Problem Formulation Guide

WHITEPAPER

2020-10-27

Overview CONTACT

Corporate Headquarters
This guide steps users through the process of developing a model suit-
3033 Beta Ave
able for the D-Wave quantum processing unit (QPU) or hybrid solvers. Burnaby, BC V5G 4M9
By reading through this guide and working through the provided exer- Canada
cises, a new user with little to no prior experience with D-Wave tech- Tel. 604-630-1428
nology will able to model small problems using the binary quadratic US Office
model (BQM). 2650 E Bayshore Rd
Palo Alto, CA 94303

Email: [email protected]

www.dwavesys.com

14-1044A-B
D-Wave Whitepaper Series
 Notice and Disclaimer
D-Wave Systems Inc. (“D-Wave”) reserves its intellectual property rights in and to this doc-
ument, any documents referenced herein, and its proprietary technology, including copyright,
trademark rights, industrial design rights, and patent rights. D-Wave trademarks used herein
include D-WAVE®, Leap™ quantum cloud service, Ocean™, Advantage™ quantum system,
D-Wave 2000Q™, D-Wave 2X™, and the D-Wave logo (the “D-Wave Marks”). Other marks used in
this document are the property of their respective owners. D-Wave does not grant any license, assign-
ment, or other grant of interest in or to the copyright of this document or any referenced documents,
the D-Wave Marks, any other marks used in this document, or any other intellectual property rights
used or referred to herein, except as D-Wave may expressly provide in a written agreement.

Copyright © D-Wave Systems Inc. Problem Formulation Guide i

 Contents
1 Introduction 1
1.1 Example 1: Traveling Salesman Problem . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Example 2: Antenna Placement Problem . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 Definitions 2
2.1 Binary Quadratic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2 Quadratic Unconstrained Binary Optimization . . . . . . . . . . . . . . . . . . . 3
2.2.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2.2 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 Ising Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3.1 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3.2 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Equivalence of QUBO and Ising . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4.1 QUBO to Ising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4.2 Ising to QUBO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Summary of Key Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Small Problems with Two Variables 6

3.1 General Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2.1 Step 1: Constraint Satisfaction Table . . . . . . . . . . . . . . . . . . . . . 6
3.2.2 Step 2: Make a System of Equations . . . . . . . . . . . . . . . . . . . . . 7
3.2.3 Step 3: Solve the System of Equations . . . . . . . . . . . . . . . . . . . . 8
3.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Problems with Multiple Constraints 9

4.1 Social Networks (Friends and Enemies) . . . . . . . . . . . . . . . . . . . . . . . 9
4.1.1 Problem Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.1.2 Combining Relationships . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
4.2 BQM Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2.1 Friend Relationship . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2.2 Enemy Relationship . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2.3 Combining All Relationships . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.3 BQM Development Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.3.1 Step 1: Write Objective and Constraints . . . . . . . . . . . . . . . . . . . 14
4.3.2 Step 2: Convert Objective and Constraints into Binary Math Expressions 15
4.3.3 Step 3: Transform Math Expressions into a BQM . . . . . . . . . . . . . . 15
4.3.4 Step 4: Combine Expressions . . . . . . . . . . . . . . . . . . . . . . . . 16
4.4 Example: Choosing Boxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.4.1 Step 1: Write Objective and Constraints . . . . . . . . . . . . . . . . . . . 17
4.4.2 Step 2: Convert Objective and Constraints into Binary Math Expressions 17
4.4.3 Step 3: Transform Math Expressions into a BQM . . . . . . . . . . . . . . 18
4.4.4 Step 4: Combine Expressions . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4.5 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Copyright © D-Wave Systems Inc. Problem Formulation Guide ii

 5 Summary of Exercises with Solutions 19
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.1.1 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2.1 QUBO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2.2 Ising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2.3 Converting Between QUBO and Ising . . . . . . . . . . . . . . . . . . . . 21
5.3 Small Problems with Two Variables . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3.1 QUBO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3.2 Ising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4 Problems with Multiple Constraints . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4.1 QUBO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.4.2 Ising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

