous answers to show that the Beckmann function
never increases after an iteration of the Frank-Wolfe algorithm (and
always decreases strictly if not at an equilibrium).

15. [74] (Proof of convergence for Frank-Wolfe.) Exercise 14 shows that the
sequence of Beckmann function values f 1 , f 2 , . . . from subsequent itera-
tions of Frank-Wolfe is nonincreasing. Starting from this point, show that
this sequence has a limit, and that the resulting limit corresponds to the
global minimum of the Beckmann function (demonstrating convergence to
equilibrium.) Your solution may require knowledge of real analysis.

16. [11] When is the Beckmann function quadratic?

17. [33] Identify conjugate directions for the following quadratic programs:

(a) f (x1 , x2 ) = x21 + x22

(b) f (x1 , x2 ) = x21 + x22 + 32 x1 x2
(c) f (x1 , x2 ) = 2x21 + 3x22 + 91 x1 x2

18. [38] The biconjugate Frank-Wolfe method chooses a target vector x∗ so

that the search direction x∗ is conjugate to the last two search directions,
rather than just the last one. Let xAON reflect the all-or-nothing solution
at the current point, x∗−1 the target vector used at the last iteration, and
x∗−2 the target vector used two iterations ago. Also let λ−1 be the step
size used for the last iteration, and define
∗ ∗ AON
− xij )t0ij
P
(i,j)∈A (λ−1 (x−1 )ij + (1 − λ−1 )(x−2 )ij − xij )(xij
µ = −P ∗ ∗ ∗ ∗ 0
(i,j)∈A (λ−1 (x−1 )ij + (1 − λ−1 )(x−2 )ij − xij )((x−2 )ij − (x−1 ))tij
(6.102)
6.7. EXERCISES 221

5 6

13 9
1 2 3
7 10

Figure 6.32: Bush for Exercise 20 with current flows.

and
∗
− xij )(xAON − xij )t0ij
P
µλ−1 (i,j)∈A ((x−1 )ij ij
ν= − ∗ . (6.103)
− xij )2 t0ij
P
1 − λ−1 (i,j)∈A ((x−1 )ij

Show that the formula

1
x∗ = xAON + νx∗−1 + µx∗−2


(6.104)
1+µ+ν
gives both a feasible target point x∗ , and one conjugate to the two previous
search directions based on the Hessian of the Beckmann function at the
current solution.
19. [41] Section 6.4.2 includes an example for Algorithm B, and Figure 6.12
shows the bush link flows at the end of a flow shifting operation. Perform
a second iteration of flow shifting on the same bush, recalculating L and U
labels, and scanning all nodes in reverse topological order. Report the new
link flows and travel times. Also report the new bush after eliminating
unused links and adding shortcuts. Does your answer depend on whether
you use the L or U labels to define “shortcuts”?
20. [43] Figure 6.32 shows a bush with the current link flows labeled. Each
link has delay function 5 + 2x2 . Calculate the L, U , M , D, and α labels
for all links and nodes in the bush, and identify the divergence node for
each bush node.
(a) Perform one step of flow shifting using Algorithm B.
(b) Perform one step of flow shifting using OBA (starting from the original
flows).
(c) Perform one step of flow shifting using LUCE (starting from the orig-
inal flows).
(d) Report the maximum excess cost before and after each of these flow
shifts. Which algorithm reduced this gap measure by the most?

21. [73]. This exercises walks through a proof of the formula (6.70) for choos-
ing the threshold  for finding proportional path flows. A set of paths
222 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

Π̂ = ∪Π̂rs is 2-consistent if there are no paths π1 , π2 , π3 , π4 satisfying the

following conditions: (1) π1 and π3 are in Π̂; (2) at least one of π2 and π4
is not in Π̂; (3) π1 and π2 connect the same OD pair; (4) π3 and π4 connect
the same OD pair; and (5) π1 and π3 use the same links as π2 and π4 ,
exactly the same number of times. For instance, in Figure 6.16, the path
set Π̂ = {[1, 2], [1, 3, 5, 6, 4, 2], [3, 5, 6, 4], [3, 5, 7, 8, 6, 4]} is not 2-consistent,
because we can choose π1 = [1, 3, 5, 6, 4, 2], π2 = [1, 3, 5, 7, 8, 6, 4, 2], π3 =
[3, 5, 6, 4], and π4 = [3, 5, 7, 8, 6, 4]. Conditions (1)–(4) are clearly satisfied.
For condition (5), look at any link in the network, and count the number
of times that link is used in π1 and π3 , and the number of times it is used
in π2 and π4 ; every link is either unused in both pairs, used in exactly one
path in both pairs, or used in both paths in each pair. Now assume that
ga ≤  ≤ gr /2, as in (6.70), and choose Π̂ to be all paths whose travel time
is within  of the shortest for its OD pair. Show that there are no four
paths π1 , . . ., π4 satisfying all of the conditions in the previous paragraph.
(Hint: argue by contradiction, and apply each of the conditions, using
the assumptions about the acceptance and rejection gaps to bound each
path’s travel time relative to the shortest path travel time.)

22. [38]. The proof of Theorem 5.4 started from the entropy-maximizing La-
grangian (5.31), which Lagrangianized both the link flow constraints (with
multipliers βij ) and the OD matrix constraints (with multipliers γrs ). Al-
ternatively, we can Lagrangianize only the link flow constraints, and re-
place (hπ /drs ) with hπ (why?), giving the equation
!
X X X
π
L̂(x, β) = hπ log hπ + βij x̂ij − δij hπ . (6.105)
π∈Π (i,j)∈A π∈Π

Show that the gradient of this alternative Lagrangian with respect to β

has components given by (6.82). That is, show that

∂ L̂
= xij − x̂ij
∂βij

where xij is computed from the current path flows hπ .

23. [59]. The dual algorithm step (6.82) can be compactly written as β ←
β + α∆β, where ∆βij = xij − x̂ij . Let f (α) denote the value of the
alternative Lagrangian (6.105) after a step of size α is taken, and the new
β and h values are calculated. Newton’s method can be used to find an
α value which approximately minimizes f (α), maximizing entropy in the
direction ∆β. The Newton step is α = −f 0 (0)/f 00 (0).

(a) Show that f 0 (0) = ||∇β L̂||2 . (See Exercise 22.)

(b) Show that f 00 (0) = ∇Tβ L̂Hβ L̂∇β L̂, where Hβ is the Hessian of the
alternate Lagrangian with respect to β.
6.7. EXERCISES 223

(2,0)
1 4 (tij,xij)
(2,100) i j
(1,100)
(1,0)
2 3
(1,0)

Figure 6.33: Network for Exercise 24.

(c) Show that

X
(xij − x̂ij )2
0
f (0) (i,j)∈A
= .
f 00 (0)
X X X
(xij − x̂ij )(xkl − x̂k` ) hπ δij
π π
δk`
(i,j)∈A (k,`)∈A π∈Π

24. [30]. In the discussion surrounding Figure 6.19, we argued that satisfying
the equilibrium conditions around a “spanning” set of PASs (one for each
block) was sufficient for establishing equilibrium on the entire network.
Consider the network in Figure 6.33, where the demand from origin 1
to destination 4 is 100 vehicles, and there are two PASs: one between
segments [1, 4] and [1, 2, 3, 4], and another between segments [2, 4] and
[2, 3, 4]. These are spanning, in the sense that by shifting flows between
these two PASs we can obtain any feasible path flow solution from any
other. They also satisfy the equilibrium conditions: for the first PAS,
because there is no flow on either segment16 ; for the second, because the
travel times are equal on the two segments. Yet the network is not at
equilibrium, since [1, 4] is the only shortest path and it is unused. Explain
this apparent inconsistency.

25. [62]. Develop one or more algorithms to find “short” segments when
generating a new PAS. These methods should require a number of steps
that grows at most linearly with network size.

26. [22]. Show that the cycle-removing procedure described in the TAPAS
algorithm maintains feasibility of the solution (flow conservation at each
node, and non-negativity of link flows), and that the Beckmann function
decreases strictly (assuming link performance functions are positive).

27. [68]. This exercise develops a technique for approximately solving equa-
tions (6.90) and (6.94), better than the heuristic given in the text.

(a) Define ∆hr (ρ) to be the amount of flow that needs to be shifted from
origin r’s flows on σ1 to σ2 , to adjust the proportion g r (σ1 )ζ /(g r (σ1 )ζ +
g r (σ2 )ζ to be exactly ρ. A negative value of this function indicates
16 There is flow on link (1,2), but not on the entire segment [1, 2, 3, 4].
224 CHAPTER 6. ALGORITHMS FOR TRAFFIC ASSIGNMENT

shifting flow in the reverse direction, from σ2 to σ1 . Show that

this function is defined over [0, 1], and its range is [∆, ∆], where
∆ = − min(i,j)∈σ2 xrij and ∆ = min(i,j)∈σ1 xrij . Furthermore show
that ∆hr (0) = ∆ and ∆hr (1) = ∆.
(b) Show that equation (6.94) is satisfied if the flow shift ∆hr (ρ) is ap-
plied to all relevant origins r ∈ Zζ .
(c) Therefore, it is enough to find a value of ρ for which
X
U (ρ) ≡ ∆hr (ρ) = 0 ,
r∈Zζ

in order to satisfy (6.90). The function U is continuous and defined

on the interval [0, 1]. Show that U (0) ≤ 0 and U (1) ≥ 0, ensuring
that a zero exists in this interval.
(d) Develop a quadratic approximation for ∆hr (ρ) based on three known
points: ∆hr (0) = ∆, ∆hr (1) = ∆, and ∆hr (ρr ) = 0, where ρr is the
current proportion g r (σ1 )ζ /(g r (σ1 )ζ + g r (σ2 )ζ ).
(e) By summing these, develop a quadratic approximation for U (ρ), and
give an explicit formula for its root.
Chapter 7

Sensitivity Analysis and

Applications

This chapter shows how a sensitivity analysis can be conducted for the traffic
assignment problem (TAP), identifying how the equilibrium assignment will
change if the problem parameters (such as the OD matrix or link performance
functions) are changed. This type of analysis is useful in many ways: it can
be used to determine the extent to which errors or uncertainty in the input
data create errors in the output data. It can be used as a component in so-
called “bilevel” optimization problems, where we seek to optimize some objective
function while enforcing that the traffic flows remain at equilibrium. This occurs
most often in the network design problem, where one must determine how to
improve network links to reduce total costs, and in the OD matrix estimation
problem, where one attempts to infer the OD matrix from link flows, or improve
upon an existing estimate of the OD matrix.
After exploring the sensitivity analysis problem using the familiar Braess
network, the first objective in the chapter is calculating derivatives of the equi-
librium link flows with respect to elements in the OD matrix. It turns out that
this essentially amounts to solving another, easier, traffic assignment problem
with different link performance functions and constraints. The remainder of
the chapter shows how these derivatives can be used in the network design and
OD matrix estimation problems, which are classic transportation examples of
bilevel programs.

7.1 Sensitivity Analysis Preliminaries

Figure 7.1 shows the Braess network. When this network was first introduced,
the demand between node 1 and node 4 was d14 = 6, and the equilibrium
solution was found to be x13 = x23 = x24 = 2 and x12 = x34 = 4, with a
travel time of 92 minutes on all three paths. What if, instead, the demand d14
took another value? Figure 7.2 presents four plots showing how the equilibrium

225
226 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

3
10x
50+x

d14 10+x d14

1 4
50+x
10x
2

Figure 7.1: Braess network with varying demand from 1 to 4.

solution varies according to the demand level. Panel (a) shows the flows x12
and x34 , panel (b) shows the flows x13 and x24 , panel (c) shows the flow x23 ,
and panel (d) shows the shortest path travel time between nodes 1 and 4, at
the corresponding equilibrium solution. You can check that when d14 = 6, the
original equilibrium solution is shown in this figure.
Instead of the OD matrix, we also could have changed the link performance
functions in the network. Now assume that d14 is fixed at its original value of 6,
but that the link performance function on link (2,3) can vary. Let t23 (x23 ) = y +
x23 , where y is the free-flow time, resulting in the network shown in Figure 7.3.
In the base solution y = 10, but conceivably the “free-flow time” could be
changed. If the speed limit were increased, y would be lower; if traffic calming
were implemented, y would be higher. In an extreme case, if the link were closed
entirely you could imagine y takes an extremely large value, large enough that
no traveler would use the path. One can also effectively decrease y by providing
incentives for traveling on this link (a direct monetary payment, a discount at
an affiliated retailer, etc.), and conceivably this incentive could be so large that
y is negative. The resulting sensitivity analysis is provided in Figure 7.4
Examining the plots in Figure 7.2 and 7.4, we see that the relationships
between the equilibrium solution (link flows and travel times) and the demand
or free-flow time are all piecewise linear. Each “piece” of these piecewise linear
functions corresponds to a particular subset of the paths being used — for in-
stance, in Figure 7.2, when the demand is lowest, only the middle path is used.
When the demand is highest, only the two outer paths are used. When the
demand is at a moderate level, all three paths are used. Within each of these
regions, the relationship between the demand and the equilibrium solution is
linear. These pieces meet at so-called degenerate solutions, where the equilib-
rium solution does not use all of the minimum travel-time paths. (For instance,
when d14 = 40/11 the equilibrium solution requires all drivers to be assigned to
the middle path, even though all three have equal travel times.)
In general networks involving nonlinear link performance functions, these
relationships cannot be expected to stay linear. However, they are still defined
by piecewise functions, with each piece corresponding to a certain set of paths
7.1. SENSITIVITY ANALYSIS PRELIMINARIES 227

(a) Flow on (1,2) and (3,4) (b) Flow on (1,3) and (2,4)

(c) Flow on (2,3) (d) Equilibrium travel time

Figure 7.2: Sensitivity analysis of the Braess network to d14 .

3
10x
50+x

6 +x 6
1 4
50+x
10x
2

Figure 7.3: Braess network with varying free-flow time on (2,3).

228 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

(a) Flow on (1,2) and (3,4) (b) Flow on (1,3) and (2,4)

(c) Flow on (2,3) (d) Equilibrium travel time

Figure 7.4: Sensitivity analysis of the Braess network to y.

7.2. CALCULATING SENSITIVITIES 229

being used, and with the pieces meeting at degenerate solutions. The goal of the
sensitivity analyses in this chapter is to identify derivatives of the equilibrium
solution (link flows and travel times) at a given point. For these derivatives
to be well-defined, we therefore assume that the point at which our sensitivity
analysis occurs is not degenerate. That is, all minimum-travel time paths have
positive flow. This assumption is not too restrictive, because there are only a
finite number of degenerate points; for instance, if we pick the demand value at
random, the probability of ending up at a degenerate point is zero.
This sensitivity analysis is still local, because the information provided by
a derivative grows smaller as we move farther away from the point where the
derivative is taken. For a piecewise function, the derivative provides no in-
formation whatsoever for pieces other than the one where the derivative was
taken.
In this chapter, we show how this kind of sensitivity analysis can be used
in two different ways. In the network design problem, this type of sensitivity
analysis can be used to determine where network investments are most valu-
able. In Figure 7.4, the fact that the equilibrium travel time increases when
y decreases (around the base solution y = 10) highlights the Braess paradox:
investing money to improve this link will actually increase travel times through-
out the network. If we were to conduct a similar analysis for other links in
the network, we would see that the equilibrium travel time would decrease with
improvements to the link. In the OD matrix estimation problem, we can use
this sensitivity analysis to help calibrate an OD matrix to given conditions.

7.2 Calculating Sensitivities

This section derives sensitivity formulas showing the derivative of the equilib-
rium link flows and travel times with respect to two parameters: (1) a change in
an entry of the OD matrix drs , and (2) a change to a parameter in the link per-
formance functions (such as the free-flow time or capacity in a BPR function).
In this section, assume that we are given some initial OD matrix or link perfor-
mance functions, and the corresponding equilibrium solution. For our purposes,
it will be most convenient if this equilibrium solution is expressed in bush-based
form, that is, with vectors x̂r showing the flow on each link corresponding to
each origin r. In this case, the non-degenerate condition requires that unused
links are not part of the equilibrium bushes, that is, if x̂rij = 0 for any link
(i, j) and any origin r, then link (i, j) does not correspond to any shortest path
starting from node r — in terms of the distance labels Lri , we have x̂rij > 0 if
and only if Lrj = Lri + tij . Let B r = {(i, j) ∈ A : x̂rij > 0} denote the equilibrium
bush for origin r.
The non-degeneracy assumption is important, because one can show that
if the change to the OD matrix or link performance functions is small and the
original equilibrium solution is non-degenerate, all of the equilibrium bushes re-
main unchanged. Equivalently, even after drivers shift flows to find the new
equilibrium, the set of used paths will remain the same as it was before. Fur-
230 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

thermore, one can show that the equilibrium solution is differentiable, and the
derivatives of the equilibrium link flows or travel times with respect to values
in the OD matrix or link performance function parameters can be interpreted
as the sensitivities of the equilibrium solution.
There are several ways to calculate the values of these derivatives: his-
torically, the first researchers used matrix-based formulas, and subsequent re-
searchers generalized these formulas using results from the theory of variational
inequalities. We adopt a different approach, using the bush-based solution rep-
resentation, because it leads to an easy solution method and is fairly straightfor-
ward. This approach is based on the fact that the equilibrium solution (travel
times tij and bush flows xrij ) must satisfy the following equations for each origin
r:

Lrj − Lri − tij (x̂ij ) = 0 ∀(i, j) ∈ B r (7.1)

Lrr =0 (7.2)
X X
x̂rhi − x̂rij =d ri
∀i ∈ N \{r} (7.3)
(h,i)∈Γ−1 (i) (i,j)∈Γ(i)
X X X
x̂rhr − x̂rrj = − drs (7.4)
(h,r)∈Γ−1 (r) (r,j)∈Γ(r) s∈Z

x̂rij = 0 / Br
∀(i, j) ∈ (7.5)

Equations (7.1)–(7.2) reflect the equilibrium condition, and equations (7.3)–

(7.4) represent flow conservation. The number of equations for each origin is no
more than the sum of the number of links and nodes in the network.
Furthermore, these conditions must remain true even as the problem data
(OD matrix and link performance functions) are perturbed. Since derivatives
of the equilibrium solution exist under the non-degeneracy assumption, we can
differentiate equations (7.1)–(7.4) to identify the relationships which must hold
r
true among these derivatives. For brevity, in this chapter we use ξij to denote
r
the derivative of xij , and Λi to denote the derivative of Li . These derivatives are
taken with respect to either an OD matrix entry or a link performance function
parameters, as described separately below.

7.2.1 Changes to the OD matrix

Assume first that we change a single entry in the OD matrix corresponding to

r
origin r̂ and destination ŝ, so ξij = dxrij /ddr̂ŝ and Λi = dLi /ddr̂ŝ . Then differ-
entiating each of equations (7.1)–(7.4) with respect to dr̂ŝ gives the following
7.2. CALCULATING SENSITIVITIES 231

equations for each r ∈ Z:

X 0
Λrj − Λri − t0ij r
ξij =0 ∀(i, j) ∈ B r (7.6)
r 0 ∈Z
Λrr = 0 (7.7)
(
X
r
X
r 1 if r = r̂ and i = ŝ
ξhi − ξij = ∀i ∈ N \r (7.8)
(h,i)∈Γ−1 (i) (i,j)∈Γ(i)
0 otherwise
(
X
r
X
r −1 if r = r̂
ξhr − ξrj = (7.9)
(h,r)∈Γ−1 (r) (r,j)∈Γ(r)
0 otherwise
r
ξij =0 / B r (7.10)
∀(i, j) ∈

where t0ij is the derivative of the link performance function, evaluated at the cur-
rent equilibrium solution x̂ (and thus treated as a constant in these equations).
Equations (7.6) enforce the fact that the equilibrium bushes must remain the
same. That is, the shortest path labels Li and travel times tij must change in
such a way that every link on the bush is part of a minimum travel time path
to its head node. Equations (7.8) and (7.9) enforce flow conservation. For all
bushes except for r̂, the total flow from the origin to each destination is the
same, so flow is allowed to redistribute among the bush links, but the flows
starting or ending at a node cannot change. For the bush corresponding to r̂, a
unit increase in demand from r̂ to ŝ must be reflected by an additional vehicle
leaving r̂ and an additional vehicle arriving at ŝ.
All together, the system of equations (7.6)–(7.9) involves variables Λri for
r
each origin r and node i, and ξij for each origin r and link (i, j). Furthermore,
for each origin, it contains an equation for each link and each node. Therefore,
this linear system of equations can be solved to obtain the sensitivity values.1
However, there is an easier way to solve for Λri and ξijr
. Using the techniques
in Section 3.3, you can show that the equations (7.6)–(7.9) are exactly the
optimality conditions to the following minimization problem:
!2
1 X 0 X
r
min tij ξij
ξr ,Λr 2
(i,j)∈A r∈Z
 
XX X X
+ Λri  r
ξhi − r
ξij − ∆ri  (7.11)
r∈Z i∈N (h,i)∈Γ−1 (i) (i,j)∈Γ(i)

s.t. Λrr = 0 ∀r ∈ Z (7.12)

r r
ξij =0 ∀(i, j) ∈
/B (7.13)
1 A careful reader will note that one of the flow conservation equations for each origin is

redundant, but this is of no consequence to what follows.

232 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

where ∆ri represents the right-hand side of equation (7.8) or (7.9), that is,
∆r̂ŝ = 1, ∆r̂r̂ = −1, and ∆ri = 0 otherwise.
This optimization problem can be put in a more convenient form by in-
terpreting Λri as the Lagrange multiplier for the flow conservation equation
corresponding to node i in bush r, and the second term in (7.11) P as the La-
r
grangianization of this equation. Furthermore, defining ξij = r∈Z ij the
ξ ,
optimization problem can be recast in the following equivalent form:
Z ξij
min t0ij ξ dξ (7.14)
ξr 0
X X
r r
s.t. ξhi − ξij = ∆ri ∀i ∈ N, r ∈ Z (7.15)
(h,i)∈Γ−1 (i) (i,j)∈Γ(i)
r
ξij =0 / Br
∀(i, j) ∈ (7.16)

This is essentially a traffic assignment problem (TAP) in bush-based form (see

Chapter 6), with the following changes:
ˆ The original link performance functions have been replaced by linear link
performance functions with slope equal to the derivative of the original
link performance function at the original equilibrium solution. (Remember
that t0ij is a constant, the value of the link performance function derivative
at the equilibrium solution.)
ˆ The equilibrium bushes for each origin are fixed at the bushes for the
original equilibrium solution.
ˆ The only entry in the OD matrix is one unit of demand from r̂ to ŝ.
ˆ There are no non-negativity conditions. This is because the solution vari-
ables ξij represent changes in the original link flows, and it is possible for
these changes to be negative as well as positive (cf. Figure 7.2).
If you have access to an implementation of a bush-based algorithm for solving
TAP, it is easy to modify the program to take account of these distinctions, and
r
to find the link flow sensitivities ξij . From here, the values of Λri can be found
dt
by solving a shortest path problem with link travel times dxijij ξij .
As a demonstration, we use the Braess network of Figure 7.1, working around
the base demand d14 = 6 and base equilibrium solution x̂12 = x̂34 = 4, x̂13 =
x̂23 = x̂24 = 2. At this level of demand, all paths are used and the equilibrium
bush contains all of the links in the original network. Furthermore, at this
equilibrium solution the derivatives of the link performance functions are t012 =
t034 = 10 and t013 = t023 = t024 = 1. The linear link performance functions based
on these derivatives are shown in Figure 7.5. The OD matrix is replaced by a
single unit of flow traveling from 1 (our r̂) to 4 (our ŝ).
Solving the problem without non-negativity constraints produces the ξij val-
ues shown in Figure 7.6. Substituting these into the link performance functions
in Figure 7.5 shows that the equilibrium is satisfied: all paths have equal travel
7.2. CALCULATING SENSITIVITIES 233

3
x 10x

1 x 1
1 4

x
10x
2

Figure 7.5: Modified traffic assignment problem for sensitivity to d14

3
11/13 2/13

1 -9/13 1
1 4
2/13 11/13

Figure 7.6: Solution to traffic assignment problem for sensitivity to d14

times of 31/13. This is exactly the slope of the piece of the equilibrium travel
time (Figure 7.2d) around d14 = 6, that is, the equilibrium travel time in the
sensitivity problem gives the derivative of the equilibrium travel time in the
original network.

7.2.2 Changes to a link performance function

Now assume that we change a parameter y (which may represent the free-flow
time, capacity, or any other parameter) in the link performance function corre-
r
sponding to link (i, j), so ξij = dxrij /dy and Λi = dLi /dy. This means that the
link performance function tij now depends on both its flow xij , and the param-
eter y, so we write tij (x, y). We will write t0ij,x to mean the partial derivative
with respect to link flow2 , and t0ij,y to mean the partial derivative with respect to
the improvement parameter y Then differentiating each of equations (7.1)–(7.4)

2 This is what we wrote as t0 in the previous section, when the link performance function
ij
only depended on xij .
234 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

with respect to y gives the following equations for each r ∈ Z:

X 0
Λrj − Λri − t0ij,x r
ξij − t0ij,y = 0 ∀(i, j) ∈ B r (7.17)
r 0 ∈Z
Λrr = 0 (7.18)
X X
r r
ξhi − ξij =0 ∀i ∈ N \{r} (7.19)
(h,i)∈Γ−1 (i) (i,j)∈Γ(i)
X X
r r
ξhr − ξrj =0 (7.20)
(h,r)∈Γ−1 (r) (r,j)∈Γ(r)
r
ξij =0 / Br
∀(i, j) ∈ (7.21)

where as before, t0ij,x and t0ij,y are evaluated at the current, equilibrium solution
x̂. Equations (7.17) enforce the fact that the equilibrium bushes must remain
the same, taking into account both the change in travel time on (i, j) due to
the change in its link performance function as well as changes in all links’ travel
times from travelers shifting paths. Equations (7.19) and (7.20) enforce flow
conservation. These equations are simpler than for the case of a change to the
OD matrix, because the total number of vehicles on the network remains the
same, and these vehicles can only shift amongst the paths in the bush. There is
no change in the flow originating or terminating at any node, and the ξ variables
must form a circulation.
As with a change in an OD matrix entry, the system of equations (7.17)–
(7.20) is a linear system involving, for each origin, variables for each node and
link. Repeating the same steps as before, this system of equations can be seen
as the optimality conditions for the following optimization problem:
Z ξij
t0ij,x ξ + t0ij,y dξ


min
r
(7.22)
ξ 0
X X
r r
s.t. ξhi − ξij =0 ∀i ∈ N, r ∈ Z (7.23)
(h,i)∈Γ−1 (i) (i,j)∈Γ(i)
r
ξij =0 / Br
∀(i, j) ∈ (7.24)

This is essentially a traffic assignment problem in bush-based form, with the

following changes:

ˆ The original link performance functions have been replaced by affine link
performance functions with slope equal to the derivative of the original link
performance function at the original equilibrium solution, and intercept
equal to the derivative of the link performance function with respect to
the parameter y.

ˆ The equilibrium bushes for each origin are fixed at the bushes for the
original equilibrium solution.

ˆ All entries in the OD matrix are zero.

7.3. NETWORK DESIGN PROBLEM 235

3
x 10x

0 1+x 0
1 4

x
10x
2

Figure 7.7: Modified traffic assignment problem for sensitivity to y23

ˆ There are no non-negativity conditions.

As a demonstration, we use the Braess network of Figure 7.3, working around

the base free-flow time y = 10 and base equilibrium solution x̂12 = x̂34 = 4,
x̂13 = x̂23 = x̂24 = 2. That is, we replace the link performance function
t23 (x23 ) = 10 + x23 with the function t23 (x23 , y) = y + x23 and see what hap-
pens when y varies. At the base value y = 10, all paths are used and the
equilibrium bush contains all of the links in the original network. Furthermore,
at this equilibrium solution the derivatives of the link performance functions
with respect to flows are t012,x = t034,x = 10 and t013,x = t023,x = t024,x = 1. For
link (2,3), we add the constant term t023,y = 1. The link performance functions
based on these derivatives are shown in Figure 7.7. The OD matrix is set equal
to zero, since there is no change in the total demand through the network.
Solving the problem without non-negativity constraints produces the ξij val-
ues shown in Figure 7.8. Substituting these into the link performance functions
in Figure 7.7 shows that the equilibrium is satisfied: all paths have equal travel
times of −9/13. Again, this is the slope of the piece of the equilibrium travel
time (Figure 7.4d) around y = 10, that is, the equilibrium travel time in the
sensitivity problem gives the derivative of the equilibrium travel time in the
original network.

7.3 Network Design Problem

In the network design problem, one must determine how best to spend funds on
improving links in the transportation network. This is a challenging optimiza-
tion problem, and in cases of any practical interest one cannot hope to identify a
globally optimal investment policy. This is mainly because in transportation sys-
tems, the planner cannot compel travelers to choose a particular path. Instead,
after any improvement is made to the links, flows will redistribute according to
the principle of user equilibrium. The goal is to find the best investment policy,
knowing and anticipating how travelers will respond once the network has been
236 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

3
1/13 -1/13

0 -2/13 0
1 4
-1/13 1/13

Figure 7.8: Solution to traffic assignment problem for sensitivity to y23

changed.
Specifically, assume that the link performance functions for each link (i, j)
now depend on the amount of money yij invested in that link (perhaps increasing
its capacity through widening, or decreasing its free-flow time) as well as on the
flow xij . One example of such a link performance function is
 β !
0 xij
tij (xij , yij ) = tij 1 + α (7.25)
uij + Kij yij

where Kij represents the capacity improvement if a single unit of money is

invested on it.
A planning agency may have many different objectives and constraints when
determining how to improve a network. This section develops and explores one
specific variation of the network design problem, but there are many other
variations which have been proposed in the literature. The version presented
here is a fairly standard one, which can be extended in a number of different
ways. In this variation, the objective of the planning agency is to minimize
the total cost, given by the sum of total system travel time T ST T (converted
to monetary units by a conversion factor Θ, which also reflects duration of the
analysis horizon and discounting) and the costs of the network improvements
themselves. The optimization problem is
X X
minf (x, y) = Θ xij tij (xij , yij ) + yij (7.26)
x,y
(i,j)∈A (i,j)∈A
X Z xij
s.t. x ∈ arg min tij (x, yij ) dx (7.27)
x∈X 0
(i,j)∈A

yij ≥ 0 ∀(i, j) ∈ A (7.28)

Most of this problem is familiar: the objective (7.26) is to minimize the sum of
total system travel time (converted to units of money) and construction cost, and
constraint (7.28) requires that money can only be spent on links (not “recovered”
7.3. NETWORK DESIGN PROBLEM 237

from them with a negative yij value). Also note that since y = 0 is a feasible
solution (corresponding to the “do-nothing” alternative), in the optimal solution
to this problem the cost savings (in the form of reduced T ST T ) must at least be
equal to the construction costs, guaranteeing that the optimal investment policy
has greater benefit than cost. The key equation here is (7.27), which requires
that the link flows x satisfy the principle of user equilibrium by minimizing
the Beckmann function. In other words, one of the constraints of the network
design problem is itself an optimization problem. This is why the network design
problem is called a bilevel program. This type of problem is also known as a
mathematical program with equilibrium constraints. This class of problems is
extremely difficult to solve, because the feasible region is typically nonconvex.
To see why, consider two feasible solutions (x1 , y1 ) and (x2 , y2 ) to the net-
work design problem. The link flows x1 are the equilibrium link flows under
investment policy y1 , and link flows x2 are the equilibrium link flows under in-
vestment policy y2 . If the feasible region were a convex set, then any weighted
average of these two solutions would themselves be feasible. Investment pol-
icy 21 y1 + 12 y2 still satisfies all the constraints on y (all link investments are
nonnegative). However, the equilibrium link flows under this policy cannot be
expected to be the average of x1 and x2 , because the influence of yij on tij
can be nonlinear and the sets of paths which are used in x1 and x2 can be
completely different. In other words, the equilibrium link flows after averaging
two investment policies need not be the average of the equilibrium link flows
under those two policies separately.
Unfortunately, solving optimization problems with nonconvex feasible re-
gions is a very difficult task. Therefore, solution methods for the network design
problem are almost entirely heuristic in nature. These heuristics can take many
forms; one popular approach is to adapt a metaheuristic method, such as those
discussed in Section C.6.
Another approach is to develop a more tailored heuristic based on specific
insights about the network design problem. This approach, being more educa-
tional, is adopted here. Specifically, we can use the sensitivity analysis from the
previous sections to identify derivatives of the objective function f with respect
to each link investment yij , and use this to move in a direction which reduces
total cost.
Specifically, notice that constraint (7.27) actually makes x a function of y,
since the solution to the user equilibrium problem is unique in link flows. That
is, the investment policy y determines the equilibrium link flows x exactly. So,
the objective function can be made a function of y alone, written f (x(y), y).
The derivative of this function with respect to an improvement on any link is
then

∂f X ∂f ∂xk`
=Θ +1 (7.29)
∂yij ∂xk` ∂yij
(k,`)∈A
238 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

or, substituting the derivative of (7.26) with respect to each link flow,
 
 
∂f  X ∂xk` ∂tk` ∂tij 
=Θ tk` (xk` , yk` ) + xk` (xk` , yk` ) + xij + 1.
∂yij  ∂yij ∂xk` ∂yij 
(k,`)∈A
(7.30)
In turn, the partial derivatives ∂x k`
∂yij can be identified using the technique of
Section 7.2.2 as the marginal changes in link flows throughout the network
when the link performance function of (i, j) is perturbed.
The vector of all the derivatives (7.30) forms the gradient of f with respect
to y. This gradient is the direction of steepest ascent, that is, the direction
in which f is increasing fastest. Since we are solving a minimization problem,
we should move in the opposite direction. Taking such a step, and ensuring
feasibility, gives the updating equation
+
y ← [y − µ∇y f ] (7.31)

where µ is a step size to be determined, and the [·]+ operation is applied to each
component of the vector. This suggests the following algorithm:

1. Initialize y ← 0.

2. Calculate the link flows x(y) by solving the traffic assignment problem
with link performance functions t(x, y).
∂f
3. For each link (i, j) determine ∂y ij
by solving the sensitivity problem de-
scribed in Section 7.2.2 and using (7.30).

4. Update y using (7.31) for a suitable step size µ.

5. Test for convergence, and return to step 2 if not converged.

Two questions are how µ should be chosen in step 4, and how convergence
should be tested in step 5. The difficulty in step 4 is that the derivatives
provided by a sensitivity analysis are only local, and in particular if µ is large
enough that (7.31) changes the set of used paths, the derivative information
is meaningless. However, if µ is small enough one will see a decrease in the
objective function if at all possible. So, one could start by testing a sequence of
µ values (say, 1, 1/2, 1/4, . . .), evaluating the resulting y values, x values, and f ,
stopping as soon as f decreases from its current value. (Note that this is a fairly
computationally intensive process, since the traffic assignment problem must be
solved for each µ to get the appropriate x value.) Other options include using a
stricter stopping criterion such as the Armijo rule, which would ensure that the
decrease is “sufficiently large”; or using bisection to try to choose the value of
µ which minimizes f in analogy to Frank-Wolfe. All of these methods require
solving multiple traffic assignment problems at each iteration.
Regarding convergence in step 5, one can either compare the progress made
in decreasing f over the last few iterations, or the changes in the investments
7.3. NETWORK DESIGN PROBLEM 239

50+x13exp(-y13) 3
10x34exp(-y34)

10+x23exp(-y23)
1 4

10x12exp(-y12)
2 50+x24exp(-y24)

Figure 7.9: Network design problem example.

y. This choice of stopping criterion is different in nature than those used in

Chapter 6 for solution methods to TAP. In that case, we can prove theoretical
convergence to the equilibrium solution, and we can design our stopping criteria
directly on the equilibrium condition. The method described above, by contrast,
is not proven to converge to the global optimum, and can get stuck in solutions
which are locally optimal but not globally so. (This often happens in nonconvex
optimization problems.) By terminating the algorithm when no more progress
is made, we are checking that we have found a local optimum, but cannot
guarantee that we have found a global one.
As stated above, the network design problem does not have a convex feasible
region, and even if one were to eliminate x by writing x as a function of y, the
resulting function can be shown to be nonconvex and have multiple local optimal
solutions. Therefore, the method described above is not guaranteed to converge
to a global optimum solution. This is perhaps a bit disappointing; but, as
stated above, at present there is no way to ensure global optimality within a
reasonable amount of computation time. Therefore, alternative heuristics can
only be compared by the quality of solutions obtained for a given quantity of
computational effort.
For an example, consider the Braess network shown in Figure 7.9, where
1
the link performance functions are shown and Θ = 20 . Notice now that the
coefficients of xij in the link performance functions now depend on the amount
of money yij invested as well, with decreasing marginal returns as more money
is spent. When y = 0, the solution to the user equilibrium problem x(y) is the
now-familiar solution to the Braess network, which divides flow evenly among
all three paths in the network. The total system travel time is 552 vehicle-
1
minutes, so the value of the objective (7.26) function is 20 (552) + 0 = 27.6.
This completes the first two steps of the algorithm.
For the third step, we must solve five sensitivity problems, one for each link,
∂f
to determine ∂y ij
. These five problems are shown in Figure 7.10, where we have
substituted the initial values y = 0 and the equilibrium link flows x. In all of
240 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

Link (1,2) Link (1,3)

 
  
–

    
  
 

– 
  


Link (2,3) Link (2,4)

 
  


 –    
   

  –

 


 –

  
Link (3,4)  


 

Figure 7.10: Five sensitivity problems for network design.

these, notice that the demand is zero, and the link performance functions are
affine. Within each problem, the link being improved has a slightly different link
performance function, accounting for the effect of the link improvement. The
other links’ performance functions only reflect their change due to shifting flows.
For instance, in the problem in the upper left, link (1,3) is not improved, so its
∂t13
link performance function is simply ξ13 ∂x 13
= 10ξ12 exp(−y13 ) = 10ξ12 since
∂t12
y13 = 0. Since link (1,2) is being improved, in addition to the term ξ12 ∂x 12
=
∂t12
ξ12 exp(−y12 ) = ξ14 , we add the constant term ∂y12 = −x14 exp(−y12 ) = −4
since x12 = 4 and y12 = 0.
The solutions (in terms of ξij ) to the five sensitivity problems are shown in
Table 7.1, as can be verified by substituting these ξ values into the networks
in Figure 7.10. Substituting these ξ values into equation (7.30), along with the
current values of the travel times and link flows, gives the gradient
   
y12 −0.85
y13   0.49 
   
∇y f =  y23  =  1.42 
   (7.32)
y24   0.49 
y34 −0.85

Each component of the gradient shows how the objective of the network de-
7.3. NETWORK DESIGN PROBLEM 241

Table 7.1: Solutions to Braess sensitivity problems.

Sensitivity problem
Link (1,2) (1,3) (2,3) (2,4) (3,4)
(1,2) 3.36 −0.17 0.15 0.014 −0.28
(1,3) −3.36 0.17 −0.15 −0.014 0.28
(2,3) 3.08 −0.15 0.31 −0.15 3.08
(2,4) 0.28 −0.014 −0.15 0.17 −3.36
(3,4) −0.28 0.014 0.15 −0.17 3.36

sign problem will change if a unit of money is spent improving a particular link.
In the derivative formula (7.30), the term in parentheses represents the marginal
change in total system travel time, and the addition of unity at the end of the
formula represents the increase in total expenditures. If the derivative (7.30)
is negative, then the reduction in total system travel time from a marginal im-
provement in the link will outweigh the investment cost. If it lies between zero
and one for a link, then a marginal investment in the link will reduce total
system travel time, but the cost of the improvement will outweigh the value of
the travel time savings. If it is greater than one, then total system travel time
would actually increase if the link is improved, as in the Braess paradox. So, in
this example, the gradient (7.32) shows that improvements on links (1,2) and
(3,4) will be worthwhile; improvements on links (1,3) and (2,4) would reduce
TSTT but not by enough to outweigh construction cost; and an improvement
on link (2,3) would actually be counterproductive and worsen congestion.

So, proceeding to step 4 of the network design algorithm, we choose a trial

value µ = 1 and apply (7.31) to obtain a candidate solution where y12 = y34 =
0.85 and all other yij values remain at zero. Re-solving equilibrium with the
new link performance functions, the new equilibrium solution is to load all
vehicles on the middle path, that is, x12 = x23 = x34 = 6 and x13 = x24 = 0.
The total system travel time is now 405 vehicle-minutes, so the objective is
1
20 (405) + 2(0.85) = 22.0. This reduces the objective from its current value of
27.6, so we accept the step size µ = 1, and return to step 2.

Notice that solving the network design problem requires solving a very large
number of traffic assignment subproblems: once for each iteration to determine
x; modified sensitivity problems for each link to calculate derivatives; and again
multiple times per iteration to identify µ. Solving practical problems can easily
require solution of thousands or even millions of traffic assignment problems. In
bilevel programs such as network design, having an efficient method for solving
traffic assignment problems is critical. Path-based and bush-based algorithms
can be efficiently “warm-started,” making them good choices for this applica-
tion.
242 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

7.4 OD Matrix Estimation

Practical application of the traffic assignment problem requires two main inputs:
information about the physical network (roadway topology, link performance
functions) as well as information about travel demand patterns (the OD matrix).
The former is much easier to obtain and validate. Standard link performance
functions (such as the BPR function) can be used only knowing the free-flow
time and capacity of links, which can be easily estimated; and in principle, this
information can be directly inferred from field measurements and traffic sensors.
The OD matrix is quite a bit harder to estimate, for several reasons. First, there
is no practical way to directly observe the complete OD matrix; at best we
can work with a sample of travelers who consent to reveal their travel patterns.
Second, at least until recently, it was difficult to obtain this information without
travelers explicitly reporting their origins and destinations — traffic sensors tell
you what is happening on a specific link, but not where the people are coming
from or where they are going. Recently, GPS and Bluetooth technologies have
become more commonplace and can in principle provide origins and destinations
automatically (a rough estimate could even be obtained from cellular phone
traces) — but there are major privacy issues associated with using such data,
as well as a number of data inference issues involved in translating these data
into an OD matrix. Third, the number of entries in an OD matrix is much larger
than the number of links in the network: the number of OD pairs is generally
proportional to the square of the number of nodes, while the number of links is
generally proportional to the number of nodes themselves. It is typical to see
practical networks which have tens of thousands of links, but millions of OD
pairs.
Therefore, it is natural to find ways to determine an OD matrix which should
be used for traffic assignment. One approach, used in the field of travel demand
modeling, is based on developing behavioral models of how households make
travel choices. Using demographic and other features, there are models to esti-
mate the total number of trips made by households over some period of time,
their destinations, their mode choices, and so forth. Another approach is to
attempt to infer an OD matrix from traffic counts on specified links in the net-
work. Both methods have advantages and disadvantages: travel demand mod-
els provide more insights about underlying behavior (critical when developing
long-range forecasts of future demand), and can directly lead to an OD matrix.
Unfortunately, the data needed to calibrate demand models is more expensive
and cumbersome to work with, often involving recruiting survey participants.
Traffic sensors, on the other hand, collect data automatically, inexpensively,
and without privacy issues, but there is a major dimensionality issue. Since
the number of OD pairs is much larger than the number of links, one cannot
hope to uniquely determine the OD matrix solely from link counts; the problem
is massively underdetermined. In a network where all nodes are zones, it is
trivial to find an OD matrix which perfectly matches link counts, by creating
an OD matrix where all trips travel from one node to an adjacent node. Even
though this matrix is entirely unrealistic, there is no deviation whatsoever from
7.4. OD MATRIX ESTIMATION 243

1510 1490
1 2
3

Figure 7.11: Observed link volumes from traffic sensors along a freeway.

the counts. Even worse, due to unavoidable errors in traffic count records (Fig-
ure 7.11), sometimes this trivial matrix will match counts much better than a
more “realistic” matrix! In this figure, it is likely that the total flow on the
freeway is approximately 1500 vehicles from node 1 to node 3; but this does
not match the counts as well as 1510 vehicles from 1 to 2, and 1490 from 2 to
3. Simply matching counts does not provide the behavioral insight needed to
identify what a “realistic” matrix is.
This section provides a method to reconcile both approaches. The idea is
that an initial OD matrix d∗ is already available from a travel demand model.
While this matrix is hopefully close to the true value, it also contains sampling
and model estimation errors and can never be fully accurate. However, there
are sensors on a subset of links Ā ⊂ A, and there is a vector x∗ of traffic volume
counts on these links. The intent is to use these traffic counts to try to improve
the initial OD matrix d∗ . The following optimization problem expresses this:
X 2
X
2
minf (d, x) = Θ (drs − d∗rs ) + (1 − Θ) xij − x∗ij (7.33)
d,x
(r,s)∈Z 2 (i,j)∈Ā
X Z xij
s.t. x ∈ arg min tij (x) dx (7.34)
x∈X(d) 0
(i,j)∈A

drs ≥ 0 ∀(r, s) ∈ Z 2 (7.35)

where Θ is a parameter ranging from zero to one and X(d) is the set of feasible
link flows when the OD matrix is d.
The objective function (7.33) is of the least-squares type, and attempts to
minimize both the deviation between the final OD matrix d and the initial esti-
mate d∗ , and the deviation between the equilibrium link flows x associated with
the OD matrix d, and the actual observations x∗ on the links with sensors. The
hope is to match traffic counts reasonably well, while not wandering into com-
pletely unrealistic OD matrices. The factor Θ is used to weight the importance
of matching the initial OD matrix estimate, and the link flows. It can reflect the
relative degree of confidence in d∗ and x∗ ; if the travel demand was obtained
from high-quality data and a large sample size, whereas the traffic count data
is old and error-prone, a Θ value close to one is appropriate. Conversely, if
the travel demand model is less trustworthy but the traffic count data is highly
reliable, a lower Θ value is appropriate. In practice, a variety of Θ values can be
chosen, and the resulting tradeoffs between matching the initial estimate and
link flows can be seen.
As indicated by the constraint (7.34), this optimization problem is also a
bilevel program, because the mapping from an OD matrix d to the resulting
244 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

link flows x involves the equilibrium process. The fact that the optimization
problem is bilevel also means that we cannot expect to find the global optimum
OD matrix, and that heuristics should be applied. A sensitivity-based heuristic,
like the one used for the network design problem, would determine how the link
flows would shift if the OD matrix is perturbed, and use this information to find
a “descent direction” which would reduce the value of the objective.
Following the same technique as in Section 7.3, the constraint (7.34) defines
x uniquely as a function of d due to the uniqueness of the link flow solution to
the traffic assignment problem. (Also note that the dependence on d appears
through the feasible region, requiring that the minimization take place over
X(d).) So, we can rewrite the objective function as a function of d alone,
by defining x(d) as the equilibrium link flows in terms of the OD matrix and
transforming the objective to f (d, x(d)). Then, the partial derivative of the
objective with respect to any entry drs in the OD matrix is given by
∂f X ∂xij
rs
= 2Θ(drs − d∗rs ) + 2(1 − Θ) (xrs − x∗rs ) rs (7.36)
∂d ∂d
(i,j)∈Ā

∂x ij
where the partial derivatives ∂drs are found from the sensitivity formulas in
Section 7.2.1.
The vector of all the derivatives (7.36) forms the gradient of f with respect
to d. So, taking a step in the opposite direction, and ensuring that the values
in the OD matrix remain non-negative provide the following update rule:
+
d ← [d − µ∇d f ] (7.37)

where µ is a step size to be determined, and the [·]+ operation is applied to each
component of the vector. This leads to the following algorithm for OD matrix
estimation:
1. Initialize d ← d∗ .
2. Calculate the link flows x(d) by solving the traffic assignment problem
with OD matrix d.
∂f
3. For each OD pair (r, s) determine ∂drs by solving the sensitivity problem
in Section 7.2.1 and using (7.36).
4. Update d using (7.37) for a suitable step size µ.
5. Test for convergence, and return to step 2 if not converged.
The comments about the step size µ and convergence criteria from the network
design problem (Section 7.3) apply equally as well here: µ can be determined
using an iterative line search procedure, choosing smaller values until the ob-
jective function decreases, and the algorithm can be terminated when it fails to
make additional substantial progress in reducing the objective.
This procedure is demonstrated using the network in Figure 7.12. In this
network, traffic counts are available on three of the links in the network, and
7.4. OD MATRIX ESTIMATION 245

1 3 3 4
4652
1 5,000 —
2 — 10,000
5 6
2150
1980
2 4

Figure 7.12: Example network for OD matrix estimation. (Observed link vol-
umes and initial OD matrix shown.)

Table 7.2: Solutions to OD matrix estimation sensitivity problems.

Sensitivity problem
Link d13 d24
(1,3) 0.733 0.067
(1,5) 0.267 −0.067
(2,4) 0.067 0.733
(2,5) −0.067 0.267
(5,6) 0.200 0.200
(6,3) 0.267 −0.067
(6,4) −0.067 0.267

an initial OD matrix is available based on a travel demand model. The link

1
performance function on every link is tij = 10 + xij /100, and Θ is given as 10 .
We begin by initializing the OD matrix to the initial matrix d∗ , and solving
a traffic assignment problem.
As the reader can verify, the equilibrium link flows on the three links with
available counts are x13 = 4733 13 , x25 = 1933 13 , and x56 = 2200. The first
sum in the objective (7.33) gives the fit of the OD matrix to the initial matrix
(zero) and the the second sum gives the fit of the equilibrium link flows to the
observed link flows (11292). When weighted with Θ = 0.1, this gives an initial
objective value of 10164. Next, we solve two sensitivity problems, one for d13
and one for d24 . (If there was reason to believe there were trips between 1 and
4, and between 2 and 3, we would solve sensitivity problems for those OD pairs
as well.) These are shown in Figure 7.13. The link performance functions are
the same in both problems; the only difference is in the demand. The solutions
to these sensitivity problems is shown in Table 7.2.
We now have all the information needed to calculate the gradient of the
objective function, using equation (7.36). Substituting the values of xij , ξij , the
246 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

1 3
  
 
5 6

 


2 4
OD pair (1,3)

1 3

 
5 6

 


2 4
 
OD pair (2,4)

Figure 7.13: Two sensitivity problems for OD matrix estimation.

7.5. HISTORICAL NOTES AND FURTHER READING 247

observed link flows, and the initial OD matrix, we calculate

" # 
∂f 
∂d13 131
∇d f = ∂f = (7.38)
∂d
5.36
24

Taking a trial step in this direction with µ = 1, the updating rule (7.37) gives
the candidate OD matrix d13 = 4869, d24 = 9995. The resulting equilibrium
link flows include x13 = 4637, x25 = 1941, and x56 = 2150, so the “fit” of
the equilibrium link flows and traffic counts has improved from 11293 to 2293.
The “fit” of the OD matrix has worsened from 0 to 17179, but with the weight
Θ = 0.1, the overall objective still decreases from 10164 to 3782. Therefore, we
accept the step size µ = 1, and return to step 3 to continue updating the OD
matrix.

7.5 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
The derivation of the equilibrium sensitivity analysis in this chapter follows
that in Boyles (2012) and Jafari and Boyles (2016). There are alternative ways
to derive the same results, using the implicit function theorem (Tobin and Friesz,
1988; Cho et al., 2000; Yang and Bell, 2007) or results from sensitivity of varia-
tional inequalities (Patriksson, 2004; Lu, 2008). In particular, for application of
the latter approach to the network design problem, see Josefsson and Patriksson
(2007).

7.6 Exercises
1. [63] Given a nondegenerate equilibrium solution in a network with a single
origin and continuous link performance functions, show that the equilib-
rium bushes remain unchanged in a small neighborhood of the current OD
matrix.

2. [65] Given a nondegenerate equilibrium solution in a network with a sin-

gle origin and differentiable link performance functions, show that the
derivatives dxij /ddrs exist at the current equilibrium solution.

3. [56] Extend the results in Exercises 1 and 2 to networks with multiple

origins.

4. [84] Show that the entropy-maximizing path flow solution is a continuous

function of the OD matrix.

5. [26] Verify that the optimality conditions for (7.11)–(7.13) include the
conditions (7.6)–(7.10).
248 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

3
10x
50+x

d14 10+y23+x d14

1 4
60+x
15x
2

Figure 7.14: Network for use in Exercises 6 and 7.

50+x13exp(-y13) 3
10x34exp(-y34)

10+x23exp(-y23)
1 4

15x12exp(-y12)
2 60+x24exp(-y24)

Figure 7.15: Network for use in Exercises 8 and 9.

6. [33] In the modified Braess network of Figure 7.14, find the sensitivity of
each link’s flow to the demand d14 .
7. [35] In the modified Braess network of Figure 7.14, find the sensitivity of
each link’s flow to the link performance function parameter y 23 . What
value of this parameter minimizes the equilibrium travel time? Suggest
what sort of real-world action would correspond to adjusting this param-
eter to this optimal value.
8. [44] In the network design problem for Figure 7.15, give the gradient of
1
the objective function at the initial solution y = 0. Assume that Θ = 20 .

9. [48] Continue Exercise 8 by performing three iterations of the algorithm

given in the text. What is the resulting total system travel time and
construction cost?
10. [25] Write out the network design optimization problem for the network
in Figure 7.16, with Θ = 1. Show that the feasible region for this problem
is not convex by constructing a counterexample.
11. [79] Design a heuristic for the network design problem, based on simulated
annealing as discussed in Section C.6.1. Compare the performance of this
7.6. EXERCISES 249

x1/(1+y1)
10 10
1 2
x2/(1+y2)

Figure 7.16: Network for use in Exercise 10.

3
50+x13 10x34

3 10+x23
1 4
2
15x12
2 60+x24

Figure 7.17: Network and observed link flows (in boxes) for Exercise 13.

heuristic with the algorithm given in the text for several networks.

12. [79] Design a heuristic for the network design problem, based on genetic
algorithms as discussed in Section C.6.2. Compare the performance of this
heuristic with the algorithm given in the text for several networks.

13. [47] In the OD matrix estimation problem of Figure 7.17, give the gradient
of the objective function at the initial solution d∗14 = 6, d∗24 = 4. What is
the value of the objective function if Θ = 21 ?

14. [49] Continue Exercise 13 by performing three iterations of the algorithm

given in the text. What is the resulting OD matrix and objective function
value?

15. [59] Generate five additional OD matrices corresponding to the network

in Figure 7.17, with different values of Θ. Which of these OD matrices
seems most reasonable to you, and why?

16. [79] Design a heuristic for the OD matrix estimation problem, based on
simulated annealing as discussed in Section C.6.1. Compare the perfor-
mance of this heuristic with the algorithm given in the text for several
networks.

17. [79] Design a heuristic for the OD matrix estimation problem, based on
genetic algorithms as discussed in Section C.6.2. Compare the perfor-
250 CHAPTER 7. SENSITIVITY ANALYSIS AND APPLICATIONS

mance of this heuristic with the algorithm given in the text for several
networks.
18. [76] Perform the following “validation” exercise: create a small network
with a given OD matrix, and find the equilibrium solution. Then, given
the equilibrium link flows, try to compute your original OD matrix using
the algorithm given in the text. Do you get your original OD matrix back?
Chapter 8

Extensions of Static
Assignment

The basic traffic assignment problem (TAP) was defined in Chapter 5 as follows:
we are given a network G = (N, A), link performance functions tij (xij ), and the
demand values drs between each origin and destination. The objective is to find
a feasible vector of path flows (or link flows) which satisfy the principle of user
equilibrium, that is, that every path with positive flow has the least travel time
among all paths connecting that origin and destination. We formulated this as
a VI (find ĥ ∈ H such that c(ĥ) · (ĥ − h) ≤ 0 for all h ∈ H) and as the solution
to the following convex optimization problem:
X Z xij
min tij (x)dx (8.1)
x,h 0
(i,j)∈A
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (8.2)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (8.3)
π∈Πrs
π
h ≥0 ∀π ∈ Π (8.4)
This formulation remains the most commonly used version of traffic assign-
ment in practice today. However, it is not difficult to see how some of the
assumptions may not be reasonable. This chapter shows extensions of the basic
TAP which relax these assumptions. This is the typical course of research: the
first models developed make a number of simplifying assumptions, in order to
capture the basic underlying behavior. Then, once the basic behavior is under-
stood, researchers develop progressively more sophisticated and realistic models
which relax these assumptions.
This chapter details three such extensions. Section 8.1 relaxes the assump-
tion that the OD matrix is known and fixed, leading to an elastic demand
formulation. Section 8.2 relaxes the assumption that the travel time on a link

251
252 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

depends only on the flow on that link (and not on any other link flows, even
at intersections). Section 8.3 relaxes the assumption that travelers have accu-
rate knowledge and perception of all travel times in a network, leading to the
important class of stochastic user equilibrium models.
For simplicity, all of these variations are treated independently of each other.
That is, the OD matrix is assumed known and fixed in all sections except Sec-
tion 8.1, and so forth. This is done primarily to keep the focus on the relevant
concept of each section, but also to guard the reader against the temptation
to assume that a model which relaxes all of these assumptions simultaneously
is necessarily better than one which does not. While realism is an important
characteristic of a model, it is not the only relevant factor when choosing a math-
ematical model to describe an engineering problem. Other important concerns
are computation speed, the existence of enough high-quality data to calibrate
and validate the model, transparency, making sure the sensitivity of the model
is appropriate to the level of error in input data, ease of explanation to decision
makers, and so on. All of these factors should be taken into account when choos-
ing a model, and you can actually do worse off by choosing a more “realistic”
model when you don’t have adequate data for calibration — the result may even
give the impression of “false precision” when in reality your conclusions cannot
be justified.

8.1 Elastic Demand

The assumption that the OD matrix is known and fixed can strain credibility,
particularly when considering long time horizons (20–30 years) or when projects
are major enough to influence travel decisions at all levels, not just route choice.
For instance, consider the “induced demand” phenomenon where major expan-
sion of roadway capacity ends up increasing the amount of demand. This is
partly due to changes in route choice (which the basic TAP accounts for), but
is also due to changes in other kinds of travel choices, such as departure time,
mode, destination, or trip frequency.
Therefore, it is desirable to develop a model which can relax the assumption
of an exogenous OD matrix known a priori. This section describes how TAP can
be extended to accommodate this relaxation. This is called the elastic demand
formulation of TAP. The elastic demand model is at once more and less useful
than basic TAP: more useful because it can provide a more accurate view of the
impacts of transportation projects; less useful because it is harder to calibrate.

8.1.1 Demand functions

The new idea in the traffic assignment problem with elastic demand is the
demand function, which relates the demand for travel between an origin and
destination to the travel time between these zones. Specifically, let Drs be a
function relating the demand between r and s to the travel time κrs on the
8.1. ELASTIC DEMAND 253

shortest path between these zones.1 Generally, Drs is a nonincreasing function

— as the travel time between r and s increases, the demand for travel between
these nodes is lower. It will also be highly useful to assume that Drs is invertible
as well, which will require it to be strictly decreasing. The inverse demand
function will give the travel time between r and s corresponding to a given
demand level.
As an example, let Drs (κrs ) = 200 exp(−κrs /100). If the travel time between
r and s is κrs = 10 by the shortest path, then the demand is drs = 181. If the
travel time was 20 minutes between these zones, the demand will be 163.7, which
is lower as fewer drivers choose to travel between r and s. The inverse demand
−1 rs
function is Drs (d ) = 100 log(200/drs ), and can be used to calculate the κrs
value corresponding to a given drs : when the demand is 181, the shortest path
−1
travel time is κrs = Drs (181) = 10, and so on.2 Or, if Drs = 50 − 21 κrs , then
−1 rs
Drs = 100 − 2d and a travel time of 10 minutes corresponds to a demand of
45 vehicles as can be seen by substituting either number into the corresponding
equation.
An attentive reader may have noticed a potential issue with the demand
function Drs = 50 − 21 κrs , namely that the demand would be negative if κrs >
100. In reality, the demand would simply equal zero if the travel time exceeded
100. We could patch this by redefining Drs as [50 − 21 κrs ]+ , but then Drs is
no longer invertible. Instead, we can allow Drs to take negative values, but
replace the relation drs = Drs (κrs ) with drs = [Drs (κrs )]+ . This allows us to
have Drs be strictly decreasing (and thus invertible), but still allow the travel
demand to be zero when costs are sufficiently high. While this “trick” may seem
a bit trivial (or at least not very useful), it will eventually allow us to formulate
the elastic demand equilibrium problem as a variational inequality and convex
program, as shown below.
The demand function can be used to define a consumer surplus CS repre-
senting the benefits of mobility in a region, defined by

X Z drs !
−1 rs rs
CS = Drs (y) dy − d κ (8.5)
(r,s)∈Z 2 0

(We
R −1are using y for the dummy variable of integration instead of d because
D (d) dd is notationally awkward.) The interpretation of this formula is as
follows. Each driver has a certain travel time threshold: if the travel time is
greater than this threshold, the trip will not be made, and if the travel time is
less than this threshold, the trip will be made. Different drivers have different
1 An alternative is to have D rs be a function of the average travel time on the used paths

between r and s, not the shortest. At equilibrium it doesn’t matter because all used paths have
the same travel time as the shortest, but in the process of finding an equilibrium the alternative
definition of the demand function can be helpful. For our purposes, though, definition in terms
of the shortest path time is more useful because it facilitates a link-based formulation.
2 In the text, we typically indicate OD pairs with a superscript, as in drs , and link variables

with a subscript, as in xij . In elastic demand, we will often need to refer to inverse functions,
and writing (Drs )−1 is clumsy. For this reason, OD pairs may also be denoted with a subscript,
−1
as in Drs . This is purely for notational convenience and carries no significance.
254 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Travel time

Consumer surplus
k
D-1(d)

d Demand
TSTT

Figure 8.1: Relationship between inverse demand function, consumer surplus,

and total system travel time.

thresholds, and the demand function represents the aggregation of these thresh-
olds: when the travel time is κ, D(κ) represents the number of travelers whose
threshold is κ or higher. If my threshold is, say, 15 minutes and the travel time
is 10 minutes, the difference (5 minutes) can be thought of as the “benefit” of
travel to me: the trip is worth 15 minutes of my time, but I was able to travel
for only 10. Adding this up for all travelers provides the total benefits of travel,
which is what CS represents. Figure 8.1 shows the connection between this
concept and equation (8.5): assume that drivers are numbered in decreasing
order of their threshold values. Then D−1 (1) gives the threshold value for the
first driver, D−1 (2) gives the threshold value for the second driver, and so forth.
At equilibrium all drivers experience a travel time of κ, so the benefit to the
first driver is D−1 (1) − κ, the benefit to the second driver is D−1 (2) − κ, and
so forth. Adding over all drivers gives equation (8.5).

8.1.2 Gartner’s transformation

Before moving to variational inequality and optimization formulations of the
elastic demand problem, we’ll take a short digression and show how the elas-
tic demand problem can be cleverly transformed into a traditional equilibrium
problem with fixed demand. This transformation works if each demand function
Drs is bounded above. Repeat the following for each OD pair (r, s). Let drs
be such an upper bound for OD pair (r, s). Create a new link (r, s) directly
connecting origin r to destination s, and make its link performance function
−1
trs (xrs ) = Drs (drs − xrs ) where xrs is the flow on that new link. If Drs is
−1
decreasing, then Drs (drs − xrs ) is increasing in xrs so this is a valid link per-
formance function.
Now, solve a fixed demand problem where the demand from each origin to
each destination is drs . An equilibrium on this network corresponds to an elastic
8.1. ELASTIC DEMAND 255

demand equilibrium on the original network as follows: the flows on the links
common to both networks represent flows on the actual traffic network; the flow
on the new direct connection links represent drivers who choose not to travel
due to excess congestion. Think of drs as the total number of people who might
possibly travel from r to s; those that actually complete their trips travel on the
original links and those who choose not to travel choose the direct connection
link. At equilibrium, all used paths connecting r to s (including the direct
connection link) have the same travel time κrs ; therefore, the flow on the direct
−1
connection link xrs must be such that Drs (drs − xrs ) = κrs , or equivalently
xrs = drs − Drs (κrs ), which is exactly the number of drivers who choose not to
travel when the equilibrium times are κrs . That is, the demand drs = drs − xrs .
The downside of this approach is that it requires creating a large number
of new links. In a typical transportation network, the number of links is pro-
portional to the number of nodes and of the same order of magnitude (so a
network of 1,000 nodes may have 3–4,000 links). However, the number of OD
pairs is roughly proportional to the square of the number of nodes, since every
node could potentially be both an origin and a destination. So, a network with
1,000 nodes could have roughly 1,000,000 OD pairs. Implementing the Gartner
transformation requires creating a new link for every one of these OD pairs,
which would result in 99.9% of the network links being the artificial arcs for the
transformation!

8.1.3 Variational inequality formulation

Because the elastic demand problem can be expressed as a version of the regu-
lar traffic assignment problem through the Gartner transformation, we immedi-
ately have a variational inequality formulation of the elastic demand equilibrium
problem. Partition the vectors of link flows and travel times into regular and
direct-connect (Gartner transformation) links, using x to represent regular link
flows, x→ flows on direct-connect links, and t and t→ similarly. Then the vari-
ational inequality is    
t(x̄) x̄ − x
· ≤0 (8.6)
t→ (x̄→ ) x̄→ − x→
−1
Since xrs = drs − drs and trs (xrs ) = Drs (drs − xrs ), the variational inequality
can be written in terms of the link flows and OD demands as
   
t(x̄) x̄ − x
· ≤0 (8.7)
−D−1 (d̄) d̄ − d
which is the customary form.

8.1.4 Optimization formulation

The Gartner transformation can also lead directly to a convex programming
formulation of the elastic demand problem, in a similar way as the variational
inequality was derived in the previous section. However, it is also instructive to
derive the convex programming formulation from first principles.
256 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

As discussed above, the demand is related to the demand function by drs =

[D (κrs )]+ . Put another way, the demand must always be at least as much as
rs

the demand function; further, if the demand is greater than zero then it must
equal the demand function. Thinking laterally, you might notice this is similar
to the principle of user equilibrium: the travel time on any path must always be
at least as large as the shortest path travel time; further, if the demand is pos-
itive then the path travel time must equal the shortest path travel time. When
deriving the Beckmann function, we showed that the latter statements could be
expressed by cπ ≥ κrs and hπ (cπ − κrs ) = 0 (together with the nonnegativity
condition hπ ≥ 0). The same “trick” applies for the relationship between de-
mand and the demand function: drs ≥ Drs (κrs ), drs (drs − Drs (κrs )) = 0, and
the nonnegativity condition drs ≥ 0.
It will turn out to be easier to express the latter conditions in terms of the
inverse demand functions D−1 , rather than the “forward” functions D, because
the convex objective function we will derive will be based on the Beckmann
function. The Beckmann function involves link performance functions (with
units of time). Since the inverse demand functions also are measured in units of
time, it will be easier to combine them with the link performance functions than
the regular demand functions (which have units of vehicles). Expressed in terms
−1 rs
of the inverse demand functions, the conditions above become κrs ≥ Drs (d ),
rs −1 rs rs rs
d (Drs (d ) − κ ) = 0, and d ≥ 0.
So, this is the question before us. What optimization problem has the fol-
lowing as its optimality conditions?

cπ ≥ κrs ∀(r, s) ∈ Z 2 , π ∈ Πrs (8.8)

π π rs 2 rs
h (c − κ ) = 0 ∀(r, s) ∈ Z , π ∈ Π (8.9)
rs −1 rs 2
κ ≥ Drs (d ) ∀(r, s) ∈ Z (8.10)
rs −1 rs rs 2
d (Drs (d ) −κ )=0 ∀(r, s) ∈ Z (8.11)
X
hπ = drs ∀(r, s) ∈ Z 2 (8.12)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (8.13)
rs 2
d ≥0 ∀(r, s) ∈ Z (8.14)

The Beckmann formulation is a good place to start, since it already includes

(8.8), (8.9), (8.12), and (8.13). So let’s start by conjecturing that the Lagrangian
takes the form
π π
P !
X Z π∈Π δij h X X
L(h, d, κ) = tij (x) dx+ κrs drs − hπ +F (d)
(i,j)∈A 0 (r,s)∈Z 2 π∈Πrs
(8.15)
where F (d) is some function involving the OD matrix. (Note also that L is
now a function of d in addition to h and κ, since the demand is a decision
variable.) You can check that the optimality conditions related to κ and h are
already included in the list of optimality conditions above. Assuming that there
8.1. ELASTIC DEMAND 257

is a nonnegativity constraint on the demand, (8.14) follows immediately as well.

What’s left is to show that the conditions ∂L/∂drs ≥ 0 and drs (∂L/∂drs ) = 0
correspond to (8.10) and (8.11).
Calculating from (8.15), we have

∂L ∂F
rs
= κrs + rs ,
∂d ∂d

∂F −1 rs
so if = −Drs
∂drs (d ), we are done (both equations will be true). Integrating,
P R drs −1
F (d) = − (r,s)∈Z 2 0 Drs (ω) dω gives us what we need. De-Lagrangianizing
the “no vehicle left behind” constraint, we obtain the optimization problem
associated with the elastic demand problem:

X Z xij X Z drs
−1
min tij (x) dx − Drs (ω) dω (8.16)
x,h,d 0 0
(i,j)∈A (r,s)∈Z 2
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (8.17)
π∈Π
X
hπ = drs ∀(r, s) ∈ Z 2 (8.18)
π∈Πrs
hπ ≥ 0 ∀π ∈ Π (8.19)
rs 2
d ≥0 ∀(r, s) ∈ Z (8.20)

8.1.5 Solution method

This section shows how the Frank-Wolfe algorithm can be used to solve the
optimization problem (8.16)–(8.20). This is certainly not the only choice, and it
is worthwhile for you to think about how other algorithms from Chapter 6 could
also be used instead of Frank-Wolfe. The implementation of this algorithm is
quite similar to how Frank-Wolfe works for the basic traffic assignment problem,
with three changes. First, since the OD matrix is a decision variable along with
the link flows, we must keep track of both d as well as x; therefore, in addition
to the target link flows x∗ we will have a target OD matrix d∗ , and in addition
when we update the link flows we must update the OD matrix as well. Each
of these pairs of flows and OD matrices should be consistent with each other,
in that the link flows x must be a feasible network loading when the demand
is d (both before and after updating), and similarly x∗ must correspond to d∗ .
Luckily, this will not be difficult.
The second change to the Frank-Wolfe algorithm is how the restricted vari-
ational inequality is solved. Instead of solving t(x̄0 ) · (x∗ − x) ≤ 0 for x̄0 =
λx∗ + (1 − λ)x, λ ∈ [0, 1], we must solve the variational inequality (8.7) in x̄0
and d̄0 . As before, the usual solution involves an “interior” λ ∈ (0, 1), in which
258 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

case we must solve the equation

X
tij (λx∗ij + (1 − λ)xij )(x∗ij − xij )
(i,j)∈A
X
−1
− Drs (λd∗rs + (1 − λ)drs )(d∗rs − drs ) = 0 (8.21)
(r,s)∈Z 2

in λ. This is simply the variational inequality (8.7) written out in terms of its
components, substituting λx∗ + (1 − λ)x for x̄0 and λd∗ + (1 − λ)d for d̄0 .
Third, the stopping criterion (relative gap or average excess cost) needs to be
augmented with a measure of how well the OD matrix matches the values from
the demand P functions. A simple measure is the total misplaced flow defined
as T M F = (r,s)∈Z 2 |drs − [Drs (κrs )]+ |. The total misplaced flow is always
nonnegative, and is zero only if all of the entries in the OD matrix are equal
to the values given by the demand function (or zero if the demand function is
negative). We should keep track of both total misplaced flow and one of the
equilibrium convergence measures (relative gap or average excess cost), and only
terminate the algorithm when both of these are sufficiently small.
Implementing these changes, the Frank-Wolfe algorithm for elastic demand
is as follows:
1. Choose some initial OD matrix d and initial link flows x corresponding to
that OD matrix.
2. Find the shortest path between each origin and destination, and calculate
convergence measures (total misplaced flow, and either relative gap or
average excess cost). If both are sufficiently small, stop.
3. Improve the solution:
(a) Calculate a target OD matrix d∗ using the demand functions: d∗rs =
[Drs (κrs )]+ for all OD pairs (r, s).
(b) Using the target matrix d∗ , find the link flows if everybody were
traveling on the shortest paths found in step 2, store these in x∗ .
(c) Solve the restricted variational inequality by finding λ such that (8.21)
is true.
(d) Update the OD matrix and link flows: replace x with λx∗ + (1 − λ)x
and replace d with λd∗ + (1 − λ)d.
4. Return to step 1.
This algorithm can also be linked to the convex programming formulation
described above. Given a current solution (x, d), it can be shown that the deriva-
tive of the Beckmann function in the direction towards (x∗ , d∗ ) is nonpositive
(and strictly negative if the current solution does not solve the elastic demand
problem), and that the solution of the restricted variational inequality (8.21)
minimizes the objective function along the line joining (x, d) to (x∗ , d∗ ). The
algebra is a bit tedious and is left as an exercise at the end of the chapter.
8.2. LINK INTERACTIONS 259

8.1.6 Example
Here we solve the small example of Figure 8.2 with the Frank-Wolfe algorithm,
using the average excess cost to measure how close we are to equilibrium. The
demand function is D(κ) = 50 − κ, so its inverse function is D−1 (d) = 50 − d.

Initialization. Arbitrarily set

 d =50, then arbitrarily load all 50 vehicles onto
the two links; say x = 50 0 .
 
Iteration 1. The link travel times are now t = 60 20 , so the shortest path
travel time κ = 20 and the demand function indicates that the demand
should be D = 50 − 20 = 30. The average excess cost is (60 × 50 − 20 ×
50)/50 = 40, and the total misplaced flow is |50 − 30| = 20. The target
demand is what the demand function indicates d∗ = 30, and this flow
should all be loaded on the bottom path, so x∗ = 0 30 . Solving the
equation

(10 + 50(1 − λ))(−50) + (20 + 30λ)(30) − (50 − (30λ + 50(1 − λ)))(−20) = 0

we obtain λ = 12/19 ≈ 0.632 so the new demand and flows are d = 37.36
and x = 18.42 18.95 .
 
Iteration 2. The link travel times are now t = 28.42 38.95 , so κ = 28.42,
D = 21.58, AEC = 5.19, and T M F = 15.78. Both convergence measures
have decreased from the first iteration,
 particularly the average excess
cost. Thus, d∗ = 21.58, x∗ = 21.58 0 , and we solve

(10+21.58λ+18.42(1−λ))(21.58−18.42)+(20+18.95(1−λ))(0−18.95)
− (50 − (21.58λ + 37.36(1 − λ)))(21.58 − 37.36) = 0

so
 λ = 0.726  and the new demand and flows are d = 25.9 and x =
20.71 5.19
 
Iteration 3 . The link travel times are now t = 30.71 25.19 , so κ = 25.19,
D = 24.09, AEC = 4.42, and T M F = 1.80. Assuming that these are
small enough to terminate, we are done.

8.2 Link Interactions

This section develops another extension of TAP, in which we relax the assump-
tion that the travel time on a link depends only on the flow on that link. There
are a few reasons why this kind of extension may be useful:
Junction interactions: At an intersection, the delay on a particular approach
often depends on flows from competing approaches. As an example, con-
sider a freeway onramp which has to yield to mainline traffic at a merge.
Because merging traffic must find an acceptably large gap in the main
260 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

10 + x

50 – κ
1 2
20 + x

Figure 8.2: Example network for elastic demand equilibrium.

lanes, the travel time on the onramp depends on the flow on the main
lanes as well as the flow on the onramp. Similar arguments hold at arte-
rial junctions controlled by two-way or four-way stops, at signalized inter-
sections with permissive phases (e.g., left-turning traffic yielding to gaps
in oncoming flow), or at actuated intersections where the green times are
determined in real-time based on available flow. The basic TAP cannot
model the link interactions characterizing these types of links.

Overtaking traffic: On rural highways, overtaking slow-moving vehicles of-

ten requires finding a (fairly large) gap in oncoming flow. If the oncoming
flow is small, the effect of slow-moving vehicles on average travel time is
negligible. However, as the oncoming flow becomes larger and larger, the
ability to overtake is diminished and traffic speeds will tend to be deter-
mined by the slowest-moving vehicle on the highway. Since traffic moving
in different directions on the same highway is modeled with different links,
a link interaction model is needed to capture this effect.

Multiclass flow: Consider a network model where there are two types of ve-
hicles (say, passenger cars and semi trucks, or passenger cars and buses).
Presumably these vehicles may choose routes differently or even have a dif-
ferent roadway network available to them (heavy vehicles are prohibited
from some streets, and buses must drive along a fixed route. This type of
situation can be modeled by creating a “two-layer” network, with the two
layers representing the links available to each class. However, where these
links represent the same physical roadway, the link performance functions
should be connected to each other (truck volume influences passenger car
speed and vice versa) even if they are not identical (truck speed need not
be the same as passenger car speed). Link interaction models therefore
allow us to model multiclass flow as well.

However, there are a few twists to the story, some of which are explored
below. Section 8.2.1 presents a mathematical formulation of the link interactions
model, but shows that a convex programming formulation is not possible except
8.2. LINK INTERACTIONS 261

in some rather unlikely cases. Section 8.2.2 explores the properties of the link
interactions model, in particular addressing the issue of uniqueness — even
though link flow solutions to TAP are unique under relatively mild assumptions,
this is not generally true when there are link interactions. Section 8.2.3 gives
us two solution methods for the link interactions model, the diagonalization
method and simplicial decomposition. Diagonalization is easier to implement,
but simplicial decomposition is generally more powerful.

8.2.1 Formulation
In the basic TAP, the link performance function for link (i, j) was a function
of xij alone, that is, we could write tij (xij ). Now, tij may depend on the flow
on multiple links. For full generality, our notation will allow tij to depend on
the flows on any or all other links in the network: the travel time is given by
the function tij (x1 , x2 , · · · , xij , · · · , xm ) or, more compactly, tij (x) using vector
notation. We assume these are given to us. Everything else is the same as
in vanilla TAP: origin-destination demand is fixed, and we seek an equilibrium
solution where all used paths have equal and minimal travel time.
Now, how to formulate the equilibrium principle? It’s not hard to see that
the variational inequality for TAP works equally well here:
c(h̄) · (h̄ − h) ≤ 0 ∀h ∈ H (8.22)
where the only difference is that the link performance functions used to calculate
path travel times C are now of the form tij (x) rather than tij (xij ). But this is
of no consequence. Path flows h̄ solve the variational inequality if and only if
c(h̄) · h̄ ≤ c(h̄) · h (8.23)
for any other feasible path flows h. That is, if the travel times were fixed at their
current values, then it is impossible to reduce the total system travel time by
changing any drivers’ route choices. This is only possible if all used paths have
equal and minimal travel time. Similarly, the link-flow variational inequality
t(x̄) · (x̄ − x) ≤ 0 , (8.24)
where x is any feasible link flow, also represents the equilibrium problem with
link interactions.
The ease of translating the variational inequality formulation for the case of
link interactions may give us hope that a convex programming formulation exists
as well. The feasible region is the same, all we need is to find an appropriate
objective function. Unfortunately, this turns out to be a dead end. For example,
the obvious approach is Z to amend the Beckmann function in some way, for
X xij
instance, changing tij (x) dx to
(i,j)∈A 0

X Z x
tij (y) dy (8.25)
(i,j)∈A 0
262 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

where the simple integral in the Beckmann function is replaced with a line
integral between the origin 0 and the current flows x. Unfortunately, this line
integral is in general not well-defined, since its value depends on the path taken
between the origin and x.
The one exception is if the vector of travel times t(x) is a gradient map (that
is, it is a conservative vector field). In this case, the fundamental theorem of
line integrals implies that the value of this integral is independent of the path
taken between 0 and x. For t(x) to be a gradient map, its Jacobian must be
symmetric. That is, for every pair of links (i, j) and (k, `), we need the following
condition to be true:
∂tij ∂tk`
= (8.26)
∂xk` ∂xij
That is, regardless of the current flow vector x, the marginal impact of another
vehicle added to link (i, j) on the travel time of (k, `) must equal the marginal
impact of another vehicle added to link (k, `) on the travel time of (i, j). This
condition is very strong. Comparing with the motivating examples used to jus-
tify studying link interactions, the symmetry condition is not usually satisfied:
the impact of an additional unit of flow on the mainline on the onramp travel
time is much greater than the impact of an additional unit of onramp flow on
mainline travel time. The impact of semi truck flow on passenger car travel
time is probably greater than the impact of passenger car flow on truck travel
time at the margin. Symmetry may perhaps hold in the case of overtaking on a
rural highway, but even then it is far from clear. So, when modeling link inter-
actions we cannot hope for condition (8.26) to hold. If it does so, consider it a
happy accident: the function (8.25) is then an appropriate convex optimization
problem.

8.2.2 Properties
This section explores the properties of the link interaction equilibrium prob-
lem defined by the variational inequality (8.22). The first question concerns
existence of an equilibrium. Because (8.22) is essentially the same variational
inequality derived for TAP, the arguments used to derive existence of an equi-
librium (based on Brouwer’s theorem) carry over directly and we have the same
result:

Proposition 8.1. If each link performance function tij (x) is continuous in the
vector of link flows x, then at least one solution exists satisfying the principle
of user equilibrium.

However, uniqueness turns out to be trickier. Consider a network of two

parallel links where the demand is 6 vehicles, and the link performance functions
are t1 = x1 + 2x2 and t2 = 2x1 + x2 . Setting the two links’ travel times equal
to each other and using x1 + x2 = 6, it is easy to see that one equilibrium is
the solution x1 = x2 = 3 when the travel time on both links is 9. However, this
is not the only equilibrium solution: if all of the drivers were to choose link 1,
8.2. LINK INTERACTIONS 263

Travel time

12 t2

6 t1

3 6 x1

Figure 8.3: Change in path travel times as x1 varies, “artificial” two-link net-
work.

then x1 = 6, x2 = 0, t1 = 6, and t2 = 12. The top link is the only used path,
but it has the least travel time so this solution also satisfies the principle of user
equilibrium. Likewise, if x1 = 0 and x2 = 6, then t1 = 12 and t2 = 6 and again
the only used path has the least travel time. Therefore, this network has three
equilibrium solutions; compare with Figure 8.3.
To make this situation less artificial, we can change the link performance
functions to represent a more realistic scenario. Assume that the rate of demand
is 1800 vehicles per hour, and that link 1 has a constant travel time of 300
seconds independent of the flow on either link. Link 2 is shorter with a free-flow
time of 120 seconds, but must yield to link 1 using gap acceptance principles.
In traffic operations, gap acceptance is often modeled with two parameters: the
critical gap tc , and the follow-up gap tf . The critical gap is the smallest headway
required in the main stream for a vehicle to enter. Given that the gap is large
enough for one vehicle to enter the stream, the follow-up gap is the incremental
amount of time needed for each additional vehicle to enter. For this example,
let tc be 4 seconds and tf be 2 seconds. Then, assuming that flows on both
links 1 and 2 can be modeled as Poisson arrivals, the travel time on link 2 can
be derived as
" r #
1 Λ x2 x2 2 8x
2
t2 (x1 , x2 ) = + −1+ −1 + 2 (8.27)
u 4 u u u Λ

where Λ is the length of the analysis period and u is the capacity of link 2
defined by
x2 exp(−x1 tc )
u= (8.28)
1 − exp(−x1 tf )
264 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Figure 8.4: Change in path travel times as x1 varies, “realistic” two-link net-
work.

Figure 8.4 shows the travel times on the two paths as x1 varies. Again, there
are three equilibria: (1) x1 = 0, x2 = 1800, where t1 = 300 and t2 = 182; (2)
x1 = 362, x2 = 1438, where t1 = t2 = 300; and (3) x1 = 892, x2 = 908, where
t1 = t2 = 300.
So, even in realistic examples we cannot expect equilibrium to be unique
when there are link interactions. The practical significance is that it raises
doubt about which equilibrium solution should be used for project evaluation
or ranking. For instance, consider a candidate project which would improve
the free-flow time on link 1 from 300 to a smaller value; this would correspond
to lowering the horizontal line in Figure 8.4. If we are at one of the equilibria
where the travel times are equal, such a change will indeed reduce the travel
times experienced by drivers. However, if we are at the equilibrium where the
top path is unused, such a change will have no impact whatsoever.
While a complete study of the methods used to distinguish among multiple
equilibria is beyond the scope of this section, a simple stability criterion is
explained here: an equilibrium solution is stable if small perturbations to the
solution would incentivize drivers to move back towards that initial equilibrium
— that is, if we reassign a few drivers to different paths, the path travel times
will change in such a way that those drivers would want to move back to their
original paths. By contrast, an unstable equilibrium does not have this property:
if a few drivers are assigned to different paths, the path travel times will change
in such a way that even more drivers would want to switch paths, and so on
until another equilibrium is found.
In the simple two-link network we’ve been looking at, stability can be iden-
tified using graphs such as those in Figures 8.3 and 8.4. The arrows on the
bottom axis indicate the direction of the path switching which would occur for
a given value of x1 . When the arrow is pointing to the left, t1 > t2 so travelers
want to switch away from path 1 to path 2, resulting in a decrease in x1 (a move
8.2. LINK INTERACTIONS 265

further to the left on the graph). When t2 > t1 , travelers want to switch away
from path 2 to path 1, resulting in an increase in x1 , indicated by an arrow
pointing to the right. At the equilibrium solutions, there is no pressure to move
in any feasible direction. So, for the first example, the only stable equilibria are
the “extreme” solutions with all travelers on either the top or bottom link. The
equilibrium with both paths used is unstable in the sense that any shift from
one path to another amplifies the difference in travel times and encourages even
more travelers to shift in that direction. In the second example, the first and
third equilibria are stable, but the second is unstable.
Based on these two examples, an intuitive explanation for the presence of
stability with link interactions can be provided. For the regular traffic assign-
ment problem with increasing link performance functions, shifting flow away
from a path π and onto another path π 0 always decreases the travel time on
π and increases the travel time on π 0 . Therefore, if the paths have different
travel times, flow will shift in a way that always tends to equalize the travel
times on the two paths. Even where there are link interactions, the same will
hold true if the travel time on a path is predominantly determined by the flow
on that path. However, when the link interactions are very strong, the travel
time on a path may depend more strongly by the flow on a different path. In
the first example, notice that each link’s travel time is influenced more by the
other link’s flow than its own. In the second example, for certain ranges of flow
the travel time on the merge path is influenced more by the flow on the priority
path. In such cases, there is no guarantee that moving flow from a higher-cost
path to a lower-cost path will tend to equalize their travel times. In the first
example, we have an extreme case where moving flow to a path decreases its
travel time while increasing the travel time of the path the flow moved away
from!
To make this idea more precise, the following section introduces the mathe-
matical concept of strict monotonicity.

Strict Monotonicity
Let f (x) be a vector-valued function whose domain and range are vectors of the
same dimension. For instance, t(x) maps the vector of link flows to the vector
of link travel times; the dimension of both of these is the number of links in the
network. We say that f is strictly monotone if for any two distinct vectors x and
y in its domain, the dot product of f (x) − f (y) and x − y is strictly positive.
For example, let f (x) be defined by f1 (x1 , x2 ) = 2x1 and f2 (x1 , x2 ) = 2x2 .
Then for any distinct vectors x and y, we have
       
(f (x) − f (y)) · (x − y) = ( 2x1 2x2 − 2y1 2y2 ) · ( x1 x2 − y1 y2 )
= 2((x1 − y1 )2 + (x2 − y2 )2 )

Since x 6= y, the right-hand side is always greater than zero, so f is mono-

tone. As another example, let the function g(x) be defined by g1 (x1 , x2 ) = 2x2
266 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT
   
and g2 (x1 , x2 ) = 2x1 . If we choose x = 0 1 and y = 1 0 , then
   
(g(x) − g(y)) · (x − y) = ( 2 −2 ) · ( −1 1 ) = −4

so g is not strictly monotone. Note that proving strict monotonicity requires a

general argument valid for any distinct vectors x and y; showing that a function
is not strictly monotone only requires a single counterexample.
Warning! It is very common for students to think that the link performance
functions are strictly monotone if they are strictly increasing functions of the
flow on each link. This is not true: in the first example in this section, all link
performance
  functions are increasing in each flow variable but if we compare
x = 0 6 and y = 5 1 , we have
       
(t(x) − t(y)) · (x − y) = ( 12 6 − 7 11 ) · ( 0 6 − 5 1 ) = −50

so these link performance functions are not strictly monotone. Roughly speak-
ing, strict monotonicity requires the diagonal terms of the Jacobian of t to be
large compared to the off-diagonal terms. The precise version of this “roughly
speaking” fact is the following:
Proposition 8.2. If f is a continuously differentiable function whose domain is
convex, then f is strictly monotone if and only if its Jacobian is positive definite
at all points in the domain.
With this definition of monotonicity in hand, we can provide the uniqueness
result we’ve been searching for:
Proposition 8.3. Consider an instance of the traffic assignment problem with
link interactions. If the link performance functions t(x) are continuous and
strictly monotone, then there is exactly one user equilibrium solution.
Proof. Since t(x) is continuous, we are guaranteed existence of at least one equi-
librium solution from Brouwer’s theorem; let x̂ be such an equilibrium and let
x̃ be any other feasible link flow solution. We need to show that x̃ cannot be an
equilibrium. Arguing by contradiction, assume that x̃ is in fact an equilibrium.
Then it would solve the variational inequality (8.24), so

t(x̃) · (x̃ − x̂) ≤ 0

Adding a clever form of zero to the left hand side, this would imply

(t(x̃) − t(x̂)) · (x̃ − x̂) + t(x̂) · (x̃ − x̂) ≤ 0 (8.29)

But since the link performance functions are strictly monotone, the first term
on the left-hand side is strictly positive. Furthermore, since x̂ is an equilibrium
the variational inequality (8.24) is true, so t(x̂) · (x̂ − x̃) ≤ 0, which implies that
the second term on the right-hand side is nonnegative. Therefore, the left-hand
side of (8.29) is strictly positive, which is a contradiction. Therefore x̃ cannot
satisfy the principle of user equilibrium.
8.2. LINK INTERACTIONS 267

8.2.3 Algorithms
This section presents two algorithms for the traffic assignment problem with link
interactions. If the link performance functions are strictly monotone, it can be
shown that both of these algorithms converge to the unique equilibrium solution.
Otherwise, it is possible that these algorithms may not converge, although they
will typically do so if they start sufficiently close to an equilibrium. In any case,
these algorithms may be acceptable heuristics even when strict monotonicity
does not hold.

Diagonalization
The diagonalization method is a variation of Frank-Wolfe, which differs only in
how the step size λ is found. Recall that the Frank-Wolfe step size is found by
solving the equation
X
tij (λx∗ij + (1 − λ)xij )(x∗ij − xij ) = 0 (8.30)
(i,j)∈A

since this minimizes the Beckmann function along the line segment connecting x
to x∗ . Since there is no corresponding objective function when there are asym-
metric link interactions, it is not clear that a similar approach will necessarily
work. (And in any case, tij is no longer a function of xij alone, so the formula
as stated will not work.)
To make this formula logical, construct a temporary link performance func-
tion t̃ij (xij ) which only depends on its own flow. This is done by assuming that
the flow on all other links is constant: t̃ij (xij ) = tij (x1 , . . . , xij , . . . , xm ). For
example, if t1 (x1 , x2 , x3 ) = x1 + x22 + x33 and the current solution is x1 = 1,
x2 = 2, and x3 = 3, then t̃1 (x1 ) = 31 + x1 , since this is what we would get if x2
and x3 were set to constants at their current values of 2 and 3, respectively.
The step size λ is then found by adapting the Frank-Wolfe formula, using
t̃ij (xij ) in place of tij (x). That is, in the diagonalization method λ solves
X
t̃ij (λx∗ij + (1 − λ)xij )(x∗ij − xij ) = 0 (8.31)
(i,j)∈A

At each iteration, new t̃ functions are calculated based on the current solu-
tion. The complete algorithm is as follows:

1. Find the shortest path between each origin and destination, and calculate
the relative gap (unless it is the first iteration). If the relative gap is
sufficiently small, stop.

2. Shift travelers onto shortest paths:

(a) Find the link flows if everybody were traveling on the shortest paths
found in step 1, store these in x∗ .
268 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

50+x13+0.5x23 3
10x34+5x24

10+x23+0.5x13
1 4

10x12
50+x24+0.5x34
2

Figure 8.5: Braess network with link interactions.

(b) If this is the first iteration, set x ← x∗ and move to step 3. Otherwise,
continue with step c.
(c) Using the current solution x, form the diagonalized link performance
functions t̃ij (xij ) for each link.
(d) Find λ which solves equation (8.31).
(e) Update x ← λx∗ + (1 − λ)x.
3. Calculate the new link travel times and the relative gap. Increase the
iteration counter i by one and return to step 1.
As an example, consider the modified Braess network shown in Figure 8.5.
At each merge node, the travel time on each of the incoming links depends on
the flow
 on both links which merge
 together. The link flow vectors are indexed
x = x12 x13 x23 x24 x34 .

Iteration 1. In the first iteration, load

 all travelers onto shortest paths, so
x = x∗ = 6 0 6 0 6 , t = 60 53 16 53 60 and the average
 

excess cost is 23.

Iteration
 2. The all-or-nothing
 loading on shortest paths given t is x∗ =
0 6 0 0 6 . The diagonalized link performance functions are ob-
tained by assuming the flows on all other links are constant at x: t̂12 =
10x12 , t̂13 = 53 + x13 , t̂23 = 10 + x23 , t̂24 = 53 + x24 , and t̂34 = 10x34 . So
we solve the equation (8.31) for λ:

(53 + 6λ)6 + 10(6(1 − λ))(−6) + (10 + 6(1 − λ))(−6) = 0

omitting terms where xij = x∗ij because they are zero in (8.31). The
 1
1 11 1

solution
 5is λ = 23/72, so we update x = 4 12 12 4 12 0 6 and
t = 40 6 53 23 1

24 15 24 53 60 (using the regular cost functions, not
the diagonalized ones.) The average excess cost is now 21.43.
8.2. LINK INTERACTIONS 269

Iteration 3. The all-or-nothing assignment is x∗ = 6 0 0 6 0 and the

 
1
diagonalized link performance functions are t̂12 = 10x12 , t̂13 = 52 24 +
23
x13 , t̂23 = 10 24 + x23 , t̂24 = 53 + x24 , and t̂34 = 10x34 . The solution

to (8.31)
 is λ = 0.284, which gives x = 4.63 1.37 2.92 1.70 4.30 ,
t = 46.3 52.8 13.6 53.9 51.5 , so the average excess cost is 6.44.

and so on until convergence is reached.

Simplicial decomposition
An alternative to diagonalization is the simplicial decomposition algorithm.
This algorithm is introduced at this point (rather than in Chapter 6) for several
reasons. First, historically it was the first provably convergent algorithm for the
equilibrium problem with link interactions. Second, although it is an improve-
ment on Frank-Wolfe, for the basic TAP it is outperformed by the path-based
and bush-based algorithms presented in that chapter. However, like those algo-
rithms, it overcomes the “zig-zagging” difficulty that Frank-Wolfe runs into (cf.
Figure 6.4).
The price of this additional flexibility is that more computer memory is
needed. Frank-Wolfe and MSA are exceptionally economical in that they only
require two vectors to be stored: the current link flows x and the target link flows
x∗ . In simplicial decomposition, we will “remember” all of the target link flows
found in earlier iterations, and exploit this longer-term memory by allowing
“combination” moves towards several of these previous targets simultaneously.
In the algorithm, the set X is used to store all target link flows found thus far.
A second notion in simplicial decomposition is that of a “restricted equilib-
rium.” Given a set X = {x∗1 , x∗2 , · · · , x∗k } and a current link flow solution x, we
say that x is a restricted equilibrium if it solves the variational inequality

t(x) · (x − x0 ) ≤ 0 ∀x0 ∈ X(x, X ) (8.32)

where X(x, X ) means the set of link flow vectors which are obtained by a convex
combination of x and any of the target vectors in X .3 Equivalently, x is a
restricted equilibrium if none of the targets in X lead to improving directions
in the sense that the total system travel time would be reduced by moving to
some x∗i ∈ X while fixing the travel times at their current values. That is,

t(x) · (x − x∗i ) ≤ 0 ∀x∗i ∈ X (8.33)

At a high level, simplicial decomposition works by iterating between adding

new target vectors to X , and then finding a restricted equilibrium using the
current vectors in X . In this light, Frank-Wolfe can be seen as a special case
of simplicial decomposition, where X only consists of the current target vector
(forgetting any from past iterations), because if there is only one target x∗ ∈ X
3 A flow vector x0 is a convex combination of x and the target vectors in X if there exist

nonnegative constants λ0 , λ1 , · · · , λk such that x0 = λ0 x + λ1 x∗1 + · · · λk x∗k .