References 22

Copyright © D-Wave Systems Inc. Problem Formulation Guide iii

1 Introduction
When working with the D-Wave systems and software, we need to formulate our problem
as a binary quadratic model, or BQM. We can formulate many classes of problems as BQMs,
such as optimization problems. We can view an optimization problem in terms of objectives
and constraints. An objective is something that we want to minimize or maximize, while
constraints are rules about which solutions are acceptable. As we explore how to formulate
our problems as binary quadratic models, we will use this terminology frequently.
Below are two examples of optimization problems that can be formulated as BQMs and
run using the D-Wave QPU or our hybrid solvers. These examples provided demonstrate
how real world problems can be broken down into objectives and constraints. While these
examples are more complex than those that we will develop in this guide, the same strategy
and techniques can help to build a BQM for each of them. For the interested reader we
provide links to where more information on the BQM formulation for each example can be
found.

1.1 Example 1: Traveling Salesman Problem

A well-known problem in computer science is the traveling salesman problem (TSP). In
this problem, we have a list of cities that must be visited (exactly once) on a sales person’s
route. The sales person would like to take the shortest route possible. We can phrase this
problem in terms of objectives and constraints.
Objective:
Minimize the total distance for the route.
Constraint:
Visit each city exactly once.
For more information on how to formulate this problem as a BQM, check out Andrew
Lucas’s paper on Ising formulations [1].

1.2 Example 2: Antenna Placement Problem

We have a set of 10 antennas that we can place anywhere in a list of 100 locations. We want
to cover as much geographic area as possible. However, some locations interfere with each
other. Where should we build our antennas?
Objective:
Maximize coverage.
Constraints:

• Build at most 10 antennas.

• Choose locations that do not cause interference.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 1

 For more information on how to formulate this problem as a BQM, take a look at the an-
tenna selection example in the Collection of Code Examples [2].

1.3 Exercise
How can we split the following problem up into an objective and constraints?
A business is trying to figure out the best way to pack a box full of items. There are five
items, with values 1, 5, 3, 4, and 2 dollars. These items have weights 2, 4, 4, 1, and 3 lbs,
respectively. The contents of the box must weigh exactly 6 lbs. What is the best way to pack
the box so that the box has the highest total value?

1.4 Resources
Many other resources are available for new users getting with D-Wave. Below is a selection
of some that may be especially useful.

Ocean SDK Documentation: https://docs.ocean.dwavesys.com/

Collection of Code Examples: https://github.com/dwave-examples

D-Wave on YouTube: https://www.youtube.com/dwavesystems

2 Definitions
2.1 Binary Quadratic Model
The binary quadratic model (BQM) class of problems consists of Ising models and quadratic
unconstrained binary optimization (QUBO) models. These two forms are mathematically
equivalent, and our software tools allow you to work in either domain. In this guide, we
will provide examples and exercises in both forms. Note that some problems may see better
results with Ising over QUBO, and vice versa.
A BQM takes the following form.
!
min ∑ ai vi + ∑ ∑ bi,j vi v j + c
i i j >i

The coefficients ai and bi,j are constant numbers we choose to define our problem, as is the
constant term c. The binary variables vi and v j are the values that we are looking for to solve
our problem. The best solution for these variables is the value for each vi that produces
the smallest value for the overall expression. Searching for the variables that minimize an
expression is called an “argmin” in mathematics.
Let’s break this expression down.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 2

 Linear Terms. The first summation, ∑ ai vi , contains linear terms, with each having just
one binary variable.

Quadratic Terms. The second summation, ∑ ∑ bi,j vi v j , contains quadratic terms, with
each term in the summation containing a product of two variables.

Constants. In the general BQM form we may or may not include constants. Since we are
looking for an argmin, any constant terms will not affect our final answer. However, it may
be useful when interpreting the output from D-Wave solvers and samplers.

2.2 Quadratic Unconstrained Binary Optimization

A QUBO model takes the following form.
!
min ∑ ai xi + ∑ ∑ bi,j xi x j + c
i i j >i

The binary variables xi and x j may take values from {0, 1}.
When we are ready to run a QUBO with the D-Wave Ocean software tool kit, one method
is to store the QUBO in a matrix format for the Ocean SDK. We can map a QUBO to a
matrix by placing linear terms along the diagonal and quadratic terms in the off-diagonal
positions.

2.2.1 Example
We can map x1 + 2x2 + 3x3 + 12x1 x2 + 13x1 x3 + 14x2 x4 to the following matrix.

 
1 12 13 0
 0 2 0 14 
 
 0 0 3 0 
0 0 0 0

In this matrix, we place coefficient ai in matrix position (i, i ) and coefficient bi,j in matrix
position (i, j).

2.2.2 Exercise
Put the following QUBO into matrix form.

x1 + x1 x2 − 3x3 + 5

Copyright © D-Wave Systems Inc. Problem Formulation Guide 3

2.3 Ising Model
An Ising model takes the following form.
!
min ∑ hi si + ∑ ∑ Ji,j si s j
i i j >i

The binary variables si and s j may take values from {−1, +1}.
When it comes time to write Ocean programs, we will need to represent our Ising model
as a matrix and a vector. We can map the quadratic terms of an Ising model to a matrix
by placing them in the off-diagonal positions, and the linear terms to a vector as a list of
coefficients as shown below.

2.3.1 Example
We can map s1 + 2s2 + 3s3 + 12s1 s2 + 13s1 s3 + 14s2 s4 to the following list and matrix.

Linear.  
h= 1 2 3 0

Quadratic.  
0 12 13 0
 0 0 0 14 
J=
 0

0 0 0 
0 0 0 0

In this matrix, coefficient Ji,j in matrix position (i, j).

2.3.2 Exercise
Put the following Ising model into matrix form.

s1 + s1 s2 − 3s3 + 5

2.4 Equivalence of QUBO and Ising

The Ising and QUBO models are mathematically equivalent — we can translate from one
model to the other using a simple substitution.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 4

2.4.1 QUBO to Ising
To translate a QUBO model to an Ising model, we use the following substitution.

si + 1
xi 7 →
2

Example.
f (x) = −22x1 − 6x2 − 14x  3 +20x1 x2+ 28x1 x3 + 9
s1 +1
= −22 2 − 6 s2 2+1 − 14 s3 2+1
     
+20 s1 2+1 s2 +1
2 + 28 s1 2+1 s3 +1
2 +9
= −11s1 − 11 − 3s2 − 3 − 7s3 − 7 + 5s1 s2
+5s1 + 5s2 + 5 + 7s1 s3 + 7s1 + 7s3 + 7 + 9
= s1 + 2s2 + 5s1 s2 + 7s1 s3

2.4.2 Ising to QUBO

To translate an Ising model to a QUBO model, we use the following substitution.

si 7→ 2xi − 1

Example.

g(s) = s1 + 2s2 + 5s1 s2 + 7s1 s3

= (2x1 − 1) + 2(2x2 − 1) + 5(2x1 − 1)(2x2 − 1) + 7(2x1 − 1)(2x3 − 1)
= 2x1 − 1 + 4x2 − 2 + 20x1 x2 − 10x1 − 10x2 + 5 + 28x1 x3 − 14x1 − 14x3 + 7
= −22x1 − 6x2 − 14x3 + 20x1 x2 + 28x1 x3 + 9

2.4.3 Exercise
Convert the following QUBO model to an Ising model.

x1 + x1 x2 − 3x3 + 5

Convert the following Ising model to a QUBO model.

s1 + s1 s2 − 3s3 + 5

2.5 Summary of Key Points

• A BQM can contain constant, linear, and quadratic terms, and no higher-order terms.

• We define the coefficients and constant in the expression.

• The solution is the value of each binary variable that minimizes the expression.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 5

3 Small Problems with Two Variables
3.1 General Process
The following steps can be used to create a two-variable BQM.

1. Make a constraint satisfaction table.

2. Create a system of equations.

3. Solve the system of equations.

The constraint satisfaction table is similar to a truth table - it covers all possible values for
our two variables and considers whether or not the constraint that we are modeling is true
or false for each choice of values.

3.2 Example
We start with a simple problem. We have two variables, say v1 and v2 , and we want a BQM
that represents the constraint v1 = v2 . For QUBO and Ising we follow the same process.

3.2.1 Step 1: Constraint Satisfaction Table

Set up a table for the two variables that covers all possible values they might take. Add
a column in the table to indicate whether or not the constraint is satisfied (true) for each
combination of variable values.