270 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

then the restricted equilibrium in X(x, X ) is simply the restricted equilibrium

along the line segment connecting x to x∗ .
In practice, it is too expensive to exactly find a restricted equilibrium at each
iteration. Instead, several “inner iteration” steps are taken to move towards a
restricted equilibrium with the current set X before looking to add another
target. In each inner iteration, the current solution x is adjusted to x + µ∆x,
where µ is a step size and ∆x is a direction which moves toward restricted
equilibrium. Smith (1984) shows that one good choice for this direction is
+
[t(x) · (x − x∗i )] (x∗i − x)
P
x∗
i ∈X
∆x = P ∗ +
(8.34)
x∗ ∈X [t(x) · (x − xi )]
i

This rather intimidating-looking formula is actually quite simple. It is nothing

more than a weighted average of the directions x∗i − x (potential moves toward
each target in X ), where the weight for each potential direction is the extent to
which it improves upon the current solution: [t(x) · (x − x∗i )] is the reduction
in total system travel time obtained by moving from x to x∗i while holding
travel times constant. If this term is negative, there is no need to move in that
direction, so the weight is simply set to zero. The denominator is simply the
sum of the weights, which serves as a normalizing factor.
The step size µ is chosen through trial-and-error. One potential strategy is
to iteratively test µ values in some sequence (say, 1, 1/2, 1/4, . . .) until we have
found a solution acceptably closer to restricted equilibrium than x.4 “Accept-
ably closer” can be calculated using the Smith gap
X
γS = ([t(x) · (x − x∗i )]+ )2 (8.35)
x∗
i ∈X

which is similar to the gap measures described in Chapter 6 in that it is zero if

and only if x is a restricted equilibrium, and positive otherwise. It reflects a re-
stricted equilibrium because the summation is only over the vectors in X : if any
of them reflect an “improvement” over the current solution, in that total travel
time would be reduced (if link travel times were held fixed), the corresponding
term in square brackets is positive, and γS > 0. It can be shown that taking a
small enough step in the direction (8.34) will reduce the Smith gap, and a line
search or trial-and-error method can be used to determine what an acceptable
step size is. Squaring the term in brackets ensures that γS is differentiable,
which plays a role in deriving the convergence rate of this method.
The Smith gap could thus be converted into a stopping criterion for the
(unrestricted) traffic assignment problem, by extending the sum to include every
possible all-or-nothing assignment, not just the ones in X , but this idea will not
be explored further in this book.
Putting all of this together, the simplicial decomposition algorithm can be
stated as:
4 Those familiar with nonlinear optimization may see parallels between this and the Armijo

rule.
8.2. LINK INTERACTIONS 271

1. Initialize the set X ← ∅

2. Find shortest paths for all OD pairs.
3. Form the all-or-nothing assignment x∗ based on shortest paths.
4. If x∗ is already in X , stop.
5. Add x∗ to X .
6. Subproblem: Find a restricted equilibrium x using only the vectors in
X.
(a) Find the improvement direction ∆x using equation (8.34).
(b) Update x ← x + µ∆x, with µ sufficiently small (to reduce γS ).
(c) Update travel times.
(d) Return to step 1 of subproblem unless γS is small enough.
7. Return to step 2.
Below we apply this algorithm to the example in Figure 8.5, choosing µ via
trial and error from the sequence 1/2, 1/4, . . . and stopping at the first value
that reduces γS . Other ways of choosing µ are also possible.
Iteration 1. We set
x∗1 = 6
 
0 6 0 6
and X = {x∗1 }. For the subproblem, the only possible solution is x = x∗1 ,
which has γS = 0 (it is trivially a restricted equilibrium) and travel times
are  
t = 60 53 16 53 60 .
Iteration 2. The new all-or-nothing assignment is
x∗2 = 0 6 0 0 6 ,
 

and X = {x∗1 , x∗2 }. For the first iteration of the subproblem, notice that
t · (x − x∗1 ) = 0 and t · (x − x∗2 ) = 138, so Smith’s formula (8.34) reduces
to
0 138 ∗
(x∗1 − x) +
 
∆x = (x2 − x) = −6 6 −6 0 0 .
138 138
Taking a step of size µ = 1/2 gives us
 
x= 3 3 3 0 6 .
The new travel times are
54 21 14 12
 
t = 30 53 60 ,
so t · x = 657, t · x∗1 = 627, and t · x∗2 = 687, so
γS = ([657 − 627]+ )2 + ([657 − 687]+ )2 = 302 + 02 = 900 .
Assume this is “small enough” to complete the subproblem.
272 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Iteration 3. The new all-or-nothing assignment is

x∗3 = 6 0 0 6 0
 

and X = {x∗1 , x∗2 , x∗3 }. For the first iteration of the subproblem, calculate
t · (x − x∗1 ) = 30, t · (x − x∗2 ) = −30, and t · (x − x∗3 ) = 159. So (8.34)
gives

30 ∗ 0 159 ∗
∆x = (x1 − x) + (x∗2 − x) + (x3 − x)
189 189 189 
= 3 −3 −2.05 5.05 −5.05 .

Taking a step of size µ = 1/2 would give

 
x = 4.5 1.5 1.98 2.52 3.48

and  
t = 45 52.5 12.7 54.3 47.4
which has γS = 1308. Assume that this is no longer “small enough” to
return to the master problem, so we begin a second subproblem iteration.
Smith’s formula (8.34) now gives ∆x = 0.2(x − x∗2 ) + 0.8(x − x∗3 ), and the
trial solution x + 21 ∆x has γS = 1195, which is an improvement. In this
case, choosing a smaller µ would work even better; for instance µ = 1/4
would reduce the Smith gap to 557. There is thus a tradeoff between
spending more time on finding the “best” value of µ, or spending more
time on finding new search directions and vectors for X . Balancing these
is an important question for implementation.

The algorithm can continue from this point or terminate if this average
excess cost is small enough.

8.3 Stochastic User Equilibrium

This section describes another extension to the basic TAP. To this point in the
text, we have been using the principle of equilibrium to determine link and path
flows, requiring all used paths between the same origin and destination to have
equal and minimal travel time. We derived this principle by assuming that all
drivers choose the least-travel time path between their origin and destination.
However, this assumption implicitly requires drivers to have perfect knowledge
of the travel times on all routes in the network. In reality, we know this is not
true: do you know the travel times on literally all routes between an origin and
destination? And can you accurately distinguish between a route with a travel
time of 16 minutes, and one with a travel time of 15 minutes and 59 seconds?
Relaxing these assumptions leads us to the stochastic user equilibrium (SUE)
model.
In SUE, rather than requiring that each driver follow the true shortest path
between each origin and destination, we assume that drivers follow the path
8.3. STOCHASTIC USER EQUILIBRIUM 273

they believe to be shortest, but allow for some perception error between their
belief and the actual travel times. An alternative, mathematically equivalent,
interpretation (explained below) is that drivers do in fact perceive travel times
accurately, but care about factors other than travel time. This section explains
the development of the SUE model. The mathematical foundation for the SUE
model is in discrete choice concepts, which are briefly reviewed in Section 8.3.1.
The specific application of discrete choice to the route choice decision is taken
up in Section 8.3.2.
These sections address the “individual” perspective of logit route choice.
The next steps to creating the SUE model are an efficient network loading
model (a way to find the choices of all drivers efficiently), and then finally the
equilibrium model which combines the network loading with updates to travel
times, to account for the mutual dependence between travel times and route
choices. These are undertaken in Sections 8.3.4 and 8.3.5, respectively. For the
most part, this discussion assumes a relatively simplistic model for perception
errors in travel times; Section 8.3.7 briefly discusses how more general situations
can be handled.

8.3.1 Discrete choice modeling

This section provides a brief overview of discrete choice concepts. The applica-
tion to route choice is in the following section. Discrete choice is a large area of
scholarly inquiry in and of itself, and so the discussion is restricted to what is
needed in the chapter.
Consider an individual who must make a choice from a set of options. For
instance, when purchasing groceries, you must choose one store from a set of
alternatives (the grocery stores in your city). When dressing in the morning,
you must choose one set of clothes among all of the clothing you own. And, more
relevant to transportation, when choosing to drive from one point to another,
you must choose one route among all of the possible routes connecting your
origin to your destination.
Mathematically, let I be a finite set of alternatives. Each alternative i ∈ I
is associated with a utility value Ui representing the amount of happiness or
satisfaction you would have if you were to choose option i. We assume that you
would choose an alternative i∗ ∈ arg maxi {Ui } which maximizes the utility you
receive. Now, the utility Ui consists of two parts: an observable utility Vi , and
an unobservable utility denoted by the random variable :

Ui = Vi + i (8.36)

The difference between observable and unobservable utility can be explained

in different ways. One interpretation is that Vi represents the portion of the
utility that is due to objective factors visible to the modeler; when choosing
a grocery store, that might include the distance from your home, the price,
the variety of items stocked, etc. The unobserved utility consists of subjective
factors that the modeler cannot see (or chooses not to include in the model),
274 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

even though they are real insofar as they affect your choice. In the grocery store
example, this might include your opinion on the taste of the store brands, the
cleanliness of the store, and so on. Then, by modeling the unobserved utility
as a random variable , we can express choices in terms of probabilities. (The
modeler does not know all of the factors affecting your choice, so they can only
speak of probabilities of choosing different options based on what is observable.)
A second interpretation is that the observed utility Vi actually represents
all of the factors that you care about. However, for various reasons you are
incapable of knowing all of these reasons with complete accuracy. (You probably
have a general sense of the prices of items at a grocery store, but very few know
the exact price of every item in a store’s inventory.) Then the random variable
i represents the error between the true utility (Vi ) and what you believe the
utility to be (Ui ). Either interpretation leads to the same mathematical model.
Depending on the distribution we choose for the random variables i , differ-
ent discrete models are obtained. A classic is the logit model, which is obtained
when the unobserved utilities i are assumed to be independent across alterna-
tives, and to have Gumbel distributions with zero mean and identical variance.
Under this assumption, the probability of choosing alternative i is given by

exp(θVi )
pi = P (8.37)
j∈I exp(θVj )

where θ is a nonnegative parameter reflecting the relative magnitude of the

observed utility relative to the unobserved utility. Notice what happens in this
formula as θ takes extreme values: if θ = 0, then all terms in the numerator and
denominator are 1, and the probability of choosing any alternative is exactly the
same. (Interpretation: the unobserved utility  is much more important than
the observed utility, so the observed utility has no impact on the decision made.
Since the unobserved utility has the same distribution for every alternative, each
is equally likely.) Or, if θ grows large, then the denominator of (8.37) will be
dominated by whichever terms have the largest observed utility Vi . If there is
some alternative i∗ for which Vi∗ > Vi for all i 6= i∗ (the observed utility for i∗ is
strictly greater than any other alternative), then as θ → ∞, the probability of
choosing i∗ approaches 1 and the probability of choosing any other alternative
approaches 0. Another important consequence of (8.37) is that the probability
of choosing any alternative is always strictly positive; there is some chance that
the unobserved utility will be large enough that any option could be chosen.
While the logit model is nice in that we have a closed-form expression (8.37)
for the probabilities of choosing any alternative, the logit assumptions are very
strong — particularly the assumption that the i are independent and identi-
cally distributed. The exercises at the end of the chapter explore some examples
demonstrating how these assumptions can lead to unreasonable results. Another
common assumption is that the  are drawn from a multivariate normal distri-
bution, which allows for correlation among the unobserved utilities for different
alternatives. This leads to the probit choice model, which is more flexible and
arguably realistic. However, unlike the logit model, the probit model does not
8.3. STOCHASTIC USER EQUILIBRIUM 275

have a closed-form expression for probabilities like (8.37). Instead, Monte Carlo
sampling methods are used to estimate choices.
The majority of this section is focused on logit-based models. While probit
models are more general and arguably more realistic, the logit model has two
major advantages from the perspective of a book like this. First, computations
in logit models can often be done analytically, simplifying explanations and
making it possible to give examples you can easily verify. This helps you better
understand the main ideas in stochastic user equilibrium and build intuition.
Second, logit models admit faster solution algorithms, algorithms which scale
relatively well with network size. This is an important practical advantage for
logit models. Nevertheless, Section 8.3.7 provides some discussion on probit and
other models and what needs to change from the logit discussion below.

8.3.2 Logit route choice

This section specializes the discrete choice framework from the previous section
to route choice in networks. Consider a traveler leaving origin r for destination
s. They must choose one of the paths π connecting r to s, that is, they must
make a choice from the set Πrs . The most straightforward way to generalize
the principle of user equilibrium to account for perception errors is to set the
observed utility equal to the negative of path travel time, so

U π = −cπ + π (8.38)

with the negative sign indicating that maximizing utility for drivers means min-
imizing travel time. Assuming that the π are independent, identically dis-
tributed Gumbel random variables, we can use the logit formula (8.37) to ex-
press the probability that path π is chosen:
exp(−θcπ )
pπ = P (8.39)
π 0 ∈Πrs exp(−θcπ )
0

The comments in the previous section apply to the interpretation of this formula.
As θ approaches 0, drivers’ perception errors are large relative to the path travel
times, and each path is chosen with nearly equal probability. (The errors are
so large, the choice is essentially random.) As θ grows large, perception errors
are small relative to path travel times, and the path with lowest travel time is
chosen with higher and higher probability. At any level of θ, there is a strictly
positive probability that each path will be taken.
For concreteness, the route choice discussion so far corresponds to the second
interpretation of SUE, where the unobserved utility represents perception errors
in the utility. The first interpretation would mean that π represents factors
other than travel time which affect route choice (such as comfort, quality of
scenery, etc.). Either of these interpretations is mathematically consistent with
the discussion in Section 8.3.1.
The fact that the denominator of (8.39) includes a summation over all paths
connecting r to s is problematic, from a computation standpoint. The number
276 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

of paths can grow exponentially with network size. Any use of stochastic user
equilibrium in a practical setting, therefore, requires a way to compute link flows
without explicitly calculating the sum in (8.39).
This is done by carefully defining which paths are in the choice set for trav-
elers. The notation Πrs in (8.39) in this book means the set of all acyclic
paths connecting origin r to destination s. Following Chapter 6, we will use the
notation Π̂rs to define the set of paths being considered by travelers in SUE;
these sets are sometimes called sets of reasonable paths. With a suitable defini-
tion of this set, using Π̂rs in place of Πrs in equation (8.39) leads to tractable
computation schemes.
Two possibilities are common: selecting an acyclic subset of links, and choos-
ing Π̂rs to contain the paths using these links only; or setting Π̂rs to consist of
literally all paths (even cyclic ones) connecting origin r to destination s. Both
of these are discussed next. The key to both of these definitions of Π̂rs is that
we can determine how many of the travelers passing through a given node came
from each of the available incoming links, without needing to know the specific
path they are on. This is known as the Markov property, and is discussed at
more length in the optional Section 8.3.3.

Totally acyclic paths

For a particular origin r and destination s, choose a set of allowable links Ârs ,
and let Π̂rs consist of all paths starting at r, ending at s, and only containing
links from Ârs . We require two conditions on the set of allowable links:

1. There is at least one path from r to s using allowable links; this ensures
that Π̂rs is nonempty.

2. The set of allowable links contains no cycle; this ensures that all paths in
Π̂rs are acyclic.

We say a set of paths Π̂rs is totally acyclic if it can be generated from an

allowable link set satisfying these conditions.
This definition is closely related to the idea of a bush from Section 5.2.3.
If Ârs contains a path from r to every destination, it is also a bush; and if
furthermore the sets Ârs are the same for all destinations s, we can do the
network loading for all the travelers leaving origin r simultaneously, rather than
separately for each destination.
Note that there are collections of acyclic paths which are not totally acyclic.
In the network in Figure 8.6, if we choose Π̂14 = {[1, 2, 3, 4], [1, 3, 2, 4]}, both
paths in this set are acyclic, but the set of allowable links needs to include every
link in the network:

Â14 = {(1, 2), (1, 3), (2, 3), (2, 4), (3, 2), (3, 4)}

This set contains the cycle [2, 3, 2], so it is not possible to generate a reasonable
path set containing [1, 2, 3, 4] and [1, 3, 2, 4] from an acyclic set of allowable links.
8.3. STOCHASTIC USER EQUILIBRIUM 277

1 4

Figure 8.6: A set of acyclic paths need not be totally acyclic.

The advantage of totally acyclic path sets is that we can define a topolog-
ical order on the nodes (see Section 2.2). With this topological order, we can
efficiently make computations using the logit formula (8.39) without having to
enumerate all the paths. This procedure is described in the next section.
We next describe two ways to form sets of totally acyclic paths. For each
link, define a positive value c0ij for each link which is constant and independent
of flow — examples include the free-flow travel time or distance on the link. For
each origin r and node i, let Lri denote the length of the shortest path from r
to i, using the quantities c0ij as the link costs. Likewise, for each destination s
and node i, let `si denote the length of the shortest path from i to s, again using
the quantities c0ij as the link costs.
Consider the following sets of paths:
1. The set of all paths, for which the head node of each link is further away
from the origin than the tail node, based on the quantities c0ij . That is,
the sets Π̂r containing all paths starting at r and satisfying Lrj > Lri for
each link (i, j) in the path.
2. The set of all paths, for which the head node of each link is closer to
the destination than the tail node, based on the quantities c0ij . That is,
the sets Π̂s containing all paths ending at s satisfying `sj < `si for each
link (i, j) in the path. (This is like the first one, but oriented toward the
destination, rather than the origin.)
3. The set of paths which satisfy both of the above criteria: Π̂rs = Π̂r ∩ Π̂s .
For instance, if c0ij reflects the physical length on each link, then for a given
origin, Π̂r would consist of all of the paths which start at that origin and al-
ways move away from it, never “doubling back.” Likewise, Π̂s would consist
of all paths which always move closer to their destination node s, without any
diversions that lead it away. The third option has paths which both continually
move away from their origin and toward their destination. Exercise 15 asks you
to show that all three possibilities for Π̂ are totally acyclic.
Figure 8.7 illustrates these three definitions. The top of the figure shows
a network with node A as origin and node F as destination, and the links
278 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

3 1 2 1 3
D E F
2 1 0

3 1 1 3

0 1 1 1 2
A B C
3 2 3

D E F D E F D E F

A B C A B C A B C

Figure 8.7: Demonstration of three alternative definitions of totally acyclic

paths. Top panel shows c0ij values on links, shortest path distances Li from
A above nodes, and shortest path distances `i to F below nodes.

are labeled with their c0ij values. The nodes are labeled with their Li values
(above each node) and `i values below. The bottom of the figure shows the
links satisfying each of the three criteria (Lj > Li ; `j < `i ; and both of these
simultaneously). The paths in these networks are the allowable paths in the
original network. Notice that in all cases, there are no cycles in these links
(even though the original network had the cycle [2,5,2]).
Of these principles, the third imposes stricter conditions on which paths
are in the reasonable set. The first and second are weaker, and includes some
paths which may not seem reasonable to you. For instance, the spiral path
in Figure 8.8 satisfies the first condition, since the distance from the origin is
always increasing. However, it does not satisfy the third condition, since at
times the distance to the destination increases as well.
However, a major advantage of the first two principles is that we can aggre-
gate travelers by origin or destination. With the first principle, the destination
of travelers can be ignored for routing purposes — if a path is reasonable for a
travel from an origin r to a node i, that path segment is reasonable for travel
to any node beyond i as well. This allows us to aggregate travelers by origin
(as in Section 5.2.3) and calculate a “one-to-all” path set for each origin, rather
than having separate path sets for each OD pair. A similar destination-based
aggregation is possible with the second principle.
8.3. STOCHASTIC USER EQUILIBRIUM 279

Figure 8.8: Is this path reasonable?

Full cyclic path set

Instead of restricting the path set to create a totally acyclic collection, an alter-
native is to have Π̂rs consist of literally all paths from origin r to destination
s, even including cycles. This will often mean these sets are infinite. For exam-
ple, consider the network in Figure 8.6. Under this definition, the (cyclic) path
[1, 2, 3, 2, 4] is part of Π̂rs , as is [1, 2, 3, 2, 3, 2, 4], and so on.
Including cyclic paths, especially paths with arbitrarily many repetitions of
cycles, may seem counterintuitive. There are several reasons why this definition
of Π̂rs is nevertheless useful. One reason is that requiring total acyclicity is in
fact quite a strong condition. In Figure 8.6, there is no totally acyclic path set
that includes both [1, 2, 3, 4] and [1, 3, 2, 4] as paths — if we want to allow one
path as reasonable, then by symmetry the other should be reasonable as well.
But any path set including both of those includes both links (2, 3) and (3, 2),
which form a cycle.
So, it is desirable to have an alternative to total acyclicity that still does
not require path enumeration. As shown in the following section, it is possible
to compute the link flows resulting from (8.39) without having to list all the
paths, if all cyclic paths are included. The intuition is that all travelers at a
given node can be treated identically in terms of which link they move to next:
in Figure 8.6, we can split the vehicles arriving at node 2 between links (2, 3)
and (2, 4) without having to distinguish whether they came via link (1,2), as if
on the path [1,2,4] or [1,2,3,4], or whether they came via link (3,2), as if on the
path [1,3,2,4], or even [1,2,3,2,4].
Furthermore, some modelers are philosophically uncomfortable with includ-
ing restrictions like those in the previous section, without evidence that those
rules really represent traveler behavior. Determining which sets of paths travel-
ers actually consider (and why) is complicated, and still not well-understood.5
One school of thought is that it is therefore better to impose no restrictions at
all, essentially taking an “agnostic” position with respect to the sets Π̂rs , rather
than imposing restrictions which may not actually represent real behavior.
5 Emerging data sources, such as Bluetooth readers, are providing more complete informa-

tion on observed vehicle trajectories. This may provide more insight on this subject.
280 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

3
1/2 1/2

1 1/2 1/2 4

1/2 1/2
2

Figure 8.9: Flows from stochastic network loading when all links have unit cost,
θ = log 2, and the full cyclic path set is allowed.

As an example, assume that every link in Figure 8.6 has the same travel
time of 1 unit, and that θ = log 2. Then there are two paths of length 2 ([1,2,4]
and [1,3,4]), two paths of length 3 ([1,2,3,4] and [1,3,2,4]), two paths of length
4 ([1,2,3,2,4] and [1,3,2,3,4]), and so on. Therefore the denominator in the logit
formula is
X 1 1 1 1 1 1
exp(−θcπ ) = + + + + + + · · · = 2 (8.40)
2 2 4 4 8 8
π∈Π̂

and the probability of choosing one of the length-2 paths is (1/2)/2 = 1/4, the
probability of choosing one of the length-3 paths is 1/8, and so on. The flows
on each link can be calculated by multiplying these path flows by the number
of times that path uses a link. For instance, to calculate the flow on link (1,2),
observe that it is used by paths [1,2,4], [1,2,3,4], [1,2,3,2,4], [1,2,3,2,3,4], and so
on, with respective probabilities 1/4, 1/8, 1/16, 1/32, etc. Thus the total flow
on this link is the sum of these, or 1/2. The flows on links (1,3), (2,4), and (3,4)
are also found to be 1/2 by the same technique. Calculating the flow on links
(2,3) and (3,2) is trickier, because some paths use these multiple times. For
example, path [1,2,3,2,3,4] uses link (2,3) twice, so even though the probability
of selecting this path is 1/25 = 1/32, it actually contributes twice this (1/16) to
the flow. It is possible to show that the flow on these links is also 1/2, giving
the final flows in Figure 8.9.
As this example shows, direct calculations involving this path set usually
involve summing infinite series. As will show in Section 8.3.4, there is an al-
ternative method that allows us to make these computations without explicitly
calculating such sums.

8.3.3 (*) The Markov property and the logit formula

(This optional section gives mathematical reasons why totally acyclic path sets,
and complete path sets, both allow for efficient calculation of the logit formula.)
8.3. STOCHASTIC USER EQUILIBRIUM 281

1 2 3

4 5

6
All links have cost 1 d13 = 40

Figure 8.10: Network for demonstrating the Markov property.

Both path set definitions above (“totally acyclic paths” and the “full cyclic
path set”) allow for the computation of the logit formula to be disaggregated
by node and by link, without having to enumerate the paths in the network.
The key to this is the Markov property. An informal statement of this property
is that if we randomly select a traveler passing through a node, and want to
know the probability that they leave that node by a particular link, there is no
information provided by knowing which link they used to arrive to that node.
For example, consider the network in Figure 8.10, where the demand is
d13 = 40 vehicles and θ = log 2. All links have unit cost. Assume first that all
paths in this network are allowed. Then the right panel of Figure 8.11 shows
the flow on each path, and the left panel shows the flow on each link.
In this network 18 vehicles pass through node 2. Suppose we pick one of
them at random, and want to know the probability that the next link in this
vehicle’s path is (2,3), as opposed to (2,5). From examining the path flows in
Figure 8.11, we can see that this probability is (8 + 4)/(8 + 4 + 4 + 2) = 2/3
(ignoring the flow on path [1,4,6,5,3] in the denominator, since these trips do
not pass through node 2). Now, suppose that these vehicles also reported the
segment of their path that led them to node 2 — that is, they also report
whether they came via segment [1,2] or segment [1,4,2]. Does this change our
answers in any way?
If we know they came from segment [1,2], then they are either on path [1,2,3]
or [1,2,5,3], and the probability that they continue on (2,3) is 8/(8 +4) = 2/3. If
we know they came from segment [1,4,2], then they are either on path [1,4,2,3]
or path [1,4,2,5,3], and the probability that they continue on (2,3) is 4/(4 + 2),
which is still 2/3. So knowing the first segment of their trip does not provide
any additional information as to the remaining segment.
The situation changes if we modify the allowable path set to only include
three paths, [1,2,3], [1,2,5,3], and [1,4,2,3].6 Here Figure 8.12 shows the corre-
sponding path flows and link flows.
Let us ask the same question of the travelers passing through node 2. With-
6 A natural way this path set might arise is to include paths that are only within a small

threshold of the shortest path cost, thus including paths with cost 2 or 3 but excluding paths
with cost 4.
282 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

12 12
1 2 3 Path Pr Flow
[1,2,3] 0.4 8
8 6 6 8 [1,2,5,3] 0.2 4
5 [1,4,2,3] 0.2 4
4
[1,4,2,5,3] 0.1 2
2 2 [1,4,6,5,3] 0.1 2

Figure 8.11: Link and path flows when all paths are allowed.

15 15
1 2 3 Path Pr Flow
[1,2,3] 1/2 10
5 5 5 5 [1,2,5,3] 1/4 5
5 [1,4,2,3] 1/4 5
4
0 0
6

Figure 8.12: Link and path flows when only a subset of paths is allowed.
8.3. STOCHASTIC USER EQUILIBRIUM 283

out knowing anything further, the probability that they continue on link (2,3)
is (10 + 5)/20 = 3/4. However, if we know they came from [1,2], then the prob-
ability that they continue on (2,3) is 10/15 = 2/3. If we know they came from
[1,4,2], then the probability that they continue on (2,3) is 1, and there is no
other option! So in this case, knowing the first segment of the path does give
us additional information about the rest of their journey.
It turns out that the Markov property will be very useful, and will allow us
to efficiently evaluate the logit formula without enumerating paths. Informally,
we can do computations using just the link flows (the left panels in Figures 8.11
and 8.12) without having to use the path flows (the right panels of these figures)
— we can get the path flows on the right from the link flows on the left. In large
networks, the link-based representation is much more compact and efficient.
A more formal statement of this property is as follows. To keep the formulas
clean, assume that there is a single origin r and destination s; in a general
network, we can apply the same logic separately to each OD pair. In logit
assignment, the path set Π̂ is said to satisfy the Markov property if there exist
values Pij for each link such that

exp(−θcπ ) Y π
δij
pπ = P π 0 = P ij (8.41)
π 0 ∈Π̂ exp(−θc ) (i,j)∈π

π
where δij is the number of times path π uses link (i, j).
That is, that the probability of a traveler selecting any path can be computed
by multiplying Pij values across its links. The Pij values can be interpreted as
conditional probabilities: given that a traveler is passing through i, they express
the probability that their path leaves that node through link (i, j).

The segment substitution property

To make the connection between totally acyclic and complete path sets and the
Markov property, we first show that both path set choices satisfy the segment
substitution property. Given any path π = [r, i1 , i2 , . . . , s], a segment σ is any
contiguous subset of one or more of its nodes. For example, the path [1, 2, 3, 2, 4]
contains [1, 2, 3], [2, 3], [2], and [2, 3, 2, 4]. It does not contain the segment
[1, 4]; even though both of those nodes are in the path, they do not appear
consecutively. Note that a segment can consist of a single link, such as [2, 3],
or even a single node, such as [2]. We use the notation ⊕ to indicate joining
segments, so [1, 2, 3, 2, 4] = [1, 2, 3] ⊕ [3, 2, 4]. If two segments are being joined,
the end node of the first must match the starting node of the second.
The set of reasonable paths Π̂ satisfies the segment substitution property
if, for any pair of reasonable paths which pass through the same two nodes,
the paths formed by exchanging the segments between those nodes are also
reasonable. That is, if π = σ1 ⊕ σ2 ⊕ σ3 is reasonable, and if there is another
reasonable path π 0 = σ10 ⊕ σ20 ⊕ σ30 with σ2 and σ20 starting and ending at the
same node, then the paths σ1 ⊕ σ20 ⊕ σ3 and σ10 ⊕ σ2 ⊕ σ30 are also reasonable.
284 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

With the allowable path set in Figure 8.12, the segment substitution property
is not satisfied. There are paths [1,2,5,3] and [1,4,2,3], both passing through
nodes 1 and 2, which can be decomposed in the following way:

[1, 2, 5, 3] = [1] ⊕ [1, 2] ⊕ [2, 5, 3] (8.42)

[1, 4, 2, 3] = [1] ⊕ [1, 4, 2] ⊕ [2, 3] (8.43)

Notice that the corresponding pairs of segments on the right-hand sides all start
and end at the same nodes. We can generate two new paths by “crossing” the
middle segments of (8.42) and (8.43): [1,2,3] and [1,4,2,5,3]. The first of these
is allowable, but the second is not.
By contrast, you can verify that segment substitution is satisfied for the
allowable path set in Figure 8.11. No matter which pairs of paths you choose,
swapping the segments results in another allowable path.
A reasonable path set cannot have the Markov property unless it satisfies
the segment substitution property, as shown in the above example. Without the
segment substitution property, we could not know how the vehicles at node i
would split without knowing the specific paths they were on. Segment substitu-
tion ensures that all travelers passing through node i are considering the same
set of outgoing links (and, indeed, the same set of path segments continuing on
to the destination).
It is fairly easy to show that both path set definitions considered above —
sets of totally acyclic paths, and the set of all paths (even cyclic ones) — satisfy
the segment substitution property; see Exercise 16.

Decomposing the logit formula

Given a reasonable path set Π̂ satisfying the segment substitution property, and
any two nodes a and b, let Σab denote the set of segments which start and end
at these nodes, and are part of a reasonable path. Define the quantity
X
Vab = exp(−θcσ ) (8.44)
σ∈Σab

where cσ = (i,j)∈σ cij is the travel time of a segment. In particular, Vrs is a

P
sum over all of the reasonable paths from origin to destination, and corresponds
to the denominator of the logit formula. Thus, the probability that a traveler
chooses a particular path π is simply exp(−θcπ )/Vrs , and the flow on this path
is drs exp(−θcπ )/Vrs . Also note the special case Vii where the start and end
nodes are the same. In this case Σii consists only of the single-node segment [i]
with zero cost, so Vii = 1.
Furthermore, note that the numerator of the logit formula can be factored
by segment, so that if π = σ1 ⊕ σ2 ⊕ σ3 , we have

exp(−θcπ ) = exp(−θ(cσ1 + cσ2 + cσ3 )) = exp(−θcσ1 ) exp(−θcσ2 ) exp(−θcσ3 ) .

(8.45)
8.3. STOCHASTIC USER EQUILIBRIUM 285

To calculate the flow on a link xij , we need to add the flow from all of the
reasonable paths which use this link. Every reasonable path using link (i, j)
takes the form σ1 ⊕ [i, j] ⊕ σ2 , where σ1 goes from the origin r to node i and σ2
goes from node j to the destination s.7 Therefore

X
π π
xij = δij h (8.46)
π∈Π̂rs
drs X π
= δij exp(−θcπ ) (8.47)
Vrs rs
π∈Π̂
drs X X
= exp(−θ(cσ1 + tij + cσ2 )) (8.48)
Vrs
σ1 ∈Σri σ2 ∈Σjs
drs X X
= exp(−θcσ1 ) exp(−θtij ) exp(−θcσ2 ) (8.49)
Vrs
σ1 ∈Σri σ2 ∈Σjs
!  
drs X X
= exp(−θcσ1 ) exp(−θtij )  exp(−θcσ2 ) (8.50)
Vrs
σ1 ∈Σri σ2 ∈Σjs

Vri exp(−θtij )Vjs

= drs (8.51)
Vrs

The third equality groups the sum over paths according to the starting and
ending segment. This equation for xij will be used extensively in the rest of
this section.
Similarly, to find the number of vehicles passing through a particular node i
(call this xi ), observe that every path through i can be divided into a segment
from r to i, and a segment from i to s. Grouping the paths according to these
segments and repeating the algebraic manipulations above gives the formula

Vri Vis
xi = drs (8.52)
Vrs

With equations (8.51) and (8.52) in hand, it is easy to show the Markov prop-
erty holds in any reasonable path set with the segment substitution property.
Define
xij exp(−θtij )Vjs
Pij = = (8.53)
xi Vis

and then multiply these values together for the links in a path, say, π =

7 If the link starts at the origin or ends at the destination, we may have i = r or j = s, in

which case σ1 or σ2 will consist of a single node, [r] or [s].

286 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

[r, i1 , i2 , · · · , ik , s]:
Y Y exp(−θtij )Vjs
Pij = (8.54)
Vis
(i,j)∈π
 
X Vi1 s Vi2 s · · · Vik s Vss
= exp −θ tij  (8.55)
Vrs Vi1 s Vi2 s · · · Vik s
(i,j)∈π
Vss
= exp(−θcπ ) (8.56)
Vrs

But Vss = 1 and Vrs is simply the denominator in the logit formula, so this
product is exactly pπ . Therefore the logit path flow assignment in the set of
reasonable paths satisfies the Markov property.

8.3.4 Stochastic network loading

The stochastic network loading problem is to determine the flows on each link
xij , given their travel times tij , according to a particular discrete choice model.
This section describes how to do so for the logit model, using formula (8.39).
For stochastic network loading, we assume that these travel times are fixed and
constant, and therefore unaffected by the path and link flows we calculate.
In the larger stochastic user equilibrium problem (where travel times can de-
pend on flows), stochastic network loading plays the role of a subproblem in an
iterative algorithm. This is analogous to how shortest paths and all-or-nothing
loadings are often used as a subproblem in the classical traffic assignment prob-
lem: although travel times do depend on link flows, to find an equilibrium we
can solve a number of shortest path problems, temporarily fixing the link costs
at particular values.
Throughout this section, we are assuming a single origin r and destination
s to simplify the notation. If there are many origins and destinations, these
procedures should be repeated for each, and the link flows added to obtain the
total link flows. (There is no harm in doing so, since we are assuming link
travel times are constant, and therefore the different OD pairs do not interact
with each other.) Depending on the choice of path set, it may be possible to
aggregate all travelers from the same origin or destination, and load them at
once. This is more efficient than doing a separate loading for each OD pair, but
requires a more limited definition of the reasonable path sets.
In principle, the stochastic network loading procedure is straightforward.
Given the link travel times tij , we can calculate the path travel times cπ as
in Section 4.1. The path flows hπ can then be calculated from the logit for-
mula (8.39), from which the link flows can be calculated by addition, again as
in Section 4.1.
While conceptually straightforward, this procedure faces the practical diffi-
culty that the logit formula requires summations over the set of all reasonable
paths Π̂rs , which can grow exponentially with network size. In a realistic-sized
8.3. STOCHASTIC USER EQUILIBRIUM 287

network, this renders the above procedure unusable, or at least computationally

taxing.
This section describes how the link flows can be calculated without explicitly
enumerating paths or using the logit formula. This is possible for both of the
Π̂ definitions discussed in the previous section: a totally acyclic set of paths, or
including the full path set, even including all cycles.
These two procedures have several features in common. As was described in
Section 8.3.3, both of these path set definitions have the Markov property. This
section derived an important formula, repeated here:

Vri exp(−θtij )Vjs

xij = drs , (8.57)
Vrs

where Vab is the sum of exp(−θcσ ) for all reasonable path segments σ starting
at node a and ending at node b. This formula is important because it allows us
to calculate the flow on each link without having to enumerate all of the paths
that use that link.
It is also possible to show (see Exercise 19) that we can replace tij with
Li + tij − Lj , where L is a vector of node-specific constants in this formula;
this reduces numerical errors in computations. It is common to use shortest
path distances at free-flow (hence the use of the notation L). Doing this for the
link and segment costs helps avoid numerical issues involved with calculating
exponentials of large values. With this re-scaling, we define the link likelihood
as
Lij = exp(θ(Lj − Li − tij )) (8.58)
and thus
Vri Lij Vjs
xij = drs . (8.59)
Vrs
It remains to describe how the Vab values can be efficiently calculated. This
section shows how this can be done for both totally acyclic path sets, and the
complete set of all paths (even cyclic ones).
In the case of a totally acyclic path set, the relevant V values can be calcu-
lated in a single pass over the network in topological order, and then the link
flows calculated in a second pass over the network in reverse topological order.
In the case of the full cyclic path set, the cycles create dependencies in the link
weights and in the link flow formulas which prevent them from being directly
evaluated. But we can still calculate them explicitly using matrix techniques.

Totally acyclic paths

The defining feature of a totally acyclic path set is that the collection of links
used by reasonable paths has no cycles. This allows us to define a topological
order on the nodes, so that each allowable link connects a lower-numbered node
to a higher-numbered one. In acyclic networks, it is often easy to perform
calculations recursively, in increasing or decreasing topological order. Stochastic
network loading is one of these cases, using Dial’s method.
288 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Define Wi and Wij to be the weight of a node and link, respectively. The
node weight Wi is a shorthand for Vri , and Wij is a shorthand for Vri Lij . These
can be calculated recursively, using the following procedure:
1. Calculate the link likelihoods Lij using equation (8.58) for all allowable
links; set Lij = 0 for any link not in the allowable set.
2. Set the current node i to be the origin r, and initialize its weight: Wr = 1.
3. For all links (i, j) leaving node i, set Wij = Wi Lij .
4. If i is the destination, stop. Otherwise, set i to be the next node in
topological order.
P
5. Calculate Wi = (h,i)∈Γ−1 (i) Whi by summing the weights of the incoming
links.
6. Return to step 3.
With the node and link weights in hand, we proceed to calculate the flows
on each link xij and the flow through each node xi . Using the link weights, we
can rewrite equation (8.52) for the flow through each node as

Vri Vis Wi Vis

xi = drs = drs . (8.60)
Vrs Ws
Combining with equation (8.59) for link flow, we have

Vrh Lhi Vis Whi

xhi = drs = xi (8.61)
Vrs Wi
so if we know the flows to some node i, we can calculate the flows to its incoming
links (h, i). So, the link flows can be calculated in reverse topological order:
1. Initialize all flows to zero: xi = 0 for all nodes, and xij = 0 for all links.
2. Set the current node i to be the destination s, and initialize its flow:
xi = drs , since all vehicles must reach the destination.
3. For all links (h, i) entering node i, set xhi = xi Whi /Wi .
4. If i is the origin, stop. Otherwise, set i to be the previous node in topo-
logical order.
P
5. Compute xi = (i,j)∈Γ(i) xij as the sum of the flows on outgoing links.
6. Return to step 3.
As an example, Dial’s method is demonstrated on the network shown in
Figure 8.13, where the demand is 2368 vehicles from node 1 to node 5 and
θ = 1. For convenience, the results of the calculations in this example are
shown in Tables 8.1 and 8.2. As a preliminary step, we calculate the shortest
8.3. STOCHASTIC USER EQUILIBRIUM 289

path from node 1 to all other nodes at free-flow conditions (assuming x = 0)

using standard techniques. Details are omitted since the technique is familiar,
and the resulting shortest path labels Li are shown in Table 8.1.
Next, we must identify the allowable links. This example will adopt the first
principle from the previous section, where the allowable links (i, j) are those for
which Li < Lj . Every link is thus allowable except for (2,3), because this link
connects node 2 (with shortest path label L2 = 2) to node 3 (L3 = 1). Therefore,
link (2,3) will be excluded from the remaining steps of Dial’s method, since no
vehicles will use this link.
The next step is to calculate the link likelihoods Lij for the allowable links
(i, j) ∈ AB . So, L12 = exp(2 − 0 − 2) = 1, L35 = exp(3 − 1 − 3) = e−1 ≈ 0.368,
and so forth. Weights are now calculated in forward topological order. The
topological ordering for the reasonable bush is 1, 3, 2, 4, and 5 in that order.
Notice that the original network has a cycle [2, 3, 2] and so no topological order
can exist on the full network. But when restricted to the allowable set, the cycle
disappears and node 3 must come before node 2 topologically. For the origin,
W1 = 1 by definition. The weights on links leaving node 1 can now be calculated:
W12 = W1 L12 = 1, and W13 = 1 × 1 = 1. Proceeding to node 3 (the next in
topological order), W3 is calculated as the sum of the weights on incoming links:
W3 = W13 = 1 and thus W32 = 1 × 1 = 1 and W35 = 1 × e−1 ≈ 0.368. Node
2 is next in topological order, and its weight is the sum of the weights on its
incoming links: W2 = W12 + W32 = 2, so W24 = 2. Node 4 is next, and we have
W4 = 2 and W45 = 2. Finally, the weight of node 5 is W5 = W35 + W45 = 2.368.
Link and node flows are now calculated in reverse topological order, starting
with node 5 and then proceeding to nodes 4, 2, 3, and 1. For node 5, the node
flow X5 is simply the demand destined to this node (since there are no outgoing
links), so X5 = 2368. This flow is now distributed among the two incoming
links (3, 5) and (4, 5) in proportion to their weights. So, x35 = X5 W35 /(W35 +
W45 ) = 2368 × 0.368/(2 + 0.368) = 368 and x45 = 2000. Proceeding upstream,
the node flow at 4 is simply the flow on link (4, 5), the only outgoing link
(since no vehicles have node 4 as their destination), and X4 = 2000. There is
only one incoming link (2, 4), so x24 = 2000. (This follows trivially from the
formula in step 3c since W4 = W24 . Thus X2 = 2000. Since there are two
incoming reasonable links to node 2 with equal weight, they receive equal flow
(again, following from the formula in 3c), setting x32 = x12 = 1000. Then
X3 = x32 + x35 = 1368, and x13 = X3 = 1368. Finally, the node flow at the
origin 1 is X1 = 1368 + 1000 = 2368, as it should be.
Dial’s method is now complete, having calculated the flows on each link. This
method is completely consistent with (8.39), if we were to restrict attention to
the reasonable paths in the network. If we were to use this formula directly,
we would first enumerate the three reasonable paths [1, 3, 5], [1, 3, 2, 4, 5], and
[1, 2, 4, 5] and calculate their costs: C [1,3,5] = 4, C [1,3,2,4,5] = C [1,2,4,5] = 3.
Equation (8.39) then gives

e−4 e−3
p[1,3,5] = = 0.155 p[1,3,2,4,5] = = 0.422
e−4 + e−3 + e−3 e−4 + e−3 + e−3
290 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

3
1 + x/100 3 + x/100

1 1 + x/100 2 + x/100 5

4 x/100
2 + x/100 5
2 1 + x/100

Figure 8.13: Network for demonstration of Dial’s method.

Table 8.1: Node calculations in Dial’s method example.

Node i Li Wi Xi
1 0 1 2368
2 2 2 1368
3 1 1 2000
4 3 2 2000
5 3 2 + e−1 2368

e−3
p[1,2,4,5] = = 0.422
e−4 + e−3 + e−3
as the path choice proportions. Multiplying each of these by the total demand
(2368) gives path flows

h[1,3,5] = 368 h[1,3,2,4,5] = 1000 h[1,2,4,5] = 1000

As you can verify, these path flows correspond to the same link flows shown in
Table 8.2.

Table 8.2: Link calculations in Dial’s method example.

Link (i, j) Lij Wij xij
(1,2) 1 1 1000
(1,3) 1 1 1368
(2,3) 0 0 0
(2,4) 1 2 2000
(3,2) 1 1 1000
(3,5) e−1 e−1 368
(4,5) 1 2 2000
8.3. STOCHASTIC USER EQUILIBRIUM 291

Full cyclic path set

We can also efficiently calculate the link flows from logit network loading, if the
set of reasonable paths contains all paths, even cyclic ones. Again we will use
formula (8.59). If we have an efficient way to compute Vab for all pairs of nodes
a and b, we can substitute them into this formula to directly obtain the link
flows.
Since the set of reasonable paths contains cycles, we cannot calculate these V
values inductively on topological order, as was done above. Rather, a different
approach is needed.
Let V be the n × n matrix whose components are Vab . Recall that Vab is
defined as the sum of exp(−θcσ ) for all segments σ starting at node a and ending
at node b. We will calculate this sum by dividing the sum into segments of the
same length.
We begin by calculating the part of Vab which corresponds to the segments of
length one (that is, the segments consisting of a single link.) If there are no par-
allel links in the network, then this is simply Lab : there is at most one such seg-
ment, which must be σ = [a, b], and if it exists exp(−θcσ ) = exp(−θtab ) = Lab .
If it does not exist, then Lab = 0, which is again the part of Vab corresponding
to segments of length one (which is empty if no such segment exists.) We can
proceed similarly if there are parallel links; see Exercise 20.
Now, let L be the n × n matrix whose components are Lab , and form the
matrix product L2 . Its components are
X
(L2 )ab = Lac Lcb (8.62)
c∈N

by the definition of matrix multiplication.8

For a given node c, the product Lac Lcb is zero unless there is a link both from
a to c, and from c to b. So, we can restrict the sum in (8.62) to be over nodes
c for which there are links (a, c) and (c, b) — which is precisely the nodes c for
which there is a segment [a, c, b] of length two. Furthermore, for such segments,

exp(−θc[a,c,b] ) = exp(−θ[tac + tcb ]) = exp(−θtac ) exp(−θtcb ) = Lac Lcb . (8.63)

So, the components of L2 are exactly the portion of the sums defining Vab for
segments of length two!
We demonstrate this using the example from Figure 8.6, recalling that d14 =
1, θ = log 2, and that all links have unit cost. In this network, the matrices L
and L2 take the form
   
0 1/2 1/2 0 0 1/4 1/4 1/2
0 0 1/2 1/2 2
0 1/4 0 1/4
L= 0 1/2
 L = 0 0 1/4 1/4 .
  (8.64)
0 1/2
0 0 0 0 0 0 0 0
8 The parentheses are intentional: (L2 ) 2
ab is the component of matrix L in row a and
column b. This is not the same as (Lab )2 , the square of the component of matrix L in row a
and column b.
292 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Looking at the first row of L2 , we see that (L2 )12 = 1/4, because there is one
segment which starts at node 1, ends at node 2, and contains two links ([1,3,2])
and exp(−θc[1,3,2] ) = 1/4. Similarly (L2 )13 = 1/4, and (L2 )14 = 1/2 because
there are two segments starting at node 1, ending at node 2, and containing two
links ([1,2,4] and [1,3,4]). The sum exp(−θc[1,2,4] ) + exp(−θc[1,3,4] ) is indeed
1/2. In this matrix, (L2 )23 = 0, because there are no segments of length two
starting at node 2 and ending at node 3.
Proceeding a step further, we have
X
(L3 )ab = (L2 )ac Lcb . (8.65)
c∈N

By the same logic, we see that this sum expresses the component of Vab corre-
sponding to segments of length three. Every segment of length three connecting
a to b consists of a segment of length two connecting a to some node c, followed
by a link (c, b). Group the sum of all segments of length three by this final link,
and note that (L2 )ac already contains the relevant portion of the sum for the
first segment.
Thus, by induction, the components of matrix Ln contains the portion of
the sum defining V corresponding to segments of length n. Therefore
V = L0 + L1 + L2 + L3 + · · · , (8.66)
where the infinite sum is needed since we allow paths with an arbitrary number
of cycles. Assuming that this sum exists, we can calculate it as follows:
V = I + L + L2 + L3 + · · · (8.67)
2
= I + L(I + L + L + · · · ) (8.68)
= I + LV . (8.69)
Therefore V − LV = I, or
V = (I − L)−1 . (8.70)
After calculating the matrix V with this formula, we can directly read off its
components Vab and use them to calculate the link flows using (8.59).
To complete the example, we have
 
1 1 1 1
0 4/3 2/3 1
V= 0 2/3 4/3 1
 (8.71)
0 0 0 1
and, for example, the flow on link (2,3) is given by
V12 L23 V34 1· 1 ·1 1
x23 = d14 =1 2 = . (8.72)
V14 1 2
Repeating this process will give the flow on every link — and unlike in the
previous section, does not involve summing an infinite series term-by-term. The
formula (8.59) handles all of the paths.
8.3. STOCHASTIC USER EQUILIBRIUM 293

8.3.5 Stochastic user equilibrium

Stochastic network loading methods, described in the previous section, are the
analogue of the all-or-nothing assignments we used to find x∗ in the method of
successive averages, or Frank-Wolfe, in Chapter 6. Recall that in those methods,
we identified x∗ by finding the link flows that would be observed if all the travel
times were held constant at the current values. In TAP, all drivers aim to take
the shortest path, so x∗ was obtained by loading all flow onto the shortest paths
at the current travel times. In logit SUE, drivers do not always take the shortest
path, but instead choose paths by (8.39). The methods in the previous section
thus calculate an “x∗ ” in the sense that it reflects the link flows which would
arise if travel times were held constant.
The full development of the SUE model requires relaxing the assumption of
constant travel times, in the same way as finding a traditional user equilibrium
requires more than a single shortest path computation. SUE is very easy to
define as a fixed point problem. Let C(h) denote the vector of path travel times
as a function of the vector of path flows (this is the same as before). However,
the logit formula directly gives us a complementary function H(c) giving path
flows as a function of path travel times, with components
exp(−θcπ )
hπ = drs P (8.73)
π 0 ∈Πrs exp(θcπ )
0

where (r, s) is the OD pair corresponding to path π. Therefore, the SUE prob-
lem can be expressed as follows: find a feasible path flow vector h∗ such that
h∗ = H(C(h∗ )). This is a standard fixed-point problem. Clearly H and C
are continuous functions if the link performance functions are continuous, and
the feasible path set is compact and convex, so Brouwer’s theorem immediately
gives existence of a solution to the SUE problem.
Notice that this was much easier than showing existence of an equilibrium
solution to the original traffic assignment problem! For that problem, there was
no equivalent of (8.73). Travelers were all using shortest paths, but if there
were two or more shortest paths there was no rule for how those ties should
be broken. As a result, we had to reformulate the problem as a variational
inequality and introduce an auxiliary function based on movement of a point
under a force. For the SUE problem, there is no need for such machinations,
and we can write down the fixed point problem immediately.
However, fixed-point theorems do not offer much help in terms of actually
finding the SUE solution. It turns out that convex programming and variational
inequality formulations exist as well, and we will get to them shortly. But first,
as a practical note, we mention that our definition of the reasonable path set
Π̂ should be specified without reference to the final travel times. The reason
is that until we have found the equilibrium solution, we do not know what
the link and path travel times will be. If the sets of reasonable paths vary
from iteration to iteration, as the flows and travel times change, there may be
problems with convergence or solution consistency. This is why the methods
described above for generating totally acyclic path sets relied on constants c0ij
294 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

which were independent of flow: constants such as physical length or free-flow

times. If using the full cyclic path set, there is no concern; at all iterations every
path is reasonable.

Path-flow formulation and the method of successive averages

Consider the following optimization problem, developed by Caroline Fisk:
X Z xij X
min θ tij (x)dx + hπ log hπ (8.74)
x,h 0
(i,j)∈A π∈Π̂
X
s.t. xij = hπ δij
π
∀(i, j) ∈ A (8.75)
π∈Π̂
X
hπ = drs ∀(r, s) ∈ Z 2 (8.76)
π∈Π̂rs

hπ ≥ 0 ∀π ∈ Π̂ (8.77)

Comparing with Beckmann’s formulation from Chapter 5, we see that it is iden-

tical except that the sums on path variables are now over the set of reasonable
paths, rather than all paths, and that the objective is first scaled by θ, and
then an additional term is added. We show below that these modifications en-
sure that the optimal path flows satisfy the logit formula (8.39). This objective
function is strictly convex in the path flows (see Exercise 23), so the SUE path
flow solution is unique. This is a different situation than the classical traffic
assignment problem, which has a unique equilibrium solution in link flows, but
not in path flows.
First, we show that the non-negativity constraint (8.77) can be ignored with-
out any problem, and in fact that at optimality the flow on each path is strictly
positive. The log hπ term in the objective is not defined at all if a path flow is
negative; if hπ = 0 we can define hπ log hπ = 0, since this is the limiting value
by l’Hôpital’s rule. But the derivative of h log h becomes infinitely steep as h
d
approaches zero: dh (h log h) = log h + 1 → −∞ as h → 0. Therefore the objec-
tive function cannot be minimized at h = 0; no matter what the change in the
other terms in the objective would be, it is better to have a very slightly posi-
tive h value than a zero one. This is a very useful property, since the optimality
conditions are much simpler if there are no non-negativity constraints.
Next, as with Beckmann’s formulation, we Lagrangianize the demand con-
straint (8.76), and substitute (8.75) in place of x, to obtain a Lagrangian in
terms of path flow variables only:
hπ δij
π
P
X Z π∈Π̂ X
L(h, κ) = θ tij (x)dx + hπ log hπ +
(i,j)∈A 0
π∈Π̂
 
X X
κrs drs − hπ  (8.78)
(r,s)∈Z 2 π∈Π̂rs
8.3. STOCHASTIC USER EQUILIBRIUM 295

Since there are no non-negativity conditions, it is enough to find the stationary

points of the Lagrangian, that is, the points where ∂L ∂h = 0. Therefore, we
require
∂L
= θcπ + 1 + log hπ − κrs = 0 , (8.79)
∂hπ
or, solving for hπ ,
hπ = exp(κrs − 1) exp(−θcπ ) . (8.80)
To find the value of the Lagrange multiplier κrs , we substitute (8.80) into the
demand constraint (8.76), and find that κrs must be chosen such that
drs
exp(κrs − 1) = P π0
. (8.81)
π 0 ∈Π̂rs exp(−θc )

Substituting into (8.80) gives the logit formula, and therefore the path flows
solving Fisk’s convex optimization problem are those solving logit stochastic
user equilibrium.
This convex optimization problem can be solved using the method of suc-
cessive averages. In this way, SUE can be solved quite simply, using a familiar
method from the basic TAP. The only change is that x∗ is calculated using
a method from the previous section, rather than by finding an all-or-nothing
assignment loading all flow onto shortest paths. The algorithm is as follows:
1. Choose an initial feasible link assignment x.
2. Update link travel times based on x.
3. Calculate target flows x∗ :
(a) For each OD pair (r, s), use a method from the previous section to
calculate OD-specific flows x∗rs .
(b) Calculate x∗ ← (r,s)∈Z 2 x∗rs
P

4. Update x ← λx∗ + (1 − λ)x for some λ ∈ (0, 1].

5. If x and x∗ are sufficiently close, terminate. Otherwise, return to step 2.
(It is often possible to do the calculations in step 3 per origin or per destination,
rather than separately for each OD pair.)
Notice that the termination criteria is slightly different than before. With
the classical traffic assignment problem, we argued that it was a bad idea to
compare the current solution to the previous solution and terminate when they
are sufficiently small. Here, we are not doing that, but are doing something
which looks similar: comparing the current solution x with the “target” solution
x∗ . This works because, unlike in the regular traffic assignment problem, the
mapping H(C) is always continuous and well-defined. This means that x∗ varies
slowly with x: small changes in link flows mean small changes in link and path
travel times, which means small changes in path flows from (8.73). With classic
traffic assignment, a small change in link flows could mean a shift in the shortest
296 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

3
10x
50+x

6 10+x 6
1 4
50+x
10x
2

Figure 8.14: Network for demonstrating method of successive averages in

stochastic assignment.

path, which would result in a dramatic change in x∗ , moving all flow from an
OD pair to that new shortest path. Furthermore, when there are ties there
are multiple possible x∗ values. So, in this case we had to introduce auxiliary
convergence measures like the relative gap or average excess cost. SUE is simpler
in that we can simply compare the current and target solutions — in fact, as
defined earlier, the relative gap and average excess cost do not make sense,
since the equilibrium principle is no longer defined by all travelers being on the
shortest path. For these reasons, the method of successive averages works much
better for SUE than for deterministic assignment.
To demonstrate this algorithm, consider the Braess paradox network shown
in Figure 8.14, where the demand from node 1 to node 4 is 6 vehicles, θ = 0.01,
and all paths are allowable. Assume that we choose the initial solution x by
performing Dial’s method on this network, using the free-flow travel times.
Table 8.3 shows the calculations; in this table, we first calculate the Li values
for nodes (based on shortest paths), then the link likelihoods Lij ; then the
node and link weights Wi and Wij , and finally the node and link flows xi and

xij . This gives initial link flows of x = 4.282 1.718 2.563 1.718 4.282 .
Recalculating the link performance functions with these flows gives new link
∗
travel times, which we put into Dial’s method again, giving the result x =
3.976 2.024 1.951 2.024 3.976 . Notice that x∗ and x are quite close to


each other! This is a very different situation than when the method of successive
averages is applied to the classical traffic assignment problem (compare with the
examples in Section 6.2.1), where x∗ was an extreme-point solution quite far
away from x.
So the link flows x are updated by averaging x∗ into the old x values, using a
weight of λ = 1/2, producing the results in the rightmost column of Figure 8.14.
The process is repeated until convergence.

Disaggregate link-flow formulation and Frank-Wolfe

For the classical TAP, the Frank-Wolfe algorithm was much faster than the
method of successive averages, because it chose λ adaptively, to maximize the
reduction in the objective function at each iteration. In theory, it is possible to
do the same thing with the convex program described above. In practice, it is
8.3. STOCHASTIC USER EQUILIBRIUM 297

Table 8.3: Method of successive averages for stochastic user equilibrium

Initialization Iteration 1
Nodes Li Wi xi Li Wi xi
1 0 1 6 0 1 6
2 0 1 4.282 42.8 1 3.976
3 10 1.670 4.282 51.7 1.964 3.976
4 10 2.341 6 94.5 2.964 6
Links Lij Wij xij Lij Wij x∗ij xij
(1,2) 1 1 4.282 1 1 3.976 4.129
(1,3) 0.670 0.670 1.718 1 1 2.024 1.871
(2,3) 1 1 2.563 0.964 0.964 1.951 2.257
(2,4) 0.670 0.670 1.718 1 1 2.024 1.871
(3,4) 1 1.670 4.282 1 1.964 3.976 4.129

harder because this program makes use of the path-flow variables h. With the
classical traffic assignment problem, we could express solutions and the objective
solely in terms of the link flows x. The addition of the h log h terms to the SUE
objective function renders this impossible. Furthermore, the number of paths
grows very quickly with network size (and if we are choosing the reasonable
path set to include fully cyclic paths, the number of paths is usually infinite).
The method of successive averages avoids this problem by not actually refer-
ring to the objective function at any point — if you review the steps above, you
see that the objective function is never calculated. Its role is implicit, guarantee-
ing that the algorithm will eventually converge, since the direction x∗ − x is one
along which the objective is decreasing. If we can find an efficient way to evalu-
ate the objective function, then we can develop an analogue to the Frank-Wolfe
method for classical assignment.
It turns out that we can reformulate the objective function in terms of the
destination-aggregated flows on each link, xsij (see Section 5.2.3), using the
Markov property of the logit loading. There is an equivalent disaggregation by
origin (if we reverse our interpretation of the Markov property; see Exercise 18.)
Recall the following results from Section 8.3.3:9

ˆ There exist values Pijs for each link (i, j) and destination s, such that
Y π
hπ = drs (Pijs )δij (8.82)
(i,j)∈π

for any path π connecting origin r to destination s.

ˆ The Pijs values can be interpreted as the conditional probability that a

vehicle arriving at node i and destined for node s will have link (i, j) as
9 Technically, this section only proved them for the case of a single origin-destination pair,

but they hold for each destination as well.

298 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

the next link in its path. Therefore,

xsij
Pijs = (8.83)
xsi

for each destination s, node i 6= s, and outgoing link (i, j).

Using these properties, we derive an equivalent formula for the additional

term to the Beckmann function.

Proposition 8.4. Let h be a feasible path flow vector satisfying the Markov
property, and let xsij and xsi be the corresponding destination-aggregated link
and node flows. Then
 
π
 
X h X X X
hπ log = xsij log xsij − xsi log xsi  (8.84)

d rs 
π∈Π̂ s∈Z (i,j)∈A i∈N
i6=s

Proof. We treat each destination s separately; summing over all destinations

gives the result.
Using (8.82) and (8.83), we have
 
 π π
 s δij
X h X Y x ij
hπ rs
= hπ log  s
 (8.85)
d x i
π∈Π̂s s
π∈Π̂ (i,j)∈A
X X xsij
= hπ π
δij log (8.86)
xsi
π∈Π̂s (i,j)∈A
!
X X xsij
= hπ δij
π
log (8.87)
xsi
(i,j)∈A π∈Πs
X xsij
= xsij log (8.88)
xsi
(i,j)∈A
X X
= xsij log xsij − xsij log xsi (8.89)
(i,j)∈A (i,j)∈A
X X
= xsij log xsij − xsi log xsi (8.90)
(i,j)∈A i∈N
i6=s

by grouping terms in the last sum by the tail node.

Furthermore, the left-hand side of equation (8.84) is very similar to the

second term in the objective function (8.74). In fact,
 π
X
π h X X
h log rs
= hπ log hπ − drs log drs (8.91)
d
s
π∈Π̂ s π∈Π̂ r∈Z
8.3. STOCHASTIC USER EQUILIBRIUM 299

of logarithms and the fact that π∈Πrs = drs . They differ

P
using properties
only by r∈Z drs log drs , which is a constant — and recall from Proposition B.2
P
that adding a constant does not affect the optimal solution to an optimization
problem.
This means that we can replace the objective function (8.74) with
 
X Z xij
1 X X X
tij (x)dx + xsij log xsij − xsi log xsi  (8.92)
 
θ

(i,j)∈A 0 s∈Z (i,j)∈A i∈N
i6=s

which does not require path enumeration. We can thus develop the following
analogue to the Frank-Wolfe algorithm:

1. Choose an initial, feasible destination-aggregated link assignment xs , and

the corresponding aggregated link flows x.

2. Update link travel times based on x.

3. Calculate target flows:

(a) For each destination s, use a method from the previous section to
calculate OD-specific flows x∗s .
(b) Calculate x∗ ← s∈Z x∗s
P

xs +
 
4. Find
 ∗ the∗ value
 of λ minimizing (8.92) along the line (1 − λ) x
λ x xs .

5. Update x ← λx∗ + (1 − λ)x .

6. If x and x∗ are sufficiently close, terminate. Otherwise, return to step 2.

8.3.6 Relationships between logit loading and most likely

path flows (*)
(This optional section draws a connection between the logit-based stochastic
equilibrium model and the concept of entropy maximization in most likely path
flows.)
The use of the h log h terms to represent logit assignment in the previous
section may have reminded you of the most likely path flows problem in deter-
ministic user equilibrium, discussed in Section 5.2.2. In fact, these two concepts
are related to each other quite closely! This optional section explores this rela-
tionship, and describes another algorithm for most likely based path flows.
300 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Recall that the most likely path flows were defined as those maximizing
entropy, and solving the following optimization problem:
X
max − hπ log(hπ /d) (8.93)
h
π∈Π̂
X
s.t. π
δij hπ = x∗ij ∀(i, j) ∈ A (8.94)
π∈Π
X
hπ = d (8.95)
π∈Π̂
hπ ≥ 0 ∀π ∈ Π (8.96)
where we have assumed there is a single origin-destination pair for simplicity.
We can transform this into an equivalent optimization that more closely re-
sembles the stochastic user equilibrium optimization problems. First, we replace
maximization by minimization by changing the sign of the objective. Second,
we can remove d from the objective, because
X X X
hπ log(hπ /d) = hπ log hπ − hπ log d (8.97)
π∈Π̂ π∈Π̂ π∈Π̂
X
= hπ log hπ − d log d (8.98)
π∈Π̂

and the constant d log d can be ignored. This gives

X
min hπ log hπ (8.99)
h
π∈Π̂
X
s.t. π
δij hπ = x∗ij ∀(i, j) ∈ A (8.100)
π∈Π
X
hπ = d (8.101)
π∈Π̂
hπ ≥ 0 ∀π ∈ Π (8.102)
We make one more change: rather than insisting that the path flows match
the equilibrium link flows exactly, we change constraint (8.100) to require that
the average travel time across all travelers be c∗ . That is, we change the con-
straint to X
hπ cπ = c∗ d . (8.103)
π∈Π̂
Arguing as in the previous section, we can ignore the non-negativity constraint (8.102),
since it will not be binding at optimality.
We then Lagrangianize the other two constraints, introducing multipliers θ
and κ for (8.103) and (8.102), respectively. The resulting Lagrangian function
is
   
X X X
L(h, θ, κ) = hπ log hπ + θ  hπ cπ − c∗ d + κ d − hπ  (8.104)
π∈Π̂ π∈Π̂ π∈Π̂
8.3. STOCHASTIC USER EQUILIBRIUM 301

without non-negativity constraints. The optimality conditions are thus

∂L
= 1 + log hπ + θcπ − κ = 0 ∀π ∈ Π̂ (8.105)
∂hπ
∂L X
= hπ cπ − c∗ d = 0 (8.106)
∂θ
π∈Π̂
∂L X
= hπ − d = 0 (8.107)
∂κ
π∈Π̂
(8.108)

The last two of these are simply the constraints (8.100) and (8.101). The first
can be solved for the path flows, giving

hπ = exp(κ − 1 − θcπ ) . (8.109)

To satisfy the demand constraint (8.101), κ must be chosen so that the sum
of (8.109) over all paths gives the total demand d.
Omitting the algebra, we must have
X
κ = log d(1 − log exp(−θcπ )) (8.110)
π 0 ∈Π

or
exp(−θcπ )
hπ = d P π
. (8.111)
π 0 ∈Π exp(−θc )

But this is just the logit formula! The Lagrange multiplier θ must be chosen to
satisfy the remaining constraint on the average path cost.
This derivation shows that the most likely path flows and stochastic user
equilibrium problems have a similar underlying structure. If we relax the re-
quirement that all travelers be on shortest paths, and simply constrain the aver-
age cost of travel, the most likely path flows coincide with a logit loading, where
the parameter θ is the Lagrange multiplier for this constraint. As θ approaches
infinity, the average cost of travel approaches its value at the deterministic user
equilibrium solution, and the stochastic user equilibrium path flows approach
the most likely path flows in the corresponding deterministic problem. This
provides another algorithmic approach for solving for most likely path flows, in
addition to those discussed in Section 6.5. In practice, this algorithm is difficult
to implement, because of numerical issues that arise as θ grows large.

8.3.7 Alternatives to logit loading

The majority of this section has focused on the logit model for stochastic
route choice, because it demonstrates the main ideas simply, and leads to
computationally-efficient solution techniques. There are several serious criti-
cisms of logit assignment. For instance, the assumption that the error terms
π are independent across paths is hard to defend if paths overlap significantly.
302 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

Common methods for creating totally acyclic route sets can also create unrea-
sonable artifacts, and allowing all cyclic paths can also be unreasonable if the
network topology creates many such paths with low travel times. See Exer-
cise 25 for some concrete examples. The main alternative to the logit model is
the probit model, in which the  terms have a multivariate normal distribution,
with a (possibly nondiagonal) covariance matrix to allow for correlation between
these terms.
The framework of stochastic user equilibrium can be generalized to other
distributions of the error terms, including correlation. The full development
of this general framework is beyond the scope of this book, but we provide an
overview and summary. The objective function (8.74) must be replaced with
X   X X Z xij
rs π
z(x) = − d E minrs (c + ) + xij tij − tij (x) dx
π∈Π 0
(r,s)∈Z 2 (i,j)∈A (i,j)∈A
(8.112)
where the expectation is taken with respect to the “unobserved” random vari-
ables , and tij in the last two terms are understood to be functions of xij . It
can be shown that this function is convex, and therefore that the SUE solution
is unique. However, evaluating this function is harder. The first term in (8.112)
involves an expectation over all paths connecting an OD pair. In discrete choice,
this is known as the satisfaction function, and expresses the expected perceived
travel time on a path chosen by a traveler. In the case of the logit model, this ex-
pectation can be computed in closed form; for most distributions it cannot, and
must be evaluated through Monte Carlo sampling or another approximation.
The method of successive averages can still be used, even without evaluating
the objective function. Step 4a needs to be replaced with a stochastic network
loading, using the current travel times and whatever distribution of  is chosen.
This often requires Monte Carlo sampling as well: for (multiple) samples of the
 terms, the shortest paths can be found using one of the standard algorithms,
and the resulting flows averaged together to form an estimate of x∗ .
Because we are using an estimate of x∗ , it is possible that the target direction
is not exactly right, and that it is not in a direction in which z(x) is decreasing.
Nevertheless, as long as it is correct “on average” (i.e., the sampling is done in
an unbiased manner), one can show that the method of successive averages will
still converge to the stochastic user equilibrium solution.

8.4 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
The original formulation of the user equilibrium traffic assignment problem
in Beckmann et al. (1956) actually modeled demand as elastic. (In this book
we chose to focus on the fixed-demand problem in previous chapters, which is
a special case.) The transformation of the elastic demand problem to a fixed-
demand problem with artificial links was reported in Gartner (1980), based on
8.5. EXERCISES 303

earlier work by Murchland (1970). While the original Frank-Wolfe algorithm

can be applied to the elastic demand equilibrium problem, the modified ver-
sion presented in this chapter is faster. It is essentially a specialized version
of the double-stage algorithm of Evans (1976) developed for a combined trip
distribution and assignment model.
The variational inequality formulation of the traffic assignment problem with
link interactions is due to Smith (1979a), Dafermos (1980), and Aashtiani and
Magnanti (1981). The diagonalization method was demonstrated in Fisk and
Nguyen (1982), and its convergence proved by Dafermos (1982). The simplicial
decomposition method is adapted from Smith (1983). The example with mul-
tiple equilibria in the simple merge network is taken from Boyles et al. (2015).
Other examples of multiple equilibria are given in Netter (1972) and Marcotte
and Wynter (2004).
Stochastic network loading using the logit model and totally acyclic path
sets was described in Dial (1971). Logit loading with the full cyclic path set
was developed by Bell (1995), Akamatsu (1996), and Akamatsu (1997).
The stochastic user equilibrium model was first proposed in Daganzo and
Sheffi (1977). Powell and Sheffi (1982) proved the convergence of the method
of successive averages for this problem, for all well-behaved distributions of the
unobserved utility term. The convex programming case for the logit model
was given in Fisk (1980), and the optimization formulation for more general
distributions was given in Daganzo (1982).

8.5 Exercises
1. [25] Verify that each of the demand functions D below is strictly de-
creasing and bounded above (for κ ≥ 0), then find the inverse functions
D−1 (d).
(a) D(κ) = 50 − κ
(b) D(κ) = 1000/(κ + 1)
(c) D(κ) = 50 − κ2
(d) D(κ) = (κ + 2)/(κ + 1)
2. [13] The total misplaced flow reflects consistency of a solution with (in
this case, that the OD matrix should be given by the demand functions).
Suggest another measure for how close a particular OD matrix and traffic
assignment are to satisfying this consistency condition. Your proposed
measure should be a nonnegative and continuous function of values related
to the solution (e.g., drs , xij , κrs , etc.), which is zero if and only if the
OD matrix is completely consistent with the demand functions. Compare
your new measure with the total misplaced flow, and comment on any
notable differences.
3. [31] Verify that the elastic demand objective function (8.16) is convex,
given the assumptions made on the demand functions.
304 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

3
10x
50+x

10+x
1 4
50+x
10x
2

Figure 8.15: Network for Exercise 5.

4. [47] Using the network in Figure 8.15, solve the elastic demand equilibrium
problem with demand given by d14 = 15−κ14 /30. Perform three iterations
of the Frank-Wolfe method, and report the average excess cost and total
misplaced flow for the solution.
5. [48] Using the network in Figure 8.15, solve the elastic demand equilibrium
problem with demand given by d14 = 10 − κ14 /30.
(a) Perform three iterations of the Frank-Wolfe method designed for elas-
tic demand (Section 8.1.5).
(b) Transform the problem to an equivalent fixed-demand problem using
the Gartner transformation from Section 8.1.2, and perform three
iterations of the original Frank-Wolfe method.
(c) Compare the performance of these two methods: after three itera-
tions, which is closer to satisfying the equilibrium and demand con-
ditions?

6. [49] Using the network in Figure 8.16, solve the elastic demand equilibrium
problem with demand functions q 19 = 1000 − 50(u19 − 50) and q 49 =
1000 − 75(u49 − 50). The cost function on the light links is 3 + (xa /200)2 ,
and the cost function on the thick links is 5 + (xa /100)2 . 1000 vehicles
are traveling from node 1 to 9, and 1000 vehicles from node 4 to node 9.
Perform three iterations of the Frank-Wolfe method and report the link
flows and OD matrix.
7. [13] Assume that (x, d) and (x∗ , d∗ ) are both feasible solutions to an
elastic demand equilibrium problem. Show that (λx + (1 − λ)x∗ , λd +
(1 − λ)d∗ ) is also feasible if λ ∈ [0, 1]. This ensures that the Frank-Wolfe
solutions are always feasible, assuming we start with x and d values which
are consistent with each other, and always choose targets in a consistent
way.
8. [57] (Calculating derivative formulas.) Let x and d be the current, feasi-
ble, link flows and OD matrix, and let x∗ and d∗ be any other feasible link
flows and OD matrix. Let x0 = λx∗ + (1 − λ)x and d0 = λd∗ + (1 − λ)d.
8.5. EXERCISES 305

7 8 9

4 5 6

1 2 3

Figure 8.16: Network for Exercise 6.

(a) Let f (x0 , d0 ) be the objective function for the elastic demand equi-
librium problem. Recognizing that x0 and d0 are functions of λ,
calculate df /dλ.
df
(b) For the elastic demand problem, show that dλ λ=0
≤ 0 if d∗ and
∗
x are chosen in the way given in the text. That is, the objective
function is nonincreasing in the direction of the “target.” (You can
assume that the demand function values are strictly positive if that
would simplify your proof.)
9. [24] Consider a two-link network. For each pair of link performance
functions shown below, determine whether or not the symmetry condi-
tion (8.26) is satisfied.
(a) t1 = 4 + 3x1 + x2 , t2 = 2 + x1 + 4x2
(b) t1 = 7 + 3x1 + 4x2 , t2 = 12 + 2x1 + 4x2
(c) t1 = 4 + x1 + 3x2 , t2 = 2 + 3x1 + 2x2
(d) t1 = 3x21 + x2 , t2 = 4 + x1 + 4x32
(e) t1 = 3x21 + 2x22 , t2 = 2x21 + 3x32
(f) t1 = 50 + x1 , t2 = 10x2
10. [34] Determine which of the pairs of link performance functions in the
previous exercise are strictly monotone.
11. [49] Consider the network in Figure 8.16 with a fixed demand of 1000
vehicles from 1 to 9 and 1000 vehicles from 4 to 9. The link performance
function on every link a arriving at a “merge node” (that is, nodes 5, 6, 8,
and 9) is 4 + (xa /150)2 + (xa0 /300)2 where a0 is the other link arriving at
the merge. Verify that the link interactions are symmetric, and perform
five iterations of the Frank-Wolfe method. Report the link flows and travel
times.
12. [49] Consider the network in Figure 8.16 with a fixed demand of 1000
vehicles from 1 to 9 and 1000 vehicles from 4 to 9. The link performance
function on every link a arriving at a “merge node” (that is, nodes 5,
306 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

1 3

A B C
2 4

Figure 8.17: Network for Exercise 13.

6, 8, and 9) is 3 + (xa /200)2 + (xa0 /400)2 if the link is light, and 5 +

(xa /100)2 +(xa0 /200)2 if the link is thick, where a0 is the other link arriving
at the merge. Show that not all link interactions are symmetric, and then
perform five iterations of the diagonalization method and report the link
flows, travel times, and relative gap.
13. [49] See the network in Figure 8.17, where the numbers indicate the label
for each link. The link performance functions are:
t1 (x) = 1 + 4x1 + 2x2
t2 (x) = 2 + x1 + 2x2
t3 (x) = 3 + 2x3 + x4
t4 (x) = 2 + 3x4

and the demand from node A to node C is 10 vehicles. For both methods
below, start with an initial solution loading all flow on links 1 and 4.
(a) Use three iterations of the diagonalization method to try to find an
equilibrium solution, and report the link flows and average excess
cost. (As before, this means finding three x∗ vectors after your initial
solution.)
(b) Use three iterations of simplicial decomposition to try to find an
equilibrium solution, and report the link flows and average excess
cost. Three iterations means that X should have three vectors in it
when the algorithm terminates (unless it terminates early because
the x∗ you find is already in X ). For each subproblem, make the
number of improvement steps one less than the size of X (so when
X has 1 vector, perform 0 steps; when it has 2 vectors, perform 1
step, and so on). For each of these steps, try the sequence of µ
values 1/2, 1/4, 1/8, . . ., choosing the first that reduces the restricted
average excess cost.

14. [64] Prove Proposition 8.2.

15. [42] Show that the three path generation methods described in Section 8.3.2
indeed yield totally acyclic path sets, and that they always include the
shortest path from the origin to the destination.
8.5. EXERCISES 307

2 H 2 I
G
2 2 2
D 1 E 1 F
2 2 2
A 2 B 2 C

Figure 8.18: Network for Exercises 21 and 22.

16. [32] Show that any set of totally acyclic paths satisfies the segment sub-
stitution property. Then show that the set of all paths (including all cyclic
paths) also satisfies this property.

17. [53] Complete the example following equation (8.40) by forming the infi-
nite sum for the flow on links (2,3) and (3,2), and showing that they are
equal to 1/2.

18. [62] Reformulate the Markov property, and the results in Section 8.3.3 in
terms of conditional probabilities for the link a vehicle used to arrive at
a given node, rather than the link a vehicle will choose to depart a given
node.

19. [34] In the logit formula (8.39), show that the same path choice probabil-
ities are obtained if each link travel time tij is replaced with Li + tij − Lj ,
where L is a vector of node-specific constants. (In practice, these are
usually the shortest path distances from the origin.)

20. [51] What would have to change in the “full cyclic path set” stochastic
network loading procedure, if there were multiple links with the same tail
and head nodes?

21. [35] Consider the network in Figure 8.18 with 1000 vehicles traveling from
node A to node I, where the link labels are the travel times. Use the first
criterion in Section 8.3.2 to define the path set, and identify the link flows
corresponding to these travel times. Assume θ = 1.

22. [35] Repeat Exercise 21 for the third definition of a path set. Assume
θ = 1.

23. [34] Show that the objective function (8.74) is strictly convex.

24. [49] Consider the network in Figure 8.16 with a fixed demand of 1000
vehicles from 1 to 9 and 1000 vehicles from 4 to 9. The light links have
delay function 3 + (xij /200)2 , and the dark links have delay function 5 +
308 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT

(xij /100)2 . Assume that drivers choose paths according to the stochastic
user equilibrium principle, with θ = 1/5. Perform three iterations of the
method of successive averages, and report the link flows. Assume all paths
are allowed.
25. [24] (Limitations of logit assignment). This problem showcases three
“problem instances” for the stochastic network loading models described
in this chapter (the logit model, and one proposed definition of allowable
paths). Throughout, assume that the first definition in Section 8.3.2 is
used to define the allowable path set. These instances motivated the de-
velopment of probit and other, more sophisticated, stochastic equilibrium
models.

(a) Consider the network in Figure 8.19(a), where the numbers by each
link represents the travel time and z ∈ (0, 1). Let p↑ represent the
proportion of vehicles choosing the
qtop path.
 In a typical probit
model, we have p↑P ROBIT = 1 − Φ z
2π−z , where Φ(·) is the stan-
dard cumulative normal distribution function. Calculate p↑LOGIT for
the logit model as a function of z, and plot p↑P ROBIT and p↑LOGIT as
z varies from 0 to 1. Which do you think is more realistic, and why?
(b) Consider networks of the type shown in Figure 8.19(b), where there is
a top path consisting of a single link, and a number of bottom paths.
The network is defined by an integer m; there are m − 1 intermediate
nodes in the bottom paths, and each consecutive pair of intermediate
nodes is connected by two parallel links with travel time 8/m. In a
√
common probit model, p↑P ROBIT = Φ(−0.435 m). What is p↑LOGIT
as a function of m? Again create plots for small values of m (say 1
to 8), indicate which you think is more realistic, and explain why.
(c) In the networks in Figure 8.19(c), identify the proportion of travelers
choosing each link if θ = 1. The left and right panels show a net-
work before and after construction of a new link. Again identify the
proportion of travelers choosing each link if θ = 1. Do your findings
seem reasonable?
8.5. EXERCISES 309

1 3
1
z
z 1−z
2
(a)
8 8
1 8 2 4 4
1 3 2
8 4 4
m=1 m=2
8
2 2 2 2
1 3 4 5 2
2 2 2 2
m=4
(b)
7 7
1 2 1 2
8
1
10 3
9
After
Before (c)

Figure 8.19: Networks for Exercise 25. The label on each link is its constant
travel time.
310 CHAPTER 8. EXTENSIONS OF STATIC ASSIGNMENT
 Part III

Dynamic Traffic
Assignment

311
Chapter 9

Network Loading

This chapter discusses network loading, the process of modeling the state of
traffic on a network, given the route and departure time of every vehicle. In
static traffic assignment, this is a straightforward process based on evaluating
link performance functions. In dynamic traffic assignment, however, network
loading becomes much more complicated due to the additional detail in the
traffic flow models — but it is exactly this complexity which makes dynamic
traffic assignment more realistic than static traffic assignment. Rather than link
performance functions, dynamic network loading models generally rely on some
concepts of traffic flow theory. There are a great many theories, and an equally
great number of dynamic network loading models, so this chapter will focus on
those most commonly used.
To give the general flavor of network loading, we start with two simple link
models, the point queue and spatial queue (Section 9.1), which describe traffic
flow on a single link. We next present three simple node models describing how
traffic streams behave at junctions (Section 9.2). With these building blocks we
can perform network loading on simple networks, showing how link and node
models interact to represent traffic flow in a modular way (Section 9.3).
However, the point queue and spatial queue models have significant limita-
tions in representing traffic flow. The most common network loading models for
dynamic traffic assignment are based on the hydrodynamic model of traffic flow,
reviewed in Section 9.4. This theory is based in fluid mechanics, and assumes
that the traffic stream can be modeled as the motion of a fluid, but it can be
derived from certain car-following models as well, which have a more behavioral
basis. The cell transmission model and link transmission model are link models
based on this theory, and both of these are discussed in Section 9.5. Section 9.6
concludes the chapter with a discussion of more sophisticated node models that
can represent general intersections.
This chapter aims to present several alternative network loading schemes as
part of the general dynamic traffic assignment framework in Figure 1.8, so that
any of them can be combined with the other steps in a flexible way.

313
314 CHAPTER 9. NETWORK LOADING

9.1 Link Model Concepts

Link models are efficient ways to represent congestion and traffic flow. Efficiency
is key, since link models must be implemented on each link in a large network
over the entire analysis period; and furthermore, due to the iterative nature of
solving for equilibrium, the network loading must be repeated many times with
different path flow values as input. This section presents the basic concepts
of link models, and demonstrates them in the point queue and spatial queue
models. More sophisticated link models will be presented later in this chapter,
in Section 9.5.
So, in this section, we are solely concerned with a single link. Call the length
of this link L, so x = 0 corresponds to the upstream end of the link, and x = L
corresponds to the downstream end. All of the link models we discuss in this
book operate in discrete time. That is, we divide the analysis period into small
time intervals of length ∆t, and assume uniform conditions within each time
interval. These time intervals are very small relative to the analysis period; at
a minimum, no vehicle should be able to traverse more than one link in a single
time interval, so ∆t can be no greater than the shortest free-flow travel time
on a link. In practice, ∆t values on the order of 5–10 seconds are a reasonable
choice. For convenience in this section, we assume that the unit of time is
chosen
 such that ∆t = 1, so we will only track the state of the network for times
t ∈ 0, 1, 2, . . . , T̄ , where T̄ is the length of the analysis period (in units of ∆t).
In any discrete time model, it is important to clarify what exactly we mean
when we index a variable with a time interval, such as N (t) or y(t). Does this
refer to the value of x at the start of the t-th time interval, the end of the
interval, the average of a continuous value through the interval, or something
else? This chapter will consistently use the following convention: when referring
to a quantity measured at a single instant in time, such as the number of vehicles
on a link, or the instantaneous speed of a vehicle, we will take such measurements
at the start of a time interval. When referring to a quantity measured over time,
such as the number of vehicles passing a fixed point on the link, we take such
measurements over the time interval. (Figure 9.1) This distinction can also
be expressed by referring, say, to the number of vehicles on a link at time t
(meaning the start of the t-th interval), or the number of vehicles that exit a
link during time t.
We load the network in increasing order of time. That is, we start with the
network state at t = 0 (usually assuming an empty condition). Then, with these
values known, we compute the network state at t = 1, then at t = 2, and so
forth. That is, for any point in time, any values at an earlier point in time can
be treated as “known” values.
The definition of the “network state” at a timestep is intentionally left a bit
vague at this point. It contains all information necessary for modeling traffic
flow (e.g., the locations of vehicles, traffic signal indications). The collection
of network states at all time steps must also have enough information to check
consistency with the equilibrium principle (cf. Section 1.3) after the network
loading is complete, and for adjusting vehicle flows if equilibrium is not satisfied.
9.1. LINK MODEL CONCEPTS 315

y(0) y(1) y(2) y(3)

3 2 5 7
5 4 7
z(0) z(1) z(2)
t=0 1 2 3

Figure 9.1: Convention for indexing discrete time intervals for two hypothetical
variables y (measured at an instant) and z (measured over time).

The exact components of the network state vary from one model to the next
(for instance, signals may be modeled in detail, approximately, or not at all),
and as you read about the link and node models presented in this chapter,
you should think about what information you would need to store in order to
perform the computations for each link and node model. At a minimum, it is
common to record the cumulative number of vehicles which have entered and
left each link at each timestep, since the start of the modeling period. These
values are denoted by N ↑ (t) and N ↓ (t), respectively; the arrows are meant as
a mnemonic for “upstream” and “downstream”, since they can be thought of
as counters at the ends of the links. So, for instance, N ↑ (3) is the number of
vehicles which have crossed the upstream end of the link by the start of the
3rd time interval (the cumulative entries), and N ↓ (5) is the number of vehicles
which have crossed the downstream end of the link by the start of the 5th time
interval (cumulative exits). We will assume that the network is empty when
t = 0 (no vehicles anywhere), and as a result N ↑ (t) − N ↓ (t) gives the number
of vehicles currently on the link at any time t.
It is possible to use different time interval lengths for different links and
nodes, and this can potentially reduce the necessary computation time. There
are also continuous time dynamic network loading models, where flows and other
traffic variables are assumed to be functions defined for any real time value, not
just a finite number of points. These are not discussed here to keep the focus
on the basic network loading ideas, and to avoid technical details associated
with infinitesimal calculations. Finally, some formulas may call for the value
of a discrete variable at a non-integer point in time, such as N ↑ (3.4), in which
case a linear interpolation can be used between the neighboring values N ↑ (3)
and N ↑ (4). If possible, the time step should be chosen to minimize or eliminate
these interpolation steps, which are time-consuming and which can introduce
numerical errors.

9.1.1 Sending and receiving flow

The main outputs of a link model are the sending flow and receiving flow,
calculated for each discrete time interval. The sending flow at time t, denoted
S(t), is the number of vehicles which would leave the link during the t-th time
316 CHAPTER 9. NETWORK LOADING

Direction of travel

Distance traveled
in t

Figure 9.2: Calculating the sending flow when there is a queue on the link (left)
and when the link is at free-flow (right).

interval (that is, between times t and t + 1) if there was no obstruction from
downstream links or nodes (you can imagine that the link is connected to a
wide, empty link downstream). You can also think of this as the flow that is
ready to leave the link during this time interval. The sending flow is calculated
at the downstream end of a link.
To visualize sending flow, Figure 9.2 shows two examples of links. In the
left panel, there is a queue at the downstream end of the link. If there is no
restriction from downstream, the queue would discharge at the full capacity
of the link, and the sending flow would be equal to the capacity of the link
multiplied by ∆t. In the right panel, the link is uncongested and vehicles are
traveling at free-flow speed. The sending flow will be less than the capacity,
because relatively few vehicles are close enough to the downstream end of the
link to exit within the next time interval. The vertical line in the figure indicates
the distance a vehicle would travel at free-flow speed during one time step, so
in this case the sending flow would be 3 vehicles. Note that the actual number
of vehicles which can leave the link during the next time step may be affected
by downstream conditions: perhaps the next link is congested, or perhaps there
is a traffic signal at the downstream end. These considerations are irrelevant
for calculating sending flow, and will be treated with node models, introduced
in Section 9.2. In any case, the sending flow is an upper bound on the actual
number of vehicles which can depart the link.
The receiving flow during time t, denoted R(t), is the number of vehicles
which would enter the link during the t-th time interval if the upstream link
could supply a very large (even infinite) number of vehicles: you can imagine
that the upstream link is wide, and completely full at jam density. You can also
think of this as the maximum amount of flow which can enter the link during
this interval, taking into account the available free space. The receiving flow is
calculated at the upstream end of the link.
To visualize receiving flow, Figure 9.3 shows two examples of links. In the
left panel, the upstream end of the link is empty. This means that vehicles can
potentially enter the link at its full capacity, and the receiving flow would equal
the link’s capacity multiplied by ∆t. The actual number of vehicles which will
enter the link may be less than this, if there is little demand from upstream —
like the sending flow, the receiving flow is simply an upper bound indicating how
9.1. LINK MODEL CONCEPTS 317

Direction of travel

Distance traveled
in t

Figure 9.3: Calculating the receiving flow when the link is at free-flow (left) and
when there is a queue on the link (right).

Point queue

Physical section

Figure 9.4: Division of a link in point queue models.

many vehicles could potentially enter the link if demand were high enough. In
the right panel, there is a stopped queue which nearly fills the entire link. Here
the vertical line indicates how far into the link a vehicle would travel at free-flow
speed during one time step. Assuming that the stopped vehicles remain stopped
throughout the t-th time interval, the receiving flow is the number of vehicles
which can physically fit into the link, in this case 2.
Each link model has a slightly different way of calculating the sending and
receiving flows, which correspond to different assumptions on traffic behavior
with the link, or to different calculation methods. The next two subsections
present simple link models. Notice how the different traffic flow assumptions in
these models lead to different formulas for sending and receiving flow.

9.1.2 Point queue

Point queue models divide each link into two sections:

ˆ A physical section which spans the length of the link and is assumed
uncongestible: vehicles will always travel over this section at free-flow
speed.

ˆ A point queue at the downstream end of the link which occupies no

physical space, but conceptually holds vehicles back to represent any con-
gestion delay on the link.

These are shown in Figure 9.4.

Point queues can be thought of in several ways. One can imagine a wide
link which necks down at its downstream end — the physical section reflects the
318 CHAPTER 9. NETWORK LOADING

Physical section
Point queue

Figure 9.5: Envisioning a point queue as vehicles stacking vertically.

wide portion of the link, and the point queue represents the vehicles which are
delayed as the capacity is reduced at the downstream bottleneck. One can also
imagine a traffic signal at the downstream end, and magical technology (flying
cars?) which allows vehicles to “stack” vertically at the signal — there can be
no congestion upstream of this “stack,” since vehicles can always fly to the top.
(Figure 9.5) One may even imagine that there is no physical meaning to either
of these, and that the physical section and point queue merely represent the
delays incurred from traveling the link at free-flow, and the additional travel
time due to congestion.
↓
The point queue discharges vehicles at a maximum rate of qmax (measured
in vehicles per unit time), called the capacity. The capacity imposes an upper
limit on the sending flow, so we always have
↓
S(t) ≤ qmax ∆t . (9.1)
However, if the queue is empty, or if only a few vehicles are in the queue,
the discharge rate may be less than this. Once the queue empties, the only
vehicles which can exit the link are ones reaching the downstream end from the
uncongested physical section. Since we assume that all vehicles in this section
travel at the free-flow speed (which we will denote uf ), this means that only the
vehicles that are closer than uf ∆t to the downstream end can possibly leave.
We can use the cumulative counts N ↑ and N ↓ to count the number of vehicles
which are close enough to the downstream end to exit in the next time step.
Since the entire physical section is traversed at the free-flow speed uf , a vehicle
whose distance from the downstream end of the link is exactly uf ∆t distance
units must have passed the upstream end of the link exactly (L−uf ∆t)/uf time
units ago. We call this a “threshold” vehicle, since any vehicle entering the link
after this one has not yet traveled far enough, while any vehicle entering the
link before this one is close enough to the downstream end to exit. The number
of vehicles between the threshold vehicle and the downstream end of the link
can thus be given by
   
L − uf ∆t L
N↑ t − − N ↓ (t) = N ↑ t + ∆t − − N ↓ (t) . (9.2)
uf uf
The sending flow is the smaller of the number of vehicles which are close
enough to the downstream end to exit, given by equation (9.2), and the capacity
of the queue. Thus
S(t) = min{N ↑ (t + ∆t − L/uf ) − N ↓ (t), qmax
↓
∆t} . (9.3)
9.1. LINK MODEL CONCEPTS 319

The receiving flow for the point queue model is easy to calculate. Since the
physical section is uncongestible, the link capacity is the only limitation on the
rate at which vehicles can enter. The capacity of the upstream end of the link
may be different than the capacity of the downstream end of the link (perhaps
due to a lane drop, or a stop sign at the end of the link), so we denote the
↑
capacity of the upstream end by qmax . (If we just use qmax without an ↑ or
↓ superscript, the same capacity applies over the whole link.) ) The receiving
flow is given by
↑
R(t) = qmax ∆t . (9.4)
In real traffic networks, queues occupy physical space and cannot be confined to
a single point. The spatial queue model in the next subsection shows one way
to reflect this.
Table 9.1 shows how the point queue model operates, depicting the state
of a link over ten time steps. The N ↑ and N ↓ columns express the number
of vehicles which have entered and left the link at each time step, as well as
the sending and receiving flows during each timestep. The difference between
N ↑ and N ↓ represents the number of vehicles on the link at any point in time.
In this example, we assume that the free-flow speed is uf = L/(3∆t) (so a
vehicle takes 3 time steps to traverse the link under free-flow conditions), the
↑ ↓
upstream capacity is qmax = 10/∆t, and the downstream capacity is qmax =
5/∆t. Initially, the sending flow is zero, because no vehicles have reached the
downstream end of the link. The sending flow then increases as flow exits, but
eventually reaches the downstream capacity. At this point, a queue forms and
vehicles exit at the downstream capacity rate. Eventually, the queue clears, the
link is empty, and the sending flow returns to zero. The receiving flow never
changes from the upstream capacity, even when a queue is present. In this
example, notice that N ↓ (t + 1) = N ↓ (t) + S(t). This happens because we are
temporarily ignoring what might be happening from downstream. Depending
on downstream congestion, N ↓ (t + 1) could be less than N ↓ (t) + S(t); but it
could never be greater, because the sending flow is always a limit on the number
of vehicles that can exit. Also notice that for all time steps, N ↑ (t + 1) ≤
N ↑ (t) + R(t), because the receiving flow is a limit on the number of vehicles
that can enter the link.
In practice, the upstream and downstream capacities are usually assumed
the same, in which case we just use the notation qmax to refer to capacities
at both ends. The network features which would make the capacities different
upstream and downstream (such as stop signs or signals) are usually better
represented with node models, discussed in the next section.

9.1.3 Spatial queue

The spatial queue model is similar to the point queue model, except that a
maximum queue length is now enforced. When the queue reaches this maximum
length, no further vehicles are allowed to enter the link. That is, the queue now
occupies physical space, and because the link is finite in length, the link can
320 CHAPTER 9. NETWORK LOADING

↑
Table 9.1: Point queue example, with uf = L/(3∆t), qmax = 10/∆t, and
↓
qmax = 5/∆t.
t N↑ N↓ R S
0 0 0 10 0
1 1 0 10 0
2 5 0 10 0
3 10 0 10 1
4 17 1 10 4
5 27 5 10 5
6 30 10 10 5
7 30 15 10 5
8 30 20 10 5
9 30 25 10 5
10 30 30 10 0

become completely blocked. This will result in queue spillback, as vehicles on

upstream links will be unable to enter the blocked link. Like the point queue
model, we assume that the queue is always at the downstream end of the link:
there is at most one uncongested physical section at the upstream end of the
link, and at most one stopped queue at the downstream end, in that order. This
is still a simplification of real traffic (where links can have multiple congested
and uncongested sections), but by allowing the length of the physical section
to shrink as the queue grows, one can model the queue spillback phenomenon
which is common in congested networks.
The sending flow for the spatial queue model is calculated in exactly the
same way as for the point queue model: the smaller of the number of vehicles
close enough to the downstream end of the link to exit in the next time step,
and the capacity of the link:

S(t) = min{N ↑ (t + ∆t − L/uf ) − N ↓ (t), qmax

↓
∆t} . (9.5)

The receiving flow includes an additional term to reflect the finite space on
the link for the queue, alongside the link capacity. Since vehicles in the queue
are stopped, the space they occupy is given by the jam density kj , expressed in
vehicles per unit length. The maximum number of vehicles the link can hold is
kj L, while the number of vehicles currently on the link is N ↑ (t) − N ↓ (t). The
receiving flow cannot exceed the difference between these:

R(t) = min{kj L − (N ↑ (t) − N ↓ (t)), qmax

↑
∆t} . (9.6)

By assuming that the queue is always at the downstream end of the link,
the spatial queue model essentially assumes that all vehicles in a queue move
together. In reality, there is some delay between when the head of the queue
starts moving, and when the vehicle at the tail of the queue starts moving —
9.2. NODE MODEL CONCEPTS 321

↑ ↓
Table 9.2: Spatial queue example, with uf = L/(3∆t), qmax = 10/∆t, qmax =
5/∆t, and kj L = 20.
t N↑ N↓ R S
0 0 0 10 0
1 1 0 10 0
2 5 0 10 0
3 10 0 10 1
4 17 1 4 4
5 21 5 4 5
6 25 10 5 5
7 30 15 5 5
8 30 20 10 5
9 30 25 10 5
10 30 30 10 0

when a traffic light turns green, vehicles start moving one at a time, with a
slight delay between when a vehicle starts moving and when the vehicle behind
it starts moving. These delays cannot be captured in a spatial queue model.
To represent this behavior, we will need a better understanding of traffic flow
theory. Section 9.4 will present this information, and we will ultimately build
more realistic link models. The point queue and spatial queue models never-
theless illustrate the basic principles of link models, and what the sending and
receiving flow represent.
Table 9.2 shows how the spatial queue model operates. It is similar to the
point queue example (Table 9.1), but we now introduce a jam density, assuming
that the maximum number of vehicles which can fit on the link is kj L = 20.
In this example, the receiving flow drops once the queue reaches a certain size.
This reflects the finite space available on the link. As a result, it takes longer for
the 30 vehicles to enter the link. The queue is still able to completely discharge
by the end of the ten time steps. As before, the difference between N ↓ (t) and
N ↓ (t + 1) is never more than S(t) (in this example, exactly equal because we
are ignoring downstream conditions), and the difference between N ↑ (t) and
N ↑ (t + 1) is never more than R(t).

9.2 Node Model Concepts

Node models complement the link models discussed in the previous section,
by representing how flows between different links interact with each other. By
focusing on sending and receiving flows from each link, one can separate the
different processes used to model traffic flow within a single link, and the inter-
actions between links at junctions. Calculations of sending and receiving flows
can be done independently, in parallel, for each link. It is even possible to use
different link models on different links (say, one link with a spatial queue model
322 CHAPTER 9. NETWORK LOADING

g j
Sgi Rij

i Rik
Shi
Ril k

h
l

Figure 9.6: Sending and receiving flows interact at nodes to produce turning
movement flows.

and another with a point queue).

All that a node model needs is the sending flow of each of the links which
enter this node, and the receiving flow for each link leaving this node. (Fig-
ure 9.6) That is, when processing node i at time t, we assume that Shi (t) and
Rij (t) have respectively been calculated for each incoming link (h, i) ∈ Γ−1 (i)
and each outgoing link (i, j) ∈ Γ(i). The task is to determine how many vehi-
cles move from each incoming link to each outgoing link during the t-th time
interval; denote this value by yhij (t). This is called the turning movement flow
from (h, i) to (i, j), or more compactly, the turning movement [h, i, j]. Let Ξ(i)
denote the set of allowable turning movements at node i; this provides a natural
way to model turn prohibitions, U-turn prohibitions, and so forth.
As an example of this notation, consider node i in Figure 9.6. If all of the
turning movements are allowed, then Ξ(i) has six elements: [g, i, j], [g, i, k],
[g, i, l], [h, i, j], [h, i, k], and [h, i, l]. This set might not include all six elements
— for instance, if the left turn from approach (g, i) to (i, j) is prohibited, then
Ξ(i) would exclude [g, i, j].
Many different node models can be used. This section discusses three simple
node models — links in series, diverges, and merges — to illustrate the general
concepts. Later in this chapter, we discuss node models that can be used for
signalized intersections, all-way stops, and other types of intersections. There
are many variations of all of these, but they share some common principles,
which are discussed here.
Any node model must satisfy a number of constraints and principles, for all
time intervals t:

1. Vehicles will not voluntarily hold themselves back. That is, it should be
impossible to increase any turning movement flow yhij (t) without violating
one of the constraints listed below, or one of the other constraints imposed
by a specific node model.