QUBO Ising
x1 x2 s1 s2 Equal

0 0 -1 -1 Yes
0 1 -1 +1 No
1 0 +1 -1 No
1 1 +1 +1 Yes

Next, replace the final column of yes/no values with small/large values, respectively.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 6

 QUBO Ising
x1 x2 s1 s2 Value

0 0 -1 -1 0
0 1 -1 +1 1
1 0 +1 -1 1
1 1 +1 +1 0

In this step, we choose a small value when the constraint is satisfied because the QPU
is designed to minimize a BQM. By choosing a small value, we line up the ground state
(lowest energy the QPU can find) with the constraint being satisfied. Note that any choice
of small and large value can be used.

3.2.2 Step 2: Make a System of Equations

Recall from the last section the general form for QUBO and Ising.

QUBO: a1 x1 + a2 x2 + b1,2 x1 x2 + c

Ising: h1 s1 + h2 s2 + J1,2 s1 s2 + c

Using each row in the constraint satisfaction table, we can write an equation. Each pair of
variable values, the general BQM model, and the final column value together determine
an equation in which the BQM coefficients are the only unknowns. This results in four
different equations (one for each row in the table).

QUBO

a1 · 0 + a2 · 0 + b1,2 · 0 · 0 + c = 0
a1 · 0 + a2 · 1 + b1,2 · 0 · 1 + c = 1
a1 · 1 + a2 · 0 + b1,2 · 1 · 0 + c = 1
a1 · 1 + a2 · 1 + b1,2 · 1 · 1 + c = 0

Ising

h1 · (−1) + h2 · (−1) + J1,2 · (−1) · (−1) + c = 0

h1 · (−1) + h2 · (+1) + J1,2 · (−1) · (+1) + c = 1
h1 · (+1) + h2 · (−1) + J1,2 · (+1) · (−1) + c = 1
h1 · (+1) + h2 · (+1) + J1,2 · (+1) · (+1) + c = 0

Copyright © D-Wave Systems Inc. Problem Formulation Guide 7

3.2.3 Step 3: Solve the System of Equations
Solve the system of equations for your coefficients.

QUBO Ising

a1 = 1 h1 = 0
a2 = 1 h2 = 0
b1,2 = −2 J1,2 = −0.5
c=0 c = 0.5

To help with solving these systems of equations there are many resources online that
are available. Some of these are WolframAlpha (wolframalpha.com) or Matrix Calculator
(matrixcalc.org).
Our final models are:

QUBO: x1 + x2 − 2x1 x2

Ising: −0.5s1 s2 + 0.5

3.3 Exercise
What do the QUBO and Ising models look like for the constraint v1 ≥ v2 ?

Copyright © D-Wave Systems Inc. Problem Formulation Guide 8

4 Problems with Multiple Constraints
When we work with problems in the real world, we often have an objective and multiple
constraints playing a role simultaneously. How do we incorporate all of these components
into a single BQM?

4.1 Social Networks (Friends and Enemies)

We work through the ”Social Networks” problem, sometimes referred to as ”Friends and
Enemies”. For more examples that discuss this problem and some applications take a
look at our demo (https://cloud.dwavesys.com/leap/resources#demos) and Jupyter
Notebook (https://cloud.dwavesys.com/leap/learning/) available on the Leap plat-
form (https://cloud.dwavesys.com/leap/).

4.1.1 Problem Setup

In a group of people, we have some existing relationships. Some pairs of people are friends,
others are enemies, and some have no relationship (maybe they’ve never met each other
before).
For example, let’s say we have the following friend and enemy relationships.
Friends:

• Katie and Anthony

• Carlos and Betty

Enemies:

• Anthony and Carlos

Suppose that we are asked to split this group of people into two sets. Maybe they’re being
divided into two teams, or being split into groups for a project. Our goal is to split the
people into two groups so that we have friends in the same group and enemies in different
groups. How can we map this to a BQM?
The first step is to think about how we can use binary variables to represent the answer to
our problem. When we consider our problem, the answer that we’re looking for is which
group each person should be assigned to. If we call these groups “A” and “B”, then an
answer should look something like:

• Katie in Group A

• Anthony in Group A

• Carlos in Group B

• Betty in Group B

Copyright © D-Wave Systems Inc. Problem Formulation Guide 9

 To do this we can assign a binary variable to each person that tells us which group they
belong to. If we’re working with a QUBO model we could say xK = 0 means Katie is in
Group A, while xK = 1 means Katie is in Group B. Similarly, if we’re working with an
Ising model we can say sK = −1 means Katie is in Group A, while sK = 1 means Katie is
in Group B.