2. Each turning movement flow must be nonnegative, that is, yhij (t) ≥ 0 for
9.2. NODE MODEL CONCEPTS 323

each [h, i, j] ∈ Ξ(i).

3. Any turning movement which is not allowable has zero flow, that is,
yhij (t) = 0 whenever [h, i, j] ∈
/ Ξ(i).
4. For each incoming link, the sum of the turning movement flows out of this
link cannot exceed the sending flow, since by definition those are the only
vehicles which could possibly leave that link in the next time step:
X
yhij (t) ≤ Shi (t) ∀(h, i) ∈ Γ−1 (i) (9.7)
(i,j)∈Γ(i)

The sum on the left may be less than Shi (t), because it is possible that
some vehicles cannot leave (h, i) due to obstructions from a downstream
link or from the node itself (such as a red signal).
5. For each outgoing link, the sum of the turning movement flows into this
link cannot exceed the receiving flow, since that is the maximum number
of vehicles that link can accommodate:
X
yhij (t) ≤ Rij (t) ∀(i, j) ∈ Γ(i) (9.8)
(h,i)∈Γ−1 (i)

The sum on the left may be less than Rij (t), because there may not be
enough vehicles from upstream links to fill all of the available space in the
link.
6. Route choices must be respected. That is, the values yhij (t) must be
compatible with the directions travelers wish to go based on their chosen
paths; we cannot reassign them on the fly in order to increase a yhij value.
7. The first-in, first-out (FIFO) principle must be respected. This is closely
related to the previous property; we cannot allow vehicles to “jump ahead”
in queue to increase a yhij value, unless there is a separate turn lane or
other roadway geometry which can separate vehicles on different routes.
(Figure 9.7).
8. The invariance principle must be respected. If the outflow from a link is
less than its sending flow, then increasing the sending flow further could
not increase the outflow beyond its current value. Likewise, if the inflow
to a link is less than its receiving flow, then increasing the receiving flow
could not increase the inflow to the link beyond its current value. In other
words, if the sending (or receiving) flow is not “binding,” then its specific
value cannot matter for the actual flows.
The invariance principle warrants additional explanation. In the first case,
assume that flow into link (i, j) is restricted by its receiving flow Rij . This
means that more vehicles wish to enter the link than can be accommodated,
so a queue must form on the upstream link. But when there is a queue on the
324 CHAPTER 9. NETWORK LOADING

B B B
B
A Path
Path B
A A

Figure 9.7: The first-in, first-out principle implies that vehicles waiting to turn
will obstruct vehicles further upstream.

hi
upstream link, its sending flow will increase to the capacity qmax ∆t. If the flow
yhij increases in response, then the queue might be cleared in the next time
interval, the sending flow would drop, a new queue would form, and so on, with
the flows oscillating between time intervals. This is unrealistic, and is an artifact
introduced by choosing a particular discretization, not a traffic phenomenon one
would expect in the field. In the second case, if flow onto link (i, j) is restricted
by the sending flow Shi , then there is more space available on link (i, j) than
vehicles wish to enter, a short time later the receiving flow will increase to the
ij
capacity qmax ∆t. If yhij would increase because of this increase in the receiving
flow, again an unrealistic oscillation would be seen. The exercises ask you to
compare some node models which violate the invariance principle to those which
do not.

9.2.1 Links in series

The simplest node to model is one with exactly one incoming link, and exactly
one outgoing link. (Figure 9.8). This may seem like a trivial node, since there
is no real “intersection” here. However, they are often introduced to reflect
changes within a link. For instance, if a freeway reduces from three lanes in
a direction to two lanes, this reduction in capacity and jam density can be
modeled by introducing a node at the point where the lane drops. In this way,
each link can have a homogeneous capacity and jam density, and we can simplify
↑
the notation — in the link models above, we distinguished between qmax and
↓
qmax at the two ends of the link. Now we can just use qmax for the entire link,
9.2. NODE MODEL CONCEPTS 325

h j
Shi Rij

Figure 9.8: Two links in series.

including both ends.

For concreteness, let the incoming link be (h, i) and the outgoing link be
(i, j). In this case, there is only one turning movement, so Ξ(i) = {[h, i, j]}, and
the only turning movement flow we need concern ourselves with is yhij (t). In
the case of two links in series, the formula is simple:

yhij (t) = min{Shi (t), Rij (t)} , (9.9)

that is, the number of vehicles moving from link (h, i) to (i, j) during the t-th
time interval is the lesser of the sending flow from the upstream link in that
time interval, and the receiving flow of the downstream link. For instance,
if the upstream link sending flow is 10 vehicles, while the downstream link
receiving flow is 5 vehicles, a total of 5 vehicles will successfully move from the
upstream link to the downstream one, because that is all there is space for. If the
upstream link sending flow is 3 vehicles and the downstream link receiving flow
is 10 vehicles, 3 vehicles will move from the upstream link to the downstream
one, because that is all the vehicles that are available.
We can check that this node model satisfies all of the desiderata from Sec-
tion 9.2. Going through each of these conditions in turn:

1. The flow yhij (t) is chosen to be the minimum of the upstream sending
flow, and the downstream receiving flow; any value larger than this would
violate either the sending flow constraint or the receiving flow constraint.

2. The sending and receiving flows should both be nonnegative, regardless of

the link model, so yhij (t) is as well.

3. There is only one turning movement, so this constraint is trivially satisfied.

4. The formula for yhij (t) ensures it cannot be greater than the upstream
sending flow. (This condition simplifies since there is only one incom-
ing and outgoing link, so the summation and “for all” quantifier can be
disregarded.)

5. The formula for yhij (t) ensures it cannot be greater than the downstream
receiving flow. (This condition simplifies in the same way.)
326 CHAPTER 9. NETWORK LOADING

h j
Shi Rij

i
Sgi
g

Figure 9.9: Prototype merge node.

6. Route choice is irrelevant when two links meet in series, since all incoming
vehicles must exit by the same link.
7. FIFO is also irrelevant, since all vehicles entering the node behave in the
same way. (This would not be the case if there was more than one exiting
link, and vehicles were on different paths.) So we don’t have to worry
about FIFO when calculating the turning movement flow.
8. To see that the formula satisfies the invariance principle, we have to check
two conditions. If the outflow from (h, i) is less than the sending flow,
this means that yhij (t) = Rij (t), and the second term in the minimum of
equation (9.9) is binding. Increasing the sending flow (the first term in
the minimum) further would not affect its value. Similarly, if the inflow to
(i, j) is less than its receiving flow, this means that yhij (t) = Shi (t), and
the first term in the minimum is binding. Increasing the receiving flow
(the second term) would not affect its value either.

9.2.2 Merges
A merge node has only one outgoing link (i, j), but more than one incoming
link, here labeled (g, i) and (h, i), as in Figure 9.9. This section only concerns
itself with the case of only two upstream links, and generalizing to the case of
additional upstream links is left as an exercise. Here Ξ(i) = {[g, i, j], [h, i, j]},
and we want to calculate the rate of flow from the upstream links to the down-
stream one, that is, the flow rates ygij (t) and yhij (t). As you might expect,
the main quantities of interest are the upstream sending flows Sgi (t) and Shi (t),
and the downstream receiving flow Rij (t). We assume that these values have
already been computed by applying a link model.
For brevity, we will omit the time index in the rest of this section — it is
implicit that all calculations are done with the sending and receiving flows at
the current time step.
There are three possibilities, one corresponding to free flow conditions at the
merge, one corresponding to congestion with queues growing on both upstream
links, and one corresponding to congestion on only one upstream link. For the
merge to be freely flowing, both upstream links must be able to transmit all of
the flow which seeks to leave them, and the downstream link must be able to
9.2. NODE MODEL CONCEPTS 327

accommodate all of this flow. Mathematically, we need Sgi + Shi ≤ Rij , and if
this is true then we simply set ygij = Sgi , and yhij = Shi .
In the second case, there is congestion (so Sgi +Shi > Rij ), and furthermore,
flow is arriving fast enough on both upstream links for a queue to form at each
of them. Empirically, in such cases the flow rate from the upstream links is
approximately proportional to the capacity on these links, that is,

ygij q gi
= max
hi
(9.10)
yhij qmax

A little thought should convince you that this relationship is plausible. Fur-
thermore, in the congested case, all of the available downstream capacity will
be used, so
ygij + yhij = Rij (9.11)
Substituting (9.10) into (9.11) and solving, we obtain
gi
qmax
ygij = gi
Rij (9.12)
qmax hi
+ qmax
with a symmetric expression for yhij .
The third case is perhaps a bit unusual. The merge is congested (Sgi + Shi >
Rij ), but a queue is only forming on one of the upstream links. This may happen
if the flow on one of the upstream links is much less than the flow on the other.
In this case, the proportionality rule allows all of the sending flow from one link
to enter the downstream link, with room to spare. This “spare capacity” can
then be consumed by the other approach. If link (g, i) is the link which cannot
send enough flow to meet the proportionality condition, so that
gi
qmax
Sgi < gi
Rij , (9.13)
qmax hi
+ qmax
then the two flow rates are ygij = Sgi and yhij = Rij − Sgi : one link sends all
of the flow it can, and the other link consumes the remaining capacity. The
formulas are reversed if it is link (h, i) that cannot send enough flow to meet its
proportionality condition.
Exercise 7 asks you to show that the second and third cases can be handled
by the single equation

q gi
 
ygij = med Sgi , Rij − Shi , gi max Rij , (9.14)
hi
qmax + qmax
where med refers to the median of a set of numbers. This formula applies
whenever Sgi + Shi > Rij . An analogous formula holds for the other approach
by swapping the g and h indices. Exercise 8 asks you to verify the desiderata
of Section 9.2 are satisfied by this equation.
For certain merges, it may not be appropriate to assign flow proportional to
the capacity of the incoming links. Rules of the road, signage, or signalization
328 CHAPTER 9. NETWORK LOADING

h j
Shi Rij

i
Rik k

Figure 9.10: Prototype diverge node.

might allocate the capacity of the downstream link differently. In general, the
share of the downstream receiving flow that is allocated to approaches (g, i) and
(h, i) can be written as ψgi and ψhi , respectively, with ψgi + ψhi = 1 and both
of them nonnegative. A more general form of the merge equation can then be
written as
ygij = med {Sgi , Rij − Shi , ψgi Rij } , (9.15)
with a similar formula for (h, i).

9.2.3 Diverges
A diverge node is one with only one incoming link (h, i), but more than one
outgoing link, as in Figure 9.10. This section concerns itself with the case of
only two downstream links. The exercises ask you to generalize to the case of
three downstream links, using the same concepts. Let these two links be called
(i, j) and (i, k), so Ξ(i) = {[h, i, j], [h, i, k]}. Our interest is calculating the rate
of flow from the upstream link to the downstream ones, that is, the flow rates
yhij and yhik . We assume that the sending flow Shi and the receiving flows Rij
and Rik have already been calculated. Unlike links in series or merges, we also
need to represent some model of route choice, since some drivers may choose
link (i, j), and others link (i, k). Let pij and pik be the proportions of drivers
choosing these two links during the t-th time interval, respectively. Naturally,
pij and pik are nonnegative, and pij + pik = 1. Like the sending and receiving
flows, these values can change with time, but to avoid cluttering formulas we
will leave the time indices off of p values unless it is unclear which time step we
are referring to.
There are two possibilities, one corresponding to free flow conditions at the
diverge, and the other corresponding to congestion. What does “free flow”
mean? For the diverge to be freely flowing, both of the downstream links must
be able to accommodate the flow which seeks to enter them. The rates at which
vehicles want to enter the two links are pij Shi and pik Shi , so if both downstream
links can accommodate this, we need pij Shi ≤ Rij and pik Shi ≤ Rik . In this
case we simply have yhij = pij Shi and yhik = pik Shi : all of the flow which wants
to leave the diverge can.
The case of congestion is slightly more interesting, and requires making
assumptions about how drivers will behave. One common assumption is that
9.3. COMBINING NODE AND LINK MODELS 329

flow waiting to enter one link at a diverge will obstruct every other vehicle on the
link (regardless of which link it is destined for). This most obviously represents
the case where the upstream link has only a single lane, so any vehicle which
has to wait will block any vehicle behind it; but this model is commonly used
even in other cases.1 When there is congestion, only some fraction φ of the
upstream sending flow can move. The assumption that any vehicle waiting
blocks every vehicle upstream implies that this same fraction applies to both of
the downstream links, so yhij = φpij Shi and yhik = φpik Shi .
So, how to calculate φ? The inflow rate to a link cannot exceeds its receiving
flow, so yhij = φpij Shi ≤ Rij and yhik = φpik Shi ≤ Rik , or equivalently φ ≤
Rij /pij Shi and φ ≤ Rik /pik Shi . Every vehicle which can move will, so
 
Rij Rik
φ = min , (9.16)
pij Shi pik Shi

Furthermore, we can introduce the uncongested case into this equation as well,
and state  
Rij Rik
φ = min , ,1 (9.17)
pij Shi pik Shi
regardless of whether there is congestion at the diverge or not. Why? If the
diverge is at free flow, then φ = 1, but Rij /pij Shi ≥ 1 and Rik /pik Shi ≥ 1.
Introducing 1 into the minimum therefore gives the correct answer for free flow.
Furthermore, if the diverge is not at free flow, then either Rij /pij Shi < 1 or
Rik /pik Shi < 1, so adding 1 does not affect the minimum value. Therefore, this
formula is still correct even in the congested case. Exercise 13 asks you to verify
the desiderata of Section 9.2 are satisfied by this equation.

9.3 Combining Node and Link Models

This section describes how node and link models are combined, to complete the
network loading process. We must also describe what happens at origins and
destination (zone) nodes. For the algorithm in this section, we assume that the
only links connected to zones are special links called centroid connectors, which
do not represent a specific roadway so much as a collection of small local streets
used by travelers entering or leaving a specific neighborhood. It is common
to give centroid connectors between origins and ordinary nodes a very large
(even infinite) jam density, and centroid connectors between ordinary nodes and
destinations a very large (even infinite) capacity. In both cases the free flow time
should be small. These considerations reflect the ideas that centroid connectors
should not experience significant congestion (or else they should be modeled as
proper links in the network), and simply convey flow from origin nodes and to
destination nodes with as little interference as possible. It is common to forbid
1 For instance, this can represent drivers attempting to “queue jump” by cutting into the

turn lane at the last moment. Or, if the turn lane is long, it may be appropriate to treat the
diverge at the point where the turn lane begins, as opposed to at the physical diverge.
330 CHAPTER 9. NETWORK LOADING

travelers from using centroid connectors except to start and end their trips, that
is, to exclude the use of centroid connectors as “shortcuts.” This can be done
either by transforming the underlying network, having origins and destinations
be distinct nodes adjacent to “one-way” centroid connectors, or by excluding
such paths when finding paths for travelers, as discussed in Chapter 10.
If centroid connectors are set up in this way, then flow entering the network
can simply be added to the upstream ends of their centroid connectors, and flow
leaving the network at destinations can simply vanish, without any constraints
in either case. The network loading algorithm can then be stated as follows:
↑ ↓
1. Initialize all counts and the time index: Nij (0) ← 0 and Nij (0) ← 0 for
all links (i, j), t ← 0.
2. Use a link model to calculate sending and receiving flows Sij and Rij for
all links.
3. Use a node model to calculate transition flows yijk for all nodes j except
for zones.
4. Update cumulative counts: for each non-zone node i, perform
↓ ↓
X
Nhi (t + 1) ← Nhi (t) + yhij (9.18)
(i,j)∈Γ(i)

for each upstream link (h, i), and

↑ ↑
X
Nij (t + 1) ← Nij (t) + yhij (9.19)
(h,i)∈Γ−1 (i)

for each downstream link (i, j).

5. Load trips: for all origins r, let Dr (t) = s∈Z drs (t) be the total demand
P
starting at this node, and for each centroid connector (r, i), let pri be the
↑
fraction of demand beginning their trips on that connector. Set Nri (t +
↑
1) ← Nri (t) + pri Dr (t) for each connector (r, i).
6. Terminate trips at each destination s: for each centroid connector (i, s),
↓ ↓
set Nis (t + 1) ← Nis (t) + Sis .
7. Increment the time: t ← t + 1. If t equals the time horizon T̄ , then stop.
Otherwise, return to step 2.
Figure 9.11 and Table 9.3 illustrate this process on a network with three links
in series, using the spatial queue link model. The link parameters are shown
in Figure 9.11; notice that link (i, j) has a larger capacity at the upstream end
than at the downstream end. The node model at i is the “links in series” model
discussed in Section 9.2.1. The node model at j is a modified version of the “links
in series” node model: when 5 ≤ t < 10, yijs = 0 regardless of the sending and
receiving flows of (i, j) and (j, s); at all other times the “links in series” model is
9.4. ELEMENTARY TRAFFIC FLOW THEORY 331

r i j s
uf = L/t uf = L/(2t) uf = L/t
qmax = 20/t q↑max = 20/t qmax = ∞
k jL = ∞ q↓max = 10/t kjL = ∞
kjL = 40

Figure 9.11: Example for network loading algorithm.

used. This might reflect a red traffic signal, or a closed drawbridge between these
time intervals, since no flow can move through the node. In this example, the
flow downstream is first interrupted when moving from the centroid connector
(r, i) to the link (i, j), whose capacity is lower than the rate at which vehicles
are being loaded at origin r. This can be seen by examining the difference
between the N ↑ and N ↓ values for link (r, i) during the initial timesteps. The
difference between these values gives the total number of vehicles on the link at
that instance in time. Since N ↑ is increasing at a faster rate than N ↓ , vehicles
are accumulating on the link, in a queue at the downstream end. There is no
↑ ↓
queue at node j, because no bottleneck exists there. Compare Nij (t) and Nij (t)
when 3 ≤ t ≤ 5. Both the upstream and downstream count values increase at
the same rate, which means there is no net accumulation of vehicles.
At t = 5, the flow through node j drops to zero, which introduces a further
bottleneck. The impacts of the bottleneck are first seen at t = 6: 40 vehicles are
now on link (i, j), up from 30. As a result, the receiving flow Rij drops to zero,
because the link is full. Therefore, the node model at i restricts any additional
inflow to link (i, j), and the queue on (r, i) grows at an even faster rate than
before, even though the number of new vehicles loaded onto the network has
dropped. At t = 10, the bottleneck at node j is released, and vehicles begin to
move again. At t = 25, the upstream and downstream counts are equal on all
links, which means that the network is empty. All vehicles have reached their
destination.

9.4 Elementary Traffic Flow Theory

This section provides an introduction to the hydrodynamic theory of traffic flow,
the basis of several widely-used link models which are more realistic than the
point or spatial queue models. In this theory, traffic is modeled as a compressible
fluid. Its primary advantage is its simplicity: it can capture many important
congestion phenomena, while remaining tractable for large networks. Of course,
vehicles are not actually molecules of a fluid, but are controlled by drivers with
heterogeneous behavior, who drive vehicles of heterogeneous size, power, and so
on. By treating vehicles as identical particles, we are ignoring such distinctions;
as an example of these limitations, we assume that no overtaking occurs within
a link (because the particles are identical, none of them has any reason to move
CHAPTER 9. NETWORK LOADING

Table 9.3: Example of network loading algorithm.

Link (r, i) Node i Link (i, j) Node j Link (j, s)
↑ ↓ ↑ ↓ ↑ ↓
t drs Rri Nri Nri Sri yrij R ij N ij N ij S ij y ijs Rjs Njs Njs Sjs
0 15 ∞ 0 0 0 0 20 0 0 0 0 ∞ 0 0 0
1 15 ∞ 15 0 15 15 20 0 0 0 0 ∞ 0 0 0
2 15 ∞ 30 15 15 15 20 15 0 0 0 ∞ 0 0 0
3 15 ∞ 45 30 15 10 10 30 0 10 10 ∞ 0 0 0
4 15 ∞ 60 40 20 10 10 40 10 10 10 ∞ 10 0 10
5 15 ∞ 75 50 20 10 10 50 20 10 0 ∞ 20 10 10
6 10 ∞ 90 60 20 0 0 60 20 10 0 ∞ 20 20 0
7 10 ∞ 100 60 20 0 0 60 20 10 0 ∞ 20 20 0
8 10 ∞ 110 60 20 0 0 60 20 10 0 ∞ 20 20 0
9 10 ∞ 120 60 20 0 0 60 20 10 0 ∞ 20 20 0
10 10 ∞ 130 60 20 0 0 60 20 10 10 ∞ 20 20 0
11 10 ∞ 140 60 20 10 10 60 30 10 10 ∞ 30 20 10
12 10 ∞ 150 70 20 10 10 70 40 10 10 ∞ 40 30 10
13 0 ∞ 160 80 20 10 10 80 50 10 10 ∞ 50 40 10
14 0 ∞ 160 90 20 10 10 90 60 10 10 ∞ 60 50 10
15 0 ∞ 160 100 20 10 10 100 70 10 10 ∞ 70 60 10
16 0 ∞ 160 110 20 10 10 110 80 10 10 ∞ 80 70 10
17 0 ∞ 160 120 20 10 10 120 90 10 10 ∞ 90 80 10
18 0 ∞ 160 130 20 10 10 130 100 10 10 ∞ 100 90 10
19 0 ∞ 160 140 20 10 10 140 110 10 10 ∞ 110 100 10
20 0 ∞ 160 150 10 10 10 150 120 10 10 ∞ 120 110 10
21 0 ∞ 160 160 0 0 10 160 130 10 10 ∞ 130 120 10
22 0 ∞ 160 160 0 0 20 160 140 10 10 ∞ 140 130 10
23 0 ∞ 160 160 0 0 20 160 150 10 10 ∞ 150 140 10
24 0 ∞ 160 160 0 0 20 160 160 0 0 ∞ 160 150 10
332

25 0 ∞ 160 160 0 0 20 160 160 0 0 ∞ 160 160 0

9.4. ELEMENTARY TRAFFIC FLOW THEORY 333

faster than another). But seen as a first-order approximation, the hydrodynamic

theory provides a simple and tractable way to model network traffic. It is
also possible to derive certain aspects of the hydrodynamic theory from more
behavioral models representing car-following.
In contrast to the link models we will ultimately use in dynamic traffic
assignment, fluid-based traffic models are formulated in continuous space and
time, so it makes sense to talk about the state of traffic at any point in time
(not just at the integer timesteps 0, 1, . . . , T̄ ). Link models which are based on
fluid models will convert these continuous quantities to discrete ones.

9.4.1 Traffic state variables

We start by modeling a roadway link as a one-dimensional object, using x to
index the distance from the upstream end of the link, and using t to index the
current time. At any point x, and at any time t, the state of traffic can be
described by three fundamental quantities: the flow q, the density k, and the
speed u. Each of these can vary over space and time, so q(x, t), k(x, t), and
u(x, t) can be seen as functions defined over all x values on the link, and all t
values in the analysis period. However, when there is no ambiguity (e.g., only
looking at a single point at a single time), we can simply write q, k, and u
without providing the space and time coordinates.
Both x and t are treated as continuous variables, as are the vehicles them-
selves, allowing derivatives of q, k, and u to be meaningfully defined. This
assumption is imported from fluid mechanics, where the molecules are so small
and numerous that there is essentially no error in approximating the fluid as a
continuum. For traffic flow, this assumption is not so trivial, and is one of the
drawbacks of hydrodynamic models.
Flow is defined as the rate at which vehicles pass a stationary point, and
commonly has units of vehicles per hour. In the field, flow can be measured
using point detectors (such as inductive loops) which record the passage of each
vehicle at a fixed location. Flow can be thought of as a temporal concentration of
vehicles. Density, on the other hand, is defined to be the spatial concentration of
vehicles at a given time, and is measured in vehicles per unit length (commonly
vehicles per kilometer or vehicles per mile). Density can be obtained from taking
a photograph of a link, noting the concentration of vehicles at different locations
at a single instant in time. The speed is the instantaneous rate at which the
vehicles themselves are traveling, and is measured in units such as kilometers per
hour or miles per hour. Speed can be directly measured from radar detectors.
These three quantities are not independent of each other. As a start, there
is the basic relationship
q = uk (9.20)
which must hold at each point and time. (You should check the dimensions of
the quantities in this formula to verify their compatibility.) If this equation is
not evident to you from the definitions of flow, density, and speed, imagine that
we want to know the number of vehicles ∆N which will pass a fixed point over a
334 CHAPTER 9. NETWORK LOADING

Flow

Density
Speed

Figure 9.12: Trajectory diagram illustrating speed, flow, and density.

small, finite time interval ∆t. If the speed of vehicles is u, any upstream vehicle
within a distance of u∆t from the fixed point will pass during the next ∆t time
units, and the number of such vehicles is

∆N = ku∆t . (9.21)

The flow rate is approximately ∆N/∆t, and the formula (9.20) then follows
from taking limits as the time increment ∆t shrinks to zero.
Figure 9.12 is a trajectory diagram showing the locations of vehicles on the
link over time — the horizontal axis denotes time, and the vertical axis denotes
space, with the upstream end of the link at the bottom and the downstream
end at the top. Speed, flow, and density can all be interpreted in terms of
these trajectories. The speed of a vehicle at any point in time corresponds
to the slope of its trajectory there. Flow is the rate at which vehicles pass a
fixed point: on a trajectory diagram, a fixed point in space is represented by
a horizontal line. Time intervals when more trajectories cross this horizontal
line have higher flow, and when fewer trajectories cross this line, the flow is
lower. Density is the spatial concentration of vehicles at a particular instant
in time: on a trajectory diagram, a specific instant is represented by a vertical
line. Where more trajectories cross this vertical line, the density is higher, and
where fewer trajectories cross, the density is lower.
However, this equation by itself is not enough to describe anything of real
interest in traffic flow. Another equation, based on vehicle conservation prin-
ciples, is described in the next subsection. The Lighthill-Whitham-Richards
model, described at the end of this section, makes a further assumption about
the relationships of three state variables. These relationships are enough to
specify the network loading problem as the solution to a well-defined system of
partial differential equations.
9.4. ELEMENTARY TRAFFIC FLOW THEORY 335

x
N=0 1 2
3
4

Figure 9.13: Vehicle trajectories are contours of N (x, t)

9.4.2 Cumulative counts and conservation

The hydrodynamic theory can be simplified by introducing a fourth variable N ,
again defined at each location x and at each time t. This new variable N is
referred to as the cumulative count, and has a similar interpretation to the N ↑
and N ↓ counts defined at the start of Section 9.1, but now applied at any point
on the link, not just the upstream and downstream ends. Imagine that each
vehicle is labeled with a number; for instance, the vehicle at the upstream end of
the link at t = 0 may be labeled as zero. The next vehicle that enters the link is
then labeled as one, the next vehicle as two, and so forth. The numbering does
not have to start at zero. What is important that each entering vehicle be given
consecutive numbers. Then N (x, t) gives the number of the vehicle at location
x at time tW (keeping in mind that we are modeling vehicles as a continuous
fluid, so N (x, t) need not be an integer). The contours of N (x, t) then give the
trajectories of individual vehicles in the traffic stream. (Figure 9.13)
It is worthwhile to point out a potential source of confusion here. In a tra-
jectory diagram like Figure 9.13, space (indexed by x) is conventionally denoted
on the vertical axis, and time (indexed by t) on the horizontal axis. This has
the unfortunate side-effect of making the “x-axis” the vertical one. Second, it
is conventional to list the spatial component before the time component, so a
point like (1, 2) refers to x = 1 and t = 2. This means that on a trajectory
diagram, the vertical component is listed first, as opposed to typical Cartesian
coordinates where the horizontal component is given first. Unfortunately, both
of these conventions are so well-established in the transportation engineering lit-
erature that it is best to simply highlight them and become comfortable using
them.
The quantity N (x, t) is called a cumulative count for the following reason: at
the upstream end of the link (x = 0), the quantity N (0, t) gives the cumulative
number of vehicles which have entered the link up to time t. Furthermore, at
any location x, the difference N (x, t0 ) − N (x, t) gives the number of vehicles
which passed point x between times t and t0 . Taking the limit as t0 approaches
336 CHAPTER 9. NETWORK LOADING

t provides the relationship

∂N
= q, (9.22)
∂t
which holds for every x and t where N is differentiable. That is, at any fixed
spatial location, the rate at which N increases in time is exactly the flow rate
q.
The cumulative counts can also be related to density with a similar argument.
At any point in time t, the difference N (x0 , t) − N (x, t) gives the number of
vehicles which lie between locations x and x0 at time t. However, we have
to be careful regarding the sign convention regarding N described above: since
vehicles are numbered in the order they enter the link, in any platoon of vehicles
N decreases as we move from the following vehicles to the lead vehicle. So, if
x0 > x, then N (x0 , t) ≤ N (x, t). Thus, taking the limit as x0 approaches x, we
must have
∂N
= −k , (9.23)
∂x
which again applies wherever N is differentiable, and where the negative sign
in the formula results from our sign convention.
In this respect, the cumulative counts N can be seen as the most basic
description of traffic flow: if we are given N (x, t) at all points x and times t,
we can calculate q(x, t) and k(x, t) by using equations (9.22) and (9.23), and
therefore u(x, t) everywhere by using (9.20).
Furthermore, wherever the flow and density are themselves continuously
differentiable functions, Clairaut’s theorem states that the mixed second partial
derivatives of N must be equal, that is,

∂2N ∂2N
= , (9.24)
∂x∂t ∂t∂x
so substituting the relationships (9.22) and (9.23) and rearranging, we have
∂q ∂k
+ = 0. (9.25)
∂x ∂t
This is an expression of vehicle conservation, that is, vehicles do not appear
or disappear at any point. This equation must hold everywhere that these
derivatives exist.2 Equations (9.22) and (9.23) are useful in another way. If
(x1 , t1 ) and (x2 , t2 ) are any two points in space and time, the difference in
cumulative count number between these points is given by the line integral
Z
N (x2 , t2 ) − N (x1 , t1 ) = q dt − k dx , (9.26)
C

where C is any curve connecting (x1 , t1 ) and (x2 , t2 ). Because vehicles are
conserved, this line integral does not depend on the specific path taken. This
2 Of course, vehicle conservation must hold even when these derivatives do not exist, it is

just that the formula (9.25) is meaningless there. We have to enforce flow conservation in a
different way at such points.
9.4. ELEMENTARY TRAFFIC FLOW THEORY 337

is helpful, because we can choose a path which is easy to integrate along. In

what follows, we often choose the straight line connecting these two points as
the integration path.
Two issues concerning cumulative counts are often confusing, and are worth
further explanation. First, the cumulative counts can only be meaningfully com-
pared within the same link. Each link maintains its own counts, and the number
associated with a vehicle may change when traveling between links. Instead of
thinking of the cumulative count as being a label permanently associated with a
vehicle, it is better to think about it as a label given to the vehicle by a specific
link, and each link maintains its labels independently of all of the other links.
Each link simply gives successive numbers to each new vehicle entering the link,
without having to coordinate its numbering scheme with other links. This is
needed to ensure that link models can function independently, and because in
complex networks it is usually impossible to assign “permanent” numbers to
vehicles such that any two vehicles entering a link consecutively have consecu-
tive numbers. Second, within any given link, only the difference in cumulative
counts is meaningful, the absolute numbers do not have specific meaning. For
instance, it may be relevant that 10 vehicles entered the link between times 5
and 6, but the specific numbers of these vehicles are not important. A common
convention is to have the first vehicle entering the link be assigned the number
0, the next vehicle the number 1, and so forth, but this is not required. In cases
where there are already vehicles on the link at the start of the modeling period,
the first vehicle entering the link may be assigned a higher number (because the
vehicles already on the link must have lower numbers, and a modeler’s aesthet-
ics may prefer nonnegative vehicle counts). But there would be nothing wrong
with a negative cumulative count either. In this regard, different choices of the
“zero point” are analogous to the different zero points in the Fahrenheit and
Celsius temperature scales: either one will give you correct answers as long as
you are consistent, and a negative number is not necessarily cause for alarm.
For example, consider a link where N (x, t) = 1000t−100x, with t measured in
hours and x measured in miles. The flow at any point and time is ∂N/∂t, which
is a constant of 1000 vehicles per hour. Likewise, the density is −∂N/∂x = 100
vehicles per mile. Using the basic relationship (9.20), the speed must be 10
miles per hour uniformly on the link.
As a more involved example, consider a link which is 1 mile long, with all
times measured in minutes and distances in miles. If we are given that
60x2
N (x, t) = 60t − 120x + , (9.27)
t+1
we can calculate the densities and flows everywhere:
 
∂N x
k(x, t) = − = 120 1 − (9.28)
∂x t+1
 2 !
∂N x
q(x, t) = = 60 1 − . (9.29)
∂t t+1
338 CHAPTER 9. NETWORK LOADING

Exercise 24 asks you to verify that the conservation relationship (9.25) is satisfied
by explicit computation.
Ordinarily, the N (x, t) map is not given — indeed, the goal of network
loading is to calculate it. For if we know N (x, t) everywhere, we can calculate
flow, density, and speed everywhere, using equations (9.22), (9.23), and (9.20).
So let’s assume that we only know the density and flow maps (9.28) and (9.29),
and try to recover information about the cumulative counts. For the given N
map, the vehicle at x = 1/2 at t = 0 has the number 0. (As discussed above,
we do not necessarily have to set the zero point at N (0, 0).) To calculate the
number of the vehicle at x = 1 and t = 1 (the downstream end of the link, one
minute later), we can use equation (9.26).
As this equation involves a line integral, we must choose a path between
(x, t) = (1/2, 0) and (1, 1). Because of the conservation relationship (9.25), we
can choose any path we wish. For the purposes of an example, we will calculate
this integral along three different paths, and verify that they give the same
answer. Figure 9.14 shows the three paths of integration.

Path A : This path consists of the line segment from (1/2, 0) to (1, 0), followed
by the segment from (1, 0) to (1, 1). Because these line segments are
parallel to the axes, this reduces the line integral to two integrals, one
over x alone, and the other over t alone. We thus have
Z Z 1 Z 1
N (1, 1) = q dt − k dx = − k(x, 0) dx + q(1, t) dt
A 1/2 0
Z 1 Z 1  2 !
1
=− 120 (1 − x) dx + 60 1 −
1/2 0 t+1
= −15 + 30 = 15 ,

and the vehicle at the downstream end of the link at t = 1 has the number
15.

Path B : This path consists of the line segment from (1/2, 0) to (1/2, 1), fol-
lowed by the segment from (1/2, 1) to (1, 1). As before, we have
Z Z 1 Z 1
N (1, 1) = q dt − k dx = q(1/2, t) dt − k(x, 1) dx
B 0 1/2
Z 1  2 ! Z 1
1/2  x
= 60 1 − dt − 120 1 − dx
0 t+1 1/2 2
= 52.5 − 37.5 = 15 .

Path C : This path is the line segment directly connecting (1/2, 0) to (1, 1).
Although this line is not parallel to either axis, the integral actually ends
up being the easiest to evaluate, because x/(t + 1) is constant along this
9.4. ELEMENTARY TRAFFIC FLOW THEORY 339

x A (1,1)
C
(1/2,0)
B

Figure 9.14: Three possible paths for the line integral between (1/2, 0) and
(1, 1).

line, and equal to 1/2. Therefore k(x, t) = 60 at all points along this line,
and q(x, t) = 45. Since dx = (1/2)dt on this line segment, we have
Z Z 1 Z 1
N (1, 1) = q dt − k dx = (45dt − 60(1/2)dt) = 15 dt
C 0 0
= 15 .

All three integrals gave the same answer (as they must), which we can verify by
checking N (1, 1) with equation (9.27). So, we can choose whichever integration
path is easiest. In this example, the integrals in Path B involved the most work.
The integral in Path C required a bit more setup, but the actual integral ended
up being very easy, since q and k were constants along the integration path.
Such a path is called a characteristic, and will be described in more detail later
in this chapter.

9.4.3 The Lighthill-Whitham-Richards model

At this point, we have two relationships between the flow, density, and speed
variables: the basic relationship (9.20) and the conservation relationship (9.25).
These first two relationships can be derived directly from the definitions of these
variables, and can describe a wide range of fluid phenomena — at this point
nothing yet has been specific to vehicle flow. The Lighthill-Whitham-Richards
(LWR) model provides a third relationship, completing the hydrodynamic the-
ory.3
Specifically, the LWR model postulates that the flow at any point is a func-
tion of the density at that point, that is,
q(x, t) = Q(k(x, t)) (9.30)
for some function Q. Equivalently, by the relationship (9.20), we can assume
that the speed at any point depends only on the density at that point. It must
3 Lighthill and Whitham published this model in 1955, as the sequel to a paper on flow

in rivers. Richards independently proposed an equivalent model in 1956. All three are now
given credit for this model.
340 CHAPTER 9. NETWORK LOADING

q
qmax
Q(k)

k
kc kj

Figure 9.15: Fundamental diagram for the LWR model.

be emphasized that this relationship, unlike (9.20) and (9.25), is an assumed

behavior and does not follow from basic principles. In dynamic network loading,
we typically assume that this function Q is uniform over space and time on a
link, an assumption we adopt in this section for simplicity. It is possible to
generalize the results in this section when Q varies over space and time.
The function Q is commonly called the fundamental diagram; an example
of such a diagram is shown in Figure 9.15. A few features of fundamental
diagrams are worth noting: they are concave functions4 , and typically assumed
to be continuous and piecewise differentiable. They also have two zeros: one
at k = 0 (zero density means zero flow, because no vehicles are present), and
another at the jam density kj , corresponding to a maximum density where there
is no flow because all vehicles are stopped. At intermediate values of density,
the flow is positive, although for a given flow value q, there can be two possible
density values corresponding to this flow, one corresponding to uncongested
conditions and the other to congested conditions.
If Q is concave, and has two zeros at 0 and kj , then there is some intermediate
point where Q(k) is maximal. This maximal value of Q is called the capacity
of the link, denoted qmax , and the critical density kc is defined to be a value
such that Q(kc ) = qmax . The k values for which Q(k) < qmax and k < kc are
referred to as subcritical, and reflect uncongested traffic flow; the k values for
which Q(k) < qmax and k > kc are supercritical, and reflected congested flow.5
Using q = uk, the speed at any point can be seen as the slope of the secant
line connecting the origin to the point on the fundamental diagram correspond-
4A function f is concave if −f is convex.
5 The definitions of subcritical and supercritical in the transportation field are exactly
opposite to how these terms are used in fluid mechanics. This is a bit annoying, but this
difference in convention reflects differences in the “default state” of flow. Traffic engineers
view the default state of traffic flow as being uncongested, when vehicles can move freely
and (as we show later) shockwaves only travel downstream. Most fluids have to be moving
rather quickly for the same state to occur, and the default state of rivers and many other fluid
systems is a slower rate of travel, where waves can move both upstream and downstream —
what traffic engineers would describe as a “congested” state.
9.4. ELEMENTARY TRAFFIC FLOW THEORY 341

ing to the density at that point. That is, in the LWR model, the density k at
a point completely determines the traffic state. The flow q at that point is ob-
tained from the fundamental diagram Q, and the speed u can then be obtained
from equation (9.31).
To summarize, the three equations relating flow, density, and speed are:
q(x, t) − k(x, t)u(x, t) = 0 (9.31)
q(x, t) − Q(k(x, t)) = 0 (9.32)
∂q ∂k
+ =0 (9.33)
∂x ∂t
and these equations must hold everywhere (with the exception that (9.33) may
not be defined if q or k is not differentiable at a point).
Together with initial conditions (such as the values of k along the link at
t = 0) and boundary conditions (such as the “inflow rates” q at the upstream
end x = 0 throughout the analysis period, or restrictions on q at the downstream
end from a traffic signal), this system of equations can in principle be solved to
yield k(x, t) everywhere. Exercise 24 asks you to verify that the N (x, t) map
used in the example in the previous section is consistent with the fundamental
diagram Q(k) = 14 k(240 − k).
The points where k is not differentiable are known as shockwaves, and often
correspond to abrupt changes in the density. Figure 9.16 shows an example of
several shockwaves associated with the changing of a traffic light. Notice that in
region A, the density is subcritical (uncongested); in region B, traffic is at jam
density; and in region C, traffic is at critical density and flow is at capacity. The
speed of a shockwave can still be determined from conservation principles, even
though the conservation equation (9.33) does not apply because the density and
flow derivatives do not exist at a shock.
Assume that kA and kB are the densities immediately upstream and immedi-
ately downstream of the shockwave (Figure 9.17). The corresponding flow rates
qA = Q(kA ) and qB = Q(kB ) can be calculated from the fundamental diagram,
and finally the speeds are obtained as uA = qA /kA and uB = qB /kB . Further-
more, let uAB denote the speed of the shockwave. Then the speed of vehicles in
region A relative to the shockwave is uA − uAB , and the rate at which vehicles
cross the shockwave from region A is (uA − uAB )kA ; this is nothing more than
equation (9.20) as viewed from the perspective of an observer moving with the
shockwave.
Likewise, the relative speed of the vehicles in region B is uB − uAB , and the
rate at which vehicles cross the shockwave and enter region B is (uB − uAB )kB .
Obviously these two quantities must be equal, since vehicles do not appear or
disappear at the shock. Equating these flow rates from the left and right sides
of the shockwave, we can solve for the shockwave speed:
qA − qB
uAB = . (9.34)
kA − kB
Notice that this calculated speed is the same regardless of whether A is the
upstream region and B the downstream region, or vice versa. This equation also
342 CHAPTER 9. NETWORK LOADING

x uf = 60 mi/hr
D kj = 240 veh/mi
C
qmax = 3600 veh/hr
B
Q(k) = k(240-k)/4
A
t

A B C D

uA = 55 uB = 0 uC = 30 uD undefined
kA = 20 kB = 240 kC = 120 kD = 0
qA = 1100 qB = 0 qC = 3600 qD = 0

Figure 9.16: Shockwaves associated with a traffic signal; vehicle trajectories

indicated in brown and shockwaves in red.

uA kA qA u B k B qB
uAB

A B

Figure 9.17: Flow conservation produces an equation for shockwave speed.

9.4. ELEMENTARY TRAFFIC FLOW THEORY 343

q C
Q(k)

k
D B

Figure 9.18: Identifying shockwave speed on the fundamental diagram.

has a nice geometric interpretation: the speed of the shockwave is the slope of
the line connecting regions A and B on the fundamental diagram (Figure 9.18).
For instance, in Figure 9.16, in region A the flow and density are 1100
vehicles per hour and 20 vehicles per mile, and in region B the flow and density
are 0 vehicles per hour and 240 vehicles per mile. Therefore, using (9.34), the
shockwave between regions A and B has a speed of (1100 − 0)/(20 − 240) =
−5 miles per hour. The negative sign indicates that the shockwave is moving
upstream. Since A represents uncongested traffic, and region B represents the
stopped queue at the traffic signal, the interpretation is that the queue is growing
at 5 miles per hour. Tracing the derivation of equation (9.34), the rate at which
vehicles enter the shockwave is (55 + 5) × 20 from the perspective of region A,
or 1200 vehicles per hour. You should check that the same figure is obtained
from the perspective of region B. Vehicles are entering the queue faster than the
upstream flow rate (1200 vs. 1100 vph) because the queue is growing upstream,
moving to meet vehicles as they arrive.
The system of equations (9.31)–(9.33) can then be solved, introducing shock-
waves as necessary to accommodate the initial and boundary conditions, us-
ing (9.34) to determine their speed. This is the LWR model. However, this
theory does not immediately suggest a technique for actually solving the sys-
tem of partial differential equations, which is the topic of the next subsection.
Notice also that the fundamental diagram determines the maximum speed
at which a shockwave can move. Because the fundamental diagram is concave,
its slope at any point can never be greater than the free-flow speed uf = Q0 (0),
nor less than the slope at jam density Q0 (kj ). The absolute values of these
slopes give the fastest speeds shockwaves can move in the downstream and
upstream directions, respectively, because shockwave speeds are the slopes of
lines connecting points on the fundamental diagram. This leads to an important
notion, the domain of dependence. Consider a point (x, t) in space and time.
Through this point, draw lines with slopes Q0 (0) and Q0 (kj ). The area between
344 CHAPTER 9. NETWORK LOADING

q x C

Fundamental diagram
(x,t) B
A

Slope Q’(kj)

C
Slope Q’(0)

k Past Future t

Figure 9.19: The fundamental diagram (left) and the domain of dependence
(right).

these lines to the left of (x, t) represents all the points which can potentially
influence the traffic state at (x, t) (labeled as region A in Figure 9.19), and the
area between the lines to the right of (x, t) represents all of the points the traffic
state at (x, t) can potentially influence (labeled as region B). In the LWR model,
points outside of these regions (the regions labeled C) are independent of what
happens at (x, t) (say, a signal turning red or green): in the past, they are either
too recent or too distant to affect what is happening at (x, t). In the future, they
are too soon or too distant to be affected by an event at (x, t). This is a crucial
fact for dynamic network loading. If you want to know what is happening at
(x, t), it is sufficient to know what has happened in the past, in the domain
of dependence. We do not need to know what is happening simultaneously at
other points in the network, and as a result we can perform network loading in
an decentralized fashion, performing calculations in parallel since they do not
depend on each other.6

9.4.4 Characteristics and the Newell-Daganzo method

By substituting equation (9.32) into (9.33), and setting aside (9.31) for the
moment, we can obtain a partial differential equation in k alone:
dQ ∂k ∂k
+ =0 (9.35)
dk ∂x ∂t
It is in this form that the LWR model is most frequently solved, obtaining k(x, t)
values everywhere. This type of partial differential equation can be approached
using the method of characteristics, which is briefly described below.
A characteristic is a curve in (x, t) space, along which the density k(x, t)
varies in a predictable way. If the fundamental diagram Q does not vary in space
6 If you have studied relativistic physics, there are many similarities with the notion of light

cones and causality.

9.4. ELEMENTARY TRAFFIC FLOW THEORY 345

and time, then straight lines form characteristics of (9.35). To demonstrate,

assume that we know the density k(x1 , t1 ) at some point, and consider a straight
line C through (x1 , t1 ) with some slope v. Then the directional derivative of k
along this line is given by
 ∂k ∂k
k 0 = ∇k · v

1 = v+ . (9.36)
∂x ∂t
Substituting the conservation equation (9.35) and rearranging, we have
 
dQ ∂k
k0 = v − . (9.37)
dk ∂x
This gives us a formula for the change in density along the line. But what is
really useful is noticing that if v = dQdk , then (9.37) vanishes and the density is
constant along the line. Put another way, the density is constant along any line
whose slope is equal to the slope of the tangent of the fundamental diagram at
that density value.7
To see this, return to the example we have been using where Q(k) = 14 k(240−
2
k) and N (x, t) = 60t−120x+ 60x t+1 . We have calculated the density and flow at all
points and times in equations (9.28) and (9.29). In our calculation, we observed
that the line integral along path C was the simplest to evaluate, because k(x, t)
was a constant 60 vehicles per mile along the line x/(t + 1) = 1/2. This path is
in fact a characteristic: the derivative of the fundamental diagram at this point
is Q0 (60) = 60 − k/2 = 30 miles per hour. Rearranging the equation of the
line, we have x = t/2 + 1/2; in other words, this is a line whose location moves
half a mile in one minute: 30 miles per hour, the same as the derivative of the
fundamental diagram.
These slopes will be different at points where the density is different, which
means that characteristic lines can potentially intersect. This indicates the pres-
ence of a shockwave separating regions of different density. One interpretation
of these characteristic lines is as “directions of influence,” since the density at a
point will determine the density at all later points along this line. In uncongested
regions, where k is subcritical, dQ/dk is positive, meaning that the character-
istics have positive slope: uncongested states will propagate downstream. In
congested regions with supercritical density, dQ/dk is negative, and the charac-
teristics have negative slope: congested states will propagate upstream.
Furthermore, by combining knowledge of characteristics with equation (9.26),
we can determine the cumulative count and the density at any point, given
sufficient initial and boundary data. This method was first suggested by G. F.
Newell, and developed further by Carlos Daganzo. Suppose we wish to calculate
the density at a point (x, t). If we knew the density at this point, then we would
know the slope of the characteristic through this point. This characteristic could
be traced back until it intersected a point (x0 , t0 ) where the cumulative count
7 Note that the speed of the characteristic is different from the speed of the vehicles them-

selves, or the speed of shockwaves (which are given by slopes of secant lines on the fundamental
diagram).
346 CHAPTER 9. NETWORK LOADING

N was known, either because it corresponds to an initial or boundary point, or

a point where N has already been calculated. Then, applying (9.26), we would
have Z
N (x, t) = N (x0 , t0 ) + q dt − k dx (9.38)
C

where C is the straight line between (x0 , t0 ) and (x, t). Since this line is a
characteristic, we have dx/dt = dQ/dk, and so
Z  
dQ
N (x, t) = N (x0 , t0 ) + q−k dt . (9.39)
C dk

Furthermore, as a characteristic, k (and therefore q) are constant. So the integral

is easy to evaluate, and we have
 
dQ
N (x, t) = N (x0 , t0 ) + q − k (t − t0 ) . (9.40)
dk

The only trouble is that we do not actually know the density at (x, t). Each
possible value of density corresponds to a slightly different cumulative count,
based on equation (9.40). The insight of the Newell-Daganzo method is that
the correct value of the cumulative count is the lowest possible value. That
is, imagine that the density at (x, t) is k, and let (xk , tk ) be the known point
corresponding to the characteristic slope of density k. Then
   
dQ
N (x, t) = inf N (xk , tk ) + q − k (t − tk ) . (9.41)
k∈[0,kj ] dk

Rigorously validating this insight requires knowledge of the calculus of vari-

ations, which is beyond the scope of this text. An intuitive justification is
that (9.40) represents an upper bound on the cumulative count imposed by a
boundary or initial point — we know the number of the vehicle passing that
boundary or initial point, and (9.40) expresses one possible value of the cumu-
lative count at the point in question, if the density took a particular value and
no shockwave intervened. However, due to the possible presence of shockwaves,
there may be another, more restrictive constraint imposed on the cumulative
count. The equation (9.41) thus finds the “most restrictive” boundary or initial
condition, which gives the correct cumulative count.
Better yet, this method becomes exceptionally easy if we assume that the
fundamental diagram takes a simple form, such as a triangular shape (Fig-
ure 9.20), where the equation is given by

Q(k) = min {uf k, w(kj − k)} . (9.42)

In this case, there are only two possible characteristic speeds: one (+uf ) cor-
responding to uncongested conditions, and the other (−w) corresponding to
congested conditions. The uncongested speed uf is the free-flow speed, and w
is known as the backward wave speed. In the case where the characteristic speed
9.4. ELEMENTARY TRAFFIC FLOW THEORY 347

q
Q(k)

u -w

k
kj

Figure 9.20: A triangular fundamental diagram, defined by three parameters

uf , w, and kj .

is uf , the vehicle speed u equals the characteristic speed uf , and both are equal
to q/k. Therefore, the line integral along the characteristic in (9.40) is

(q − kuf ) (t − t0 ) = (q − uk) (t − t0 ) = 0 (9.43)

because q = uk.
In other words, the vehicle number is constant along characteristics at free-
flow speed. For the congested characteristic with slope −w, we have
q 
(q − k(−w)) (t − t0 ) = w + k (t − t0 ) = kj w(t − t0 ) (9.44)
w
since k + wq = kj , as can be seen in Figure 9.20. This expression can also be
written as kj (x0 − x). Since these are the only two characteristics which can
prevail at any point, equation (9.41) gives

N (x, t) = min {N (xU , tU ), N (xC , tC ) + kj (xC − x)} (9.45)

where (xU , tU ) is the known point intersected by the uncongested characteristic,

and (xC , tC ) is the known point intersected by the congested characteristic.
Trapezoidal fundamental diagrams, such as that in Figure 9.21, are also
commonly used, given by the equation

Q(k) = min {uf k, qmax , w(kj − k)} . (9.46)

In this case, there is a third possible characteristic speed, corresponding to the

flat region where flow is at capacity. Since the derivative of the fundamental
diagram at this point is zero, this characteristic has zero speed, represented by
a horizontal line on a space-time diagram. Tracing this characteristic back to
a known point (x, tR ), the change in vehicle number between this known point
and the unknown point (x, t) is just

qmax (t − tR ) (9.47)
348 CHAPTER 9. NETWORK LOADING

q
qmax Q(k)

-w
u

k
kj

Figure 9.21: A trapezoidal fundamental diagram, defined by four parameters

uf , qmax , w, and kj .

since dx = 0 in the line integral (9.26). (Note that the location of this third
known point is the same as the location of the point we are solving for, since the
characteristic is stationary.) This adds a third term to the minimum in (9.45),
giving

N (x, t) = min {N (xU , tU ), N (x, tR ) + qmax (t − tR ), N (xC , tC ) + kj (xC − x)}

(9.48)
for trapezoidal fundamental diagrams.
The trapezoidal fundamental diagram requires four parameters to calibrate:
the free-flow speed uf , the capacity qmax , the jam density kj , and the backward
wave speed −w. The first three of these are fairly straightforward to estimate
from traffic engineering principles. The backward wave speed is a bit trickier;
empirically it is often a third to a half of the free-flow speed.
As a demonstration of this method, consider a link which is 1 mile long.
The fundamental diagram is shown in the left panel of Figure 9.22, and has the
equation  
1
Q(k) = min k, 120 − k (9.49)
2
when flow is measured in vehicles per minute, and density in vehicles per mile.
Initially, vehicles on the link flow at an uncongested 48 veh/min — this state has
existed for a long time in the past, and vehicles continue entering the link at this
rate. However, at the downstream end of the link there is an obstruction which
prevents any vehicles from passing, such as a red light or an incident blocking
all lanes. This will cause a queue of stopped vehicles to form. Assume that we
want to know how many vehicles lie between the obstruction and three given
points: half a mile upstream of the obstruction, 30 seconds after it begins; an
eighth of a mile upstream of the obstruction, 30 seconds after it begins; and an
eighth of a mile upstream of the obstruction, one minute after it begins. We do
not know, a priori, whether these points lie within the queue or in the portion
of the link which is still uncongested.
9.4. ELEMENTARY TRAFFIC FLOW THEORY 349

q x (mi)
(veh/min) 1 I G
7/8
3/4 C B
E
80 1/2 A
3/8
k H
(veh/mi) D F t (min)
80 240 1/8 1/4 1/2 3/4 1

Figure 9.22: Example of the Newell-Daganzo method.

The first step is to establish a coordinate system. As always, we set x = 0

at the upstream end of the link. For this problem, it will be convenient to set
t = 0 at the time when the obstruction begins, and to count vehicles starting
from the first vehicle stopped at the obstruction, that is, N (1, 0) = 0. This
way, N (x, t) will immediately give the number of vehicles between point x and
the obstruction at time t. Next, we use the given data from the problem to
construct initial and boundary conditions where we already know N (x, t). Since
no vehicles can pass the obstruction, we know that N (1, t) = 0 for all t. Since
the link is initially uncongested at a flow rate of 48 veh/min, the fundamental
diagram (9.49) gives the initial density to be 48 veh/mi. Therefore, the initial
condition is N (x, 0) = 48 − 48x, and the vehicle number at the origin of the
coordinate system is 48. Since vehicles continue to enter the link at a rate of 48
veh/min, we have N (0, t) = 48 + 48t along the upstream boundary of the link.
The three points where we must calculate N (x, t) are labeled as A, B, and C in
the right panel of Figure 9.22.
We start with point A, half a mile upstream of the obstruction and 30
seconds after it begins. There are two possible characteristics at this point,
one with slope +1 (corresponding to uncongested conditions) and one with
slope − 21 (corresponding to congested conditions). We can trace back these
characteristics until they reach a point where N (x, t) is known, in this case
an initial or boundary condition. These points of intersection are labeled D
and E in Figure 9.22. From the initial condition, we know that N (D) = 48
and N (E) = 12. Along the uncongested characteristic, there is no change in
the cumulative count, while along the congested characteristic the cumulative
count increases at a rate of kj = 240 veh/mi for each mile traveled. Therefore,
equation (9.43) tells us that N (A) = 48+0 = 48 if point A is uncongested, while
equation (9.44) tells us that N (A) = 12+ 14 240 = 72 if point A is congested. The
correct value is the smaller of the two: N (A) = 48, and this point is uncongested
(the queue has not yet reached this point).
We next move to point B, an eighth of a mile upstream of the obstruction
and 60 seconds after it begins. Tracing back the two possible characteristics
from point B leads us to the points labeled F and G, and from the boundary
350 CHAPTER 9. NETWORK LOADING

conditions we know N (F ) = 54 and N (G) = 0. So N (B) is the lesser of

N (F ) + 0 = 54 and N (G) + 81 240 = 30. This means that there are 30 vehicles
between point B and the obstruction, and since the congested characteristic
produced the lower value, point B lies within the queue.
At point C, the two characteristics lead to the points labeled H and I, where
N (H) = 30 from an initial condition and N (I) = 0 from a boundary condition.
So N (C) = min{30 + 0, 0 + 18 240} = 30. In this case, both characteristics led
to the same value of N (C). This indicates that the shockwave passes exactly
through point C.
An alternative approach for this problem would be to explicitly calculate
the location of the shockwave, determine the densities in each region, and ap-
ply (9.26) directly. For this problem, this approach would be simpler than the
Newell-Daganzo method. However, if there were multiple shockwaves intro-
duced into the problem (say, from multiple red/green cycles), it would become
very tedious to track the locations of all of the shockwaves and determine which
region the points lie in. The Newell-Daganzo method can be applied just as
easily in such a case, once the boundary conditions are determined.

9.5 LWR-based Link Models

The hydrodynamic traffic flow model developed by Lighthill, Whitham, and
Richards forms the basis for several popular link models. The LWR model
is simple enough to be usable in large-scale dynamic network loading, while
capturing enough key properties of traffic flow for its results to be meaningful.
Through shockwaves, we can capture how congestion grows and shrinks over
time. These shockwaves allow us to model queue spillback (when a congestion
shockwave reaches the upstream end of a link) and to account for delays in queue
startup (unlike the spatial queue model). This section describes two link models
based on the LWR model — the cell transmission model, which is essentially an
explicit solution scheme for the LWR system of partial differential equations,
and the link transmission model, which uses the Newell-Daganzo method to
directly calculate sending and receiving flows. Lastly, we show how the point
and spatial queue models can be seen as special cases of LWR-based link models,
with a suitable choice of the fundamental diagram.

9.5.1 Cell transmission model

In the cell transmission model, in addition to discretizing time into intervals
of length ∆t, we also discretize space, dividing the link into cells of length ∆x.
These two discretizations are not chosen independently. Rather, they are related
by
∆x = uf ∆t , (9.50)
that is, the length of each cell is the distance a vehicle would travel in time ∆t
at free flow. The reasons for this choice are discussed at the end of this section.
9.5. LWR-BASED LINK MODELS 351

t
3 3 3 2 2 1 2 3
3 4 4 4 2 2 4 5
4 4 4 4 5 4 3 1

Figure 9.23: Discretizing space and time into cells, n(x, t) values below.

With this discretization in mind, we will use the notation n(x, t) to describe
the number of vehicles in cell x at time t, where x and t are both integers —
we must convert the continuous LWR variables k and q into discrete variables
for dynamic network loading, which we will call n and y.
If the cell size is small, we can make the approximation that

n(x, t) ≈ k(x, t)∆x , (9.51)

essentially assuming that the density within the cell is constant. Further define
y(x, t) to be the number of vehicles which enter cell x during the t-th time
interval. Making a similar assumption, we can make the approximation

y(x, t) ≈ q(x, t)∆t . (9.52)

In a space-time diagram showing vehicle trajectories, such as Figure 9.23, n(x, t)

and y(x, t) respectively correspond to the number of trajectories crossing the
vertical line at t between locations x and x + ∆x, and the number of trajectories
crossing the horizontal line at x between times t and t + ∆t.
The cell transmission model provides methods for solving for n(x, t) for all
integer values of x and t; these can then be converted to density values k(x, t)
through (9.51). These density values can then be used to calculate flows q(x, t)
through the fundamental diagram, and u(x, t) values through equation (9.20),
finally providing an approximate solution to the system of partial differential
equations (9.31)–(9.33). Recall that the goal of a link model is to determine
352 CHAPTER 9. NETWORK LOADING

y(1,t) y(2,t)
3 0
5 4 7
n(0,t) n(1,t) n(2,t)
x=0 1 2 3

Figure 9.24: Where discrete values are calculated in the cell transmission model.

sending and receiving flow at each time step. For this reason, we will be con-
tent with determining how the n(x, t + ∆t) values can be calculated, given the
n(x, t) values (which are already known), and the y(x, t) values, which must be
calculated.
Since q(x, t) = Q(k(x, t)), substitution into equations (9.51) and (9.52) give
 
n(x, t)
y(x, t) ≈ Q ∆t . (9.53)
∆x

Substituting the particular form of the fundamental diagram gives an equation

for y(x, t) in terms of n(x, t). Using the trapezoidal diagram of Figure 9.21, we
have
  
∆t n(x, t)
y(x, t) ≈ min uf n(x, t) , qmax ∆t, w∆t kj − . (9.54)
∆x ∆x

Using the fact that ∆x/∆t = uf , the first term in the minimum is simply n(x, t).
The third term can be simplified by defining n̄(x) = kj ∆x to be the maximum
number of vehicles which can fit into a cell and δ = w/uf to be the ratio between
the backward wave speed and free-flow speed. Then, factoring out 1/∆x from
the term in parentheses and again using ∆x/∆t = uf , the third term simplifies
to δ(n̄(x) − n(x, t)).
There is one more point which is subtle, yet incredibly important. Being a
“flow” variable, y(x, t) is calculated at a single point (over a time interval), while
n(x, t) is calculated at a single time (over a longer spatial interval). As shown
in Figure 9.24, the x in y(x, t) refers to a single location, while the x in n(x, t)
refers to an entire cell. So, when we are calculating the flow across the (single)
point x, which is the boundary between two cells, do we look at the adjacent
cell upstream n(x − 1, t), or the adjacent cell downstream n(x, t)?
The correct answer depends on the fundamental diagram, and the meaning of
characteristics. In uncongested conditions, corresponding to the increasing part
of the fundamental diagram and the first term in the minimum, the traffic state
moves from upstream to downstream (because the characteristic has positive
speed). In congested conditions, corresponding to the decreasing part of the
fundamental diagram and the third term in the minimum, the traffic state moves
from downstream to upstream (because the characteristic has negative speed.)
9.5. LWR-BASED LINK MODELS 353

So, if traffic is uncongested at the (single) point x, we need to refer to the

upstream cell, while if traffic is congested at the (single) point x, we must refer
to the downstream cell. So, the final expression for the cell transmission model
flows is
y(x, t) = min{n(x − ∆x, t), qmax ∆t, δ (n̄(x) − n(x, t))} . (9.55)
This expression also has a nice intuitive interpretation. The number of vehicles
moving from cell x − 1 to cell x is limited either by the number of vehicles in the
upstream cell (the first term), the capacity of the roadway (the second term),
or by the available space in the downstream cell (the third term).
Expression (9.55) is the discrete equivalent of the differential equation (9.32).
We now derive the discrete form of the partial differential equation giving the
∂q
conservation law (9.33). The derivative ∂x can be approximated as

∂q 1
(x, t) ≈ (y(x + ∆x, t) − y(x, t)) (9.56)
∂x ∆t∆x
∂k
and the derivative ∂t can be approximated as

∂k 1
(x, t) ≈ (n(x, t + ∆t) − n(x, t)) . (9.57)
∂t ∆t∆x
Substituting into (9.33), we have

1
(y(x + ∆x, t) − y(x, t) + n(x, t + ∆t) − n(x, t)) = 0 (9.58)
∆t∆x
or, in a more convenient form,

n(x, t + ∆t) = n(x, t) + y(x, t) − y(x + ∆x, t) (9.59)

This also has a simple intuitive interpretation: the number of vehicles in cell
x at time t + 1 is simply the number of vehicles in cell x at the previous time
t, plus the number of vehicles which flowed into the cell during the t-th time
interval, minus the number of vehicles which left.
Together, the equations (9.55) and (9.59) define the cell transmission model
for trapezoidal fundamental diagrams. There are only two pieces of “miss-
ing” information, at the boundaries of the link. Refer again to Figure 9.24.
How should y(0, t) and y(L, t) be calculated? For y(0, t), the first term in for-
mula (9.55) involves n(−∆x, t), while for y(L, t), the third term in the formula
involves n(L + ∆x, t), and both of these cells are “out of range.” The answer is
that these boundary flows are used to calculate the sending and receiving flows
for the link, and a node model will then give the actual values link inflows y(0, t)
and link outflows y(L, t).
Remember that the sending flow is the maximum number of vehicles which
could leave the link if there was no obstruction from downstream. In terms
of (9.55), this means that the third term in the minimum (which corresponds to
downstream congestion) is ignored. Then, the first two terms in the minimum
354 CHAPTER 9. NETWORK LOADING

(which only refer to cells on the link) are the possible restrictions on the flow
leaving the link, so
S(t) = min{n(C, t), qmax ∆t} . (9.60)

Likewise, the receiving flow is the maximum number of vehicles which could
enter the link if there were a large number of vehicles wanting to enter from
upstream. In terms of (9.55), this means that the first term in the minimum
(which corresponds to the number of vehicles wanting to enter) is ignored. The
second two terms in the minimum refer to cells on the link, and

R(t) = min{qmax ∆t, δ (n̄(0) − n(0, t))} . (9.61)

It remains to explain the choice of the discretization (9.50). An intuitive

explanation for linking the cell length and time discretizations in this way is
that this choice limits vehicles to moving at most one cell between time steps.
In fact, in uncongested conditions, all vehicles in a cell will move to the next cell
downstream in the next time interval, simplifying calculations — for instance,
the simplification in (9.54) only works because of the choice made in (9.50).
The underlying mathematical reason has to do with the speed of characteris-
tics, which for the trapezoidal fundamental diagram lie between −w and +uf .
Empirically, w < uf , so the fastest moving characteristic (in either direction) is
one with speed uf . More important than vehicles moving at most one cell be-
tween time steps is that characteristics cannot move more than one cell between
time steps, in either the upstream or downstream directions. This condition
is needed for stability of finite-difference approximations to partial differential
equations, and corresponds to the Courant-Friedrich-Lewy (CFL) condition for
explicit solution methods such as the cell transmission model.
Table 9.4 provides a demonstration of how the cell transmission model works.
In this example, a link is divided into three cells, implying that it takes three
time steps for a vehicle to traverse the link at free-flow. The fundamental
diagram is such that at most 10 vehicles can move between cells in one time
step, at most 30 vehicles can fit into one cell at jam density, and the ratio
between the backward and forward characteristic speeds is δ = 2/3. There is a
red light at the downstream end of the link which turns green at t = 10, and
remains green thereafter. In this table, d(t) represents the number of vehicles
that wish to enter the link during the t-th timestep (perhaps the sending flow
from an upstream link), while R(t) is the receiving flow for the link, calculated
from (9.61). The middle columns of the table show the main cell transmission
model calculations: the number of vehicles in each cell at the start of each
timestep, N (x, t), and the number of vehicles moving into each cell during the
t-th time step. These values are calculated from (9.59) and (9.55), respectively,
along with the initial condition that the link is empty, that is, N (x, 0) = 0 for all
x. The rightmost columns of the table show the link’s sending flow, calculated
from (9.60), and the actual flow which leaves the link, denoted y(3, t). This
latter value is constrained to be zero as long as the light at the downstream end
of the link is red.
9.5. LWR-BASED LINK MODELS 355

Notice that the table has non-integer values: we do not need to round cell oc-
cupancies and flows to whole values, since the LWR model assumes vehicles are
a continuously-divisible fluid. Preserving non-integer values also ensures that
the cell transmission model remains accurate no matter how small the timestep
∆t is (in fact, its accuracy should increase as this happens). Insisting that flows
and occupancies be rounded to whole numbers can introduce significant error if
the timestep is small, unless one is careful with implementation.
Table 9.5 shows only the cell occupancies at each timestep, color-coded ac-
cording to the density in the cells. In this example, the link is initially at
free-flow, until the first vehicles encounter the red light and must stop. A queue
forms, and a shockwave begins moving backward. When this shockwave reaches
the cell at the upstream end of the link, the receiving flow of the link decreases,
and the inflow to the link is limited. When the light turns green, a second
shockwave begins moving backward as the queue clears. Once this shockwave
overtakes the first, vehicles can begin entering the link again. For a few time
steps, the inflow is greater than d(t), representing demand which was blocked
when the receiving flow was restricted by the queue and which was itself queued
on an upstream link (the “queue spillback” phenomenon). Unlike the point
queue and spatial queue link models, the cell transmission model tells us what
is happening in the interior of a link, not just at the endpoints. This is both
a blessing and a curse: sometimes this additional information is helpful, while
other times we may not be concerned with such details. The link transmission
model, described next, can simplify computations if we do not need information
on the internal state of a link.

9.5.2 Link transmission model

The link transmission model allows us to calculate sending and receiving flows
for links with any trapezoidal fundamental diagram.8 In contrast to the cell
transmission model, it only involves calculations at the ends of the links — de-
tails of what happen in the middle of the link are ignored. As a consequence, the
link transmission model does not require us to keep track of information within
the link. However, it does require us to keep track of information on the past
state of the link, whereas the cell transmission model calculations only involve
quantities at the current time step, and the previous time step. The link trans-
mission model can also overcome the “shock spreading” phenomenon, where
backward-moving shockwaves in the cell transmission model can diffuse across
multiple cells, even though they are crisp in the LWR model. (See Exercise 29.)
Assume that the trapezoidal fundamental diagram is parameterized as in
Figure 9.21. There are three characteristic speeds, +uf at free-flow, 0 at capac-
ity flow, and −w at congested flow. As with the point queue and spatial queue
models, we apply the Newell-Daganzo method to calculate sending and receiv-
ing flows. We will only need to refer to cumulative counts N at the upstream
8 In fact, it can be generalized to any piecewise linear fundamental diagram without too

much difficulty; see Exercise 34.

356 CHAPTER 9. NETWORK LOADING

Table 9.4: Cell transmission model example, on a link with three cells, qmax ∆t =
10, n̄ = 30, and δ = 23
Cell 0 Cell 1 Cell 2
t d(t) R(t) y(0, t) N (0, t) y(1, t) N (1, t) y(2, t) N (2, t) S(t) y(3, t)
0 10 10 10 0 0 0 0 0 0 0
1 10 10 10 10 10 0 0 0 0 0
2 10 10 10 10 10 10 10 0 0 0
3 10 10 10 10 10 10 10 10 10 0
4 10 10 10 10 10 10 6.7 20 10 0
5 9 10 9 10 10 13.3 2.2 26.7 10 0
6 8 10 8 9 5.9 21.1 0.7 28.9 10 0
7 7 10 7 11.1 2.5 26.3 0.2 29.6 10 0
8 6 9.6 6 15.6 1 28.5 0.1 29.9 10 0
9 5 6.3 5 20.6 0.4 29.4 0 30 10 0
10 4 3.2 3.2 25.2 0.1 29.8 0 30 10 10
11 3 1.2 1.2 28.3 0.1 29.9 6.7 20 10 10
12 2 0.4 0.4 29.4 4.5 23.3 8.9 16.7 10 10
13 1 3.1 3.1 25.3 7.4 18.9 9.6 15.6 10 10
14 0 6 1.3 21.0 8.9 16.7 9.9 15.2 10 10
15 0 10 0 13.4 9.5 15.7 10 15.1 10 10
16 0 10 0 3.9 3.9 15.3 10 15 10 10
17 0 10 0 0 0 5.8 9.2 15 10 10
18 0 10 0 0 0 0 0 14.2 10 10
19 0 10 0 0 0 0 0 4.2 4.2 4.2
20 0 10 0 0 0 0 0 0 0 0
9.5. LWR-BASED LINK MODELS 357

Table 9.5: Cell occupancies from the example in Table 9.4, green is zero density
and red is jam density.
358 CHAPTER 9. NETWORK LOADING

and downstream ends of the link, that is, at N (0, ·) and N (L, ·), respectively.
In keeping with the notation introduced in Section 9.1, we will refer to these as
N ↑ and N ↓ . If there is no obstruction from a downstream link or node, then
the end of the link will be uncongested, and the characteristic at this end will
either have slope +uf or slope zero. The Newell-Daganzo method thus gives

N ↓ (t + ∆t) = min{N ↑ ((t + ∆t) − L/uf ), N ↓ (t) + qmax ∆t} (9.62)

and the equation for the sending flow is obtained as the difference between
N ↓ (t + 1) and N ↓ (t):

S(t) = min{N ↑ ((t + ∆t) − L/uf ) − N ↓ (t), qmax ∆t} . (9.63)

For the receiving flow, we have to take into account the two relevant charac-
teristic speeds of 0 and −w, since the receiving flow is calculated assuming an
inflow large enough that the upstream end of the link is congested (or at least
at capacity). The stationary characteristic corresponds to the known point
N (0, t), while the backward-moving characteristic corresponds to the known
point N (L, t − L/w). Thus, applying the last two terms of equation (9.46)
would give

N ↑ (t + ∆t) = min{N ↑ (t) + qmax ∆t, N ↓ ((t + ∆t) − L/w) + kj L} (9.64)

and

R(t) = min{qmax ∆t, (N ↓ ((t + ∆t) − L/w) + kj L) − N ↑ (t)} . (9.65)

It is possible to show that equations (9.63) and (9.65) ensure that the number
of vehicles on the link is always nonnegative, and less than kj L.
The link transmission model is demonstrated on an example similar to the
one used for the cell transmission model; the only difference is that the ratio
of backward-to-forward characteristics has been adjusted from 2/3 to 3/4. In
particular, L/uf = 3∆t, and L/w = 4∆t, so forward-moving characteristics
require three time steps to cross the link, and backward-moving characteristics
require four time steps. The total number of vehicles which can fit on the
link is kj L = 90. Otherwise, the example is the same: the demand profile is
identical, and a red light prevents outflow from the link until t = 10. The
results of the calculations are shown in Table 9.6. The rightmost column shows
the number of vehicles on the link, which is the difference between the upstream
and downstream cumulative counts at any point in time. Notice that inflow to
the link is completely blocked during the 12th and 13th time intervals, even
though the number of vehicles on the link is less than the jam density of 90.
This happens because the queue has started to clear at the downstream end, but
the clearing shockwave has not yet reached the upstream end of the link. The
vehicles at the upstream end are still stopped, and no more vehicles can enter.
In contrast, the spatial queue model would allow vehicles to start entering the
link as soon as they began to leave.
9.5. LWR-BASED LINK MODELS 359

Table 9.6: Link transmission model example, with L/uf = 3∆t, L/w = 4∆t,
and kj L = 90.
t d(t) R(t) Inflow N ↑ (t) N ↓ (t) S(t) Outflow Vehicles on link
0 10 10 10 0 0 0 0 0
1 10 10 10 10 0 0 0 10
2 10 10 10 20 0 0 0 20
3 10 10 10 30 0 10 0 30
4 10 10 10 40 0 10 0 40
5 9 10 9 50 0 10 0 50
6 8 10 8 59 0 10 0 59
7 7 10 7 67 0 10 0 67
8 6 10 6 74 0 10 0 74
9 5 10 5 80 0 10 0 80
10 4 5 4 85 0 10 10 85
11 3 1 1 89 10 10 10 79
12 2 0 0 90 20 10 10 70
13 1 0 0 90 30 10 10 60
14 0 10 5 90 40 10 10 50
15 0 10 0 95 50 10 10 45
16 0 10 0 95 60 10 10 35
17 0 10 0 95 70 10 10 25
18 0 10 0 95 80 10 10 15
19 0 10 0 95 90 5 5 5
20 0 10 0 95 95 0 0 0
360 CHAPTER 9. NETWORK LOADING

q
qmax
Q(k)

Figure 9.25: Fundamental diagram in the point queue model.

9.5.3 Point and spatial queues and the LWR model (*)
(This optional section shows how the previously-introduced point and spatial
queue models can be seen as special cases of the LWR model.)
The first link models introduced in this chapter were the point and spatial
queue models, in Section 9.1. These were presented as simple link models to
illustrate concepts like the sending and receiving flow, rather than realistic de-
pictions of traffic flow. Nevertheless, it is possible to view the point and spatial
queue models as special cases of the LWR model by making an appropriate
choice of the fundamental diagram, as shown in this section. Applying the
Newell-Daganzo method with this fundamental diagram gives us a second way
to derive the expressions for sending and receiving flow for these models.
The point queue model is equivalent to assuming that the flow-density rela-
tionship is as shown in Figure 9.25. This diagram is unlike others we’ve seen,
because there is no jam density. This represents the idea that the point queue
occupies no physical space: no matter how many vehicles are in queue, there
is nothing to prevent additional vehicles from entering the link and joining the
queue. It is also the simplest diagram which we have seen so far, and is defined
by only two parameters: the free-flow speed uf and the capacity qmax . (Even
the triangular fundamental diagram in Figure 9.20 required a third parameter,
either −w or kj .)
The Newell-Daganzo method leads to a simple expression for the sending
and receiving flows in a point queue model. To calculate the sending flow S(t),
we need to examine the downstream end of the link, so x = L. Since we are
solving in increasing order of time, we already know N ↓ (0), N ↓ (∆t), . . . , N ↓ (t)
(the number of vehicles which have left the link at each time interval). Likewise,
we know how many vehicles have entered the link at earlier points in time, so
we know N ↑ (0), N ↑ (∆t), . . . , N ↑ (t). For the sending flow, we are assuming that
there are no obstructions from downstream. If this were the case, then we can
calculate N ↓ (t + ∆t) using the Newell-Daganzo method, and

S(t) = N ↓ (t + ∆t) − N ↓ (t) . (9.66)

In the point queue model, there are two possible wave speeds, uf (corre-
9.5. LWR-BASED LINK MODELS 361

x
N(L,t) N(L,t+1)
L

uf

N(0,t) t
N(0,t-ufL) N(0,t+1)

Figure 9.26: Point queue model characteristics for sending and receiving flow.

sponding to free-flow conditions) and zero (corresponding to flow moving at

capacity). Figure 9.26 shows how these characteristics can be traced back to
known data, either at the upstream end x = 0, or at the downstream end x = L.
Therefore, tR = t, xU = 0, and tU = (t + ∆t) − L/uf . We can think of this as
a special case of a “trapezoidal” diagram where the jam density kj is infinite.
Applying equation (9.48), we see that if kj = ∞ then the minimum must occur
in one of the first two terms, so

N ↓ (t + ∆t) = min{N ↑ ((t + ∆t) − L/uf ), N ↓ (t) + qmax ∆t} (9.67)

and
S(t) = min{N ↑ ((t + ∆t) − L/uf ) − N ↓ (t), qmax ∆t} . (9.68)
The two terms in the minimum in (9.68) correspond to the case when the queue
is empty, and when there are vehicles in queue. In the first term, since there
is no queue, we just need to know how many vehicles will finish traversing the
physical section of the link between t and t + ∆t; this is exactly the difference
between the total number of vehicles which have entered by time t + ∆t − L/uf
and the total number that have left by time t. When there is a queue, the
vehicles exit the link at the full capacity rate.
In these expressions, it is possible that (t + ∆t) − L/uf is not an integer,
that is, it does not line up with one of the discretization points exactly. In this
case the most accurate choice is to interpolate between the known time points
on either side (remember that we chose ∆t so that L/uf ≥ 1). If you are willing
to sacrifice some accuracy for efficiency, you can choose to round to the nearest
integer, or to adjust the length of the link so that L/uf is an integer.
For the receiving flow, we look at the upstream end of the link. We can treat
the same points as known — N ↑ (0), N ↑ (∆t), . . . , N ↑ (t) and N ↓ (0), N ↓ (∆t), . . . , N ↓ (t).
Since the fundamental diagram for the point queue model has no decreasing por-
tions, the known data at the downstream end can never be relevant. (A line
connecting one of these points to the unknown point N ↑ (t + ∆t) must have
362 CHAPTER 9. NETWORK LOADING

q
qmax
Q(k)

k
kj

Figure 9.27: Fundamental diagram in the spatial queue model.

negative slope, see Figure 9.26.) Furthermore, for the receiving flow, the char-
acteristic with positive slope +uf , corresponding to upstream conditions, is
irrelevant because we are assuming an unlimited number of vehicles are avail-
able to move from upstream — and therefore its term in (9.48) will never be the
minimum. We are only left with the middle term, corresponding to capacity, so

N ↑ (t + ∆t) = N ↑ (t) + qmax ∆t (9.69)

and
R(t) = qmax ∆t . (9.70)
In terms of the fundamental diagram, the spatial queue model takes the
form in Figure 9.27. This diagram requires three parameters to calibrate: the
free-flow speed uf , the capacity qmax , and the jam density kj . Notice, however,
that the fundamental diagram is discontinuous, and immediately drops from
qmax to zero once jam density is reached. This implies that backward-moving
shockwaves can have infinite speed in the spatial queue model — a physical
interpretation is that when vehicles at the front of the queue begin moving,
vehicles at the rear of the queue immediately start moving as well. In reality,
there is a delay before vehicles at the rear of the queue begin moving, and this
can be treated as an artifact arising from simplifying assumptions made in the
spatial queue model.9
There are thus three possible characteristic speeds: +uf at free-flow, 0 at ca-
pacity flow, and −∞ when the queue reaches jam density. The Newell-Daganzo
method is applied in much the same way as was done for the point queue model.
In particular, the sending flow expression is exactly the same, because the two
characteristics with nonnegative velocity are the same. We thus have

N (L, t + 1) = min{N ↑ ((t + ∆t) − L/uf ), N ↓ (t) + qmax ∆t} (9.71)

and
S(t) = min{N ↑ ((t + ∆t) − L/uf ) − N ↓ (t), qmax ∆t} . (9.72)
9 Alternatively, connected and autonomous vehicles may be able to exhibit such behavior
if an entire platoon of vehicles coordinates its acceleration.
9.5. LWR-BASED LINK MODELS 363

x
N(L,t) N(L,t+1)
L

uf

t
N(0,t-ufL) N(0,t) N(0,t+1)

Figure 9.28: Spatial queue model characteristics for sending and receiving flow;
the upstream-moving wave is an approximation of the vertical component of the
fundamental diagram.

For the receiving flow, we have to take into account the new shockwave
speed. Dealing with an infinite speed can be tricky, since, taken literally, would
mean that the upstream cumulative count N ↑ (t + ∆t) could depend on the
downstream cumulative count at the same time N ↓ (t + ∆t). Since we are solv-
ing the model in forward order of time, however, we do not know the value
N ↓ (t + ∆t) when calculating N ↑ (t + ∆t). In an acyclic network, we could sim-
ply do the calculations such that N ↓ (t+∆t) is calculated first before N ↑ (t+∆t),
using the concept of a topological order. In networks with cycles — virtually
all realistic traffic networks — this will not work. Instead, what is best is to
approximate the “infinite” backward wave speed with one which is as large as
possible, basing the calculation on the most recent known point N ↓ (t) (Fig-
ure 9.28). Effectively, this replaces the infinite backward wave speed with one
of speed L/∆t. Equation (9.48) thus gives

N ↑ (t + ∆t) = min{N ↑ (t) + qmax ∆t, N ↓ (t) + kj L} (9.73)

and
R(t) = min{qmax ∆t, (N ↓ (t) + kj L) − N ↑ (t)} . (9.74)
Equation (9.74) will ensure that the number of vehicles on the link will never
exceed kj L, assuming that this is true at time zero, as you are asked to show in
Exercise 4.

9.5.4 Discussion
This chapter has presented four different link models: point queues, spatial
queues, the cell transmission model, and the link transmission model. Although
not initially presented this way, all four can be seen as special cases of the
Lighthill-Whitham-Richards model. The point and spatial queue models can
364 CHAPTER 9. NETWORK LOADING

be derived from particularly simple forms of the fundamental diagram (as well
as from physical first principles, as in Section 9.1), while the cell transmission
model and link transmission model are more general methods which can handle
more sophisticated fundamental diagrams (typically triangular or trapezoidal in
practice). The cell transmission model directly solves the LWR system of partial
differential equations by discretizing in space and time, and applying a finite-
difference approximation. The link transmission model is based on the Newell-
Daganzo method. The primary distinction between these methods is that the
Newell-Daganzo method only requires tracking the cumulative counts N at the
upstream and downstream ends of each link in time, while the cell transmission
model also requires tracking the number of vehicles n at intermediate cells within
the link. However, the cell transmission model does not require storing any
values from previous time steps, and can function entirely using the number of
vehicles in each cell at the current time. The Newell-Daganzo method requires
that some past cumulative counts be stored, for the amount of time needed
for a wave to travel from one end of the link to the other. Which is more
desirable depends on implementation details, and on the specific application
context — at times it may be useful to know the distribution of vehicles within
a link (as the cell transmission model gives), while for other applications this
may be an irrelevant detail. One final advantage of the Newell-Daganzo method
is that the values it gives are exact. In the cell transmission model, backward-
moving shockwaves will tend to “spread out” as a numerical issue involved in the
discretization; this will not happen when applying the Newell-Daganzo method.
The exercises explore this issue in more detail. On the other hand, the cell
transmission model is easier to explain to decision-makers, and its equations
have intuitive explanations in terms of vehicles moving within a link and the
amount of available space. The Newell-Daganzo method is a “deeper” method
requiring knowledge of partial differential equations, and seems more difficult
to convey to nontechnical audiences.

The point queue and spatial queue models have their places as well, despite
their strict assumptions. The major flaw in the point queue model, from the
standpoint of realism, is its inability to model queue spillbacks which occur
when links are full. On the other hand, by ignoring this phenomenon, the
point queue model is much more tractable, and is amenable even to closed-form
expressions of delay and sensitivity to flows. It is also more robust to errors in
input data, because queue spillback can introduce discontinuities in the network
loading. There are cases where this simplicity and robustness may outweigh the
(significant) loss in realism induced by ignoring spillbacks. The spatial queue
model can represent spillbacks, but will tend to underestimate its effect due to
its assumption of infinitely-fast backward moving shockwaves. Nevertheless, it
can also lead to simpler analyses than the link transmission model.
9.6. FANCIER NODE MODELS 365

9.6 Fancier Node Models

Section 9.2 introduced node modeling concepts, and how the sending and re-
ceiving flows from adjacent links are mapped to transition flows showing how
many vehicles move from each incoming link to each outgoing link in a sin-
gle time step. In analogy with Section 9.5, we now expand our discussion of
node models beyond the simple intersection types presented thus far. A gen-
eral intersection can have any number of incoming and outgoing links. General
intersections can represent signal-controlled intersections, stop-controlled inter-
sections, roundabouts, and so forth. Node models for general intersections are
less standardized, and less well-understood, than node models for merges and
diverges, and representations of these can vary widely in dynamic traffic as-
signment implementations. This section presents four alternative models for
general intersections: the first for simple traffic signals, the second for inter-
sections where drivers strictly take turns or have equal priority for all turning
movements (such as an all-way stop), and the third where approaches have dif-
ferent priority levels and crossing conflicts must be considered, such as a two-way
stop or signal with permitted phasing. A fourth node model is presented which
allows one to use arbitrary models for turning movement capacity and intersec-
tion delay. As you read through this section, it would be helpful to think about
how these models can be adapted or extended to represent more sophisticated
signal types, roundabouts, and other types of junctions. Many of the ideas in
these models are based on the ideas described for merges and diverges.

9.6.1 Basic signals

For the purposes of this section, a “basic signal” is one which (1) only has pro-
tected phases (no permitted turns which must yield to oncoming traffic) and (2)
all turn movements corresponding to the same approach move simultaneously
(for instance, where there are no turn lanes). See Figure 9.29 for an illustra-
tion. In this case, the node can be modeled as a diverge intersection, where the
“upstream” link varies over time, depending on which approach has the green
indication. Flows from other approaches (which have red indications) are set to
zero. Following the notation for diverges, we use phij to reflect the proportion of
the sending flow from approach (h,P i) which wishes to leave via link (i, j): these
values must be nonnegative, and (i,j):(h,i,j)∈Ξ(i) phij = 1 for all approaches
(h, i). The algorithm is as follows:

1. Let (h∗ , i) ∈ Γ−1 (i) be the approach which has the green indication at the
current time.

2. Calculate the fraction of flow which can move:

 
Rij
φ= min , 1 . (9.75)
(i,j):(h∗ ,i,j)∈Ξ(i) ph∗ ij Sh∗ i
366 CHAPTER 9. NETWORK LOADING

Figure 9.29: Phasing plan for a basic signal.

3. Calculate the transition flows for each turning movement:

(
φShi phij if h = h∗ and (h, i, j) ∈ Ξ(i)
yhij = (9.76)
0 otherwise

In this implementation, one must be a little bit careful if the green times
in the signal are not multiples of the time step ∆t. It is possible to round the
green times so that they are multiples of ∆t, but this approach can introduce
considerable error over the analysis period: for instance, assume that ∆t is equal
to six seconds, and a two-phase intersection has green times of 10 seconds and
14 seconds, respectively. Rounding to multiples of the time step would give both
phases twelve seconds each, which seems reasonable enough; but over a three-
hour analysis period, the phases would receive 75 and 105 minutes of green time
in reality, as compared to 90 minutes each in simulation. In highly congested
situations, this can introduce considerable error. This issue can be avoided if,
instead of rounding, one gives the green indication to the approach which would
have green in reality at that time. In the example above, the intersection has a
cycle length of 24 seconds. So, when t = 60 seconds, we are 12 seconds into the
third cycle; and at this point the green indication should be given to the second
phase. In this way, there is no systematic bias introduced into the total green
time each approach receives.

9.6.2 Equal priority movements

An all-way stop intersection is characterized by turn-taking: that is, vehicles
have the opportunity to depart the intersection in the order in which they arrive.
No turning movement has priority over any other, but the turning movements
from different approaches interact with each other and may compete for space
on the same outgoing link. This is different from the basic signal model, where
the phasing scheme ensures that at most one approach is attempting to use an
outgoing link at any given point of time.
Intersections with equal priority movements have characteristics of both di-
verges and merges. Like a diverge, if a vehicle is unable to turn into a down-
9.6. FANCIER NODE MODELS 367

1 1
(a) (b)
60 60

2 60 3 2 60 3

60 30 30
4 (a) 4 (b)

Figure 9.30: Oriented capacities for a three-leg intersection, where all ap-
proaches have qmax = 60. (a) All approaches map to a unique outgoing link.
(b) The flow on approach (4,2) is split between two outgoing links.

stream link because of an obstruction, we assume that the vehicle obstructs all
other vehicles from the same approach, respecting the FIFO principle. This
means that the outflows for all of the turning movements corresponding to any
approach must follow the same proportions as the number of drivers wishing
to use all of these movements. Similar to a merge, we assume that if there are
high sending flows from all the approaches, the fraction of the receiving flow
allocated to each approach is divided up proportionally. However, instead of
allocating the receiving flow Rij to approach (h, i) based on the full capacity
hi
qmax , we instead divide up the receiving flow based on the oriented capacity
hij hi
qmax = qmax phij , (9.77)

where phij is the proportion of the flow from approach (h, i) which wishes to
exit on link (i, j). (If turn movement [h, i, j] is not in the allowable set Ξ(i),
then phij = 0.)
Multiplying the capacity by this proportion reflects the fact that an upstream
approach can only make use of an available space on a downstream link if there
is a vehicle wishing to turn. In Figure 9.30(a), each incoming link uses a unique
exiting link, and thus can claim its full capacity. In Figure 9.30(b), half the
vehicles on link (4,2) want to turn right and half wish to go straight, whereas
all the vehicles on link (3,2) wish to turn right. Link (3,2) therefore has twice
as many opportunities to fill available space on link (2,1), and thus its rightful
share is twice that of link (4,2). For any two approaches [h, i, j] and [h0 , i, j]
using the same outgoing link, we thus require that

yhij q hij
= max
h0 ij
. (9.78)
yh0 ij qmax

assuming that both approaches are fully competing for the link (i, j). If an ap-
proach has a small sending flow, it may use less of its assigned receiving flow
than equation (9.78) allocates, and this unused receiving flow may be used by
other approaches.
368 CHAPTER 9. NETWORK LOADING

We also define the oriented sending flow

Shij = Shi phij (9.79)

to reflect the demand for the turning movement [h, i, j]. Following the same
principles as the merge model, if the oriented sending flow from an approach is
less than its proportionate share of a downstream link’s receiving flow, its unused
share will be divided among the other approaches with unserved demand still
remaining, in proportion to their oriented capacities. If the oriented sending
flow for a turn movement is greater than the oriented receiving flow for that
movement, then by the FIFO principle applied to diverges, it will restrict flow
to all other downstream links by the same proportion, and for any two turning
movements (h, i, j) and (h, i, j 0 ) from the same approach we must have
yhij Shij phij
= = . (9.80)
yhij 0 Shij 0 phij 0

We can rearrange this equation to show that the ratio yhij /Shij (the ratio of
actual flow and desired flow for any turning movement) is uniform for all the
turning movements approaching from link (h, i) — this ratio plays the same role
as φ in a diverge.
The presence of multiple incoming and outgoing links causes another com-
plication, in that the flows between approaches are all linked together. If an
approach is restricted by the receiving flow of a downstream link, flow from
that approach is restricted to all other downstream links. This means that the
approach may not fully consume its “rightful share” of another downstream
link, thereby freeing up additional capacity for a different approach. Therefore,
we cannot treat the approaches or downstream links separately or even sequen-
tially in a fixed order, because we do not know a priori how these will be linked
together.
However, there is an algorithm which generates a consistent solution de-
spite these mutual dependencies. In this algorithm, each approach link can be
demand-constrained, or supply-constrained by a downstream link. If an approach
is demand-constrained, its oriented sending flow to all downstream links is less
than its rightful share, and therefore all of the sending flow can move. If an
approach is supply-constrained by link (i, j), then the approach is unable to
move all of its sending flow, and the fraction which can move is dictated by link
(i, j). (That is, receiving flow on (i, j) is the most restrictive constraint for the
approach). The algorithm must determine which links are demand-constrained,
and which are supply-constrained by a downstream link.
To find such a solution, we define two sets of auxiliary variables, S̃hij to
reflect the amount of unallocated sending flow for movement [h, i, j], and R̃ij to
reflect the amount of unallocated receiving flow for outgoing link (i, j). These
are initialized to the oriented sending flows and link receiving flows, and reduced
iteratively as flows are assigned and the available sending and receiving flows are
used up. The algorithm also uses the notion of active turning movements; these
are turning movements whose flows can still be increased. A turning movement
9.6. FANCIER NODE MODELS 369

[h, i, j] becomes inactive either when S̃hij drops to zero (all vehicles that wish to
turn have been assigned), or when R̃ij 0 drops to zero for any outgoing link (i, j 0 )
that approach (h, i) is using (that is, for which phij 0 > 0). Allocating all of the
receiving flow for one outgoing link can thus impact flow on turning movements
which use other outgoing links, because of the principle that vehicles wishing to
turn will block others, as expressed in equation (9.80). The set of active turning
movements will be denoted by A; a turning movement remains active until we
have determined whether it is supply-constrained or demand-constrained.
At each stage of the algorithm, we will increase the flows for all active turning
movements. We must increase these flows in a way which is consistent both
with the turning fractions (9.80), and with the division of receiving flow for
outgoing links given by equation (9.78), and we will use αhij to reflect the rate
of increase for turning movement (h, i, j). The absolute values of these α values
do not matter, only their proportions, so you can scale them in whatever way is
most convenient to you. Often it is easiest to pick one turning movement (h, i, j)
and fix its α value either to one, or to its oriented sending flow. The turning
proportions from (h, i) then fix the α values for all other turning movements
from the same approach. You can then use equation (9.78) to determine α
values for turning movements competing for the same outgoing link, then use
the turning fractions for the upstream link on those turning movements, and so
on until a consistent set of α values has been determined.
The algorithm then increases the active turning movement flows in these
proportions until some movement becomes inactive, because its sending flow or
the receiving flow on its outgoing link becomes exhausted. The process is then
repeated with the smaller set of turning movements which remain active, and
continues until all possible flows have been assigned. This algorithm is a bit
more involved than the node models seen thus far, and you may find it helpful
to follow the example below as you read through the algorithm steps.

1. Initialize by calculating oriented capacities and sending flows using equa-

tions (9.77) and (9.79); by setting yhij ← 0 for all [h, i, j] ∈ Ξ(i); by set-
ting S̃hij ← Shij for all [h, i, j] ∈ Ξ(i) and R̃ij ← Rij for all (i, j) ∈ Γ(i);
and by declaring active all turning movements with positive sending flow:
A ← {[h, i, j] ∈ Ξ(i) : Shij > 0}.

2. Identify a set of αhij values for all active turning movements which is con-
sistent with the turning fractions (αhij /αhij 0 = phij /phij 0 for all turning
movements from the same incoming link) and oriented capacities (αhij /αh0 ij =
hij h0 ij
qmax /qmax for all turning movements to the same outgoing link).

3. For each outgoing link (i, j) ∈ Γ(i), identify the rate αij at which its
receiving flow will be reduced, by adding
P αhij for all active turn movements
whose outgoing link is (i, j): αij ← [h,i,j]∈A αhij .

4. Determine the point at which some turning movement will become inac-
370 CHAPTER 9. NETWORK LOADING

10 1
1

2 1 60 3

1/2 1/2
60
4

Figure 9.31: Example of equal priorities algorithm, showing sending flows and
turning proportions. All links have capacity of 60.

tive, by calculating
( ( ) ( ))
S̃hij R̃ij
θ = min min , min . (9.81)
[h,i,j]∈A αhij (i,j)∈Γ(i):αij >0 αij

5. Increase flows for active turning movements, and update unallocated send-
ing and receiving flows: for all [h, i, j] ∈ A update yhij ← yhij + θαhij ,
S̃hij ← S̃hij − θαhij , and R̃ij ← R̃ij − θαhij .
6. Update the set of active turning movements, by removing from A any
turning movement [h, i, j] for which S̃hij = 0 or for which R̃ij 0 = 0 for any
(i, j 0 ) ∈ Γ(i) which is being used (phij 0 > 0).
7. If there are any turning movements which are still active (A 6= ∅), return
to step 3. Otherwise, stop.
As a demonstration, consider the intersection in Figure 9.31, where all links
(incoming and outgoing) have the same capacity of 60 vehicles per time step,
and the sending flows Shi and turning proportions phij are shown. None of
the downstream links is congested, so their receiving flows are equal to the
capacity. (In the figure, two sets of numbers are shown for each approach; the
“upstream” number is the sending flow and the “downstream” number(s) are
the proportions.) The oriented capacities can be seen in Figure 9.30(b). For
this example, the six turning movements will be indexed in the following order:
Ξ(2) = {[1, 2, 3], [1, 2, 4], [3, 2, 1], [3, 2, 4], [4, 2, 1], [4, 2, 3]} (9.82)
All vectors referring to turning movements will use this ordering for their com-
ponents.
Step 1 of the algorithm initializes the oriented sending flows using equa-
tion (9.79),    
S = Shij = 0 10 60 0 30 30 (9.83)
and the oriented capacities using equation (9.77)
 hij   
qmax = qmax = 0 60 60 0 30 30 . (9.84)
9.6. FANCIER NODE MODELS 371

The step also initializes the turning movement flows and auxiliary variables:
 
y← 0 0 0 0 0 0 , (9.85)
   
S̃ ← Shij = 0 10 60 0 30 30 , (9.86)
and    
R̃ = R̃21 R̃23 R̃24 = 60 60 60 . (9.87)
The set of active turn movements is A = {[1, 2, 4], [3, 2, 1], [4, 2, 1], [4, 2, 3]}.
Step 2 of the algorithm involves calculation of a set of consistent αhij values.
One way of doing this is to start by setting α423 ← S423 = 30. The turning
fractions from (4, 2) then require that α421 = α423 = 30. The allocation rule
for outgoing link (2, 1) then forces S321 = 60: the oriented capacity for [3, 2, 1]
is twice that of [4, 2, 1], and the α values must follow the same proportion.
Turning movement [1, 2, 4] is independent of all of the other turning movements
considered thus far, so we can choose its α value arbitrarily; say, α124 ← S124 =
10. (You should experiment around with different ways of calculating these α
values, and convince yourself that the final flows are the same as long as the
proportions of α values for interdependent turning movements are the same.)
We thus have    
α = αhij = 0 10 60 0 30 30 . (9.88)
The α values for inactive turning movements have been set to zero for clarity;
their actual value is irrelevant because they will not be used in any of the steps
that follow.
With these flow increments, flow on outgoing links (2, 1), (2, 3), and (2, 4)
will be α21 = 90, α23 = 30, and α24 = 10, as dictated by Step 3.
In Step 4, we determine how much we can increase the flow at the rates
given by α until some movement becomes inactive. We have

S̃124 S̃321 S̃421 S̃423 R̃21 R̃23 R̃24

10 60 30 30 60 60 60

θ = min 10 , 60 , 30 , 30 , 90 , 30 , 10 (9.89)

or θ = 2/3.
We can now adjust the flows, as in Step 5. We increase the flow on each
active turning movement by 23 αhij , giving

y ← 0 6 23 40 0 20 20 .
 
(9.90)

We subtract these flow increments from the auxiliary sending and receiving
flows, giving
S̃ ← 0 3 31 20 0 10 10
 
(9.91)
and
53 31 .
 
R̃ ← 0 40 (9.92)

Step 6 updates the set of active turning movements. With the new S̃ and R̃
values, we see that [3, 2, 1] and [4, 2, 1] have become inactive, since there is no
372 CHAPTER 9. NETWORK LOADING

remaining receiving flow on link (2, 1). Furthermore, this inactivates movement
[4, 2, 3]: even though there are still travelers that wish to turn in this direction,
and space on the downstream link (S̃423 and R̃23 are still positive), they are
blocked by travelers waiting to use movement [4, 2, 1]. So, there is only one
active movement remaining, A ← {[1, 2, 4]}, and we must return to step 3.
In Step 3, we must recalculate the αij values because some of the turning
movements are inactive. With the new set A, we have α21 = α23 = 0 and
α24 = 10. The new step size is

S̃124 R̃24
n
3 13 53 13
o 1
θ = min 10 ,
= . (9.93)
10 3
We then increase the flows, increasing y124 by 10 × 13 to 10, decreasing S̃124
to zero, and decreasing R̃24 to 50. This change inactivates movement [1, 2, 4].
Since there are no more active turning movements, the algorithm terminates,
and the final vector of flows is
 
y = 0 10 40 0 20 20 . (9.94)

9.6.3 Intersections with priority

Intersections which allow crossing conflicts are more complex to model than
the intersection types described above. These include intersections with stop
control only on some of the approaches, but not all, or signalized intersections
with permitted movements that must yield to another traffic stream. One way
to model this type of intersection is to introduce a set C of conflict points. Like
outgoing links, we associate a receiving flow Rc indicating the maximum number
of vehicles which can pass through this conflict point during a single time step
∆t. For each conflict point c ∈ C, let Ξ(c) denote the turning movements which
make use of this conflict point. Then, for each turning movement [h, i, j] ∈ Ξ(c),
c
we define a (strictly positive) priority parameter βhij . These priority parameters
are interpreted through their ratios: the share of the receiving flow allocated
to turning movement [h, i, j] relative to that allocated to turning movement
[h0 , i, j 0 ] is in the same proportion as the ratio
c
βhij phij /βhc 0 ij 0 ph0 ij 0 . (9.95)
These play an analogous role to the oriented capacities defined in the pre-
vious subsection, in suggesting how the conflict point receiving flow should be
divided among the competing approaches. They are used more generally, how-
ever, to reflect priority rules. For instance, a through movement often has
priority over a turning movement which crosses it. Even if the capacity and
sending flows of the through lane and turn lane are the same, the through lane
should have access to a greater share of the crossing point’s receiving flow. If, at
full saturation, ten through vehicles move for every turning vehicle, then the β
value for the through movement should be ten times the β value for the turning
movement.
9.6. FANCIER NODE MODELS 373

The node model is simplified if we assume that the β values are all strictly
positive. You might find this unrealistic: in the example above, if the turning
movement must yield to the through movement, then at full saturation perhaps
no turning vehicles could move. In practice, however, priority rules are not
strictly obeyed as traffic flows near saturation. Polite through drivers may
stop to let turning drivers move, or aggressive turning drivers might force their
way into the through stream. (Think of what would happen at a congested
freeway if vehicles merging from an onramp took the “yield” sign literally!)
The requirement of strictly positive β values thus has some practical merits, as
well as mathematical ones. The exercises explore ways to generalize this node
model, including strict priority, and cases where different turning movements
may consume different amounts of the receiving flow (for instance, if they are
moving at different speeds).
We can now adapt the algorithm for equal priority for the case of intersec-
tions with different priorities. The algorithm is augmented by adding receiving
flows Rc and auxiliary receiving flows R̃c for each conflict point, and we extend
some of the computations to include the set of conflict points. First, we require
that the ratio of α values for two turning movements using the same crossing
point follow the ratio of the β values, assuming saturated conditions:
c
αhij βhij phij
= c ∀c ∈ C; [h, i, j], [h0 , i, j 0 ] ∈ Ξ(c) . (9.96)
αh0 ij 0 βh0 ij 0 ph0 ij 0

Note that the β values are multiplied by the relevant turning fraction for the
movements. As with the oriented capacity, this respects the fact that the more
flow is attempting to turn in a particular direction, the more opportunities or
gaps will be available for it to claim.
Second, we must calculate the inflow rates to conflict points given the α
values from active turning movements:
X
αc = αhij . (9.97)
[h,i,j]∈Ξ(c)∩A

Third, the calculation of the step size θ must now include obstructions from
conflict points:
( ( ) ( ) ( ))
S̃hij R̃ij R̃c
θ = min min , min , min . (9.98)
[h,i,j]∈A αhij (i,j)∈Γ(i):αij >0 αij c∈C:αc >0 αc

With these modifications, the algorithm proceeds in the same way as before.
Specifically,
1. Initialize by calculating oriented capacities and sending flows using equa-
tions (9.77) and (9.79); by setting yhij ← 0 for all [h, i, j] ∈ Ξ(i); by
setting S̃hij ← Shij for all [h, i, j] ∈ Ξ(i), R̃ij ← Rij for all (i, j) ∈ Γ(i),
and R̃c ← Rc for each c ∈ C; and by declaring active all turning move-
ments with positive sending flow: A ← {[h, i, j] ∈ Xi(i) : Shij > 0}.
374 CHAPTER 9. NETWORK LOADING

60 1 Capacity 60
Receiving flow 45
1/4 3/4
Receiving flow 60
1 27 3
2
Capacity 30
Receiving flow 30

Capacity 60 1
Receiving flow 60 54
4

Figure 9.32: Example of intersection priority algorithm, showing sending flows

and turning proportions. Link capacity and receiving flows shown.

2. Identify a set of αhij values for all active turning movements which is con-
sistent with the turning fractions (αhij /αhij 0 = phij /phij 0 for all turning
movements from the same incoming link), oriented capacities (αhij /αh0 ij =
hij h0 ij
qmax /qmax for all turning movements to the same outgoing link), and con-
flict points based on equation (9.96).

3. For each outgoing link (i, j) ∈ Γ(i) and conflict point c ∈ C, identify the
rate αij at which its receiving flow will be reduced, by adding
P αhij for all
active turn movements whose outgoing link is (i, j): αij ← [h,i,j]∈A αhij
for links, and equation (9.97) for conflict points.

4. Determine the point at which some turning movement will become inac-
tive, by calculating θ using equation (9.98).

5. Increase flows for active turning movements, and update unallocated send-
ing and receiving flows: for all [h, i, j] ∈ A update yhij ← yhij + θαhij ,
S̃hij ← S̃hij − θαhij ; for all (i, j) ∈ Γ(i) update R̃ij ← R̃ij − θαij , and for
all c ∈ C update R̃c ← R̃c − θαc .

6. Update the set of active turning movements, by removing from A any

turning movement [h, i, j] for which S̃hij = 0, for which R̃ij 0 = 0 for any
(i, j 0 ) ∈ Γ(i) which is being used (phij 0 > 0), or for which R̃c = 0 for any
movement [h, i, j 0 ] ∈ Ξ(c) and conflict point c.

7. If there are any turning movements which are still active (A 6= ∅), return
to step 3. Otherwise, stop.

As a demonstration, consider the intersection in Figure 9.32, where all links

(incoming and outgoing) have the same capacity of 60 vehicles per time step,
and the sending flows Shi and turning proportions phij are shown. (In the figure,
two sets of numbers are shown for each approach; the “upstream” number is the
sending flow and the “downstream” number(s) are the proportions.) Links (2, 4)
9.6. FANCIER NODE MODELS 375

and (2, 3) have receiving flows of 60 vehicles, while link (2, 1) has a receiving flow
of only 45 vehicles. There is one conflict point, indexed c, which is marked with a
circle in Figure 9.32). Conflict point c has a receiving flow of 60, and the turning
movement [1, 2, 3] must yield to the through movement [4, 2, 1], as reflected by
the ratio β421 /β123 = 5. For this example, the six turning movements will be
indexed in the following order:

Ξ(2) = {[1, 2, 3], [1, 2, 4], [3, 2, 1], [3, 2, 4], [4, 2, 1], [4, 2, 3]} (9.99)

All vectors referring to turning movements will use this ordering for their com-
ponents.
Step 1 of the algorithm initializes the oriented sending flows using equa-
tion (9.79),    
S = Shij = 45 15 27 0 54 0 (9.100)
and the oriented capacities using equation (9.77)
 hij   
qmax = qmax = 30 30 30 0 60 0 . (9.101)

The step also initializes the turning movement flows and auxiliary variables:
 
y← 0 0 0 0 0 0 , (9.102)
   
S̃ ← Shij = 45 15 27 0 54 0 , (9.103)
and    
R̃ = R̃21 R̃23 R̃24 R̃c = 45 30 60 60 . (9.104)
The set of active turn movements is A = {[1, 2, 3], [1, 2, 4], [3, 2, 1], [4, 2, 1]}.
Step 2 of the algorithm involves calculation of a set of consistent αhij values.
One way of doing this is to start by setting α123 ← S123 = 15. The turning frac-
tions from (1, 2) then require that α124 = α123 = 45. The allocation rule (9.96)
for conflict point c forces S421 = 300: the β value for [4, 2, 1] is five times that
of the β value for [1, 2, 3], and the turning fraction for [4, 2, 1] is a third higher.
Since S421 = 300, we must have S321 = 150 because the oriented capacity of
[3, 2, 1] is half that of [4, 2, 1]. This gives the flow increments
   
α = αhij = 45 15 150 0 300 0 . (9.105)

With these flow increments, flow on outgoing links (2, 1), (2, 3), (2, 4), and
the conflict point c will be α21 = 450, α23 = 45, α24 = 15, and αc = 345, as
dictated by Step 3.
In Step 4, we determine how much we can increase the flow at the rates
given by α until some movement becomes inactive. We have

S̃123 S̃124 S̃321 S̃421 R̃21 R̃23 R̃24 R̃c

15 45 27 54 45 60 60 60

θ = min 15 , 45 , 150 , 300 , 450 , 45 , 15 , 345 (9.106)
1
or θ = 10 .
376 CHAPTER 9. NETWORK LOADING

We can now adjust the flows, as in Step 5. We increase the flow on each
1
active turning movement by 10 αhij , giving
 
y ← 4.5 1.5 15 0 30 0 . (9.107)

We subtract these flow increments from the auxiliary sending and receiving
flows, giving  
S̃ ← 40.5 13.5 12 0 24 0 (9.108)
and  
R̃ ← 0 25.5 58.5 25.5 . (9.109)

Step 6 updates the set of active turning movements. With the new S̃ and
R̃ values, we see that [3, 2, 1] and [4, 2, 1] have become inactive, since there is
no remaining receiving flow on link (2, 1). So, there are two active movements
remaining, A ← {[1, 2, 3], [1, 2, 4]}, and we must return to step 3.
In Step 3, we must recalculate the αij values because some of the turning
movements are inactive. With the new set A, we have α21 = 0, α23 = 45,
α24 = 15, and αc = 45. The new step size is

S̃123 S̃124 R̃23 R̃24 R̃c

 40.5 13.5 25.5 58.5 25.5 17
θ = min 45 , 15 , 45 , 15 , 45 =
30
. (9.110)

17
We then increase the flows, increasing y123 by 45 × 30 to 30 and y124 by 15 × 17
30
to 10, decreasing S̃123 to 15, S̃124 to 5, R̃23 to 0, R̃24 to 50, and R̃c to 0. This
change inactivates movement [1, 2, 3]; and movement [1, 2, 4] is then inactivated
because these movement’s flows are blocked by vehicles waiting to take [1, 2, 3].
Since there are no more active turning movements, the algorithm terminates,
and the final vector of flows is
 
y = 30 10 15 0 30 0 . (9.111)

9.7 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
The idea of dividing networks into link models and node models (which op-
erate independently of each other, possibly with different models) was suggested
by Nie et al. (2008) and Yperman (2007). The sending flow and receiving flow
concepts date to Daganzo (1995b) and Lebacque (1996) (who used the terms
“demand” and “supply”), although the presentation in this chapter more closely
follows that of Yperman (2007). The point queue and spatial queue models were
described in Vickrey (1969) and Zhang and Nie (2006).
The list of node model desiderata in this chapter is that of Tampère et al.
(2011); the invariance principle specifically is discussed more in Lebacque and
Khoshyaran (2005). The diverge and merge equations are that of Daganzo
9.8. EXERCISES 377

(1995a); for an alternative merge model (which does not satisfy the invariance
principle), see Jin and Zhang (2003).
The hydrodynamic traffic flow theory described in Section 9.4.3 was indepen-
dently developed by Lighthill and Whitham (1955) and Richards (1956). Newell
(1993a), Newell (1993b), and Newell (1993c) recognized that the cumulative ve-
hicle counts N , and the analysis of characteristics resulting from a triangular
fundamental diagram, greatly simplify the solution of the model, a theory com-
pleted by Daganzo (2005a) and Daganzo (2005b). Interestingly, an equivalent
model (used for soil erosion) was separately developed by Luke (1972). There
are alternative means of solving the LWR model not presented in this book,
through recognizing it as a Hamilton-Jacobi system of partial differential equa-
tions (LeVeque, 1992; Evans, 1998) which can be solved using the Lax-Hopf
formula or viability theory (Lax, 1957; Hopf, 1970; Claudel and Bayen, 2010);
or for the purposes of a link model, representing sending and receiving flows
using a “double queue,” one at each end of the link (Osorio and Bierlaire, 2009;
Osorio et al., 2011), a representation suitable for a stochastic version of the
LWR model.
The cell transmission model was reported in Daganzo (1994) and Daganzo
(1995a), essentially a Gudanov scheme for solving the LWR system (Godunov,
1959; Lebacque, 1996). The link transmission model was developed by Yperman
(2007); see also Gentile (2010).
The more sophisticated node models reported later in the chapter are adapted
from Tampère et al. (2011) and Corthout et al. (2012).
Finally, network loading can be accomplished by entirely different means
than that reported in this chapter, without the use of explicit link and node mod-
els discretized in space and time. For instance, the discretization can be done in
the space of vehicle trajectories (Bar-Gera, 2005). In mathematical terms, this
involves converting from Eulerian coordinates (x and t) to Lagrangian coordi-
nates (with the cumulative count N in place of either x or t). For more on this
alternative, and reformulations of the LWR model with this change of variables,
see Laval and Leclercq (2013).
Another common alternative is to use traffic simulation to perform the
network loading. Examples include the software packages VISSIM (Fellen-
dorf, 1994), AIMSUN (Barcelo, 1998), DynaMIT (Ben-Akiva et al., 1998),
VISTA (Ziliaskopoulos and Waller, 2000), DYNASMART (Mahmassani, 2000),
Dynameq (Mahut et al., 2003), and DynusT.

9.8 Exercises
1. [14] Table 9.7 shows cumulative inflows and outflows to a link with a
capacity of 10 vehicles per time step, and a free-flow time of 2 time steps.
Use the point queue model to calculate the sending and receiving flow for
each time step.

2. [14] Repeat Exercise 1 with the spatial queue model. Assume that the jam
378 CHAPTER 9. NETWORK LOADING

Table 9.7: Upstream and downstream counts for Exercises 1 and 2.

t N ↑ (t) N ↓ (t)
0 0 0
1 5 0
2 10 0
3 15 2
4 20 4
5 20 6
6 20 10
7 20 15
8 20 20
9 20 20

density is such that at most 20 vehicles can fit on the link simultaneously.
.

3. [33] In the point queue model, if the inflow and outflow rates qin and qout
are constants with qin ≥ qout , show that the travel time experienced by
n qin
the n-th vehicle is L/uf + qin qout − 1 . The same result holds for the
spatial queue model, if there is no spillback.

4. [24] In the spatial queue model, show that if the total number of vehicles
on a link is at most kj L at time t, then the total number of vehicles on
the link at time t + 1 is also at most kj L.

5. [13] In a merge, approach 1 has a sending flow of 50 and a capacity of

100. Approach 2 has a sending flow of 100 and a capacity of 100. Report
the number of vehicles moving from each approach if (a) the outgoing link
has a receiving flow of 150; (b) the outgoing link has a receiving flow of
120; and (c) the outgoing link has a receiving flow of 100.

6. [32] Extend the merge model of Section 9.2.2 to a merge node with three
incoming links.

7. [43] Show that the formula (9.14) indeed captures both case II and case
III of the merge model presented in Section 9.2.2.

8. [32] Show that the merge model of Section 9.2.2 satisfies all the desiderata
of Section 9.2.

9. [41] Consider an alternative merge model for the congested case, which
allocates the receiving flow proportional to the sending flows of the in-
coming links, rather than proportional to the capacities of the incoming
links as was done in Section 9.2.2. Show that this model does not satisfy
the invariance principle.
9.8. EXERCISES 379

10. [13] In a diverge, the incoming link has a sending flow of 120, 25% of the
vehicles want to turn onto outgoing link 1, and the remainder want to turn
onto outgoing link 2. Report the number of vehicles moving to outgoing
links 1 and 2 if their respective receiving flows are (a) 80 and 100; (b) 80
and 60; (c) 10 and 40.

11. [10] Extend the diverge model of Section 9.2.3 to a diverge node with
three outgoing links.

12. [21] Develop a model for a diverge node with two outgoing links, in which
flows waiting to enter one link do not block flows entering the other link.
When might this model be more appropriate?

13. [25] Show that the diverge model of Section 9.2.3 satisfies all the desiderata
of Section 9.2.

14. [10] In the network loading procedure in Section 9.3, we specified that
centroid connectors starting at origins should have high jam density, and
those ending at destinations should have high capacity. Would anything
go wrong if centroid connectors starting at origins also had high capacity?
What if centroid connectors ending at destinations had high jam density?

15. [12] Draw trajectory diagrams which reflect the following situations: (a)
steady-state traffic flow, no vehicles speeding up or slowing down; (b)
vehicles approaching a stop sign, then continuing; (c) a slow semi truck
merges onto the roadway at some point in time, then exits at a later point
in time. Draw at least five vehicle trajectories for each scenario.

16. [33] The relationship q = uk can be used to transform the fundamen-

tal diagram (which relates density and flow) into a relationship between
density and speed, and vice versa. For each of these three speed-density
relationships, derive the corresponding fundamental diagram.

(a) The Greenshields (linear) model: u = uf (1 − k/kj ).

(b) The Greenberg (logarithmic) model: u = C log(k/kj ) where C is a
constant.
(c) The Underwood (exponential) model: u = uf exp(−k/kc ).
(d) The Pipes model: u = uf (1 − (k/kj )n ) where n is a constant. (The
Greenshields model is a special case when n = 1.)
(e) What features of the Greenberg and Underwood models make them
less suitable for dynamic network loading? (Hint: Draw plots of these
speed-density diagrams.)

17. [51] Which of the following statements are true with the LWR model and
a concave fundamental diagram?

(a) Speed uniquely defines the values of flow and density.

380 CHAPTER 9. NETWORK LOADING

(b) It is possible for higher density to be associated with higher speed.

(c) No shockwave can move downstream faster than the free-flow speed.
(d) Flow uniquely defines the values of speed and density.
(e) With a triangular fundamental diagram, traffic speed is constant for
subcritical densities.
(f) Density uniquely defines the values of speed and flow.

18. [44] The relationship q = uk can transform the fundamental diagram

(which relates density and flow) into a relationship between speed and
flow. The speed-flow relationship is traditionally plotted with the speed
on the vertical axis and flow on the horizontal axis.

(a) Show that for any concave fundamental diagram and any flow value q
less than the capacity, there are exactly two possible speeds u1 and u2
producing the flow q, one corresponding to subcritical (uncongested)
conditions and the other corresponding to supercritical (congested)
conditions.
(b) Derive and plot the speed-flow relationship corresponding to the Green-
shields model of Exercise 16. Express this relationship with two
functions u1 (q) and u2 (q) corresponding to uncongested and con-
gested conditions, respectively; these functions should have a domain
of [0, qmax ] and intersect at capacity.
(c) Derive and plot the fundamental diagram corresponding to the High-
way Capacity Manual speed-flow relation for basic freeway segments.
The capacity of such a segment is given by qmax = 1800 + 5uf . In this
equation, speeds are measured in km/hr and flows in veh/hr/lane:

uf if q ≤ 3100 − 15u0
u1 (q) =  2.6
uf − 23u0 −1800 q+15uf −3100 if 3100 − 15u0 ≤ q ≤ qmax
28 20uf −1300
(9.112)
u2 (q) = 28q (9.113)

19. [22] For each of these fundamental diagrams, derive the speed-density
function (that is, the travel speed for any given density value), and provide
a sketch.

(a) Q(k) = Ck(kj − k), where C is a constant and kj is the jam density.
(b) Q(k) = min{uf k, w(kj − k)}
(c) Q(k) = min{uf k, qmax , w(kj − k)}

20. [44] Consider a long, uninterrupted freeway with a capacity of 4400 vehi-
cles per hour, a jam density of 200 vehicles per mile, and a free-flow speed
of 75 miles per hour. Initially, freeway conditions are uniform and steady
9.8. EXERCISES 381

with a subcritical flow of 2000 vehicles per hour. An accident reduces the
roadway capacity to 1000 veh/hr for thirty minutes. Draw a shockwave di-
agram to show the effects of this accident, reporting the space-mean speed,
volume, and density in each region of your diagram, and the speed and
direction of each shockwave. Assume that the fundamental diagram takes
the shape of the Greenshields model (Exercise 16), and that a stopped
queue discharges at capacity.

21. [46] Consider a roadway with a linear-speed density relationship (cf. Ex-
ercise 16) whose capacity is 2000 veh/hr and free-flow speed is 40 mi/hr.
Initially, the flow is 1000 veh/hr and uncongested. A traffic signal is red
for 45 seconds, causing several shockwaves. When the light turns green,
the queue discharges at capacity.

(a) Sketch a time-space diagram, indicating all of the shockwaves which

are formed.
(b) Calculate the speed and direction of each shockwave from your dia-
gram.
(c) What is the minimum green time needed to ensure that no vehicle
has to stop more than once before passing the intersection? (Neglect
any yellow time, reaction time, etc. Assume that when the signal is
green, people move immediately, and that when it is red, people stop
immediately.)

22. [56] Consider a single-lane roadway with a triangular fundamental dia-

gram, a free-flow speed of 60 mi/hr, a backward wave speed of 30 mi/hr,
and a jam density of 200 veh/mi. Initially, traffic flow is uncongested, and
the volume is half of capacity. A slow-moving truck enters the roadway at
time t = 0, and travels at 20 mi/hr. This vehicle turns off of the roadway
one mile later.

(a) What is the capacity of the roadway?

(b) At time t = 2 minutes, you are a quarter of a mile behind the truck.
Use the Newell-Daganzo method to determine how many vehicles are
between you and the truck.
(c) In total, how many shockwaves are generated by the slow-moving
truck? Sketch them on a trajectory diagram.

23. [11] Show that a shockwave connecting two uncongested (subcritical) traf-
fic states always moves downstream, while a shockwave connecting two
congested (supercritical) traffic states always moves upstream. This is re-
lated to the observation in the chapter that “uncongested states propagate
downstream, and congested states propagate upstream.”

24. [36] This exercise asks you to fill in some details of the example in Sec-
1
tion 9.4 where the fundamental diagram was Q(k) = 240 k(240 − k) and
382 CHAPTER 9. NETWORK LOADING

60 vpm D
42 vpm
80 vpm 40 vpm
A B 10 vpm E F
20 vpm
42 vpm

Figure 9.33: Network for Exercise 25

60x2
the cumulative count map was N (x, t) = 60t − 120x + t+1 . Times are
measured in minutes, and distances in miles.

(a) Calculate the capacity, jam density, and free-flow speed associated
with this fundamental diagram.
(b) Verify that the conservation relationship (9.25) is satisfied by the flow
and density maps q(x, t) and k(x, t).
(c) Verify that the density and flow maps are consistent with the given
fundamental diagram.
(d) Calculate the speed u(x, t) at each point and time. Are vehicles ac-
celerating, decelerating, or maintaining a constant speed?

25. [45] Consider the network in Figure 9.33, where each link has a free-flow
time of 5 minutes and a capacity shown on the figure, and vehicles split
equally at each diverge (that is, pABC = pABD = pBCD = pBCE = 1/2
at all times). Vehicles enter the network at a rate of 80 veh/min for
20 minutes, and then the inflow rate drops to zero. Perform dynamic
network loading, using point queues for the link models. For each link in
the network, plot the cumulative counts N ↑ and N ↓ over time, as well as
the sending flow and receiving flow over time. At what time does the last
vehicle leave the network?

26. [25] Write the formula for the fundamental diagram Q(k) in the cell trans-
mission model example depicted in Table 9.4.

27. [13] A link is divided into four cells; on this link the capacity is 10 vehicles
per time step, each cell can hold at most 40 vehicles, and the ratio of
backward wave speed to free-flow speed is 0.5. Currently, the number of
vehicles in each cell is as in Table 9.8 (Cell 1 is at the upstream end of
the link, Cell 4 at the downstream end.) Calculate the number of vehicles
that will move between each pair of cells in the current time interval (that
is, the y12 , y23 , and y34 values.), and the number of vehicles in each cell
at the start of the next time interval. Assume no vehicles enter or exit the
link. .
9.8. EXERCISES 383

Table 9.8: Current cell occupancies for Exercise 27.

Cell 1 Cell 2 Cell 3 Cell 4
8 10 30 5

Table 9.9: Upstream and downstream counts for Exercise 28.

t N ↑ (t) N ↓ (t)
0 0 0
1 5 0
2 10 0
3 15 2
4 16 4
5 17 6
6 20 10
7 20 15

28. [23] Table 9.9 shows cumulative inflows and outflows to a link with a
capacity of 10 vehicles per time step, a free-flow time of 2 time steps, and
a backward wave time of 4 time steps. At jam density, there are 20 vehicles
on the link. Use the link transmission model to calculate the sending flow
S(7) and the receiving flow R(7).

29. [44] (Exploring shock spreading.) A link is seven cells long; at most 15
vehicles can fit into each cell, the capacity is 5 vehicles per timestep, and
w/uf = 1/2. Each time step, 2 vehicles wish to enter the link, and will do
so if the receiving flow can accommodate. There is a traffic signal at the
downstream end of the link. During time steps 0–9, and from time step
50 onward, the light is green and all of the link’s sending flow can leave.
For the other time steps, the light is red, and the sending flow of the link
is zero.

(a) Use the cell transmission model to propagate flow for 80 time steps,
portraying the resulting cell occupancies in a time-space diagram (time
on the horizontal axis, space on the vertical axis). At what time
interval does the receiving flow first begin to drop; at what point does
it reach its minimum value; and what is that minimum value? Is there
any point at which the entire link is at jam density?
(b) Repeat, but with w/uf = 1.
(c) Repeat, but instead use the link transmission model (with the same
time step) to determine how much flow can enter or leave the link.

30. [68] Consider the network in Figure 9.34. The figure shows each link’s
length, capacity, jam density, free-flow speed, and backward wave speed.
384 CHAPTER 9. NETWORK LOADING

0.125 km 0.125 km
11520 vph 11520 vph
320 veh/km
320 veh/km C 90 kph
90 kph
30 kph 30 kph

A B D E
Length: 0.25 km 0.125 km 0.25 km
Capacity: 17280 vph 5760 vph 10800 vph
Jam density: 480 veh/km 220 veh/km 160 veh/km
Free-flow speed: 90 kph 45 kph 90 kph
Backward wave speed: 30 kph 15 kph 30 kph

Figure 9.34: Network for Exercise 30

The inflow rate at node A is 4320 veh/hr for 0 ≤ t < 30 (t measured

in seconds), 8640 veh/hr for 25 ≤ t < 120, and 0 veh/hr thereafter. The
splitting proportion towards node C is 32 for 0 ≤ t < 50, 16 for 50 ≤ t < 90,
and 21 for t ≥ 90.

(a) Use a point queue model to propagate the vehicle flow with the time
step ∆t = 5 s. Plot the turning movement flows yABC , yABD , yBDE ,
and yBCE from t = 0 until the last vehicle has left the network. (q12
is the rate at which flow leaves the downstream end of link 1 to enter
the upstream end of link 2).
(b) Use the cell transmission model to propagate the vehicle flow with the
time step ∆t = 5 s. Plot the same flow rates as in the previous part.
(c) Use the link transmission model to propagate the vehicle flow with
the time step ∆t = 10 s. Plot the same flow rates as in the previous
part.
(d) Comment on any differences you see in these plots for the three flow
models.

31. [21] Assuming that a cell initially has between 0 and n̄ vehicles, show
the cell transmission model formula (9.55) ensures that it will have be-
tween 0 and n̄ vehicles at all future time steps, regardless of upstream or
downstream conditions.

32. [21] On a link, we must have N ↑ (t) ≥ N ↓ (t) at all time steps. Assuming
this is true for all time steps before t, show that the link transmission
model formulas (9.63) and (9.65) ensure this condition holds at t as well.

33. [42] Generalize the cell transmission model formula (9.55) to handle an
arbitrary piecewise-linear fundamental diagram (not necessarily triangular
or trapezoidal).
9.8. EXERCISES 385

Lamar
SB sending flow: 300
Proportion to Guadalupe: 30%
NB receiving flow: 100

Lamar Guadalupe
NB sending flow: 300 NB sending flow: 200
Proportion to Guadalupe: 5% Proportion to NB Lamar: 100%
SB receiving flow: 500 SB receiving flow: 500

Saturation flows:
NB Lamar -> Guadalupe: 50
NB Lamar -> Lamar: 200
SB Lamar -> Guadalupe: 200
SB Lamar -> Lamar: 200
Guadalupe -> NB Lamar: 200

15 s 45 s

Figure 9.35: Intersection for Exercise 35

34. [65] Generalize the link transmission model formulas (9.63) and (9.65) to
handle an arbitrary piecewise-linear fundamental diagram.

35. [35] Figure 9.35 represents the intersection of Lamar and Guadalupe,
showing sending and receiving flows, saturation flows, turning movement
proportions, and the signal timing plan (assume no lost time due to clear-
ance intervals or startup delay). Note that the receiving flow on north-
bound Lamar is quite low, because of congestion spilling back from a
nearby signal just downstream. No U-turns are allowed, and drivers may
not turn left from Guadalupe onto southbound Lamar.

(a) Find the transition flows yijk for all five turning movements at the
current time step, using the “smoothed signal” node model.
(b) The southbound receiving flow on Guadalupe is now reduced to 50
due to congestion further downstream. Find the updated transition
flow rates for all turning movements.

36. [51] Extend the “basic signal” node model to account for turns on red,
where a vehicle facing a red indication may make a turn in the direction
nearest to them (usually right-on-red in countries that drive on the right,
386 CHAPTER 9. NETWORK LOADING

Figure 9.36: Intersection for Exercise 39.

left-on-red in countries that drive on the left). Vehicles turning on red

must yield to traffic which has a green indication.
37. [53] Extend the “basic signal” node model to the case where there are
turn lanes, and not all turn lanes from an approach have the same green
time.
38. [61] Extend the “basic signal” node model to account for permitted turns
on green (a turning movement which has a green indication, but must
yield to oncoming traffic).
39. [36] Consider the four-legged intersection shown in Figure 9.36. The set of
turning movements is Ξ = {[1, 2, 3], [1, 2, 4], [3, 2, 1], [3, 2, 4], [5, 2, 4]}. For
the current time step, the sending and receiving flow values are S12 = 6,
S32 = 6, S52 = 1, R12 = 14, R23 = 14, and R24 = 14. Half of the
drivers approaching from link (1,2) want to turn left, and half want to
go straight. Half of the drivers approaching from link (3,2) want to turn
right, and half want to go straight. Finally, the intersection geometry
and signal timing are such that the capacities of the incoming links are
max max max
q12 = 200, q32 = 2000, and q52 = 20.
Apply the “equal priorities” algorithm, and report the transition flows (y
values) for each turning movement.
40. [36] Repeat Exercise 39, but with these values of the sending and receiving
flows: S12 = 18, S32 = 18, S52 = 3, R12 = 42, R23 = 42, and R24 = 12.
41. [68] Modify the “partial stop control” node model to allow cases of ab-
solute priority (if movement [h, i, j] has absolute priority over [h0 , i, j 0 ] at
conflict point c, then αhc 0 ij 0 /αhij
c
would be zero.) You will need to decide
how priority will be granted for any combination of sending flows wishing
to use a conflict point (including flows of equal absolute priority, and when
not all sending flows are present), and your formulas can never divide by
zero.
42. [89] Show that all of the node models in Section 9.6 satisfy all of the
desiderata from Section 9.2.
Chapter 10

Time-Dependent Shortest
Paths

This chapter discusses how travelers make choices when traveling in networks
whose state varies over time. Two specific choices are discussed: how drivers
choose a route when link costs are time-varying (Sections 10.1 and 10.2), and
how drivers choose a departure time (Section 10.3). This chapter is the comple-
ment of the previous one. In network loading, we assumed that the travelers’
choices were known, and we then determined the (time-varying) flow rates and
congestion pattern throughout the network. In this chapter, we take this con-
gestion pattern as known, and predict the choices that travelers would make
given this congestion pattern. In particular, by taking the congestion level as
fixed, we can focus the question on an individual traveler and do not need to
worry about changes in congestion based on these choices just yet.

10.1 Time-Dependent Shortest Path Concepts

The time-dependent shortest path problem involves finding a path through a
network of minimum cost, when the cost of links varies with time. As with
the static shortest path problem, the “cost” of a link can include travel time,
monetary costs, a combination of these, or any other disutility which can be
added across links; and by a “shortest” path we mean one with least cost. If a
traveler enters link (i, j) at time t, they will experience a cost of cij (t). Unlike
the static shortest path problem, however, we must always keep track of the
travel time on a link even if the cost refers to a separate quantity: because the
network state is dynamic, we must always know the time at which a traveler
enters a link. The travel time experienced by a traveler entering link (i, j) at
time t is denoted by τij (t). This problem can be formulated either in discrete
time (where t and τij (t) are limited to being integer multiples of the timestep
∆t) or in continuous time (where t can take any real value within a stated
range).

387
388 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

There are several different variations of the time-dependent shortest path

problem, all stemming from dynamic link costs and travel times. One can
consider time-dependent shortest paths where waiting at intermediate nodes is
allowed, or one can forbid it. In public transit networks, waiting at intermediate
nodes is logical, but in road networks, one would not expect drivers to volun-
tarily stop and wait in the middle of a route. One can also restrict attention to
problems where the link travel times satisfy the first-in, first-out (FIFO) prop-
erty, where it is impossible for one to leave a link earlier by entering later, that
is, for any link (i, j) and any distinct t1 < t2 , we have

t1 + τij (t1 ) ≤ t2 + τij (t2 ) . (10.1)

If this is true, more efficient algorithms can be developed. As an example, if the

cost of a link is its travel time (cij (t) = τij (t)), there is no benefit to waiting in
a FIFO network. In discrete time, equation (10.1) is equivalent to requiring

τij (t + ∆t) − τij (t) ≤ ∆t , (10.2)

that is, a link’s travel time cannot decrease by more than ∆t in one time step.
In continuous time, if τij (t) is a piecewise differentiable function, we must have
d
− τij (t) ≤ −1 , (10.3)
dt
everywhere that τij (t) has a derivative, which expresses the same idea.
One can also distinguish time-dependent shortest path problems by whether
the departure time is fixed, whether the arrival time is fixed, or neither. In
the first case, the driver has already decided when they will leave, and want to
find the route to the destination with minimum cost when leaving at that time,
regardless of the arrival time at the destination. In the second case, the arrival
time at the destination is known (perhaps the start of work), and the traveler
wants to find the route with minimum cost arriving at the destination at that
time (regardless of departure time). In the third case, both the departure and
arrival times are flexible (as with many shopping trips), and the traveler wants
to find a minimum cost route without regard to when they leave or arrive. In
this way, we can model the departure time choice decision simultaneously with
the route choice decision. Section 10.3 develops this approach further, showing
how we can incorporate penalty “costs” associated with different departure and
arrival times. Strictly speaking, one can imagine a fourth variant where both
departure and arrival time are fixed, but this problem is not always well-posed;
there may be no path from the origin to the destination with those exact depar-
ture and arrival times. In such cases, we can allow free departure and/or arrival
times, but severely penalize departure/arrival times that differ greatly from the
desired times.
In comparison with the static shortest path problem, the fixed departure
time variant is like the one origin-to-all destinations shortest path problem, and
the fixed arrival time variant is like the all origins-to-one destination shortest
path problem.
10.1. TIME-DEPENDENT SHORTEST PATH CONCEPTS 389

To be specific, in this chapter, we will first focus on time-dependent shortest

paths with fixed departure times and free arrival times, in Section 10.2, but all
of these approaches can be adapted to solve the fixed arrival time/free departure
time version without much difficulty. The origin r and departure time t0 will be
specified, and we will find the paths to all other nodes along which the sum of
the costs cij is minimal, given the changes in the costs which will occur during
travel. We will study two specific variants of the problem. In the first, time is
continuous, but the network is assumed to follow the FIFO principle and the
link costs must be the travel times. In the other variants, we do not require the
FIFO assumption, and the link costs need not be the same as the travel times;
but in exchange we will restrict ourselves to discrete time. Section 10.3 will
then treat the case of free departure and arrival times, where only the origin r
will be specified.

10.1.1 Time-expanded networks

For discrete-time shortest path problems, we can form what is known as the
time-expanded network. This technique transforms a time-dependent shortest
path problem into a static shortest path problem, that can be solved using algo-
rithms for this simpler problem (see Section 2.4). However, the time-expanded
network contains many more links and nodes than the original network, and
can impose computational burdens.
If the time step is ∆t, then the possible arrival times at any node are
0, ∆t, 2∆t, . . . , T ∆t where T is the time horizon under consideration. 1 In
the time-expanded network, we replicate each node T + 1 times, one for each
possible arrival time. So, if the original network has n nodes, the time-expanded
network has nT nodes. Each node in the time-expanded network is written in
the form i : t, denoting the physical node i at the time interval t∆t. This label-
ing is shown in Figure 10.1. The original network is called the physical network
when we need to distinguish it from the corresponding time-expanded network.
Links in the time-expanded network represent both the physical connection
between nodes, as well as the time required to traverse the link. In particular,
for each link (i, j) in the original network, and for each time interval t∆t, if the
travel time is τij (t∆t) = k∆t, we create a link (i : t, j : (t + k)) in the time
expanded network, assuming that t + k ≤ T , and set the cost of this link equal
to cij (t). A little bit of care must be taken when links would arrive at a node
later than the time horizon. If the time horizon is large enough, this should not
be a significant issue. In this chapter, we will not create a time-expanded link
if its head node falls outside of the time horizon. Exercises 4 and 5 introduce
two other ways to treat boundary issues associated with the time horizon.
The advantage of the time-expanded network is that it reduces the time-
dependent shortest path problem to the static one: solving the one-to-all static
1 In transportation systems, it is often reasonable to assume that after a large enough

amount of time, all congestion will dissipate and travel times can be treated as constants
equal to free-flow time. All routing after this point can be done with a static shortest path
algorithm.
390 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

Physical node
i j k l m

i Time 4 i:4 j:4 k:4 l:4 m:4

Time 3 i:3 j:3 k:3 l:3 m:3

j

Time 2 i:2 j:2 k:2 l:2 m:2

k m

Time 1 i:1 j:1 k:1 l:1 m:1

l
Time 0 i:0 j:0 k:0 l:0 m:0

(a) (b)

Figure 10.1: (a) Original network (b) Corresponding time-expanded network

with five physical nodes and five time intervals.

shortest path problem from node i at time t0 corresponds exactly to solving the
fixed-departure time-dependent shortest path problem. Furthermore, if all link
travel times are positive, the time-expanded network is acyclic, with the time
labels forming a natural topological order. Shortest paths on acyclic networks
can be solved rather quickly, so this is a significant advantage. Even if some
link travel times are zero (as may occur with some artificial links or centroid
connectors), the time-expanded network remains acyclic unless there is a cycle
of zero-travel time links in the network; and if that is the case, it is often
possible to collapse the zero-travel time cycle into a single node. Time-expanded
networks can also unify some of the variants of the time-dependent shortest path
problem. If waiting is allowed at a node i, we can represent that with links of
the form (i : t, i : (t + 1)) connecting the same physical node to itself, a time
step later. The FIFO principle in a time-dependent network means that two
links connecting the same physical nodes will never cross (although they may
terminate at the same node).
A disadvantage is that the number of time intervals T may be quite large.
In dynamic traffic assignment, network loading often uses a time step on the
order of a few seconds. If this same time step is used for time-dependent shortest
paths, a typical planning period of a few hours means that T is in the thousands.
Given a physical network of thousands of nodes, the time-expanded network
can easily exceed a million nodes. Even though the time-expanded network is
acyclic, this is a substantial increase in the underlying network size, increasing
both computation time and amount of computer memory required. With clever
implementations, it is often possible to avoid explicitly generating the entire
10.1. TIME-DEPENDENT SHORTEST PATH CONCEPTS 391

i1

r i2

Figure 10.2: Bellman’s principle illustrated. If the dashed path were a shorter
route from s to i2 , then it would also form a “shortcut” for the route from s to
d.

time-expanded network, and only generate links and nodes as needed.

10.1.2 Bellman’s principle in time-dependent networks

The number of paths between any two points in a network can be very large,
growing exponentially with the network size. Therefore, any approach based on
enumerating all paths and comparing their cost will not scale well to realistic
networks. So, efficient algorithms for finding shortest paths, whether time-
dependent or not, must be more clever. The key insight is called Bellman’s
principle. In the static shortest path problem, Bellman’s principle states that
any segment of a shortest path between two nodes must itself be a shortest path
between the endpoints of that segment. Otherwise, the shortest path between
the endpoints of the segment could be spliced into the original path, reducing
its cost (Figure 10.2).
In a time-dependent shortest path problem, the same general idea applies
but we must be slightly more careful about how the principle is defined. In
a FIFO network where link costs are equal to the travel time, the principle
holds identically: since it is always better to arrive at nodes as soon as possible,
any “shortcut” between two nodes in a path can be spliced into the full path,
thereby reducing its total travel time. When the FIFO principle does not hold,
or if link costs are different than travel time, this may not be true. Figure 10.3
shows two counterexamples. In the first, because the FIFO principle is violated
and waiting is not allowed, the path segment [1, 2, 3] has a shorter travel time
between nodes 1 and 3 than path [1, 3]. However, when the link (3, 4) is added
to these paths, the path [1, 2, 3, 4] has a higher travel time than path [1, 3, 4]
because link (3, 4) is entered at different times, and so its travel time is different.
In the second, path [1, 3] has a lower cost than path [1, 2, 3], even though path
[1, 3, 4] has higher cost than path [1, 2, 3, 4]. In the second case, this is true even
though arriving at any node later can only increase cost, and the same result
392 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

Scenario 1 Scenario 2


4 3 4 3
 c(2)=5
 c(3)=10

 
 c=1 
c=1

c=1
1 2 1 2



Costs and times are constant when no time index is given.

Figure 10.3: Two counterexamples to the naı̈ve application of Bellman’s prin-

ciple in time-expanded networks.

holds even if waiting is allowed.

In these cases, the correct approach is to apply Bellman’s principle to the
time-expanded network, an approach which also works in FIFO networks. Bell-
man’s principle does apply to the time-expanded network, corresponding to the
following principle in the physical network: Let π be a shortest path be-
tween nodes r and s when departing at time t0 . If i and h are two
nodes in this path, and if the arrival times at these nodes are ti and
th , respectively, then the segment of π between i and h must be a
shortest path between these nodes when departing at ti and arriving
at th . All of the algorithms in this chapter implicitly use this principle, by
allowing us to construct shortest paths a single link at a time: if we already
know the shortest path πi from the origin at t0 to some other node i at time ti ,
then any other shortest path passing through node i at time ti can be assumed
to start with πi .

10.2 Time-Dependent Shortest Path Algorithms

This section provides two algorithms for the time-dependent shortest path prob-
lem with fixed departure times. In the first, time can be modeled as either
discrete or continuous, but the network must satisfy the FIFO principle and the
cost of a link must be its travel time. In the second and third, time must be dis-
crete, but FIFO need not hold and the cost of a link may take any value. The
10.2. TIME-DEPENDENT SHORTEST PATH ALGORITHMS 393

second algorithm finds the time-dependent shortest path from a single origin
and departure time to all destinations, while the third algorithm finds the time-
dependent shortest paths from a single origin and all departure times to a single
destination. There are many other possible variants of time-dependent shortest
path algorithms, some of which are explored in the exercises — in particular,
Exercise 6 asks you to develop a time-dependent shortest path algorithm for
all origins and departure times simultaneously, which often arises in dynamic
traffic assignment software. Nevertheless, the algorithms here should give the
general flavor of how they function. Which one of these algorithms is best, or
whether another variant is better, depends on the particular dynamic traffic
assignment implementation. All of these algorithms use labels with similar (or
even identical) names, but the meanings of these labels are slightly different in
each.

10.2.1 FIFO networks

For this algorithm to apply, assume that the FIFO principle holds, in either its
discrete or continuous form. Also assume that the cost of a link is simply its
travel time, so cij (t) = τij (t) for all links (i, j) and times t. We are also given
the origin r and departure time t0 . Because the FIFO principle holds, waiting
at an intermediate node is never beneficial, and so it suffices to find the earliest
time we can reach a node and ignore all later times. This allows us to adapt
Dijkstra’s algorithm (Section 2.4.3) for the static shortest path problem, a label-
setting approach. Two labels are defined for each node: Li gives the earliest
possible arrival time to node i found so far, when departing origin r at time t0 .
The backnode label qi gives the previous node in a path corresponding to this
earliest known arrival time. By the dynamic version of Bellman’s principle in
FIFO networks, this suffices for reconstructing the shortest path from r to i.
For nodes i where we have not yet calculated the earliest possible arrival time,
we will set Li to ∞, and for nodes i where the backnode is meaningless (either
because we have not yet found a path there, or because i is the origin), we set
qi to −1. We also maintain a set of finalized nodes F . If node i belongs to F ,
this means that we are sure that we have found the earliest possible arrival time
to i over all possible paths.
The algorithm functions as follows:

1. Initialize by setting Li ← ∞ for all nodes i, except for the origin, where
Lr ← t0 .

2. Initialize the set of finalized nodes F ← ∅, and the backnode vector q ←

−1.

3. Choose an unfinalized node i with the lowest Li value. (At the first iter-
ation, this will be the origin r.)

4. Finalize node i: F ← F ∪ {i}.

394 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

5. For each link (i, j) ∈ Γ(i) such that Li + τij (Li ) is within the time horizon,
perform the following steps:
(a) Update Lj ← min {Lj , Li + τij (Li )}.
(b) If Lj changed in the previous step, update qj ← i.
6. If all nodes are finalized (F = N ), terminate. Otherwise, return to step 3.
As an example of this algorithm, consider the network in Figure 10.4. The
time-dependent travel times are shown in this figure. The FIFO assumption is
satisfied: for links (1, 3) and (2, 4) the travel times are constant; for links (1, 2)
and (3, 4) the travel times are increasing (so arriving earlier always means leaving
earlier); and for link (2, 3) the travel time is decreasing, but at a slow enough
rate that you cannot leave earlier by arriving later, cf. (10.3). Assume that the
initial departure time from node 1 is at t0 = 2, and that the time horizon is large
enough that Li + τij (Li ) is always within the time horizon whenever step 5 is
encountered. The steps of the algorithm are explained below, and summarized
in Table 10.1. This table shows the state of the algorithm just before step 3 is
performed, counting the first time through as iteration zero.
Initially, no nodes are finalized, all cost labels are initialized to ∞ (except
for the origin, which is assigned 2, the departure time), and all backnode labels
are initialized to −1. The unfinalized node with the least L value is node 1,
which is selected as the node i to scan. At the current time of 2, link (1, 2) has
a travel time of 6, and link (1, 3) has a travel time of 10. Following these links
would result in arrival at nodes 2 and 3 at times 8 and 12, respectively. Each of
these is less than their current values of ∞, so the cost and backnode labels are
adjusted accordingly. At the next iteration, node 2 is the unfinalized node with
the least L value, so i = 2. At time 8, link (2, 3) has a travel time of 1, and link
(2, 4) has a travel time of 5. Following these links, one would arrive at nodes
3 and 4 at times 9 and 13, respectively. Both of these values are less than the
current L values for these nodes, so their L and q labels are changed. At the
next iteration, node 3 is the unfinalized node with the least L value, so i = 3.
At this time, link (3, 4) would have a travel time of 4 21 , and choosing it means
arriving at node 4 at time 13 12 . This is greater than the current value (L4 = 13),
so no labels are adjusted. Finally, node 4 is chosen as the only unfinalized node.
Since it has no outgoing links (Γ(4) is empty), there is nothing to do in step 5,
and since all nodes are finalized the algorithm terminates.
At this point, we can trace back the shortest paths using the backnode
labels: the shortest paths to nodes 2, 3, and 4 are [1, 2], [1, 2, 3], and [1, 2, 4],
respectively; and following these paths one arrives at the nodes at times 8, 9,
and 13.

10.2.2 Discrete-time networks, one departure time

This algorithm applies in any discrete-time network, regardless of whether or
not the FIFO principle holds, and regardless of whether cij (t) = τij (t) or not.
We are given the origin r and departure time t0 , and work in the time-expanded
10.2. TIME-DEPENDENT SHORTEST PATH ALGORITHMS 395

3
t
10 2

max 5 − 2t , 1

1 4

4+t 5

Figure 10.4: Network and travel time functions for demonstrating the FIFO
time-dependent shortest path algorithm.

Table 10.1: FIFO time-dependent shortest path algorithm for the network in
Figure 10.4, departing node 1 at t0 = 2.
Iteration F i L1 L2 L3 L4 q1 q2 q3 q4
0 ∅ — 2 ∞ ∞ ∞ −1 −1 −1 −1
1 {1} 1 2 8 12 ∞ −1 1 1 −1
2 {1, 2} 2 2 8 9 13 −1 1 2 2
3 {1, 2, 3} 3 2 8 9 13 −1 1 2 2
4 {1, 2, 3, 4} 4 2 8 9 13 −1 1 2 2
396 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

network. Each node i : t in the time-expanded network is associated with two

labels. The label Lti denotes the cost of the shortest path from r to i known
so far, when departing at time t0 and arriving at time t. The backnode label
qit provides the previous node on a path corresponding to cost Lti . As before,
Lti = ∞ signifies that no path to i at time t is yet known, and qit = −1 signifies
that the backnode is meaningless. Since the FIFO principle may not hold, it
is not always advantageous to arrive at a node as early as possible. To reflect
this, we work in the time-expanded network, where we can naturally identify
the best time to arrive at nodes. We presume that all links have strictly positive
travel time, so that the time-expanded graph is acyclic and each link connects
a node of earlier time to a node of later time. If there are zero-travel time links
in the physical network, but no cycles of zero-travel time links, then step 4 is
assumed to proceed in topological order by physical node. In this case, we can
adapt an algorithm for finding static shortest paths (e.g., from Section 2.4), in
acyclic networks by applying it to the time-expanded network in the following
way:

1. Initialize q ← −1 and L ← ∞.

2. Set Ltr0 ← 0.

3. Initialize the current time to the departure time, t ← t0

4. For each time-expanded node i : t for which Lti < ∞, and for each time-
expanded link (i : t, j : t0 ), perform the following steps:
0
n 0 o
(a) Set Ltj = min Ltj , Lti + cij (t) .
0 0
(b) If Ltj changed in the previous step, update qjt ← i : t.

5. If t = T , then terminate. Otherwise, move to the next time step (t ← t+1)

and return to step 4.

At the conclusion of this algorithm, we have the least-cost paths for each
possible arrival time at each destination. To find the least-cost path to a par-
ticular destination s (at any arrival time), you can consult the Lts labels at all
times t, and trace back the path for the arrival time t with the least Lts value.
This algorithm is demonstrated on the network in the right panel of Fig-
ure 10.3, and its progress is summarized in Table 10.2. The table shows the
state of the algorithm just before step 4 is executed. For brevity, this table
only reports Lti and qit labels for nodes and arrival times which are reachable
in the network (that is, i and t values for which Lti < ∞ at the end of the
algorithm). All other cost and backnode labels are at ∞ and −1 throughout
the entire duration of the algorithm.
Initially, all cost labels are set to ∞ and all backnode labels to −1, except
for the origin and departure time: L01 = 0. The algorithm then sets t = 0, and
10.2. TIME-DEPENDENT SHORTEST PATH ALGORITHMS 397

Table 10.2: Discrete time-dependent shortest path algorithm for the network in
the right panel of Figure 10.3, departing node 1 at t0 = 0.
t L01 L12 L23 L33 L34 L44 q10 q21 q32 q33 q43 q44
0 0 ∞ ∞ ∞ ∞ ∞ −1 −1 −1 −1 −1 −1
1 0 1 ∞ 1 ∞ ∞ −1 1 −1 1 −1 −1
2 0 1 2 1 ∞ ∞ −1 1 2 1 −1 −1
3 0 1 2 1 7 ∞ −1 1 2 1 3 −1
4 0 1 2 1 7 11 −1 1 2 1 3 3

scans over all physical nodes which are reachable at this time2 Only node 1 can
be reached at this time, and the possible links are (1 : 0, 2 : 1) and (1 : 0, 3 : 3).
Following either link incurs a cost of 1, which is lower than the (infinite) values
of L12 and L33 , so the cost and backnode labels are updated.
The algorithm then sets t = 1. Only node 2 is reachable at this time, and
the only link is (2 : 1, 3 : 2). Following this link incurs a cost of 1; in addition
to the cost of 1 already involved in reaching node 2, this gives a cost of 2 for
arriving at node 3 at time 2. The cost and time labels for 3 : 2 are updated.
Since the costs and times are different, notice that arriving at node 3 at a later
time (3 vs. 2) incurs a lower cost (1 vs. 2). This is why we need to track labels
for different arrival times, unlike the algorithm in the previous section.
The next time step is t = 2. Only node 3 is reachable at this time (from the
path [1, 2, 3]), and the only link is (3 : 2, 4 : 3). Following this link incurs a cost
of 5, resulting in a total cost of 7, and the labels for 4 : 3 are updated. Time
t = 3 is next, and again only node 3 is reachable at this time — but from the
path [1, 3]. The only link is (3 : 3, 4 : 4), and following this link incurs a cost of
10, for a total cost of 11. Labels are updated for 4 : 4. There are no further label
changes in the algorithm (all nodes have already been scanned at all reachable
times), and it terminates as soon as t is increased to the time horizon.
After termination, the Lt3 labels show that we can reach node 3 either with
a cost of 1 (arriving at time 3) or a cost of 2 (arriving at time 2). The least-cost
path thus arrives at time 3, and it is [1, 3]. The Lt4 labels show that we can
reach node 4 either with a cost of 7 (arriving at time 3) or a cost of 11 (arriving
at time 4). The least-cost path arrives at time 3, and it is [1, 2, 3, 4]. (The
least-cost path to node 2 is [1, 2], since there is only one possible arrival time
there). Notice that the naı̈ve form of Bellman’s principle is not satisfied: the
least-cost path to node 2 is not a subset of the least-cost path to node 3. This
was why we needed to keep track of different possible arrival times to nodes
— the least-cost path to node 3 arriving at time 2 is indeed a subset of the
least-cost path to node 4 arriving at time 3.

2 Reachability is expressed by the condition Lt < ∞; a node which cannot be reached at

i
this time will still have its label set to the initial value. Any node which can be reached at
this time will have a finite Lti value, since step 4a will always reduce an infinite value.
398 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

10.3 Departure Time Choice

Travelers often have some flexibility when choosing their departure or arrival
times, and may choose to leave earlier or later in order to minimize cost. For
example, commuters with flexible work hours may want to time their commutes
to avoid congestion, or when dynamic congestion charges are lower. Both flexible
departures and arrivals can be incorporated into the shortest path algorithms
described in the previous section — in fact, you may have already noticed that
the algorithms in Sections 10.2.1 and 10.2.2 allow flexible arrival times. This
section explores these choices more systematically, showing how the departure
time can also be made flexible, and ways to represent different departure time
behaviors.

10.3.1 Artificial origins and destinations

In the time-expanded network, departure and arrival time choice can be modeled
by adding artificial “super-origin” and “super-destination” nodes which reflect
the start and end of a trip without regard to the time. A super-origin is con-
nected to time-expanded nodes which correspond to allowable departure times,
and a super-destination is connected to time-expanded nodes which correspond
to allowable arrival times. Initially, we will assign these links a cost of zero,
which means the traveler is indifferent among any of these departure or arrival
times. Section 10.3.2 describes how arrival and departure time preferences can
be modeled.
Figure 10.5 shows how super-origins and super-destinations can be added
to the physical network of Figure 10.1(a), assuming that nodes i and k are
origins, and nodes j and l are destinations. The figure is drawn as if departures
were allowed only for times 0, 1, and 2, but arrivals are allowed at any time.
The network is similar to the time-expanded network in Figure 10.1(b), but
now includes four artificial nodes, and artificial links corresponding to allowable
departure and arrival times.
Once these artificial links and nodes are added to the network, the algorithm
from Section 10.2.2 can be applied directly, with only very minor changes. In
what follows, r is the origin and the algorithm finds least-cost paths to all
destinations, for any allowable departure and arrival time:

1. Initialize q ← −1 and L ← ∞ for all nodes in the time-expanded network.

2. Set Lr ← 0 for the super-origin r, and for each artificial link (r, r : t) set
Ltr ← cr,depart (t) and qrt ← r, where cr,depart (t) is the cost on the artificial
link for departing node r at time t.

3. Initialize the current time t to the earliest possible departure time (the
lowest t index for which an artificial link (r, r : t) exists).

4. For each time-expanded node i : t for which Lti < ∞, and for each time-
expanded link (i : t, j : t0 ), perform the following steps:
10.3. DEPARTURE TIME CHOICE 399

Physical node
i j k l m

Time 4 i:4 j:4 k:4 l:4 m:4

Time 3 i:3 j:3 k:3 l:3 m:3

Time 2 i:2 j:2 k:2 l:2 m:2

Time 1 i:1 j:1 k:1 l:1 m:1

Time 0 i:0 j:0 k:0 l:0 m:0

i j k l

Figure 10.5: Time-expanded network with artificial origins and destinations.

Artificial links shown with light arrows for clarity.

0
n 0 o
(a) Set Ltj = min Ltj , Lti + cij (t) .
0 0
(b) If Ltj changed in the previous step, update qjt ← i : t.

Likewise, for each artificial link (i : t, i) reaching a super-destination node

i with cost ci,arrive (t), perform the following steps:

(a) Set Li = min {Li , Lti + ci,arrive (t)}.

(b) If Li changed in the previous step, update qi ← i : t.

5. If t = T , then terminate. Otherwise, move to the next time step (t ← t+1)

and return to step 4.

The algorithm initializes labels differently in step 2; step 3 starts at the earliest
possible departure time rather than the fixed time t0 ; and step 4 is expanded
to update labels both at adjacent time-expanded nodes and super-destinations.
All other steps work in the same way.
To demonstrate this algorithm, consider the network in Figure 10.6, where
the time horizon is T = 20 and the destination is node 4, and where the cost of a
link is equal to its travel time. Table 10.3 shows the cost and forwardnode labels
at the conclusion of the algorithm. Each iteration of the algorithm generates
one row of this table, starting with T = 1 and working up to T = 20. Whenever
Lti = ∞ is seen in Table 10.3, there is no way to reach the destination node
within the time horizon, if leaving node i at time t.
400 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

1 + 2t 5

1 max {10 − t, 0} 4

5 1+t

Figure 10.6: Network and travel time functions for demonstrating the all de-
parture times time-dependent shortest path algorithm.

Table 10.3 also shows the labels for the super-origin and super-destination,
below the labels for the time-expanded nodes. The backnode label for the super-
destination tells us that the least-cost path arrives at node 4 at time 6; the q46
label tells us the least-cost path there comes through node 3 at time 1; q31 tells
us the least-cost path there comes through node 1 at time 0, and q10 brings
us to the super-origin. Therefore, we should depart the origin at time 0, and
follow the path [1, 3, 4] to arrive at the destination at time 6, with a total cost
of c4 = 6.

10.3.2 Arrival and departure time preferences

While there is often flexibility in departure or arrival time, usually travelers are
not completely indifferent about when they depart or arrive. For instance, there
may be a well-defined arrival deadline (start of work, or check-in time before a
flight), and a strong desire to arrive before this deadline. Departing extremely
early would ensure arriving before the deadline, but carries opportunity costs
(by departing later, the traveler would have more time to do other things). Both
situations can be modeled by attaching a cost to when a traveler departs the
origin, and to when they arrive at the destination.
A common way to model arrival costs is with the schedule delay concept.
In this model, travelers have a preferred arrival time t∗ at the destination,
and arriving either earlier or later than t∗ is undesirable. The most general
formulation involves a function f (t) denoting the cost (or disutility) of arriving
at the destination at time t. This function is typically convex and has a minimum
at t∗ . One example of such a function is

f (t) = α[t∗ − t]+ + β[t − t∗ ]+ , (10.4)

where [·]+ = max {·, 0} expresses the positive part of the quantity in brackets.
In (10.4), the first bracketed term then represents the amount by which the
10.3. DEPARTURE TIME CHOICE 401

Table 10.3: Labels at conclusion of time-dependent shortest path algorithm with

departure time choice, to node 4 in Figure 10.6.
t Lt1 Lt2 Lt3 Lt4 q1t q2t q3t q4t
20 0 5 5 10 1 1 : 15 2 : 20 3 : 15
19 0 5 5 10 1 1 : 14 2 : 19 3 : 14
18 0 5 5 10 1 1 : 13 2 : 18 3 : 13
17 0 5 5 10 1 1 : 12 2 : 17 3 : 12
16 0 5 5 10 1 1 : 11 2 : 16 3 : 11
15 0 5 5 10 1 1 : 10 2 : 15 3 : 10
14 0 5 5 ∞ 1 1 : 9 2 : 14 −1
13 0 5 5 12 1 1 : 8 2 : 13 2 : 6
12 0 5 5 10 1 1 : 7 2 : 12 3 : 7
11 0 5 5 11 1 1 : 6 2 : 11 2 : 5
10 0 5 5 ∞ 1 1 : 5 2 : 10 −1
9 0 5 ∞ 8 1 1:4 −1 3:4
8 0 5 ∞ ∞ 1 1:3 −1 −1
7 0 5 5 ∞ 1 1:2 1:2 −1
6 0 5 ∞ 6 1 1:1 −1 3:1
5 0 5 ∞ ∞ 1 1:0 −1 −1
4 0 ∞ 3 ∞ 1 −1 1:1 −1
3 0 ∞ ∞ ∞ 1 −1 −1 −1
2 0 ∞ ∞ ∞ 1 −1 −1 −1
1 0 ∞ 1 ∞ 1 −1 1:0 −1
0 0 ∞ ∞ ∞ 1 −1 −1 −1

(L, q) labels at super-origin 1: (0, −1)

(L, q) labels at super-destination 4: (6, 4 : 6)
402 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

traveler arrived early, compared to the preferred time, and the second bracketed
term represents the amount by which the traveler arrived late. The coefficients
α and β then weight these terms and convert them to cost units; generally
α < β to reflect the fact that arriving early by a certain amount of time, while
undesirable, is usually not as bad as arriving late by that same amount of time.
Another possible function is nonlinear, taking a form such as

f (t) = α([t∗ − t]+ )2 + β([t − t∗ ]+ )2 . (10.5)

In this function, the penalty associated with early or late arrival grows faster
and faster the farther the arrival time from the target. This may occur if, for
instance, being ten minutes late is more than ten times as bad as being one
minute late. Special cases of these functions arise when α = β (the function
becomes symmetric), or when α = 0 (there is no penalty for early arrival, but
only for late arrival). This function is also differentiable everywhere, in contrast
to equation (10.4) which is not differentiable at t∗ . For certain algorithms this
may be advantageous.
To each of these schedule delay functions f (t), one can add the cost of the
path arriving at time t (the sum of the link costs cij along the way) to yield the
total cost of travel. We assume that travelers will choose both the departure
time and the path to minimize this sum. The algorithm from Section 10.3.1 can
be used to find both this ideal departure time and the path. The only change is
that artificial links connecting time-expanded destination nodes (s : t) to super-
destinations s now have a cost of f (t), rather than zero. At termination, the
t∗
departure time t∗0 minimizing Lr0 corresponds to the least total cost, and the
forwardnode labels trace out the path.
To demonstrate this algorithm, again consider the network in Figure 10.6,
but with the arrival time penalty function

f (t) = [10 − t]+ + 2[t − 10]+ , (10.6)

which suggests that the traveler wishes to arrive at time 10, and that late arrival
is twice as costly as early arrival.
As before, the time horizon is T = 20 and the destination is node 4, and
the cost of each link is equal to its travel time. Table 10.4 shows the cost and
backnode labels at the conclusion of the algorithm, which runs in exactly the
same way as before except that the artificial destination links (4 : t, 4) have a
cost equal to f (t). At the conclusion of the algorithm, we can identify the total
cost on the shortest path from node 1 to node 4, now including the penalty for
arriving early or late at the destination.
Table 10.4 also shows the labels for the super-origin and super-destination,
below the labels for the time-expanded nodes. The backnode label for the
super-destination tells us that the least-cost path arrives at node 4 at time 9;
the q49 label tells us the least-cost path there comes through node 3 at time 4;
q34 tells us the least-cost path there comes through node 1 at time 1, and q11
brings us to the super-origin. Therefore, we should depart the origin at time 1,
and follow the path [1, 3, 4] to arrive at the destination at time 9, with a total
10.3. DEPARTURE TIME CHOICE 403

Table 10.4: Labels at conclusion of arrival-penalty time-dependent shortest

path, to node 4 in Figure 10.6.
t Lt1 Lt2 Lt3 Lt4 q1t q2t q3t q4t
20 0 5 5 10 1 1 : 15 2 : 20 3 : 15
19 0 5 5 10 1 1 : 14 2 : 19 3 : 14
18 0 5 5 10 1 1 : 13 2 : 18 3 : 13
17 0 5 5 10 1 1 : 12 2 : 17 3 : 12
16 0 5 5 10 1 1 : 11 2 : 16 3 : 11
15 0 5 5 10 1 1 : 10 2 : 15 3 : 10
14 0 5 5 ∞ 1 1 : 9 2 : 14 −1
13 0 5 5 12 1 1 : 8 2 : 13 2 : 6
12 0 5 5 10 1 1 : 7 2 : 12 3 : 7
11 0 5 5 11 1 1 : 6 2 : 11 2 : 5
10 0 5 5 ∞ 1 1 : 5 2 : 10 −1
9 0 5 ∞ 8 1 1:4 −1 3:4
8 0 5 ∞ ∞ 1 1:3 −1 −1
7 0 5 5 ∞ 1 1:2 1:2 −1
6 0 5 ∞ 6 1 1:1 −1 3:1
5 0 5 ∞ ∞ 1 1:0 −1 −1
4 0 ∞ 3 ∞ 1 −1 1:1 −1
3 0 ∞ ∞ ∞ 1 −1 −1 −1
2 0 ∞ ∞ ∞ 1 −1 −1 −1
1 0 ∞ 1 ∞ 1 −1 1:0 −1
0 0 ∞ ∞ ∞ 1 −1 −1 −1

(L, q) labels at super-origin 1: (0, −1)

(L, q) labels at super-destination 4: (9, 4 : 9)

cost of c4 = 9. Of this cost, 8 units are due to travel time (difference between
arrival and departure times), and 1 unit is due to the arrival time penalty from
equation (10.6) with t = 9.
Departing earlier, at t = 0, on the same path would reduce the travel time
to 6, but increase the early arrival penalty to 4. The total cost of leaving at
t = 0 is thus higher than departing one time step later. Leaving at t = 2 and
following the same path increases the travel time cost to 10. Since this means
arriving at time 12, there is a late penalty cost of 4 added, resulting in a total
travel cost of 14. By comparing all possible paths and departure times, you can
verify that it is impossible to have a total cost less than 9.
Considering costs associated with departure time, rather than arrival time, is
done in essentially the same way, by assigning a nonzero cost to the artificial links
connecting the super-origin r to the time-expanded nodes r(t). It is thus possible
to have only departure time penalties, only arrival time penalties, both, or
neither, depending on whether the artificial origin links and artificial destination
links have nonzero costs.
404 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

10.4 Dynamic A∗
Because the time-expanded network transforms the time-dependent shortest
problem into the classical, static shortest path problem from Section 2.4, any
of the algorithms described there can be applied. The time-expanded network
is acyclic, so it is fastest to use the algorithm from Section 2.4.1 — and indeed
this is all that the algorithm in Section 10.2.2 is, using the time labels as a
topological order.
Section 2.4.4 also presented the A∗ algorithm, which provides a single path
from one origin to one destination, rather than all shortest paths from one origin
to all destinations, or all origins to one destination. By focusing on a single
origin and destination, A∗ can often find a shortest path much faster than a
one origin-to-all destinations algorithm. The tradeoff is that the algorithm has
to repeated many times, once for every OD pair, rather than once for every
origin or destination. In static assignment, one-to-all or all-to-one algorithms
are preferred because, because in many cases, running A* for each OD pair
takes more time than running a one-to-all algorithm for each origin.
In dynamic traffic assignment with fixed departure times, however, the num-
ber of “origins” in the time-expanded network is multiplied by the number of
departure times. If one were to write a full time-dependent OD matrix, the
number of entries in this matrix is very large: in a network with 1000 centroids
and 1000 time steps, there are 1 billion entries, one for every origin, every des-
tination, and departure time. This is much larger than the number of vehicles
that will be assigned, so almost every entry in this matrix will be zero. In such
cases, A∗ can work much better, only being applied to origins, destinations, and
departure times with a positive entry in the matrix.
As discussed in Section 2.4.4, an effective estimate for A∗ in traffic assign-
ment problems is to use the free-flow travel costs. As a preprocessing step at
the start of traffic assignment, you can use an all-to-one static shortest path
algorithm to find the least-cost travel cost gjs from every node j to every des-
tination s at free flow. For the remainder of the traffic assignment algorithm,
you can then use gjs as the estimates for A∗ . This is quite effective in practical
networks.

10.5 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
Several authors have discussed how the time-dependent shortest path prob-
lem differs from the static case; both label-setting (Dreyfus, 1969) and label-
correcting (Cooke and Halsey, 1969; Ziliaskopoulos and Mahmassani, 1994)
methods exist. Chabini (1999) showed how the time index as a topological
order in the time-expanded network. The use of free-flow times as estimates in
time-dependent A∗ is from Boyles (2013).
Section 2.5 described how the shortest path problem can be generalized to
10.6. EXERCISES 405

Table 10.5: Link costs for Exercises 1 and 10.

Entry time Link 1 Link 2 Link 3 Link 4 Link 5
0 5 1 3 1 1
1 6 1 3 6 1
2 4 2 3 5 2
3 5 3 3 4 4
4 7 5 3 3 2
5 8 8 3 2 1

include stochastic link costs, where the link costs are drawn from some prob-
ability distribution and the path with minimum expected cost is sought. You
may recall that this problem was not too difficult to address in the static case.
However, if travel times are both time-dependent and stochastic, more care is
needed (Hall, 1986; Fu and Rilett, 1998), because Bellman’s principle need not
hold. Examples of algorithms to handle this issue are given in Hall (1986) and
Miller-Hooks and Mahmassani (2000).

10.6 Exercises
1. [13] Table 10.5 shows time-dependent costs on five links, for different entry
times. Which links have costs satisfying the FIFO principle?
2. [23] Prove that waiting is never beneficial in a FIFO network where link
costs are equal to travel time.
3. [34] Prove that there is an acyclic time-dependent shortest path in a FIFO
network where link costs are equal to travel time.
4. [53] In this chapter, we assumed it is impossible to travel beyond the time
horizon, by not creating time-expanded links if they arrive at a down-
stream node after T̄ . Another alternative is to assume that travel beyond
T̄ is permitted, but that travel times and costs stop changing after that
point and take constant values. (Perhaps free-flow times after the peak
period is over.)
(a) Modify the time-expanded network concept to handle this assump-
tion. (The network should remain finite.)
(b) Modify the algorithm in Section 10.2.2 to work in this setting.
5. [63] Another way to handle the time horizon is to assume that the link
travel times and costs are periodic with length T̄ . (For instance, T̄ may
be 24 hours, and so entering a link at t = 25 hours would be the same
as at t = 1 hour.) First repeat Exercise 4 with this assumption. Then
prove that the modified algorithm you create will converge to the correct
time-dependent shortest paths for any possible departure time.
406 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

Table 10.6: Backnode labels for Exercise 9.

Entry time Node 1 Node 2 Node 3 Node 4
6 −1 4 3 2
5 −1 4 3 −1
4 −1 −1 −1 −1
3 −1 4 −1 −1
2 −1 4 −1 −1
1 −1 −1 −1 −1

Table 10.7: Cost labels for Exercise 9.

Entry time Node 1 Node 2 Node 3 Node 4
6 0 1 5 5
5 0 2 4 ∞
4 0 ∞ ∞ ∞
3 0 2 ∞ ∞
2 0 1 ∞ ∞
1 0 ∞ ∞ ∞

6. [31] Modify the algorithm in Section 10.3.1 to become an all-to-one algo-

rithm, that finds least-cost paths from all origins and all departure times
to one destination (arrival at any time is permitted).

7. [34] Show that the classical form of Bellman’s principle (Section 2.4 holds
in a FIFO, time-dependent network where the link costs are equal to the
link travel time.)

8. [32] Find the time-dependent shortest path when departing node 1 in the
network shown in the left panel of Figure 10.3, departing at time 0.

9. [23] Tables 10.6 and 10.7 show the backnode and cost labels for a time-
dependent shortest path problem, where the destination is node 4, and
the time horizon is 6. (Figure 10.7 shows the network topology.) What is
the shortest path from node 1 to node 4, when departing at time 1?

10. [22] In the same network as Exercise 9, fill in the forwardnode and cost
labels for time 0. The link costs are as in Table 10.5, and waiting is not
allowed at nodes.

11. [45] Consider the network in Figure 10.8.

(a) Verify that the travel times satisfy the FIFO principle.
(b) Find the shortest paths between nodes 1 and 4 when departing at
t = 0, t = 2, and t = 10.
10.6. EXERCISES 407

Figure 10.7: Network for Exercises 9 and 10.

3
10 5

1 max{8 – t/2, 1} 4

2+t 2+t
2

Figure 10.8: Network for Exercise 11.

(c) For what departure times would the travel times on paths [1, 2, 4] and
[1, 3, 4] be equal?

12. [23] Solve the time-dependent shortest path problem on the network in
Figure 10.6 using the algorithm in Section 10.2.1. Comment on the amount
of work needed to solve the problem using this algorithm, compared to the
way it was solved in the text.
13. [51] In the schedule delay equation (10.4), we typically assume 0 ≤ α ≤
1 ≤ β for peak-hour commute trips. What counterintuitive behavior would
occur if any of these three inequalities were violated?
14. [12] Find the optimal departure times and paths from nodes 2 and 3 in the
network in Figure 10.6, with equation (10.6) as the arrival time penalty.
(You can answer this question directly from Table 10.3.)
15. [36] Find the optimal departure times and paths from nodes 1, 2, and 3 in
the network in Figure 10.6, if the arrival time penalty function is changed
to f (t) = ([12 − t]+ )2 + 2([t − 12]+ )2 .
16. [0] According to the penalty function in Exercise 15, what is the desired
arrival time?
17. [88] Implement all of the algorithms described in this chapter, and test
them on transportation networks with different characteristics (number of
408 CHAPTER 10. TIME-DEPENDENT SHORTEST PATHS

origins, destinations, nodes, links, time intervals, ratio of links to nodes,

with or without waiting, FIFO or non-FIFO, etc.). What algorithms per-
form best under what circumstances? For circumstances most resembling
real-world transportation networks, what performs best?
Chapter 11

Dynamic User Equilibrium

The previous two chapters presented separate perspectives on dynamic traffic

modeling. Chapter 9 described the network loading problem, in which driver be-
havior (as represented by path choices and possibly departure times) was known,
and where we sought to represent the resulting traffic and congestion patterns
on the network. Chapter 10 described the time-dependent shortest path and
departure time choice problems, in which the network state (as represented by
travel times and costs) was known, and where we sought to identify how drivers
would behave. This chapter synthesizes these two perspectives through the con-
cept of dynamic user equilibrium, defined as driver choices and network traffic
which are mutually consistent, given the network loading and driver behavior
assumptions. The dynamic user equilibrium principle, and how the network
loading and driver behavior models can be connected, are the subjects of Sec-
tion 11.1. Section 11.2 then presents several algorithms that can be used to solve
for dynamic equilibria. Such algorithms are almost always heuristics, since re-
alistic network loading models are not amenable to exact analysis, and indeed
Section 11.3 shows that dynamic user equilibrium need not exist; and if it ex-
ists, it need not be unique. This section also provides examples to show that
the dynamic user equilibrium solution need not minimize total travel time in a
network, and that providing additional capacity to a network can increase total
travel time, a dynamic analogue of the Braess paradox.

11.1 Towards Dynamic User Equilibrium

This section has two major goals, aimed at reconciling the network loading prob-
lem of Chapter 9 and the time-dependent shortest path algorithms of Chap-
ter 10. In particular, we will need to solve these problems sequentially, and
repeatedly, as shown in Figure 11.1. The first goal of this section is to provide
the mechanics to link the output of each problem to the input needed by the
other. Following these, we will be equipped to define dynamic user equilibrium
formally. Fixed point and variational inequality formulations are given as well.

409
410 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

11.1.1 Flow Representation

The first question to address is how to represent the choices of all the travelers
on the network. In describing network loading, Chapter 9 took an aggregate
approach. All of the link models in that chapter can be seen as fluid models,
where vehicles are infinitely divisible rather than discrete entities. However, the
algorithms in Chapter 10 took a disaggregate approach, and found paths and
departure times for a single vehicle. There are several ways to reconcile these
perspectives.
The method adopted in this chapter is to store the number of vehicles de-
parting on each path during each time interval. Denote by hπt the number of
vehicles which start traveling on path π during the t-th interval. These values do
not need to be integers; all that matters is that they are nonnegative, and that
all paths connecting the same origin and destination sum to the total demand
between these zones for each time interval. The hπt values can be collected into
a single matrix H, whose dimensions are equal to the number of paths and time
intervals. With this notation, the feasible set can be written as
( )
T̄ ×|Π|
X
π rs 2
H̄ = H ∈ R+ : ht = dt ∀(r, s) ∈ Z , t . (11.1)
π∈Πrs

The main advantage of this notation is that the behavior is clear: by tracking
some auxiliary variables in the network loading (as described in Section 11.1.3,
at any point in time we can see exactly which vehicles are on which links, and can
trace these to the paths the vehicles must follow). These paths can be connected
directly with the paths found in a time-dependent shortest path algorithm. A
disadvantage of this approach is that the number of paths grows exponentially
with the network size. In practice, “column generation” schemes are popular,
in which paths π are only identified when found by a shortest path algorithm,
and hπt values only need be calculated and stored for paths to which travelers
have been assigned.
This is not the only possible way to represent travel choices. A link-based
representation requires fewer variables. For each turning movement [h, i, j],
t
each destination s, and each time interval t, let αhij,s denote the proportion of
travelers arriving at node i, via approach (h, i), during the t-th time interval,
who will exit onto link (i, j) en route to destination s. This approach mimics the
flow splitting rules often found in traffic microsimulation software. The number
of variables required by this representation grows with the network size, but
not at the exponential rate required by a path-based approach. One can show
that the path-based and link-based representation of route choice are equivalent
in the sense that there exist α values which represent the same network state
as any feasible set of H values, regardless of the network loading model, and
that one can identify the H values corresponding to a given set of α values
(see Exercise 1). The primary disadvantage of this representation is that the
behavior is less clear: one cannot trace the path of any vehicle throughout the
network deterministically (although one can do so stochastically, making a turn
11.1. TOWARDS DYNAMIC USER EQUILIBRIUM 411

Travel
times

1. Network loading 2. Time-dependent

shortest paths

Updated path
flows Target paths

3. Adjust route choices

toward equilibrium

Figure 11.1: Iterative dynamic traffic assignment framework, with details from
previous chapters.

at each node by treating the α values as probabilities, identifying the arrival

times at successive nodes using the procedure in Section 11.1.2).
Both of these methods adopt the aggregate, continuous-flow perspective of
Chapter 9. Yet another way to represent travel choices is to adopt the indi-
vidual perspective of the algorithms in Chapter 10, and explicitly model each
vehicle as a discrete agent with one specific path. This is the most behavioral
way to represent choices: each individual vehicle is assigned one path, with no
divisibility or fractional flows. Downsides of this approach are scalability —
if the demand doubles but the network topology is unchanged, the individual
vehicle approach would require twice as many variables, whereas the continuous
methods would not require any more variables — and increased difficulty in
implementing the network loading algorithms. It is certainly possible to imple-
ment discrete versions of merges, diverges, the cell transmission model, and so
forth, but one must be careful about rounding. For instance, if the time step
is chosen so that the capacity of a link is less than half a vehicle per timestep,
consistently rounding numbers to the nearest integer would mean no vehicles
can ever exit. Stochastic rounding, or accumulating a continuous flow value
which can then be rounded, can address these difficulties.

11.1.2 Travel time calculation

Once network loading is complete, a time-dependent shortest path algorithm
can be applied to determine the cost-minimizing routes for travelers. To do so,
we need the travel times τij (t) on each link (i, j) for travelers entering at each
time t. The network loading models do not provide this directly, but rather give
the cumulative counts N at the upstream and downstream ends of each link,
at each time interval. However, this information will suffice for calculating the
↑
travel times. For link (i, j) at time t, the upstream count Nij (t) gives the total
number of vehicles which have entered the link by time t, while the downstream
↓
count Nij (t) gives the total number of vehicles which left the link by time t.
↑ ↓
Assume for a moment that Nij (t) and Nij (t) are strictly increasing, continu-
412 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

N
Upstream

Downstream

(t)
t
t

Figure 11.2: Obtaining travel times from plots of the cumulative counts N ↑ and
N ↓.

ous functions of t. If this is the case, we can define inverse functions Tij↑ (n) and
Tij↓ (n), respectively giving the times when the n-th vehicle entered the link and
left the link. The travel time for the n-th vehicle is then the difference between
these: Tij↓ (n) − Tij↑ (n). Graphically, this can be seen as the horizontal difference
between the upstream and downstream N curves. (Figure 11.2). Then, to find
the travel time for a vehicle entering the link at time t, we simply evaluate this
↑
difference for n = Nij (t):

τij (t) = Tij↓ (Nij

↑
(t)) − Tij↑ (Nij
↑
(t)) = Tij↓ (Nij
↑
(t)) − t (11.2)

since Tij↑ and Nij

↑
are inverse functions.
↑ ↓
A little bit of care must be taken because Nij (t) and Nij (t) are not strictly
increasing functions of time unless there is always a positive inflow and outflow
rate for link (i, j). Furthermore, we often introduce a time discretization. For
both of these reasons, the inverse functions Tij↑ (n) and Tij↓ (n) may not be well-
defined. We thus need to modify equation (11.2) in a few ways:

ˆ We must ensure that τij (t) is always at least equal to the free-flow travel
time on the link. The danger is illustrated in Figure 11.3, where no vehicles
enter or leave the link for an extended period of time. The horizontal
distance between the N ↑ and N ↓ curves at their closest point is small (the
dashed line in the figure), but this does not reflect the actual travel time
of any vehicle. In reality, a vehicle entering the link when it is completely
empty, and when there is no downstream bottleneck, would experience
free-flow conditions on the link.

↓
ˆ If the link has no outflow for an interval of time, then Nij (t) will be
constant over that interval. This frequently happens with traffic signals.
↓
In this case, there are multiple values of time where Nij (t) = n. The
↓
correct way to resolve this is to define Tij (n) to be the earliest time for
11.1. TOWARDS DYNAMIC USER EQUILIBRIUM 413

N
Upstream

Downstream

Figure 11.3: Cumulative counts are flat when there is no inflow or outflow.

↓
which Nij (t) = n:
n o
Tij↓ (n) = min t : Nij
↓
(t) = n . (11.3)
t

ˆ In discrete time, the time at which the n-th vehicle departs may not line
↓
up with a multiple of ∆t, so there may be no known point where Nij (t) is
exactly equal to n. In this case, it is appropriate to interpolate between
↓
the last time point where Nij (t) < n, and the first time point where
↓
Nij (t) ≥ n.

With these modifications to how Tij↓ is calculated, the formula (11.2) can be
used to calculate the travel times on each link and arrival time.
The travel time on a path π for a traveler departing at time t, denoted C π (t),
can then be calculated sequentially. If the path is π = [r, i1 , i2 , . . . , s], then the
traveler departs origin r at time t, and arrives at node i1 at time t + τri1 (t). The
travel time on link (i1 , i2 ) is then τi1 i2 (t + τri1 (t)), so the traveler arrives at i2
at time t + τri1 (t) + τi1 i2 (t + τri1 (t)), and so forth. Writing out this formula can
be a bit cumbersome, but calculating it in practice is quite simple: it is nothing
more than accumulating the travel times of the links in the path, keeping track
of the time at which each link is entered.

11.1.3 Determining splitting proportions

The network loading requires knowledge of when and where vehicles enter the
network, and the splitting proportions p at diverges and general intersections.
In dynamic traffic assignment, this is reflected indirectly, through the variables
hπ (t), denoting the number of travelers departing path π at time t. We do not
specifically give the proportions p as a function of time, because we do not know
when any traveler will reach any node before actually performing the network
loading. Path choice is a behavioral parameter associated with travelers, so it
is easier to reconcile path choices with behavior if they are expressed this way,
rather than by simply specifying the turning fractions at each node.
414 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

N
Upstream
(total)
Path 2 sending flow
S(t)
Path 2
Path 1 sending flow Downstream
(total) t
Path 1 t1 t2 t

Figure 11.4: Disaggregating the sending flow by path.

That said, the network loading does in fact need phij (t) values for each
turning movement [h, i, j] at a diverge or general intersection, for each time
period t. These are obtained by examining the vehicles comprising the sending
flow Shi (t), and calculating the fraction of these vehicles whose path includes
link (i, j) as the next link. One way to do this is to disaggregate the cumulative
count values N ↑ (t) and N ↓ (t) calculated at the upstream and downstream ends
of each link. For each path π in the network, and for every link (h, i) and time
↑ ↓
t, define Nhi,π (t) and Nhi,π (t) to be the total number of vehicles using path π
which have respectively entered and left link (h, i) by time t. Clearly we have

↑ ↑ ↓ ↓
X X
Nhi (t) = Nhi,π (t) Nhi (t) = Nhi,π (t) ∀(h, i) ∈ A, t . (11.4)
π∈Π π∈Π

Then, the sending flow for each link and time interval can be disaggregated
in the same way, with Shi,π (t) defined as the number of vehicles in the sending
flow which are using path π. Assume that we are calculating sending flow for
link (h, i) at time t, and have determined Shi (t). At this point in time, the total
↓
number of vehicles which has left this link is Nhi (t), Therefore, the vehicles in
↓ ↓
the sending flow are numbered in the range Nhi (t) to Nhi (t) + Shi (t). Using the
inverse functions (11.2), the times at which these vehicles entered link (h, i) are
↓ ↓
in the range T ↑ (Nhi (t)) to T ↑ (Nhi (t) + Shi (t)). Denote these two times by t1
and t2 , respectively. Then the disaggregate sending flow Shi,π (t) is the number
of vehicles on path π which entered link (h, i) between t1 and t2 , that is,

↑ ↑
Shi,π (t) = Nhi,π (t2 ) − Nhi,π (t1 ) . (11.5)

This procedure is illustrated in Figure 11.4.

The proportion of travelers in the sending flow Shi wishing to turn onto link
(i, j) can now be calculated. Let Π[h,i,j] (t) be the set of paths which arrive at
11.1. TOWARDS DYNAMIC USER EQUILIBRIUM 415

node (h, i) at time t and immediately continue onto link (i, j). Then
P
π∈Π[h,i,j] (t) Shi,π (t)
phij (t) = (11.6)
Shi (t)
for use in diverge or general intersection node models. Then, after a node model
produces the actual flows yhij (t) between links, the disaggregated cumulative
counts must be P updated as well.
Let yhi (t) = (i,j)∈Γ(i) yhij (t) be the total flow which leaves link (h, i) during
↑
time t. Then all vehicles which entered the link between t1 and Thi (N ↓ (t)hi +
yhi ) can leave the link (use t3 to denote this latter time), so we update the
downstream aggregate counts to
↓ ↑
Nhi,π (t + ∆t) = Nhi,π (t3 ) (11.7)

and the upstream counts to

↑ ↓
Nij,π (t + ∆t) = Nhi,π (t + ∆t) . (11.8)

where (h, i) is the link immediately upstream of (i, j) on path π.

This formula is an approximation, since it assumes that the vehicles on
different paths in the sending flow are uniformly distributed. In reality, vehicles
towards the start of the sending flow may have a different mix of paths than
vehicles towards the end of the sending flow. This means that the p value
(calculated based on the entire sending flow) and the actual vehicles which
are moved (which are from the start of the sending flow) may not be entirely
consistent. In practice, this violation is typically small. If one wishes to obtain
an exact formula, an iterative process can be used to ensure that the proportions
p are consistent with the proportion of vehicles between t1 and t3 .

11.1.4 Principle of dynamic user equilibrium

With the procedures in the previous two subsections, we can now iteratively
perform network loading and find time-dependent shortest paths. The dynamic
traffic assignment problem is to find a mutually consistent solution between
these two models. Let H be a matrix of path flows indicating the number of
vehicles departing on each path at each time, and let T be a matrix of path
travel times, indicating the travel time for each path and departure time. Then
the network loading can be concisely expressed by

T = N (H) (11.9)

where N is a function encompassing whatever network loading is performed,

mapping the path flows to path travel times. Likewise, let the assumed behav-
ioral rule be denoted by B, which indicates the allowable path flow matrices H
when the travel times are T . For instance, one may assume that departure times
are fixed, but only minimum-travel time paths must be used; that minimum-
travel time paths must be used but the departure times must minimize total
416 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

cost (travel time plus schedule delay); or some other principle. In general, there
may be multiple H matrices which satisfy this principle, as when multiple paths
have the same, minimal travel time, in which case B actually denotes a set of
H matrices. So, we can express the behavioral consistency rule by
H ∈ B(T ) . (11.10)
We are now in a position to define the solution to the dynamic traffic assign-
ment problem as a fixed point problem: the path flow matrix H is a dynamic
user equilibrium if
H ∈ B(N (H)) . (11.11)
That is, the path flow matrix must be consistent with the driver behavior as-
sumptions, when the travel times are obtained from that same path flow matrix.
The most common behavioral rule is that departure times are fixed, but only
minimum-travel time paths will be used. In this case the principle can be stated
more intuitively as all used paths connecting the same origin to the same des-
tination at the same departure time must have equal and minimal travel time.
If departure times can be varied to minimize total cost, then the principle can
be stated as all used paths connecting the same origin to the same destination
have the same total cost, regardless of departure time.
The dynamic user equilibrium solution can also be stated as a solution to a
variational inequality. For the case of fixed departure times, the set of feasible
path flow matrices is given by
( )
T̄ ×|Π|
X
π rs 2
H̄ = H ∈ R+ : ht = dt ∀(r, s) ∈ Z , t , (11.12)
π∈Πrs

that is, matrices with nonnegative entries where the sum of all path flows con-
necting the origin r to destination s at departure time t is the corresponding
value in the time-dependent OD matrix drst . Then, the dynamic user equilibrium
solution Ĥ satisfies
N (Ĥ) · (Ĥ − H) ≤ 0 ∀H ∈ H̄ , (11.13)
where the product · is the Frobenius product, obtained by treating the matrices
as vectors and calculating their dot product.
In the case of departure time choice, we can define S = S(H) to be the
matrix of total costs for each path and departure time, where S is obtained by
composing the schedule delay function (10.4) with the network loading mapping
N . In this case, the set of feasible path flow matrices is given by
( )
T̄ ×|Π|
X X
π rs 2
Ĥ = H ∈ R+ : ht = d ∀(r, s) ∈ Z , (11.14)
π∈Πrs t

where drs is the OD matrix giving total flows between each origin and desti-
nation throughout the analysis period. The variational inequality in this case
is
S(Ĥ) · (Ĥ − H) ≤ 0 ∀H ∈ Ĥ . (11.15)
11.2. SOLVING FOR DYNAMIC EQUILIBRIUM 417

Although these fixed point and variational inequality formulations encode

the equilibrium principle, and concretely specify the problem at hand, little else
can be proven concretely except in the case of simple link and node models
(such as point queues). The vast majority of theoretical results on variational
inequalities and fixed point problems relies on regularity conditions, such as
continuity or monotonicity of the mappings N or H. For realistic link models
and node models, these mappings are not continuous. For example, queue
spillback imposes discontinuities on the network loading mapping. This means
that we usually cannot prove that dynamic user equilibrium always exists or is
unique, and indeed later in this chapter we present counterexamples where this
need not be the case. However, the variational inequality can be used to define
gap measures that can measure the degree of compliance with the equilibrium
principle. In practice, whether the increased realism of a dynamic traffic model
outweighs these disadvantages is case-dependent, and these considerations enter
into how the appropriate modeling tools are chosen.

11.2 Solving for Dynamic Equilibrium

This section describes how dynamic user equilibrium solutions can be found.
All of the algorithms in this section follow the same general framework, based
on iterating between network loading, finding time-dependent shortest paths,
and a path updating procedure, and they share common termination criteria.
Three specific path updating procedures are discussed here: the convex combi-
nations method, simplicial decomposition, and gradient projection. These are
all heuristics, without making stronger assumptions on the properties of the
network loading. This is demonstrated more fully in the next section, where we
show that dynamic traffic assignment does not always share the neat solution
properties of static assignment. These methods are described only for the case
of fixed departure times, but it is not difficult to extend any of them to the case
of departure time choice with schedule delay.

11.2.1 General framework

The general framework for almost all dynamic traffic assignment algorithms
involves iterating three steps: network loading, as discussed in Chapter 9, time-
dependent shortest paths, as discussed in Chapter 10, and updating the path
flow matrix H, as discussed here. The algorithm takes the following form:
1. Initialize the path flow matrix H to a feasible value in H̄.
2. Perform the network loading using path flows H, obtaining path travel
times T = N (H).
3. Identify the time-dependent shortest paths.
4. Check termination criteria; if sufficiently close to dynamic user equilib-
rium, stop.
418 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

5. Update the path flow matrix H and return to step 2.

A few steps in this algorithm warrant further explanation. The choice of the
initial solution is arbitrary. One reasonable choice is to identify one or more
low travel-time paths between each OD pair based on free-flow travel times,
and to divide the demand in the OD matrix between them. The effort and time
involved in identifying the initial solution should be balanced against the benefit
obtained by a more careful choice, as opposed to starting with a simpler choice
and just running additional iterations.
The termination criteria in step 4 can take several forms. The most the-
oretically sound termination criterion is a gap measure, comparing the path
flows H and travel times T with the assumed behavioral rule for travelers. Such
measures are preferred to simpler criteria, such as stopping when the path flows
do not change much from iteration to iteration — if the latter occurs, there is
no way to tell whether this is because we are near an equilibrium solution, or
because the algorithm is unable to make any more progress and has gotten stuck
near a non-equilibrium solution. For the case of fixed demand, the behavioral
rule is that travelers must take a least travel-time path between their origin
and destination, for their departure time. After performing the time-dependent
shortest path step, we can calculate the travel time on such a path for each OD
∗
pair (r, s) and departure time t; call this value τrs,t . If all travelers in the net-
work were on such paths, then the total travel time experienced by all travelers
in the network would be given by
X X
∗
SP T T = drs
t τrs,t . (11.16)
(r,s)∈Z 2 t

This is often known as the shortest path travel time. This is contrasted with
the total system travel time, which is the actual total travel time spent by all
travelers on the network:
XX
T ST T = htπ τπt = H · T . (11.17)
π∈Π t

Clearly T ST T ≥ SP T T , and T ST T = SP T T only if all travelers are on

least-time paths, which corresponds to dynamic user equilibria. Therefore, the
gap between these values is a quantitative measure of how close the path flows
H are to satisfying the equilibrium principle. Two common gap values involve
normalizing the difference T ST T − SP T T in two ways. The relative gap γ
normalizes this difference by SPTT,
T ST T − SP T T
γ= , (11.18)
SP T T
while the average excess cost normalizes this difference by the total number of
travelers on the network,
T ST T − SP T T
AEC = P P rs . (11.19)
(r,s)∈Z 2 t dt
11.2. SOLVING FOR DYNAMIC EQUILIBRIUM 419

The algorithm can be terminated whenever either of these measures is suffi-

ciently small. An advantage of the average excess cost is that it is measured
in time units, and has the intuitive interpretation as being the average differ-
ence between a traveler’s actual travel time, and the shortest path travel time
available to them. Similar expressions can be derived for departure time choice,
with schedule delay.
Finally, the last step (updating the path flow matrix) requires the most
care, and this section presents three alternatives. The convex combinations
method is the simplest and most economical in terms of computer memory. The
simplicial decomposition method converges faster, by storing previous solutions
and using them to find a better improvement direction for H. The third method
is gradient projection, which involves computation of an approximate derivative
of path travel times with respect to path flow. This derivative enables the use
of Newton’s method to identify the amount of flow to shift between paths.