Group A Group B

QUBO 0 1
Ising -1 +1

Now that we have figured out how to align our answers with binary valued variables,
we need to figure out how to write our minimization function. A good way to tackle this
problem is to think about individual relationships instead of the entire set of people as a
whole. For two friends with a given relationship, which team assignments are good and
which are bad?

Katie Anthony Good/Bad?

A A Good
A B Bad
B A Bad
B B Good

This table should look familiar now — it’s similar to what we did for our two variable rela-
tionships in the beginning of this guide! In fact, this is exactly the same as the QUBO/Ising
truth tables we used to create a BQM that represents the constraint v1 = v2 .

4.1.2 Combining Relationships

At this point we are able to write a BQM for each relationship. How can we combine them
into one BQM? The answer is surprisingly simple: addition.
Once we have a BQM for every relationship, we can add all of the BQMs together to make
a single, comprehensive BQM for the entire population.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 10

4.2 BQM Formulation
4.2.1 Friend Relationship
For our friend relationships, we saw in a previous section (Small Problems with Two Vari-
ables) that we can represent this relationship (x1 = x2 ) by the following BQMs.

QUBO: x1 + x2 − 2x1 x2

Ising: −0.5s1 s2 + 0.5

This gives us the following BQMs for our friend relationships.

QUBO Ising

Katie and Anthony x A + xK − 2x A xK −0.5s A sK + 0.5

Carlos and Betty x B + xC − 2x B xC −0.5s B sC + 0.5

4.2.2 Enemy Relationship

For our enemy relationship, we need to figure out what BQM models the relationship. We
do this in the same way we did our two-variable BQMs in the earlier section.
Constraint Satisfaction Table:

QUBO Ising
x1 x2 s1 s2 Different?

0 0 -1 -1 No
0 1 -1 +1 Yes
1 0 +1 -1 Yes
1 1 +1 +1 No

Copyright © D-Wave Systems Inc. Problem Formulation Guide 11

 Replacing the final column of yes/no values with small/large values.

QUBO Ising
x1 x2 s1 s2 Different?

0 0 -1 -1 1
0 1 -1 +1 0
1 0 +1 -1 0
1 1 +1 +1 1

Building our system of equations.

QUBO

a1 · 0 + a2 · 0 + b1,2 · 0 · 0 + c = 1
a1 · 0 + a2 · 1 + b1,2 · 0 · 1 + c = 0
a1 · 1 + a2 · 0 + b1,2 · 1 · 0 + c = 0
a1 · 1 + a2 · 1 + b1,2 · 1 · 1 + c = 1

Ising

h1 · (−1) + h2 · (−1) + J1,2 · (−1) · (−1) + c = 1

h1 · (−1) + h2 · (+1) + J1,2 · (−1) · (+1) + c = 0
h1 · (+1) + h2 · (−1) + J1,2 · (+1) · (−1) + c = 0
h1 · (+1) + h2 · (+1) + J1,2 · (+1) · (+1) + c = 1

Solving our system of equations.

QUBO Ising

a1 = −1 h1 = 0
a2 = −1 h2 = 0
b1,2 = 2 J1,2 = 0.5
c=1 c = 0.5

Copyright © D-Wave Systems Inc. Problem Formulation Guide 12

 Our final models are:

QUBO: − x1 − x2 + 2x1 x2 + 1

Ising: 0.5s1 s2 + 0.5

For the pair of enemies in our set of people, we now have:

QUBO Ising

Anthony and Carlos − x A − xC + 2x A xC + 1 0.5s A sC + 0.5

4.2.3 Combining All Relationships

Now that we have a BQM for each relationship, we add them together to get our BQM for
the entire problem.

QUBO:
( x A + xK − 2x A xK ) + ( x B + xC − 2x B xC ) + (− x A − xC + 2x A xC + 1)

This simplifies down to the following.

x B + xK + 2x A xC − 2x A xK − 2x B xC + 1

Ising:
(−0.5s A sK + 0.5) + (−0.5s B sC + 0.5) + (0.5s A sC + 0.5)

This simplifies down to the following.