11.2.2 Convex combinations method

The convex combinations method is the simplest way to update the path choice
matrix. In this method, after computing time-dependent shortest paths, one
identifies a “target matrix” H ∗ which would give the path flows travelers would
choose if the path travel times were held constant at their current values T .
For the case of fixed departure times, for each OD pair (r, s) and departure
time t, all of the demand drs t is loaded onto its time-dependent shortest path
in H ∗ , and all other path flows are set to zero. For the case of departure time
choice, for each OD pair (r, s) the departure time t∗ and corresponding path
with minimal schedule delay is identified, and all drs travelers associated with
this OD pair are assigned to that path and departure time, with all other path
flows set to zero. This is often called an all-or-nothing assignment, because it
involves choosing a single alternative to assign an entire class of travelers.
The basic form of the convex combinations method updates H using the
formula
H ← λH ∗ + (1 − λ)H , (11.20)
where λ ∈ [0, 1] is a step size showing how far to move in the direction of the
target. If λ = 0, then the path flow matrix is unchanged, whereas if λ = 1,
the path flow matrix is set equal to the target. Intermediate values produce a
solution which is a weighted average of the current solution H and the target H ∗ .
There are a number of ways to choose the step sizes λ. The simplest alternative
is to use a decreasing sequence of values, choosing for the i-th iteration the step
size λi = 1/(i + 1), so the step sizes form the
P sequence 1/2, P 1/3, 1/4, and so on.
It is common to use sequences such that λi = +∞ and λ2i < ∞, because
we do not want the sequence of H values to converge too quickly (or else it
may stop short of the equilibrium), but if the step size drops too slowly, the
algorithm is likely to “overshoot” the equilibrium and oscillate.
More sophisticated variations are possible as well. Rather than using a fixed
step size, λ can be chosen to minimize, say, the relative gap or average excess
420 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

H5*

H7*
H6*

H4*

H
H1*

H2*

H3*

Figure 11.5: Simplicial decomposition can move towards non-corner points.

cost. While this likely decreases the number of iterations required to reach a
small gap, the computation at each iteration is increased. In particular, each
λ value tested requires an entire network loading step to determine the corre-
sponding gap, which can be a significant time investment. It is also possible to
vary the λ value for different path flows and departure times. Some popular
heuristics are to use larger λ values for OD pairs which are further from equi-
librium, or to vary λ for different departure times — since travel times for later
departure times depend on the path choices for earlier departure times, it may
not make sense to invest much effort in equilibrating later departure times until
earlier ones have stabilized.

11.2.3 Simplicial decomposition

The simplicial decomposition algorithm extends the convex combinations method
by remembering the target matrices H ∗ from earlier iterations. This requires
more computer memory, but storing these previous target matrices leads to a
more efficient update of the path flow vector. In particular, choosing an all-or-
nothing assignment as a target matrix essentially restricts the choice of search
direction to the corner points of the feasible region. As a result, when the
equilibrium solution is reached, the convex combinations method will “zig-zag,”
taking a number of short, oblique steps rather than a more direct step towards
the equilibrium problem. (Figure 11.5). Simplicial decomposition is able to do
so by combining multiple target points into the search direction. In this algo-
rithm, the set H is used to store all of the target path flow matrices found thus
far.
A second notion in simplicial decomposition is that of a “restricted equilib-
rium.” Given a set H = {H1∗ , H2∗ , · · · , Hk∗ } , we say that a time-dependent path
flow matrix Ĥ is a restricted equilibrium relative to H if it solves the variational
11.2. SOLVING FOR DYNAMIC EQUILIBRIUM 421

inequality
N (Ĥ) · (Ĥ − Hi∗ ) ≤ 0 ∀Hi∗ ∈ H . (11.21)
This is different from the variational inequality (11.13) because the only possible
choices for H are the matrices in the set H, rather than any of the feasible H
matrices satisfying (11.12).
Effectively, H is a restricted equilibrium if none of the targets in H lead
to improving directions in the sense that the total system travel time would be
reduced by moving to some Hi∗ ∈ H while fixing the travel times at their current
values.
At a high level, simplicial decomposition works by iterating between adding
a new target matrix to H, and then finding a restricted equilibrium using the
current matrices in H. In practice, it is too expensive to exactly find a restricted
equilibrium at each iteration. Instead, several “inner iteration” steps are taken
to move towards a restricted equilibrium with the current set H before looking
to add another target. In each inner iteration, the current solution H is adjusted
to H + µ∆H, where µ is a step size and ∆H is a direction which moves toward
restricted equilibrium. One good choice for this direction is
+
[N (H) · (H − Hi∗ )] (Hi∗ − H)
P
Hi∗ ∈H
∆H = + (11.22)
[N (H) · (H − Hi∗ )]
P
Hi∗ ∈H

This rather intimidating-looking formula is actually quite simple. It is nothing

more than a weighted average of the directions Hi∗ − H (potential moves toward
each target in H), where the weight for each potential direction is the extent to
which it improves upon the current solution: [N (H) · (H − Hi∗ )] is the reduction
in total system travel time obtained by moving from H to Hi∗ while holding
travel times constant. If this term is negative, there is no need to move in that
direction, so the weight is simply set to zero. The denominator is simply the
sum of the weights, which serves as a normalizing factor.
The step size µ is chosen through trial-and-error. One potential strategy is
to iteratively test µ values in some sequence (say, 1, 1/2, 1/4, . . .) until we have
found a solution acceptably closer to restricted equilibrium than H.1 “Accept-
ably closer” can be calculated using the restricted average excess cost

N (H) · H − minHi∗ ∈H {N (H) · Hi∗ }

AEC 0 = P P rs (11.23)
(r,s)∈Z 2 t dt

which is similar to the average excess cost, but instead of using the shortest
path travel time uses the best available target vector in H.
Unlike the version of simplicial decomposition used for the static traffic as-
signment problem (Chapter 6), there is no guarantee that a sufficiently small
choice of µ will result in a reduction of restricted average excess cost. However,
in practice, this rule seems to work acceptably.
1 Those familiar with nonlinear optimization may see parallels between this and the Armijo

rule.
422 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

Putting all of this together, for simplicial decomposition, step 5 of the dy-
namic traffic assignment algorithm in Section 11.2.1 involves performing all of
these steps as a “subproblem”:

5. Subproblem: Find an approximate restricted equilibrium H using only

the vectors in H.

(a) Find the improvement direction ∆H using equation (11.22).

(b) Update H ← H + µ∆H, with µ sufficiently small (to reduce AEC 0 ).
(c) Perform network loading to update travel times.
(d) Return to step (a) of subproblem unless AEC 0 is small enough.

Furthermore, the set H must be managed. It is initialized to be empty, and

whenever time-dependent shortest paths are found, a new all-or-nothing assign-
ment (the one which would have been chosen as the sole target in the convex
combinations method) is added to H.

11.2.4 Gradient projection

The gradient project method updates the path flow matrix H by using Newton’s
method. Consider first the simpler problem of trying to shift flows between two
paths π1 and π2 (for the same departure time) to equalize their travel times.
Write τ1 (h1 , h2 ) and τ2 (h1 , h2 ) to denote the travel times on the two paths as a
function of the flows h1 and h2 on the two paths. If we were to shift ∆h vehicles
from path 1 to path 2, the difference in travel times between the paths is given
by
g(∆h) = τ1 (h1 − ∆h, h2 + ∆h) − τ2 (h1 − ∆h, h2 + ∆h) . (11.24)
We want to choose ∆h so that the difference g(∆h) between the path travel
times is equal to zero.
Using one step of Newton’s method, an approximate value of ∆h is given by

g(0)
∆h = − . (11.25)
g 0 (0)

The numerator g(0) is simply the difference in travel times between the two
paths before any flow is shifted. To calculate the denominator, we need to know
how the difference in travel times will change as flow is shifted from π1 to π2 .
Computing the derivative of (11.24) with the help of the chain rule, we have
   
0 ∂t1 ∂t2 ∂t1 ∂t2
g (0) = + − + . (11.26)
∂h2 ∂h1 ∂h1 ∂h2

This formula is not easy to evaluate exactly, but we can make a few approxima-
tions. The first term in (11.26) reflects the impact the flow on one path has on
the other path’s travel time, while the second term reflects the impact the flow
on one path has on its own travel time. Typically, we would expect the second
11.2. SOLVING FOR DYNAMIC EQUILIBRIUM 423

effect will be larger in magnitude than the first effect (although exceptions do
exist). So, as a first step we can approximate the derivative as
 
0 ∂t1 ∂t2
g (0) ≈ − + . (11.27)
∂h1 ∂h2
The next task is to calculate the derivative of a path travel time with re-
spect to flow along this path. Unlike in static traffic assignment, there is no
closed-form expression mapping flows to travel times, but rather a network
loading procedure must be used. For networks involving triangular fundamen-
tal diagrams (which can include the point queue model, cf. Section 9.5.3), the
derivative of the travel time on a single link (i, j) with respect to flow entering
at time t can be divided into two cases. In the first case, suppose that the
vehicle entering link (i, j) at time t exits at t0 = t + τij (t), and that the link is
demand-constrained at that time (that is, all of the sending flow Sij (t0 ) is able
to move). In this case, even if a marginal unit of flow is added at this time,
the link will remain demand-constrained, and all of the flow will still be able
to move. No additional delay would accrue, so the derivative of the link travel
time is
dτij (t)
= 0. (11.28)
dh
In the second case, suppose that at time t0 flows leaving link (i, j) are con-
strained either by the capacity of the link or by the receiving flow of a down-
stream link. In this case, not all of the flow is able to depart the link, and a
queue has formed at the downstream end of (i, j). The clearance rate for this
queue is given by yij (t0 ), so the incremental delay added by one more vehicle
joining the queue is 1/yij (t0 ), and

τij (t) 1
= . (11.29)
dh yij (t0 )

(The denominator of this expression cannot be zero, since t0 is the time at which
a vehicle is leaving the link.)
Therefore, we can calculate the derivative of an entire path’s travel time
inductively. Assume that π = [r, i1 , i2 , . . . , s], and that ti gives the travel time
for arriving at each node i in the path. Then dtπ /dhπ is obtained by summing
expressions (11.28) and (11.29) for the uncongested and congested links in this
path, taking care to use the correct time indices:
dtπ X 1
≈ [yij (tj ) < Sij (tj )] , (11.30)
dhπ yij (tj )
(i,j)∈π

where the brackets are an indicator function equal to one if the statement in
brackets is true, and zero if false.
We are almost ready to give the formula for (11.27), but we can make one
more improvement to the approximation. Two paths may share a certain num-
ber of links in common. Define the divergence node of two paths to be node
424 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

h k

j
l

i
m

Figure 11.6: Node j is the divergence node of these two paths.

where the paths diverge for the first time. (Figure 11.6). Prior to the divergence
node, the two paths include the same links, so there will be no effect of shifting
flow from one path to the other. Therefore, the sum in (11.30) need only be
taken for links beyond the divergence node. Even if the paths rejoin at a later
point, they may do so at different times, so we cannot say that there is no effect
of shifting flow between common links downstream of a divergence node. So, if
d(π1 , π2 ) is the divergence node, then we have

dtπ X 1
≈ [yij (tj ) < Sij (tj )] , (11.31)
dhπ yij (tj )
(i,j)∈π:(i,j)>d(π1 ,π2 )

where the “>” notation for links indicates links downstream of a node in a path.
This expression can then be used in (11.27) and (11.25) to find the approximate
amount of flow which needs to be shifted from π1 to π2 to equalize their costs.
The procedure for updating H can now be described as follows:

5. Subproblem: Perform Newton updates between all paths and the target
path found.

(a) For each path π with positive flow, let ρ be the least-travel time path
connecting the same OD pair and departure time.
(b) Calculate ∆h using equations (11.31), (11.27), and (11.25), using π
as the first path and ρ as the second.
(c) If ∆h < hπ , then update hρ ← hρ +∆h and hπ ← hπ −∆h. Otherwise,
set hρ ← hρ + hπ and hπ ← 0.

As stated, the path travel times and derivatives are not updated after each shift
is performed. Doing so would increase accuracy, but greatly increase computa-
tion time since a complete network loading would have to be performed.
11.3. PROPERTIES OF DYNAMIC EQUILIBRIA 425

11.3 Properties of Dynamic Equilibria

The dynamic traffic assignment problem is much less well-behaved than the
static traffic assignment problem, where a unique equilibrium provably exists
under natural conditions. This is because the more realistic network loading
procedures used in dynamic traffic assignment lack the regularity properties
which make static assignment more convenient mathematically. Recall that the
dynamic user equilibrium path flows solve the variational inequality

N (Ĥ)(˙ Ĥ − H) ≤ 0 ∀H ∈ H̄ . (11.32)

Showing existence of a solution to this variational inequality would typically rely

on showing that N is continuous. However, when queue spillback is modeled, the
travel times are discontinuous in the path flow variables. Showing uniqueness
of a solution to the variational inequality typically involves showing that N
has some flavor of monotonicity, but common node models can be shown to
violate monotonicity even when we restrict attention to networks involving only
a single diverge and merge. Thus, the failure of existence or uniqueness results
is directly traced to the features that make dynamic traffic assignment a more
realistic model. As an analyst, you should be aware of this tradeoff.
This section catalogs a few examples of dynamic traffic assignment problems
highlighting some unusual or counterintuitive results. We begin with an example
where the choice of node model results in no dynamic user equilibrium solution
existing. Two examples of equilibrium nonuniqueness are given, first where
multiple user equilibrium solutions exist, and second where literally all feasible
path flow matrices are user equilibrium solutions, even though they vary widely
in total system travel time. We conclude with a dynamic equivalent to the Braess
paradox which was developed by Daganzo, in which increasing the capacity on
a link can make network-wide conditions arbitrarily worse, because of queue
spillback.

11.3.1 Existence: competition at merges

The network in Figure 11.7 has two origin-destination pairs (A to B, and C to D).
Aside from these four centroids, there are two nodes representing intersections.
Turns are not allowed at these intersections, so Ξ(1) = {[A, 1, 2], [C, 1, D]} and
Ξ(2) = {[1, 2, B], [C, 2, D]}. As a result, each OD pair has only two routes
available to it: [A, B] and [A, 1, 2, B] for A to B, and [C, 1, D] and [C, 2, D] for
C to D. For the sake of convenience, call these four routes “top,” “bottom,”
“left,” and “right,” respectively. The figure shows the travel times with each
link. Two links in the network also carry a toll, expressed in the same units as
travel time, and drivers choose routes to minimize the sum of travel time and
toll. Capacities and jam densities are large enough that congestion will never
arise, and all links will remain at free flow.
Nodes 1 and 2 have a distinctive node model, representing absolute priority
of one approach over the other. For Node 1, all flow on the movement [A, 1, 2]
426 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

3 (Toll 0.5)

A B
2 D 1
1 1
1
1 2

1 2
(Toll 0.5) C

Figure 11.7: Network for example with no equilibrium solution. Link travel
times shown.

must yield to flow on [C, 1, D]. This can be expressed with the following node
model:
yC1D = min {SC1 , R1D } (11.33)
(
0 if yC1D > 0
yA12 = (11.34)
min {SA1 , R12 } otherwise

Likewise, for Node 2, all flow on the movement [C, 2, D] must yield to flow on
[1, 2, B]. That is, traffic from C to D has priority at node 1; and traffic from
A to B has priority at node 2. These relationships are indicated on the figure
with triangles next to the approach that must yield.
Each OD pair has one unit of demand, that departs during the same time
interval. We now show that there is no assignment of demand to paths that
satisfies the equilibrium principle, requiring any used path to have minimal
travel time. We begin first by showing that any equilibrium solution cannot
split demand among multiple paths; each OD pair must place all of its flow on
just one path or the other. With the given demand, the cost on the bottom path
is either 3 (if there is no flow on the left path from C to D) or 4 (if there is), in
either case, the cost is different from that on the top path (3.5). All demand will
thus use either the top path or the bottom path, whichever is lower. Similarly,
for the OD pair from C to D, the cost on the right path is either 3 or 4, whereas
that on the left is either 3.5 or 4.5. Both paths cannot have equal travel time,
so at most one path can be used.
Thus, there are only four solutions that can possibly be equilibria: all flow
from A to B must either be on the top path or the bottom path; and all flow
from C to D must either be on the left path or the right path. First consider
the case when all flow from A to B is on the top path. The left path would then
have a cost of 3.5, and the right path a cost of 3. So all flow from A to B must
be on the right path. But then the travel times on the top and bottom paths are
11.3. PROPERTIES OF DYNAMIC EQUILIBRIA 427

3.5 and 3, respectively, so this solution cannot be an equilibrium (the travelers

from A to B would prefer the bottom path, contradicting our assumption).
So now consider the other possibility, when all flow from A to B is on the
bottom path. The travel times on the left and right paths are now 3.5 and 4,
so all travelers from C to D would pick the left path. But then the travel times
on the top and bottom paths are 3.5 and 4, respectively, so this solution is not
an equilibrium either.
Therefore, there is no assignment of vehicles to paths that satisfies the prin-
ciple of user equilibrium. This network is essentially the Ginger/Harold “match-
ing pennies” game from Section 1.3 encoded into dynamic traffic assignment.
Game theorists often resolve this type of situation by proposing a mixed-strategy
equilibrium, in which the players randomize their actions. One can show that
a mixed-strategy equilibrium exists in this network (Exercise 8), but this equi-
librium concept is usually not applied to transportation planning. This is for
both computational reasons (when there are many players, as in practical plan-
ning networks, computing mixed-strategy equilibria is hard) and for modeling
reasons (typically travelers do not randomize their paths with the intent of
“outsmarting” other travelers).
Mathematically, the reason no equilibrium exists is because the node model
specified by (11.33) and (11.34) is not continuous in the sending and receiving
flows.
This is similar to static traffic assignment, where equilibrium existence can-
not be guaranteed if the link performance functions are not continuous. In
dynamic traffic assignment, when building node models that capture absolute
priority (as at yields, two-way stops, or signals with permitted turns), ensuring
continuity is difficult. In part, this is the motivation for using ratios of α values
in Chapter 9 to reflect priorities at merges and general intersections.
But even if the link and node models are chosen to be continuous, dynamic
equilibrium need not exist. Repeating the arguments made in Chapter 5.2 would
also require the travel time calculations to be continuous in the link cumulative
entrances and exits N ↑ and N ↓ . The use of a minimum to disambiguate multiple
possible values of travel times in Equation (11.3) can mean that small changes in
a link’s cumulative entries or exits could cause a large change in the travel time
(imagine how the length of the dashed line in Figure 11.3 would change if one of
the corner points were shifted slightly). Continuity could be assured only if the
cumulative counts were strictly increasing, that is, if vehicles were constantly
flowing into and out of each link. Therefore, although one can construct dynamic
traffic assignment models which guarantee existence of equilibria, these require
quite strong and significant restrictions on the network loading and travel time
calculation procedures.

11.3.2 Uniqueness: Nie’s merge

The network in Figure 11.8 consists of a single diverge and merge. One can
visualize this network as a stylized version of a freeway lane drop (Figure 11.9),
where the reduction from two lanes to one lane reduces the capacity from 80
428 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

Link 2
Link 1 40 Link 4
80 Link 3 40
40

Figure 11.8: Nie’s merge network. All capacity values in vehicles per minute.

Figure 11.9: A physical interpretation of Nie’s merge.

vehicles per minute to 40 vehicles per minute, and where drivers either merge
early (choosing the top lane at the diverge point) or late (waiting until the lane
drop itself). The inflow rate from upstream is 80 vehicles per minute. Since
the capacity of the downstream exit is only 40 vehicles per minute, the excess
vehicles will form one or more queues in this network.
To simplify the analysis, assume that the time horizon is short enough that
none of these queues will grow to encompass the entire length of the link. With
this assumption, queue spillback can be ignored, and we can focus on the issue
of route choice. Furthermore, under these assumptions, S1 is always 80 vehicles
per minute, and R2 , R3 , and R4 are always 40 vehicles per minute. We work
with a timestep of one minute, and will express all flow quantities in vehicles
per minute.2
In this network, there is only one choice of routes (the top or bottom link
at the diverge). We will restrict our attention to path flow matrices H where
the proportions of vehicles choosing the top and bottom routes are the same
for all departure times, and show that there are multiple equilibrium solutions
even when limiting our attention to such H matrices. There may be still more
equilibria where the proportion of vehicles choosing each path varies over time.
Therefore, the ratio h124 /h134 is constant for all time intervals, which means
that the splitting fractions p12 and p13 at the diverge point are also constant,
and equal to h124 /40 and h134 /40, respectively.
Using the diverge and merge models from Section 9.2.3, we can analyze the
queue lengths and travel times on each link as a function of these splitting
2 This is larger than what is typically used in practice, but simplifies the calculations and

does not affect the conclusions of this example.

11.3. PROPERTIES OF DYNAMIC EQUILIBRIA 429

fractions.
It turns out that three distinct dynamic user equilibria exist:

Equilibrium I: If p12 = 1 and p13 = 0, then the diverge formula (9.17) gives
the proportion of flow which can move as
 
40 40 1
φ = min 1, , = . (11.35)
80 0 2

Therefore, the transition flows at each time step are y12 = 40 and y13 = 0.
Since y12 < S12 , a queue will form on the upstream link, and its sending
flow will remain at S1 = 80. Therefore, once the first vehicles reach the
merge, we will have S2 = 40 and S3 = 0. Applying these proportions,
together with R4 = 40, the merge formula gives y24 = 40 and y34 = 0.
There will be no queue on link (2,4), so both link (2,4) and link (3,4) are
at free-flow. This solution is an equilibrium: the two paths through the
network only differ by the choice of link 2 or link 3, and both of these have
the same travel time since they are both at free-flow. In physical terms,
this corresponds to all drivers choosing to merge early; the queue forms
upstream of the point where everyone chooses to merge, and there is no
congestion downstream.

Equilibrium II: If p12 = 0 and p13 = 1, the solution is exactly symmetric to

that of Equilibrium I. A queue will again form on the upstream link, now
because all drivers are waiting to take the bottom link. In physical terms,
this corresponds to all drivers choosing to merge early, but to the lane
which is about to drop. They then all merge back when the lane actually
ends. This solution does not seem especially plausible, but it does satisfy
the equilibrium condition.

Equilibrium III: If p12 = p13 = 1/2, drivers wish to split equally between
the top and bottom links. The proportion of flow which can move at the
diverge is  
40 40
φ = min 1, , = 1, (11.36)
40 40
so all vehicles can move and there is no queue at the diverge: y12 = y13 =
40. Once these vehicles reach the merge, we will have S2 = S3 = 40 and
R4 = 40. The merge formula (9.12) then gives y24 = y34 = 20, so queues
will form on both merge links. However, since the inflow and outflow rates
of links 2 and 3 are identical, the queues will have identical lengths, and
so the travel times on these links will again be identical. Therefore, this
solution satisfies the principle of dynamic user equilibrium. In physical
terms, this is the case when no drivers change lanes until the lane actually
ends, and queueing occurs at the merge point.

By examining the intermediate solutions, we can calculate the travel times on

both paths for different values of the fraction p12 . (We do not need to state the
430 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

Path travel time

Path [1,3,4]

III
I II

Path [1,2,4]

1/2 1
p12

Figure 11.10: Travel time on top and bottom paths as a function of splitting
fraction, with three equilibrium solutions marked.

values of the other fraction, since p13 = 1 − p12 .) This is shown in Figure 11.10,
and the three equilibria correspond to the crossing points of the paths. Note
that all three of the equilibria share the same equilibrium travel time. Although
the queues in Equilibrium III are only half as long as those in Equilibria I and
II, being split between two links, the outflow rates of these queues are also only
half as great (20 veh/min instead of 40 veh/min).
This example shows that the dynamic user equilibrium solution is not unique,
even in an extremely simple network. This nonuniqueness has practical conse-
quences. The effect of a potential improvement to link 2 or 3 will depend
crucially on the number of travelers on the link, which varies widely in all
three equilibria. One criterion for distinguishing which of these equilibria is
more likely is stability, which explores what would happen if the equilibria were
slightly perturbed. As shown in Figure 11.10, if one begins at Equilibrium I and
perturbs p13 to a small value (reducing p12 by the same amount), we see that
the travel time on the top path increases, where as the travel time on the bot-
tom path decreases. It may seem odd that increasing flow on the bottom path
decreases its travel time — what is happening is that congestion at the diverge
decreases (lowering the travel time of both paths), but congestion forms on the
top link at the merge (increasing the travel time just of the top path). Super-
imposing these effects produces the result in Figure 11.10. As a result, travelers
will switch from the (slower) top path to the (faster) bottom one, moving us
even further away from Equilibrium I. Therefore, Equilibrium I is not stable.
The same analysis holds for Equilibrium II.
11.3. PROPERTIES OF DYNAMIC EQUILIBRIA 431

Equilibrium III, on the other hand, is stable. If a few travelers switch from
the top to the bottom path, the travel time on the top path decreases and that on
the bottom path increases. Therefore, travelers will tend to switch back to the
top path, restoring the equilibrium. The same holds if travelers switch from the
bottom path to the top path. This gives us reason to believe that Equilibrium
III is more likely to occur in practice than Equilibrium I or II. This type of
analysis is much more complicated in larger networks, and for the most part is
completely unexplored. Coming to a better understanding of the implications
of nonuniqueness in large networks, as well as techniques for addressing this, is
an important research question.

11.3.3 Nie’s merge redux

Modifying the merge network from the previous section, we can produce an even
more pathological result. We now eliminate the lane drop entirely, but preserve
the diverge/merge network, by setting the capacity of the downstream link to
80 vehicles per minute, the same as the upstream link. This might represent
a roadway section where lane changing is prohibited, perhaps in a work zone
or with a high-occupancy/toll lane. Now, at the merge point S2 + S3 ≤ R4 no
matter what the flow pattern is on the network. This means that the merge will
always be freely-flowing, and there will be no queues at the merge. Links 2, 3,
and 4 will be at free-flow regardless of the path flows H.
However, queues can still occur at the diverge, where the analysis is the same
as before. If p12 = p13 = 1/2, no queues will form at the diverge point, whereas
for any other values of p12 and p13 formula (11.35) will predict queues on the
upstream link, reaching their maximum length if p12 = 1 or p13 = 1. However,
all of these solutions satisfy the principle of equilibrium because the only delay
occurs on the upstream link, which is common to both paths. No matter what
the values of p are, there is zero relative gap or average excess cost, since no
traveler can reduce his or her travel time by choosing a different route. Choosing
a different route could influence the travel time of those further upstream, but
under the assumption that drivers act only to minimize their own travel time,
this influence on other travelers is of no concern.
Figure 11.11 shows the queue lengths and total system travel time as p
varies. The system-optimal solution is unique: if p12 = p13 = 1/2 there are no
queues in the system and all vehicles experience free-flow travel times. For any
other values of p12 and p13 , a queue will form at the diverge and some delay
is experienced. Yet all possible p values satisfy the equilibrium principle, since
the delay is upstream of the diverge and drivers cannot choose another route to
minimize their own travel time. Furthermore, as the time horizon grows longer,
the difference in total travel time between the worst of the equilibria (either
p12 = 0 or p12 = 1) and the system optimum solution can grow arbitrarily
large. This is in contrast with the bounded “price of anarchy” which can often
be found for static traffic assignment (cf. Section 5.4.3).
This effect less of a practical concern than the original version of Nie’s merge
— there are good behavioral reasons to doubt that a significant imbalance of
432 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

Total system travel time

Link 1
queue length

1/2 1
p12

Figure 11.11: Queue length and total travel time for different splitting propor-
tions p in the modified Nie’s merge network. Note that all p values represent
equilibria.

travelers will choose one alternative over another identical one. One argument
is from entropy principles (Section 5.2.2): if travelers have the same behavior
assumption, it is unlikely they would all choose one route over another with equal
travel time. Furthermore, the assumption of a triangular fundamental diagram
implies that travel speeds remain at free-flow for all subcritical densities. In
practice the speed will drop slightly due to variations in preferred speeds and
difficulties in overtaking at higher density, so drivers would likely prefer the
route chosen by fewer travelers.
However, from the standpoint of modeling the fact that all feasible solutions
are equilibria poses significant challenges.
Literally any solution will have zero gap, and if an all-or-nothing assignment
is chosen as the initial solution (as is sometimes done in implementations),
dynamic traffic assignment software will report that a perfect equilibrium has
been reached.
This example shows that initial solutions for dynamic traffic assignment
should be carefully chosen, perhaps by spreading vehicles over multiple paths,
or breaking ties stochastically in shortest path algorithms to avoid assigning all
vehicles to the same path in the first all-or-nothing assignment.

11.3.4 Efficiency: Daganzo’s paradox

The previous section showed that the worst user equilibrium solution can be
arbitrarily worse than the system optimal solution, in terms of total travel
time. However, there was still one user equilibrium solution that was also system
optimal (the case where p12 = p13 = 1/2). This need not be in the case. Here we
present an example where the only user equilibrium solution can be arbitrarily
worse than system optimum.
Furthermore, it is “paradoxical” in the sense that increasing the capacity of
the only congested link on the network can make the problem worse, and that
11.3. PROPERTIES OF DYNAMIC EQUILIBRIA 433

Link 2
Link 1 5 Link 4
20
L/uf = 4 60
1
qmax = 20 Link 3A Link 3B 20
1 1 ∞
kjL = ∞ 20 10
50 25

Figure 11.12: Network for Daganzo’s paradox.

reducing the capacity on this link can improve system conditions! In this sense,
it is a dynamic equivalent of the Braess paradox from Section 4.3.
Many have criticized the Braess paradox on the grounds that the link per-
formance functions used in static assignment are unrealistic. In the example
shown below, queue spillback (a feature unique to dynamic network loading) is
actually the critical factor in the paradox.
See the network in Figure 11.12, where time is measured in units of time
steps ∆t. Like the networks in the two previous sections, it consists of a single
merge and diverge. However, the free-flow travel times and capacities on the
top and bottom links are now different: the top route is longer, but has a higher
capacity, while the bottom route is shorter at free-flow, but has a bottleneck
limiting the throughput on this route — link 3B has only half the capacity of
3A. We will use the spatial queue model of Section 9.1.3 to propagate traffic
flow, although the same results would be obtained with an LWR-based model
or anything else which captures queue spillback and delay. The input demand
is constant, at 20 vehicles per time step.
The capacity on the top route is high enough to accommodate all of the
demand; if all of this demand were to be assigned to this route, the travel time
would be 10 minutes per vehicle. Assigning all vehicles to the top route is neither
the user equilibrium nor the system optimum solution, but it does give an upper
bound on the average vehicle travel time in the system optimal assignment —
it is possible to do better than this if we assign some vehicles to the bottom,
shorter route.
To derive the user equilibrium solution, notice that initially all vehicles will
choose the bottom path, since it has the lower travel time at free flow. A queue
will start forming on link 3A, since the output capacity is only 10 vehicles per
time step (because of the series node model from Section 9.2.1, and the capacity
of link 3B) and vehicles are entering at double this rate. As the queue grows,
the travel time on link 3A will increase as well. With the spatial queue model
and these inflow and outflow rates, you can show that the travel time for the
n-th vehicle entering the link is
 
n qin n
1+ −1 =1+ (11.37)
qin qout 20
434 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

as long as the queue has not spilled back (see Exercise 9.3.
Based on equation (11.37), when the 60th vehicle enters the link, it will
spend four time units on link 3A, and thus its travel time across the entire
bottom path would be the same as if it had chosen the top path. At this point
there are 40 vehicles on link 3A, as can be seen by drawing the N ↑ and N ↓
curves for this link. From this point on, vehicles will split between the top and
bottom paths to maintain equal travel times.
So far, so good; the first 60 vehicles that enter the network have a travel
time of less than 10 minutes, and all the rest have a travel time of exactly 10
minutes. Now see what happens if we increase the capacity on link 3B, with
the intent of alleviating congestion by improving the bottleneck capacity. If
the capacity on 3B increases from 10 to 12, the story stays the same, except
that it is the first 90 vehicles that have a travel time of less than 10 minutes.
We can see this by setting (11.37) equal to four (the time needed to equalize
travel times on the top and bottom paths), but with qin = 12 instead of 10, and
solving for n. Network conditions indeed have improved. But if the capacity
increases still further, to 15, then equation (11.37) tells us that it is only after
180 vehicles have entered the system that travelers would start splitting between
the top and bottom links. By tracing the N ↑ and N ↓ curves, at this point in
time 120 vehicles will have exited link 3A, meaning the queue length would be
60 vehicles. But the jam density of the bottom link only allows it to
hold 50 vehicles. The queue will thus spill upstream of the diverge node,
and in this scenario, no vehicles will opt to take the top path. By the time
a driver reaches the diverge point, the number of vehicles on the bottom link
is 50, giving a travel time on 3A of 3 13 minutes, and a total travel time of 9 13
minutes from origin to destination. This is less than the travel time on the top
path, so drivers prefer waiting in the queue to taking the bypass route.
As a result, all drivers will choose the bottom path, and the queue on the
origin centroid connector will grow without bound, as will the travel times
experienced by vehicles entering the network later and later. By increasing the
length of time vehicles enter the network at this rate, we can make the delays
as large as we like.
We thus see that even with dynamic network loading, increasing the capacity
on the only bottleneck link can make average travel times worse — and in fact
arbitrarily worse than the system optimum solution, where the average travel
time is less than 10 minutes. The reason for this phenomenon is the interaction
between queue spillback and selfish route choice: in the latter scenario it would
be better for the system for some drivers to choose the top route, even though
it would worsen their individual travel time.
Of course, if the capacity on 3B were increased even further, all the way
to 20 vehicles per time step, delays would drop again since there would be no
bottleneck at all. Exercise 11 asks you to plot the average vehicle delay in this
network as the capacity on link 3B varies from 0 to 25 vehicles per time step.
11.4. HISTORICAL NOTES AND FURTHER READING 435

11.3.5 Conclusion
The purpose of these examples is to show that dynamic user equilibrium is
complex. Guaranteeing existence or uniqueness of dynamic equilibrium requires
making strong assumptions on traffic flow propagation. However, for some prac-
tical applications, using a more realistic traffic flow model is more important
than mathematical properties of the resulting model. Many people find comfort
in the fact that we can at best solve for an equilibrium approximately, and thus
dismiss the question of whether an equilibrium “truly” exists as akin to asking
how many angels can dance on the head of a pin.
We emphasize that existence and uniqueness are not simply mathematical
abstractions, and that they have significant implications for practice: if an equi-
librium does not exist, should we really be ranking projects based on equilibrium
analysis? If multiple equilibria exist, what should we plan for? Can we even find
them all? At the same time, we acknowledge that using static equilibrium to
sidestep these difficulties is often unacceptable. For many applications, the as-
sumptions in link performance functions are simply too unrealistic. Such is the
nature of mathematical modeling in engineering practice. No tool is right for ev-
ery task, but rather experienced practitioners know how to match the available
tools to the job at hand. By reading this book, we hope that you have gained
the insight to understand the advantages and disadvantages of both static and
dynamic traffic assignment models, and to make educated decisions about the
right tool for a particular project.
Finally, more research is needed on these topics, understanding the signifi-
cance of equilibrium nonexistence and nonuniqueness, particularly in practical
networks and not just in “toy” networks such as the ones in this section. As
researchers, we would be delighted for you to contribute to work in this field.

11.4 Historical Notes and Further Reading

(These sections are incomplete in this beta version of the text, and will be sub-
stantially expanded in the complete first edition.)
Formulating dynamic traffic assignment models is more difficult than do-
ing so for static traffic assignment. For this reason, this chapter has eschewed
highly detailed mathematical formulations of the equilibrium principle in fa-
vor of simply expressing the dynamic user equilibrium principle and proposing
heuristics.
It is particularly difficult to formulate this problem if we seek a user equilib-
rium, rather than a system optimum, and if if there are multiple destinations.
For this reason, the first DTA models (Merchant and Nemhauser, 1978a,b) were
restricted to a single destination and a system optimum was sought. These au-
thors used an “exit function” link model which is no longer common; using more
recent link models, it is possible to find the system-optimum solution with a sin-
gle destination by solving a linear program (Ziliaskopoulos, 2000). When there
are multiple destinations, ensuring that flow to one destination does not over-
436 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

take that headed to another imposes nonconvexity on the feasible flows (Carey,
1992).
For this reason, variational inequality formulations are more common than
mathematical optimization. Examples of these include Friesz et al. (1993), Wie
et al. (1995), and Chen and Hsueh (1998). Optimal control approaches have
also been proposed (Friesz et al., 1989; Ran et al., 1993), as have formulations
as a nonlinear complementarity problem (Ban et al., 2012). Fixed point ap-
proaches are also common — Bar-Gera (2005) and Bellei et al. (2005) are just
two examples — but as with the static assignment problem, are more useful for
specifying the problem then for solving it.
Using link or path flows as the main decision variable is the most intuitive
choice, and therefore the most common in the literature. However, the use of
splitting proportions is becoming more common (Nezamuddin and Boyles, 2015;
Long et al., 2012; Gentile, 2016).
For the convex combinations and simplicial decomposition algorithms, refer
to the references in Section 6.6. For gradient projection as specialized to dy-
namic network loading, see Nezamuddin and Boyles (2015) and Gentile (2016).
The equilibrium existence counterexample in Section 11.3.1 is from Waller
(2006). The uniqueness counterexample in Section 11.3.2 is from Nie (2010b),
and its special case in Section 11.3.3 is from Boyles et al. (2013). The efficiency
counterexample in Section 11.3.4 is from Daganzo (1998). One consequence
of these counterexamples is that queue spillback significantly complicates the
finding and interpretation of dynamic user equilibria; see also the discussion in
Boyles and Ruiz Juri (2019).

11.5 Exercises
1. [37] (Equivalence of link-based and path-based flow representations). Given
t
splitting proportions αhij,s for each destination s, time interval t, and turn-
ing movement [h, i, j], show how “equivalent” path flow values hπt can be
found. Then, if given path flows hπt , show how “equivalent” αhij,s t
values
can be found. (“Equivalent” means that the link cumulative counts N ↑
and N ↓ would be the same for all time steps after performing network
loading, possibly with a small error due to time discretization that would
shrink to zero as ∆t → 0.)

2. [37] Consider the four-link network shown in Figure 11.13, and perform
network loading using the cell transmission model. (Each link is one cell
long.) During the first time interval, 10 vehicles enter Link 1 on the top
path; during the second time interval, 5 vehicles enter Link 1 on the top
path and 5 on the bottom path; and during the third time interval, 10
vehicles enter Link 1 on the bottom path. No other vehicles enter the
network.
Interpolating as necessary, what time do the first and last vehicles on the
top path exit cell 4? the first and last vehicles on the bottom path? What
11.5. EXERCISES 437

Link 2
1
5
30
Link 1 Link 4
1
L/uf = 1 1
qmax = 20 5
Link 2
30
kjL = 30 1 1
w/uf = 1 5
30
1

Figure 11.13: Network for Exercise 2.

is the derivative of the travel time on the top path for a vehicle leaving at
the start of interval 2?

3. [12] In the course of the convex combinations algorithm, assume that the
H and H ∗ matrices are as below, and λ = 1/3. What is the new H
matrix?    
6 12 18 0
H= H∗ = (11.38)
30 24 0 54

4. [23] Consider a network with only one origin-destination pair connected

by four paths, with three departure time intervals. Departure times are
fixed. At some point in the simplicial decomposition algorithm, H contains
the following three matrices:
   
20 0 0 0 0 20 0 0 20 0 0 0
 0 10 0 0 0 10 0 0  0 0 0 10 (11.39)
0 0 30 0 0 30 0 0 0 0 30 0

and the current path flow and travel time matrices are:
   
14 6 0 0 20 20 24 27
H =0 8 0 2 T (H) = 30 34 37 40 (11.40)
0 9 21 0 44 35 36 40

What are the unrestricted and restricted average excess costs of the current
solution? What is the matrix ∆H?

5. [35] Assume that there is a single OD pair, two paths, and three departure
times; 15 vehicles depart during the first, 10 during the second, and 5
during the third. Let Hπt denote the number of vehicles departing on
path π at time t, and that the path travel times are related to the path
438 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

flows by the following equations:3

T11 = 15
T21 = 13 + 3H11 + H21
2
T12 = 15 + H11
2
T22 = 13 + H11 + 3H12 + H21 + H22
2 1 2
T13 = 15 + H12 + H11
2
2
T23 = 13 + H12 + 3H13 + H22 + H23

Find the path flows obtained from three iterations of the convex combi-
nations method with step sizes λ1 = 1/2, λ2 = 1/4, and λ3 = 1/6 (so you
should find a total of four H ∗ matrices, counting the initial matrix, and
take three weighted averages). Your initial matrix should be the shortest
paths with zero flow. What is the resulting AEC? Break any ties in favor
of path 1..
6. [35] Using the same network and demand as in Exercise 5, now assume
that you are solving the same problem with simplicial decomposition, and
at some stage H contains the following two matrices:
   
15 0 15 0
 0 10 10 0
5 0 0 5
and that the current solution is
 
15 0
H =4 6
3 2
(a) What is the average excess cost of the current solution?
(b) What is the restricted average excess cost?
(c) What is the search direction from H?
(d) Give the updated H matrix and new restricted average excess cost
after taking a step with µ = 0.1.
(e) If we terminate the subproblem and return to the master algorithm,
what matrix (if any) do we add to H?
7. [35] Again using the network and demand from Exercise 5, now start with
the initial solution  
10 5
H =  5 5
0 5
3 In practice these would come from performing network loading and calculating path travel

times as described in this chapter; these functions are provided here for your convenience.
11.5. EXERCISES 439

Using the gradient projection method, identify a new H and T matrix.

Try both exact and quasi-Newton steps, and see which gives the greater
reduction in AEC.
8. [46] Find the mixed-strategy equilibrium in the network in Figure 11.7.
(For each of the two vehicles, indicate the probability it will choose each
of the two paths available to it; these probabilities should be such that
the expected travel times are the same for both options.)
9. [35] Consider a network with one origin, one destination, two possible
departure time intervals (1 and 2), and two paths (A and B). A total of 16
vehicles depart during time interval 1, and 4 depart during time interval
2. Let tπτ refer to the travel time on the π-th path when departing at the
τ -th time interval, with hπτ defined similarly for path flows. Suppose the
travel times are related to the path flows as follows:
tA A A
1 = 0.5h1 + 5h2 + 6

tB B B
1 = 0.5h1 + 3h2 + 10

tA A A
2 = 0.3h1 + 0.6h2 + 0.8

tB B B
2 = 0.2h1 + 0.4h2 + 2

and consider the following four path flow matrices:

       
0 16 12 4 8 8 16 0
H1 = H2 = H3 = H4 =
4 0 2 2 2 2 0 4
(M. Netter, 1972)4
(a) Which of the above path flow matrices satisfy the principle of user
equilibrium?
(b) Which of the above path flow matrices represent efficient equilibria?
(c) Perturb H3 in the following way: adjust hA 2 from 2 to 2.1 (so h2
B
A
becomes 1.9), changing all of the travel times. Then, adjust h1 and
hB A B
1 until t1 = t1 (restoring equilibrium for the first departure time).
Then, adjust hA B A B
2 and h2 until t2 = t2 (restoring equilibrium for
the second departure time), and so on, alternating between the two
departure times, until you reach a new equilibrium solution. Which
equilibrium solution do you end up at? Is H3 stable?
10. [33] Consider an instance of the Daganzo paradox network (Figure 11.12),
but with travel times τ2 = 10, τ3A = 4 and τ3B = 1, capacities of c3B = 10
and c1 = c2 = c3A = c4 = 50 vehicles per time step, and an inflow rate
of Q = 50 vehicles per time step. Assume that the maximum number of
vehicles which can fit on link 3A is (kj L)3A = 300 vehicles.
4 If you are wondering how the travel time on a path can be influenced by drivers leaving

at a later time, recall that FIFO violations can occur due to phenomena such as express lanes
opening, allowing “later” vehicles to overtake “earlier” ones and delay them in the network.
440 CHAPTER 11. DYNAMIC USER EQUILIBRIUM

(a) How many vehicles should be on link 3A at any time to maintain

equilibrium?
(b) Does this queue fit on the link?
(c) The capacity on 3A is now increased to c3A = 25. At the new equi-
librium, have drivers’ travel times increased, decreased, or stayed the
same?
(d) The capacity is now increased to c3A = 50. At the new equilibrium,
have drivers’ travel times increased, decreased, or stayed the same
(relative to c3A = 20)?
11. [55] In the network of Figure 11.12, find the equilibrium solution in terms
of the capacity on link 3B. Plot the average travel time as this capacity
ranges between 0 and 25 vehicles per time step. Assume a time horizon of
20 time steps. (You may do this either by deriving a closed-form expression
for the average travel time, or by using one of the algorithms in this chapter
to solve for the approximate equilibrium solution.)
Appendix A

Mathematical Concepts

This appendix reviews the mathematical background needed to understand the

mathematical formulations in this book. This involves understanding certain
properties of matrices, sets, and functions. The descriptions are relatively terse,
because there are countless textbooks and online resources with additional ex-
planations and examples. It is also a very incomplete overview, only covering
concepts which will be used later in the book and largely assuming that the
reader has seen this material at some point in the past. The section also col-
lects a number of useful results related to matrices, sets, and functions. The
proofs of most of these are left as exercises.

A.1 Indices and Summation

We commonly work with attributes of links or nodes, such as the total number
of trips starting or ending at a node, or the travel time on a link. In such cases,
it makes sense to use the same variable letter for the number of trips produced
at a node (say P ), and to indicate the productions at a specific node with a
subscript or superscript (P1 or P 3 .) Unfortunately, the superscript notation
can be ambiguous, since the superscript in P 3 can be interpreted both as an
exponent, and as the index referring to a particular node. In most cases it will be
clear which meaning is intended, and parentheses can be used in cases which are
unclear: (P 3 )2 is the square of the productions at node 3. To avoid confusion,
subscripts are generally preferred to superscripts, but superscripts can make the
π
notation more compact when there are multiple indices. An example is δij from
Section 1.5, where the variable δ is indexed both by a link (i, j) and a path π. We
have striven to be consistent with which indices appear as subscripts and which
appear as superscripts. However, in a few occasions, being absolutely rigid in
this regard would create cluttered formulas. In such cases, we have sacrificed
rigor for readability, and hope that you are not thrown off by an index typically
appearing as a subscript in the superscript position, and vice versa.
Subscript or superscript indices also allow formulas to be written more com-

441
442 APPENDIX A. MATHEMATICAL CONCEPTS

pactly. A common example is the summation notation, such as

5
X
Pi = P1 + P2 + P3 + P4 + P5 . (A.1)
i=1

The left-hand side of Equation (A.1) is used as shorthand for the right-hand
side. More formally, the left-hand side instructs us to choose all the values of i
between 1 and 5 (inclusive); collect the terms Pi for all of these values of i; and
to add them together. A variant of this notation is
X
Pi , (A.2)
i∈N

which expresses the sum of the productions from all nodes in the set N . Here
i ranges over all elements in the set N , rather than between two numbers as
in (A.1). We can also add conditions to this range by using a colon. If we only
wanted to sum over nodes whose products are less than, say, 500, we can write
X
Pi . (A.3)
i∈N :Pi <500

When it is exceptionally clear what values or set i is ranging over, we can simply
write X
Pi , (A.4)
i

but this “abbreviated” form should be used sparingly, and avoided if there is
any ambiguity at all as to what values i should take in the sum.
When there are multiple indices, a double summation can be used:
3 X
X 3
xij = x12 + x22 + x32 + x13 + x23 + x33 . (A.5)
i=1 j=2

You can think of a double sum either as a “nested” sum:

 
X3 X 3 3
X X3 X3
xij =  xij =
 (xi2 + xi3 ) , (A.6)
i=1 j=2 i=1 j=2 i=1

which expands to the same thing as the right-hand side of (A.5), or as summing
over all combinations of i and j such that i is between 1 and 3, and j is between
2 and 3. Triple sums, quadruple sums, and so forth behave in exactly the same
way, and are common when there are many indices.
The summation index is often called a dummy variable, because the indices
do not have an absolute meaning. Rather, they are only important P5 insofar
as they point to the correct numbers to add up. For instance, i=1 Pi and
P4
P
i=0 i+1 are exactly the same, because both expressions have you add up P1
A.1. INDICES AND SUMMATION 443

P5 P5
through P5 . Likewise, i=1 Pi and j=1 Pj are exactly the same, the fact that
we are counting from 1 to 5 using the variable j instead of i is of no consequence.
Related to this, it is wrong toPrefer to a summation index outside of the sum
5
itself. A formula such as xi + i=1 yi is incorrect. For the formula to make
sense,Pxi needs to refer to one specific value of i. But using i as the index in the
5
sum i=1 means that i must range over all the values between 1 and 5. Does
yi refer to the index of summation (which ranges from 1 to 5), or to the one
specific value of i used outside the sum? If you want to refer to a specific node,
as well as to index over all nodes, you can add a prime to one of them, as in
5
X
xi + yi0 , (A.7)
i0 =1

or you can either use a different letter altogether, as in

5
X
xi + yj . (A.8)
j=1

Both conventions are common.

In transportation network analysis, we frequently have to sum over all of the
links which are in the forward or reverse star of a node (Section 2.3). If the link
flows are denoted by a variable x, the total flow entering node i is the sum of
the link flows in its reverse star. This can be written in several ways:
X X
xhi xhi (A.9)
(h,i)∈Γ−1 (i) (h,i)∈A

The notation on the right can be a bit confusing at first glance, since it looks
like we are summing the flow of every link (the whole set A), rather than just
the links entering node i. The critical point is that in this formula, i refers
to one specific node which was previously chosen and defined outside of the
summation. The only variable we are summing over in Equation (A.9) is h. In
this light, i is fixed, and the right-most sum is over all the values of h such that
(h, i) is a valid link (that is, (h, i) ∈ A). These are exactly the links which form
the reverse star of i.
Almost all of the sums we will see in this book involve only a finite number
of terms. These sums are much easier to work with than infinite sums, and
have the useful properties listed below. (Some of these properties do not always
apply to sums involving infinitely many terms.)
1. You can factor constants out of a sum:
X X
cxi = c xi , (A.10)
i i

no matter what values i ranges over. This follows from the distributive
property for sums: a(b + c) = ab + ac. (A “constant” here is any term
which does not depend on i.)
444 APPENDIX A. MATHEMATICAL CONCEPTS

2. If what you are summing is itself a sum, you can split them up:
X X X
(xi + yi ) = xi + yi . (A.11)
i i i

This follows from the commutative property for sums: a + b = b + a, so

you can rearrange the order in which you add up the terms.
3. You can exchange the order of a double summation:
XX XX
xij = xij (A.12)
i j j i

no matter what values i and j range over. This also follows from the
commutative property.
None of these properties
P are anything new; they simply formalize how we can
operate with the notation using the basic properties of addition. These
properties can also be combined. It is common to exchange the order of a
double sum, and then factor out a term, to perform a simplification:
!
XX XX X X
cj xij = cj xij = cj xij . (A.13)
i j j i j i

The last step is permissible because cj is a “constant” relative to a sum over

i. We could not do this step at the beginning, because cj is not aPconstant
relative to a sum over j. This kind of manipulation is helpful if, say, i xij has
a convenient form or a previously-calculated value.
Occasionally it is useful
P to refer to products over an index. In analogy with
the summation notation used
Q for sums over an index, products over an index
are written with the notation . As an example
5
Y
Pi = P1 P2 P3 P4 P5 (A.14)
i=1

Exercise 3 asks you to investigate which of the properties of summation notation

carry over to the product notation.

A.2 Vectors and Matrices

√
A scalar is a single real number, such as 2, −3, 2, or π. The notation x ∈ R
means that x is a real number (a scalar). A vector is a collection of scalars;
the dimension
 of a vector is the number
 of scalars
 it contains. For instance,
3 −5 is a two-dimensional vector, 6 0 π is a three-dimensional vector,
and so forth. It is possible to write vectors either horizontally, with its com-
ponent scalars in a row, or vertically, with its component scalars in a column.
For the most part it doesn’t matter whether vectors are written in a row or
A.2. VECTORS AND MATRICES 445

in a column; the exception is in formulas involving multiplication between vec-

tors and matrices, as described below, where row and column vectors must be
distinguished. Vectors are usually denoted with boldfaced lower-case letters,
n
like
 x; the notation  x ∈ R indicates that x is an n-dimensional vector.1 (So,
2 3
3 −5 ∈ R and 6 0 π ∈ R .) Individual components of vectors are often
denoted with subscripts or superscripts, as in x1 or x1 . The zero vector is a
vector with zeros for all of its components, and is denoted 0.
Two vectors of the same dimension can be added together  by adding
 the 
corresponding  components of each vector. If x = 1 2 and y = 3 4 ,

then x + y = 4 6 . Multiplying a vector by a scalar  means multiplying
 each

component of the vector by the scalar, so 3x = 3 6 and −y = −3 −4 .
The dot product of two vectors of the same dimension is defined as
X
x·y = xi yi , (A.15)
i

where the sum is taken over all vector components; with the example above,
x · y = 1 × 3 + 2 × 4 = 11.
The dot product of two√vectors is a scalar. The magnitude of a vector x
is given by its norm |x| = x · x. This norm provides a measure of distance
between two vectors; the distance between x and y is given by |x − y|.
The dot product can also be written

x · y = |x||y| cos θ (A.16)

where θ is the angle between the vectors x and y if they are both drawn from
a common point. In particular, x and y are perpendicular if x · y = 0.
A collection of vectors x1 , . . . , xn is linearly independent if the only solution
to the equation a1 x1 + · · · + an xn = 0 is a1 = · · · = an = 0. Otherwise, these
vectors are linearly dependent.
A matrix is a rectangular array of scalars. If a matrix has m rows and n
columns, it is called an m × n matrix, and is an element of Rm×n . A matrix is
square if it has the same number of rows and columns. In this book, matrices
are denoted by boldface capital letters, such as X or Y. In the examples that
follow, let X, Y, and Z be defined as follows:
     
1 2 0 −1 −1 1 −2
X= Y= Z= .
3 4 −3 5 3 −5 8

Elements are matrices are indexed by their row first and column second, so
x11 = 1 and x12 = 2.
Addition and scalar multiplication of matrices works in the same way as
with vectors, so    
1+0 2−1 1 1
X+Y = =
3−3 4+5 0 9
1 There are some exceptions; for instance, shortest-path labels are traditionally denoted

with an upper-case L.
446 APPENDIX A. MATHEMATICAL CONCEPTS

and  
5 10
5X = .
15 20
The transpose of a matrix A, written AT , is obtained by interchanging the
rows and columns, so that if A is an m × n matrix, AT is an n × m matrix. As
examples we have
 
  −1 3
1 3
XT = ZT =  1 −5 .
2 4
−2 8

Matrix multiplication is somewhat less intuitive. Two matrices can only be

multiplied together if the number of columns in the first matrix is the same as
the number of rows in the second. (The reason for this will become clear when
the operation is defined.) For instance, you can multiply a 2 × 3 matrix by a
3 × 3 matrix, but you can’t multiply a 2 × 3 matrix by another 2 × 3 matrix.
This immediately suggests that the order of matrix multiplication is important,
since two matrices may have compatible dimensions in one order, but not in
the other. Multiplying an m × n matrix by a n × p matrix creates an m × p
matrix, defined as follows. Let A ∈ Rm×n and B ∈ Rn×p . Then the product
C = AB ∈ Rm×p has elements defined as
n
X
cij = aik bkj . (A.17)
k=1

If you imagine that each row of the first matrix is treated as a vector, and that
each column of the second matrix is treated as a vector, then each component
of the product matrix is the dot product of a row from the first matrix and
a column from the second. For this dot product to make sense, these row
and column vectors have to have the same dimension, that is, the number of
columns in the first matrix must equal the number of rows in the second. Using
the matrices defined above, we have
   
1 × 0 + 2 × −3 1 × −1 + 2 × 5 −6 9
XY = = .
3 × 0 + 4 × −3 3 × −1 + 4 × 5 −12 17

Observe that the element in the first row and first column of the product matrix
is the dot product of the first row from X and the first column from Y; the
element in the first row and second column of the product is the dot product of
the first row from X and the second column of Y; and so forth. You should be
able to verify that  
5 −9 14
XZ = .
9 −17 26
A matrix and a vector can be multiplied together, if you interpret a row
vector as an 1×n matrix or a column vector as a m×1 matrix. A dot product of
two vectors x and y of equal dimension can be written as a matrix multiplication:
A.2. VECTORS AND MATRICES 447

x · y = xT y if they are both column vectors, or as xyT if they are both row
vectors. Here, the distinction between row and column vectors is important,
because matrices can only be multiplied if their dimensions are compatible. In
matrix multiplications, the convention used in this text is that vectors such as x
are column vectors, and a row vector is denoted by xT . Again, this distinction is
only relevant in matrix multiplication, and for other purposes row and column
vectors can be treated interchangeably.
It is worth repeating that matrix multiplication is not commutative, so that
usually XY 6= YX (and in fact both products may not even exist, depending
on the dimensions of X and Y), although there are some exceptions. You may
wonder why this seemingly-complicated definition of matrix multiplication is
used instead of other, seemingly simpler, approaches. One reason is that this
definition actually ends up representing real-world calculations more frequently
than other definitions. For instance, it can be used to compactly write a set of
equations, as is common in optimization problems (see Section B.3).
A square matrix A is symmetric if aij = aji (in other words, for a symmetric
matrix A = AT ), and it is diagonal if aij = 0 unless i = j (that is, all its
elements are zero except on the diagonal from upper-left to lower-right). A very
special diagonal matrix is the identity matrix, which has 1’s along the diagonal
and 0’s everywhere else. The notation I denotes an identify matrix of any size,
so we can write
 
  1 0 0 0
1 0 0 1 0 0
I= or I= 0 0 1 0 .

0 1
0 0 0 1

In practice, the dimension is usually obvious because of the requirements of

matrix multiplication. The identity matrix has the unique property that AI =
IA = A for any matrix A.
A square matrix A is invertible if there is another square matrix (call it
A−1 ) such that AA−1 = A−1 A = I. Multiplying by an inverse matrix should
be thought of as the equivalent of “matrix division.” Just as it is not possible
to divide by all numbers (division by zero is undefined), not all matrices have
inverses. One can show that a square matrix is invertible if and only if the
vectors forming its rows are linearly independent. The matrices seen in this
book will all be invertible.
Computing the matrix inverse is a bit tedious, and is rarely needed in trans-
portation network analysis. However one special case is worth mentioning:
Proposition A.1. A diagonal matrix A is invertible if and only if all its diag-
onal entries are nonzero; in this case, its inverse A−1 is also a diagonal matrix,
whose diagonal entries are the reciprocal of the diagonal entries in A.
So, for instance,
 −1  
5 0 1/5 0
= .
0 −3 0 −1/3
448 APPENDIX A. MATHEMATICAL CONCEPTS

Finally, a symmetric matrix A ∈ Rn×n is positive definite if the matrix

product xT Ax is strictly positive for any nonzero vector x ∈ Rn , and positive
semidefinite if xT Ax ≥ 0 for any x ∈ Rn whatsoever. As examples, consider
the matrices
     
10 1 0 0 1 10
A= B= C= .
1 5 0 1 10 5
 
The matrix A is positive definite, because for any vector x = x1 x2 , the
matrix product is
  
 10 1 x1
xT Ax = x1 x2 = 10x21 + 2x1 x2 + 5x22 .

1 5 x2

The expression on the right is always positive, because it can be rewritten as

9x21 + 4x22 + (x1 + x2 )2 .

None of those three terms can be negative, and since x 6= 0, at least one of these
terms is strictly positive.  
The matrix B is not positive definite, since if x = 1 0 then xT Bx = 0,
which is not strictly positive. However, it is positive semidefinite, since the
matrix product is
  
 0 0 x1
xT Bx = x1 x2 = x22

0 1 x2

which is surely nonnegative.

The
 matrix C is neither positive definite nor positive semidefinite, since if
x = 1 −1 then
  
 1 10 1
xT Cx = 1 −1

= −14 < 0 .
10 5 −1

Notice that for a matrix to be positive definite or semidefinite, we have to check

whether a condition holds for all possible nonzero vectors. If the condition fails
for even one vector, the matrix is not positive definite or semidefinite.
Checking positive definiteness or semidefiniteness can be tedious. One test
is to find all the eigenvalues of the matrix; if they are all positive, the matrix is
positive definite, and if all are non-negative it is positive semidefinite. However,
diagonal matrices arise fairly often in transportation network analysis, and this
case is easy.
Proposition A.2. A diagonal matrix is positive definite if and only if all its
diagonal entries are strictly positive. A diagonal matrix is positive semidefinite
if and only if all its diagonal entries are nonnegative.
There are also times where we will apply the concept of positive definiteness
to non-symmetric matrices. The idea is the same – we want xT Ax to be strictly
A.3. SETS 449

positive for all nonzero x – but the eigenvalue test does not apply, and there
are non-symmetric matrices which have strictly positive eigenvalues but do not
satisfy xT Ax > 0 for all nonzero x. However, we can form the symmetric part
of the matrix A by calculating 12 (A + AT ). It is easy to show that this is
always
a symmetric matrix, and that xT Ax > 0 if and only if xT 21 (A + AT ) x > 0.
So, when we refer to a non-symmetric matrix being positive definite, what we
will mean is that its symmetric part 12 (A + AT ) is positive definite.
The determinant of a square matrix is occasionally useful in transportation
network analysis (but less so than in other applied mathematics fields). For
a 1 × 1 matrix, its determinant is simply the value of the single entry in the
matrix. For an n × n matrix, the determinant can be computed by the following
procedure: select any row or column of the matrix; for each entry in this row
or column, compute the determinant of the (n − 1) × (n − 1) matrix resulting
from deleting both the row and column of this entry; and alternately add and
subtract the resulting determinants. For our purposes, the determinant can be
used to concisely express other matrix properties. For instance, one can show
that a square matrix is invertible if and only if its determinant is not zero,
and that it is positive definite if and only if the determinants of all the square
submatrices including the element in the first row and column are positive. It
also appears in the characterization of totally unimodular matrices, which are
discussed in Section C.5.2 as a property of certain integer optimization problems
which makes them easier to solve.

A.3 Sets
A set is a collection of any type of objects, denoted by a plain capital letter, such
as X or Y . In transportation network analysis, we work with sets of numbers,
sets of nodes, sets of links, sets of paths, sets of origin-destination pairs, and
so forth. Sets can contain either a finite or infinite number of elements. As
examples, let’s work with the sets

X = {1, 2, 4} Y = {1, 3, 5, 7} ,

using curly braces to denote a set, and commas to list the elements. Set mem-
bership is indicated with the notation ∈, so 1 ∈ X, 1 ∈ Y , 2 ∈ X, but 2 ∈/ Y.
The union of two sets X ∪ Y is the set consisting of elements either in X or in
Y (or both), so
X ∪ Y = {1, 2, 3, 4, 5, 7} .
The intersection of two sets X ∩ Y is the set consisting only of elements in both
X and Y , so
X ∩ Y = {1} .
A set is a subset of another set, denoted by ⊆, if all of its elements also belong
to the second set. With these sets X * Y . Even though the element 1 is in both
X and Y , the elements 2 and 4 are in X but not Y . We do have X ∩ Y ⊂ X
and X ∩ Y ⊂ Y , however.
450 APPENDIX A. MATHEMATICAL CONCEPTS

In this book, sets will take one of three forms:

1. Sets that consist of a finite number of elements, which can be listed as

with X and Y above. This includes the set of nodes in a network, the set
of acyclic paths in a network, and so forth. For a finite set, the notation
|X| indicates the number of elements in X.

2. Sets which are intervals on the real line, such as [−1, 1] or (0, 2]. These
intervals are sets containing all real numbers between their endpoints; a
square bracket next to an endpoint means that the endpoint is included
in the set, while parentheses mean that the endpoint is not included.
Intervals usually contain infinitely many elements.

3. Sets which contain all objects satisfying one or more conditions. For in-
stance, the set {x ∈ R : x2 < 4} contains all real numbers whose square is
less than four; in this case it can be written simply as the interval (−2, 2).
A more complicated set is {(x, y) ∈ R2 : x+y ≤ 3, |x| ≤ |y|}. This set con-
tains all two-dimensional vectors which (if x and y are the two components
of the vector) satisfy both the conditions x + y ≤ 3 and |x| ≤ |y|. It can
also be thought of as the intersection of the sets {(x, y) ∈ R2 : x + y ≤ 3}
and {(x, y) ∈ R2 : |x| ≤ |y|}. If there are no vectors which satisfy all of
the conditions, the set is empty, denoted ∅.

We use the common mathematical conventions that R is the set of all real
numbers, and Z is the set of all integers. If we want to restrict attention to non-
negative real numbers or integers (i.e., positive or zero), the notations R+ and
Z+ are used. Superscripts (e.g., R3 or Z5 ) indicate vectors of a particular dimen-
sion whose elements belong to a particular set: R3 is the set of 3-dimensional
vectors of real numbers, and Z5 the set of 5-dimensional integer vectors. To refer
to a matrix of a particular size, we indicate both dimensions in the superscript;
for instance, Z3×5
+ is the set of matrices with 3 rows and 5 columns, each of
whose elements is a non-negative integer.
Sets of scalars or vectors can be described in other ways. Given any vector
x ∈ Rn , the ball of radius r is the set

Br (x) = {y ∈ Rn : |x − y| < r]} , (A.18)

that is, the set of all vectors whose distance to x is less than r, where r is some
positive number. A ball in one dimension is an interval; a two-dimensional
ball is a circle; a three-dimensional ball is a sphere; a four-dimensional ball a
hypersphere, and so on.
Given some set of real numbers X, the vector x is a boundary point of X if
every ball Br (x) contains both elements in X and elements not in X, no matter
how small the radius r. Notice that the boundary points of a set need not belong
to the set: 2 is a boundary point of the interval (−2, 2). A set is closed if it
contains all of its boundary points. A set X is bounded if every element of X
is contained in a sufficiently large ball centered at the origin, that is, if there is
A.3. SETS 451

X
d
a
b
e

Figure A.1: Some examples of the projection operation.

some r such that x ∈ Br (0) for all x ∈ X. A set is compact if it is both closed
and bounded.
These facts will prove useful:
Proposition A.3. Let f (x1 , x2 , · · · , xn ) be any linear function, that is, f (x) =
a1 x1 + a2 x2 + · · · + an xn for some constants ai , and let b be any scalar.
(a) The set {x ∈ Rn : f (x) = b} is closed.
(b) The set {x ∈ Rn : f (x) ≤ b} is closed.
(c) The set {x ∈ Rn : f (x) ≥ b} is closed.
Proposition A.4. Let X and Y be any sets of scalars or vectors.
(a) If X and Y are closed, so are X ∩ Y and X ∪ Y
(b) If X and Y are bounded, so are X ∩ Y and X ∪ Y
(c) If X and Y are compact, so are X ∩ Y and X ∪ Y .
Combining Propositions A.3 and A.4, we can see that any set defined solely
by linear equality or weak inequality constraints (any number of them) is closed.
If X is a closed set of n-dimensional vectors, and y ∈ Rn is any n-dimensional
vector, the projection of y onto X, written projX (y), is the vector x in X which
is “closest” to y in the sense that |x − y| is minimal. In Figure A.1, we have
projX (a) = b, projX (c) = d, and projX (e) = e.
This subsection concludes with a discussion of what it means for a set to be
convex. This notion is very important, and is described in more detail than the
other concepts, starting with an intuitive definition.
If X is convex, geometrically this means that line segments connecting points
of X lie entirely within X. For example, the set in Figure A.2 is convex, while
452 APPENDIX A. MATHEMATICAL CONCEPTS

Figure A.2: A convex set.

Figure A.3: A nonconvex set with a “bite” taken out of it.

those in Figure A.3 and Figure A.4 are not. Intuitively, a convex set cannot
have any “holes” punched into it, or “bites” taken out of it.
Mathematically, we write this as follows:
Definition A.1. A set X ⊆ Rn is convex if, for all x1 , x2 ∈ X and all λ ∈ [0, 1],
the point λx2 + (1 − λ)x1 ∈ X.
If this definition is not clear, notice that one way to express the line segment
between any two points is λx2 + (1 − λ)x1 , and that you cover the entire line
between x1 and x2 as λ varies between 0 and 1, regardless of how close or far
apart these two points are located.
Example A.1. Show that the one-dimensional set X = {x : x ≥ 0} is convex.
Solution. Pick any x1 , x2 ≥ 0 and any λ ∈ [0, 1]. Because x1 , x2 , and λ are
all nonnegative, so are λx2 and (1 − λ)x1 , and therefore so is λx2 + (1 − λ)x1 .
Therefore λx2 + (1 − λ)x1 belongs to X as well. 
Pn
Example A.2. Show that the hyperplane X = {x ∈ Rn : i=1 ai xi − b = 0} is
convex.
A.3. SETS 453

Figure A.4: A nonconvex set with a “hole” in it.

Pn
Solution. This set is the same as {x ∈ Rn : i=1 ai xi = b} Pick any
x, y ∈ X and any λ ∈ [0, 1]. Then
n
X n
X n
X
ai (λyi + (1 − λ)xi ) = λ ai yi + (1 − λ) ai xi
i=1 i=1 i=1
= λb + (1 − λ)b
=b

so λy + (1 − λ)x ∈ X as well. 
Sometimes, more complicated arguments are needed.
 
Example A.3. Show that the two-dimensional ball B = x y ∈ R2 : x 2 + y 2 ≤ 1
is convex.
Solution. Pick any vectors
 a, b ∈ B and any λ ∈ [0, 1]. We will write the
components of these as a = ax ay and b = bx by . The point λb+(1−λ)a
is the vector λbx + (1 − λ)ax λby + (1 − λ)ay . To show that it is in B, we
must show that the sum of the squares of these components is no greater than
1.

(λbx + (1 − λ)ax )2 + (λby + (1 − λ)ay ))2

= λ2 (b2x + b2y ) + (1 − λ)2 (a2x + a2y ) + 2λ(1 − λ)(ax bx + ay by )
≤ λ2 + (1 − λ)2 + 2λ(1 − λ)(ax bx + ay by )

because a, b ∈ B (and therefore a2x + a2y ≤ 1 and b2x + b2y ≤ 1). Notice that
ax bx + ay by is simply the dot product of a and b, which is equal to |a||b| cos θ,
where θ is the angle between the vectors a and b. Since |a| ≤ 1, |b| ≤ 1 (by
definition of B), and since cos θ ≤ 1 regardless of θ, ax bx + ay by ≤ 1. Therefore

λ2 + (1 − λ)2 + 2λ(1 − λ)(ax bx + ay by )

≤ λ2 + (1 − λ)2 + 2λ(1 − λ) = (λ + (1 − λ))2 = 1
454 APPENDIX A. MATHEMATICAL CONCEPTS

so the point λb + (1 − λ)a is in B regardless of the values of a, b, or λ. Thus

B is convex. 
 
Example A.4. Show that the set C = x y ∈ R2 : x2 + y 2 > 1 is not con-
vex.
   
 Solution.
 Let b1 = 2 0 , b2 = −2 0 , λ = 1/2. Then λb2 +(1−λ)b1 =
0 0 ∈ / C even though b1 , b2 ∈ C and λ ∈ [0, 1]. Therefore C is not convex.

Proving that a set is convex requires showing that something is true for all
possible values of x1 , x2 ∈ X, and λ ∈ [0, 1], whereas disproving convexity only
requires you to pick one combination of these values where the definition fails.
Lastly, this result will prove useful in showing that a set is convex:
Proposition A.5. If X and Y are convex sets, so is X ∩ Y .

A.4 Functions
A function is a mapping between sets. If the function f maps set X to set Y ,
then f associates every element of X with some single element of Y . (Note
that not every element of Y needs to be associated with an element of X.) The
set X is known as the domain of f . Examples include the familiar functions
f (x) = x2 and g(x, y) = x2 + y 2 . The function f maps R to R, while g maps
R2 to R. An example of a function which maps R2 to R2 is the vector-valued
function  
3x1 + 2x2
h(x1 , x2 ) =
−x1 x2
The inverse of a function f , denoted f −1 , “undoes” the mapping f in the sense
that
√ if f (x) = y, then f −1 (y) = x. As an example, if f (x) = x3 , then f −1 (x) =
3
x. Not every function has an inverse, and inverse functions may need to be
restricted to subsets of X and Y .
The composition of two functions f and g, denoted f ◦g involves substituting
function g into function f . If f (x) = x3 and g(x) = 2x + 4, then f ◦ g(x) =
(2x + 4)3 . This can also be written as the function f (g(x)).
A function is continuous if, for any point x̂ ∈ X, the limit limx→x̂ f (x) exists
and is equal to f (x̂). Intuitively, continuous functions can be drawn without
lifting your pen from the paper. A function is differentiable if, for any point
x ∈ X, the limit
f (x + ∆x) − f (x)
lim (A.19)
∆x→0 |∆x|
exists; that limit is then called the derivative of f , and gives the slope of the
tangent line at any point. It can be shown that any differentiable function must
be continuous.
Proposition A.6. Let f and g be continuous functions. Then we have the
following:
A.4. FUNCTIONS 455

ˆ The multiple αf is a continuous function for any scalar α.

ˆ The sum f + g is a continuous function.
ˆ The product f g is a continuous function.
ˆ The composition f ◦ g is a continuous function.
Furthermore, all of these results hold if “continuous” is replaced by “differen-
tiable.”
Proposition A.7. Let X be a nonempty, convex set of n-dimensional vectors.
Then the projection function projX (x) is defined and continuous for all x ∈ Rn .
Differentiability is more complicated when dealing with functions of multiple
variables (that is, functions whose domain is R2 , R3 , and so forth). The basic
notion is the partial derivative, in which all variables except one are assumed
constant, and an ordinary derivative is taken with respect to the remaining
variable. For instance, if f (x, y) = x2 + 3xy + y 3 , the partial derivatives with
respect to x and y are
∂f ∂f
= 2x + 3y = 3x + 3y 2 .
∂x ∂y
Second partial derivatives are found in the same way, taking partial derivatives
of ∂f ∂f
∂x and ∂y . If all of the second partial derivatives of a function are continuous
at a point, the order of differentiation does not matter and
∂2f ∂2f
= (A.20)
∂x∂y ∂y∂x
at that point.
All of these partial derivatives can be organized into different kinds of vec-
tors and matrices. The gradient of a function is the vector of all its partial
derivatives. For f (x, y) = x2 + 3xy + y 3 , this is
   
∂f
2x + 3y
∇f (x, y) =  ∂x  =  .
∂f 2
∂y 3x + 3y

A point x where ∇f (x) = 0 is called a stationary point. (If f is a function of a

single variable, this is just a point where the derivative vanishes.) The Hessian
of a function is the matrix of all of its second partial derivatives. For this same
function, we have
 2   
∂ f ∂2f
2 2 3
Hf (x, y) =  ∂x2 ∂x∂y 
= 
∂ f ∂2f
∂y∂x ∂y 2 3 6y

By equation (A.20), if the Hessian matrix contains continuous functions, it is

symmetric.
456 APPENDIX A. MATHEMATICAL CONCEPTS

Now consider a vector-valued function of multiple variables f : Rn → Rm .

(That is, f has n input variables and produces an m-dimensional vector as
output.) The Jacobian of f is a matrix of all of its first partial derivatives. For
a concrete example, let
  2
x + y2 + z2
 
g1 (x, y, z)
g(x, y, z) = = .
g2 (x, y, z) xyz

The Jacobian is the matrix

   
∂g1 ∂g1 ∂g1
2x 2y 2z
Jg(x, y, z) =  ∂x ∂y ∂z 
= 
∂g2 ∂g2 ∂g2
∂x ∂y ∂z yz xz xy

Be sure to note the difference between Hessians and Jacobians. The Hessian is
defined for scalar-valued functions, and contains all the second partial deriva-
tives. The Jacobian is defined for vector-valued functions, and contains all the
first partial derivatives.
Finally, there is an important notion of function convexity. Confusingly,
this is a different idea than set convexity discussed in Section A.3, although
there are some relationships and similar ideas between them. Set and function
convexity together play a pivotal role in optimization and network equilibrium
problems, so function convexity is discussed at length here. Geometrically, a
convex function lies below its secant lines. Remember that a secant line is the
line segment joining two points on the function. As we see in Figure A.5, no
matter what two points we pick, the function always lies below its secant line.
On the other hand, in Figure A.6, not every secant line lies above the function:
some lie below it, and some lie both above and below it. Even though we can
draw some secant lines which are above the function, this isn’t enough: every
possible secant must lie above the function. For this concept to make sense, the
domain X of the function must be a convex set, an assumption which applies
for the remainder of this section
The following definition makes this intuitive notion formal:
Definition A.2. A function f : X → R is convex if, for every x1 , x2 ∈ X and
every λ ∈ [0, 1],

f ((1 − λ)x1 + λx2 ) ≤ (1 − λ)f (x1 ) + λf (x2 ) (A.21)

and strictly convex if

f ((1 − λ)x1 + λx2 ) < (1 − λ)f (x1 ) + λf (x2 ) (A.22)

for all distinct x1 , x2 ∈ X, λ ∈ (0, 1)

Essentially, x1 and x2 are the two endpoints for the secant line. Since this
entire line segment must be above the function, we need to consider every point
between x1 and x2 . This is what λ does: as λ varies between 0 and 1, the
points λx2 + (1 − λ)x1 cover every point between x1 and x2 . You can think of
A.4. FUNCTIONS 457

Figure A.5: A convex function lies below all of its secants.

Figure A.6: A nonconvex function does not lie below all of its secants.
458 APPENDIX A. MATHEMATICAL CONCEPTS

Figure A.7: All of the relevant points for the definition of convexity.

λx2 + (1 − λ)x1 as a “weighted average,” where λ is the weight put on x2 . For

λ = 0, all the weight is on x1 . For λ = 1, all the weight is on x2 . For λ = 1/2,
equal weight is put on the two points, so the weighted average is the midpoint.
λ = 1/3 corresponds to the point a third of the way between x1 and x2 .
So we need to say that, at all such intermediate x values, the value of f
is lower than the y-coordinate of the secant. The value of the function at
this point is simply f ((1 − λ)x1 + λx2 ). Because the secant is a straight line,
its y-coordinate can be seen as a weighted average of the y-coordinates of its
endpoints, that is, f (x1 ) and f (x2 ). This weighted average can be written as
(1 − λ)f (x1 ) + λf (x2 ), so requiring the function to lie below the secant line is
exactly the same as enforcing condition (A.21) for all possible secant lines: that
is, for all x1 , x2 ∈ X and all λ ∈ [0, 1].
Figure A.7 explains this in more detail. Along the horizontal axis, the secant
endpoints x1 and x2 are shown, along with an intermediate point λx2 +(1−λ)x1 .
The y-coordinates are also shown: at the endpoints, these are f (x1 ) and f (x2 ).
At the intermediate point, the y-coordinate of the function is f (λx2 + (1 −
λ)x1 ), while the y-coordinate of the secant is λf (x2 ) + (1 − λ)f (x1 ). Because
the function is convex, the former can be no bigger than the latter. Time
spent studying this diagram is very well spent. Make sure you understand what
each of the four points marked on the diagram represents, and why the given
mathematical expressions correctly describe these points. Make sure you see
what role λ plays: as λ increases from 0 to 1, the central vertical line moves
from x1 to x2 . (What would happen if we picked x1 and x2 such that x1 > x2 ?
What if x1 = x2 ?)
A.4. FUNCTIONS 459

Example A.5. Is the function f (x) = |x|, x ∈ R convex? Is it strictly convex?

Solution. To see if f is convex, we need to see if (A.21) is true; to see if
it is strictly convex, we need to check (A.22). Furthermore, these inequalities
have to be true for every x1 , x2 ∈ R, and every λ ∈ [0, 1]. It is not enough
to simply pick a few values randomly and check the equations. So, we have to
work symbolically. In this case,

f ((1 − λ)x1 + λx2 ) = |(1 − λ)x1 + λx2 |

≤ |(1 − λ)x1 | + |λx2 | by the triangle inequality
= (1 − λ)|x1 | + λ|x2 | because λ, 1 − λ ≥ 0
= (1 − λ)f (x1 ) + λf (x2 )

Therefore (A.21) is satisfied, so f is convex. To show that it is strictly convex,

we would have to show that the inequality

|(1 − λ)x1 + λx2 | ≤ |(1 − λ)x1 | + |λx2 |

can be replaced by a strict inequality <. However, we can’t do this: for example,
if x1 = 1, x2 = 2, λ = 0.5, the left side of the inequality (|1/2 + 2/2| = 3/2) is
exactly equal to the right side (|1/2| + |2/2| = 3/2). So f is not strictly convex.

Note that proving that f (x) is convex requires a general argument, where
proving that f (x) was not strictly convex only required a single counterexample.
This is because the definition of convexity is a “for all” or “for every” type of
argument. To prove convexity, you need an argument that allows for all possible
values of x1 , x2 , and λ, whereas to disprove it you only need to give one set of
values where the necessary condition doesn’t hold.
Example A.6. Show that every linear function f (x) = ax + b, x ∈ R is convex,
but not strictly convex.
Solution.

f ((1 − λ)x1 + λx2 ) = a((1 − λ)x1 + λx2 ) + b

= a((1 − λ)x1 + λx2 ) + ((1 − λ) + λ)b
= (1 − λ)(ax1 + b) + λ(ax2 + b)
= (1 − λ)f (x1 ) + λf (x2 )

So we see that inequality (A.21) is in fact satisfied as an equality. That’s fine,

so every linear function is convex. However, this means we can’t replace the
inequality ≤ with the strict inequality <, so linear functions are not strictly
convex. 
Sometimes it takes a little bit more work, as in the following example:
Example A.7. Show that f (x) = x2 , x ∈ R is strictly convex.
460 APPENDIX A. MATHEMATICAL CONCEPTS

Solution. Pick x1 , x2 so that x1 6= x2 , and pick λ ∈ (0, 1).

f ((1 − λ)x1 + λx2 ) = ((1 − λ)x1 + λx2 )2
= (1 − λ)2 x21 + λ2 x22 + 2(1 − λ)λx1 x2
What to do from here? Comparing term by term, because λ ∈ (0, 1), we know
that λ2 < lambda and (1 − λ)2 < 1 − λ. Therefore (1 − λ2 )x21 < (1 − λ)x21
and λ2 x22 < λx22 . This takes care of two of the terms, all we have left is Since
x1 6= x2 , (x1 − x2 )2 > 0. Expanding, this means that x21 + x22 > 2x1 x2 . This
means that
(1 − λ)2 x21 + λ2 x22 + 2(1 − λ)λx1 x2 < (1 − λ)2 x21 + λ2 x22 + (1 − λ)(λ)(x21 + x22 )
= (1 − λ)x21 + λx22 (after some algebra)
= (1 − λ)f (x1 ) + λf (x2 )
which proves strict convexity. 
This last example shows that proving convexity can be difficult and nonin-
tuitive, even for simple functions like x2 . The good news is that there are often
simpler conditions that we can check. These conditions involve the first and
second derivatives of a function.
Proposition A.8. Let f : X → R be a differentiable function, where X is a
subset of R. Then f is convex if and only if
f (x2 ) ≥ f (x1 ) + f 0 (x1 )(x2 − x1 )
for all x1 , x2 ∈ X.
Proposition A.9. Let f : X → R be twice differentiable, where X is a subset
of R, and let f be twice differentiable on X. Then f is convex if and only if
f 00 (x) ≥ 0 for all x ∈ X
Equivalent conditions for strict convexity can be obtained in a natural way,
changing ≥ to > and requiring that x1 and x2 be distinct in Proposition A.8.
Proposition A.9 also changes slightly in this case; f 00 (x) > 0 is sufficient for strict
convexity but is not necessary. You are asked to prove these statements in the
exercises. Essentially, Proposition A.8 says that f lies above its tangent lines
(Figure A.8), while Proposition A.9 says that f is always “curving upward.” (A
convex function lies above its tangents, but below its secants.)
These conditions are usually easier to verify than that of Definition A.2.
Example A.8. Show that f (x) = x2 is strictly convex using Proposition A.8
Solution. Pick any x1 , x2 ∈ R with x1 6= x2 . We have f 0 (x1 ) = 2x1 , so we
need to show that
x22 > x21 + 2x1 (x2 − x1 )
Expanding the right-hand side and rearranging terms, we see this is equivalent
to
x21 − 2x1 x2 + x22 > 0
A.4. FUNCTIONS 461

Figure A.8: A convex function lies above its tangents.

or
(x1 − x2 )2 > 0

which is clearly true since x1 6= x2 . Thus f is strictly convex. 

Example A.9. Show that f (x) = x2 is strictly convex using Proposition A.9

Solution. f 00 (x) = 2 > 0 for all x ∈ R, so f is strictly convex.  Much

simpler! If f is differentiable (or, better yet, twice differentiable) checking these
conditions is almost always easier.
Furthermore, once we know that some functions are convex, we can use this
to show that many other combinations of these functions must be convex as
well.

Proposition A.10. If f and g are convex functions, and α and β are positive
real numbers, then αf + βg is convex as well.

Proposition A.11. If f and g are convex functions, then f ◦ g is convex as

well.

Some common convex functions are |x|, x2 , ex , and ax + b. So, Proposi-

tion A.10 tells us that 3x2 + 4|x| is convex. It also tells us that any quadratic
function ax2 + bx + c is convex as long as a > 0. Proposition A.11 says that
2
the composition of two convex functions is convex as well. For instance, ex is
convex, and x4 = (x2 )2 is convex as well.
What about functions of more than one variable? The “shortcut” conditions
in Propositions A.8 and A.9 only apply if the domain of f is one-dimensional.
It turns out that very similar conditions can be given for the multivariable case.
The multi-dimensional equivalent of the first derivative is the gradient, and the
equivalent of the second derivative is the Hessian.
The equivalent conditions on convexity are
462 APPENDIX A. MATHEMATICAL CONCEPTS

Proposition A.12. Let f : X → R be a function whose gradient exists every-

where on X. Then f is convex if and only if

f (x2 ) ≥ f (x1 ) + ∇f (x1 ) · (x2 − x1 )

for all x1 , x2 ∈ X.

Proposition A.13. Let f : X → R be a function whose Hessian exists every-

where on X. Then f is convex if and only if H(f ) is positive semidefinite for
all x ∈ X ; and f is strictly convex if H(f ) is positive definite for all x ∈ X.

Unfortunately, neither of these is as easy to check as the single-dimension

equivalents. In particular, it is rather tedious to check whether or not a ma-
trix is positive semidefinite or not. Fortunately, in many important cases in
transportation network analysis, the Hessian is diagonal and Proposition A.2
applies.
As a final note, one useful link between convex sets and convex functions is
the following result:

Proposition A.14. If g : Rn → R is a convex function, then the set X = {x :

g(x) ≤ 0 is a convex set.

A.5 Exercises
1. [22] For this exercise, let x11 = 5, x12 = 6, x13 = 7, x21 = 4, x22 = 3, and
x23 = 9. Evaluate each of these sums.
P2 P3
(a) i=1 j=1 xij
P2 P3
(b) i=1 j=2 xij
P3
(c) j=1 x1j
P2 P3
(d) j=1 i=1 xji
P3 P1
(e) j=1 k=0 xk+1,j
P2 P3
(f) i=1 j=i xij
Q P
2. [22] Repeat Exercise 1, but with products instead of sums .
P
3. [33] Section A.1 lists three properties of the summation notation Q. For-
mulate and prove analogous properties for the product notation . You
can assume that the product involves only a finite number of factors.

4. [22] Prove Proposition A.1.

5. [24] Prove Proposition A.2.

6. [53] Prove that if a matrix is invertible, its inverse matrix is unique.

A.5. EXERCISES 463

7. [26] For each of the sets below, identify its boundary points and indicate
whether or not it is closed, whether or not it is bounded, and whether or
not it is convex.
(a) (5, 10]
(b) [4, 6]
(c) (0, ∞)
(d) {x : |x| > 5}
(e) {x : |x| ≤ 5}
(f) {(x, y) : 0 ≤ x ≤ 4, −3 ≤ y ≤ 3}
(g) {(x, y) : 4x − y = 1}
8. [34] Prove Proposition A.3.
9. [34] Prove Proposition A.4.
10. [53] Identify the projections of the following points on the corresponding
sets. You may find it helpful to draw sketches.
(a) The point x = 3 on the set [0, 1].
1
(b) The point x = 2 on the set [0, 1].
(c) The point (2, 5) on the unit circle x2 + y 2 = 1.
(d) The point (2, 5) on the line x + y = 1.
(e) The point (2, 5) on the line segment between (0, 1) and (1, 0).
(f) The point (2, 3) on the line segment between (0, 1) and (1, 0).
(g) The point (1, 2, 3) on the sphere x2 + y 2 + z 2 = 2.
11. [20] Prove Proposition A.5.
12. [23] Find the inverses of the following functions.
(a) f (x) = 3x + 4
(b) f (x) = 5ex − 1
(c) f (x) = 1/x
13. [57] Prove Proposition A.6 from first principles, using the definitions of
continuity and differentiability in the text.
14. [65] Prove Proposition A.7, using the formal definition of continuity.
15. [22] Calculate the gradients of the following functions:
(a) f (x1 , x2 , x3 ) = x21 + 2x2 x3 + x23
xy
(b) f (x, y) = x2 +y 2 +1

(c) f (x1 , x2 , d1 , d2 ) = 3x1 + 5x22 + 3(d1 + 4)2 + 5(d2 − 1)3

464 APPENDIX A. MATHEMATICAL CONCEPTS

16. [25] Calculate the Jacobian matrices of the following functions:

 
y
(a) f (x, y) =
−x
 2 
x1 + x2 + x3
(b) f (x1 , x2 , x3 ) =
3x2 − 5x3
 
3x + y
(c) f (x, y) = 3y + x
xy
17. [26] Calculate the Hessian matrices of functions from Exercise 15.
18. [35] Determine which of the following functions are convex, strictly convex,
or neither. Justify your answer rigorously.
(a) f (x) = log x where x ≥ 1.
(b) f (x) = −x3 where x ≤ 0.
(c) f (x) = sin x where x ∈ R.
2
(d) f (x) = ex −2x−3 where x ∈ R.
(
0 if x ≤ 0
(e) f (x) =
x if x ≥ 0

19. [35] Determine which of the following sets is convex. Justify your answer
rigorously.
(a) X = {(x1 , x2 ) ∈ R2 : 4x1 − 3x2 ≤ 0}
(b) X = {x ∈ R : ex ≤ 4}
(c) X = {(x1 , x2 , x3 ) ∈ R3 : x1 = 0}
(d) X = {(x1 , x2 , x3 ) ∈ R3 : x1 6= 0}
(e) X = {1}
20. [31] Show that the integral of an increasing function is a convex function.
21. [54] Show that a differentiable function f of a single variable is convex if
and only if f (x) + f 0 (x)(y − x) ≤ f (y) for all x and y.
22. [54] Show that a twice-differentiable function f of a single variable is
convex if and only if f 00 is everywhere nonnegative.
23. [68] Prove the claims of Exercises 21 and 22 for functions of multiple
variables.
24. [32] Page 460 states that “f 00 (x) > 0 is sufficient for strict convexity
but not necessary.” Give an example of a strictly convex function where
f 00 (x) = 0 at some point.
Appendix B

Optimization Concepts

This appendix provides a brief introduction on formulating optimization prob-

lems. We provide examples of linear, nonlinear, and integer formulations. The
focus in this appendix is on the basic terminology and formulation of these prob-
lems. Specifically, we provide a detailed explanation of representing optimiza-
tion formulations using index and matrix notations. The following appendix
goes into more detail on solution methods.

B.1 Components of an Optimization Problem

An optimization problem is a mathematical model that can help decision mak-
ers arrive at decisions which optimize a specific goal, given the constraints they
face. Every optimization problem has three components: an objective func-
tion, decision variables, and constraints. When one talks about formulating
an optimization problem, it means translating a “real-world” problem into the
mathematical equations and variables which comprise these three components:

Objective function: A mathematical function which represent the goal of the

decision maker, to be either maximized or minimized.

Decision variables: Variables representing the factors or choices that can be

controlled by the decision maker, either directly or indirectly.

Constraints: Equations or inequalities representing restrictions on the values

the decision variables can take.

The objective function, often1 denoted f or z, reflects a single quantity

to be either maximized or minimized. Examples in the transportation world
include “minimize congestion”, “maximize safety”, “maximize accessibility”,
1 But not always! Often problem-specific notation will be used, such as c for cost, and so

forth. The notation given in this section is what is traditionally used if we are referring to a
generic optimization problem outside of a specific context.

465
466 APPENDIX B. OPTIMIZATION CONCEPTS

“minimize cost”, “maximize pavement quality”, “minimize emissions,” “maxi-

mize revenue,” and so forth. You may object to the use of a single objective
function, since real-world problems typically involve many different and conflict-
ing objectives from different stakeholders, and this objection is certainly valid.
There are several reasons why the scope of this book is restricted to single objec-
tives. From a historical perspective, if we don’t know how to optimize a single
objective function, then we have no hope of being able to optimize multiple
objectives simultaneously, since the latter build on the former. From a peda-
gogical perspective, the methods of multi-objective optimization are much more
complex, and the basic concepts of optimization are best learned in a simpler
context first. From a mathematical perspective, the definition of “simultane-
ously optimize” is very tricky, since there is probably no plan which will, say,
simultaneously minimize congestion and agency cost — therefore multiobjective
optimization is “fuzzier,” and this fuzziness can be confusing when first taught.
So, don’t be alarmed by this restriction to a single objective, but do keep it in
the back of your head if you use optimization beyond this book.
The decision variables (often, but not always, denoted as the vector x) reflect
aspects of the problem that you (or the decision maker) can control. This can
include both variables you can directly choose, as well as variables which you
indirectly influence by the choice of other decision variables. For example, if you
are a private toll road operator trying to maximize your profit, you can directly
choose the toll, so that is a decision variable. You can’t directly choose how
many people drive on the road — but because that’s influenced by the toll you
chose, the toll road volume should be a decision variable as well. However, you
should avoid including extraneous decision variables. Every decision variable
in your formulation should either directly influence the objective function, or
influence another decision variable that affects the objective function.
Constraints represent any kind of limitation on the values that the decision
variables can take. The most intuitive types of constraints are those which di-
rectly and obviously limit the choices you can make: you can’t exceed a budget,
you are required by law to provide a certain standard of maintenance, you are
not allowed to change the toll by more than $1 from its current value, and so
forth. One very frequent mistake is to omit “obvious” constraints (e.g., the
toll can’t be negative). An optimization formulation must be complete and not
leave out any constraint, no matter how obvious it may seem to you. Real-world
optimization problems are solved by computers, for which nothing is “obvious.”
The second type of constraint is required to ensure consistency among the de-
cision variables. Following the toll road example from the previous paragraph,
while you can influence both the toll (directly) and the roadway volume (indi-
rectly), the roadway volume must be consistent with the toll you chose. This
relationship must be reflected by an appropriate constraint; say, a demand func-
tion d(τ ) giving the demand d for the tollway in terms of the toll τ , which can
be estimated in a variety of econometric ways. Constraints are the primary way
these linkages between decision variables can be captured.
A particular choice of decision variables x is feasible if it satisfies all of the
constraints, and it is optimal if it maximizes or minimizes the objective function
B.1. COMPONENTS OF AN OPTIMIZATION PROBLEM 467

(whichever is appropriate for the given problem). For some problems, we may
need to distinguish local and global optima. Section B.6.1 discusses this in
greater detail. In short, we usually want to find a global optimum, but for some
complicated optimization problems a local optimum may be the best we can do.
The following example explains these definitions:
Example B.1. A plant produces two types of concrete (Type 1 and Type 2). The
owner of the plant makes a profit of $90 for each truckload of Type 1 concrete
produced and a profit of $120 for each truckload of Type 2 concrete produced.
Three materials are required to produce concrete: cement, aggregate, and water.
The plant requires 30 units of cement, 50 units of aggregate, and 60 units of
water to produce one truckload of Type 1 concrete. The plant requires 40 units
of cement, 20 units of aggregate, and 90 units of water to produce one truckload
of Type 2 concrete. The plant is supplied 2000 units of cement, 2500 units of
aggregate, and 4000 units of water daily. How many truckloads of Type 1 and
Type 2 cement should the plant produce to maximize daily profit?
In any optimization problem, the first step is to identify the decision variables
or the variables which can be controlled by the decision maker. In this problem,
the decision variables are the daily truckloads of Type 1 and Type 2 concrete
produced by the plant.
Let x represent the daily truckloads of Type 1 concrete produced by the
plant and y represent the daily truckloads of Type 2 concrete produced by the
plant.
The second step is to identify the constraints or restrictions on the decision
variables. In this problem there are restrictions on the total volume of cement,
aggregate, and water supplied to the plant daily which limits the total amount
of Type 1 and Type 2 concrete which can be produced.
Each truckload of Type 1 concrete requires 30 units of cement. Therefore,
to produce x truckloads of Type 1 concrete requires 30x units of cement. Each
truckload of Type 2 concrete requires 40 units of cement. Therefore, to produce
y truckloads of Type 2 concrete requires 40y units of cement. The plant is
supplied 2000 units of cement daily. Therefore, the constraint on the total
cement consumed by the plant can be written as below:

30x + 40y ≤ 2000 (B.1)

The plant is supplied 2500 units of aggregate and 4000 units of water daily.
Along similar lines, the constraint on the total volume of aggregate and water
consumed by the plant can be written as below:

50x + 20y ≤ 2500 (B.2)

60x + 90y ≤ 4000 (B.3)

In addition, the daily truckloads of Type 1 and Type 2 concrete produced

has to be greater than or equal to zero. The non-negativity constraint can be
represented as:
x ≥ 0, y ≥ 0 (B.4)
468 APPENDIX B. OPTIMIZATION CONCEPTS

Each Type 1 concrete truckload generates a profit of $90. Therefore, x

truckloads of Type 1 concrete generates $90x profit. Each Type 2 concrete
truckload generates a profit of $120. Therefore, y truckloads of Type 2 concrete
generates a profit of $120y. The goal of the decision maker is to maximize this
daily profit which can be written as:

max 90x + 120y (B.5)

The entire problem can be summarized as below. The following set of equa-
tions represents an optimization formulation or a mathematical programming
formulation for the concrete plant profit maximizing problem.

min 90x + 120y

x,y
s.t. 50x + 20y ≤ 2500
60x + 90y ≤ 4000
x, y ≥0

Here x and y are written underneath max to indicate that these are the
decision variables. In cases where it is obvious what the decision variables are,
we sometimes omit writing them below the max or min in the objective. More
complicated optimization problems might use letters to name things other than
decision variables, and in those cases it is helpful to explicitly write down which
variables the decision maker can affect. The abbreviation “s.t.” stands for
“subject to” or “such that” and indicates the constraints.
The above example has linear objective functions and constraints. Opti-
mization formulations can also have nonlinear functions as shown in the exam-
ple below. The objective in the above example is to maximize the profit. When
the decision maker is concerned with controlling costs rather than profit, the
objective function often involves minimization.