−0.5s A sK − 0.5s B sC + 0.5s A sC + 1.5

To double-check our work, we can evaluate our BQM values for all possible assignments
for these four people. The rows with the smallest value tell us the best assignment of people
to teams.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 13

 Anthony Betty Carlos Katie Value

A A A A 1
A A A B 2
A A B A 1
A A B B 2
A B A A 2
A B A B 3
A B B A 0
A B B B 1
B A A A 1
B A A B 0
B A B A 3
B A B B 2
B B A A 2
B B A B 1
B B B A 2
B B B B 1

4.3 BQM Development Process

To successfully map problems to BQMs, we work through a series of steps.

4.3.1 Step 1: Write Objective and Constraints

This should be written out in terms of your problem domain (i.e. not using math!).
The objective of a problem is what we are looking to minimize or maximize. Look for those
key words in your problem to determine if your problem has one, or if you are just looking
to satisfy a set of constraints.
A constraint in a problem is a rule that we have to follow. An answer that doesn’t satisfy a
given constraint is called infeasible —- it’s not a good answer, and generally we won’t be
able to use it.
For example, the previous problem might be described with the following:

Objective: None

Constraint 1: Friends on the same team

Constraint 2: Enemies on different teams

Note that there may be more than one way to interpret your problem as objectives and
constraints.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 14

4.3.2 Step 2: Convert Objective and Constraints into Binary Math
Expressions
The first step is to think about binary variables. What answer are you seeking to your
problem? What “Yes” or “No” questions can you ask that will give you that answer? Once
we think of our problem in these terms, we can assign a binary variable for each question.
Next, we transform our objective and constraints into math expressions using these binary
variables. Often this can be done with our truth table approach if can find a way to break
our problem down into two- or three-variable relationships.

4.3.3 Step 3: Transform Math Expressions into a BQM

Different types of expressions will require different strategies. If you used the truth table
approach in step 2, you may not need to make any adjustments!
Some common transformations are the following.

Squared Terms. Our QUBO and Ising models don’t have room for squared binary vari-
ables. It turns out that we don’t actually need them.

QUBO: Our binary variables in a QUBO are 0 and 1. Note that 02 = 0 and 12 = 1, so we
can replace any term ‘xi2 with xi .
Ising: Our binary variables in an Ising model are -1 and +1. Note that (−1)2 = 1 and
(+1)2 = 1, so we can replace any term s2i with the constant 1.

Maximization to Minimization. If our objective function is a maximization function (for

example, you might be maximizing profit), we can convert this to a minimization by mul-
tiplying the entire expression by -1.
Example:
arg max(3v1 + 2v1 v2 ) = arg min(−3v1 − 2v1 v2 )

Equality to Minimization. If you have a constraint that is an equality, we can convert it to a

minimization expression by moving all arguments and constants to one side of the equality
and squaring the non-zero side of the equation. This will leave us with an expression that
is satisfied at its smallest value (0) and unsatisfied at any larger value (¿0).
Example:
Original constraint:
v1 + v2 = 1

Move all arguments and constants to one side:

v1 + v2 − 1 = 0

Square the expression:

( v1 + v2 − 1)2

Copyright © D-Wave Systems Inc. Problem Formulation Guide 15

 Note that the smallest value a squared expression can equal is zero, so the smallest value
(v1 + v2 − 1)2 can equal is zero. This expression will equal zero when v1 + v2 − 1 = 0,
which is exactly the constraint that we are looking to satisfy.

4.3.4 Step 4: Combine Expressions

Once we have written all of the components (objective and constraints) as BQM expres-
sions, we make our final BQM by adding all of the components together.
In this step, we have the choice to weight the components if we choose. This weighting is
done using a Lagrange parameter. Generally, we will assign individual Lagrange parame-
ters to each constraint to prioritize valid solutions (ones in which all of our constraints are
satisfied).
BQM = min(objective + γ(constraint))

Example:
Suppose that we have two companies that can complete the same job, each for a different
price. Company 1 can complete the job for $12, while Company 2 can complete the job for
$8. We want to have the job completed for the smallest cost.
We can model this problem as a QUBO. The objective is to minimize cost and the constraint
is that we must choose exactly one company. The binary variables x1 and x2 can be used to
indicate whether we will hire Company 1 or Company 2 to complete the job. Using these
binary variables, we write the objective as 12x1 + 8x2 and the constraint as ( x1 + x2 − 1)2 .
These can be combined to make the following QUBO.
 
min (12x1 + 8x2 ) + γ( x1 + x2 − 1)2

If we choose γ = 1, we see the following values for our QUBO.

x1 x2 Value

0 0 1
0 1 8
1 0 12
1 1 21

Note that the smallest value of 1 corresponds to not choosing either company to complete
the job! Clearly this is not the optimal solution to our original problem, so we need to
adjust the Lagrange parameter to ensure that our constraint (choose exactly one company)
is satisfied.
Setting a larger Lagrange parameter will enforce our constraint. The following table demon-
strates the QUBO values for γ = 20.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 16

 x1 x2 Value

0 0 20
0 1 8
1 0 12
1 1 40

Now we see that the smallest value corresponds to the optimal solution of hiring Company
2 to complete the job.
Note that you may need to try a few different values to identify the best Lagrange parame-
ter value for your specific BQM. A good starting value is to set γ equal to your best estimate
of the value of your objective function. If you find that your constraints are not satisfied in
the solutions returned, you may need to increase γ. On the other hand, if your constraints
are all satisfied but your solutions are not close to optimal, you may need to decrease γ.

4.4 Example: Choosing Boxes

We’re given three boxes with different weights. We want to choose the two boxes with the
smallest sum.

Box 1 Box 2 Box 3

Weight 15 20 25

This is a simple problem – we know that the answer is to choose boxes 15 and 20. But how
do we phrase this problem as a BQM?

4.4.1 Step 1: Write Objective and Constraints

Objective: We are looking for the smallest sum, so our objective is “minimize the sum of
the boxes chosen”.

Constraint: We are allowed to choose two boxes, so our constraint is “choose exactly two
boxes”.

4.4.2 Step 2: Convert Objective and Constraints into Binary Math

Expressions
Binary Variables. First, we need to define our binary variables. The answer that we are
looking for is which boxes we should choose. For each box, we can ask “do we choose this
box?”. This points us to how we should define our binary variables.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 17

 Use Box i Don’t use Box i

QUBO xi = 1 xi = 0
Ising s i = +1 s i = −1

Once we have defined our binary variables, we can convert our objective and constraint
into math expressions.

Objective. We consider our objective for both QUBO and Ising forms.
QUBO:
To figure out the sum of the boxes that are chosen, we can use a weighted sum: 15x1 +
20x2 + 25x3 . In this sum, the boxes that are chosen will have xi = 1 and the boxes that are
not chosen will have xi = 0. In other words, the value of boxes that are not chosen will be
multiplied by zero and so we will only be adding up the value of the boxes that are chosen.
Our objective function becomes:

min 15x1 + 20x2 + 25x3

Ising:
Using our binary variables, we can convert our +1/-1 to 1/0 using the Ising to QUBO
translation shown earlier, which maps +1 7→ 1 and −1 7→ 0. Our objective function can
then be written as:
      
s +1 s +1 s +1
min 15 1 + 20 2 + 25 3
2 2 2

Constraint. We consider our constraint for both QUBO and Ising forms.
QUBO:
Our constraint “choose exactly two boxes” means that we need exactly two of our binary
variables to have value 1, and the remaining binary variable will have value 0. In other
words, our constraint can be written as:

x1 + x2 + x3 = 2

Ising:
Our constraint “choose exactly two boxes” means that we need exactly two of our binary
variables to have value +1, and the remaining binary variable will have value -1. In other
words, our constraint can be written as:

s1 + s2 + s3 = 1

4.4.3 Step 3: Transform Math Expressions into a BQM

Our objective function is fine as written, so we only need to modify our constraint.

Copyright © D-Wave Systems Inc. Problem Formulation Guide 18

 QUBO:
To modify our constraint, we need to use the method for equalities.
Original constraint:
x1 + x2 + x3 = 2

Move everything to one side:

x1 + x2 + x3 − 2 = 0

Square the expression:

( x1 + x2 + x3 − 2)2

Ising:
Following the same method, we can rewrite our constraint as follows.