Example B.2. A business has the option of setting up concrete plants at two
locations. The daily cost operating a plant at Location 1 per unit of production
has two components: a fixed cost of 20 and an additional cost which increases
by 0.5 for every unit of production. Similarly, the daily cost of operating a plant
at Location 2 per unit production has a fixed cost of 12 and an additional cost
which increases by 1.2 for every unit of production. The business has committed
to supplying at least 12 units of concrete daily. What is the total amount to be
produced at each location to minimize cost and satisfy demand?

There are two decision variables in this problem: the production in plant
at Location 1 and production in plant at Location 2. Let x denote the units
of production in plant at Location 1 and y represent the units of production in
plant at Location 2.
The business has agreed to supply at least 12 units of concrete daily. There-
fore, the sum of production in plants at Location 1 and Location 2 must be
greater than equal to 12.
B.1. COMPONENTS OF AN OPTIMIZATION PROBLEM 469

x + y ≥ 12 (B.6)
Also, the production at both plants cannot be negative which is represented
as:

x ≥ 0, y ≥ 0 (B.7)
The daily cost of operating a plant at Location 1 per unit of production
is 20 + 0.5x. Since the plant produces x units of concrete, the total cost of
operating a plant at Location 1 is (20 + 0.5x)x. Similarly, the total cost of
operating a plant at Location 2 is (12 + 1.2y)y. The goal of the business is to
minimize the total cost of operation which is given as:

min (20 + 0.5x)x + (12 + 1.2y)y (B.8)

The optimization formulation can thus be summarized as below:

min 20x + 0.5x2 + 12y + 1.2y 2

x,y
s.t. x+y ≥ 12
x, y ≥0

Optimization models are widely used in resource allocation where a fixed

resource has to be allocated among various tasks to maximize or minimize ob-
jectives. A common variant of the resource allocation for civil engineers is the
budget allocation problem. In the budget allocation problem, a fixed budget
has to be distributed among various projects. The objective can vary depend-
ing on the nature of the decision maker. For example, private entities might
be interested in maximizing returns or profit whereas public entities might be
interested in maximizing social welfare or equity.
Example B.3. A business is seeking to invest in two public-private partnership
projects. Project 1 yields an expected return of $10 and a standard deviation of
$5 for each dollar invested. Project 2 has an expected return of $12 and a
standard deviation of $7 for each dollar invested. For any investment portfolio,
the business’s utility is described as U = µ − 21 σ, where µ represents the average
return on investment and σ the standard deviation of return. The business has
a total of $1000 to invest. Assume that the returns on investments on both
projects are independent. Determine the amount of money to be invested in
both projects so that the business can maximize its utility?
There are two decision variables in this problem: the amount of money
invested in Project 1 (x), and the amount of money invested in Project 2 (y).
We know the total amount of money available is $1000. Therefore:
x + y ≤ 1000 (B.9)
Also, the business cannot invest negative money in either project:
x ≥ 0, y ≥ 0 (B.10)
470 APPENDIX B. OPTIMIZATION CONCEPTS

Project 1 yields an average return of $10 for each dollar invested. Therefore,
with an investment of $x, the average return is $10x. Project 2 yields an average
return of $12 for each dollar invested. Therefore, with an investment of $y, the
average return is $12y. The total average return on investment is given as
10x + 12y.
The standard deviation of return on project 1 is $5 for each dollar invested
and for project 2 is $7 for each dollar invested.
p Therefore, the standard deviation
of return on investments is given as 25x2 + 49y 2 . p
The business’s utility is given as: U = µ− 12 σ = (10x+12y)− 21 25x2 + 49y 2 .
The optimization formulation can be summarized as follows:
p
min 10x + 12y − 0.5 25x2 + 49y 2
x,y
s.t. x+y ≤ 1000
x, y ≥0

B.2 Index Notation

In the previous section, we used the symbols x and y to represent the decision
variables. Each symbol corresponds to a scalar. In this section, we will de-
velop an optimization formulation for a problem with more decision variables
and constraints. You will notice that using scalar based symbols will become
cumbersome as the size of the formulation increases. We will then introduce the
set and index notations which will help represent the optimization formulation
in a compact manner using more convenient notation. First, let us consider the
following problem.
Example B.4. A business has five factories located at Pittsburgh, Boston,
Austin, Los Angeles, and Miami producing two types of window frames. It
costs $10 and $26 to produce one unit of window frame of Type 1 and Type
2 respectively. The business has committed to supplying at least 300 units of
window frame of Type 1 and at least 200 units of window frame of Type 2 each
week. Due to the number of workers employed at each location, the maximum
number of window frames which can be produced at Pittsburgh, Boston, Austin,
Los Angeles, and Miami are 100, 125, 100, 125, and 50 respectively. The cost
of producing one unit of window frame of Type 1 at Pittsburgh, Boston, Austin,
Los Angeles, and Miami is $10, $10, $25, $30, and $30 respectively. The cost
of producing one unit of window frame of Type 2 at Pittsburgh, Boston, Austin,
Los Angeles, and Miami is $40, $40, $15, $20, and $20 respectively. Determine
the number of window frames of Type 1 and Type 2 which the business has to
produce at each location with the objective of minimizing cost?
Let us first identify the decision variables. In this problem, the business has
to decide how many units of window frames of Type 1 and Type 2 to produce
at the five locations: Pittsburgh, Boston, Austin, Los Angeles, and Miami.
Therefore, there are 10 decision variables.
B.2. INDEX NOTATION 471

Let a, b, c, d, and e represent the number of units of window frames of

Type 1 to be produced per week at Pittsburgh, Boston, Austin, Los Angeles,
and Miami respectively. Let p, q, r, s, and t represent the number of units
of window frames of Type 2 to be produced per week at Pittsburgh, Boston,
Austin, Los Angeles, and Miami respectively.
The business has committed to supplying at least 300 units of window frames
of Type 1 per week. Therefore, the total amount of window frames of Type 1
produced per week must be greater than or equal to 300.

a + b + c + d + e ≥ 300 (B.11)
The business has committed to supplying at least 200 units of window frames
of Type 2 per week. Therefore, the total amount of window frames of Type 2
produced per week must be greater than or equal to 200.

p + q + r + s + t ≥ 200 (B.12)
The Pittsburgh factory can produce a maximum of 100 units of window
frames per week. Therefore, the sum of the total number of window frames
of Type 1 produced per week at Pittsburgh and the total number of window
frames of Type 2 produced per week at Pittsburgh must be less than or equal
to 100.

a + p ≤ 100 (B.13)
The factory at Boston can produce a maximum of 125 units of window frames
per week. Therefore,

b + q ≤ 125 (B.14)
The factory at Austin can produce a maximum of 100 units of window frames
per week. Therefore,

c + r ≤ 100 (B.15)
Along similar lines, the factories at Los Angeles and Miami are constrained
to produce a maximum of 125 and 50 window frames respectively.

d + s ≤ 125 (B.16)
e + t ≤ 50 (B.17)

The number of window frames of both types produced per week at all loca-
tions has to be greater than or equal to zero. The non-negativity constraints
are represented as:

a ≥ 0, b ≥ 0, c ≥ 0, d ≥ 0, e ≥ 0 (B.18)
472 APPENDIX B. OPTIMIZATION CONCEPTS

p ≥ 0, q ≥ 0, r ≥ 0, s ≥ 0, t ≥ 0 (B.19)
The cost of producing a units of window frames of Type 1 and p units of
window frames of Type 2 at Pittsburgh is 10a+40p. Along similar lines, the cost
of producing specific number of window frames of both types at the other four
locations can be determined. The total production cost is the sum of production
costs at each location which is given as: 10a + 10b + 25c + 30d + 30e + 40p +
40q + 15r + 20s + 20t. The objective is to minimize the total production costs.
The optimization formulation can thus be summarized as shown below

min 10a + 10b + 25c + 30d + 30e + 40p + 40q + 15r + 20s + 20t
a,...,e,p,...,t
s.t. a+b+c+d+e ≥ 300
p+q+r+s+t ≥ 200
a+p ≤ 100
b+q ≤ 125
c+r ≤ 100
d+s ≤ 125
e+t ≤ 50
a, b, c, d, e ≥0
p, q, r, s, t ≥0

This formulation is correct, but a bit unwieldy. To someone looking just

at the formulation, it is hard to tell which variables refer to what. Index no-
tation, presented in the following subsections, can improve the presentation of
formulations like this one.

B.2.1 Single index notation

This section introduces sets and indices, and explains how to use notation based
on them to present formulations concisely. First imagine that the business had
factories at ten locations and was producing five types of window frames. We
would need 50 different symbols to denote the window frame types produced
at all locations! The formulation can be represented in a simpler way by using
set and index notation. This section first presents the single index notation. A
more compact multiple index notation is provided in the following section.
A set is a collection of similar objects. Let I denote the set of all locations.
For the above example, let

I = {Pittsburgh, Boston, Austin, Los Angeles, Miami} . (B.20)

In the above example, we are given the names of locations of five factories.
In some cases, we may not know or we may not be interested in the names of
the locations. All we may know or care about is that there are five factories
located in five different places. In such case, the set I can be represented as
B.2. INDEX NOTATION 473

I = {1, 2, 3, 4, 5 .} (B.21)
The above notation can become cumbersome if the number of locations is
high. So a shorter form representation is

I = {1, . . . , 5 .} (B.22)
The above notation can be further generalized for any value n denoting
number of locations as

I = {1, . . . , n .} (B.23)
An index is used to refer to any element of the set. The symbol ∈ represents
“an element of”. Therefore, i ∈ I denotes an index i which is an element of the
set I. Thus when I = {Pittsburgh, Boston, Austin, Los Angeles, Miami}, i can
refer to any of Pittsburgh, Boston, Austin, Los Angeles, or Miami. Indices can
be combined with symbols to represent decision variables and input parameters
in a concise manner.
In the previous example, we used the symbols a, b, c, d, and e to represent
the number of units of window frames of Type 1 to be produced per week at
Pittsburgh, Boston, Austin, Los Angeles, and Miami respectively. Using the set-
index notation, let xi denote the number of units of window frames per week of
Type 1 to be produced at location i ∈ I. Thus a corresponds to xPittsburgh , b
corresponds to xBoston , and so on.
The business has to produce at least 300 units of window frames of Type 1
per week. This constraint can be represented as:

xPittsburgh + xBoston + xAustin + xLos Angeles + xMiami ≥ 300 (B.24)

The above constraint can be succinctly represented using the summation

operator Σ as
X
xi ≥ 300 . (B.25)
i∈I

If the set I = {1,. . . , 5}, then the above constraint can also be represented
using either of the following equations:
5
X
xi ≥ 300 (B.26)
i=1
X
xi ≥ 300 . (B.27)
1≤i≤5

Similarly, let yi denote the number of units of window frames per week of
Type 2 to be produced at location i ∈ I. As before, the constraint on minimum
474 APPENDIX B. OPTIMIZATION CONCEPTS

number of units of Type 2 window frames produced can be written as any of

the following::
X
yi ≥ 200 (B.28)
i∈I
5
X
yi ≥ 200 (B.29)
i=1
X
yi ≥ 200 (B.30)
1≤i≤5

Now, look at the constraints which limits the number of window frames
produced at each location. Consider the case where the city names are writ-
ten explicitly, so I = {Pittsburgh, Boston, Austin, Los Angeles, Miami}. The
constraints are:

xPittsburgh + yPittsburgh ≤ 100 (B.31)

xBoston + yBoston ≤ 125 (B.32)
xAustin + yAustin ≤ 100 (B.33)
xLos Angeles + yLos Angeles ≤ 125 (B.34)
xMiami + yMiami ≤ 50 (B.35)
This will become very cumbersome to write as the number of locations and
window frame types increase. Let ui represent the maximum amount of window
frame which can be produced at location i ∈ I. For example, uPittsburgh = 100,
uLos Angeles = 125. The symbol ∀ represents “for all”. Therefore, ∀i ∈ I implies
for all the elements in the set I. All of these constraints can then be written as
a single equation:

xi + yi ≤ ui ∀i ∈ I (B.36)
The above equation denotes for each element i ∈ I, the constraints xi + yi ≤
ui holds. Similarly the non-negativity constraints can be written as:

xi ≥ 0 ∀i ∈ I (B.37)
yi ≥ 0 ∀i ∈ I (B.38)
When I = {1,. . . , 5}, the production limit at each location and non-negativity
constraints can also be represented using either of the two following set of equa-
tions:

xi + yi ≤ ui ∀i = 1, . . . , 5 (B.39)
xi ≥ 0 ∀i = 1, . . . , 5 (B.40)
yi ≥ 0 ∀i = 1, . . . , 5 (B.41)
B.2. INDEX NOTATION 475

or:

xi + yi ≤ ui ∀1 ≤ i ≤ 5 (B.42)
xi ≥ 0 ∀1 ≤ i ≤ 5 (B.43)
yi ≥ 0 ∀1 ≤ i ≤ 5 (B.44)

The objective function for the formulation was:

min 10a + 10b + 25c + 30d + 30e + 40p + 40q + 15r + 20s + 20t (B.45)

Using the index-set notation, the objective function can be written

min 10xPittsburgh + 10xBoston + 25xAustin + 30xLos Angeles + 30xMiami

+ 40yPittsburgh + 40yBoston + 15yAustin + 20yLos Angeles + 20xMiami . (B.46)

Let vi represent the cost of producing one window frame of Type 1 at location
i ∈ I and wi represent the cost of producing one window frame of Type 2 at
location i ∈ I. For example vMiami = 30, wMiami = 20. The objective function
can now be succinctly represented as:
X
min (vi xi + wi yi ) (B.47)
i∈I

When I = {1,. . . , 5} the objective function can also be represented as:

5
X
min (vi xi + wi yi ) (B.48)
i=1
or: X
min (vi xi + wi yi ) (B.49)
1≤i≤5

Therefore, the formulation can be rewritten as

X
min (vi xi + wi yi )
xi ,yi
i∈I
X
s.t. xi ≥ 300
i∈I
X
yi ≥ 200
i∈I
xi + yi ≤ ui ∀i ∈ I
xi ≥ 0 ∀i ∈ I
yi ≥ 0 ∀i ∈ I
P5 P
or equivalently, with i=1 or 1≤i≤5 on the summations.
476 APPENDIX B. OPTIMIZATION CONCEPTS

In addition to making the formulation more compact, the set index notation
also makes the formulation easier to understand, and easier to change (if there
were more cities, all we would have to change is the definition of the set I or
the number 5 to whatever the new number of cities is). Notice also that now it
is important to specify that xi and yi are the decision variables: ui , vi , and wi
now represent given problem data which we cannot change.

B.2.2 Multiple index notation

This section introduces a notation that further simplifies the example presented
in the previous subsection . Previously we had set

I = {Pittsburgh, Boston, Austin, Los Angeles, Miami}

or I = {1, . . . , 5}. Two sets of decision variables xi , yi were used to represent

window frames of Type 1 and 2 produced at all locations i ∈ I. If the business
was producing ten types of window frames, then even the single index nota-
tion becomes cumbersome, as we would need ten different subscripted symbols,
xi , yi , zi , . . . and so on.
This issue can be addressed using multiple indices. Let us introduce the set
J to model the two types of window frames. Similar to the set of locations, J
can be defined in two ways: J = {Type 1, Type 2} or J = {1, 2}. Let j be an
index referring to any element in J , i.e., j ∈ J .
Instead of using two symbols with subscripts for location, we will define
the decision variable using one symbol with two subscripts. Let xij denote the
number of type j ∈ J window frames produced per week at location i ∈ I.
When the sets were I = {Pittsburgh, Boston, Austin, Los Angeles, Miami}
and J = {Type 1, Type 2}, in the single index notation two equations (one for
Type 1, other for Type 2) were used to represent the total production must be
greater than equal to demand.

X
xi ≥ 300 (B.50)
i∈I
X
yi ≥ 200 (B.51)
i∈I

Using the double index notation, the two equations can be summarized into
one equation as shown below, by introducing bj to be the demand for window
frames of type j ∈ J . For this example, bType 1 = 300 and bType 2 = 200.
X
xij ≥ bj ∀j ∈ J (B.52)
i∈I

Pay close attention to the two set element references, ∀j ∈ J on the right
hand side and i ∈ I underneath the summation
P operator. The ∀j ∈ J on the
right hand side ensures that the equation i∈I () is repeated for each element
j ∈ J as shown below.
B.2. INDEX NOTATION 477

xPittsburgh,j + xBoston,j + xAustin,j + xLos Angeles,j + xMiami,j ≥ bj ∀j ∈ J (B.53)

Since the set J has two elements, the equation is repeated twice representing
the demand needing to be met for Type 1 and Type 2 window frames. In the
single index notation xi + yi ≤ ui ∀i ∈ I is used to represent the constraint
on maximum window frames which can be produced at each location. In the
double index notation, the left hand side can be made more compact using a
summation operator as shown below.
X
xij ≤ ui ∀i ∈ I (B.54)
j∈J

Pay close attention to the indices over which the summation (j ∈ J ) is

happening and the elements on the right hand side (i ∈ I). The above expression
repeats the following equation for each element or each location i ∈ I.

xi,Type 1 + xi,Type 2 ≤ ui ∀i ∈ I (B.55)

In the single index notation, two sets of equations were used to represent
the non-negativity conditions - xi ≥ 0 ∀i ∈ I and yi ≥ 0 ∀i ∈ I. This can be
concisely represented using a single equation as follows:

xij ≥ 0 ∀i ∈ I, j ∈ J (B.56)
You are enforcing xij to be ≥ 0 for each element or location i ∈ I as well
as window frame type j ∈ J . The order in which you reference the elements
and sets on the right hand side after ∀ does not matter, i.e., the following two
equations represent the same non-negativity constraints.

xij ≥ 0 ∀i ∈ I, j ∈ J (B.57)
xij ≥ 0 ∀j ∈ J , i ∈ I (B.58)

Using
P the single index notation, the objective function was represented as
min i∈I (vi xi + wi yi ). InP
the double index notation the objective function
can be represented as min i∈I (vi xi,Type 1 + wi x,:Type 2 ). We can make the
representation even more compact and more intuitive by defining vij as the cost
of producing one window frame of type j ∈ J at location i ∈ I. The objective
function can then be defined as:
XX
min vij xij (B.59)
i∈I j∈J

Note
P thatP the order inPwhichP we sum the objective function does not matter,
i.e, i∈I v x
j∈J ij ij = j∈J i∈I vij xij . (See Section A.1).
The formulation in double index notation is finally:
478 APPENDIX B. OPTIMIZATION CONCEPTS

5 X
X 2
min vij xij
xij
i=1
X j=1
s.t. xij ≥ bj ∀j ∈ J
i∈I
X
xij ≤ ui ∀i ∈ I
j∈J
xij ≥ 0 ∀i ∈ I, j ∈ J

B.2.3 Additional information on index notation

This section provides a brief overview of some mathematical notation which is
commonly seen in the optimization literature.
The set R contains all real numbers. The set of all non-negative real numbers
is commonly represented as R+ . Another way to represent the non-negativity
constraint xij ≥ 0 ∀i ∈ I, j ∈ J is shown below:

xij ∈ R+ ∀i ∈ I, j ∈ J (B.60)
In the above equation each decision variable xij is restricted to lie in the
set of non-negative real numbers. If the decision variables can be positive or
negative real numbers or zero, then the above equation can be modified as:

xij ∈ R ∀i ∈ I, j ∈ J (B.61)
2
The set Z contains the set of all integers. The set of all non-negative integers
is commonly represented as Z+ . In certain type of optimization problems, called
integer programs, the decision variables are restricted to be integers or non-
negative integers which can be represented as follows:

xij ∈ Z ∀i ∈ I, j ∈ J (B.62)
xij ∈ Z+ ∀i ∈ I, j ∈ J (B.63)

In some cases, there will be certain constraints defined on a subset of indices.

Let us assume in the above example, all the window frames produced on the east
coast of the United States are first transported to a warehouse before delivery
to retailers. The warehouse can store only 150 window frames each week. This
restriction can be represented by the following constraint equation:

xBoston,Type 1 + xBoston,Type 2 + xMiami,Type 1 + xMiami,Type 2 ≤ 150 (B.64)

The above constraint can be more concisely represented by defining a subset

K(I) = {i ∈ I : i lies on East Coast}. In this expression : corresponds to
2 The use of the letter ‘Z’ is from the German word “Zahlen,” meaning “numbers.”
B.3. VECTOR AND MATRIX NOTATIONS 479

“such that”. Another way to present this information is using the | symbol.
K(I) = {i ∈ I|i lies on East Coast}. Given the definition of the subset K(I)
( which can also be represented as KI ), the constraint can be more succinctly
presented as:
X X
xkj ≤ 150 (B.65)
k∈K(I) j∈J

The set notation can also be used to represent various mathematical expres-
sions of decision variables in a clean manner. Let I = {1, . . . , 10}. We want to
represent the expression x2 + x4 + x6 + x8 + x10 in a concise manner. One way
to do this would be to define K(I) = {i ∈ I : i is even} and then write:
X
x2 + x4 + x6 + x8 + x10 = xk (B.66)
k∈K(I)

Another way to represent this would be to indicate the condition expression

in the summation operator itself

X X
x2 + x4 + x6 + x8 + x10 = xi = xi (B.67)
{i∈I:i is even} {i∈I|i is even}

Along similar lines:

X X
x3 + x4 + x5 = xi = xi (B.68)
{i∈I:3≤i≤5} {i∈I|3≤i≤5}
X X
x7 + x8 + x9 + x10 = xi = xi (B.69)
{i∈I:i≥7} {i∈I|i≥7}

(B.70)

B.3 Vector and Matrix Notations

Vector and matrix notation, introduced in Section A.2, are also very widely
used to represent optimization formulations concisely. This section explains
these conventions. To start, when we write equations or inequalities involving
vectors, we mean that they apply to every component of those vectors. For
example, x ≥ b, means that every element in x is greater than or equal to the
corresponding element in b.


 x1 ≥ b1

x2 ≥ b2

x ≥ b =⇒ . (B.71)

 ..


xn ≥ bn

480 APPENDIX B. OPTIMIZATION CONCEPTS

Thus x ≥ 0 implies each element of x is greater than or equal to zero.

Similarly,


 a11 x1 + a12 x2 + · · · + a1n xn = b1

a21 x1 + a22 x2 + · · · + a2n xn = b2

Ax = b =⇒ . (B.72)

 ..


am1 x1 + am2 x2 + · · · + amn xn = bn


Also,
n
X
bT x = b1 x1 + b2 x2 + . . . + bn xn = bi x i (B.73)
i=1
is another way to express the dot product of two vectors x and b.
Given the above information, we now reformulate the window frame opti-
mization formulation using vectors and matrices. Let x = (x1 , . . . , x10 ) repre-
sent a column vector of decision variables, i.e.,
 
x1
 x2 
x=  ... 
 (B.74)
x10
In the vector x, let x1 , x2 , x3 , x4 , and x5 represent the five decision vari-
ables corresponding to the amount of window frame of Type 1 produced at
the five locations and x6 , x7 , x8 , x9 , and x10 represent the five decision vari-
ables corresponding to the amount of window frame of Type 2 produced at the
five locations. Let c = (c1 , . . . , c10 ) represent a column vector of costs where
c1 , c2 , c3 , c4 , and c5 represent the cost of producing one unit of Type 1 window
frame at the five different locations and c6 , c7 , c8 , c9 , and c10 represent the cost
of producing one unit of Type 2 window frame at the five different locations.
   
c1 10
 c2   10 
   
 c3   25 
   
 c4   30 
   
 c5   30 
c= 
=
   (B.75)
 c6   40 

 c7   40 
   
 c8   15 
   
 c9   20 
c10 20
The objective function in this case is:

min 10x1 +10x2 +25x3 +30x4 +30x5 +40x6 +40x7 +15x8 +20x9 +20x10 (B.76)

Let cT represent the transpose of the vector c, so cT is a row vector. The

objective function can then be compactly represented as:
B.3. VECTOR AND MATRIX NOTATIONS 481

10
X
min cT x = ci xi (B.77)
i=1

The production of window frame of Type 1 must be greater than 300 and
window frame of Type 2 must be higher than 200.

x1 + x2 + x3 + x4 + x5 ≥ 300 (B.78)
x6 + x7 + x8 + x9 + x10 ≥ 200 (B.79)

Let us represent the above set of constraints in a compact manner. Define

a matrix A as:
 
1 1 1 1 1 0 0 0 0 0
A= (B.80)
0 0 0 0 0 1 1 1 1 1

Let b be a column vector of demands.

   
b1 300
b= = (B.81)
b2 200
The constraints can now be represented as
   
x1 + x2 + x3 + x4 + x5 300
Ax = ≥ =b (B.82)
x6 + x7 + x8 + x9 + x10 200

or in a compact form, simply as:

Ax ≥ b (B.83)

The constraints on the number of window frames produced at each location is

given as:

x1 + x6 ≤ 100 (B.84)
x2 + x7 ≤ 125 (B.85)
x3 + x8 ≤ 100 (B.86)
x4 + x9 ≤ 125 (B.87)
x5 + x10 ≤ 50 (B.88)

Define a column vector u and matrix U as:

 
100
 125 
 
u= 100 (B.89)


 125 
50
482 APPENDIX B. OPTIMIZATION CONCEPTS

 
1 0 0 0 0 1 0 0 0 0

 0 1 0 0 0 0 1 0 0 0 

U=
 0 0 1 0 0 0 0 1 0 0 
 (B.90)
 0 0 0 1 0 0 0 0 1 0 
0 0 0 0 1 0 0 0 0 1
Thus the production at each location being lesser than the capacity can be
written in a compact manner as:

Ux ≤ u (B.91)

The nonnegativity constraints can be represented as x ≥ 0. Thus the final

optimization formulation can be given as:

min cT x
x
s.t. Ax ≥b
Ux ≤u
x ≥0

Another way to represent the non-negativity constraints is x ∈ R10 + . In the

general case, when the decision variable vector x has n elements, the nonneg-
ativity constraints can be represented as x ∈ Rn+ . If the decision variables are
restricted to the set of positive integers, then x ∈ Zn+ .

B.4 Examples of Basic Optimization Problems

This section provides several examples of optimization problems. The first is a
classic optimization problem known as the transportation problem.
Example B.5. (Transportation problem.) A timber company has three mills
which produces wooden frames and five markets. The three mills can produce
20, 40, and 30 units of wooden frames respectively on a daily basis. The daily
demands for the wooden frames at the five markets are 15, 30, 20, 15, and 10
respectively. The cost to transport the wooden frames from the three mills to the
five markets are shown below:
Mill Location/Market 1 2 3 4 5
1 22 13 24 31 47
2 35 24 11 27 38
3 44 33 25 11 26
Formulate an optimization problem for satisfying demands at the markets while
minimizing the transportation costs from the mills.
Let I and J denote the set of mill and market locations, respectively. The
decision variables in this problem are xij , each of which represents the volume
of wooden frames to be sent from location i ∈ I to location j ∈ J . For example,
B.4. EXAMPLES OF BASIC OPTIMIZATION PROBLEMS 483

x23 corresponds to the volume of wooden frames to be sent from Mill 2 to Market
3.
Let ui represent the amount of wooden frames which can be produced at
mill i ∈ I and dj represent the amount of wooden frames needed at market
j ∈ J . For example, u3 = 30 and d4 = 15.
At each mill, the total volume of wooden frames transported must be less
than or equal to the capacity of the mill. For example, at Mill 2:
x21 + x22 + x23 + x24 + x25 ≤ 40 (B.92)
This can be represented generally as:
X
xij ≤ ui ∀i ∈ I (B.93)
j∈J

Similarly, demand at each market must be met. The demand constraint can
be represented as: X
xij ≥ dj ∀j ∈ J (B.94)
i∈I
In addition, the volume of wooden frames transported between mill and
market locations cannot be negative.
xij ≥ 0 ∀i ∈ I, ∀j ∈ J (B.95)
Let cij represent the cost to transport a wooden frame from mill i ∈ I to
location j ∈ J . For example c24 = 27. The objective is to minimize the total
transportation costs which is:
XX
min cij xij (B.96)
i∈I j∈J

The final formulation for the transportation problem can be summarized as:
XX
min cij xij
xij
i∈I
X j∈J
s.t. xij ≤ ui ∀i ∈ I
j∈J
X
xij ≥ di ∀j ∈ J
i∈I
xij ≥ 0 ∀i ∈ I, ∀j ∈ J

This formulation is an example of a linear program. In a linear program, the

objective functions and constraints are linear functions of the decision variables,
and the solution can be any real number satisfying the constraints (it does not
have to be an integer). Notice that in the above example, we are not restricting
the number of wooden frames to be transported between each pair of mill and
market to be integer. However, the solution to this specific linear program will
always yield integer solutions as long as the input data is integer, although this
is not true in general (see Section C.5.2).
We now modify this problem by considering an additional factor.
484 APPENDIX B. OPTIMIZATION CONCEPTS

Example B.6. The trucks used in transporting wooden frames from Mill 1 are
very old, the frames may be damaged because of their poor suspension systems.
To prevent this, additional packing material is needed, and the amount depends
on the destination market location. The table below shows the amount of packing
material per frame for each mill and market combination (Mills 2 and 3 have
newer trucks that do not require special packaging.)

Mill Location/Market 1 2 3 4 5
1 3 7 3 1 0
2 0 0 0 0 0
3 0 0 0 0 0

If Mill 1 has 21 units of packing material available each day, formulate the
problem of meeting demands while minimizing transportation costs.

Let rij represent the amount of packaging needed to transport a wooden

frame from mill i ∈ I to market j ∈ J . For example, r12 = 7 and r34 = 0. Let
Ri represent the total packing material available at mill i each day, so R1 = 21.
R2 and R3 can be assigned the value zero, because they do not need to have
any packing material available. The linear transportation problem formulation
can be modified by adding an additional resource constraint as shown below:
XX
min cij xij
xij
i∈IX
j∈J
s.t. xij ≤ ui ∀i ∈ I
j∈J
X
xij ≥ di ∀j ∈ J
Xi∈I
rij xij ≤ Ri ∀i ∈ I
j∈J
xij ≥ 0 ∀i ∈ I, ∀j ∈ J

It turns out that adding the packing material constraint changes the opti-
mization problem in such a way that the optimal solutions may not be integers.
If this condition is important, we must enforce it with an additional constraint,
by replacing

xij ≥ 0 ∀i ∈ I, ∀j ∈ J (B.97)
with
xij ∈ Z+ ∀i ∈ I, ∀j ∈ J (B.98)
The formulation now becomes an integer program. The methods needed
to solve integer programs are different than the methods used to solve linear
program, and integer programs are much harder to solve. Sometimes, it may be
adequate to solve the problem as a linear program and then convert its optimal
solution to an integral one, say, by rounding — if the values of the decision
B.4. EXAMPLES OF BASIC OPTIMIZATION PROBLEMS 485

variables are in the hundreds or thousands, the effect of rounding is likely small.
However, for some integer programs this can lead to very poor solutions.
We now return to the original transportation problem formulation of Exam-
ple B.5 without the packing material constraint. In that example, the objective
functions and constraints are all linear functions of the decision variables. In
many real world applications, it might not be possible to use linear functions to
represent the objective function or constraints. The next example modifies the
transportation costs to reflect “diseconomies of scale,” where the unit cost of
shipping increases with the quantity (perhaps the most efficient trucks are used
first, but as more and more frames are shipped, you have to start using older
and less fuel-efficient trucks).
Example B.7. Assume now that the unit cost of transporting wooden frames
between mill i ∈ I and market j ∈ J is cij + xij . (For example, if 10 wooden
frames are being transported between mill i ∈ I and location j ∈ J , then the
cost of transporting each frame between the two locations is cij + 10, and the
total transportation cost would be (cij + 10) × 10). Formulate an optimization
problem to meet the demands at the markets while minimizing transportation
costs.
In this formulation, if xij wooden frames are being transported then the total
transportation cost between mill i ∈ I and market j ∈ J is (cij + xij ) × xij .
The constraints are the same, but the objective function now changes:
XX
min (cij + xij )xij
xij
i∈I
X j∈J
s.t. xij ≤ ui ∀i ∈ I
j∈J
X
xij ≥ di ∀j ∈ J
i∈I
xij ≥ 0 ∀i ∈ I, ∀j ∈ J

In this formulation, the objective function is no longer linear in the decision

variables. The above formulation is an example of a nonlinear program. In a
nonlinear program, either the objective function or any of the constraints is
nonlinear. If we wanted to enforce the decision variables to be integers, we
would need to replace

xij ≥ 0 ∀i ∈ I, ∀j ∈ J (B.99)
with
xij ∈ Z+ ∀i ∈ I, ∀j ∈ J (B.100)
in the final formulation. This leads to a nonlinear integer program, which
again requires different solution methods.
In the linear variant of the transportation problem, we assume a unit cost
of transportation between each mill and market location. For example, the unit
cost of transportation between mill 2 and market 3 is 11. Therefore, the total
486 APPENDIX B. OPTIMIZATION CONCEPTS

cost of transporting 10 wooden frames between mill 2 and market 3 is 110.

The total cost of transporting 50 wooden frames between mill 2 and market
3 is 550. However in the real world, there is often discounts available in bulk
transport. For example, the unit cost of transportation might be 11 up to 30
wooden frames and 5 for the wooden frames above 30. This is called economy
of scale. Therefore, the total cost of transporting 50 wooden frames would be
30 × 11 + 20 × 5 = 430. The presence of economy of scale in transportation
makes the linear variant of the transportation problem an
Integer programs are also used in modeling selections or incorporating “yes
or no” decisions. In such problems often a decision variable which can only take
the values 0 (for “no”) or 1 (for “yes”) is used to model the selection decisions.
This is called a binary variable.

Example B.8. The timber company is now interested in establishing factories

to produce wooden doors. The potential sites for establishing the factories are
the same as the current locations of the timber mills. There is a fixed cost
(which has been amortized to a daily cost) of 300, 350, and 450 associated with
establishing the factories at the three locations. Factories built at each site would
respectively have daily production capacities of 40, 80, and 60 doors. The daily
demands for the wooden doors at the five markets are 35, 15, 50, 20, and 40
respectively. The unit cost of transporting wooden frames is assumed to be the
same as the unit cost of transporting wooden frames which are given in the
linear transportation example. Formulate an optimization problem to meet the
demands at the markets while minimizing the facility location and transportation
costs.

As in Example B.5, let I and J represent the set of potential factory loca-
tions and markets respectively. There are two sets of decision variables. The
first set of decision variables yi takes the value 1 if a facility is located at i ∈ I
and 0 otherwise. For example the values y1 = 1, y2 = 0, and y3 = 1 would mean
that factories are opened at sites 1 and 3. The second set of decision variables
xij represents the volume of demand at market j ∈ J served by a facility at
location i ∈ I.
Let dj represent the demand for wooden doors at market j ∈ J . Therefore
d1 = 35, d2 = 15 and so on. Let ui represent the capacity of the factory at
location i ∈ I. Note that u1 = 40, u2 = 80, u3 = 60.
If a factory is located at i ∈ I, then the total volume of wooden doors
supplied to all markets cannot exceed the production capacity of the factory.
If a factory is not located at i ∈ I, then the total volume of total volume of
wooden doors supplied to all markets must be zero. This constraint can be
represented as
X
xij ≤ ui yi ∀i ∈ I . (B.101)
j∈J

Note that when yi = 1, the right hand side becomes ui (total production at
an open factory cannot exceed its capacity). When yi = 0 the right hand side
B.4. EXAMPLES OF BASIC OPTIMIZATION PROBLEMS 487

becomes zero (nothing can be produced at a factory which was never opened).
Similar to the transportation problem, the total volume supplied to each market
location must meet the demand:

X
xij ≥ di ∀j ∈ J . (B.102)
i∈I

In addition to the non-negativity constraints on the variable xij , there is

also an additional binary restriction on the facility location decision variable yi .

xij ≥ 0 ∀i ∈ I, ∀j ∈ J (B.103)
yi ∈ {0, 1} ∀i ∈ I (B.104)

The objective function has two cost components: the cost of establishing the
facility and the transportation costs. The expression for the transportation costs
is the same as in the linear transportation problem example. Let fi represent
P a facility at location i ∈ I. The total facility location
the cost of establishing
cost can be given as i∈I fi yi . Therefore the objective function is:

X XX
min fi yi + (cij + xij )xij (B.105)
i∈I i∈I j∈J

The final formulation is

X XX
min fi yi + (cij + xij )xij
xij ,yi
i∈I
X i∈I j∈J
s.t. xij ≤ ui yi ∀i ∈ I
j∈J
X
xij ≥ di ∀j ∈ J
i∈I
xij ≥ 0 ∀i ∈ I, ∀j ∈ J
yi ∈ {0, 1} ∀i ∈ I

In this formulation, the facility location decision variable is restricted to be

either 0 or 1 and cannot be any real number. Therefore, the above formulation is
an example of an integer program. Binary variables are also common in project
or portfolio selection problems, which appear often in civil engineering. The
next example shows such a problem.

Example B.9. A business is seeking to invest in two private-public partnership

projects. There is an option investing $400, $200, $300, $150, and $250 in
projects 1 through 5 respectively. The expected and standard deviation of return
on the five projects is shown in the table below.
488 APPENDIX B. OPTIMIZATION CONCEPTS

Project Expected Value Stdev

1 5000 300
2 2000 150
3 3500 125
4 2000 225
5 3750 75
The returns on the five projects have a correlation coefficient of +0.1. For any
investment portfolio, the business balances risk and expected return with the
utility function U = µ − 21 σ where µ is the average return on investment and σ
the standard deviation of return. The business has a total of $1000 to invest.
Formulate an optimization problem for determining the investment projects to
maximize its utility.
Let I represent the set of projects. In this case, let I = {1, 2, 3, 4, 5}.
The decision variables in this problems should reflect which projects are
selected for investment. This can be done using binary decision variables xi
which take the value 1 if project i ∈ I is selected, and 0 if not.
The business has a total budget of $1000. Let B represent the total budget
and bi represent the amount which needs to be invested in each project i ∈ I.
For example b4 = 150 and b5 = 250. The budget constraint can be enforced as:
X
bi xi ≤ B (B.106)
i∈I

Now let us tackle the objective function. Let µi represent the average return
of investment in project i ∈ I. The average return on investment µ is given as:

X
µ = 5000x1 + 2000x2 + 3500x3 + 2000x4 + 3750x5 = µi xi (B.107)
i∈I

Given the standard deviation of returns σi for each i ∈ I and the correlation
coefficient ρij = 0.1 for all i ∈ I and j ∈ I. The standard deviation of return is
given as:

p
σ= 3002 x1 + . . . + 2252 x4 + 752 x5 + 0.1 × 300 × 150 + . . .
sX X X
= σi2 xi + ρij σi σj xi xj (B.108)
i∈I i∈I j∈I:j6=i

The optimization formulation is

s
X 1 X 2 X X
min µi xi − σi xi + ρij σi σj xi xj
xi 2
i∈I i∈I i∈I j∈I:j6=i
X
s.t. bi xi ≤ B
i∈I
xi = {0, 1} ∀i ∈ I
B.5. MORE EXAMPLES OF OPTIMIZATION FORMULATIONS 489

The above formulation is a nonlinear integer program. The nonlinearity

arises from two sources: both the square root and the xi xj interaction terms in
the objective function. The integrality arises from the binary decision variables.

B.5 More Examples of Optimization Formula-

tions
The previous section was a tour of common types of optimization problems:
linear programs, integer programs, and nonlinear programs. This section pro-
vides additional examples of how optimization problems might be formulated in
a transportation engineering setting. As you read through these, think about
how they relate to the concepts defined in the previous section (which ones are
linear, integer, etc.).

Example B.10. (Transit frequency setting.) You are working for a public
transit agency in a city, and must decide the frequency of service on each of the
bus routes. The bus routes are known and cannot change, but you can change
how the city’s bus fleet is allocated to each of these routes. (The more buses
assigned to a route, the higher the frequency of service.) Knowing the ridership
on each route, how should buses be allocated to routes to minimize the total
waiting time?

Solution. To formulate this as an optimization problem, we need to

identify an objective, decision variables, and constraints. For this problem,
as we do this we will be faced with other assumptions which must be made.
With this and other such problems, there may be more than one way to write
down an optimization problem, what matters is that we clearly state all of the
assumptions made. Another good guiding principle is to start with the simplest
model which captures the important behavior, which can then be refined by
relaxing assumptions or replacing simple assumptions with more realistic ones.
We need some notation, so let R be the set of bus routes. We know the
current ridership on each route dr . Here we will make our first assumption:
that the demand on each route is inelastic and will not change based on the
service frequency. Like all assumptions, it is not entirely true but in some cases
may be close enough to the truth to get useful results; if not, you should think
about what you would need to replace this assumption, which is always a good
exercise. We are also given the current fleet size (which we will denote N ), and
must choose the number of buses associated with each route (call this nr ) —
thus the decision variables in this problem are the nr values.
The objective is to minimize the total waiting time, which is the sum of the
waiting time for the passengers on each route. How long must passengers wait
for a bus? If we assume that travelers arrive at a uniform rate, then the average
waiting time will be half of the service headway. How is the headway related
to the number of buses on the route nr ? Assuming that the buses are evenly
490 APPENDIX B. OPTIMIZATION CONCEPTS

dispersed throughout the time period we are modeling, and assuming that each
bus is always in use, then the headway on route r will be the time required
to traverse this route (Tr ) divided by the number of buses nr assigned to this
route. So, the average delay per passenger is half of the headway, or Tr /(2nr ),
and the total passenger delay on this route is (dr Tr )/(2nr ). This leads us to the
objective function
X dr Tr
D(n) = (B.109)
2nr
r∈R

in which the total delay is calculated by summing the delay associated with each
route.
What constraints do we have? Surely we must run at least one bus on each
route (or else we would essentially be canceling a route), and in reality as a
matter of policy there may be some lower limit on the number of buses assigned
to each route; for route r, call this lower bound Lr . Likewise, there is some
upper bound Ur on the number of buses assigned to each route as well. So, we
can introduce the constraint Lr ≤ nr ≤ Ur for each route r.
Putting all of these together, we have the optimization problem

min D(n) = r∈R d2n r Tr

P
n r

s.t. nr ≥ Lr ∀r ∈ R
P nr ≤ Ur ∀r ∈ R
r∈R r ≤ N
n

Example B.11. (Scheduling maintenance.) You are responsible for schedul-
ing routine maintenance on a set of transportation facilities (such as pavement
sections or bridges.) The state of these facilities can be described by a condi-
tion index which ranges from 0 to 100. Each facility deteriorates at a known,
constant rate (which may differ between facilities). If you perform maintenance
during a given year, its condition will improve by a known amount. Given an
annual budget for your agency, when and where you should perform mainte-
nance to maximize the average condition of these facilities? You have a 10 year
planning horizon.
Solution. In contrast to the previous example, where the three compo-
nents of the optimization problem were described independently, from here on
problems will be formulated in a more organic way, describing a model built
from the ground up. (This is how optimization models are usually described in
practice.) After describing the model in this way, we will identify the objective
function, decision variables, and constraints to write the optimization problem
in the usual form. We start by introducing notation based on the problem
statement.
Let F be the set of facilities, and let ctf be the condition of facility f at the
end3 of year t, where t ranges from 1 to 10. Let df be the annual deterioration on
3 This word is intentionally emphasized. In this kind of problem it is very easy to get
B.5. MORE EXAMPLES OF OPTIMIZATION FORMULATIONS 491

facility f , and if the amount by which the condition will improve if maintenance
is performed. So, if no maintenance is performed during year t, then

ctf = ct−1
f − df ∀f ∈ F, t ∈ {1, 2, . . . , 10} (B.110)

and if maintenance is performed we have

ctf = ct−1
f − df + if ∀f ∈ F, t ∈ {1, 2, . . . , 10} (B.111)

Both of these cases can be captured in one equation with the following trick:
let xtf equal one if maintenance is performed on facility f during year t, and 0
if not. Then

ctf = ct−1
f − df + xtf if ∀f ∈ F, t ∈ {1, 2, . . . , 10} (B.112)

Finally, the condition can never exceed 100 or fall below 0, so the full equation
for the evolution of the state is

100
 if ct−1
f − df + xtf if > 100
ctf = 0 if ct−1
f − df + xtf if < 0 (B.113)
 t−1
 t
cf − df + xf if otherwise

for all f ∈ F and t ∈ {1, 2, . . . , 10}. Of course, for this to be usable we need to
know the current conditions of the facilities, c0f .
The annual budget can be represented this way: let kf be the cost of per-
forming maintenance on facility f , and B t the budget available in year t. Then
X
kf xtf ≤ B t ∀t ∈ {1, . . . 10} . (B.114)
f ∈F

For the objective, we needPthe average

P10 t condition across all facilities and all
1
years; this is simply 10|F | f ∈F t=1 cf . The obvious decision variables are
the maintenance variables xtf , but we also have to include ctf because these are
influenced by the maintenance variables. As constraints, we need to include
the state evolution equations (B.113), the budget constraints (B.114), and, less
obviously the requirement that xtf be either 0 or 1. Putting it all together, we
have the optimization problem
1
P P10 t
max 10|F | f ∈F t=1 cf
x,c
t t
P
s.t. f ∈F kf xf ≤ B ∀t ∈ {1, . . . 10}
ctf is given by (B.113) ∀f ∈ F, t ∈ {1, 2, . . . , 10}
xtf ∈ {0, 1} ∀f ∈ F, t ∈ {1, 2, . . . , 10}


confused about what occurs at the start of period t, at the end of period t, during the middle
of period t, etc.
492 APPENDIX B. OPTIMIZATION CONCEPTS

1 2 3 ...

Figure B.1: Coordinates and labeling of intersections for Example B.12

In this example, pay close attention to the use of formulas like (B.114), which
show up very frequently in optimization. It is important to make sure that every
“index” variable in the formula is accounted for in some way. Equation (B.114)
involves the variables xtf , but for which facilities f and time periods t? The
facility index f is summed over, while the time index t is shown at right as
∀t ∈ {1, . . . 10}. This means that a copy of (B.114) exists for each time period,
and in each of these copies, the left-hand side involves a sum over all facilities at
that time. Therefore the one line (B.114) actually includes 10 constraints, one
for each year. It is common to forget to include things like ∀t ∈ {1, . . . 10}, or to
try to useP an index of summation outside of the sum as well (e.g., an expression
like Bf − f ∈F kf xtf ), which is meaningless. Make sure that all of your indices
are properly accounted for!
In this next example, the objective function is less obvious.

Example B.12. (A facility location problem.) In a city with a grid network,

you need to decide where to locate three bus terminals. Building the terminal
at different locations costs a different amount of money. Knowing the home
locations of customers throughout the city who want to use the bus service, where
should the terminals be located to minimize the construction cost and walking
distance customers have to walk? Assume that each customer will walk from
their home location to the nearest terminal.

Solution. This problem could easily become very complicated if we take

into account the impact of terminal locations on bus routes, so let’s focus on
what the problem is asking for: simply locating terminals to minimize walking
distance from customers’ home locations.
Number each of the intersections in the grid network (Figure B.1) from 1 to
I, the total number of intersections. Assume that the terminals will be located
at these intersections, and let the variables L1 , L2 , and L3 denote the numbers of
the intersections where terminals will be built. Let C(i) be the cost of building a
terminal at location i, so the total cost of construction is C(L1 )+C(L2 )+C(L3 ).
B.5. MORE EXAMPLES OF OPTIMIZATION FORMULATIONS 493

Let P be the set of customers, and let Hp denote the intersection that is the
home location of customer p.
How can we calculate the walking distance between two intersections (say,
i and j)? Figure B.1 shows a coordinate system superimposed on the grid.
Let x(i) and y(i) be the coordinates of intersection i in this system. Then the
walking distance between points i and j is

d(i, j) = |x(i) − x(j)| + |y(i) − y(j)| . (B.115)

This is often called the Manhattan distance between two points, after one of the
densest grid networks in the world.
So what is the walking distance D(p) for customer p? The distance from
p to the first terminal is d(Hp , L1 ), to the second terminal is d(Hp , L2 ), and
to the third is d(Hp , L3 ). The passenger will walk to whichever is closest, so
D(p, L1 , L2 ,P L3 ) = min{d(Hp , L1 ), d(Hp , L2 ), d(Hp , L3 )} and the total walking
distance is p∈P D(p, L1 , L2 , L3 ).
For this problem, the decision variables and constraints are straightforward:
the only decision variables are L1 , L2 , and L3 and the only constraint is that
these need to be integers between 1 and I. The tricky part is the objective
function: we are instructed both to minimize total cost as well as total walking
distance. We have equations for each of these, but we can only have one objective
function. In these cases, it is common to form a convex combination of the two
objectives, introducing a weighting parameter λ ∈ [0, 1]. That is, let
 
X
f (L1 , L2 , L3 ) = λ[C(L1 ) + C(L2 ) + C(L3 )] + (1 − λ)  D(p, L1 , L2 , L3 )
p∈P
(B.116)
Look at what happens as λ varies. If λ = 1, then the objective function
reduces to simply minimizing the cost of construction. If λ = 0, the objective
function is simply minimizing the total walking distance. For a value in between
0 and 1, the objective function is a weighted combination of these two objectives,
where λ indicates how important the cost of construction is relative to the
walking distance.
For concreteness, the optimization problem is
hP i
min λ[C(L1 ) + C(L2 ) + C(L3 )] + (1 − λ) p∈P D(p, L1 , L2 , L3 )
L1 ,L2 ,L3
s.t. Lf ∈ {1, 2, . . . , I} ∀f ∈ {1, 2, 3}


In the final example in this section, finding a mathematical representation
of a solution is more challenging.

Example B.13. (Shortest path problem.) Figure B.2 shows a road network,
with the origin and destination marked. Given the travel time on each roadway
link, what is the fastest route connecting the origin to the destination?
494 APPENDIX B. OPTIMIZATION CONCEPTS

Origin

Destination

Figure B.2: Roadway network for Example B.13

Solution. We’ve already presented some algorithms for solving this prob-
lem in Section 2.4, but here we show how it can be placed into the general
framework of optimization problems.
Notation first: let the set of nodes be N , and let r and s represent the origin
and destination nodes. Let the set of links be A, and let tij be the travel time
on link (i, j). So far, so good, but how do we represent a route connecting two
intersections?
Following Example B.11, introduce binary variables xij ∈ {0, 1}, where xij =
1 if link (i, j) is part of the route, and xij = 0 if link (i, j) is not part of the
route. The travel time of a P route is simply the sums of the travel times of the
links in the route, which is (i,j)∈A tij xij .
We now have an objective function and decision variables, but what of the
constraints? Besides the trivial constraint xij ∈ {0, 1}, we need constraints
which require that the xij values actually form a contiguous path which starts
at the origin r and ends at the destination s. We do this by introducing a flow
conservation constraint at each intersection. For node i, recall that Γ(i) denotes
the set of links which leave node i, and Γ−1 (i) denotes the set of links which
enter node i.
Consider any contiguous path connecting intersection r and s, and examine
any node i. One of four cases must hold:

Case I: Node i does not lie on the path at all. Then all of the xij values
should be zero for (i, j) ∈ Γ(i) ∪ Γ−1 (i).

Case II: Node i lies on the path, but is not the origin or destination. Then
xij = 1 for exactly one (i, j) ∈ Γ(i), and for exactly one (i, j) ∈ Γ−1 (i).

Case III: Node i is the origin r. Then all xij values should be zero for (i, j) ∈
Γ−1 (i), and xij = 1 for exactly one (i, j) ∈ Γ(i).

Case IV: Node i is the destination s. Then all xij values should be zero for
(i, j) ∈ Γ(i), and xij = 1 for exactly one (i, j) ∈ Γ−1 (i).
B.6. GENERAL PROPERTIES OF OPTIMIZATION PROBLEMS 495

P
An elegant way to combine these cases is to look at the differences (i,j)∈Γ(i) xij −
P
(h,i)∈Γ−1 (i) xhi . For cases I and II, this difference will be 0; for case III, the
difference will be +1, and for case IV, the difference will be −1.
So, this leads us to the optimization formulation of the shortest path prob-
lem:
P
min (i,j)∈A tij xij
x 
P P 1
 if i = r
s.t. (i,j)∈Γ(i) x ij − −1
(h,i)∈Γ (i) x hi = −1 if i=s ∀i ∈ N

0 otherwise

xij ∈ {0, 1} ∀(i, j) ∈ A
(B.117)

A careful reader will notice that if the four cases are satisfied for a solution,
then the equations (B.117) are satisfied, but the reverse may not be true. Can
you see why, and is that a problem?
There is often more than one way to formulate a problem: for instance,
we might choose to minimize congestion by spending money on capacity im-
provements, subject to a budget constraint. Or, we might try to minimize the
amount of money spent, subject to a maximum acceptable limit on congestion.
Choosing the “correct” formulation in this case may be based on which of the
two constraints is harder to adjust (is the budget constraint more fixed, or the
upper limit on congestion?). Still, you may be troubled by this seeming impre-
cision. This is one way in which modeling is “as much art as science,” which
is not surprising — since the human and political decision-making processes
optimization tries to formalize, along with the inherent value judgments, (what
truly is the objective?) are not as precise as they seem on the surface. One
hallmark of a mature practitioner of optimization is a flexibility with different
formulations of the same underlying problem, and a willingness to engage in
“back-and-forth” with the decision maker as together you identify the best for-
mulation for a particular scenario. In fact, in some cases it may not matter.
The theory of duality (which is beyond the scope of this book) shows that these
alternate formulations sometimes lead to the same ultimate decision, which is
comforting.

B.6 General Properties of Optimization Prob-

lems
After seeing some concrete examples of optimization formulations, we now define
a general form for optimization problems. This form is useful because it allows
us to talk about properties of optimization problems, or methods for solving
them, in a way that is independent of any specific context (shipping, investment,
routing, etc.). Any of the optimization problems discussed in this book can be
represented in this form, perhaps after some conversions or transformation.
496 APPENDIX B. OPTIMIZATION CONCEPTS

The general form is

min f (x)
x
s.t. gi (x) ≥ bi ∀i = 1, 2, .., m
hi (x) = 0 ∀i = 1, 2, .., l

If any variables are restricted to be integers, those constraints are listed in

addition to the ones given above.
In the above problem x = (x1 , x2 , ..., xn ) is a vector of decision variables.
Decision variables are factors which are under the control of the decision maker.
The function f (x) is the objective function which represents the cost or utility
associated with a decision x. The equations gi (x) ≥ bi and hi (x) = 0 are
constraint functions which impose restrictions on the range of values the decision
variables can take. Let X = {x ∈ Rn : gi (x) ≥ bi ∀i = 1, 2, ..m; hi (x) =
0 ∀i = 1, 2, ..l}. The set X is called the feasible region which represents the
set of values which the decision variables can take or the set of values which
satisfies all constraint functions. Any value x ∈ X is a feasible solution to the
optimization problem. The goal of the mathematical problem is to determine
the optimal solution which is defined as the feasible solution with the smallest
objective function value, i.e, to find x∗ ∈ X such that f (x∗ ) ≤ f (x) ∀x ∈ X.
Depending on the mathematical properties of the objective and constraint
functions there are different categories of optimization problems. Linear opti-
mization problems are mathematical programs where the functions f (x), gi (x) ∀i =
1, 2, ..m, hi (x) ∀i = 1, 2, ..m are linear. If any of the objective or constraint
functions are nonlinear, then we have nonlinear optimization problems. In-
teger programs are mathematical programs where the decision variables have
restricted to a set of integers. Depending on the nature of the mathematical
program, there are different types of solution algorithms with varying degrees
of efficiency.

B.6.1 Local and global solutions

Objective functions can have both local and global maximum or minimum so-
lutions (Figure B.3). Let f be a function defined on a region X ⊆ Rn . Recall
from Section A.3 that a ball centered on x of radius  is the set containing all
points of R whose distance to x is less than .
Definition B.1. A point x∗ ∈ X is a local minimum of a function f (x) if there
is a ball B centered on x with positive radius such that

f (x) ≥ f (x∗ ) ∀x ∈ B ∩ X

Definition B.2. A point x∗ ∈ X is a global minimum of a function f (x) if

f (x) ≥ f (x∗ ) ∀x ∈ X

Strict local and global minima are obtained by replacing the inequalities in
the above definitions with strict inequalities:
B.6. GENERAL PROPERTIES OF OPTIMIZATION PROBLEMS 497

Figure B.3: Local, strict local, and global minima.

Definition B.3. A point x∗ ∈ X is a strict local minimum of a function f (x)

if there is a ball B centered on x with positive radius such that
f (x) > f (x∗ ) ∀x ∈ B ∩ X, x 6= x∗ )
Definition B.4. A point x∗ ∈ X is a global minimum of a function f (x) if
f (x) > f (x∗ ) ∀x ∈ X, x 6= x∗
The corresponding terms for maximization problems are defined in analogous
ways, by reversing the signs of the inequalities.

B.6.2 Objective function transformations

The following results are very useful in showing cases when two superficially
different optimization problems may in fact be the same. The first result shows
that it is easy to convert any maximization problem to a minimization prob-
lem (or vice versa) by negating the objective function. The second shows that
constants may be freely added or subtracted to objective functions without
changing the optimal solutions. The third shows that multiplication or division
by a positive constant does not change the optimal solutions.
Proposition B.1. If the feasible set is X, the solution x̂ is a global maximum
of f if and only if x̂ is a global minimum of −f for the same feasible set X.
Proof. For the first part, assume that x̂ is the global maximum of f on the set
X. Then f (x̂) ≥ f (x) for all x ∈ X. Multiplying both sides by −1 reverses
the sign of the inequality, giving −f (x̂) ≤ −f (x) for all X, so x̂ is a global
minimum for −f . For the second part, assume that x̂ is the global minimum of
−f , so −f (x̂) ≤ −f (x) for all x ∈ X. Again multiplying both sides by −1 gives
f (x̂) ≥ f (x) for all x ∈ X, so x̂ is a global maximum of f .
498 APPENDIX B. OPTIMIZATION CONCEPTS

This result is convenient, because we do not need to develop two different

theories for maximization and minimization problems. Instead, we can focus
just on one. In this book, we develop results for minimization problems, a fairly
common convention in engineering.4 Whenever you encounter a maximization
problem, you can convert it to a minimization problem by negating the objective
function, then using the minimization results and procedures. (Of course, this
choice is completely arbitrary; we could just as well have chosen to develop
results only for maximization problems and in fact some other fields do just
this.)

Proposition B.2. Let x̂ be an optimal solution when the objective function is

f (x) and the feasible set is X. Then for any constant b not depending on x,
x̂ is also an optimal solution when the objective function is f (x) + b and the
feasible set is X.

Proof. By Proposition B.1, with no loss of generality we can assume that the
optimization problem is a minimization problem and that x̂ is a global minimum.
Therefore f (x̂) ≤ f (x) for all x ∈ X. We can add the constant b to both sides
of the equation, so f (x̂) + b ≤ f (x) + b for all x ∈ X. Therefore x̂ is also a
global minimum (and thus optimal) when the objective function is f (x) + b.

Proposition B.3. Let x̂ be an optimal solution when the objective function is

f (x) and the feasible set is X. Then for any constant c > 0 not depending on
x, x̂ is also an optimal solution when the objective function is cf (x) and the
feasible set is X.

Proof. See Exercise 6.

B.6.3 Existence of solutions

It is possible to formulate optimization problems that do not have solutions
— either because the constraints are contradictory and there is no feasible so-
lution satisfying all of them, or because of the way the objective function and
constraints are defined. Optimization problems of the first type are called infea-
sible. For example, the problem “minimize x subject to x ≤ 0” has no optimal
solution — whatever x you choose, you can find an x which is even more neg-
ative. Optimization problems of this type, where the objective can be made as
negative as you like, are called unbounded. Some problems lack solutions even
without falling into either of these categories; for example, “minimize x subject
to x > 0” has no optimal solution, for whatever x you choose, x/2 is feasible but
smaller. This is why we tend to avoid strict inequalities in optimization prob-
lems. But there are still problematic optimization formulations even without
strict inequalities, such as “minimize 1/x subject to x ≥ 1.”
4 Economists often use maximization problems as their standard convention.
B.6. GENERAL PROPERTIES OF OPTIMIZATION PROBLEMS 499

Practically speaking, if you are in a situation like the ones described above,
you should take another look at your optimization problem: perhaps one con-
straint can be relaxed (maybe penalized in the objective, rather than strictly
excluding solutions), or perhaps you missed a constraint that should have been
there. Systems with unbounded or truly infeasible solutions are rare. This
subsection provides a mathematical perspective on the topic, giving conditions
on the constraints and objective function which can guarantee existence of a
solution.
Keeping with our standard convention, we consider a problem of the form
“minimize f (x) subject to x ∈ X,” where the set X contains all solutions satis-
fying all constraints. Weierstrass’ Theorem identifies gives sufficient conditions,
under which an optimal solution exists.
Theorem B.1. (Weierstrass’ Theorem.) Let f be a continuous, real-valued
function defined on X, and let X be a non-empty, closed, and bounded set. The
optimization problem {min f (x) : x ∈ X} has a minimum solution.
(For definitions of the terms in this theorem, see Appendix A.)
Example B.14. Consider the optimization formulation min(x − 5)2 , x ∈ X =
{x : x ∈ R, 3 ≤ x ≤ 7}. Does a minimum exist?
Solution. The function (x−5)2 is continuous and real valued. The feasible
region is non-empty, closed, and bounded. Therefore, the above optimization
problem has a minimum. By plotting the function, we can see that the minimum
occurs at x = 5. 
Example B.15. Consider the optimization formulation min(x − 5)2 , x ∈ X =
{x : x ∈ R, 5 < x ≤ 7}. Does a minimum exist?
Solution. In the above variant, the feasible region is not closed. Given any
y ∈ X, we will be able to find another z ∈ X, such that f (z) < f (y). For
example, if we pick y = 5.001, we can always find another z = 5.0001 whose
objective function value is smaller. Therefore, no minimum exists in this case.

Note that Weierstrass’ Theorem provides sufficient and not necessary con-
ditions (more on this in next subsection). The continuity, closed, bounded
assumption are not required for minima to exist. We can have situations where
minima exist without the above conditions being satisfied.
Example B.16. Consider the optimization formulation min(x − 5)2 , x ∈ X =
{x : x ∈ R, x > 3}. Does a minimum exist?
Solution. In the above optimization problem, the feasible region is not
closed or bounded. However, we do know that the minimum exists at x = 5. 
Example B.17. Consider the objective function
(
1 x=0
f (x) = .
5 otherwise
The feasible region is X = {x : x ∈ R}. Does a minimum exist?
500 APPENDIX B. OPTIMIZATION CONCEPTS

Solution. In this example, the function is discontinuous, the feasible region

is not closed or bounded. However, we still have a minimum at x = 0. 

B.6.4 Necessary and sufficient conditions

The previous subsection presented Weierstrass’ Theorem, which gave sufficient
conditions for the existence of a minimum solution. Some students are unfa-
miliar with the distinction between “necessary” and “sufficient” conditions, so
we explain them here in the context of this theorem. Let us consider two state-
ments A and B. Saying that A is a necessary condition for B means that B is
true only if A is true, written B =⇒ A (read “B implies A”). For example,
every square is also a rectangle. This means that a shape can be a square only
if it is a rectangle; being a rectangle is necessary for being a square, so we can
identify A with “the shape is a rectangle,” and B with “the shape is a square.”
Notice that A being true does not imply that B is true; there are rectangles
which are not squares.
In the above situation, we can also say that B is a sufficient condition for A,
meaning that if A is true, B must also be true. For example, “being a square”
is sufficient for “being a rectangle.” In a statement like B =⇒ A, the arrow
points from the sufficient condition to the necessary one. When B is a sufficient
condition for A, A can be true without B being true.
Used alone, the terms “necessary” or “sufficient” mean that A and B are
not exactly the same thing, but that one includes cases that the other does not
have. For example, there are some rectangles which are not squares. If A and
B do mean exactly the same thing, we say that A is necessary and sufficient
for B; A is true if and only if (abbreviated “iff”) B is true. This is written
A ⇐⇒ B, and means that both A =⇒ B and B =⇒ A are true. An
example are the statements “n is an even integer” and “n = 2k for some integer
k.” Each statement is true exactly when the other is true; in other words, they
are equivalent and we can replace one statement by the other whenever we like.
In Weierstrass’ Theorem, statement A is “the function f is continuous and
real-valued, and X is non-empty, closed, and bounded,” and statement B is
“the function has a minimum value.” Condition A is sufficient, because any
time A is true, B is true as well. However, condition A is not necessary, because
there are examples where A is false but B is still true (Examples B.14 and B.15
above).

B.7 Exercises
1. [11] Why is it generally a bad idea to use strict inequalities (‘>’, ‘<’) in
mathematical programs?

2. [12] Mathematically specify an optimization problem which has no op-

timal solution, and an optimization problem which has multiple optimal
solutions.
B.7. EXERCISES 501

Table B.1: Data for Exercise 7

Bridge Initial condition Deterioration rate Repairability Maintenance cost
f c0f df if kf
1 85 5 3 2
2 95 10 1 4
3 75 1 1 3
4 80 6 6 1
5 100 3 3 1

3. [24] For the following functions, identify all stationary points and global
optima.

(a) f (x) = x4 − 2x3

(b) f (x1 , x2 ) = 2x21 + x22 − x1 x2 − 7x2
(c) f (x1 , x2 ) = x61 + x62 − 3x21 x22
(d) f (x1 , x2 , x3 ) = (x1 − 4)2 + (x2 − 2)2 + x23 + x1 x2 + x1 x3 + x2 x3

4. [42] Express the constraint “x ∈ Z” (that is, x must be an integer) in the

form g(x) = 0 for some continuous function g. (This shows that nonlinear
programming can’t be any easier than integer programming!)

5. [30] Can you ever improve the value of the objective function at the opti-
mal solution by adding constraints to a problem? If yes, give an example.
If no, explain why not.

6. [21] Prove Proposition B.3.

7. [27] Write out the objective and all of the constraints for the maintenance
scheduling problem of Example B.11, using the bridge data shown in Ta-
ble B.1, where the maintenance cost is expressed in millions of dollars.
Assume a two-year time horizon and an annual budget of $5 million.

8. [44] Reformulate the transit frequency-setting problem of Example B.10

so that the objective is to minimize the total amount of money spent,
while achieving a pre-specified level of service (maximum delays).

9. [53] You are responsible for allocating bridge maintenance funding for a
state, and must develop optimization models to assist with this for the
current year. Political realities and a strict budget will require you to
develop multiple formulations for the purposes of comparison.
Let B denote the set of bridges in the state. There is an “economic value”
associated with the condition of each bridge: the higher the condition, the
higher the value to the state, because bridges in worse condition require
more routine maintenance, impose higher costs on drivers who must drive
slower or put up with potholes, and carry a higher risk of unforeseen fail-
ures requiring emergency maintenance. To represent this, there is a value
502 APPENDIX B. OPTIMIZATION CONCEPTS

function Vb (xb ) associated with each bridge b ∈ B, giving the economic

value of this bridge if xb is spent on maintenance this year. You have a
total budget of X to spend statewide. The state is also divided into n
districts, and each bridge belongs to one district. Let Di denote the set of
bridges which belong to district i; different districts may contain different
numbers of bridges.
In the following, be sure to define any additional notation you introduce.
There is more than one possible formulation that achieves the stated goals,
so feel free to explain any parts of your formulations which may not be
self-evident.

(a) Maximize average benefits: Formulate an optimization model

(objective function, decision variables, and constraints) to maximize
the average economic value of bridges across the state.
(b) Equitable allocation: The previous model may recommend spend-
ing much more money in some districts than others, which is politically
infeasible. Formulate an optimization model based on maximizing the
total economic value to the state, but ensuring that each district re-
ceives a relatively fair share of the total budget X.
(c) Equitable benefits: There is a difference between a fair alloca-
tion of money, and a fair allocation of benefits (say, if the functions
Vb differ greatly between districts). Defining benefit as the difference
between the economic value after investing in maintenance, and the
“do-nothing” economic value, formulate an optimization model which
aims to achieve a relatively fair distribution of benefits among dis-
tricts.

10. [34] Consider the mathematical program

max 3x1 + 5x2 + 6x3

s.t x1 + x2 + x3 = 1
2x1 + 3x2 + 2x3 ≤ 5
x2 + x3 ≤ 1/2
x1 , x2 , x3 ≥ 0

(a) Are all of the constraints needed?

(b) Solve this problem graphically.

11. [57] In an effort to fight rising maintenance costs and congestion, a state
transportation agency is considering a toll on a certain freeway segment
during the peak period. Imposing a toll accomplishes two objectives at
once: it raises money for the state, and also reduces congestion by making
some people switch to less congested routes, travel earlier or later, carpool,
take the bus, and so forth. Suppose that there are 10000 people who would
B.7. EXERCISES 503

want to drive on the freeway if there was no toll and no congestion, but
the number of people who actually do is given by

x = 10000e(15−6τ −t)/500

where τ is the roadway toll (in dollars) and t is the travel time (in minutes).
(That is, the higher the toll, or the higher the travel time, the fewer people
drive.) The travel time, in turn, is given by
  x 4 
t = 15 1 + 0.15
c
minutes, where c = 8000 veh/hr is the roadway capacity during rush
hour. Regulations prohibit tolls exceeding $10 at any time. Citizens are
unhappy with both congestion and having to pay tolls. After conducting
multiple surveys, the agency has determined that citizen satisfaction can
be quantified as
s = 100 − t − τ /5
(a) Formulate two nonlinear programs based on this information, where
the objectives are to either (a) maximize revenue or (b) maximize
citizen satisfaction.
(b) Solve both of these problems, reporting the optimal values for the
decision variables and the objective function.
(c) Comment on the two solutions. (e.g., do they give a similar toll value,
or very different ones? How much revenue does the state give up in
order to maximize satisfaction?)
(d) Name at least two assumptions that went into formulating this prob-
lem. Do you think they are realistic? Pick one of them, and explain
how the problem might be changed to eliminate that assumption and
make it more realistic.
12. [59] You are asked to design a traffic signal timing at the intersection of
8th & Grand. Assuming a simple two-phase cycle (where Grand Avenue
moves in phase 1, and 8th Street in phase 2), no lost time when the
signal changes, and ignoring turning movements, the total delay at the
intersection can be written as
λ1 (c − g1 )2 λ (c − g2 )2
  + 2 
2c 1 − µλ11 2c 1 − µλ22

where g1 and g2 are the effective green time allotted to Grand Avenue and
8th Street, c = g1 + g2 is the cycle length, λ1 and µ1 are the arrival rate
and saturation flow for Grand Avenue, and λ2 and µ2 are the arrival rate
and saturation flow for 8th Street.
All signals downtown are given a sixty-second cycle length to foster good
progression, and the arrival rate and saturation flow are 2200 veh/hr and
504 APPENDIX B. OPTIMIZATION CONCEPTS

3600 veh/hr for Grand Avenue, respectively, and 300 veh/hr and 1900
veh/hr for 8th Street. Furthermore, no queues can remain at the end of
the green interval; this means that µi gi must be at least as large as λi c
for each approach i.
(a) Why does the constraint µi gi ≥ λi c imply that no queues will remain
after a green interval?
(b) Formulate a nonlinear program to minimize total delay.
(c) Simplify the nonlinear program so there is only one decision variable,
and solve this nonlinear program using the bisection method of Sec-
tion 3.3.2 (terminate when b − a ≤ 1).
(d) Write code to automate this process, and perform a sensitivity analysis
by plotting the effective green time on Grand Ave as λ1 varies from
500 veh/hr to 3000 veh/hr, in increments of 500 veh/hr. Interpret
your plot.
(e) Identify two assumptions in the above model. Pick one of them, and
describe how you would change your model to relax that assumption.
13. [26] Find the global minima of the following functions in two ways: the
bisection method, and using Newton’s method to directly find a stationary
point (if the Newton step leaves the feasible region, move to the boundary
point closest to where Newton’s method would go). Run each method
for five iterations, and see which is closer to the optimum, making the
comparison based on the value of the objective function at the final points.
(a) f (x) = − arctan x, x ∈ [0, 10]
(b) f (x) = x sin(1/(100x)), x ∈ [0.015, 0.04]
(c) f (x) = x3 , x ∈ [5, 15]

14. [51] Give a network and a feasible solution x to the mathematical formu-
lation of the shortest path problem in Example B.13 where the links with
xij = 1 do not form a contiguous path between r and s, as alluded to at
the end of the example.
Appendix C

Optimization Techniques

This appendix contains additional information on optimization problems, and

specific techniques for solving them. In turn, we present solution methods
for one-dimensional optimization, linear optimization problems, unconstrained
nonlinear optimization, constrained nonlinear optimization, and optimization
problems with integrality constraints. We close with an introduction to meta-
heuristics, general-purpose techniques which can be applied to any optimization
problem, but without guarantees of finding an optimal solution.

C.1 More Line Search Algorithms for One-Dimensional

Optimization
Section 3.3.2 presented the bisection method as a way of solving optimization
problems with a single decision variablex; a convex, differentiable objective
function f ; and a single, “interval” constraint of the form a ≤ x ≤ b. (Actually,
we can relax the assumption that f is convex to f being unimodal, that is,
if every local minimum is also a global minimum. Convexity is sufficient, but
not necessary, for the function to be unimodal.) Most optimization problems
have more than one variable, but solution algorithms often rely on simpler
“subproblems” which are often one-dimensional. The Frank-Wolfe method from
Section 6.2.2 is an example of this.
This section presents two other line search methods that can be applied
to similar problems. The golden section method requires more iterations than
bisection, but does not requiring taking the derivative of f . It is useful if f
is not differentiable, or if the derivative is cumbersome to evaluate. Newton’s
method, on the other hand, requires the stronger assumption that f is twice
differentiable (and requires calculating both first and second derivatives), but
usually requires fewer iterations than bisection.

505
506 APPENDIX C. OPTIMIZATION TECHNIQUES

C.1.1 Golden section method

The golden section method does not use derivatives to narrow down where the
minimum lies. Instead, the logic is as follows: assume we know the value of
the function at two intermediate points (call them c and d), as well as at the
endpoints a and b. These points are in the order a < c < d < b. If f is
unimodal, then there are only two possibilities: either f (a) ≥ f (c) ≥ f (d), or
f (c) ≤ f (d) ≤ f (b). The reason is that a unimodal function must decrease
before its minimum is reached, and increase afterwards. So if f (c) ≥ f (d), then
the minimum must be to the right of c, and the function must be decreasing
everywhere to the left of c, so f (a) ≥ f (c) ≥ f (d). If f (c) ≤ f (d), then using a
similar argument we know that f (c) ≤ f (d) ≤ f (b). In the first case, we know
that the minimum is somewhere in the interval [c, b], and in the second case we
know it is somewhere in the interval [a, d]. In either case, we have reduced the
width of the interval where the minimum can lie, and we iterate by choosing
two more intermediate points in the reduced interval, on till convergence.
This rule works no matter how the two intermediate points are chosen. To
make the method as efficient as possible, we want to re-use points already found
if possible. For instance, in the first case above, the interval is reduced to [c, b].
The point d is within this interval, and can serve as one of the two points we use
at the next iteration. We also want to reduce the interval as much as possible
from one iteration to the next. Since we don’t know in advance which end we
will trim from, c should be as far from the left endpoint as d is from the right,
that is, c − a = b − d. To be able to re-use the point d in the next iteration,
we want d to be the same fraction of the distance from c to b as c is from a to
b. (Figure C.1). The same logic applies if we are in the second case and must
eliminate the upper end.
Using these principles
√ to solve for c and d, we see that c should be to the
right of a by (3 − 5)/2 ≈ 38.2% of the √ total length of the interval [a, b], and
that d should be to the right of a by ( 5 − 1)/2 ≈ 61.8% of the length of [a, b].
Both of our principles are satisfied: c is 38.2% to the right of a, and d is 38.2%
to the left of b; and point d lies 38.2% of the way from c to b, just as c is 38.2% of
the way from a to b. This means that the lengths of the intervals [c, b] and [a, b]
have the golden ratio of 1 to 1.618. (This ratio appears often in nature, art,
design, and mathematics, being very closely related to the Fibonacci sequence.)
With this in mind, here are the steps of the algorithm:
√
Step 0: Initialize. Let θ = (3 − 5)/2, set the iteration counter k = 0, and
initialize a0 = a, b0 = b, c0 = a + θ(b − a), d0 = b − θ(b − a).
Step 1: Determine which end to eliminate. If f (ck ) ≥ f (dk ), go to step
2; otherwise, go to step 3.
Step 2: Eliminate lower end. Update ak+1 = ck , bk+1 = bk , ck+1 = dk , and
dk+1 = ck + θ(bk − ck ). Skip to step 4.
Step 3: Eliminate upper end. Update ak+1 = ak , bk+1 = dk , ck+1 = ak +
θ(ck − ak ), and dk+1 = ck . Proceed to step 4.
C.1. MORE LINE SEARCH ALGORITHMS FOR ONE-DIMENSIONAL OPTIMIZATION507

Figure C.1: Golden section method.

Step 4: Iterate. Increase the counter k by 1 and check the termination crite-
rion. If bk − ak < , then terminate; otherwise, return to step 1.

Example C.1. Find the minimum of the function f (x) = (x − 1)2 + ex in the
interval [0, 2], within a tolerance of  = 0.01.
Solution. In the initialization phase, we set k = 0, a0 = 0, b0 = 2,  = 0.01,
and

c0 = a0 + θ(b0 − a0 ) = 0.764
d0 = b0 − θ(b0 − a0 ) = 1.236

We now proceed to step 1. Since f (c1 ) = 2.2025 < f (d1 ) = 3.4975, we decide
to eliminate the upper end and perform step 3.

a1 = a0 = 0b1 = d0 = 1.236
c1 = a0 + θ(c0 − a0 ) = 0.4722d1 = c0 = 0.764

The interval is still wider than the tolerance , so we return to the first step.
Now, f (c2 ) = 1.882 < f (d2 ) = 2.2025, so we again eliminate the upper end by
performing step 3.

a2 = a1 = 0
b2 = d1 = 0.764
c2 = a1 + θ(c1 − a1 ) = 0.2918
d2 = c1 = 0.4721
508 APPENDIX C. OPTIMIZATION TECHNIQUES

Table C.1: Demonstration of the golden section algorithm with f (x) = (x −

1)2 + ex , x ∈ [0, 2].
k ak bk ck dk f (ck ) f (dk )
0 0 2 0.764 1.236 2.20254 3.49751
1 0 1.236 0.472152 0.764 1.88206 2.20254
2 0 0.764 0.291848 0.472152 1.84038 1.88206
3 0 0.472152 0.180362 0.291848 1.86946 1.84038
4 0.180362 0.472152 0.291848 0.360688 1.84038 1.84304
5 0.180362 0.360688 0.249247 0.291848 1.84669 1.84038
6 0.249247 0.360688 0.291848 0.318118 1.84038 1.8395
7 0.291848 0.360688 0.318118 0.334391 1.8395 1.84012
8 0.291848 0.334391 0.308100 0.318118 1.83956 1.8395
9 0.308100 0.334391 0.318118 0.324348 1.8395 1.83963
10 0.308100 0.324348 0.314306 0.318118 1.83948 1.8395
11 0.308100 0.318118 0.311926 0.314306 1.8395 1.83948
12 0.311926 0.318118 0.314306 0.315753 1.83948 1.83949

We proceed as follows until bk − ak < , with the remaining steps shown in

Table C.1. The optimal value is a13 +b
2
13
= 0.3150. 
We used this same example for bisection in Section 3.3.2. Golden section
required more iterations than bisection, to find a solution with the same tol-
erance. This is because bisection reduces the interval width by half at each
iteration, whereas golden section reduces it by θ ≈ 38%. On the other hand, we
did not have to calculate any derivatives. The value of the objective function f
was enough.

Example C.2. Find the minimum of the function f (x) = 1 + e−x sin(−x) in
the interval [0, 3] with a tolerance of 0.01.

Solution. See Table C.2 below. The optimal value is a13 +b 2

13
= 0.7860.
(Using calculus, the optimal solution is actually π/4 ≈ 0.7854.) 

C.1.2 Newton’s method

You may recall Newton’s method from calculus: to find a zero of some function
g, the method starts with an initial guess, which is iteratively updated by finding
where the linear approximation of g at that point has a zero. Under the right
conditions, Newton’s method is almost miraculously fast, exhibiting quadratic
convergence (roughly speaking, the number of correct digits doubles at each
iteration). There are a few downsides: it can only be applied if g is differentiable,
and Newton’s method may not converge at all if you start too far away from
the root, or if the shape of the function makes Newton’s method “overshoot”
and end up further away than it started. Nevertheless, Newton’s method and
its variants are very commonly used in optimization. In this appendix, we will
C.1. MORE LINE SEARCH ALGORITHMS FOR ONE-DIMENSIONAL OPTIMIZATION509

Table C.2: Demonstration of the golden section algorithm with f (x) = 1 +

e−x sin(−x), x ∈ [0, 3].
k ak bk ck dk f (ck ) f (dk )
0 0 3 1.146 1.854 0.710349 0.849629
1 0 1.854 0.708228 1.146 0.679624 0.710349
2 0 1.146 0.437772 0.708228 0.726369 0.679624
3 0.437772 1.146 0.708228 0.875457 0.679624 0.680064
4 0.437772 0.875457 0.604968 0.708228 0.689418 0.679624
5 0.604968 0.875457 0.708228 0.77213 0.679624 0.67766
6 0.708228 0.875457 0.77213 0.811575 0.67766 0.67782
7 0.708228 0.811575 0.747707 0.77213 0.678073 0.67766
8 0.747707 0.811575 0.77213 0.787178 0.67766 0.677604
9 0.77213 0.811575 0.787178 0.796507 0.677604 0.677643
10 0.77213 0.796507 0.781442 0.787178 0.677608 0.677604
11 0.781442 0.796507 0.787178 0.790752 0.677604 0.677612
12 0.781442 0.790752 0.784999 0.787178 0.677603 0.677604

use it in several ways; and several of the algorithms mentioned in the main text
make use of it as well.
Here we are specifically concerned with solving a one-dimensional optimiza-
tion problem with an interval constraint. If the objective function f is convex,
then it is enough to find a point where the derivative f 0 vanishes. So, we simply
apply Newton’s method to the derivative, with g ≡ f 0 , to try to find x̂ such that
f 0 (x̂) = 0. Newton’s method uses the derivative of g, which ends up being the
second derivative f 00 . (If f is not twice-differentiable, Newton’s method cannot
be applied.) We have to make one minor modification to Newton’s method: the
line search cannot leave the feasible region [a, b], so we truncate the search at
these boundary points. This can actually be helpful, since it prevents Newton’s
method from diverging. There are still cases where Newton’s method can fail;
an example is given in Example C.5 below.
Unlike bisection or golden section, Newton’s method is not an “interval re-
duction” method, where we gradually shrink the range of possible values where
the optimum can lie. So we need a different way to measure convergence. It
is common to stop when f 0 is “close enough” to zero; we will let 0 denote this
value.
The steps of Newton’s method for line search are:
Step 0: Initialize. Set the iteration counter k = 0, and initialize x0 to any
point in [a, b]. (If you have a good guess as to the minimum point, it can
greatly speed things up.)
Step 1: Check convergence. If |f 0 (xk )| < 0 , then terminate.
Step 2: Calculate recommended shift. Create a candidate point x̃ = xk −
f 0 (xk )/f 00 (xk ).
510 APPENDIX C. OPTIMIZATION TECHNIQUES

Step 3: Ensure feasibility. Project candidate onto the feasible region xk+1 =
proj[a,b] (x̃) by setting xk+1 = a if x̃ < a, xk+1 = b if x̃ > b, and xk+1 = x̃
otherwise.
Step 4: Iterate. Increase k by 1, and return to step 1.
Example C.3. Find the minimum of the function f (x) = (x − 1)2 + ex in
the interval [0, 2] using Newton’s method, with 0 = 0.0337. (In this and the
next example, 0 is chosen to make the tolerance comparable to the  value used
for bisection and golden section; for this function, when |f 0 (x)| < 0.0337, x is
within 0.01 of its optimal value.)
Solution. Start by computing the formulas for the first and second
derivative of f , since we will be using these often: f 0 (x) = 2(x − 1) + ex ,
and f 00 (x) = 2 + ex .
In the initialization phase, we set k = 0. For an initial guess, choose x0 = 1.
(This makes for a fair comparison, since this is the starting point for bisection.)
The first and second derivatives are equal to e and e+2, respectively, so the new
candidate point is x̃ = 1 − e/(e + 2) = 0.4239. This lies within the boundary
[0, 2], so we accept the candidate point as the next solution: x1 = 0.4239.
At this new point, the first and second derivatives equal 0.3756 and 3.528, so
the next candidate is 0.4239 − 0.3756/3.528 = 0.3174. We accept the candidate
as the new point, so x2 = 0.3174. At x2 , the derivative is f 0 (x2 ) = 0.0084 < 0 ,
so we terminate and report 0.3174 as the optimal solution. 
Notice that Newton’s method achieved in only two iterations the level of
precision bisection reached in eight, and golden section reached in thirteen! At
this point, the solution given by Newton’s method differs from the true optimum
x by roughly 2 × 10−3 . One more iteration of Newton’s method would reduce
this error to 1 × 10−6 , and yet another would reduce it to 3 × 10−13 . This is
what we mean when we say its convergence rate is miraculous!
Example C.4. Find the minimum of the function f (x) = f (x) = 1+e−x sin(−x)
in the interval [0, 3] using Newton’s method, with 0 = 0.00645. (Again, this
choice of 0 ensures that when Newton’s method terminates, x is within 0.01 of
its optimal value.)
Solution. Start by computing the formulas for the first and second deriva-
tive of f , since we will be using these often: f 0 (x) = e−x (sin x − cos x), and
f 00 (x) = 2e−x cos x.
In the initialization phase, we set k = 0. For an initial guess, choose x0 = 1.5.
(This makes for a fair comparison, since this is the starting point for bisection.)
The first and second derivatives are equal to 0.2068 and 0.03157, respectively, so
the new candidate point is x̃ = 1.5 − 0.2068/0.03157 = −5.051. This is outside
of the feasible interval [0, 3], so we project the candidate back onto the feasible
region, choosing x1 = 0.
At this new point, the first and second derivatives equal −1 and 2, so the
next candidate is 0 − (−1)/2 = 0.5. We accept the candidate as the new point,
so x2 = 0.5. Another two iterations are needed: x3 = 0.5 − (−0.2415)/(1.065) =
C.2. LINEAR PROGRAMMING 511

0.7268, and x4 = 0.7822. At x4 , the derivative is |f 0 (x4 )| = 0.0021 < 0 , so we

terminate and report 0.7822 as the optimal solution. 
Newton’s method is not infallible, however, as shown by the next example.
Example C.5. Apply Newton’s method to the function f (x) = x tan−1 x −
1 2
2 log(1 + x ) over x ∈ [−2, 2], starting with the initial guess x0 = 1.3917452.
(This formula looks unwieldy, but plot the function! It actually looks quite nice,
and you can spot the minimum immediately by inspection.)
Solution. The first and second derivatives are f 0 (x) = tan−1 x and f 00 (x) =
1/(x2 + 1). Calculating f 0 and f 00 at the initial guess x0 , we find the next point
to be x1 = −1.3914752 — exactly the opposite of where we started! You can
guess what happens next. Calculating from x1 , we find the next point to be
x2 = +1.3914752, and so on ad infinitum. 
Again, this is a “nice” function — it is even strictly convex! Yet Newton’s
method fails to find the optimal solution. You can verify that if we start with
an initial guess closer to zero, Newton’s method will converge, and if the initial
guess is further away, Newton’s method will diverge to whatever the endpoints
are of the feasible interval.

C.2 Linear Programming

This section briefly outlines the properties of a linear optimization problem
(commonly called a linear program), terminology associated with solutions of
linear programs, and two solution methods: a graphical method suitable for
small problems, and the simplex method which is widely used for practical
problems. Recall that a linear program is an optimization where the objective
function and constraints are all linear in the decision variables; the constraints
may either be equalities (=) or weak inequalities (≥, ≤).

C.2.1 Basic properties and graphical solution method

This section will predominantly use the following example:
min 10x + 26y
x,y
s.t. 11x + 3y ≥ 21
6x + 20y ≥ 39
x+y ≤9
x, y ≥0
The feasible region for this problem is shown in Figure C.2 as the area shaded
in yellow.
The optimal solution to this linear program is the point (x, y) in the feasible
region which minimizes 10x+26y. In this case, the optimal solution is (1.5, 1.5),
with a corresponding objective function value of 54. It is not immediately
obvious how we found this point, since there is an infinite number of feasible
solutions.
512 APPENDIX C. OPTIMIZATION TECHNIQUES

10

6 x+y ≤9

5
y

11x + 3y ≥ 21
4

3 Z = 10x + 26y

2
(1.5,1.5)
1
6x + 20y ≥ 39

0
0 1 2 3 4 5 6 7 8 9 10
x

Figure C.2: An illustrative linear program.

C.2. LINEAR PROGRAMMING 513

This point happens to lie at a “corner” of the feasible region. This is not
a coincidence. If a linear program has an optimal solution, and if its feasible
region has a corner, then there is an optimal solution that lies at such a corner
point. As a result, we can usually confine our attention to the corner points, of
which there are only a finite number. In this particular case, the corner point
feasible solutions are (0, 9), (0, 7), (1.5, 1.5), (6.5, 0) and (9, 0). Checking in each
in turn, we see that the objective function is minimized at (1.5, 1.5), so this is
the optimal solution.
This provides us with a method for solving a linear program, if the linear
program has an optimal solution and its feasible region has at least one corner.

ˆ Plot all the constraints and identify the feasible region.

ˆ Determine the feasible corner points.

ˆ Evaluate the objective function at the feasible corner points.

ˆ The “best” feasible corner point is the optimal solution

For small problems this can be done graphically, as in Figure C.2. For larger
problems, it is difficult to identify all the corner points by inspection, and a
different method is needed.
This intuitive description is enough for our purposes; for readers wanting a
more technical definition of a “corner,” we can define it as a point that is not the
midpoint of any line segment contained in the feasible region. These points are
also known as vertices, or extreme points. A point along an edge of the feasible
region can be drawn as the midpoint of a line segment drawn along this edge;
a point in the interior of the feasible region is the midpoint of many different
line segments. But a line segment drawn in the feasible region can only have a
corner as one of its endpoints, never the midpoint.
Why do we need the caveats “if a linear program has an optimal solution”
and “if its feasible region has a corner”? Consider the examples below. Modify
the constraints so the optimization problem has the following form:

min 10x + 26y

x,y
s.t. x+y ≤9
y ≤5
y ≥ 0.

This produces the feasible region shown in Figure C.3. There is no restriction
on how small x can be, and as x tends to −∞, so does the objective function.
Therefore there is no optimal solution; the problem is said to be unbounded
since there is no limit to how much the objective function can be minimized.
(The same would hold true for a maximization problem if it is possible for the
objective function to be made arbitrarily large.)
514 APPENDIX C. OPTIMIZATION TECHNIQUES

10

6
y

5
y≤5

4 x+y ≤9

2 Z = 10x + 26y

y≥0
-4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10
-1 x

Figure C.3: An unbounded linear program.

C.2. LINEAR PROGRAMMING 515

Or, consider this optimization problem:

min 10x + 26y

x,y
s.t. x+y ≤9
y ≤5
x ≥ 9.5
y ≥ 0.

There is no solution which obeys all the constraints (see Figure C.4). The
constraints x + y ≤ 9 and y ≥ 0 mean that x cannot be greater than 9; however
there is also the constraint x ≥ 9.5. Such a problem is called infeasible, because
there is not even a feasible solution (let alone an optimal one). As a rule of
thumb, unbounded or infeasible problems often mean that you have missed
something in your formulation. In the real world, it is not possible to produce
“infinitely good” solutions (surely there is some limitation; this would be a
constraint you have missed), and there is usually some possible course of action,
even if it is very unpleasant (your constraints are too restrictive).
Here is an optimization problem which has an optimal solution, but its
feasible region has no corners (Figure C.5):

min x+y
x,y
s.t. x + y ≥ 0.

Any solution on the line x + y = 0 is feasible and optimal. So this problem has
infinitely many optimal solutions, but no corner points. These examples show
that the statement “at least one corner point is optimal, so we only have to look
at corner points” needs a few technical qualifications (that there is in fact an
optimal solution, and a corner point). In practice these kinds of counterexamples
are rare.
Also note that this statement does not say that only corner points may be
optimal, even with the qualifications that there are corner points and optimal
solutions. There very well may be a non-corner point which is optimal. But in
such cases there is also an optimal corner point, so it is still acceptable to just
check the corner points. We will still find an optimal solution that way. An
example of such a problem is

min 12x + 40y

x,y
s.t. 11x + 3y ≥ 21
6x + 20y ≥ 39
x+y ≤9
x, y ≥ 0.

(See Figure C.6) Any feasible point on the line segment between (1.5, 1.5)
and (6.5, 0) has the same, optimal objective function value of 78. So a solution
like (3.5, 0.75) is optimal even though it is not a corner point. However, the
corner points (1.5, 1.5) and (6.5, 0) are both optimal as well, so we can still find
516 APPENDIX C. OPTIMIZATION TECHNIQUES

10

6
y

5 x ≥ 9.5
y≤5

4 x+y ≤9

2 Z = 10x + 26y

y≥0
-4 -3 -2 -1 1 2 3 4 5 6 7 8 9 10
-1 x

Figure C.4: An infeasible linear program.

C.2. LINEAR PROGRAMMING 517

5
y
4

2 Z =x+y

x+y ≥0 1
x

-5 -4 -3 -2 -1 1 2 3 4 5
-1

-2

-3

-4

-5

Figure C.5: An linear program with no corner points.

518 APPENDIX C. OPTIMIZATION TECHNIQUES

10

6 x+y ≤9

5
y

11x + 3y ≥ 21
4

2 Z = 12x + 40y

1
6x + 20y ≥ 39

0
0 1 2 3 4 5 6 7 8 9 10
x

Figure C.6: Multiple optimal solution

C.2. LINEAR PROGRAMMING 519

an optimal solution if we just check corners. (For most purposes it is fine to

identify a single optimal solution, Identifying all optima is usually unnecessary.)
This graphical method is good for building intuition and illustrating the main
ideas behind solving linear programs. However, it becomes difficult as problems
become larger. A problem with three decision variables would require a 3D
drawing; a problem with four decision variables would require drawing in 4 di-
mensions, which is beyond the scope of most of us mere mortals. Practical-sized
problems commonly involve thousands or even millions of decision variables, so
a different technique is needed. However, the idea of focusing on corner points
can be translated to problems of this scale. To do so requires a systematic pro-
cedure for generating and checking corner points. This procedure is called the
simplex algorithm. The mathematical tools for doing this rely on linear algebra,
since corner points can be identified by solving systems of linear equations (in
the examples in this section, each corner point was the intersection of two lines).
The remainder of this section is focused on presenting this method.

C.2.2 Standard form of a linear program

The simplex algorithm is easier to explain if the linear program is converted to
a standard form. This standard form is not restrictive, and you can transform
any linear program into an equivalent one which is in standard form. A linear
program is in standard form if:

ˆ The objective function is to be minimized.

ˆ All decision variables have a non-negativity constraint (that is, xi ≥ 0 is

a constraint for every decision variable xi ).

ˆ All other constraints are equality (=) constraints.

In the previous section, we considered this linear program:

min 10x + 26y

x,y
s.t. 11x + 3y ≥ 21
6x + 20y ≥ 39
x+y ≤9
x, y ≥0

This is a minimization problem where all variables must be non-negative; but

the remaining constraints are inequalities, not equalities. We can fix this by
introducing three new variables s1 , s2 , and s3 , and rewriting the optimization
problem:
min 10x + 26y
x,y,s1 ,s2 ,s3
s.t. 11x + 3y − s1 = 21
6x + 20y − s2 = 39
x + y + s3 =9
x, y, s1 , s2 , s3 ≥ 0.
520 APPENDIX C. OPTIMIZATION TECHNIQUES

This problem is equivalent to the original one in the sense that any solution to
the standard-form problem can easily be translated to a solution in the original
problem (just ignore the new variables); and a solution is feasible to one problem
if and only if it is feasible to the other problem. Given a solution (x, y) to the
original problem, let s1 = 11x+3y −21, s2 = 6x+20y −30, and s3 = 9−(x+y).
If that solution is feasible, these new variables must all be non-negative, and the
three equality constraints in the standard-form problem will be satisfied. The
reverse is true as well; for instance, since 11x+3y = 21+s1 in a feasible solution
to the standard-form problem, and since s1 ≥ 0, we must have 11x + 3y ≥
21 and the first constraint in the original problem is also satisfied. The new
variables introduced (s1 , s2 , and s3 ) are often called slack variables, because
they show how much the left-hand side of the constraint can be changed before
it is violated.
This technique can be used to ensure that all constraints are equalities.
Similarly, non-standard problems violating the other requirements can also be
converted into equivalent, standard-form linear programs. If the objective func-
tion is maximization, rather than minimization, we can multiply the objective
function by −1. (Minimizing −f is the same as maximizing f ; see Proposi-
tion B.1 in Section B.6.2.) If there is a decision variable which is “free” (it does
not have to be non-negative), it can be replaced with two new decision variables.
For instance, if the variable y can be either negative or positive, everywhere y
appears in the optimization problem we can replace it with y + − y − , where y +
and y − are two new decision variables with non-negativity constraints (y + ≥ 0,
y − ≥ 0). This works because we represent any value (positive or negative) as a
subtraction of two positive numbers.
It is common to write linear programs in matrix-vector form, since this
is more convenient when there are a large number of decision variables and
constraints. In matrix notation, the linear program in standard form can be
written as:
min c · x
x
s.t. Ax =b
x ≥0

where x is the vector of decision variables, c is a vector representing the

coefficients in the objective function, b is a vector representing the right-hand
side of each equality constraint, and A is the matrix of coefficients in all of
the equality constraints. (Section A.2 provides a review of vector and matrix
operations if you are unfamiliar with them.) For the above example in standard
form, these vectors and matrices are as follows:
 
x
y
 
s1 
x= 
s2 
s3
C.2. LINEAR PROGRAMMING 521

 
10
26
 
0
c= 
0
0

 
11 3 −1 0 0
A =  6 20 0 −1 0
1 1 0 0 1

 
21
b = 39
9

C.2.3 Simplex algorithm

Usually, a linear program has an infinite number of feasible solutions. However,
as discussed above, it is enough to look at corner points of the feasible region,
since at least one of them will be optimal (assuming that the problem has an
optimal solution, and that there is a corner point, conditions we will take as
given in this subsection). The simplex algorithm is designed to systematically
search these corner points. Starting from one corner point, it moves to an
adjacent corner point with an equal or smaller objective function value. Since
there are a finite number of corner points, and since linear programs have the
property that you can reach an optimal corner from any other by a sequence
of adjacent corners with nonincreasing objective function values, this procedure
is guaranteed to find an optimal solution. A corner point can be identified
by solving systems of equations representing a subset of the constraints, so the
simplex method is presented in terms of linear algebra. This subsection provides
the steps of the simplex method, along with some intuitive justification for them,
but we do not provide rigorous proofs of correctness, nor do we handle all of
the “corner cases” that can arise when applying the simplex method. A reader
interested in these details is referred to a standard text in linear programming,
such as Bertsimas and Tsitsiklis (1997).
Assume we are given a standard-form linear program, where the decision
variable x is a vector of dimension n, the objective function coefficient c is a
vector of dimension n, b is a vector of dimension m, and the coefficient matrix
A is a matrix of dimension m × n. We can assume there are at least as many
variables as equality constraints (n ≥ m); otherwise some of the constraints are
redundant and can be removed. (Indeed, if the number of independent con-
straints is equal to the number of variables, there is only one feasible solution!)
Let Ai be the vector corresponding to the i-th column of the matrix A.
522 APPENDIX C. OPTIMIZATION TECHNIQUES

 
ai1
 ai2 
Ai =  . 
 
 .. 
aim

min c·x
x
s.t. Ax =b
x ≥0

A basis B refers to a set of m linearly independent columns of A. These

columns are represented by the square matrix AB , which is of size m × m.
We will use N B (for “nonbasic”) to represent all of the other columns of A.
By limiting ourselves to the decision variables xi corresponding to the columns
in the basis, we now have exactly the same number of decision variables and
constraints. Fixing all of the other decision variables to zero, we can now solve
the m equality constraints in terms of these m basic variables. The resulting
solution xB corresponds to an intersection point of the constraints. If xB ≥
0, then this solution satisfies all of the constraints, and is therefore feasible.
The resulting value of x, whose basic components solve the constraints (xB =
AB −1 b), and whose nonbasic components are zero (xN = 0) form a corner
point solution.
Every corner point can be identified in this way. So the question is to find
some basis B corresponding to an optimal solution. Starting from some initial
basis, the simplex method looks for an adjacent basis (one which is the same
except for one column) with a lower value of the objective function. If one exists,
we update the basis and repeat the process to try to find a better one. If no
adjacent basis is better, we terminate with the optimal solution. The simplex
method can determine whether an adjacent basis is better by computing the
reduced cost, the rate of change in the objective function if a nonbasic column
were to enter the basis in place of one of the existing columns. Since the objective
function is minimization, a negative reduced cost will represent a better solution.
Using matrix algebra, we can easily compute reduced costs, and new bases, using
formulas shown below.
The steps of the simplex method are as follows:
1. Initialization:
ˆ Determine a basis AB comprising of m linearly independent columns
of A.
ˆ Calculate xB = AB −1 b.
ˆ If xB ≥ 0, continue; otherwise, this basis is infeasible. Choose a
different basis and repeat this step. If you have already tried every
basis, the linear program is infeasible. Terminate the algorithm and
report this.
C.2. LINEAR PROGRAMMING 523

ˆ Set xN = 0.
2. Compute reduced costs:
ˆ Compute the reduced cost of each nonbasic index j ∈ N as c̄j =
cj −cB T AB −1 Aj , where cB is the column vector of objective function
coefficients for the basic variables.
ˆ If c̄j ≥ 0 for all j ∈ N , then the current solution is optimal. Termi-
nate the algorithm and return the solution.
ˆ Otherwise, choose some nonbasic variable j ∈ N with c̄j < 0. The
column representing this variable will enter the basis.
ˆ Set dj = 1, and di = 0 for all i ∈ N − {j}.
3. Identify descent direction:
ˆ Compute the basic direction dB = −AB −1 Aj .
ˆ If di ≥ 0 for all i ∈ B, the linear program is unbounded. Terminate
the algorithm and report this.
ˆ Otherwise, di < 0 for at least one i ∈ B. Compute the maximum
step size  
−xi
θ = min . (C.1)
i∈B:di <0 di
ˆ Let k be the index of a variable achieving this minimum, with θ =
−xi /di . The column representing this variable will leave the basis
(to be replaced with column j).
4. Compute new solution:
ˆ Set xnew
i = xi + θdi for all variables i (both basic and nonbasic).
ˆ Update the basis AB by replacing Ak with Aj .
ˆ Update the index sets by removing k from B, and adding j. Likewise,
remove j from N and add k.
ˆ Update the vectors cb , xb , xN appropriately based on xnew
i .
ˆ Return to step 2.
As an example, we show how the simplex method will operate on the follow-
ing standard-form linear program (Figure C.7 plots the feasible region).

min 10x + 26y

x,y,s1 ,...,s4
s.t. 11x + 3y − s1 = 21
6x + 20y − s2 = 39
x + y + s3 =9
1
− x + y + s4 =6
3
4x − 5y + s5 = 20
x, y, s1 , s2 , s3 , s4 ≥0
524 APPENDIX C. OPTIMIZATION TECHNIQUES

10

−1
7 3 x +y ≤6

5 x+y ≤9
y

4
11x + 3y ≥ 21

3 4x − 5y ≤ 20
Z = 10x + 26y

2
(1.5,1.5) 6x + 20y ≥ 39
1

0
0 1 2 3 4 5 6 7 8 9 10
x

Figure C.7: Feasible Region of the LP

C.2. LINEAR PROGRAMMING 525

Initialization: Set B = {1, 2, 3, 4, 7} and N = {5, 6}. Then

     
21 11 3 −1 0 0 10
39
 
6
 20 0 −1 0

26
 
1 1 0 0 0
b=  9  AB =  −1  cB =  0 
  
6  1 0 0 0  0
3
20 4 −5 0 0 1 0
   
x 2.25
y  6.75     
s3 0
xB = AB −1 b = 
   
s
 1
 =  24.00 xN = =
s2 

109.5
 s4 0
s5 44.75
The values of xB are non-negative, so this basis is a feasible corner point.
The objective function Z = cB T xB = 198. The solution corresponds to the red
dot in Figure C.7.
Compute reduced costs: N = {5, 6}.

c¯5 = c5 − cB T AB −1 A5 = −14
c¯6 = c6 − cB T AB −1 A6 = −12

As both reduced costs are negative, we can pick either of them. Let us pick
j = 5. Set d5 = 1, d6 = 0.
Identify descent direction:
Now B = {1, 2, 3, 4, 7}. Compute:

 
−0.75
−0.25
−1
 
dB = −AB A5 =  −9.00

−9.50
1.75
 
2.25 6.75 24.0 109.5 24
θ = min , , , =
0.75 0.25 9.0 9.5 9
The index k = 3.
Compute new solution:
     
2.25 −0.75 0.25
 6.75  −0.25  6.0833 
     
 24  −9.00  0 
xnew
     
109.5 + θ −9.50 = 84.1667
= x + θd =      
 0   1.00   2.6667 
     
 0   0   0 
44.75 1.75 49.4167
526 APPENDIX C. OPTIMIZATION TECHNIQUES

The new B = {1, 2, 5, 4, 7} and N = {3, 6}.

   
x 0.25
y  6.083     
    s1 0
xB =  s
 3
 =  2.6667  x N = =
s2 

84.1667
 s4 0
s5 49.4167
   
11 3 0 0 0 10
6
 20 0 −1 0

26
 
AB =  1
 −1 1 1 0 0  cB =  0 
 
 1 0 0 0  0
3
4 −5 0 0 1 0
The objective function Z = cB T xB = 160.66588. The basic feasible solution
corresponds to the yellow dot in Figure C.7.
Compute reduced costs: N = {3, 6}.

c¯3 = c3 − cB T AB −1 A3 = 1.5556
c¯6 = c6 − cB T AB −1 A6 = −21.333

As the reduced cost c6 < 0, pick j = 6. Set d6 = 1, d3 = 0.

Identify descent direction:
Now B = {1, 2, 5, 4, 7}. Compute:

 
0.25
 −0.9167 
−1
 
dB = −AB A6 = 
 0.6667 

−16.8333
−5.5833
 
6.083 84.1667 49.4167 84
θ = min , , =
0.967 16.8333 5.5833 16.8333
The index k = 4.
Compute new solution:
     
0.25 0.25 1.5
 6.0833   −0.9167   1.5 
     
 0   0.0   0 
xnew
     
84.1667 + θ −16.8333 =  0 
= x + θd =      
 2.6667   0.6667   6.0 
     
 0   1.0   5.0 
49.4167 −5.5833 21.5
The new B = {1, 2, 5, 6, 7} and N = {3, 4}.
C.2. LINEAR PROGRAMMING 527

   
x 1.5
y  1.5     
    s1 0
xB = s
 3
 =  6.0  xN = =
s4 
 
 5.0  s2 0
s5 21.5
   
11 3 0 0 0 10
6 20 0 0 0 26
   
AB =  1
 −1 1 1 0 0 cB =  0 
 
 1 0 1 0  0
3
4 −5 0 0 1 0
The objective function Z = cB T xB = 54.0. The basic feasible solution
corresponds to the green dot in Figure C.7.
Calculate the reduced cost: N = {3, 4}.

c¯3 = c3 − cB T AB −1 A3 = 0.2178
c¯4 = c4 − cB T AB −1 A4 = 1.2673

As both the reduced costs are positive, we have reached the optimal solution
which is x = 1.5, y = 1.5 with an objective function value of 54.
The above procedure provides the basic framework for the simplex algorithm.
There are a few issues that can arise when applying this algorithm. First, the
issue of degeneracy, which occurs when a basic variable has the value of zero.
Usually it is the nonbasic variables that are equal to zero; but sometimes we need
to set a basic variable to zero to solve the system of equations. This can cause a
few problems, and a poorly-designed implementation of the simplex algorithm
may get stuck in an infinite loop in the presence of degeneracy, rotating among
the same set of columns over and over again with no change in the objective
function or the solution, just changing which zero variables count as basic and
nonbasic. Degeneracy can be addressed by a good tiebreaking rule that can
prevent these infinite loops. One such rule is Bland’s rule: among all the indices
j for which c̄j is a negative reduced cost, choose the first of them (smallest j)
to enter the basis; and among all the indices k which achieve the minimum
in (C.1), choose the first of them (smallest k) to leave the basis.
There are also a few steps which were not unambiguously specified in the
algorithm above.

Identifying an initial basis: Choosing an initial basis which corresponds to

a feasible solution can be difficult, especially if there are many columns.
One solution to this problem is the “big-M method,” which generates an
auxiliary variable yi for each of the m equality constraints. To each of
the constraints, we either add yi (if the right-hand side bi is positive) or
subtract yi (if bi is negative). In the objective function, we add M yi for
each of these auxiliary variables, and we add a non-negativity constraint
528 APPENDIX C. OPTIMIZATION TECHNIQUES

for each yi . By adding these variables, it is easy to get an initial solution by

setting yi = |bi | for each constraint. If M is a large enough constant, this
solution will have a very poor objective function value, and any feasible
corner point solution to the original problem will have a lower objective.
So, we can start the simplex method at this solution, and as it moves to
better and better solutions it will eventually remove all of the artificial
variables from the basis. The resulting optimal solution will be feasible to
the original problem as well.
Selecting a nonbasic variable to enter the basis: In the description of the
simplex method, we left the choice j of an entering variable arbitrary; we
could choose any nonbasic variable with a negative reduced cost. If there is
more than one choice, we have to decide how to proceed, and this choice
can affect the number of iterations required. In the above example, in
the first iteration, we had c¯5 = −14 and c¯6 = −12, either of them could
have been chosen. We chose the nonbasic variable corresponding to in-
dex 5 (s3 ) to enter the basis. The simplex algorithm then traveled to the
basic feasible solution corresponding to the yellow dot (Figure C.7), and
then reached the optimal solution in the next round. Therefore, when we
picked s3 , the simplex needed two more iterations to reach the optimal
solution. If we had instead chosen the nonbasic variable corresponding to
index 6 (s4 ) to enter the basis, the algorithm would have traveled to the
orange dot, then the black dot and finally reached the optimal solution.
Therefore, the choice of index 6 would have resulted in one additional it-
eration. Unfortunately, there is no known way to predict in advance how
many iterations will be required. There are two common methods which
perform acceptably in practice: either pick the nonbasic variable with the
most negative reduced cost; or among the nonbasic variables with negative
reduced costs, pick the one with the smallest index.

C.3 Unconstrained Nonlinear Optimization

This section discusses optimization problems of the form min f (x), where x is an
n-dimensional real vector, with no constraints whatsoever. We begin by deriving
optimality conditions, and then present methods for solving such problems.

C.3.1 Optimality conditions

In the main text, Section 3.3 gave optimality conditions for optimization prob-
lems with a convex objective function, linear equality constraints, and non-
negativity constraints on the variables. Not all optimization problems fall into
these categories, and this section provides some information on these cases.
Recall that the purpose of optimality conditions is to provide mathematical
equations or inequalities that can be used to check whether an optimal solution
has been found or not. In theory, it may be possible to obtain an optimal solu-
tion by solving such a system of equations and inequalities directly. However,
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 529

this is often very difficult. Instead, the optimality conditions are mostly used in
solution algorithms to know when an optimal solution has been found (or if we
are close to optimality), and to provide guidance on how to improve a solution
if it is not optimal. Throughout this discussion, we make heavy reference to
“necessary” and “sufficient” conditions, explained in Section B.6.4.
This subsection deals with unconstrained nonlinear minimization problems
of the form:
min {f (x), x ∈ Rn }
The objective function may or may not be convex. Some of the results require
assumptions on differentiability, which we will state as needed. These results
will be stated without proof; readers wanting more explanation are referred to
the books by Bertsekas (2016) and Bazaraa et al. (2006).
This section presents several necessary and sufficient optimality conditions
for unconstrained nonlinear minimization problems. Optimality conditions are
important because they help identify if a given solution is optimal or not. This
can help in algorithm development to check if we can stop the algorithm or
proceed further. In certain specific cases, the optimality conditions can also help
solve for the optimal solution or arrive at a set containing optimal solutions.
Definition C.1. If the function f is differentiable, a stationary point is a value
of x∗ such that ∇f (x∗ ) = 0.
In many cases, stationary point is either a local minimum or local maximum.
However, this is not always the case; for instance if f (x) = x3 , then x∗ = 0 is a
stationary point, but the function has neither a minimum or maximum there. A
stationary point which is neither a local minimum or a local maximum is called
a saddle point.
Theorem C.1. (First-order necessary conditions for local minima.) If f is
differentiable at a point x∗ which is a local minimum, then x∗ is also a stationary
point.
This result is a necessary condition; it is “first-order” because it refers to
the first derivative. Therefore, if x∗ is a local minimum, then it is a stationary
point. But stationary points need not be local minima; they could also be local
maxima or saddle points, for instance.
Example C.6. Determine the stationary points of the function f (x) = (x−5)2 .
Does a minimum exist?
Solution.
∇f (x∗ ) = 2(x∗ − 5) = 0 =⇒ x∗ = 5.
Thus x∗ = 5 is the stationary point. By plotting the graph, one can easily
determine that the stationary point corresponds to a local as well as global
minimum. 
Example C.7. Determine the stationary points of the function f (x) = x3 +
x2 − x + 1. Does a minimum exist?
530 APPENDIX C. OPTIMIZATION TECHNIQUES

Solution.

∇f (x∗ ) = 3(x∗ )2 + 2x∗ − 1 = 0 =⇒ x∗ = 1/3, −1.

If you plot the graph, you will notice that x∗ = 1/3 corresponds to a local
minimum and x∗ = −1 corresponds to a local maximum. 

Example C.8. Determine the stationary points of the function f (x1 , x2 ) =

4(x1 − 7)2 + 5x22 (x2 − 10).

Solution.
∂f
= 8(x∗1 − 7) = 0 =⇒ x∗1 = 7
∂x∗1

∂f
= 3(x∗2 )2 − 20x∗2 = 0 =⇒ x∗2 = 0, 20/3.
∂x∗2
The two stationary points are (7, 0) and (7, 20/3). However, we do not have
enough information to determine if the stationary points are minima, maxima,
or saddle points. 
If the first order necessary condition does not provide clarity on whether a
stationary point is a minimum or not, a second-order necessary condition can
be used if the function is twice differentiable:

Theorem C.2. (Second-order necessary conditions for local minima.) If f is

twice differentiable at a point x∗ which is a local minimum, then (i) x∗ is a
stationary point, and (ii) the Hessian matrix Hf (x∗ ) is positive semidefinite.

The theorem can be extended as follows: if x∗ is a local minimum, then x∗

is stationary and the Hessian is negative semidefinite there; if x∗ is a saddle
point, then x∗ is stationary and the Hessian is neither positive semidefinite nor
negative semidefinite. Again, these are necessary conditions. They must be
true of every optimal solution, but not every point satisfying these conditions
is optimal.

Example C.9. Consider the function f (x) = (x − 5)5 . Identify the stationary
point and check if the Hessian is positive semidefinite.

Solution.
∇f (x∗ ) = 5(x∗ − 5)4 = 0 =⇒ x∗ = 5.
The stationary point is x∗ = 5. The Hessian at the stationary point is f 00 (x∗ ) =
20(x∗ − 5)3 , which is nonnegative (actually zero) and thus positive semidefinite.
However if you plot the function, you will notice that x∗ = 5 does not correspond
to a local minima, but instead a saddle point. 

Example C.10. Consider the stationary point (7, 0) of the function f (x1 , x2 ) =
4(x1 − 7)2 + 5x22 (x2 − 10). Is (7, 0) is a saddle point?
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 531

Solution. At (7, 0) we have

∂2f
=8
∂x21
2
∂ f
=0
∂x1 ∂x2
∂2f
=0
∂x2 ∂x1
∂2f
= 6x2 − 20 ,
∂x22

so the Hessian is given by

 
8 0
Hf = .
0 −20

The Hessian is neither positive nor negative semidefinite. Therefore, the sta-
tionary point (7, 0) is a saddle point. 
A minor modification of the second-order necessary conditions gives a suffi-
cient condition on optimality.

Theorem C.3. (Second-order sufficiency conditions for a strict local mini-

mum.) For a function f twice-differentiable at x∗ , x∗ is a strict local minimum
if (i) x∗ is a stationary point and (ii) the Hessian matrix H(x∗ ) is positive
definite.

Likewise, if H is negative definite, then we know x∗ is a strict local maximum.

As a sufficient condition, any point satisfying these conditions must be a local
minimum. It is not necessary; for instance if f (x) = x2 , then x∗ = 0 is a strict
local minimum, but the Hessian is not positive definite there.

Example C.11. Verify if the stationary points of the function f (x) = x3 +

x2 − x + 1 are local minima. Does a minimum exist?

Solution. The stationary points of the function are x∗ = 1/3 and x∗ = −1.

Hf (x∗ ) = 6(x∗ ) + 2

The Hessian is positive definite at x∗ = 1/3 and negative definite at x∗ = −1.

Therefore, x∗ = 1/3 is a strict local minimum and x∗ = −1 is a strict local
maximum. 

Example C.12. Consider the stationary point (7, 20/3) of the function f (x1 , x2 ) =
4(x1 − 7)2 + 5x22 (x2 − 10). Is (7, 20/3) is a strict local minimum?
532 APPENDIX C. OPTIMIZATION TECHNIQUES

Solution. At (7, 20/3) we have

∂2f
=8
∂x21
∂2f
=0
∂x1 ∂x2
∂2f
=0
∂x2 ∂x1
∂2f
= 6x2 − 20 ,
∂x22

so the Hessian is  
8 0
Ff = .
0 20

This matrix is positive definite, so (7, 20/3) is a strict local minimum. 

Recall from Section 3.3 that if f is a convex function, then any local minimum
must also be a global minimum, and that if f is strictly convex, any local
minimum is the (unique) strict global minimum.

C.3.2 Solution framework

Unconstrained optimization problems are commonly solved using the follow-
ing algorithmic framework; the details of steps 1–3 are filled in later in this
subsection.

Step 0: Initialize. Set the iteration counter k = 0, and initialize x0 by choos-

ing any vector in Rn . (If you have a good guess as to the minimum point,
it can greatly speed things up.)

Step 1: Check convergence. Evaluate one of the termination criteria (Sec-

tion C.3.3.) If it is satisfied, return xk as an approximately-optimal solu-
tion, and stop.

Step 2: Determine a descent direction dk . (Section C.3.4; this step is where

gradient descent and Newton’s method differ.)

Step 3: Determine a step size αk . (Section C.3.5.)

Step 4: Iterate. Update the solution, xk+1 ← xk + αk dk , then increase k by

1, and return to step 1.

The following subsections spell out choices for Steps 1, 2, and 3. These
choices are independent of each other, and you can combine any choice for one
step with any choice for another.
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 533

C.3.3 Convergence criteria

There are several stopping criteria available to decide when to terminate. It is
always best to select a stopping criterion to directly measure what we are trying
to do. In this case, we know that the minimum of f must occur at a stationary
point, which means ∇f = 0. So, a natural stopping criteria is to terminate
when ∇f is “close enough” to zero, that is, when ||∇f || < .
One possible difficulty with this choice is that you need a good intuition for
what values of ||∇f || correspond to “good enough” solutions for your particular
application, and this is not always easy. So, there are other convergence criteria
that can also be used:

ˆ You can stop when the solution stabilizes, ||xk − xk−1 || < .

ˆ You can stop when the objective function stabilizes, |f (xk )−f (xk−1 )| < .

ˆ You can normalize the previous two inequalities to reflect relative stability
(e.g., stop when the objective function decreases by less than 1% between
iterations).

ˆ You can stop after a pre-specified number of iterations.

ˆ You can stop after a certain amount of computation time has elapsed.

These methods are more intuitive to apply than ||∇f || < . The downside
is that there is no guarantee that you are close to an optimal solution when
they are satisfied. This is particularly obvious for the last criteria; but even
for the earlier ones, there is no way to tell the difference between a solution
“stabilizing” for a good reason (you are close to the minimum) or for a bad
reason (the algorithm is stuck somewhere suboptimal but can’t make progress).
You can also use a combination of these methods; for instance, stopping
when the objective changes by less than 1%, or after one hour of run time
(whichever comes first).

C.3.4 Descent direction

There are two main ways to select the direction dk ; the idea in both cases is
that the gradient points in the direction where the function increases as quickly
as possible. Therefore, the negative of the gradient is the direction where the
function decreases as quickly as possible. Since we are trying to minimize the
objective, we will use the latter in some way.
In gradient descent, we simply use the gradient itself, choosing dk = −∇f (xk ).
This method is simple and requires relatively little computation.
A more sophisticated choice is Newton’s method, which is based on the fol-
lowing logic: we can take a quadratic approximation to the objective function
at the current point. We can identify the minimum point of that quadratic
approximation more easily than the original function f (which may be much
more complicated), and then choose the search direction dk to move towards
534 APPENDIX C. OPTIMIZATION TECHNIQUES

that minimum point. We will derive this direction, and then explain how it is
related to Newton’s method as you learned it in calculus, or as we used it in the
previous section.
The quadratic approximation to f at xk is its second-order Taylor series,
based on its gradient and Hessian:

1
f (x) ≈ f (xk ) + (x − xk )T ∇f (xk ) + (x − xk )T Hf (xk )(x − xk )T . (C.2)
2

Let f˜ denote the right-hand side of equation (C.2). The minimum point of the
quadratic approximation is the point where ∇f˜ vanishes, that is, where

∇f˜(x) = ∇f (xk ) + Hf (xk )(x − xk ) = 0 . (C.3)

Solving for x, we obtain

x = xk − (Hf (xk ))−1 ∇f (xk ) . (C.4)

So, the search direction is

dk = x − xk = −(Hf (xk ))−1 ∇f (xk ) . (C.5)

Why is this method related to Newton’s method? Newton’s method aims to

find the zero of a function by finding the zero of its linear approximation. In this
case, we are trying to find the minimum point of f , which occurs where ∇f = 0.
So, when we use Newton’s method, we find where the linear approximation
of ∇f vanishes. But if the gradient of a function is linear, then the original
function is quadratic, and the gradient will be zero at the minimum point of
this quadratic. Also recall the previous section, where in a one-dimensional
setting we used Newton’s method to try to find the point where f 0 (x) = 0, and
ended up reducing x by f 0 (x)/f 00 (x). The formula dk = −(Hf (xk ))−1 ∇f (xk ) is
analogous: the gradient is the generalization of the first derivative; the Hessian
is the generalization of the second derivative; and multiplying by the inverse
of a matrix is the generalization of dividing by a scalar (just as multiplying
by 2−1 is the same as dividing by 2). Indeed, in the single-dimensional case,
equation (C.5) simplifies to

dk = −f 0 (xk )/f 00 (xk ) , (C.6)

so the new point xk+1 = xk + dk is given by exactly the same formula as in Step
2 of Newton’s method for line search in Section C.1.2
Newton’s method usually yields a “better” search direction, and fewer iter-
ations are required to reach the minimum. However, it can be computationally
expensive to calculate all the elements in the Hessian matrix, and to calculate its
inverse. For this reason, there are a variety of “quasi-Newton” methods, which
replace H −1 in equation (C.5) by another matrix which is easier to calculate,
but is approximately the same.
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 535

C.3.5 Step size

There are several procedures available to determine the step size αk . The sim-
plest is to use a constant step size, where at each iteration αk = α, for some
pre-determined constant α. If the predetermined value is too big, the algorithm
may overshoot the minimum, slowing convergence or even diverging. If the step
size is too small, the algorithm may take a long time to converge. If you are
determining the search direction using Newton’s method, a choice of α = 1 may
actually perform quite well, since this will move you exactly to the minimum
point of the quadratic approximation. This choice will probably not work well
for gradient descent, since the “units” of the gradient are not the same as for
the decision variables. So, if you use this approach, you should have a good
intuition about the objective function.
A more sophisticated method is exact line search, where you choose the
value of αk so that f (xk + αk dk ) is minimized. This is a one-dimensional
optimization problem, which can be solved using the methods of the previous
section.1 While this choice yields the greatest improvement in the objective
function at each iteration, it can be computationally expensive to compute this
minimum. (You would have to use the line search method at every iteration.)
A compromise between the two is a backtracking line search , which also
tries to minimize f (xk + αk dk ), but in a loose way. Rather than hoping to find
the exact minimum, we choose a value of αk which is “good enough.” Select
a positive initial step size η, and two other constants β and γ, both strictly
between 0 and 1. We then test the sequence of αk values αk = η, βη, β 2 η, . . .,
stopping as soon as this inequality is satisfied:

f (xk + αk dk ) − f (xk ) < γαk ∇f (xk )T dk (C.7)

The first value η, βη, β 2 η, . . . which satisfies this equation is chosen for αk . The
intuition in formula (C.7) is that the left-hand side shows how much f changes
if we take a step of size αk . We are trying to solve a minimization problem,
so hopefully f decreases, and the left-hand side is negative. On the right-hand
side, αk ∇f (xk )T dk is how much we would expect f to decrease based on its
linear approximation. Likewise, since dk is a direction in which f decreases, the
right-hand side is also a negative number. We stop at the first choice of αk for
which the actual decrease in the objective function is at least a certain fraction
γ of what we would expect from the linear approximation; this is exactly what
the condition (C.7) checks. Typical values of the constants in this method are
η = 1, β = 1/2, and γ = 1/10, but you should experiment with different values
for your specific problem. This rule is often called the Armijo rule.

C.3.6 Examples
Example C.13. Apply the unconstrained optimization algorithm to the func-
tion f (x) = x2 − 10x + 20. Terminate when |xk − xk−1 | < 0.01, use gradient
1 You will need to impose an upper bound on α to do this, but in practice this is not
k
usually very hard.
536 APPENDIX C. OPTIMIZATION TECHNIQUES

descent for the direction, and use a constant step size of α = 1. Repeat with a
constant step size of α = 0.1.
Solution. From basic calculus we know that the minimum of the above
function is at x∗ = 5. For this example, we pick an initial value of x0 = 15. At
any point xk , the descent direction is dk = −∇f (xk ) = 2xk − 10. Then, using
a constant step size of α = 1, we have the following:
At the initial point, the descent direction is d0 = −20, so x1 = 15 − 1 × 20 =
−5. Proceeding to the next iteration, we check the termination criterion. Since
|x1 −x0 | > 0.01, we continue. The new descent direction is d1 = −20. Therefore,
x2 = −5 + 1 × 20 = 15. Since |x2 − x1 | > 0.01, we proceed to the next iteration.
But we’ve returned to our initial point! Notice that the solutions will continue
to oscillate between 15 and −5, due to the large step size.
Repeating using a smaller constant step size of α = 0.1 produces conver-
gence: from the initial point, the descent direction is d1 = −20 and x1 =
15 − 0.1 × 20 = 13. We have |x1 − x0 | > 0.01, so we continue. At this new
point, the descent direction is d1 = −16. Therefore, x2 = 13 − 0.1 × 16 = 11.4.
Subsequent iterations are shown in Table C.3. 
In this example, the constant step size plays a major role in how quickly
we converge (if at all). For reference, Table C.4 shows how many iterations are
required for different choices. Similarly, the initial value chosen, and the param-
eters η, β, γ of an inexact line search play an important role in convergence.
Example C.14. Apply the unconstrained optimization algorithm to the func-
tion f (x, y) = (x − 1)4 + 5(y − 2)4 + xy. Terminate when ||xk − xk−1 || < 0.01,
and use gradient descent. First use the algorithm with a constant step size of
α = 0.0025, then solve again using backtracking line search with η = 1, β = 0.1,
and γ = 0.1.
Solution. Notice that there are two decision variables; we will use the
vector x = (x, y) to describe both decision variables together. For a specific
iteration, we will let their values be given by xk = (xk , yk ).
For this example, we pick an initial value of (x1 , y1 ) = (4, 4). At any point
k, the descent direction is
−4(xk − 1)3 − yk
 
dk = −∇f (xk , yk ) = .
−20(yk − 2)3 − xk
So from the initial point, the descent direction is
−4(x0 − 1)3 − y0
   
−112
d0 = −∇f (x0 , y0 ) = =
−20(y0 − 2)3 − x0 −164
and the new point is
       
x1 4 −112 3.72
x1 = = + 0.0025 = .
y1 4 −164 3.59
p
For convergence, we check if (x1 − x0 )2 + (y1 − y0 )2 < 0.01. This is not true,
so we increase k to 1 and move to the next iteration.
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 537

Table C.3: Gradient descent applied to f (x) = x2 − 10x + 20, with a constant
step size αk = 0.1
k xk f (xk )
0 15 100
1 13.0 64.0
2 11.4 40.96
3 10.12 26.2144
4 9.096 16.7772
5 8.2768 10.7374
6 7.6214 6.8719
7 7.0972 4.3980
8 6.6777 2.8147
9 6.3422 1.8014
10 6.0737 1.1529
11 5.8590 0.7379
12 5.6872 0.4722
13 5.5498 0.3022
14 5.4398 0.1934
15 5.3518 0.1238
16 5.2815 0.0792
17 5.2252 0.0507
18 5.1801 0.0325
19 5.1441 0.0208
20 5.1153 0.0133
21 5.0922 0.0085
22 5.0738 0.0054
23 5.0590 0.0035
24 5.0472 0.0022
25 5.0378 0.0014

Table C.4: Number of iterations for gradient descent to minimize f (x) = x2 −

10x + 20 with different step sizes
αk Iterations needed
0.1 27
0.2 14
0.3 9
0.4 7
0.5 3
0.6 7
0.7 10
538 APPENDIX C. OPTIMIZATION TECHNIQUES

The new descent direction is

−4(x1 − 1)3 − y1
   
−84.04
d1 = −∇f (x1 , y1 ) = =
−20(y1 − 2)3 − x1 −84.11

and the new point is

       
x 3.72 −84.08 3.51
x2 = 2 = + 0.0025 = .
y2 3.59 −84.11 3.38
p
Since (x2 − x1 )2 + (y2 − y1 )2 > 0.01, we increase k to 2 and continue. The
algorithm converges after 64 iterations, at the point (1.66, 2.23). The objective
function has a value of 3.91.
Solving again using backtracking line search, we have the same initial points
and directions: x0 = (4, 4) and d0 = (−112, −164). However, we have to do a
little more work to determine the step size α0 . We start by testing α0 = η = 1.
With this choice, the new solution would be
     
4 −112 −108
+1 = .
4 −164 −160

The value of the objective function here is very large, approximately 3.58 × 109 ,
so the left-hand side of (C.7) is 3.58 × 109 − 177 ≈ 3.58 × 109 . For the right
hand side, we calculate
 
t
  −112
γα0 ∇f (x0 ) d0 = (0.1)(1) 112 164 = −3944 .
−164

Inequality (C.7) is clearly false (the left-hand side is very positive, the right-
hand side is negative), so we try again with α0 = βη = 1/10. This choice
corresponds to the solution
     
4 1 −112 −7.2
+ = .
4 10 −164 −12.4

The objective function has a value of 2.19 × 105 , so the left-hand side of (C.7)
is still approximately 2.19 × 105 , while the right-hand side is
 
t 1   −112
γα0 ∇f (x0 ) d0 = 112 164 = −394.4 .
100 −164

which is again false. Trying again with α0 = β 2 η = 1/100, the new solution is
x = (2.88, 2.36), and the left and right-hand sides of (C.7) are now −157.6 and
−39.4, respectively. So we accept this choice: α0 = 1/100, and x1 = (2.88, 2.36).
Proceeding similarly, you can verify that the next step sizes are α1 = 1/10,
α2 = 1/10, α3 = 1, and α4 = 0.1, with the algorithm terminating after that
step at the solution x5 = (0.236, 1.774), with objective value 0.772. Notice that
each iteration of backtracking line search required more work, but in the end
we only had to perform five iterations, rather than sixty-four. 
C.3. UNCONSTRAINED NONLINEAR OPTIMIZATION 539

Example C.15. Apply the unconstrained optimization method to the function

f (x, y) = (x − 1)4 + 5(y − 2)4 + xy. Terminate when ||xk − xk−1 || < 0.01, and
use Newton’s method with a constant step size of α = 1.
Solution. We will use the same initial value x0 = (4, 4). As before, the
gradient at an arbitrary point is
4(xk − 1)3 + yk
 
∇f (xk , yk ) = ,
20(yk − 2)3 + xk
and the Hessian is
12(xk − 1)2
 
1
Hf (xk , yk ) = .
1 60(yk − 2)2
So, at the initial point we have
4(4 − 1)3 + 4
   
112
∇f (4.0, 4.0) = =
20(4 − 2)3 + 4 164
and  
108 1
Hf (4.0, 4.0) = .
1 240
The search direction from Newton’s method is
 
−1 −1.0307
d0 = (Hf (x1 )) ∇f (x0 ) = .
−0.6790
Using the constant step size α = 1, the new point is
       
x1 4 −1.0307 2.9692
= +1 =
y1 4 −0.6790 3.3209
p
Since (x1 − x0 )2 + (y1 − y0 )2 > 0.01, we increment k to 1 and move to the
next iteration.
At (x1 , y1 ) = (2.9692, 3.3209), the gradient and Hessian are
4(2.9692 − 1)3 + 3.3209
   
38.8672
∇f (2.9692, 3.3209) = =
20(3.3209 − 2)3 + 2.9692 49.069
and  
46.5353 1
Hf (2.9692, 3.3209) = ,
1 104.696
respectively. So the search direction is
 
−0.7178
d1 = −(Hf (x1 ))−1 ∇f (x1 ) =
−0.4618
and        
x2 2.9692 −0.7178 2.2513
= +1 = ,
y2 3.3209 −0.4618 2.8591
and so forth. The algorithm terminates in iteration 16, to the solution x =
(0.2399, 1.756), with an objective function value of 0.7728. 
540 APPENDIX C. OPTIMIZATION TECHNIQUES

C.4 Constrained Nonlinear Optimization

The main text derived optimality conditions for a specific class of nonlinear opti-
mization problems including the static traffic assignment problem (Section 3.3).
Likewise, Sections 6.2.1 and 6.2.2 introduced the method of successive averages
and Frank-Wolfe algorithms for solving traffic assignment. This section discusses
how these techniques can be applied to other kinds of nonlinear optimization
problems involving linear constraints.
In this section, we will consider optimization problems of the form

min f (x)
x=(x1 ,...,xn )
s.t. gi (x) ≤ 0 ∀i = 1, 2, . . . , m
hj (x) = 0 ∀j = 1, 2, . . . , ` ,

where the gi and hj are functions representing the inequality and equality
constraints, respectively. We will concisely denote the feasible region by X.
That is, X is the set

X = {x ∈ Rn : gi (x) ≤ 0 ∀i = 1, 2, . . . , m; hj (x) = 0 ∀j = 1, 2, . . . , `} .

We first derive optimality conditions that apply to a general constrained non-

linear optimization problem, and then specialize them to the case where all of
the constraints are linear and the objective function is convex.

C.4.1 Optimality conditions

As with unconstrained optimization, there are first- and second-order necessary
and sufficient conditions characterizing the local minima of constrained opti-
mization problems. The main distinction is that optimal solutions can now lie
along the boundary of the feasible region, as well as in its interior. It is pos-
sible to create examples where constraints behave irregularly. To address this,
we will need to define constraint qualifications, which are essentially regularity
conditions on the constraints. We begin with some definitions leading up to
these.
Definition C.2. For any feasible solution x ∈ X, an active constraint is an
inequality constraint which is satisfied with equality. That is, the i-th inequality
constraint is active if gi (x) = 0.
Definition C.3. For any feasible solution x ∈ X, the active constraint set is
the set of all active constraints at x: A(x) = {i : gi (x̂) = 0}.
We now introduce several constraint qualifications, which we will discuss
below.
Definition C.4. The linear independence constraint qualification holds at a
feasible solution x ∈ X if the gradient vectors ∇gi (x) for active constraints
i ∈ A(x), and for all equality constraints ∇hj (x), are linearly independent.
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 541

Definition C.5. The Mangasarian-Fromovitz constraint qualification holds at

a feasible solution x ∈ if (i) the gradient vectors ∇hj (x̂) are linearly independent,
and (ii) there exists a direction vector d ∈ Rn such that ∇gi (x̂)T d < 0 for every
i ∈ A(x) and ∇hj (x̂)T d = 0 for all j = 1, . . . , `.

The linear independence constraint qualification is stronger, and you can

show that whenever it holds, the Mangasarian-Fromovitz constraint qualifica-
tion also holds. There are other constraint qualifications as well. If the linear
independence constraint qualification holds, then most other constraint quali-
fication conditions hold as well. In most nonlinear optimization formulations
representing traffic assignment, all of the constraints are linear, and a simpler
constraint qualification holds:

Definition C.6. The linearity constraint qualification holds if all of the func-
tions gi and hj defining the constraints are affine functions.

Perhaps the most famous necessary optimality conditions for constrained

optimization are the Karush-Kuhn-Tucker (KKT) conditions. They make use
of auxiliary variables µ and λ, often called Lagrange multipliers.

Theorem C.4. (First-order necessary conditions for a local minimum.) Let

f, gi , hj be continuously differentiable functions. If (i) x∗ is a local minimum of
f , and (ii) any of the above constraint qualifications hold at x∗ , then there exist
scalars µi , i = 1, 2, . . . , m and λj , j = 1, . . . , `, such that the following system
holds:
m
X `
X
∇f (x∗ ) + µi ∇gi (x∗ ) + λj ∇hj (x∗ ) = 0 (C.8)
i=1 j=1

hj (x∗ ) = 0 ∀j = 1, 2, . . . , ` (C.9)
gi (x∗ ) ≤ 0 ∀i = 1, 2, . . . , m (C.10)
µi gi (x∗ ) = 0 ∀i = 1, 2, . . . , m (C.11)
µi ≥ 0 ∀i = 1, 2, . . . , m (C.12)

The above conditions can also be written in matrix form. Let µ and λ be
the vectors whose components are µ1 , µ2 , . . . , µm and λ1 , λ2 , . . . , λ` , and likewise
let g(x) and h(x) be vectors with components gi (x) and hj (x). Then we can
rewrite the KKT conditions in terms of the Jacobians of g and h as

∇f (x∗ ) + Jg(x∗ )µ + Jh(x∗ )λ = 0

h(x∗ ) = 0
g(x∗ ) ≤ 0
µT g(x∗ ) = 0
µ≥0
542 APPENDIX C. OPTIMIZATION TECHNIQUES

Another way of representing the first order necessary condition is using the
Lagrangian function
m
X l
X
L(x, µ, λ) = f (x∗ ) + µi gi (x∗ ) + λj hj (x∗ )
i=1 j=1

We can rewrite the KKT functions in a simpler way using the Lagrangian.
Taken as a vector, the partial derivatives of L with respect to x (written as
∇x L) form the left-hand side of the first KKT condition (C.8). Similarly, the
partial derivatives of L with respect to λ form the left-hand side of the second
KKT condition (C.9). So we can write the KKT conditions as:

∇x L(x∗ , µ, λ) = 0
∇λ L(x∗ , µ, λ) = 0

These are necessary conditions, meaning that they must be satisfied at any
optimal solution. Without additional restrictions, they are not sufficient, mean-
ing that there may be non-optimal points which also satisfy those conditions.
Still, we can identify all the points which satisfy the KKT conditions to gener-
ate a set of “candidate solutions;” if an optimal solution exists it must be one
of them. We give some examples of how to do this in the following subsection.
For large-scale problems this is not a practical approach (the following subsec-
tion describes some methods that can be used), but it may be reasonable for
problems with only a few decision variables and constraints, or where there is a
special structure which further simplifies these conditions.
With additional conditions on the objective function and constraints, we
can give stronger results based on the KKT conditions: For instance, under
the linear independence constraint qualification, there are unique vectors of
Lagrange multipliers µ and λ satisfying the KKT conditions at a local minimum.
The KKT conditions can also be sufficient, under additional restrictions. For
example, by imposing convexity conditions on f and the gi , and linearity on the
hj , we have this result:
Theorem C.5. (First-order sufficient conditions for global minima.) Let f and
all gi be continuously differentiable convex functions, and let all hj be linear
functions. Then if there are x∗ ∈ Rn , µ ∈ Rm , and λ ∈ R` satisfying the
following conditions, x∗ is a global minimum of f subject to x ∈ X.
m
X `
X
∇f (x∗ ) + µi ∇gi (x∗ ) + λj ∇hj (x∗ ) = 0 (C.13)
i=1 j=1

hj (x∗ ) = 0 ∀j = 1, 2, . . . , ` (C.14)
∗
gi (x ) ≤ 0 ∀i = 1, 2, . . . , m (C.15)
∗
µi gi (x ) = 0 ∀i = 1, 2, . . . , m (C.16)
µi ≥ 0 ∀i = 1, 2, . . . , m (C.17)
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 543

In this result, the conditions on gi and hj imply that the feasible region X
is a convex set; since f is also a convex function, we have a convex optimization
problem, and therefore any local minimum is also global. This is what allowed
us to convert the “local” necessary condition into a global sufficient condition.
As with unconstrained problems, we can also formulate a second-order suf-
ficient condition, based on second partial derivatives.

Theorem C.6. (Second-order sufficient conditions for a local minimum.) If

the objective f and all gi and hj defining the constraints are twice continuously
differentiable; the Hessian of L with respect to x is positive definite at x∗ , µ,
and λ; and µi > 0 for any i ∈ / A(x∗ ); then any solution satisfying the KKT
conditions is a strict local minimum of f .

This condition can be relaxed slightly. Rather than requiring that the ma-
trix Hxx consisting of all second partial derivatives ∂ 2 L/∂xi ∂xj be positive
definite (requiring dT Hxx (x∗ )d > 0 for all nonzero d ∈ Rn ), it is enough if
dT Hxx (x∗ )d > 0 holds for any nonzero d ∈ Rn such that ∇hj (x∗ )T d = 0 for
all j, and ∇gi (x∗ )T d = 0 for all active i ∈ A(x∗ ).

C.4.2 Examples of applying KKT conditions

This section shows how the KKT conditions can be directly solved to identify
candidate solutions, and how to then identify an optimal solution among these
candidates. Again, this approach is not practical for large-scale problems, but
is good for small problems, for building intuition about how the conditions
work, and on occasion in problems with a special structure (for instance, if the
conditions can be solved in closed form).

Example C.16. Solve the following optimization problem:

min x2
s.t. (x1 − 1)2 + x22 ≤ 1
x1 ≥ 2

Are the KKT conditions satisfied here?

Solution. In this formulation, there is only a single feasible point x =

(2, 0), so this solution must be optimal (both locally and globally). To examine
the KKT conditions at this point, we start by writing the two constraints can
be written as

g1 (x1 , x2 ) = (x1 − 1)2 + x22 − 1 ≤ 0g2 (x1 , x2 ) = −x1 + 2 ≤ 0,

and so their gradients are

   
2(x1 − 1) −1
∇g1 (x1 , x2 ) = ∇g2 (x1 , x2 ) = .
2x2 0
544 APPENDIX C. OPTIMIZATION TECHNIQUES

The gradient of the objective function is

 
0
∇f (x1 , x2 ) = .
1

From the first-order necessary conditions, the minimum should satisfy

m
X `
X
∇f (x∗ ) + µi ∇gi (x∗ ) + λj ∇hj (x∗ ) = 0
i=1 j=1

For this particular example, this equation reduces to

∇f (x∗ ) + µ1 ∇g1 (x∗ ) + µ2 ∇g2 (x∗ ) = 0

or    ∗   
0 2(x1 − 1) −1
+ µ1 + µ2 = 0.
1 2x∗2 0
At (x∗1 , x∗2 ) = (2, 0), we have
     
0 2 −1
+ µ1 + µ2 = 0.
1 0 0

But there are no values of u1 and u2 for which the above equation is satisfied.
So it is impossible to satisfy the KKT conditions at this point, even though it
is optimal. How can this be? The answer is that the constraint qualifications
are not satisfied at this point. The first constraint is not linear, so the linearity
constraint qualification fails. The gradients of the two constraints are (2, 0)
and (−1, 0), which are linearly dependent, so the linear independence constraint
qualification also fails. Similarly, you can show that the Mangasarian-Fromovitz
constraint qualification fails at this point. 
This example highlights the importance of constraint qualifications. How-
ever, in traffic assignment, a majority of the formulations will have linear con-
straints, in which case the linear constraint qualification holds. In such cases,
you do not have to worry further.
Note that convexity alone does not imply constraint qualification. In the
above example both g1 and g2 are convex functions, so both the objective func-
tion and constraints are convex functions, so this is a convex optimization prob-
lem.

Example C.17. Solve the following optimization problem, where a, b, and c

are positive real numbers.

min x1 + x2 + x3
x21 x2 x2
s.t. + 2+ 3 =1
a b c
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 545

Solution. We have f (x1 , x2 , x3 ) = x1 + x2 + x3 and a single equality

x2 x2 x2
constraint h(x1 , x2 , x3 ) = a1 + b2 + c3 − 1. Calculating gradients, we have
   2x1 
1 a
∇f (x1 , x2 , x3 ) = 1 ∇h(x1 , x2 , x3 ) =  2xb 2  ,
2x3
1 c

and the condition

∇f (x∗1 , x∗2 , x∗3 ) + λ∇h(x∗1 , x∗2 x∗3 ) = 0

becomes    2x∗ 
1
1 a∗
2x
1 + λ  2  = 0.

b 
1 2x∗3
c

Solving, we get x∗1 = a

− 2v , x∗2 b
= − 2v , and x∗3 = − 2vc
.
∗ ∗ ∗
We also have the condition h(x1 , x2 , x3 ) = 0. For this problem, we compute
as follows:
(x∗1 )2 (x∗ )2 (x∗ )2
∇h(x∗1 , x∗2 , x∗3 ) = 0 =⇒ + 2 + 3 =1
a b c
a b c
=⇒ + 2 + 2 =1
4v 2 √ 4v 4v
± a+b+c
=⇒ v =
2
Therefore, we have two possible solutions for (x∗1 , x∗2 , x∗3 ):
 
a b c
√ ,√ ,√
a+b+c a+b+c a+b+c
and  
a b c
−√ , −√ , −√ .
a+b+c a+b+c a+b+c
Checking both, the objective function is minimized at the second point, so this
is optimal. 

Example C.18. Write down the KKT conditions for the following optimization
problem.
min x21 + x22
s.t. 11x1 + 3x2 ≥ 21
6x1 + 20x2 ≥ 39
x1 + x2 ≤ 9
x1 ≥ 0
x2 ≥ 0
546 APPENDIX C. OPTIMIZATION TECHNIQUES

Solution. Since all constraints are linear, the linear constraint qualification
is satisfied and the KKT conditions are indeed necessary. We can rewrite the
optimization problem in the form needed for the KKT conditions:

min x21 + x22

s.t. g1 (x1 , x2 ) = −11x1 − 3x2 + 21 ≤0
g2 (x1 , x2 ) = −6x1 − 20x2 + 39 ≤0
g3 (x1 , x2 ) = x1 + x2 − 9 ≤0
g4 (x1 , x2 ) = −x1 ≤0
g5 (x1 , x2 ) = −x2 ≤ 0.

We calculate gradients as follows:

     
2x1 −11 −6
∇f (x1 , x2 ) = ∇g1 (x1 , x2 ) = ∇g2 (x1 , x2 ) =
2x2 −3 −20
     
1 −1 0
∇g3 (x1 , x2 ) = ∇g4 (x1 , x2 ) = ∇g5 (x1 , x2 ) = .
1 0 −1
With these gradients, the first KKT condition (C.8) reduces to the two
equations

2x∗1 − 11µ1 − 6µ2 + µ3 − µ4 = 0

2x∗2 − 3µ1 − 20µ2 + µ3 − µ5 = 0 .

The second KKT condition (C.9) is ignored since there are no equality con-
straints. The third KKT condition (C.10) ensures feasibility:

g1 (x∗1 , x∗2 ) ≤ 0 =⇒ −11x∗1 − 3x∗2 + 21 ≤ 0

g2 (x∗1 , x∗2 ) ≤ 0 =⇒ −6x∗1 − 20x∗2 + 39 ≤ 0
g3 (x∗1 , x∗2 ) ≤ 0 =⇒ x∗1 + x∗2 − 9 ≤ 0
g4 (x∗1 , x∗2 ) ≤ 0 =⇒ −x∗1 ≤ 0
g5 (x∗1 , x∗2 ) ≤ 0 =⇒ −x∗2 ≤ 0 .

The fourth ensures complementarity (C.11), that the µi values must be zero
unless the constraint is active:

µ1 g1 (x∗1 , x∗2 ) = 0 =⇒ µ1 (−11x∗1 − 3x∗2 + 21) = 0

µ2 g2 (x∗1 , x∗2 ) = 0 =⇒ µ2 (−6x∗1 − 20x∗2 + 39) = 0
µ3 g3 (x∗1 , x∗2 ) = 0 =⇒ µ3 (x∗1 + x∗2 − 9) = 0
µ4 g4 (x∗1 , x∗2 ) = 0 =⇒ µ4 (−x∗1 ) = 0
µ5 g5 (x∗1 , x∗2 ) = 0 =⇒ µ5 (−x∗2 ) = 0

Finally, we require that the Lagrange multipliers all be non-negative:

µ1 , µ2 , µ3 , µ4 , µ5 ≥ 0 .
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 547

 If desired, the optimum solution can be found by finding values of x and µ

satisfying all these conditions. One way to do this is to start by identifying all 32
combinations of active/inactive constraints. For each, set µi = 0 for the inactive
constraints, and solve equations (C.8) and (C.11) for the remaining variables as
a system of equations. If the solution satisfies the feasibility conditions (C.10)
and has non-negative Lagrange multipliers, it satisfies all KKT conditions, and
is a candidate solution. Among these, the solutions with the lowest objective
function value are the optima.
We illustrate this procedure in the following example, which has fewer con-
straints:
Example C.19. Solve the following optimization problem.
min x21 + 2x22
s.t. x1 + 2x2 ≤6
x2 ≥ x1 + 2

Solution. Converting to the form required by the KKT conditions, we

have f (x1 , x2 ) = x21 + 2x22 , and constraints g1 (x1 , x2 ) = x1 + 2x2 − 6 and
g2 (x1 , x2 ) = x1 − x2 + 2, with gradients
     
2x1 1 1
∇f (x1 , x2 ) = ∇g1 (x1 , x2 ) = ∇g2 (x1 , x2 ) = .
4x2 2 −1
The KKT conditions for this problem are
2x∗1 + u1 + µ2 = 0
4x∗2 + 2u1 − µ2 = 0

x∗1 + 2x∗2 − 6 ≤ 0
x∗1 − x∗2 + 2 ≤ 0

µ1 (x∗1 + 2x∗2 − 6) = 0
µ2 (x∗1 − x∗2 + 2) = 0

µ1 , µ2 ≥ 0 .
The functions defining the constraints are linear, so the KKT conditions are
indeed necessary. We now find candidate solutions by considering all combina-
tions of active and inactive constraints.
Case I: Both constraints inactive:
In this case µ1 = µ2 = 0, so the equality conditions reduce to
2x∗1 = 0
4x∗2 = 0
.
548 APPENDIX C. OPTIMIZATION TECHNIQUES

The solution is (x∗1 , x∗2 ) = (0, 0). However (0, 0) violates the constraint g2 (x1 , x2 ) =
x∗1 − x∗2 + 2 ≤ 0. So, there is no optimum solution where both constraints are
inactive.
Case II: Only the second constraint is active:
In this case µ1 = 0. The equality conditions reduce to
2x∗1 + µ2 = 0
4x∗2 − µ2 = 0
µ2 (x∗1 − x∗2 + 2) = 0 .
The first two equations give x∗1 = −µ2 /2 and x∗2 = µ2 /4. Because the second
constraint is active, we know x∗1 − x∗2 + 2 = 0. Therefore, the third equation
will be satisfied automatically. Substituting x∗1 = −µ2 /2 and x∗2 = µ2 /4 into
the active constraint, we have µ2 = 8/3, and therefore x∗1 = 4/3 and x∗2 = 2/3.
This solution satisfies both constraints, and the Lagrange multipliers are
non-negative, so it is a candidate for optimality.
Case III: Only the first constraint is active:
In this case µ2 = 0, and the equality conditions are:
2x∗1 + µ1 = 0
4x∗2 + 2µ1 = 0
µ1 (x∗1 + 2x∗2 − 6) = 0

Proceeding in the same way, the first two equations require x∗1 = −µ1 /2 and
x∗2 = −µ2 /2. Likewise, because we assume the first constraint is active, we can
replace the third equation with x∗1 + 2x∗2 = 6. But with the values of x∗1 and x∗2
for this case, this simplifies to −µ1 = 6. This violates the requirement µ1 ≥ 0,
and therefore there cannot be an optimal solution corresponding to this case.
Case IV: Both constraints are active:
In this case, we know both constraints are satisfied with equality:
x∗1 + 2x∗2 = 6
x∗1 − x∗2 = −2
.
The only solution to these equations is x∗1 = 2/3 and x∗2 = 8/3. It remains
to see whether there are Lagrange multipliers satisfying the rest of the KKT
conditions. Substituting into the first two conditions, we have
4
µ1 + µ2 = −2x∗1 = −
3
∗ 32
2µ1 − µ2 = −4x2 = − .
3
Solving this system, we find µ1 < 0, violating the non-negativity condition,
and establishing that the solution where both constraints are active cannot be
optimal.
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 549

We have exhausted all combinations of active and inactive constraints, and

are left with a single candidate solution (from Case II). Therefore, this one is
the optimum, and the solution to the problem is x∗1 = − 43 and x∗2 = 23 . If there
were multiple candidate solutions, we would evaluate the objective function at
each to see which one(s) are minima. 

C.4.3 Algorithms for linear constraints

The remainder of this section will describe four algorithms for solving nonlinear
optimization problems with linear constraints. The first two algorithms are the
method of successive averages (MSA) and Frank-Wolfe (FW) algorithm. Both
of these are called convex combination methods. The third and fourth meth-
ods, gradient projection (GP) and manifold suboptimization (MS), modify the
direction of steepest descent to ensure feasibility. The algorithms are presented
in increasing order of complexity and efficiency. All of these algorithms will
converge to a local optimum. If the objective is convex, this local optimum will
also be globally optimal.
We focus here on nonlinear optimization problems with linear constraints, as
the traffic assignment problem have this structure. These four algorithms have
all been applied to traffic assignment as well. In Chapter 6, these algorithms
are presented specifically in this context. Here we treat them more generally,
as they can be used for other nonlinear optimization problems as well.
There are many ways to modify these algorithms, and the best implementa-
tion often depends on the specific problem structure and the computing envi-
ronment being used. For example, all four methods have the common structure
of “find an improving direction” and “choose a step size in that direction.”
This section mainly emphasizes the former, since the algorithms differ more in
this regard. However, there are alternative choices for the latter as well. Our
presentations of FW, PG, GP, and MS in this subsection choose the step size
which minimizes the objective function along the chosen direction. This indeed
results in the greatest improvement possible at each iteration. However, this
can be computationally demanding, and it might be more efficient to take a
faster but less exact step. For instance, the algorithms in Chapter 6 often use
a single iteration of Newton’s method in place of finding an exactly-optimal
step size, since this tradeoff is usually worthwhile in traffic assignment. If you
are comparing the presentations of the algorithms in that chapter which are
specialized to traffic assignment to the more general presentations here, such
differences should be noted, but they are matters of implementation and not
essential algorithmic distinctions.
The standard form of the nonlinear optimization problem with linear con-
straints is as follows. The decision variables are x = x1 x2 . . . xn . The ob-
jective function f is a convex function of x. The feasible region is given as
X = {x ∈ Rn : Ax − b ≤ 0 ∀i ∈ {1, 2, . . . , m}; Cx − d = 0 ∀i ∈ {1, 2, . . . , `}} .
With linear constraints such as these, the feasible region is convex and closed.
We further assume that it is bounded, as is the case in traffic assignment and
many other applications.
550 APPENDIX C. OPTIMIZATION TECHNIQUES

Therefore, this is the form of the optimization problems we will consider in

this section:
min f (x)
s.t. Ax − b ≤ 0 ∀i ∈ {1, 2, . . . , m}
Cx − d = 0 ∀i ∈ {1, 2, . . . , `}

C.4.4 Convex combination methods

All convex combinations methods can be described by the following framework
(see Algorithm 1). They have three main steps: (i) determining a feasible
descent direction, (ii) determining the step size, and (iii) updating the solution.
The convergence criterion for feasible direction methods can be chosen in a
similar manner as for unconstrained optimization algorithms.
Initialize:
k ← 1: iteration counter
x1 : initial value
: tolerance
while ConvergenceCriterion >  do
Determine feasible descent direction: pk
Determine step size: αk ≥ 0 such that f (xk + αk pk ) < f (xk )
and xk + αk pk ∈ X
Update: xk+1 ← xk + αk pk
k ←k+1
end
return xk
Algorithm 1: Feasible directions method
At iteration k, when the current solution xk , the feasible descent direction
pk is obtained by taking pk = yk − xk , where yk is a feasible solution ob-
tained by solving an approximation of the original nonlinear problem. Both
MSA and FW obtain the feasible direction by replacing the objective function
with its first-order (linear) approximation at xk . Because the constraints are
linear, the resulting first-order approximation will be the linear program shown
below, which can be solved using the simplex algorithm. (In some applica-
tions, the first-order approximation can be solved more efficiently by exploiting
a particular structure of the problem. For example, in traffic assignment, one
can show that an all-or-nothing assignment to shortest paths is optimal to this
subproblem, which can be found using one of the algorithms in Section 2.4.

min ∇f (xk )T y
s.t. Ax − b ≤ 0 ∀i ∈ {1, 2, . . . , m}
Cx − d = 0 ∀i ∈ {1, 2, . . . , `}

The difference between MSA and FW is how the step size is chosen In MSA,
at each iteration k, the step size is commonlyPchosen to be αP k = 1/(k + 1),
although any sequence of step sizes satisfying αk = ∞ and αk2 < ∞ will
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 551

work. In FW, the step size αk is obtained by solving the following optimization
problem in λ. This optimization problem can be solved using any of the line
search methods from Sections 3.3.2 or C.1.

min f (xk + αpk )

α
s.t. 0≤α≤1

FW picks the step size in a more intelligent manner than MSA and typically
converges faster. For this reason, MSA is rarely used in solving nonlinear opti-
mization with linear constraints. However, in certain complex traffic assignment
problems, evaluating the objective function f can be computationally expensive.
In such cases, a line search method in FW can require large amounts of time,
and taking faster MSA steps outweighs the greater precision that each FW step
would have.
Example C.20. Determine the optimal solution of the following optimization
problem using MSA and FW.
min 2(x1 − 1)2 + (x2 − 2)2
s.t. x1 + 4x2 − 2 ≤0
−x1 + x2 ≤0
x1 , x2 ≥0

Solution. The KKT conditions for this problem can be solved to yield
the optimal solution x∗1 = 0.7878, x∗2 = 0.3030. The MSA and FW algorithms
will converge to this solution over successive iterations; the advantage of these
methods is that solving the KKT conditions for a large problem can be very
difficult, whereas MSA and FW scale better with problem size. For this objective
function, the gradient at any point (x1 , x2 ) is
 
4(x1 − 1)
∇f (x1 , x2 ) = .
2(x2 − 2)
MSA: Assume the initial solution is x11 = 0, x12 = 0. The search direction
is obtained by solving the following linear program.
min −4y1 − 4y2
s.t. y1 + 4y2 − 2 ≤0
−y1 + y2 ≤0
y1 , y2 ≥0
The optimal solution to this
 problem
 is y11 = 2, y21 = 0. Therefore, the search
1 1 1
direction is p = y − x = 2 0 . Therefore
1
x21 = 0 + (2) = 1
2
1
x22 = 0 + (0) = 0
2
552 APPENDIX C. OPTIMIZATION TECHNIQUES

Table C.5: Convergence of the method of successive averages.

k x1 x2 Error
1 0 0 0.7124
2 1 0 0.1368
3 0.8 0.1333 0.0289
4 0.7 0.2 0.01831
5 0.96 0.16 0.0501
6 0.866 0.2 0.0168
7 0.800 0.2285 0.005
20 0.780 0.28 0.0005
53 0.7849 0.2943 8.33793 ×10−5

At the start of the second iteration we have x21 = 1, x22 = 0. Therefore, the
search direction is obtained by solving the following linear program.

min −4y2
s.t. y1 + 4y2 − 2 ≤0
−y1 + y2 ≤0
y1 , y2 ≥0

The optimal solution to this linear program is y12 = 0.4, y22 = 0.4. In iteration
2, the step size is α2 = 31 . Therefore,
1
x31 = 1 + (0.4 − 1) = 0.8
3
3 1
x2 = 0 + (0.4 − 0) = 0.1333
3
Table C.5 shows the progress of MSA over further iterations. The error in this
table is calculated as the sum of square deviation from the optimal solution
which we know to be x∗1 = 0.7878, x∗2 = 0.3030.
FW: Again assume the initial solution to x11 = 0, x12 = 0. The search
direction is obtained by solving the same linear program as in MSA.

min −4y1 − 4y2

s.t. y1 + 4y2 − 2 ≤0
−y1 + y2 ≤0
y1 , y2 ≥0

The optimal solution

 to the LP is again y11 = 2, y21 = 0, so the direction is the
1
same: p = 2 0 . In FW, the step size is obtained by solving the following
minimization problem:
2 2
min 2 (0 + α1 (2 − 0) − 1) + (0 + α1 (0 − 0) − 2)
α1
s.t. 0≤α≤1
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 553

Table C.6: Convergence of the Frank-Wolfe algorithm.

k x1 x2 Error
1 0 0 0.7124
2 1 0 0.1368
3 0.4545 0.3636 0.1147
4 0.7979 0.2828 0.0005
20 0.7919 0.2947 8.6 ×10−5

The optimal solution is α1 = 0.5, and the solution is updated as

x21 = 1 + 0.5 × (2 − 0) = 1
x22 = 0 + 0.5 × (0 − 0) = 0

To this point, FW is proceeding in the same way as MSA.

At the start of the second iteration we have x21 = 2, x22 = 0. Therefore, the
search direction is obtained by solving the following linear program.
min −4y2
s.t. y1 + 4y2 − 2 ≤0
−y1 + y2 ≤0
y1 , y2 ≥0

The optimal solution to the linear program is y12 = 0.4, y22 = 0.4, as before.
However, at this point FW chooses the step size differently than MSA. Solving
the optimization problem
2
min 2 (1 + α2 (0.4 − 1) − 1))2 + (0 + α2 (0.4 − 0) − 2)
α2
s.t. 0 ≤ α2 ≤ 1
gives the step size α2 = λ = 0.9091. Therefore, we obtain the new solution

x31 = 1 + 0.9091 × (0.4 − 1) = 0.4545

x32 = 0 + 0.9091 × (0.4 − 0) = 0.3634 .

Table C.6 shows the convergence of FW over successive iterations. Compar-

ing with Table C.5, we see that FW converges in fewer iterations than MSA.
FW achieves an error of 0.0005 in 4 iterations; MSA required 20 iterations to
achieve this level of precision. The tradeoff is that FW required more effort
at each iteration to determine the step size α. In problems such as this, the
tradeoff is clearly in favor of FW. 

C.4.5 Gradient projection

The gradient projection method operates by taking a step in the direction of
steepest descent (ignoring feasibility), and then jumping to the nearest feasible
554 APPENDIX C. OPTIMIZATION TECHNIQUES

point if the new solution is infeasible. Recall that projX (x) means “the point in
X closest to x.” (If x ∈ X already, then projX (x) = x.) If the set X is convex,
then the projection operation is uniquely defined, and a continuous function
of x. In general, projection cannot be easily computed. When the constraints
are linear (as we assume in this section), it is considerably easier. For some
problems, it can be exceptionally easy. For instance, in traffic assignment, we
can reformulate all constraints to be simple non-negativity constraints of the
form hπ ≥ 0, and “projection” simply means “any negative path flow should be
set to zero.”
The steps of gradient projection are shown in Algorithm 2.

Initialize:
k ← 1: iteration counter
x1 : initial value
: tolerance
while ConvergenceCriterion >  do
dk ← −∇f (xk )
αk ← arg minα {f (projX (xk + αdk )), 0 ≤ α ≤ ᾱ}
xk+1 ← xk + αk d
k ←k+1
end
return xk
Algorithm 2: Gradient projection method

Example C.21. Re-solve the previous optimization problem using the gradient
projection method, with ᾱ = 1.

min 2(x1 − 1)2 + (x2 − 2)2

s.t. x1 + 4x2 − 2 ≤0
−x1 + x2 ≤0
x1 , x2 ≥0

1 1
Solution. We start from the same  initial
 solution x1 = 0, x2 = 0. Here the
search direction is d1 = −∇f (x1 ) = 4 4 . To find the step size, we solve the
optimization problem

min f (projX (x1 + αd1 )

α
s.t. 0≤α≤1

The projection operation can be more easily seen on a plot of the feasi-
ble region, as in Figure C.8. For a given value of α, x1 + αd1 is the point
(4α, 4α). If α ≤ 1/10, this coincides with the line x1 = x2 , which is part
of the feasible region. As a result, projection does not change the point and
projX (4α, 4α) = (4α, 4α). Once α > 1/10, the point (4α, 4α) violates the con-
straint x1 + 4x2 ≤ 2, and is infeasible. For these points, we have to project
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 555

0
≤
2
x
+
1
x
−
1.5

x2 1 α = 1/4 → (1, 1)

0.5

x1 +
4x2 ≤
projX (1, 1) = 2
(14/17, 5/17)
0
0 0.5 1 1.5 2
x1

Figure C.8: Feasible region and projection.

them onto the nearest feasible point. As an example, if α = 1/4, the point
(4α, 4α) = (1, 1) is infeasible. The closest feasible point is (14/17, 5/17); this
can be seen geometrically in Figure C.8 by drawing a line through (1, 1) perpen-
dicular to the line corresponding to that constraint. The expression mapping a
value of α in this range to the closest feasible point can be found algebraically,
and the formula is shown below. If α > 23 , the point on the line x1 + 4x2 lies
below the x2 -axis, which is also infeasible. For α values this large, the closest
feasible point is simply (2, 0). Therefore we have

(4α, 4α)
 α ∈ [0, 1/10]
1
projX (x1 + αd1 ) = 17 (48α + 2, −12α + 8) α ∈ [1/10, 2/3] .

(2, 0) α ≥ 2/3


You can check that this formula is continuous.


556 APPENDIX C. OPTIMIZATION TECHNIQUES

Evaluating the objective function in terms of α, we thus have

f (projX (x1 + αd1 )) =


2 2
2(4α − 1) + (4α −
2)
 α ∈ [0, 1/10]
1 2 1

2
2 17 (48α + 2) − 1 + 17 (−12α + 8) − 2 α ∈ [1/10, 2/3] .

6 α ≥ 2/3


Again, you can check that this function is continuous and convex. Using a line
search method, the minimum is found for α = 0.236, which corresponds to the
point (0.7878, 0.3030). As shown above, this is actually the optimal solution,
found in a single iteration! Typically, GP only converges to the optimal solution
in the limit, but as this example shows it can be extremely efficient in terms of
iterations. The drawback is that each iteration requires a significant amount of
work. But in problems where the projection can be calculated very efficiently
(as in traffic assignment), GP can be an excellent algorithmic choice.

C.4.6 Manifold suboptimization

Manifold suboptimization works by modifying the direction of steepest descent
(related to the gradient of the objective) to ensure that it remains within the
feasible set. This method also involves projection, but unlike gradient projec-
tion, the projection is implicit through matrix multiplications which ensure that
whichever constraints are exactly satisfied at the current solution remain sat-
isfied. There are several names that are used to describe this method, which
unfortunately can cause confusion. Some authors refer to this method as the
projected gradient method (distinguishing it from gradient projection, because
the feasibility is used to identify the improving direction, rather than moving in
a direction and then checking feasibility), and others even call it gradient projec-
tion (for historical reasons). We believe that calling it manifold suboptimization
avoids this confusion, and by doing so follow the convention of Bertsekas (2016).
Intuitively, the idea is to distinguish between “active” constraints (where
the left-hand side is exactly equal to the right-hand side) and “inactive” ones
(where the left-hand side is strictly less than the right-hand side). The gradient
is modified so that any active constraints continue to hold with equality; inactive
constraints do not impose any such restriction, since their left-hand sides can
vary some without violating feasibility. Equality (=) constraints are always
active. Inequality (≤) constraints may or may not be active. As the algorithm
proceeds, which constraints are active and which are inactive may vary, and
some care must be taken to manage this set properly and to ensure feasibility of
the step size (the left-hand side of an inactive constraint can vary some, but if
it changes too much it may become violated). In this subsection, we present the
steps of the algorithm. Readers wishing a derivation of these steps, or a proof
that these steps indeed maintain feasibility and lead to an optimal solution, are
referred to the books by Bazaraa et al. (2006) and Bertsekas (2016).
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 557

We again consider the optimization problem

min f (x)
s.t. Ax ≤ b ∀i = 1, 2, .., m
Cx = d ∀i = 1, 2, .., l

and write the feasible region X as

X = {x : Ax ≤ b, Cx = d} .

For any feasible solution x̂, let Aa be the matrix containing the coefficients
of the set of active inequality constraints, that is,

Aa x̂ = ba .

Define E as the matrix containing the coefficients of all active constraints:

 
Aa
E= .
C

It may be that the rows of the matrix E are not linearly independent; this
indicates that one or more of the active constraints is redundant (they are
implied by some of the others). These redundant constraints may be removed.

Algorithm 3 provides the steps for the manifold suboptimization method.

Note that the convergence criterion is different from the MSA and FW methods.
In this algorithm, ᾱ is a parameter giving a “default” maximum step size (which
558 APPENDIX C. OPTIMIZATION TECHNIQUES

may be further restricted to ensure feasibility).

Initialize:
k ← 1: iteration counter
x1 : initial value
OptimalityF lag ← F ALSE
while OptimalityF lag = F ALSE do
P ← I − ET (EET )−1 E
dk ← −P∇f (xk )
if dk 6= 0 then
β ← max{ᾱ, xk + αdk ∈ X}
αk ← arg minα {f (xk + αdk ), 0 ≤ α ≤ β}
xk+1 ← xk + αk dk
k ←k+1
end
if d = 0 then
Q = −(EET )−1 E∇f (xk )
if qj ≥ 0 ∀j corresponding to Aa then
OptimalityF lag ← T RU E
end
else
Delete row from E corresponding to most negative qj
end
end
end
return xk
Algorithm 3: Manifold suboptimization

Example C.22. Re-solve the previous optimization problem using the manifold
suboptimization method, using ᾱ = 1.

min 2(x1 − 1)2 + (x2 − 2)2

s.t. x1 + 4x2 − 2 ≤0
−x1 + x2 ≤0
x1 , x2 ≥0

Solution. We start from the same initial solution x11 = 0, x12 = 0. At this
point, the active constraints are −x1 + x2 ≤ 0, x1 ≥ 0, and x2 ≥ 0. We only use
−x1 + x2 ≤ 0 and x2 ≥ 0 as they together imply x1 ≥ 0. (Mathematically, the
three vectors corresponding
   to the coefficients
 in these constraints are linearly
dependent: −1 1 , 0 1 , and 1 0 .)
 
−1 1
E=
0 −1
 
0 0
P = I − ET (EET )−1 E =
0 0
C.4. CONSTRAINED NONLINEAR OPTIMIZATION 559

 
0
d1 = −P∇f (x1 ) =
0
Since d = 0, we calculate Q:
 
T −1 −4
Q = −(EE ) E∇f (x1 ) =
−8

Since the second row is the most negative, we delete the second row of the
matrix E.  
E = −1 1
 
T T −1 0.5 0.5
P = I − E (EE ) E =
0.5 0.5
 
4
d1 = −P∇f (x1 ) =
4
The new solution can be obtained as:

x21 = 0 + 4α
x22 = 0 + 4α

We need to find the step size α which minimizes 2(0+4α −1)2 +(0+4α −2)2
while retaining feasibility. Note that

x1 + 4x2 − 2 ≤ 0 =⇒ α ≤ 0.1 ,

which is less than ᾱ = 1, so β = 0.1. Therefore, we obtain the step size by

solving the optimization problem

min 2(0 + 4α − 1)2 + (0 + 4α − 2)2

α
s.t. 0 ≤ α ≤ 0.1

whose solution is α = 0.1. Therefore, the new solution to the original problem
is

x21 = 0 + 4α = 0.4
x22 = 0 + 4α = 0.4

If we do one more iteration of MS, we get x31 = 0.7878, x32 = 0.3030, which is
optimal. MS converges to the optimal solution much faster than MSA or FW
(and about as fast as GP), because its steps are closer to the steepest descent
direction. Section 6.3 has more discussion of these reasons in the specific context
of traffic assignment. 
560 APPENDIX C. OPTIMIZATION TECHNIQUES

C.5 Integer Programming

All of the optimization formulations presented above assume that the deci-
sion variables are continuous in nature. In some cases, this is not a tenable
assumption. For example, when determining the location to establish a ware-
house, a fractional solution splitting the warehouse location between two nodes
is not meaningful. Integer programs are a special class of optimization problems,
where some or all decision variables must take integer values. Integer programs
have wide applications:

Vehicle routing : Determine routes for a specified number of trucks, start-

ing from a depot location to service a given number of customers with
demands, subject to truck capacity constraints at minimal cost.

Facility location : Determine where to locate facilities in a network, satisfying

customer demand, with an objective such as minimizing transportation
costs, distances, or maximum distance to facilities.

Transit network design : Determine the optimal routes and schedules of a

bus fleet to satisfy customer demands at minimal costs.

As always, there is no such thing as a free lunch; for reasons discussed below,
integer problems are significantly harder to solve, and often it is impractical to
find a provably optimal solution.
There are several types of integer programs. An integer linear program
(ILP) is a linear program where all the variables are restricted to be integers.
A mixed integer linear program (MILP) is a linear program where only some of
the decision variables are assumed to be integer. A MILP takes the form

min c·x+d·y
x,y
s.t. Ax + Ey ≤ b
x ≥ 0, y ≥ 0
y ∈ Zp

where x, c ∈ Rn , y, d ∈ Rp , A ∈ Rm×n , E ∈ Rm×p , and b ∈ Rm with m,

n, and p respectively denoting the number of constraints, continuous decision
variables, and integer decision variables. In an ILP, m = 0 and the terms
involving x in this formulation disappear. A special case of an ILP is the binary
integer linear program (BILP) where the decision variables are restricted to be
either 0 or 1, i.e. y ∈ {0, 1}p .
Integer programs are tougher to solve than linear programs as optimal solu-
tions need not occur at the corner points of a convex polyhedron. Indeed, the
feasible region itself is not a convex set; y1 = 1 and y2 = 2 may be feasible
solutions, but their average 21 y1 + 12 y2 = 1.5 is not. Solving the corresponding
linear program and then rounding the solution may not produce a good solution
either. (One notable exception is discussed in Section C.5.2.)
C.5. INTEGER PROGRAMMING 561

10

7
Optimal LP Solution
6
Optimal IP Solution
y2 5 11y1 + 3y2 ≥ 21

2
Z = 10y1 + 26y2
(1.5,1.5)
1
6y1 + 20y2 ≥ 39
0
0 1 2 3 4 5 6 7 8 9 10
y1

Figure C.9: Difference between LP and IP solution.

Consider the following ILP with two decision variables y1 and y2 :

min 10y1 + 26y2

y1 ,y2
s.t. 11y1 + 3y2 ≥ 21
6y1 + 20y2 ≥ 39
y1 , y2 ≥ 0
y1 , y2 ∈ Z

Without the integer restrictions on the variables, the optimal solution of

the linear program would be y1 = 1.5, y2 = 1.5, with an optimal objective
function value of 54 (see figure C.9). When the LP solutions are rounded up
(y1 = 2, y2 = 2) or rounded down (y1 = 1, y2 = 1) the resulting points are
sub-optimal or infeasible for the integer program. The optimal solution for the
integer program is in fact y1 = 4, y2 = 1, with an optimal objective function
value of 66.
562 APPENDIX C. OPTIMIZATION TECHNIQUES

Still, while ignoring the integrality constraints and solving the resulting lin-
ear program may not lead to an optimal solution, this linear program relaxation
is still very useful, because it provides a lower bound on the optimal value of the
objective function. In the above example, the optimal solution without integral-
ity constraints had an objective function value of 54. Restricting the feasible
set more by adding the integer constraints cannot improve this. Therefore, even
if we don’t know the optimal integer solution of the optimization problem, we
know its objective function value cannot be lower than 54.
Furthermore, any feasible integer solution provides an upper bound on the
optimal value for the objective. For instance, the rounded solution y1 = y2 = 2
is feasible, and has an objective function value of 72. Even if we don’t know
whether this solution is optimal or not, this solution tells us that the optimal
objective function value can’t be any higher than 72. We therefore have a
corresponding pair of upper and lower bounds. Solving the LP relaxation, and
calculating the objective function at the feasible solution (2, 2) tell us that the
optimal objective function value (often denoted Z ∗ ) must lie between 54 and
72.
A common framework for solving integer programs is to work to bring these
bounds closer together. Upper bounds can be tightened by finding better feasi-
ble solutions (with lower objective function values that are closer to the optimal
solution). Lower bounds can be tightened by solving “partial relaxations” where
only some of the integrality constraints are enforced; these will have higher ob-
jective function values than a full relaxation, which are also closer to the optimal
solution (but from the other side). Ultimately, we determine a decreasing se-
quence of upper bounds ZU1 B ≥ ZU2 B ≥ · · · Z ∗ and an increasing sequence of
lower bounds ZLB 1 2
≤ ZLB . . . · · · ≤ Z ∗ . When the difference between the bounds
is small enough (say, within ), we know that the feasible solution correspond-
ing to the best-known upper bound is within  of being optimal to the original
integer program.
The success of such a framework depends on the strength of the bounds.
The tighter the upper and lower bounds are, the greater the guarantee we can
provide for the solution we find. Effectively solving these problems often requires
intelligently exploiting the problem structure, and expert domain knowledge, in
order to provide good feasible solutions (for upper bounds), and good partial
relaxations (for lower bounds). The branch and bound algorithm (described in
Section C.5.1 is one framework for doing this. Before presenting this method, we
will provide another concrete example of an integer program for facility location,
and how it relates to the selection of good bounds.
Consider the case of locating a certain number of facilities in a region (say,
warehouses, or fire stations). Let I denote the set of locations that demand
what the facility provides, and J the set of potential locations where facilities
can be built. Let cij denote the cost of meeting the demand at location i ∈ I
from facility j ∈ J. Let fj represent the fixed cost of locating a facility at site
j ∈ J. There are two sets of decision variables corresponding to location and
assignment decisions. The location-related decision variable yj takes the value
1 if a facility is located at site j and 0 otherwise. The decision variable xij
C.5. INTEGER PROGRAMMING 563

denotes the fraction of demand for center i met by facility j. The xij variables
can be continuous (its demand can be provided by multiple facilities), whereas
the yj values must be integer (a facility cannot be “half-built” The optimization
model for the UFLP is shown below:

O-UFLP :
XX X
min cij xij + fj yj (C.18)
x,y
i∈I j∈J j∈J
X
s.t. xij = 1 ∀i ∈ I (C.19)
j∈J
X
xij ≤ |I|yj ∀j ∈ J (C.20)
i∈I
yj ∈ {0, 1} ∀j ∈ J (C.21)
xij ≥ 0 ∀i ∈ I, ∀j ∈ J (C.22)

The objective of the UFLP is to locate facilities and assign demand points to
facilities so that the overall facility location costs and transportation costs are
minimized, as shown in equation (C.18). Constraint (C.19) ensures all demands
are met. Constraint (C.20) ensures that the demand points are assigned to open
facilities only. (If yj = 0, no facility is located at site j, so no demand can be
served from there; if yj = 1, then potentially all demand could be served from
there.)
Now consider a new optimization model consisting of the same objective
function (C.18), along with constraints (C.19), (C.21), and (C.22), but with
constraint (C.20) replaced by: xij ≤ yj ∀i ∈ I, ∀j ∈ J.

S-UFLP :
XX X
min cij xij + fj yj (C.23)
x,y
i∈I j∈J j∈J
X
s.t. xij = 1 ∀i ∈ I (C.24)
j∈J

xij ≤ yj ∀i ∈ I, ∀j ∈ J (C.25)
yj ∈ {0, 1} ∀j ∈ J (C.26)
xij ≥ 0 ∀i ∈ I, ∀j ∈ J (C.27)

Let P and Q denote the feasible regions of the linear relaxations for models
O-UFLP and S-UFLP, respectively. We can show that Q ⊆ P , that is, any
solution feasible to S-UFLP is also feasible to O-UFLP. Any point which satisfies
xij ≤ yj must also satisfy xij ≤ |I|yj , as can be seen by summing up xij
along all values of i. Now, when restricted to integer values of y, the set of
feasible solutions is the same in both cases. However, when considering their
564 APPENDIX C. OPTIMIZATION TECHNIQUES

Figure C.10: Stronger and weaker IP formulations.

LP relaxations, there are solutions which are feasible to O-UFLP which are not
feasible to S-UFLP. For example, if y1 = 1/2, a solution of the form x11 = 1
and xi1 = 0 for i = {2, . . . , n} satisfies xi1 ≤ |I|y1 but not x11 ≤ y1 . As a
result of this, the optimal solution to a relaxation based on S-UFLP will have a
greater (or equal) objective function value than the same relaxation based on O-
UFLP, producing a tighter (and more useful) lower bound on the original integer
program. We therefore say that S-UFLP is a stronger (or tighter ) formulation
for UFLP. This discussion can be generalized in the following result:
Proposition C.1. Consider two feasible regions P and Q for the linear relax-
ation of the same minimization integer program Z = min{cT x : x ∈ X ∩ Zn }
P Q
with Q ⊆ P . If zLP = min{cT x : x ∈ P } and zLP = min{cT x : x ∈ Q}, then
P Q
zLP ≤ zLP
The intuition is shown in the figure C.10. Note that Q ∩ Zn = P ∩ Zn .
However since Q ⊆ P , the corresponding LP solution will be larger. Stronger
formulations leading to tighter linear programming bounds will result in faster
solutions, as long as the time need to solve the tighter linear programming
relaxation is not significantly higher.
C.5. INTEGER PROGRAMMING 565

∗
ZLP 1 = 54
y1 = 1.5, y2 = 1.5

y2 ≤ 1 y2 ≥ 2

∗ ∗
ZLP 3 = 57.67 ZLP 2 = 65.64
y1 = 3.17, y2 = 1 y1 = 1.36, y2 = 2

y1 ≤ 3 y1 ≥ 4 y1 ≤ 1 y1 ≥ 2

∗ ∗ ∗
LP7 ZLP 6 = 59.5 ZLP 5 = 96.67 ZLP 4 = 72
Infeasible y1 = 4, y2 = 0.75 y1 = 1, y2 = 3.33 y1 = 2, y2 = 2

y2 ≤ 0 y2 ≥ 1

∗ ∗
ZLP 9 = 65 ZLP 8 = 66
y1 = 6.5, y2 = 0 y1 = 4, y2 = 1

y1 ≤ 6 y1 ≥ 7

∗
LP11 ZLP 10 = 70
Infeasible y1 = 7, y2 = 0

Figure C.11: Branch and bound example.

C.5.1 Branch and bound algorithm

Branch and bound is one of the most popular algorithms used to solve integer
programs, and is the basis for many commercial integer programming solvers.
It is essentially an intelligent enumeration algorithm, based on the principle of
“divide and conquer.” The original integer program is replaced by a series of
easier-to-solve relaxations, whose solutions are used to further guide the search
toward the optimal solution of the integer program. The branch and bound
algorithm is illustrated using the same integer programming example (see figure
C.11) used before:
min 10y1 + 26y2
y1 ,y2
s.t. 11y1 + 3y2 ≥ 21
6y1 + 20y2 ≥ 39
y1 , y2 ≥ 0
y1 , y2 ∈ Z
We will first show how the algorithm works on this specific example, and then
provide the general method below.
The branch and bound algorithm starts by initializing the upper bound
566 APPENDIX C. OPTIMIZATION TECHNIQUES

ZU B = ∞ and lower bound ZLB = −∞, since we do not yet know any infor-
mation about the optimal solution. The linear relaxation of the above integer
program is solved. Let XLP 1 denote the feasible region of the LP relaxation.
Thus XLP 1 = {y ∈ R2+ : 11y1 + 3y2 ≥ 21, 6y1 + 20y2 ≥ 39}. The optimal value
∗
of the linear programming objective function is ZLP 1 = 54, and the solution is
y1 = 1.5, y2 = 1.5. The best lower bound is now updated to to 54, ZLB = 54.
Now, the optimal solution either has y2 as an integer greater than or equal
to 2, or as an integer less than or equal to 1. We consider each option in turn.
(We could have also branched on y1 ; it would be instructive for you to solve the
problem that way.) We generate two linear programs by adding the constraints
y2 ≥ 2 to one, and and y2 ≤ 1 to the other. The feasible region of the second
linear program XLP 2 = XLP 1 ∩ {y2 ≥ 2} and the feasible region of the third
linear program XLP 3 = XLP 1 ∩ {y2 ≤ 1}. Solving the two linear programs we
∗ ∗
get ZLP 2 = 65.64 with the solution y1 = 1.36, y2 = 2 and ZLP 3 = 57.67 with
the solution y1 = 3.17, y2 = 1.
Let us first examine LP2. Since y2 is already an integer in its optimal
solution, we don’t need to branch again on y2 . However, y1 is not an integer, and
in the optimal solution it must be either greater than or equal to 2, or less than or
equal to 1. So the constraints y1 ≥ 2 and y2 ≤ 1 are added to the feasible regions
of LP2 to generate new linear programs LP4 and LP5, with the respective
feasible regions XLP 4 = XLP 2 ∩ {y1 ≥ 2} and XLP 5 = XLP 2 ∩ {y1 ≤ 1}.
Solving LP4 we get the optimal objective value to be 74 with the solution
y1 = 2, y2 = 2. The solution to this linear programming relaxation is an integer
∗
solution which is feasible for the original integer program. Therefore ZLP 5 is as
∗
an upper bound to the optimal integer solution. Moreover since ZLP 5 < ZU B
we update ZU B = 74; this is the first upper bound we have the optimal objective
function value.. Since this is an integer solution, we do not search further in
this direction. This is called “fathoming by optimality.”
Solving LP5 we get the optimal objective value to be 96.67 with the solution
y1 = 1, y2 = 3.33. Since the LP objective function is greater than the current
best upper bound 74, there is no point in further exploring in this direction also
and the search is stopped here. This is called “fathoming by bound.” Notice
that this cuts off a considerable portion of the feasible region. We know we
do not have to explore it any further, since the best possible objective function
values there are at least 96.67, whereas we have already found a solution (2, 2)
with an objective of 74.
Let us now look at LP3. Since y2 has integer solution we don’t need to
branch on y2 . But with respect to y1 , the optimal solution is either an integer
at least 4, or at most 3. So the constraints y1 ≥ 4 and y1 ≤ 3 is added to the
feasible region of LP3 to get new linear programs LP6 and LP7, with respective
feasible regions XLP 6 = XLP 3 ∩ {y1 ≥ 4} and XLP 7 = XLP 3 ∩ {y2 ≤ 3}.
LP7 turns out to be infeasible, so the search is stopped along this direction.
This is called “fathoming by infeasibility.” LP6, on the other hand, has an
optimal objective function value of 59.5 with y1 = 4, y2 = 0.75. We branch
again on y2 , adding the constraints y2 ≥ 1 and y2 ≤ 0 to the feasible region of
LP6 to obtain new linear programs LP8 and LP9 with feasible regions XLP 8 =
C.5. INTEGER PROGRAMMING 567

XLP 6 ∩ {y2 ≥ 1} and XLP 9 = XLP 3 ∩ {y2 ≤ 0}.

LP8 has an integer solution y1 = 4, y2 = 1, with an objective function value

of 66. This solution is feasible for the original integer program. Moreover since
∗
ZLP 8 < ZU B , we have found a better feasible solution, and the upper bound is
updated to 66: ZU B = 66. We fathom the search in this direction by integrality.

Looking at the solution of LP9, we generate two linear programs, LP10 and
LP11, by adding the constraints y1 ≥ 7 and y2 ≤ 6 to the feasible region of LP9.
The feasible region of the new linear programs are XLP 10 = XLP 9 ∩ {y1 ≥ 7}
and XLP 11 = XLP 9 ∩ {y1 ≤ 6}. LP11 is infeasible, and therefore we fathom by
infeasibility. LP10 has an integer solution, but the objective function value is
greater than the current upper bound. Therefore we fathom by bound as well as
integrality. There are no more directions left to search. Therefore the optimal
solution corresponds to the current best upper bound of 66.

The general steps of branch and bound are shown in Algorithm 4.

Initialization: Set ZU B ← −∞ and ZLB ← ∞. Add the LP relaxation

of the original integer program to a queue Q.;
while Q is not empty do
Select an element CurrLP from Q and remove it;
Attempt to solve CurrLP;
if CurrLP is infeasible then
Continue while loop (fathom by infeasibility);
else
Let x be the optimal solution to CurrLP, and Z the objective
value;
if x satisfies all of the original integrality constraints then
ZU B ← min {ZU B , Z};
If ZU B changed, mark x as the best known solution;
Continue while loop (fathom by optimality);
else
if Z ≥ ZU B then
Continue while loop (fathom by bound);
else
ZLB ← max {ZLB , Z};
Choose a fractional variable to branch on;
Generate two new LPs by adding ≥ and ≤ constraints to
the feasible region of CurrLP;
Add the two new LPs to Q;
end
end
end
end
Algorithm 4: Branch and bound algorithm.
568 APPENDIX C. OPTIMIZATION TECHNIQUES

3)

(c
34
r1

,r
3,

34
1
(c

)
1 (c23 , r23 ) 4

(c

)
24
1

,r
2
,r

24
12

(c
) 2

Figure C.12: Network for shortest path example

C.5.2 Total unimodularity

In general, we cannot exactly solve an integer program only by solving its linear
programming relaxation (although we can use these LP relaxations in a method
such as branch and bound). There is one notable exception, which shows up
frequently enough in network optimization problems to be worth knowing.

Definition C.7. A matrix A is totally unimodular if the determinant of every

square submatrix of A has the value −1, 0, or 1.

Consider an integer linear program Z = min{cT x : x ∈ X ∩ Zn } where

X = {x ∈ Rn : Ax ≤ b, x ≥ 0}, and the corresponding linear programming
relaxation ZLP = min{cT x : x ∈ X}. If the matrix A is totally unimodu-
lar, then the LP relaxation has integral solutions. The vertex points of X are
integer points. That is, if the coefficients representing the constraints form a
totally unimodular matrix, we can ignore the integer constraints altogether. The
resulting linear program will have an integer optimal solution which is feasible
to the original integer program. In such a case, we can simply apply a linear
programming method to find the optimal solution.
The concept of total unimodularity is explained using a modification to the
standard shortest path problem from Section 2.4. In this modification, each
link is associated with two attributes: a cost for traversing it, and the amount
of some other resource consumed when traveling it (time, battery charge, etc.)
Consider the network shown in figure C.12. The network has four nodes and
five arcs. Let c12 and r12 denote the cost and the amount of resource consumed
by traversing arc (1, 2). The goal of the shortest path problem is to determine
the minimum cost path from 1 to 4.
Let x12 ,x13 ,x23 ,x24 , and x34 denote the decision variables. These are equal
C.6. METAHEURISTICS 569

to 1 if the corresponding arc lies on the shortest path, and 0 otherwise. The
shortest path formulation for the above network is:
min c12 x12 + c13 x13 + c23 x23 + c24 x24 + c34 x34
s.t. x12 + x13 =1
−x12 + x23 + x24 =0
−x13 − x23 + x34 =0
−x24 − x34 = −1
x12 , x13 , x23 , x24 , x34 ∈ {0, 1}

The constraint matrix is

 
1 1 0 0 0
−1 0 1 1 0
A= 
0 −1 −1 0 1
0 0 0 −1 −1

All the square submatrices of A have determinant 0, 1, or −1. Therefore we can

solve the shortest path problem by solving the corresponding linear program-
ming relaxation. (The shortest path problem could be solved using branch and
bound; but it is better to solve it as a linear program, and even better to solve
it using a specialized method like those in Section 2.4.)
The resource-constrained shortest path problem is a variant where the short-
est path is sought, but with an upper bound on the total amount of resources
consumed. An example might be finding the fastest route in a network with
toll roads, where the total monetary cost does not exceed a given budget. We
represent this with an additional constraint in the shortest path formulation:

r12 x12 + r13 x13 + r23 x23 + r24 x24 + r34 x34 ≤ R .

In this case the constraint matrix becomes

 
1 1 0 0 0
−1 0 1 1 0
 
A=  0 −1 −1 0 1 
0 0 0 −1 −1
r12 r13 r23 r24 r34

Depending on the values of the resources consumed, the constraint matrix need
not be totally unimodular. In this case the linear programming relaxation does
not always have an integer solution. For this reason, the resource-constrained
shortest path problem is much harder to solve than the traditional shortest path
problem.

C.6 Metaheuristics
In optimization, there is often a tradeoff between how widely applicable a solu-
tion method is, and how efficient or effective it is at solving specific problems.
570 APPENDIX C. OPTIMIZATION TECHNIQUES

For any one specific problem, a tailor-made solution process is likely much faster
than a generally-applicable method, but generating such a method requires more
effort and specialized knowledge, and is less “rewarding” in the sense that the
method can only be more narrowly applied. Some of the most general techniques
are heuristics, which are not guaranteed to find the global optimum solution,
but tend to work reasonably well in practice. In practice, they only tend to
be applied for very large or very complicated problems which cannot be solved
exactly in a reasonable amount of time with our current knowledge.
Many engineers are initially uncomfortable with the idea of a heuristic. Af-
ter all, the goal of an optimization problem is to find an optimal solution, so
why should we settle for something which is only approximately “optimal,” of-
ten without any guarantees of how approximate the solution is? First, for very
complicated problems a good heuristic can often return a reasonably good so-
lution in much less time than it would take to find the exact, global optimal
solution. For many practical problems, the cost improvement from a reasonably
good solution to an exactly optimal one is not worth the extra expense (both
time and computational hardware) needed, particularly if the heuristic gets you
within the margin of error based on the input data.
Heuristics are also very, very common in psychology and nature. If I give
someone a map and ask them to find the shortest-distance route between two
points in a city by hand, they will almost certainly not formulate an mathe-
matical model and solve it to provable optimality. Instead, they use mental
heuristics (rules of thumb based on experience) and can find paths which are
actually quite good. Many of the heuristics are inspired by things seen in nature.
An example is how ant colonies find food. When a lone wandering ant
encounters a food source, it returns to the colony and lays down a chemical
pheromone. Other ants who stumble across this pheromone begin to follow it to
the food source, and lay down more pheromones, and so forth. Over time, more
and more ants will travel to the food source, taking it back to the colony, until
it is exhausted at which point the pheromones will evaporate. Is this method
the optimal way to gather food? Perhaps not, but it performs well enough for
ants to have survived for millions of years!
Another example is the process of evolution through natural selection. The
human body, and many other organisms, function remarkably well in their habi-
tats, even if their biology is not exactly “optimal.”2 One of the most common
heuristics in use today, and one described below, is based on applying principles
of natural selection and mutation to a “population” of candidate solutions to
an optimization problem, using an evolutionary process to identify better and
better solutions over time. This volume describes two heuristics: simulated an-
nealing, and genetic algorithms, both of which can be applied to many different
optimization problems.

2 For instance, in humans the retina is “backwards,” creating a blind spot; in giraffes, the

laryngeal nerve takes an exceptionally long and roundabout path; and the descent of the testes
makes men more vulnerable to hernias later in life.
C.6. METAHEURISTICS 571

C.6.1 Simulated annealing

Simulated annealing is a simple heuristic that makes an analogy to metallurgy,
but this analogy is best understood after the heuristic is described.
As a better starting point for understanding simulated annealing, consider
the following “local search” heuristic. (For now, descriptions will be a bit vague;
more precise definitions will follow soon.) Local search proceeds as follows:

1. Choose some initial feasible solution x ∈ X, and calculate the value of the
objective function f (x).

2. Generate a new feasible solution x0 ∈ X which is close to the current

solution x.