( s1 + s2 + s3 − 1)2

4.4.4 Step 4: Combine Expressions

Now that we have written our objective and constraint in BQM form, we can combine them
together to make our final model using addition and adding in a Lagrange parameter. Both
of these can be expanded and simplified to prepare for input to an Ocean python program.
QUBO:  
min (15x1 + 20x2 + 25x3 ) + γ( x1 + x2 + x3 − 2)2

Ising:
       
s1 + 1 s2 + 1 s3 + 1 2
min 15 + 20 + 25 + γ ( s1 + s2 + s3 − 1)
2 2 2

4.4.5 Exercise
Write a BQM for the following problem.
A business is trying to figure out the best way to pack a box full of items. There are five
items, with values 1, 5, 3, 4, and 2 dollars. These items have weights 2, 4, 4, 1, and 3 lbs,
respectively. The contents of the box must weigh exactly 6 lbs. What is the best way to pack
the box so that the box has the highest total value?

5 Summary of Exercises with Solutions

Below you will find solutions to the exercises included in this ”Learn to Formulate Prob-
lems” guide. For further discussion of these solutions, please visit our community forum
in Leap (https://support.dwavesys.com/hc/en-us/community/topics).

Copyright © D-Wave Systems Inc. Problem Formulation Guide 19

5.1 Introduction
Split the following problem into an objective and constraints.
A business is trying to figure out the best way to pack a box full of items. There are five
items, with values 1, 5, 3, 4, and 2 dollars. These items have weights 2, 4, 4, 1, and 3 lbs,
respectively. The contents of the box must weigh exactly 6 lbs. What is the best way to pack
the box so that the box has the highest total value?

5.1.1 Solution
This problem can be split into an objective and constraints as follows.

Objective: Maximize total value

Constraint: Total weight of contents is exactly 6 lbs

5.2 Definitions
5.2.1 QUBO
Put the following QUBO into matrix form.

x1 + x1 x2 − 3x3 + 5

Solution:  
1 1 0
 0 0 0 
0 0 −3

5.2.2 Ising
Put the following Ising model into matrix form.

s1 + s1 s2 − 3s3 + 5

Solution:
Linear:  
h= 1 0 −3

Quadratic  
0 1 0
J= 0 0 0 
0 0 0

Copyright © D-Wave Systems Inc. Problem Formulation Guide 20

5.2.3 Converting Between QUBO and Ising
QUBO to Ising:
3 1 3 1 17
s1 + s2 − s3 + s1 s2 +
4 4 4 4 4
Ising to QUBO:
−2x2 − 6x3 + 4x1 x2 + 8

5.3 Small Problems with Two Variables

5.3.1 QUBO
What is the QUBO model for the constraint x1 ≥ x2 ?
Solution:
x2 − x1 x2

5.3.2 Ising
What is the Ising model for the constraint s1 ≥ s2 ?
Solution:
1 1 1 1
− s1 + s2 − s1 s2 +
4 4 4 4

5.4 Problems with Multiple Constraints

Write a BQM for the following problem.
A business is trying to figure out the best way to pack a box full of items. There are five
items, with values 1, 5, 3, 4, and 2 dollars. These items have weights 2, 4, 4, 1, and 3 lbs,
respectively. The contents of the box must weigh exactly 6 lbs. What is the best way to pack
the box so that the box has the highest total value?

5.4.1 QUBO
Solution:
( x1 + 5x2 + 3x3 + 4x4 + 2x5 ) + γ(2x1 + 4x2 + 4x3 + x4 + 3x5 − 6)2

5.4.2 Ising
Solution:
   2
1 5 3 15 1 3
s1 + s2 + s3 + 2s4 + s5 + + γ s1 + 2s2 + 2s3 + s4 + s5 + 1
2 2 2 2 2 2

Copyright © D-Wave Systems Inc. Problem Formulation Guide 21

 References
1 A. Lucas, “Ising formulations of many np problems,” Frontiers in Physics 2, 5 (2014).
2 Collection of code examples, https://github.com/dwave-examples (2020).

Copyright © D-Wave Systems Inc. Problem Formulation Guide 22