3. Calculate f (x0 )

4. If f (x0 ) ≤ f (x), the new solution is better than the old solution. So
update the current solution by setting x equal to x0 .

5. Otherwise, the old solution is better, so leave x unchanged.

6. Return to step 2 and repeat until we are unable to make further progress.

One way to visualize local search is a hiker trying to find the lowest elevation
point in a mountain range. In this analogy, the park boundaries are the feasible
set, the elevation of any point is the objective function, and the location of the
hiker is the decision variable. In local search, starting from his or her initial
position, the hiker looks around and finds a nearby point which is lower than
their current point. If such a point can be found, they move in that direction
and repeat the process. If every neighboring point is higher, then they stop and
conclude that they have found the lowest point in the park.
It is not hard to see why this strategy can fail; if there are multiple local
optima, the hiker can easily get stuck in a point which is not the lowest (Fig-
ure C.13). Simulated annealing attempts to overcome this deficiency of local
search by allowing a provision for occasionally moving in an uphill direction, in
hopes of finding an even lower point later on. Of course, we don’t always want
to move in an uphill direction and have a preference for downhill directions, but
this preference cannot be absolute if there is any hope of escaping local minima.
Simulated annealing accomplishes this by making the decision to move or
not probabilistic, introducing a temperature parameter T which controls how
likely you are to accept an “uphill” move. When the temperature is high,
the probability of moving uphill is large, but when the temperature is low,
the probability of moving uphill becomes small. In simulated annealing, the
temperature is controlled with a cooling schedule. Initially, the temperature
is kept high to encourage a broad exploration of the feasible region. As the
temperature decreases slowly, the solution is drawn more and more to lower
areas. The cooling schedule can be defined by an initial temperature T0 , a final
temperature Tf , the number of search iterations n at a given temperature level,
572 APPENDIX C. OPTIMIZATION TECHNIQUES

f(x)

Figure C.13: Moving downhill from the current location may not lead to the
global optimum.

and a scaling factor k ∈ (0, 1) which is applied every n iterations to reduce the
temperature.
Finally, because the search moves both uphill and downhill, there is no
guarantee that the final point of the search is the best point found so far. So,
it is worthwhile to keep track of the best solution x∗ encountered during the
algorithm. (This is analogous to the hiker keeping a record of the lowest point
observed, and returning to that point when done searching.)
So, the simulated annealing algorithm can be stated as follows:

1. Choose some initial feasible solution x ∈ X, and calculate the value of the
objective function f (x).

2. Initialize the best solution to the initial one x∗ ← x.

3. Set the temperature to the initial temperature: T ← T0

4. Repeat the following steps n times:

(a) Generate a new feasible solution x0 which neighbors x.

(b) If f (x0 ) < f (x∗ ), it is the best solution found so far, so update x∗ ←
x0 .
(c) If f (x0 ) ≤ f (x), it is a better solution than the current one, so update
x ← x0 .
(d) Otherwise, update x ← x0 with probability exp(−[f (x0 ) − f (x)]/T )

5. If T > Tf , then reduce the temperature (T ← kT ) and return to step 4.

6. Report the best solution found x∗

The key step is step 4d. Notice how the probability of “moving uphill”
depends on two factors: the temperature, and how much the objective function
will increase. The algorithm is more likely to accept an uphill move if it is only
slightly uphill, or if the temperature is high. The exponential function captures
C.6. METAHEURISTICS 573

these effects while keeping the probability between 0 and 1. A few points require
explanation.
How should the cooling schedule be chosen? Unfortunately, it is
hard to give general guidance here. Heuristics often have to be “tuned” for
a particular problem: some problems do better with higher temperatures and
slower cooling (k values closer to 1, n larger), others work fine with faster cooling.
When you use simulated annealing, you should try different variations of the
cooling schedule to identify one that works well for your specific problem.
How should an initial solution be chosen? It is often helpful if the
initial solution is relatively close to the optimal solution. For instance, if the
optimization problem concerns business operations, the current operational plan
can be used as the initial solution for further optimization. However, it’s easy
to go overboard with this. You don’t want to spent so long coming up with
a good initial solution that it would have been faster to simply run simulated
annealing for longer starting from a worse solution. The ideal is to think of a
good, quick rule of thumb for generating a reasonable initial solution; failing
that, you can always choose the initial solution randomly.
How do I define a neighboring solution for step 4a? Again, this
is problem-specific, and one of the decisions that must be made when apply-
ing simulated annealing. A good neighborhood definition should involve points
which are “close” to the current solution in some way, but ensure that feasi-
ble solutions are connected in the sense that any two feasible solutions can be
reached by a chain of neighboring solutions. For the examples in Section B.4,
some possibilities (emphasis on possibilities, there are other choices) are:

Transit frequency setting problem: The decision variables n are the num-
ber of buses assigned to each route. Given a current solution n, a neighbor-
ing solution is one where exactly one bus has been assigned to a different
route.

Scheduling maintenance: The decision variables are x, indicating where

and when maintenance is performed, and c, indicating the condition of
the facilities. In this problem, the constraints completely define c in terms
of x, so for simulated annealing it is enough to simply choose a feasible
solution x, then calculate c using the state evolution equations. In this
problem, there is no reason to perform less maintenance than the bud-
get allows each year. If the initial solution is chosen in this way, then a
neighboring solution might be one where one maintenance activity on a
facility is reassigned to another facility that same year. If the resulting so-
lution is infeasible (because the resulting c values fall below the minimum
threshold), then another solution should be generated.

Facility location problem: The decision variables are the intersections that
the three terminals are located at, L1 , L2 , and L3 . Given current values for
these, in a neighboring solution two of the three terminals are at the same
location, but one of the three has been assigned to a different location.
574 APPENDIX C. OPTIMIZATION TECHNIQUES

Shortest path problem: The decision variables specify a path between the
origin and destination. Given a current path, a neighboring path is one
which differs in only intersection. (Can you think of a way to express this
mathematically?)

How do I perform a step with a given probability? Most program-

ming languages have a way to generate a uniform random variable between 0
and 1. If we want to perform a step with probability p, generate a random
number from the continuous uniform (0, 1) distribution. If this number is less
than p, perform the step; otherwise, do not.

Example with facility location

This section demonstrates simulated annealing, on an instance of the facility
location problem from Section B.4. In this problem instance, the road grid
consists of ten north-south and ten east-west streets. The cost of locating a
terminal at each of the locations is shown in Figure C.14 (these numbers were
randomly generated). There are 30 customers, randomly located throughout
the grid, as shown in Figure C.15. In these figures, the coordinate system (x, y)
is used to represent points, where x represents the number of blocks to the
right of the upper-left corner, and y represents the number of blocks below the
upper-left corner.
The cooling schedule is as follows: the initial temperature is T0 = 1000,
the final temperature is Tf = 100, k = 0.75, and the temperature is reduced
every 8 iterations. (These values were chosen through trial-and-error, and work
reasonably well for this problem.) The first several iterations are shown below.
The algorithm begins with an initial solution, generated randomly. Suppose
this solution locates the three facilities at coordinates (5,1), (8,6), and (6,1),
respectively; the total cost associated with this solution is 158.0 cost units. A
neighboring solution is generated by picking one of the facilities (randomly) and
assigning it to another location (randomly). Suppose that the third facility is
reassigned to location (1,0), so the new candidate solution locates the facilities
at (5,1), (8,6), and (1,0). This solution has a cost of 136.4 units, which is lower
than the current solution. So, simulated annealing replaces the current solution
with the new one.
In the next iteration, suppose that the first facility is randomly chosen to
be reassigned to (8,2), so the candidate solution is (8,2), (8,6), and (1,0). This
solution has a cost of 149.1, which is higher than the current cost of 136.4. The
algorithm will not always move to such a solution, but will only do so with
probability calculated as in Step 4c:

p = exp(−[f (x0 ) − f (x)]/T ) = exp(−[149.1 − 136.4]/100) = 0.881

This probability is high, because (being one of the early iterations) the temper-
ature is set high. If the temperature were lower, the probability of accepting
this move would be lower as well.
C.6. METAHEURISTICS 575

Figure C.14: Cost of locating facilities at each intersection.

576 APPENDIX C. OPTIMIZATION TECHNIQUES

Figure C.15: Locations of customers.

C.6. METAHEURISTICS 577

Figure C.16: Progress of simulated annealing. Solid line shows current solution
cost, dashed line best cost so far.

Supposing that the move is accepted, the algorithm replaces the current
solution with the candidate and continues as before. If, on the other hand, the
move is rejected, the algorithm generates another candidate solution based on
the same current solution (5,1), (8,6), and (1,0). The algorithm continues in the
same way, reducing the temperature by 25% every 8 iterations.
The progress of the algorithm until termination is shown in Figure C.16. The
solid line shows the cost of the current solution, while the dashed line tracks the
cost of the best solution found so far. A few observations are worth making.
First, in the early iterations the cost is highly variable, but towards termination
the cost becomes more stable. This is due to the reduction in temperature
which occurs over successive iterations. When the temperature is high, nearly
any move will be accepted so one expects large fluctuations in the cost. When
the temperature is low, the algorithm is less likely to accept cost-increasing
moves, so fewer fluctuations are seen. Also notice that the best solution was
found shortly after iteration 700. The final solution is not the best, although it
is close.
The best facility locations found by simulated annealing are shown in Fig-
ure C.17, with a total cost of 92.1.

C.6.2 Genetic algorithms

Another heuristic is the genetic algorithm. In contrast to simulated annealing,
where there is a single search happening through the feasible region (recall
the analogy with the hiker), genetic algorithms maintain a larger population of
solutions which reflect a diverse set of points within the feasible region. Genetic
algorithms work by attempting to improve the population from one iteration to
the next.
The process of improvement is intended to mimic the processes of natural
578 APPENDIX C. OPTIMIZATION TECHNIQUES

Figure C.17: Locations of facilities found by simulated annealing (cost 92.1).

selection, reproduction, and mutation which are observed in biology. For the
sake of our purposes, natural selection means identifying solutions in the popu-
lation which have good (low) values of the objective function. The hope is that
there are certain aspects or patterns in these solutions which make them good,
which can be maintained in future generations. Given an initial population as
the first generation, subsequent generations are created by choosing good solu-
tions from the previous generation, and “breeding” them with each other in a
process called crossover which mimics sexual reproduction: new “child” solu-
tions are created by mixing characteristics from two “parents.” Lastly, there
is a random mutation element, where solutions are changed externally with a
small probability. If all goes well, after multiple generations the population will
tend towards better and better solutions to the optimization problem.
At a high level, the algorithm is thus very straightforward and presented
below. However, at a lower level, there are a number of careful choices which
need to be made about how to implement each step for a particular problem.
Thus, each of the steps is described in more detail below. The high-level version
of genetic algorithms is as follows:

1. Generate an initial population of N feasible solutions (generation 0), set

generation counter g ← 0.

2. Create generation g + 1 in the following way, repeating each step N times:

(a) Choose two “parent” solutions from generation g.

(b) Combine the two parent solutions to create a new solution.
C.6. METAHEURISTICS 579

(c) With probability p, mutate the new solution.

3. Increase g by 1 and return to step 2 unless done.

Although not listed explicitly, it is a good idea to keep track at every stage
of the best solution x∗ found so far. Just like in simulated annealing, there is
no guarantee that the best solution will be found in the last generation. So,
whenever a new solution is formed, calculate its objective function value, and
record the solution if it is better than any found so far.
Here are explanations of each step in more detail. It is very important to
realize that there are many ways to do each of these steps and that what is
presented below is one specific example intended to give the general flavor while
sparing unnecessary complications at this stage.
Generate an initial population of feasible solutions: In contrast
with simulated annealing, where we had to generate an initial feasible solution,
with genetic algorithms we must generate a larger population of initial feasible
solutions. While it is still desirable to have these initial feasible solutions be rea-
sonably good, it is also very important to have some diversity in the population.
Genetic algorithms work by combining characteristics of different solutions to-
gether. If there is little diversity in the initial population the difference between
successive generations will be small and progress will be slow.
Selection of parent solutions: Parent solutions should be chosen in a
way that better solutions (lower objective function values) are more likely to
be chosen. This is intended to mimic natural selection, where organisms better
adapted for a particular environment are more likely to reproduce. One way
to do this is through tournament selection, where a number of solutions from
the old generation are selected randomly, and the one with the best objective
function value is chosen as the first parent. Repeating the “tournament” again,
randomly select another subset of solutions from the old generation, and choose
the best one as the second parent. The number of entrants in the tournament
is a parameter that you must choose.
Combining parent solutions: This is perhaps the trickiest part of ge-
netic algorithms: how can we combine two feasible solutions to generate a new
feasible solution which retains aspects of both parents? The exact process will
differ from problem to problem. Here are some ideas, based on the example
problems from Section B.4:

Transit frequency setting problem: The decision variables are the num-
ber of buses on each route. Another way of “encoding” this decision is
to make a list assigning each bus in the fleet to a corresponding route.
(Clearly, given such a list, we can construct the nr values by counting
how many times a route appears.) Then, to generate a new list from two
parent lists, we can assign each bus to either its route in one parent, or
its route in the other parent, choosing randomly for each bus.

Scheduling maintenance: As described above, optimal solutions must al-

ways exhaust the budget each year. So, along the lines of the transit
580 APPENDIX C. OPTIMIZATION TECHNIQUES

frequency setting problem, make a list of the number of maintenance ac-

tions which will be performed in each year, noting the facility assigned
to each maintenance action in the budget. To create a new solution, for
each maintenance action in the budget choose one of the facilities assigned
to this action from the two parents, selecting randomly. Once this list is
obtained, it is straightforward to calculate x and c.

Facility location problem: For each facility in the child solution, make its
location the same as the location of that facility in one of the two parents
(chosen randomly).

Mutating solutions: Mutation is intended to provide additional diversity

to the populations, avoiding stagnation and local optimal solutions. Mutation
can be accomplished in the same way that neighbors are generated in simulated
annealing. The probability of mutation should not be too high — as in nature,
most mutations are harmful — but enough to escape from local optima. The
mutation probability p can be selected by trial and error, determining what
works well for your problem.

Example with facility location

This section demonstrates genetic algorithms on the same facility location prob-
lem used to demonstrate simulated annealing. The genetic algorithm is imple-
mented with a population size of 100, over ten generations. Tournaments of size
3 are used to identify parent solutions, and the mutation probability is 0.05.
The initial population is generated by locating all the facilities are completely
at random.
To form the next generation, 100 new solutions need to be created, each
based on combining two solutions from the previous generation. These two
parents are chosen using the tournament selection rule, as shown in Figure C.18.
The two winners of the tournament locate the three facilities at (2,3), (1,7),
(8,7); and (1,4), (7,6), (5,0) respectively. These combine in the following way:

1. The first facility is located either at (2,3) or (1,4); randomly choose one
of them, say, (2,3).

2. The second facility is located either at (1,7) or (7,6); randomly choose one
of them, say, (1,7).

3. The third facility is located either at (8,7) or (5,0); randomly choose one
of them, say, (5,0).

This gives a new solution (2,3), (1,7), (5,0) in the next generation, as shown
in Figure C.18. With 5% probability, this solution will be “mutated.” If this
solution is selected for mutation, one of the three facility is randomly reassigned
to another location. For instance, the facility at (1,7) may be reassigned to
(0,4). This process is repeated until all 100 solutions in the next generation
have been created.
C.6. METAHEURISTICS 581

Population of 100 solutions

Tournament 1 Tournament 2
(5,1) (8,6) (6,1) : cost 158.0 (6,7) (2,1) (0,0) : cost 125.3
(8,1) (1,7) (3,9) : cost 167.4 (1,4) (7,6) (5,0) : cost 112.6
(2,3) (1,7) (8,7) : cost 118.7 (8,8) (2,1) (5,0) : cost 130.7

Parent 1 Parent 2
(2,3) (1,7) (8,7) (1,4) (7,6) (5,0)

Child
(2,3) (1,7) (5,0)

with probability 0.05

Child
(2,3) (0,4) (5,0)

Figure C.18: Generation of a new solution: reproduction and mutation.

582 APPENDIX C. OPTIMIZATION TECHNIQUES

Figure C.19: Progress of genetic algorithm. Solid line shows average generation
cost, dashed line best cost so far, crosses cost of individual solutions.

Figure C.19 shows the progress of the algorithm over ten generations. The
solid line shows the average cost in each generation. The dashed line shows the
cost of the best solution found so far, and the crosses show the cost of each of
the solutions comprising the generations. Since lower-cost alternatives are more
likely to win the tournaments and be selected as parents, the average cost of
each generation decreases. Figure C.20 shows the locations of the terminals in
the best solution found.
C.6. METAHEURISTICS 583

Figure C.20: Locations of facilities found by genetic algorithm (cost 94.7).

584 APPENDIX C. OPTIMIZATION TECHNIQUES
Appendix D

Algorithms and Complexity

Appendices B and C introduced the concepts of formulating optimization prob-

lem, provided examples of different types and categories of formulations, and
presented algorithms for solving them. This appendix focuses more on what
specifically an algorithm is, and how complexity analysis can be used to com-
pare the performance of alternative algorithms for solving a problem.

D.1 Algorithms
The word algorithm appears often in applied mathematics and computer science
to denote a concrete procedure for accomplishing a task, specified in enough de-
tail that it can be implemented by a computer (often in a programming language
like C or Python). We adopt Knuth’s (1997) criteria for operationalizing this
definition: an algorithm must be finite (always stopping in a finite number of
steps), definite (each step is unambiguous), effective (each operation is simple
enough to be done in a finite amount of time), and map inputs to at least one
output. An algorithm can have any finite number of inputs, even zero: an al-
gorithm with no inputs would compute the same thing every time it is run,
but that still might be useful (e.g., to answer a difficult question). However,
it must have at least one output; an algorithm must report something when it
terminates.
For example, a step such as “add integers x and y” is definite and effective.
A step such as “multiply π by 2” is definite, but not effective since π is an
irrational number with an infinite number of digits (there is no way to multiply
all the digits by 2). Such a step can be made effective by specifying “round π
to 15 digits and multiply it by 2.” A step such as “pick an integer between 1
and 10” is effective, but it is not definite (how exactly should I pick such an
integer?) One way to make it definite is to specify “randomly pick an integer
between 1 and 10, with each possibility equally likely.”
Another example of a “definite but not effective” step is “add 1 to x if there
are one million consecutive zeros in the decimal expansion of π.” It is definite:

585
586 APPENDIX D. ALGORITHMS AND COMPLEXITY

either there are that many consecutive zeros or there are not. However, we
don’t already know whether that is true, and there is no simple way to find out.
We could spend a lot of time computing digits of π and maybe we will find this
many zeros; but maybe we won’t, and we don’t know if it’s because there aren’t
any, or if we just haven’t computed far enough. With our current knowledge
of mathematics, there is no way to answer that question in a finite amount of
time, so this step is not effective.
Many algorithms contain three main components:

Initialization: These are the first steps performed. The initialization steps
prepare for the main algorithm, assembling the inputs, and providing ini-
tial or default values to variables which will be used later on.

Body: The steps in the body form the main core of the algorithm. Many
algorithms are iterative, and the steps in the body are repeated multiple
times.

Stopping criterion: An algorithm must be finite, and not run forever. How-
ever, we often don’t know how many iterations are required in advance.
Therefore, we periodically check a stopping criterion; once it is satisfied,
we can stop. There may also be a few final processing and “clean up”
steps needed at the end of the algorithm.

You were probably taught an algorithm for adding two positive integers when
you were a child (although that word was probably not used). One way to do
this is to work from right to left, adding the digits one at a time. If the sum
of the digits is 9 or less, then that is the corresponding digit in the sum; if it
is 10 or more, subtract 10 to get the digit in the sum, and “carry” by adding 1
to the next digit to the left. For example, to add 158 and 371, you would start
by adding 8 + 1 = 9 for the rightmost digit; 5 + 7 = 12, so enter 2 for the next
digit and carry a 1 for the next digit: 1 + 3 + 1 = 5, so the sum is 529.
This can be specified as an algorithm in the following way. For simplicity, we
assume that the two integers we are adding (a and b) have the same number of
digits n, and they are specified as a = an an−1 . . . a2 a1 and b = bn bn−1 . . . b2 b1 .
(In the previous example, a = 158, and a3 = 1, a2 = 5, and a3 = 8; and likewise
for b = 371.) We also assume that we know how to add two single digits, perhaps
with an “addition table” telling us the value of adding 0 + 0, 0 + 1, and so on,
up through 9 + 9. The steps below describe the addition algorithm, calculating
the digits in the sum s.
Note the use of the ← operator to mean assignment: calculate the expression
on the right-hand side, and store the result in the variable on the left-hand side.
For instance, c ← 0 sets the variable c to zero. Many programming languages
use the = sign to denote this (you would type c = 0). In this text, we will
reserve = to refer to mathematical equality, the statement that the left-hand
side and right-hand side are currently equal to each other. So a statement like
x = x + 1 can never be true (no number is equal to one plus itself). However, a
statement like x ← x + 1 is perfectly fine; it means add one to the current value
D.1. ALGORITHMS 587

in x, and then store that new value in x. Some programming languages use
context to determine whether = is meant in the sense of assignment or equality,
while others use different operators for the two (for instance, C and Python
use = for assignment, and == for equality). To avoid confusion, we will use the
different characters ← for assignment, and = for equality.

1. Initialization: set the carry digit c ← 0, and the current digit i ← 1.

2. Compute si ← c + ai + bi .
3. If si ≥ 10, set c ← 1 and si ← si − 10; otherwise, set c ← 0.
4. If i < n, set i ← i + 1 and return to step 2.
5. If c = 1, set sn+1 = 1.
6. Return the integer s and terminate.

In this algorithm, step 1 is the initialization, and steps 2–3 are the body.
Step 4 contains the stopping criterion, expressed in a “negative” way (“repeat
as long as i < n” rather than “stop repeating if i ≥ n.”) Step 5 contains a final
step before we terminate (if there is a carry digit after adding the n-th place,
we need to add a 1 at the left-hand side of the sum, which will have one more
digit than the original numbers.)
In general, when reading an algorithm, you should work out a small example
on a separate piece of paper to see how the steps work. Trying to mentally
think through all the steps adds unnecessary difficulty, especially for complex
algorithms. It is much easier to see how it works on a small example or two, and
then study the specific steps to see how they fit together.1 So, let’s see how this
algorithm works for the numbers a = 158 and b = 371. These are three-digit
numbers, so n = 3. After the initialization (step 1), c = 0 and i = 1, meaning
we are working with the first (right-most) digit and there is nothing to carry
so far. Step 2 adds the two right-most digits: s1 = c + a1 + b1 = 9, giving the
right-most digit in the sum. This is less than 10, so the carry digit is set to zero
(it is already zero, so nothing changes). The current digit is i = 1, which is less
than 3; so in step 4 we increase it and return to step 2 with i = 2. Repeating
with the second digit, we compute s2 = c + a2 + b2 = 12. This is greater than
10, so in step 3, set the carry digit c to 1, and reduce the sum digit by 10, so
s2 = 2, the second digit from the right in the sum. In step 4, the current digit
is still less than the number of digits, so we increase i and return to step 2 with
i = 3. With the third digit, we compute s3 = c + a3 + b3 = 5; since this is less
than 10, we set the carry digit to zero. Now i = n, so we meet the stopping
criterion and move to step 5. Since the carry digit is zero, we are done and
return the sum s = 529 (collecting the individual digits computed above). Step
5 would have handled the case when there is a carry digit in the end, meaning
1 This is similar to studying a play or movie. The best way to do this is to first watch a

performance, and then turn to the script to study it in more detail. Starting with the script is
much harder, and not really how a movie is meant to be experienced. So too with algorithms.
588 APPENDIX D. ALGORITHMS AND COMPLEXITY

the sum has more digits than the addends; for instance, the sum of 999 and 111
is 1110, and this step would have added the ‘1’ at the very left.
This algorithm satisfies all the criteria. It is finite, since we repeat the body
once for every digit, and any integer has a finite number of digits. It is definite,
since every step is spelled out precisely without any room for ambiguity. It is
effective, since every step can be done in a finite amount of time. It takes two
inputs (the numbers a and b), and produces one output (the sum s).
An analogy is often made between algorithms, and cooking recipes. For
example, cooking pasta can be expressed by the following recipe:

1. Bring a large pot of water to a rolling boil.

2. Add the pasta, and stir until the water returns to a boil.

3. Wait 9–11 minutes (or until pasta has desired consistency).

4. Drain pasta in a colander.

Which of the criteria of an algorithm are satisfied by a recipe like this? Which
of them are violated, and are they violated in a “weak” way (where you could
satisfy our definition by adding some more explanation to the step) or a “strong”
way (a fundamental violation that can’t be easily fixed)?

D.2 Pseudocode
The previous section gave the formal definition of an algorithm. Sometimes,
specifying an algorithm formally — either writing out all the steps in full detail
in English (as in the addition example above) or with source code in a pro-
gramming language — may be unnecessarily verbose or confusing. Pseudocode
is a less formal way of specifying the steps of an algorithm. Pseudocode can
be thought of as a midway point between intuitive English descriptions (which
express the general idea, but are usually ambiguous) and computer code (un-
ambiguous, but harder to read and understand). The goal of pseudocode is to
express the main ideas of an algorithm while avoiding unnecessary pedantry. It
communicates the algorithm in enough detail that a programmer can write ac-
tual code in a specific language, but without the “fiddly bits” inherent in specific
languages (declaring variables, calling library functions, details of data structure
implementations, and so forth). Algorithm 5 is an example of pseudocode, and
illustrates some common conventions.
This pseudocode is for an algorithm that finds the maximum of n numbers.
As an example, let us find the maximum of five numbers: 7, 3, 5, 9, and 1. In
the initialization, the five numbers are assigned to the list or set or array A, so
that A[1] = 7, A[2] = 3, . . . , A[5] = 1.2 The variable max value is initialized to
a large negative number.
2 Note that some programming languages index arrays from 0, rather than 1; using pseu-

docode lets us abstract from this. The assumption is that a programmer working in such a
language would make the necessary changes.
D.2. PSEUDOCODE 589

Initialize:
A = {A[1], A[2], . . . , A[n]}: Set, list, or array containing the n numbers
max value ← −∞
for i ← 1 to n do
if A[i] > max value then
max value ← A[i]
end
end
Algorithm 5: Finding the maximum of n numbers.

The for loop is a control structure which repeatedly executes a set of instruc-
tions a fixed number of times, as specified by an iterator variable. In this case, i
is the iterator variable which takes values from 1 to n. The variable i takes the
values 1, 2, 3, 4, and 5 in turn; for each of these values, the set of instructions
indented under the for loop is executed. Inside the for loop is an if block. This
is another control structure which executes the indented instructions when (and
only when) a certain condition is satisfied.
Each iteration corresponds to one time that the instructions in the for loop
are executed. We now illustrate a few iterations. At first, the iterator i takes
the value 1. The if statement then checks whether A[1] = 7 > max value =
−∞. Since 7 > −∞, the instructions under the if statement are executed, and
max value now takes the value 7. Next, the iterator variable i takes the value
2. The if statement checks whether A[1] = 3 > max value = 7. This is not
true, so the instructions under the if statement are not executed, and max value
remains at 7. After five iterations, we will find that max value is equal to 9,
which is indeed the maximum of the given five numbers. The stopping criterion
for this algorithm is when the for loop has been executed for all the numbers in
the array.
As another example, the pseudocode in Algorithm 6 represents the algo-
rithm to find the root of a differentiable function f (x) using Newton’s method.
This pseudocode indicates some conventions commonly found in optimization
algorithms. Like the other algorithms seen so far, Newton’s method repeats a
“body” of steps iteratively until the stopping criterion is met. The variable k
keeps track of the number of iterations, and xk is the solution at iteration k.
Initialize:
x0 : initial guess
: tolerance
x1 ← x0 − ff0(x 0)
(x0 )
k←1
while |xk − xk−1 | >  do
xk+1 ← xk − ff0(x k)
(xk )
k ←k+1
end
Algorithm 6: Finding the root of a function f (x).
590 APPENDIX D. ALGORITHMS AND COMPLEXITY

In this implementation, the while loop control structure is used to control the
number of repetitions. In the while loop, a set of instructions is repeated as long
a condition is met. The condition in the while loop corresponds to the stopping
criterion, expressed in a “negative” way (repeat as long as the stopping criterion
is false). In Newton’s method, the algorithm stops when the change in solution
across consecutive iterations |xk − xk−1 | is less than a pre-specified tolerance
level . The tolerance level is chosen depending on the level of accuracy desired.
We illustrate a few iterations of this algorithm for the function f (x) = x2 −3,
with the initial guess x0 = 1. The tolerance  is set at 0.001, which represents
the desired level of accuracy in the output. x1 is now calculated using the
formula to be equal to 2 and iteration counter k is set to 1. The condition in
the while loop |x1 − x0 | is equal to 1 which is greater than epsilon. Therefore,
we now calculate x3 to be equal to 1.75 and increase the value of k to 2. This
process is repeated until the difference between xk and xk−1 is less than 0.001.
The stopping criterion we use (|xk − xk−1 | ≤ ) is one of several widely used
stopping criteria. Another popular criterion is to stop the iterations when the
iteration counter hits a pre-specified maximum number N .
An attentive reader will notice that to be a valid algorithm, it must be finite,
meaning that the convergence criterion must eventually be true. If the stopping
criterion is a maximum number of iterations N , this will clearly happen at some
point. For the stopping criterion |xk − xk−1 | ≤ , it is not obvious that it will
eventually be satisfied (and indeed, for some functions f and initial guesses it
may never be). Analysis of algorithms is the sub-field of computer science which
studies algorithms to identify conditions for finiteness (will it eventually stop?),
proofs of correctness (are the outputs always the ones we want?), and the time
and computer memory they need before stopping. For the most part, this book
will not dwell on these details except where relevant for applying transportation
network models.
In this case, the advantage of the |xk −xk−1 | ≤  stopping criterion is that it is
directly related to the precision in the solution; after a fixed number of iterations
N we may or may not be close to the actual root of the function; or we may have
found the solution many iterations ago but kept on going unnecessarily until we
hit N iterations. However, the k = N stopping criterion has the advantage
of guaranteeing that it will eventually be satisfied. We can combine the two
stopping criteria to gain the advantages of both, by changing the condition in
the while loop, so it reads “while |xk − xk−1 | >  and k < N .” This lets us stop
as soon as we have found the solution, and also guarantees the algorithm will
not run forever.

D.3 Algorithmic Efficiency

Many different algorithms can be devised to solve a particular problem. For
example, there are several methods available in the literature to sort a list
of numbers. How can we choose the “best” one for a particular application?
Algorithm efficiency can be characterized in several ways: run time, memory
D.3. ALGORITHMIC EFFICIENCY 591

use, and so on. In this section, we focusing solely on characterizing efficiency in

terms of run time; similar methods can be applied to characterize memory use
and other resources.
One way to do that would be to implement each algorithm in a programming
language, and measure how long each one takes. A difficulty with this approach
is that this kind of performance comparison depends on the operating system,
programming language, and the hardware specification of the computer used
to test the algorithms. You would also need to decide what kinds of tests
to run: how large should the list of numbers be, how many tests are needed,
and so forth. This section presents an alternative approach based on worst-
case asymptotic analysis, identifying the critical bottlenecks in the algorithm
as the problem becomes harder and harder. By “worst-case,” we mean that
our intention is to find a conservative performance guarantee; for particular
inputs the algorithm may perform better than our analysis shows, but it will
never perform worse. An alternative is to examine “average-case” performance
under a typical range of inputs, but this is trickier because what is “typical”
for one person’s application of an algorithm may be completely different than
another’s. Worst-case performance lets us focus on the algorithm itself, and not
one individual use case or implementation.
A small example illustrates what we mean by worst-case analysis. Imagine
you are the owner of a casino, and you have developed a new game in which a
gambler wagers a certain amount of money. A 6-sided die is then rolled. If the
value is 1 or 2, the gambler loses 50% of the amount wagered; if the outcome is
3 or 4, the entire wager is lost; and if the outcome is 5 or 6, the gambler wins
an amount of money equal to the wager (so their initial stake is doubled). The
process of determining the amount won can be represented by Algorithm 7.

Initialize:
Initialize money bet to the wager:
Obtain outcome by rolling a 6-sided die.
if outcome=1 or outcome=2 then
money returned ← money bet/2
end
else if outcome=3 or outcome=4 then
money returned ← money bet
end
else if outcome=5 or outcome=6 then
money returned ← money bet × 2
end
return money returned
Algorithm 7: Gambling algorithm for demonstrating worst-case analysis.

We want to measure how long it will take for this algorithm to run. We
can either measure this in terms of “clock time” (imagine using a stopwatch to
time how long it takes from start to finish), or by counting the number of oper-
ations that need to be performed. Algorithmic analysis usually uses the latter
592 APPENDIX D. ALGORITHMS AND COMPLEXITY

approach, because it keeps the focus on the algorithm rather than a specific
hardware. The exact same algorithm will run much faster on a supercomputer
than on your phone. As computers continue to advance, it is more useful to say
“this algorithm requires X steps” than “it takes Y seconds to run.”
This requires defining what we mean by an “operation.” Some computations
take much more time than others; computing the square root of a number, or
the cosine of a number, takes significantly longer than changing the sign of a
number from positive to negative, for instance. We will define a basic operation
to be (i) a simple arithmetic computation (add, subtract, multiply, divide); (ii)
testing whether one value is equal to, greater than, or less than another; and
(iii) storing a value in a variable using the assignment operator ←.3
In Algorithm 7, there are two basic operations associated with initialization.
Depending on the value of outcome, it takes a different number of basic opera-
tions to determine which if block is executed. If outcome = 1, we only require
one comparison to see that the money returned should be half the bet (we don’t
have to check whether outcome = 2, because the first if statement has an “or” in
it; as soon as we know outcome = 1 we can jump to the statement in the block.)
If outcome = 2, we need two comparisons: the first checks whether outcome
is 1 (false), the second checks whether outcome is 2 (true), and only then we
execute the statement. If outcome = 3, we need three comparisons; two failed
comparisons testing whether outcome is 1 or 2, and a third successful one. In
the worst case, outcome = 6, and we need six comparisons to determine which
block of code to execute. Within the if blocks, there are either zero arithmetic
operations (if the outcome is 3 or 4) or one arithmetic operation (all other out-
comes) multiplying or dividing the wager by 2. There is always one assignment
operation in each block.
Putting this together, how many basic operations does the algorithm need?
In the best case, outcome = 1, and a total of five basic operations are needed
(two initialization, one comparison, one arithmetic, and one assignment). In the
worst case, outcome = 6, and ten operations are needed (two initialization, six
comparison, one arithmetic, one assignment). The average case can be treated
by representing the operation count as a random variable. Its expected value can
be calculated by identifying all possible operation counts under every outcome,
multiplying by the probability of that outcome, and adding. The resulting
number is 7.2, representing the long-run average of the number of operations
needed.
Which one of these is most useful? The average case might seem to be best,
but its calculation was more involved, and it would have to be recomputed if a
different kind of die were used where all six outcomes were not equally likely.
By contrast, the best case and worst case were much simpler to compute, and
they would stay the same even if the die were to be weighted more to some
outcomes than others. Between these two, the worst case is more useful for
planning purposes. The best case is optimistic to a fault; all it tells you is that
3 By “value” in this section, we mean an integer, or a floating-point number rounded to a

finite number of digits.

D.3. ALGORITHMIC EFFICIENCY 593

you’ll need at least five basic operations (but nothing about how bad it could
get). This is akin to a shady investment pitch which only tells you how much
you could earn, and nothing about how much you stand to lose. The worst case,
by contrast, is a more solid basis for planning and allocating resources. You will
never need more than ten operations (and often will need less), so if you plan
for this you will never exceed your resource budget.
In this case, the performance of the algorithm requires a constant number
of steps. What if it is repeated n times (either the same gambler playing the
game multiple times, or several gamblers playing)? The worst case is that ten
operations are needed each time; in this case 10n basic operations are needed
to run the game. Again, this is a conservative estimate. The only time 10n
operations are needed is if a 6 is rolled all n times the game is played, which is
unlikely. However, we can absolutely guarantee that you will never need more
than this.
As another example, return to Algorithm 5 for finding the largest of n num-
bers. In the for loop, for each value of i we perform at most two basic operations:
one comparison operation (testing whether A[i] > max value) and one assign-
ment operation (max value ← A[i]). Since i takes values from 1 to n, these two
operations will be repeated n times, giving a total of 2n operations. There is
also one basic operation associated with initialization, for a total of 2n + 1 basic
operations in the worst case. (The best case has fewer than this; for example, if
the A[1] is the largest of the n numbers, then only n + 2 operations are needed,
because max value is only assigned once.)
To further simplify the worst-case analysis, we introduce an asymptotic no-
tation which highlights the bottleneck as the problem grows harder. For finding
the maximum of n numbers, the one operation needed for initialization becomes
insignificant if n is large. If n is equal to a million, say, what really matters are
the (up to) two million operations needed in the body of the algorithm, not the
one needed for initialization. So we would like to focus on the 2n part and not
the +1 part. Even here, we can simplify further. The most important part is
that it grows with n (and not n2 , say): doubling n will also double the number
of steps needed. The factor of 2 is less important.
To see why, imagine that someone proposes an alternative algorithm for
finding the maximum of n numbers, and worst case analysis shows that it may
require up to n2 +4n+5 operations. As n grows large, the n2 part will dominate
the remaining terms. Furthermore, if n were to double, n2 would increase by a
factor of four. This is evidence that Algorithm 5 is a better choice. As n gets
larger and larger, the number of operations needed by Algorithm 5 increases in
direct proportion, whereas that needed by the hypothetical alternative grows
with the square of n. The fact that the first algorithm is 2n, as opposed to n is
much less important than that it is linear in n and not quadratic.
Another reason to focus on the n and not the 2 is to abstract certain imple-
mentation details. In Algorithm 5, do we count an initialization step associated
with the first line, setting the values in the array A? Maybe it is passed di-
rectly as an address in computer memory, to values that already exist; in this
case there is no new operation. On the other hand, if we have to assign them
594 APPENDIX D. ALGORITHMS AND COMPLEXITY

from other values in memory, we need an additional n assignment operations,

for a total of 3n + 1. Or maybe we have to read them from a file, requiring n
operations to read them from the drive and n assignment operations to store
them in A, for a total of 4n + 1. In all of these cases, the most important part
is that the computational requirements scale proportionally to n, since this will
tell you how the run time will change if you give it a larger or smaller problem
instance. Whether the coefficient is 2, 3, or 4 is of secondary importance.
We will introduce an asymptotic notation that lets us simplify in this way;
we want a concise way to express that as n grows large, the number of steps
required by Algorithm 5 grows in direct proportion to n, without fiddling around
with the constants in the full expression 2n + 1. The notation will say that
the number of steps required by Algorithm 5 is O(n) (read “big-oh of n”).
Technically, we say that a function f (n) is O(g(n)) if there are constants A and
n0 such that |f (n)| ≤ Ag(n) whenever n > n0 . Practically, we can identify
such a function g(n) by selecting the dominant term in f (n) as n → ∞, and
removing any constant coefficient it has. For example, n2 + 4n + 5 is O(n2 ),
and 4 · 2n + 6n3 − 8 log n is O(2n ). The initial version of the gambling algorithm
(Algorithm 7) required at most 10 steps; we can say this is O(1).
Some common asymptotic notations that arise are:

ˆ O(1), for algorithms that run in constant time (independent of input size);

ˆ O(log n), sometimes called logarithmic time or sublinear time (the run
time increases more slowly than the problem size);

ˆ O(n), or linear time (run time scales in direct proportion to the problem
size);

ˆ O(n2 ), or quadratic time;

ˆ O(n3 ), O(n4 ), and other polynomial time algorithms of the form O(nc )
for some constant c.

ˆ O(2n ), or exponential time;

ˆ O(n!), or factorial time.

These are listed in increasing order of difficulty. Generally, algorithms up to

polynomial time are considered to be reasonably efficient. Algorithms with ex-
ponential time complexity are difficult to solve for large instances in reasonable
amount of time.
Some people think that this classification is not very useful, because comput-
ers continue to get faster over time. However, the difference between complexity
classes is absolutely critical, and will remain so for the foreseeable future. Imag-
ine you have two algorithms for a network optimization problem with n nodes;
one runs in quadratic time, O(n2 ), and the other runs in exponential time,
O(2n ). Also imagine that using either algorithm, on your current computer you
can just barely solve a problem with 1000 nodes in an acceptable amount of
D.3. ALGORITHMIC EFFICIENCY 595

time. After some time, you upgrade your computer so that it is twice as power-
ful as the one you had before. Roughly speaking, the quadratic-time algorithm √
on this better machine can handle a network which is larger by a factor of 2,
so roughly 1400 nodes. The exponential-time algorithm, on the other hand, can
only handle a network with one more node, since 21001 is twice as large as 21000 !
Faster hardware will not overcome the disadvantages of an exponential-time
algorithm, at least for large problems.
However, just because an algorithm is exponential doesn’t mean it is use-
less. For one, the O(·) notation is asymptotic, and tells you what will happen
as n tends to infinity. An exponential time algorithm may be perfectly fine
working with the problem sizes you are dealing with; it just means that it does
not scale well, and for large enough problems you will run into difficulties. It
is also worst case, and there are algorithms which may have exponential worst
case complexity, yet behave more like polynomial time in practice. The simplex
method for solving linear programs (Section C.2.3), and the branch and bound
method for solving mixed-integer linear programs (Section C.5.1) are examples
of algorithms which are exponential in the worst case, yet offer better perfor-
mance in practice, especially when implemented well and tailored to a specific
application.
For one more example of complexity calculation, Algorithm 8 presents the
“bubble sort” algorithm for sorting an array of n numbers into increasing order.
The algorithm has two loops: an outer loop using the iterator i, and an inner
loop with the iterator j. When i = 1, there are 4(n − 1) operations associated
with the j loop. When i = 2, there are 4(n−2) operations, and so on. Therefore
in total there are 4(n−1)+4(n−2)+. . .+1 = 2n(n−1) operations. The bubble
sort is therefore O(n2 ) in complexity, and has quadratic time. This implies that
if the input size doubles, the running time quadruples. There are faster sorting
methods with O(n log n) complexity; if you have to sort a large array of numbers
this will perform much better than bubble sort. However, bubble sort is much
easier to implement, and may perform acceptably if you only have to sort a
small array.

Initialize:
A: Set, list, or array containing the n numbers
for i ← 1 to n − 1 do
for j ← 1 to n − i do
if A[j] > A[j + 1] then
temp ← A[j]
A[j] ← A[j + 1]
A[j + 1] ← temp
end
end
end
Algorithm 8: Bubble sort
596 APPENDIX D. ALGORITHMS AND COMPLEXITY

D.4 Complexity Classes

The previous section introduced asymptotic notation, and how it is used to

characterize the worst-case run time of an algorithm, measured by counting the
number of basic operations needed. This is useful for comparing two algorithms
for the same problem. As the input size of a problem becomes larger and larger,
any O(1) algorithm will eventually run faster than any O(n) algorithm, which
will run faster than any O(n2 ) algorithm, which will run faster than any O(2n )
algorithm, and so forth. But can we say anything about how hard a problem is?
In other words, I might prefer a O(n2 ) algorithm to an O(2n ) one. But is O(n2 )
the best I can do? Or with some more ingenuity and work, could we come up
with something even better?
Complexity classes aim to understand hardness in terms of the underlying
problems. As explained in more detail below, one such class of problems is
called P ; these are problems for which a polynomial-time algorithm exists to
solve it. A larger class of problems is called N P ; for some of these problems,
we do not know of any polynomial-time algorithm that will always solve them.
Some N P problems are called N P -complete, which means that any polynomial-
time algorithm solving them would also solve any other problem in N P . Such
an algorithm would be extremely valuable, and decades of work has gone into
trying to find one. This search has thus far proved fruitless.
?
This is the core of the famous P = N P question, arguably the most im-
portant open problem in theoretical computer science, and one carrying a $1
million bounty from the Clay Mathematics Institute. Answering this question
either way (providing such an algorithm, or definitively proving that none exists)
would be a tremendous accomplishment. Indeed, even though this core question
has as yet gone unanswered, the search for its answer has yielded many other
important findings in computer science. Aaronson (2016) provides an excellent
survey of the history of this question, attempts to answer it, and the value this
search has provided
Many students learning about complexity classes initially find them an eso-
teric theoretical conceit. However, there is great practical significance! If your
supervisor asks you to solve a particular problem, they may not be happy if
the best you can come up with is an exponential time algorithm, since it will
scale badly for large-scale problems. However, if you can show this problem
is N P -complete, then you can point out that it’s not just you who can’t find
a better algorithm — the brightest minds in computer science have also been
unable to find a better algorithm for this problem! In other words, it’s not you,
it’s the problem that’s inherently difficult. If you can suggest a simplification
of the problem that brings it into P , then this directly suggests a faster way to
solve it.
This section provides a brief overview of complexity classes, emphasizing
applications to transportation network optimization problems.
D.4. COMPLEXITY CLASSES 597

D.4.1 Decision problems vs. optimization

In computer science, complexity classes are usually defined in terms of decision
problems, where the answer is either “yes” or “no.” This book primarily dis-
cusses optimization problems, where the answer involves a solution in terms of
decision variables and an objective function. However, optimization problems
can be transformed into decision problems. Here are a few examples of how to
do this:

Shortest path problem: Consider a directed network G = (N, A), where N

is the set of nodes, and A the set of directed links, with non-negative integer
costs cij for each (i, j) ∈ A. Let r and s denote the origin and destination node.
Optimization problem: Find the path π between r and s minimizing its cost
cπ = (i,j)∈π cij .
P
Decision problem: For a given value C, is there a path π between r and s
whose cost cπ does not exceed C?

Traveling salesperson: Consider a complete network G = (N, A), where N

is the set of nodes and A the set of links, with integer costs cij .
Optimization problem: Determine a minimum cost tour which visits each
node exactly once. (In other words, find a permutation of nodes (p1 , p2 , . . . , pn )
Pn−1
which minimizes i=1 cpi pi+1 + cpn p1 .)
Decision problem: For a given value C, is there a tour visiting each node
exactly once whose cost does not exceed C? (In other words, is there a permu-
Pn−1
tation of nodes such that i=1 cpi pi+1 + cpn p1 ≤ C?)

Knapsack problem: Consider a set of objects I with positive integer sizes

ai and profits pi for each i ∈ I, and a positive integer U denoting the capacity.
Optimization
P problem:P Find a subset of objects JI ⊂ I maximizing total
profit j∈JI pj such that j∈JI aj ≤ U .
P profit level P , is there a set of objects JI ⊂ I
Pproblem: For a given
Decision
such that j∈JI pj ≥ P and j∈JI aj ≤ U .

Resource-constrained shortest path problem: Consider a directed net-

work G = (N, A), where N is the set of nodes, and A the set of directed links,
with positive integer costs cij and resources rij for each (i, j) ∈ A. Let r and s
denote the origin and destination node, and R the resource budget.
Optimization problem: Find the path P π between origin node s and destina-
tion node t with minimum cost cπ = (i,j)∈π cij , subject to the resource usage
P
constraint (i,j)∈π rij ≤ R.
Decision version:
P For a given value C, is there a path π between s and t such
that cπ ≤ C and (i,j)∈π rij ≤ R?
Longest acyclic path problem: Consider a directed network G = (N, A),
where N is the set of nodes, and A the set of directed links, with positive integer
costs cij for each (i, j) ∈ A. Let r and s denote the origin and destination node.
598 APPENDIX D. ALGORITHMS AND COMPLEXITY

Optimization P problem: Find the acyclic path π between r and s with maxi-
mal cost cπ = (i,j)∈π cij .
Decision problem: For a given value C, is there an acyclic path π between r
and s whose cost cπ is at least C?
From the standpoint of solution, a polynomial time algorithm for the op-
timization problem also leads to a polynomial time algorithm for the decision
problem, and vice versa. For example, if we have a polynomial time algorithm
for the optimization version of the shortest path problem, we can use it to solve
the decision problem: just find the shortest path using the optimization version;
if its cost is less than or equal to C, then the answer to the decision problem is
“yes.” Otherwise it is “no.” Or, if we have a polynomial time algorithm for the
decision version of the problem and want to find the exact cost of the shortest
path (not just whether it is less than C or not), we can repeatedly solve the de-
cision version problem for different versions of C using a bisection-like approach
(cf. Section 3.3.2). This requires solving the decision problem multiple times,
but a polynomial number of times; substituting one polynomial into another
still results in a polynomial, so the overall method is still polynomial time (even
if the polynomial has a higher degree).

D.4.2 The classes P , N P , N P -hard, and N P -complete

This subsection introduces four classes of problems that arise often in optimiza-
tion and elsewhere in computer science: P , N P , N P -hard, and N P -complete.
They are formally defined in terms of decision problems like the ones in the pre-
vious subsection, where the answer to a posed question is either “yes” or “no.”
These four classes are not mutually exclusive, so a given problem may exist in
more than one of these classes at the same time — for example, the standard
shortest path problem is in both P and N P . They are also not exhaustive;
there are decision problems which do not lie in any of these four classes.
We start by defining the classes P (“polynomial”) and N P (“nondeterminis-
tic polynomial,” and not “non-polynomial” as is often guessed by people seeing
the acronym for the first time). The definitions given here are not fully rigor-
ous, since we should also state exactly what our model of computation is, what
we count as a basic operation, the exact format of the problem input, and so
on. Readers interested in these details should consult Sipser (2012) or another
standard text in the theory of computation. The definitions below will suffice if
you are using common hardware, programming languages, and data formats.
Definition D.1. A decision problem belongs to the class P if, given an instance
of size n, the answer can be obtained using an algorithm whose worst-case run-
ning time is polynomial, that is, O(nk ) for some constant k.
Definition D.2. A decision problem belongs to the class N P if, given an in-
stance of size n whose answer is “yes,” there is a polynomial-time algorithm to
confirm that the answer is in fact “yes.”
Worded intuitively, a problem in P is easy to solve; a problem in N P is easy
to check, in the sense that someone claiming the answer is “yes” can provide
D.4. COMPLEXITY CLASSES 599

evidence that you can verify in polynomial time. A few examples are shown
below; keep in mind that all of these refer to the decision versions of these
problems described in the previous subsection, not the optimization versions.
The decision version of the shortest path problem with non-negative link
costs is in P , since Dijkstra’s algorithm (and others) can find the shortest path
in polynomial time. So we simply compute the shortest path; if its cost is less
than or equal to C, the answer to the decision problem is “yes” (there is a path
of cost at most C), otherwise the answer is “no” (no such path exists).
The decision version of the traveling salesperson problem is in N P . If the
answer is “yes” (there is a tour visiting each node exactly once whose cost is
less than or equal to C), someone can convince you of this by providing the
tour itself as evidence. You can check in polynomial time whether this tour is
valid (it visits every node once and only once) and whether its total cost does
not exceed C. We have said nothing about how someone might find such a tour
in the first place; simply that given such a tour as supposed evidence that the
answer is “yes,” you can check this evidence in polynomial time. Likewise, we
have not said anything about “no” instances of the problem, in which all tours
have a cost greater than C. Indeed, there does not seem to be an easy way to
demonstrate that a given instance of the traveling salesperson problem is “no.”
Just providing a tour whose cost is more than C does not prove that there isn’t
some other tour with a lower cost. But this doesn’t matter; the class N P is
only concerned with “yes” instances of decision problems.4
Similarly, the decision versions for the knapsack problem, longest path prob-
lem, and resource-constrained shortest path problems are in N P , since in any
case where the answer is “yes” you can provide a feasible knapsack assignment
or path as easily-checked evidence that the answer really is “yes.”
Every decision problem in P is also in N P . Given a “yes” instance of the
problem, I can check whether the answer is indeed “yes” by simply answering
the question myself using a polynomial-time algorithm (which exists because
the problem is in P ). So the decision version of the shortest path problem is in
both P and N P .
We do not know if the reverse is true, that every problem in N P is also in
P . For instance, we know that the traveling salesperson problem is in N P , but
we do not know if it is in P or not. We do not currently know of any polynomial
time algorithm for solving it, but this does not mean that none exists. Perhaps
there is such an algorithm, but if so nobody has been clever enough to find it
yet. Most computer scientists believe that no such algorithm exists, but nobody
has been able to prove that either. Answering the question either way would be
a tremendous advance in computer science.
The reason is that algorithms used to solve one kind of problem can often be
used to solve other kinds of problems as well. Many important problems (even
problems with no obvious “network” in them) can be reduced to the traveling
salesperson problem in the sense that an algorithm for the latter can also be used
to solve the former, by constructing a network, choosing costs in the right way,
4 The complexity class co-N P is the analogous class for “no” instances.
600 APPENDIX D. ALGORITHMS AND COMPLEXITY

and then translating the resulting tour back to the original problem. So, finding
a good way to solve the traveling salesperson problem would directly provide
good ways to solve many other problems as well. This idea is formalized in the
following definition:
Definition D.3. A decision problem A can be polynomially reduced to decision
problem B if (a) there is a procedure that can transform an instance of A to an
instance of B in polynomial time; and (b) the answer for an instance of A is
“yes” if and only if the answer for the corresponding instance of “B” under this
transformation is also “yes.”
The implication is that if we know a polynomial time algorithm for B, then
we can use it to solve A. The number of steps needed to do this is the sum of the
number of steps needed to construct an instance of problem B from the given
instance of A, and the number of steps needed to answer that instance of B. If
there is a polynomial reduction from A to B, and if B is solvable in polynomial
time, then the total number of steps required is also polynomial. Intuitively,
the problem A is no harder than B, since any method for B can also be used to
solve A. (The problem A could be easier than B, in the sense that a specialized
method for A might be a faster approach than translating the problem to B
and solving that; we just know it can’t be harder.)
We give a few examples of these kinds of transformations for decision prob-
lems in the next subsection. The main reason for introducing polynomial re-
ductions is to define two additional complexity classes: N P -hard, and N P -
complete.
Definition D.4. A decision problem X belongs to the class N P -hard if every
problem in the class N P has a polynomial reduction to X.
In other words, any problem in class N P -hard is at least as difficult to solve
as any problem in N P , since any problem in class N P can be converted to an
N P -hard problem. N P -hard problems are not necessarily in N P ; they might
be strictly more difficult in the sense that we might not be able to verify a “yes”
answer to an N P -hard problem in polynomial time.
Definition D.5. A decision problem belongs to the class N P -complete if it is
both in classes N P and N P -hard.
The problems in N P -complete can be thought of as the most difficult prob-
lems in N P , because any other N P problem can be converted to an N P -
complete problem which will have the same answer. The consequence of this
is that if someone can develop a polynomial time algorithm for any problem
in N P -complete, then we can answer all problems in N P in polynomial time
as well; therefore the classes P and N P will be the same: P = N P . On the
other hand, proving that no polynomial time algorithm exists to answer an
N P -complete problem would show that there are some problems in N P which
are not in P , so P 6= N P . This makes the class N P -complete very important.
The decision version of the traveling salesperson problem is an example of an
D.4. COMPLEXITY CLASSES 601

Figure D.1: P vs NP vs NP-complete vs NP-hard

N P -complete problem. The proof of this statement is difficult and beyond the
scope of this book.

We summarize the discussion in this subsection in Figure D.1, which illus-

trates the relationships between the four complexity classes discussed. The left
panel shows how the classes overlap if P 6= N P ; the right panel shows what
it would look like if P = N P . Again, determining whether P and N P are in
fact equal is perhaps the most important open question in computer science.
Most experts believe that P 6= N P is the likelier answer to this question, but
this is not a universally-held opinion, and in any case belief is not proof (there
are plenty of true mathematical results whose discovery came as a surprise to
expert mathematicians).

As a final word, we comment on some differences between decision problems

(the kind used to formally define the complexity classes in this section) and op-
timization problems (the kind arising more when doing transportation network
analysis in practice). For instance, the decision version of the traveling salesper-
son problem is in N P (a given tour can be verified to be feasible and have cost
less than C in polynomial time), but the optimization version is not (there is no
easy way to prove that a given tour is in fact minimal across all feasible tours).
Sometimes, researchers in operations research use this terminology a bit loosely,
and refer to a problem as being “N P -complete” (or similarly) without explic-
itly saying whether they are referring to the decision or optimization versions
of these problems. This is unfortunate, and if you are conducting research in-
volving the complexity of transportation network problems we recommend that
you become familiar with the distinction between the two.
602 APPENDIX D. ALGORITHMS AND COMPLEXITY

D.4.3 Examples demonstrating N P -completeness

One of the primary uses of complexity classes is to establish that a given problem
is intrinsically hard. In particular, if you can show a problem is N P -complete,
then expecting to develop a polynomial time algorithm is more than simply
an ambitious goal; it would be a groundbreaking advance in computer science.
For practical application, we are therefore content with heuristics that provide
reasonably good solutions in a reasonable amount of time, rather than exact
methods, since we do not know of any that scale well with problem size (at
least in the worst case). This subsection provides some examples of how you
can show a decision problem is N P -complete, using the idea of polynomial
reductions introduced above.
The typical method for showing a decision problem X is N P -complete has
two steps:
ˆ Showing that X is in NP. That is, given an instance of X where the answer
is “yes,” it is possible to demonstrate this in polynomial time.
ˆ Providing a polynomial reduction from any other N P -complete problem
to X. This shows that X is at least as difficult as any problem in N P
(and therefore N P -hard); together with the proof that X is also in N P ,
this establishes N P -completeness.
The second of these is more challenging and can require more creativity than the
first. Notice that this method requires that we already know some problems are
N P -complete. Currently, hundreds of decision problems are known to be N P -
complete, and you can use any of them as the starting point for the second step
in a proof of N P -completeness. (The very first proof of N P -completeness for a
decision problem was much harder, since the above scheme could not be followed.
Instead, one had to give a general argument that any N P problem could be
reduced to that problem!) Garey and Johnson (1979) is a good reference listing
a variety of problems known to be N P -complete.
For example, the following decision problems are known to be N P -complete:
Hamiltonian cycle problem: Given a directed network G = (N, A), is there
a cycle which visits every node exactly once?
Partition problem: Given a set S of positive integers (possibly with multiple
copies of each number), can the set be partitioned into two subsets S1 and
S2 such that the sum of the integers in S1 equals the sum of the integers
in S2 ?
Using these, we can prove that the traveling salesperson problem, knapsack
problem, and resource-constrained shortest path decision problems introduced
above are all N P -complete. Let us look at the NP-completeness proof of the
decision problems discussed above.
Example D.1. Show that the decision version of the traveling salesperson prob-
lem (given in Section D.4.1) is N P -complete.
D.4. COMPLEXITY CLASSES 603

Solution. The first step is to show that this problem is in N P ; we have

already shown in the previous section that we can check that a tour is valid,
and that its total cost is at most C, in linear time, so we can verify a “yes”
instance in polynomial (in fact, linear) time.
The second step involves a polynomial reduction from a known N P -complete
problem. In this case, we select the Hamiltonian cycle problem, and produce
a reduction that would allow us to answer the Hamiltonian cycle problem by
solving a suitable traveling salesperson problem. (The order of the reduction
is critical; showing that we could answer the traveling salesperson problem by
solving a suitable Hamiltonian cycle problem would not be helpful. That would
show that the Hamiltonian cycle problem is at least as hard as the traveling
salesperson problem; but we already know that the Hamiltonian cycle problem
is N P -complete. Our goal is to say something about the traveling salesperson
problem, so we have to proceed in the other direction and show that we can use
it to answer a Hamiltonian cycle question.)
So, assume we are given an instance of the Hamiltonian cycle problem with
a network G = N (A, A), and let n = |N | denote the number of nodes in the
network. A Hamiltonian cycle visits every node exactly once before returning to
the starting node, so such a cycle will contain n links. Construct a new network
G0 = (N 0 , A0 ) in the following manner: the node sets are identical (N 0 = N ),
but G0 is a complete graph, meaning there is a link (i, j) between every pair of
nodes. The costs in G0 are chosen in the following way: cij = 1 if (i, j) ∈ A,
and cij = 2 if (i, j) 6∈ A. That is, if the link exists in the original network, it has
a cost of 1; if the link did not, it has a cost of 2. This transformation can be
accomplished in polynomial time, since the number of links in G0 is n(n − 1).
Now, consider the decision version of the traveling salesperson problem in
G0 , with C = n. If the answer to this problem is “yes,” then there is a tour
visiting every node and involving exactly n links, all with a cost of 1. All of these
links therefore exist in the original network G, and correspond to a Hamiltonian
cycle, so the answer to the Hamiltonian cycle problem is also “yes.”
Or, if the answer to the traveling salesperson problem is “no,” then every
tour in G0 has at least one link with a cost of 2. No such links exist in the
original network G, so there is no Hamiltonian cycle either: the answer to that
problem is also “no.”
This completes the reduction; we have already shown that it can be accom-
plished in polynomial time, and that the traveling salesperson problem is in
N P . Therefore it is N P -complete. 
Essentially, we showed that any procedure which can solve the traveling
salesperson problem could also solve the Hamiltonian path problem, and there-
fore it cannot be any easier.
Example D.2. Show that the decision version of the knapsack problem (given
in Section D.4.1) is N P -complete.
Solution. We start by showing that a “yes” instance can be verified in
polynomial time to establish that the problem is in N P . If the answer is “yes,”
one can prove this by furnishing a subset JI whose total profit is at least P ,
604 APPENDIX D. ALGORITHMS AND COMPLEXITY

and whose total size is no greater than U . The total profit and total size can be
calculated in linear time by simply summing the profit and size of each object
in JI .
The remainder of the argument establishes a polynomial reduction from an
N P -complete problem. For this example, we will reduce from the partition
decision problem, with a given set of positive integers S. From this, we will
construct an instance of the knapsack decision problem whose answer (“yes” or
“no”) will always be the same as that for the given partition problem. Let S =
{b1 , b2 , . . . , bn } (possibly with repetitions). Create an instance of the knapsack
problem where the set of objects is the set of integers I = {1, 2, . . . , |S|}. The
size of each object, and its profit, are both set equal to the corresponding integer
in S, so ai = piP = bi . The available capacity is half of the sum of the integers
in S, so U =P21 i∈S bi . The minimum profit P is set to this same amount as
well: P = 12 i∈S bi . This reduction can be done in linear time.
Assume that the answer to the P knapsack problem isP“yes.” Then there
is a subset of objects JI such that a ≤ U and j∈JI pj ≥ P . But
Pj∈JI j
U = P = 12 i∈S bi , so this implies JI aj = 12 i∈S bi . Furthermore, each
P P

aj = bj , so j∈JI bj = 21 i∈S bi . That is, the integers represented by JI form

P P
exactly half of the total sum of the integers in S. Therefore, the remaining
integers not in JI must also form exactly half the total sum of the integers in
S, so the sets JI and I − JI have equal sums, and the answer to the partition
problem is “yes.”
Now assume that the answerP to the knapsack problem is “no.” Then every
subset of objects JI either has j∈JI aj > U (total size exceeds capacity) or
1
P P
j∈JI pj < P (total profit falls short of target). Since U = P = 2 i∈S bi and
ai = pi = bi , this means that no subset JI has the property that its integers add
up to exactly half of the total of the integers in S. This means that the answer
to the partition problem must also be “no,” since if there P were sets S1 and S2
whose sums were equal, they would both have to equal 12 S bi , and JI = S1
would be a solution to the knapsack problem.
We can therefore answer the partition problem (known to be N P -complete)
by solving the knapsack problem, so the knapsack problem must be at least as
hard. Since we have shown that the knapsack problem is in N P , it is therefore
N P -complete as well. 

Example D.3. Show that the decision version of the resource-constrained short-
est path problem (given in Section D.4.1) is N P -complete.

Solution. This problem is in N P , since a “yes” instance can be verified in

linear time once an appropriate path is provided as evidence; simply check that
the path is valid, and compute two sums over its links to verify that its total
cost is at most C, and that its total resource consumption is at most R.
To show N P -completeness, we reduce the partition decision problem to the
resource-constrained shortest path problem. Given the set S = {b1 , b2 , . . . , bn },
we construct
P a network G = (N, A) with node and link sets defined as follows.
Let U = bi ∈S bi , and V = 2nU . We create n + 1 nodes; and between each
D.4. COMPLEXITY CLASSES 605

Figure D.2: Network transformation to reduce the partition problem to resource-

constrained shortest paths.

consecutive pair of nodes i and i + 1, we create two parallel links. In each pair
of parallel links, the top link has a cost of V , and a resource consumption of 0.
The bottom link has a cost of V − bi , and a resource consumption of bi . (See
Figure D.2.) Finally, for this instance of the resource-constrained shortest path
decision problem, set the cost target C = nV − (U/2), the resource limit to U/2,
and select nodes 1 and n + 1 as the origin and destination. The number of links
and nodes is linear in the size of the original set S, so this transformation can
be done in polynomial time.
Assume that the resulting resource-constrained shortest path problem is a
“yes” instance. Let π be a path whose cost is at most C, and whose resource
consumption is at most R. This path consists of n links, each of which is either
the top or bottom link in the n pairs of parallel links in the network. Let
S1 contain the values of S corresponding to “bottom” links, and S2 P the values
corresponding to “top” links. Then thePtotal cost of this path is nV − bi ∈S1 bi ,
and its total resource consumption isP bi ∈S1 bi . Since this is a “yes”Pinstance,
the cost target is achieved and nV − bi ∈S1 bi ≤ C = nVP− (U/2), so bi ∈S1 ≥
P the resource constraint is satisfied, so bi ∈S1 bi ≤ R = U/2.
U/2. Likewise,
Therefore bi ∈S1 is exactly U/2, that is, the sum of the entries in set S1 is
exactly half the total of the sum of the entries in S. We can thus conclude
that the sum of the values in sets S1 and S2 are equal, and this is also a “yes”
instance of the partition problem.
Or, if this is a “no” instance of the resource-constrained shortest path prob-
lem, then every path either exceeds the cost target or the resource limitation.
Following the same construction Pas in the previousP paragraph, this means that
for every set S1 we have either bi ∈S1 bi < U/2 or bi ∈S1 bi > U/2. Therefore,
there is no subset of the values in S whose sum is exactly half of the total sum
of the values in S; so it is not possible to partition S into two sets S1 and S2
whose sums are exactly equal. Therefore this is a “no” instance of the partition
problem as well. 
You may have noticed that this proof was fairly similar to that used to
show that the knapsack problem is N P -complete. In fact, once Example D.2
was done, and we established that the knapsack problem was N P -complete, we
could have done Example D.3 in a faster way by reducing the knapsack problem
to the resource-constrained shortest path problem (rather than starting from the
606 APPENDIX D. ALGORITHMS AND COMPLEXITY

partition problem anew). Verifying this would be a very instructive exercise.

Example D.4. Show that the decision version of the longest acyclic path prob-
lem (given in Section D.4.1) is N P -complete.
Solution. For a “yes” instance, verifying that a given path is feasible and
has cost at least C can be done in linear time, so this problem is in N P .
To show N P -completeness, we reduce the Hamiltonian cycle problem to
the longest acyclic path problem. The transformation proceeds as follows: let
G = (N, A) be the given network for the Hamiltonian cycle problem (where n
is the number of nodes). Create a network G0 with n + 1 nodes in the following
way: select any node i in N , and split it into two nodes i+ and i− . Any link
(i, j) in the forward star of node i appears in G0 as a link from i+ to j. Any
link (h, i) in the reverse star of node i appears in G0 as a link from h to i− . All
other links in G appear in G0 as they are, connecting the same nodes. Select
i+ as the origin, and i− as the destination, and finally select the cost threshold
C = n to create an instance of the longest acyclic path problem in G0 .
If this is a “yes” instance for the longest acyclic path problem, then there
is a path π whose cost is at least n. Since the path is acyclic, it cannot revisit
any node; and since there are n + 1 nodes in G0 the path π must pass through
all of them. This acyclic path can be translated to a Hamiltonian cycle on G
by replacing i+ and i− in π by the single node i.
If this is a “no” instance for the longest acyclic path problem, then every
path π in G0 has cost less than n, meaning it does not pass through every
node. In this case, the original Hamiltonian cycle problem must also be a “no”
instance, since any Hamiltonian cycle can be translated to a path in G0 from i+
to i− passing through every node. 
Two final comments on this subject. First, pay attention to the problem
specifics. Minor changes can have a significant impact on the complexity of
a problem. For example, the Hamiltonian cycle problem (determining whether
there is a path visiting each node exactly once) is N P -complete; but the Eulerian
cycle problem (is there a path visiting each link exactly once) can actually be
solved in linear time. The shortest path problem (with positive link costs) can
be solved in polynomial time; the shortest path problem with negative link costs
and the longest (acyclic) path problem are both N P -complete.
Second, just because a problem is N P -hard or N P -complete does not mean
the problems are impossible to solve, or completely impractical. Such problems
are often solved either using heuristics (which are inexact, but may be good
enough), exact methods which are usually fast enough in practice (even if in the
worst case the number of steps is exponential), or for small problem instances
where an exponential time algorithm may still be acceptable. As some con-
crete examples, integer optimization problems (Section C.5) are NP-hard. But
several algorithms have been developed that often solve real world instances
efficiently. Knapsack problems can be solved by pseudo-polynomial algorithms,
a complexity class where the number of operations is bounded by the value of
the input, and not its absolute size. These problems are also commonly solvable
in practice.
Bibliography

?
Aaronson, S. (2016). P = N P . In Open Problems in Mathematics, pp. 1–122.
Springer.
Aashtiani, H. Z. and T. L. Magnanti (1981). Equilibria on a congested trans-
portation network. SIAM Journal on Algebraic and Discrete Models 2, 213–
226.
Ahuja, R. K., T. Magnanti, and J. Orlin (1993). Network Flows. Englewood
Cliffs, NJ: Prentice-Hall.
Akamatsu, T. (1996). Cyclic flows, Markov process and stochastic traffic as-
signment. Transportation Research Part B 30, 369–386.
Akamatsu, T. (1997). Decomposition of path choice entropy in general transport
networks. Transportation Science 31, 349–362.
Andreatta, G. and L. Romeo (1988). Stochastic shortest paths with recourse.
Networks 18, 193–204.
Ban, X. J., J.-S. Pang, H. X. Liu, and R. Ma (2012). Modeling and solving
continuous-time instantaneous dynamic user equilibria: A differential com-
plementarity systems approach. Transportation Research Part B 46 (3), 389–
408.
Bar-Gera, H. (2002). Origin-based algorithm for the traffic assignment problem.
Transportation Science 36 (4), 398–417.
Bar-Gera, H. (2005). Continuous and discrete trajectory models for dynamic
traffic assignment. Networks and Spatial Economics 5, 41–70.
Bar-Gera, H. (2006). Primal method for determining the most likely route flows
in large road networks. Transportation Science 40 (3), 269–286.
Bar-Gera, H. (2010). Traffic assignment by paired alternative segments. Trans-
portation Research Part B 44 (8–9), 1022–1046.
Bar-Gera, H. and D. Boyce (1999). Route flow entropy maximization in origin-
based traffic assignment. In Proceedings of the 14th International Symposium
on Transportation and Traffic Theory, Jerusalem, pp. 397–415.

607
608 BIBLIOGRAPHY

Barcelo, J. (1998). Parallelization of microscopic traffic simulation for ATT

systems analysis. In P. Marcotte and S. Nguyen (Eds.), Equilibrium and
Advanced Transportation Modelling. Kluwer Academic Publishers.

Bazaraa, M. S., H. D. Sherali, and C. M. Shetty (2006). Nonlinear Programming:

Theory and Algorithms. Hoboken, NJ: Wiley-Interscience.

Beckmann, M. J., C. B. McGuire, and C. B. Winston (1956). Studies in the

Economics of Transportation. New Haven, CT: Yale University Press.

Bell, M. G. H. (1995). Alternatives to Dial’s logit assignment algorithm. Trans-

portation Research Part B 29B, 287–295.

Bell, M. G. H. and Y. Iida (1997). Transportation Network Analysis. West

Sussex, England: John Wiley & Sons.

Bellei, G., G. Gentile, and N. Papola (2005). A within-day dynamic traffic

assignment model for urban road networks. Transportation Research Part
B 39, 1–29.

Bellman, R. (1957). Dynamic Programming. Princeton, NJ: Princeton Univer-

sity Press.

Bellman, R. (1958). On a routing problem. Quarterly of Applied Mathemat-

ics 16 (1), 87–90.

Ben-Akiva, M., M. Bierlaire, H. Koutsopoulos, and R. Mishalani (1998). Dyna-

MIT: a simulation-based system for traffic prediction. In DACCORS Short
Term Forecasting Workshop, The Netherlands.

Bertsekas, D. P. (2003). Convex Analysis and Optimization. Cambridge, MA:

Athena Scientific.

Bertsekas, D. P. (2016). Nonlinear Programming. Cambridge, MA: Athena

Scientific.

Bertsekas, D. P., E. M. Gafni, and R. G. Gallager (1984). Second derivative al-

gorithms for minimum delay distributed routing in networks. In IEEE Trans-
actions on Communication, COM-32, pp. 911–919.

Bertsimas, D. and J. N. Tsitsiklis (1997). Introduction to Linear Optimization.

Cambridge, MA: Athena Scientific.

Birge, J. R. and F. Louveaux (1997). Introduction to Stochastic Programming.

New York: Springer Verlag.

Boyce, D., B. Ralevic-Dekic, and H. Bar-Gera (2004). Convergence of traffic

assignments: how much is enough? Journal of Transportation Engineer-
ing 130 (1), 49–55.
BIBLIOGRAPHY 609

Boyles, S. D. (2006). Reliable routing with recourse in stochastic, time-

dependent transportation networks. Master’s thesis, The University of Texas
at Austin.

Boyles, S. D. (2009). Operational, Supply-Side Uncertainty in Transportation

Networks: Causes, Effects, and Mitigation Strategies. Ph. D. thesis, The
University of Texas at Austin.

Boyles, S. D. (2012). Bush-based sensitivity analysis for approximating subnet-

work diversion. Transportation Research Part B 46, 139–155.

Boyles, S. D. (2013). Statewide mesoscopic simulation for Wyoming. Technical

Report FHWA-WY-13/05F, Wyoming Department of Transportation.

Boyles, S. D., C. L. Melson, J. C. Duthie, and T. Rambha (2013). Diverge

models and dynamic traffic equilibria. Presented at the Annual Meeting of the
Institute for Operations Research and the Management Sciences, Minneapolis,
MN.

Boyles, S. D. and T. Rambha (2016). A note on detecting unbounded instances

of the online shortest path problem. Networks 31, 270–276.

Boyles, S. D., T. Rambha, and C. Xie (2015). Equilibrium analysis of low-

conflict network designs. Transportation Research Record 2467, 129–139.

Boyles, S. D. and N. Ruiz Juri (2019). Queue spillback and demand uncertainty
in dynamic network loading. Transportation Research Record 2673, 38–48.

Braess, D. (1969). Über ein Paradoxon aus der Verkehrsplanung. Un-

ternehmensforschung 12, 258–268.

Brouwer, L. E. J. (1910). Über eineindeutige, stetige transformationen von

Flächen in sich. Mathematische Annalen 69, 176–180.

Carey, M. (1992). Nonconvexity of the dynamic assignment problem. Trans-

portation Research Part B 26 (2), 127–133.

Chabini, I. (1999). Discrete dynamic shortest path problems in transportation

applications: complexity and algorithms with optimal run time. Transporta-
tion Research Record 1645, 170–175.

Chen, H.-K. and C. F. Hsueh (1998). A model and an algorithm for the dynamic
user-optimal route choice problem. Transportation Research Part B 32B, 219–
234.

Chiu, Y.-C., J. Bottom, M. Mahut, A. Paz, R. Balakrishna, T. Waller, and

J. Hicks (2010). A Primer for Dynamic Traffic Assignment. Prepared by the
Transportation Network Modeling Committee of the Transportation Research
Board (ADB30).
610 BIBLIOGRAPHY

Cho, H.-J., T. E. Smith, and T. L. Friesz (2000). A reduction method for local
sensitivity analyses of network equilibrium arc flow. Transportation Research
Part B 34 (1), 31–51.

Claudel, C. G. and A. M. Bayen (2010). Lax-Hopf based incorporation of inter-

nal boundary conditions into Hamilton-Jacobi equation. part i: theory. IEEE
Transactions on Automatic Control 55, 1142–1157.

Cooke, L. and E. Halsey (1969). The shortest route through a network with
time-dependent internodal transit times. Journal of Mathematical Analysis
and Applications 14, 492–498.

Correa, J. R., A. S. Schulz, and N. E. Stier-Moses (2004). Selfish routing in

capacitated networks. Mathematics of Operations Research 29, 961–976.

Corthout, R., G. Flötteröd, F. Viti, and C. M. J. Tampère (2012). Non-unique

flows in macroscopic first-order intersection models. Transportation Research
Part B 46, 343–359.

Dafermos, S. C. (1980). Traffic equilibrium and variational inequalities. Trans-

portation Science 14, 42–54.

Dafermos, S. C. (1982). Relaxation algorithms for the general asymmetric traffic

equilibrium problem. Transportation Science 16, 231–240.

Daganzo, C. F. (1982). Unconstrained extremal formulation of some transporta-

tion equilibrium problems. Transportation Science 16, 332–360.

Daganzo, C. F. (1994). The cell transmission model: a dynamic representation

of highway traffic consistent with the hydrodynamic theory. Transportation
Research Part B 28B (4), 269–287.

Daganzo, C. F. (1995a). The cell transmission model, part II: network traffic.
Transportation Research Part B 29B (2), 79–93.

Daganzo, C. F. (1995b). A finite difference approximation of the kinematic wave

model of traffic flow. Transportation Research Part B 29B, 261–276.

Daganzo, C. F. (1998). Queue spillbacks in transportation networks with a route

choice. Transportation Science 32 (1), 3–11.

Daganzo, C. F. (2005a). A variational formulation of kinematic waves: ba-

sic theory and complex boundary conditions. Transportation Research Part
B 39 (2), 187–196.

Daganzo, C. F. (2005b). A variational formulation of kinematic waves: solution

methods. Transportation Research Part B 39 (10), 934–950.

Daganzo, C. F. and Y. Sheffi (1977). On stochastic models of traffic assignment.

Transportation Science 11 (3), 253–274.
BIBLIOGRAPHY 611

Dantzig, G. (1963). Linear Programming and Extensions. Princeton, NJ: Prince-

ton University Press.
Deo, N. and C. Pang (1984). Shortest path algorithms: taxonomy and annota-
tion. Networks 14, 275–323.

Di, X., H. X. Liu, J.-S. Pang, and X. Ban (2013). Boundedly rational user equi-
libria (BRUE): Mathematical formulation and solution sets. Transportation
Research Part B 57, 300–313.
Dial, R. B. (1971). A probabilistic multipath traffic assignment model which
obviates path enumeration. Transportation Research 5, 83–111.

Dial, R. B. (2006). A path-based user-equilibrium traffic assignment algorithm

that obviates path storage and enumeration. Transportation Research Part
B 40 (10), 917–936.
Dijkstra, E. W. (1959). A note on two problems in connexion with graphs.
Numerische Mathematik 1, 269–271.

Dorigo, M. and T. Stützle (2004). Ant Colony Optimization. Cambridge, MA:

MIT Press.
Downs, A. (1962). The law of peak-hour expressway congestion. Traffic Quar-
terly 16, 393–409.

Dreyfus, S. (1969). An appraisal of some shortest-path algorithms. Operations

Research 17, 395–412.
Eiger, A., P. B. Mirchandani, and H. Soroush (1985). Path preferences and
optimal paths in probabilistic networks. Transportation Science 19 (1), 75–
84.

Evans, L. C. (1998). Partial Differential Equations. Providence, RI: Graduate

Studies in Mathematics, American Mathematical Society.
Evans, S. P. (1976). Derivation and analysis of some models for combining trip
distribution and assignment. Transportation Research 10 (1), 37–57.

Facchinei, F. and J. S. Pang (2003). Finite-dimensional variational inequalities

and complementarity problems. Springer Verlag.
Fan, Y., R. Kalaba, and J. Moore (2005). Arriving on time. Journal of Opti-
mization Theory and Applications 127 (3), 497–513.

Fellendorf, M. (1994). VISSIM: A miroscopic simulation tool to evaluate ac-

tuated signal control including bus priority. Presented at the 64th Annual
Meeting of the Institute of Transportation Engineers, Dallas, TX.
Feo, T. A. and M. G. C. Resende (1995). Greedy randomized adaptive search
procedures. Journal of Global Optimization 6, 109–133.
612 BIBLIOGRAPHY

Fisk, C. (1980). Some developments in equilibrium traffic assignment. Trans-

portation Research 14B, 243–255.
Fisk, C. and S. Nguyen (1982). Solution algorithms for network equilibrium
models with asymmetric user costs. Transportation Science 16, 361–381.
Florian, M., I. Constantin, and D. Florian (2009). A new look at projected gradi-
ent method for equilibrium assignment. Transportation Research Record 2090,
10–16.
Florian, M. and D. Hearn (1995). Network equilibrium models and algorithms.
In M. O. Ball, T. L. Magnanti, C. L. Monma, and G. L. Nemhauser (Eds.),
Network Routing, Handbooks in Operations Research and Management Sci-
ence, Volume 8, pp. 485–549. Oxford, UK: Elsevier Science.
Ford, L. R. (1956). Network flow theory. Technical report, Report P-923, Rand
Corporation.
Frank, M. and P. Wolfe (1956). An algorithm for quadratic programming. Naval
Research Logistics Quarterly 3, 95–110.
Friesz, T., D. Bernstein, T. E. Smith, R. L. Tobin, and B. W. Wie (1993). A
variational inequality formulation of the dynamic network user equilibrium
problem. Operations Research 41 (1), 179–191.
Friesz, T., J. Luque, R. Tobin, and B. Wie (1989). Dynamic network traffic
assignment considered as a continuous time optimal control problem. Opera-
tions Research 37, 893–901.
Fu, L. and L. R. Rilett (1998). Expected shortest paths in dynamic and stochas-
tic traffic networks. Transportation Research Part B 32 (7), 499–516.
Fudenberg, D. and J. Tirole (1991). Game Theory. Cambridge, MA: MIT Press.
Gallager, R. G. (1977). A minimum delay routing algorithm using distributed
computation. In IEEE Transactions on Communication, COM-25, pp. 73–85.
Gao, S. (2005). Optimal Adaptive Routing and Traffic Assignment in Stochastic
Time-Dependent Networks. Ph. D. thesis, Massachusetts Institute of Tech-
nology, Cambridge, MA.
Garey, M. R. and D. S. Johnson (1979). Computers and Intractability: A Guide
to the Theory of N P -Completeness. W. H. Freeman and Company.
Gartner, N. H. (1980). Optimal traffic assignment with elastic demands: a
review. Part II: algorithmic approaches. Transportation Science 14 (2), 192–
208.
Gentile, G. (2010). The general link transmission model for dynamic network
loading and a comparison with the DUE algorithm. In L. G. H. Immers,
C. M. J. Tampère, and F. Viti (Eds.), New Developments in Transport Plan-
ning: Advances in Dynamic Traffic Assignment. Edward Elgar Publishing.
BIBLIOGRAPHY 613

Gentile, G. (2014). Local user cost equilibrium: a bush-based algorithm for

traffic assignment. Transportmetrica Part A 10, 15–54.

Gentile, G. (2016). Solving a dynamic user equilibrium model based on splitting

rates with gradient projection algorithm. Transportation Research Part B 92,
120–147.

Glover, F. (1989). Tabu search – part 1. ORSA Journal on Computing 1,

190–206.

Glover, F. (1990). Tabu search – part 2. ORSA Journal on Computing 2, 4–32.

Godunov, S. K. (1959). A difference scheme for numerical solution of discon-

tinuous solution of hydrodynamic equations. Matematicheskii Sbornik 47,
271–306.

Hall, R. W. (1986). The fastest path through a network with random time-
dependent travel times. Transportation Science 20 (3), 182–188.

Hart, P. E., N. J. Nilsson, and B. Raphael (1968). A formal basis for the heuristic
determination of minimum cost paths. IEEE Transactions on Systems Science
and Cybernetics SSC4 4 (2), 100–107.

Hopf, E. (1970). On the right weak solution of the Cauchy problem for a quasi-
linear equation of first order. Indiana University Mathematics Journal 19,
483–487.

Jafari, E. and S. D. Boyles (2016). Improved bush-based methods for network

contraction. Transportation Research Part B 83, 298–313.

Jayakrishnan, R., W. T. Tsai, J. N. Prashker, and S. Rajadhyaksha (1994). A

faster path-based algorithm for traffic assignment. Transportation Research
Record 1443, 75–83.

Jin, W. L. and H. M. Zhang (2003). On the distribution schemes for determining

flows through a merge. Transportation Research Part B 37 (6), 521–540.

Josefsson, M. and M. Patriksson (2007). Sensitivity analysis of separable traf-

fic equilibrium equilibria with application to bilevel optimization in network
design. Transportation Research Part B 41, 4–31.

Kakutani, S. (1941). A generalization of Brouwer’s fixed point theorem. Duke

Mathematical Journal 8 (3), 457–459.

Khani, A. and S. D. Boyles (2015). An exact algorithm for the mean-standard

deviation shortest path problem. Transportation Research Part B 81, 252–
266.

Kirkpatrick, S., C. D. Gelatt, and M. P. Vecchi (1983). Optimization by simu-

lated annealing. Science 220 (4598), 671–680.
614 BIBLIOGRAPHY

Knight, F. H. (1924). Some fallacies in the interpretation of social cost. Quar-

terly Journal of Economics 38, 582–606.

Knuth, D. E. (1997). The Art of Computer Programming, Volume 1: Funda-

mental Algorithms (3rd ed.). Reading, MA: Addison-Wesley.

Larsson, T., J. T. Lundgren, C. Rydergren, and M. Patriksson (2001). Most

likely traffic equilibrium route flows analysis and computation. In F. Gian-
nessi, A. Maugeri, and P. M. Pardalos (Eds.), Equilibrium Problems: Nons-
mooth Optimization and Variational Inequality Models, Nonconvex Optimiza-
tion and its Applications, Volume 58, pp. 129–159. Boston, MA: Springer.

Larsson, T. and M. Patriksson (1992). Simplicial decomposition with disag-

gregated representation for the traffic assignment problem. Transportation
Science 26, 4–17.

Laval, J. A. and L. Leclercq (2013). The Hamilton-Jacobi partial differential

equation and the three representations of traffic flow. Transportation Research
Part B 52, 17–30.

Lawphongpanich, S. and D. W. Hearn (1984). Simplicial decomposition of

the asymmetric traffic assignment problem. Transportation Research Part
B 18 (2), 123–133.

Lax, P. D. (1957). Hyperbolic systems of conservation laws II. In P. Sarnak

and A. Majda (Eds.), Communications on Pure and Applied Mathematics,
pp. 537–566. Wiley Periodicals.

Lebacque, J. P. (1996). The Godunov scheme and what it means for first order
traffic flow models. In Proceedings of the 13th International Symposium on
Transportation and Traffic Theory, London, pp. 647–678.

Lebacque, J. P. and M. M. Khoshyaran (2005). First-order macroscopic traffic

flow models: intersection modeling, network modeling. In Proceedings of the
16th International Symposium on Transportation and Traffic Theory, College
Park, MD, pp. 365–386.

LeVeque, R. (1992). Numerical Methods for Conservation Laws. Basel, Switzer-

land: Birkhüser.

Lighthill, M. and G. Whitham (1955). On kinematic waves. II. A theory of traffic

flow on long crowded roads. Proceedings of the Royal Society of London. Series
A, Mathematical and Physical Sciences 229, 317–345.

Long, J., W. Y. Szeto, H.-J. Huang, and Z. Gao (2012). An intersection-

movement-based dynamic traffic assignment problem. Presented at the 4th
International Symposium on Dynamic Traffic Assignment, Martha’s Vine-
yard, MA.
BIBLIOGRAPHY 615

Lou, Y., Y. Yin, and S. Lawphongpanich (2010). Robust congestion pricing

under boundedly rational user equilibrium. Transportation Research Part
B 44, 15–28.

Loui, R. P. (1983). Optimal paths in graphs with stochastic or multidimensional

weights. Communications of the ACM 26 (9), 670–676.

Lu, S. (2008). Sensitivity of static traffic user equilibria with perturbations in

arc cost function and travel demand. Transportation Science 42 (1), 105–123.

Luke, J. C. (1972). Mathematical models for landform evolution. Journal of

Geophysical Research 77, 2460–2464.

Mahmassani, H. S. (2000). DYNASMART-P, volume ii, user guide. Technical

report, STO67-85-PII, Center for Transportation Research, The University of
Texas at Austin.

Mahmassani, H. S. and G.-L. Chen (1987). On boundedly rational user equilib-

rium in transportation systems. Transportation Science 21, 89–99.

Mahut, M., M. Florian, and N. Tremblay (2003). Space-time queues and dy-
namic traffic assignment: a model, algorithm and applications. Presented at
the 82nd Annual Meeting of the Transportation Research Board, Washington,
DC.

Marcotte, P. and L. Wynter (2004). A new look at the multiclass network

equilibrium problem. Transportation Science 38 (3), 282–292.

Merchant, D. and G. Nemhauser (1978a). A model and an algorithm for the

dynamic traffic assignment problems. Transportation Science 12, 183–199.

Merchant, D. and G. Nemhauser (1978b). Optimality conditions for a dynamic

traffic assignment model. Transportation Science 12, 200–207.

Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller

(1953). Equation of state calculations by fast computing machines. Journal
of Chemical Physics 21, 1087.

Miller-Hooks, E. D. (2001). Adaptive least-expected time paths in stochastic,

time-varying transportation and data networks. Networks 37 (1), 35–52.

Miller-Hooks, E. D. and H. S. Mahmassani (2000). Least expected time paths

in stochastic, time-varying transportation networks. Transportation Sci-
ence 34 (2), 198–215.

Mitradjieva, M. and P. O. Lindberg (2013). The stiff is moving — conjugate

direction Frank-Wolfe methods with application to traffic assignment. Trans-
portation Science 47 (2), 280–293.
616 BIBLIOGRAPHY

Montemanni, R. and L. Gambardella (2004). An exact algorithm for the ro-

bust shortest path problem with interval data. Computers and Operations
Research 31 (10), 1667–1680.

Murchland, J. D. (1970). Braess’s paradox of traffic flow. Transportation Re-

search 4, 391–394.

Murthy, I. and S. Sarkar (1996). A relaxation-based pruning technique for a class

of stochastic shortest path problems. Transportation Science 30 (3), 220–236.

Netter, M. (1972). Affectations de trafic et tarification au coût marginal social:

critique de quelques idées admises. Transportation Research 6, 411–429.

Newell, G. F. (1993a). A simplified theory of kinematic waves in highway traffic,

part I: General theory. Transportation Research Part B 27 (4), 281–287.

Newell, G. F. (1993b). A simplified theory of kinematic waves in highway traf-

fic, part II: Queueing at freeway bottlenecks. Transportation Research Part
B 27 (4), 289–303.

Newell, G. F. (1993c). A simplified theory of kinematic waves in highway traffic,

part III: Multi-destination flows. Transportation Research Part B 27 (4), 305–
313.

Nezamuddin and S. D. Boyles (2015). A continuous DUE algorithm using the

link transmission model. Networks and Spatial Economics 15, 465–483.

Nie, Y. and X. Wu (2009). Shortest path problem considering on-time arrival

probability. Transportation Research Part B 43 (6), 597–613.

Nie, Y. M. (2010a). A class of bush-based algorithms for the traffic assignment

problem. Transportation Research Part B 44, 73–89.

Nie, Y. M. (2010b). Equilibrium analysis of macroscopic traffic oscillations.

Transportation Research Part B 44, 62–72.

Nie, Y. M., J. Ma, and H. M. Zhang (2008). A polymorphic dynamic network

loading model. Computer-Aided Civil and Infrastructure Engineering 23 (2),
86–103.

Osorio, C. and M. Bierlaire (2009). An analytic finite capacity queueing net-

work model capturing the propagation of congestion and blocking. European
Journal of Operational Research 196, 996–1007.

Osorio, C., G. Flötteröd, and M. Bierlaire (2011). Dynamic network loading:

a stochastic differentiable model that derives link state distributions. Trans-
portation Research Part B 45, 1410–1423.

Pape, U. (1974). Implementation and efficiency of Moore-algorithms for the

shortest route problem. Mathematical Programming 7, 212–222.
BIBLIOGRAPHY 617

Patriksson, M. (1994). The Traffic Assignment Problem — Models and Methods.

Utrecht, Netherlands: VSP.
Patriksson, M. (2004). Sensitivity analysis of traffic equilibria. Transportation
Science 38, 258–281.

Peeta, S. and A. K. Ziliaskopoulos (2001). Foundations of dynamic traffic as-

signment: the past, the present, and the future. Networks and Spatial Eco-
nomics 1, 233–265.
Pham, D. T., A. Ghanbarzadeh, E. Koc, S. Otri, S. Rahim, and M. Zaidi
(2005). The bees algorithm. Technical report, Manufacturing Engineeirng
Centre, Cardiff University.
Pigou, A. C. (1920). The Economics of Welfare. London: Macmillan and Co.
Polychronopoulos, G. H. and J. N. Tsitsiklis (1996). Stochastic shortest path
problems with recourse. Networks 27 (2), 133–143.

Powell, W. B. and Y. Sheffi (1982). The convergence of equilibrium algorithms

with predetermined step sizes. Transportation Science 16, 45–55.
Provan, J. S. (2003). A polynomial-time algorithm to find shortest paths with
recourse. Networks 41 (2), 115–125.
Psaraftis, H. N. and J. N. Tsitsiklis (1993). Dynamic shortest paths in acyclic
networks with Markovian arc costs. Operations Research 41 (1), 91–101.
Ran, B., D. E. Boyce, and L. J. LeBlanc (1993). A new class of instantaneous
dynamic user-optimal traffic assignment models. Operations Research 41 (1),
192–202.

Rechenberg, I. (1973). Evolutionsstrategie. Stuttgart, Germany: Holzmann-

Froboog.
Richards, P. (1956). Shock waves on the highway. Operations Research 4 (1),
42–51.
Rockafellar, R. T. (1997). Convex Analysis. Princeton, NJ: Princeton University
Press.
Rossi, T. F., S. McNeil, and C. Hendrickson (1989). Entropy model for consis-
tent impact fee assessment. Journal of Urban Planning and Development 115,
51–63.

Roughgarden, T. A. (2002). Selfish Routing. Ph. D. thesis, Cornell University.

Schwefel, H.-P. (1977). Numerische Optimierung von Computor-Modellen mit-
tels der Evolutionsstrategie: mit einer vergleichenden Einführung in die Hill-
Climbing- und Zufallsstrategie. Basel, Switzerland; Stuttgart, Germany:
Birkhäuser.
618 BIBLIOGRAPHY

Sen, S., R. Pillai, S. Joshi, and A. K. Rathi (2001). A mean-variance model

for route guidance in advanced traveler information systems. Transportation
Science 35 (1), 37–49.
Shahabi, M., A. Unnikrishnan, and S. D. Boyles (2013). An outer approximation
algorithm for the robust shortest path problem. Transportation Research Part
E 58, 52–66.
Shahabi, M., A. Unnikrishnan, and S. D. Boyles (2015). Robust optimization
strategy for the shortest path problem under uncertain link travel cost distri-
bution. Computer-Aided Civil and Infrastructure Engineering 30, 433–448.

Sheffi, Y. (1985). Urban Transportation Networks. Englewood Cliffs, NJ:

Prentice-Hall.
Sipser, M. (2012). Introduction to the Theory of Computation (3rd ed.). Cengage
Learning.
Sivakumar, R. A. and R. Batta (1994). The variance-constrained shortest path
problem. Transportation Science 28 (4), 309–316.
Smith, M. J. (1979a). The existence, uniqueness, and stability of traffic equilib-
ria. Transportation Research 13B, 295–304.
Smith, M. J. (1979b). Traffic control and route-choice; a simple example. Trans-
portation Research 13B, 289–294.
Smith, M. J. (1983). The existence and calculation of traffic equilibria. Trans-
portation Research Part B 17B (4), 291–303.
Tampère, C. M. J., R. Corthout, D. Cattrysse, and L. H. Immers (2011). A
generic class of first order node models for dynamic macroscopic simulation
of traffic flows. Transportation Research Part B 45, 289–309.
Tobin, R. L. and T. L. Friesz (1988). Sensitivity analysis for equilibrium network
flow. Transportation Science 22 (4), 242–250.
Vickrey, W. S. (1969). Congestion theory and transport investment. American
Economic Review 59 (2), 251–261.
Waller, S. T. (2006). Personal communication.
Waller, S. T. and A. K. Ziliaskopoulos (2002). On the online shortest path
problem with limited arc cost dependencies. Networks 40 (4), 216–227.

Wardrop, J. (1952). Some theoretical aspects of road traffic research. Proceedings

of the Institute of Civil Engineers, Part II 1, 325–378.
Wie, B. W., R. L. Tobin, T. L. Friesz, and D. Bernstein (1995). A discrete
time, nested cost operator approach to the dynamic network user equilibrium
problem. Transportation Science 29, 79–92.
BIBLIOGRAPHY 619

Wolsey, L. A. (1998). Integer Programming. Hoboken, NJ: Wiley-Interscience.

Xie, J., Y. Nie, and X. Yang (2013). Quadratic approximation and convergence
of some bush-based algorithms for the traffic assignment problem. Trans-
portation Research Part B 56, 15–30.

Xing, T. and X. Zhou (2011). Finding the most reliable path with and with-
out link travel time correlation: a Lagrangian substitution based approach.
Transportation Research Part B 59, 22–44.
Yang, H. and M. G. H. Bell (2007). Sensitivity analysis of network traffic equi-
libria revisited: the corrected approach. In B. Heydecker (Ed.), Mathematics
in Transport, pp. 373–411. Elsevier.
Yperman, I. (2007). The Link Transmission Model for Dynamic Newtork Load-
ing. Ph. D. thesis, Katholieke Universiteit Leuven, Belgium.
Yu, G. and J. Yang (1998). On the robust shortest path problem. Computers
and Operations Research 25 (6), 457–468.

Zhang, H. M. and Y. Nie (2006). A hybrid traffic simulation model. Presented at

the 85th Annual Meeting of the Transportation Research Board, Washington,
DC.
Zhang, Y., Z.-J. M. Shen, and S. Song (2016). Parametric search for the bi-
attribute concave shortest path problem. Transportation Research Part B 94,
150–168.
Ziegelmann, M. (2001). Constrained Shortest Paths and Related Problems. Ph.
D. thesis, Universität des Saarlandes.
Ziliaskopoulos, A. K. (2000). A linear programming model for the single desti-
nation system optimum dynamic traffic assignment problem. Transportation
Science 34 (1), 37–49.
Ziliaskopoulos, A. K. and H. S. Mahmassani (1994). Time-dependent, shortest-
path algorithm for real-time intelligent vehicle highway system applications.
Transportation Research Record 1408, 94–104.

Ziliaskopoulos, A. K. and S. T. Waller (2000). An Internet-based geographic

information system that integrates data, models and users for transportation
applications. Transportation Research Part C 8, 427–444.