Exterior Calculus in Graphics

Course notes for a SIGGRAPH 2023 course

Stephanie Wang, Mohammad Sina Nabizadeh, and Albert Chern

Version 1.00: June 6, 2023

Latest version available at
https://stephaniewang.page/ExteriorCalculusInGraphics/

Permission to make digital or hard copies of part or all of this work for personal or classroom
use is granted without fee provided that copies are not made or distributed for profit or
commercial advantage and that copies bear this notice and the full citation on the first
page. Copyrights for third-party components of this work must be honored. For all other
uses, contact the Owners/Authors.
Copyright is held by the owners/authors.
SIGGRAPH ’23 Courses, August 06-10, 2023, Los Angeles, CA, USA
ACM 979-8-4007-0145-0/23/08.
10.1145/3587423.3595525
 Abstract

The demand for a more advanced multivariable calculus has rapidly increased in
computer graphics research, such as physical simulation, geometry synthesis, and
differentiable rendering. Researchers in computer graphics often have to turn to
references outside of graphics research to study identities such as the Reynolds
Transport Theorem or the geometric relationship between stress and strain tensors.
This course presents a comprehensive introduction to exterior calculus, which covers
many of these advanced topics in a geometrically intuitive manner. The course
targets anyone who knows undergraduate-level multivariable calculus and linear
algebra and assumes no more prerequisites. Contrary to the existing references,
which only serve the pure math or engineering communities, we use timely and
relevant graphics examples to illustrate the theory of exterior calculus. We also
provide accessible explanations to several advanced topics, including continuum
mechanics, fluid dynamics, and geometric optimizations. The course is organized into
two main sections: a lecture on the core exterior calculus notions and identities with
short examples of graphics applications, and a series of mini-lectures on graphics
topics using exterior calculus.
 Preface

Disclaimer
These notes are still being developed and they contain mistakes. Please visit the course
project page at https://stephaniewang.page/ExteriorCalculusInGraphics/ for
the newest version. We encourage the readers to examine the statements and proofs
in these notes and determine the correctness themselves. If you come across any
errors / omissions, please email the authors while providing the page number and
notes version (available on the title page).

Prerequisites
This is a comprehensive beginner-level course to introduce exterior calculus to
students, researchers, educators, and developers. The course prerequisites include
only undergraduate-level linear algebra and multivariable calculus.

Course Rationale
The course aims to extend the mathematical tool set for the audiences to gain short
cuts and geometric intuitions through a variety of important advancements in modern
computer graphics. The scope where exterior calculus applies is not limited to geom-
etry processing which has been taught in computer graphics [Grinspun et al.(2006),
Crane et al.(2013)]. In particular, we cover examples including differentiable render-
ing, optimization techniques, continuum mechanics, optimal transports, and shape
synthesis. This course serves to fulfill the following goals:

I. Fill in the missing episodes of the mathematical language for graphics. Multi-
variable calculus and linear algebra lie at the center of continuous mathematics,
6

which is the pillar of computer graphics that aims at synthesizing our continuous
world. Furthermore, differential calculus has become increasingly important in
the recent developments of major areas of computer graphics such as render-
ing (e.g. differentiable rendering) and simulations (e.g. constitutive models).
However, the computation of differentials can appear less than intuitive using
the traditional tools from multivariable
˜ calculus. For instance, how does one
differentiate a surface integral Σ f dA regarding perturbations in the integra-
tion domain Σ? How does one differentiate a Dirac-δ function relative to its
placement? These questions naturally arise from the sensitivity analysis of
occlusion geometry in rendering and the differentiation of discontinuity and
collision in simulations. While the standard multivariable calculus curriculum
covers partial derivatives using coordinates, double and triple integrals, gradient
operators, and a few variations of Stokes Theorems, there remains a gap between
elementary vector calculus and cutting-edge graphics research. This course
provides a bridge to cross this gap by building on top of freshman calculus
and linear algebra and explaining various aspects of exterior calculus using
examples in graphics.

II. Provide a reference for exterior calculus. Most existing texts on differen-
tial forms and exterior calculus are written for the community of differen-
tial geometers. While a few texts do provide engineering-related exposi-
tions [Flanders(1963), Frankel(2011)], most coverage of continuum mechan-
ics does not include physical simulations or rendering references. More and
more graphics papers (e.g. [Wang and Chern(2021)] in geometry optimization,
[Sellán et al.(2021), Yang et al.(2021)] in physics simulations, [Nicolet et al.(2021),
Zhao et al.(2020a)] for inverse rendering) use the language of geometric calcu-
lus; there has been an increasing demand for a comprehensive reference for
exterior calculus in the graphics community. This course serves to fulfil the
demand by publishing an accompanying course notes that document a list of
exterior calculus notions and identities.

III. Illustrate continuous, geometric intuition to differential forms. Past appearance

of exterior calculus has been limited to courses covering discrete differential
geometry and geometry processing [Grinspun et al.(2006), Crane et al.(2013),
Crane(2018)]. These material focuses on a discrete and computational aspect of
exterior calculus. In particular, differential forms are described as data assigned
to points, edges, faces, etc. Expanding these discrete ideas to continuous ones is
not immediately obvious. This course introduces geometric interpretations for
the continuous exterior calculus that best illustrate the structure dependency.
With these interpretations in mind, one gains more accurate intuitions to the
important concepts in simulations and optimization methods.

Course Contents
The course is organized into two main sections. The first section consists of core
technical exposition of exterior calculus with short applications in graphics along the
 7

way. The second section is a few application-focused topics formulated with exterior
calculus.

Part I: Fundamentals
The course starts with the linear algebra of vectors and covectors. Vectors are
arrows that represent direction, whereas covectors are contour planes that represent
slopes. These two concepts are distinct. The significant application for vectors and
covectors is the optimization method of gradient descent. How does one translate
the information of slope of an objective function to an update direction in an
iterative optimization? It turns out that the method is not unique. Leveraging the
distinction between vectors and covectors can lead to efficient optimization schemes
(see Example 3.9.)
The course moves on to differential forms and introduces three pictures of
differential forms:
(i) Alternating multilinear form. A direct generalization of covectors from contour
planes to lines and points. This picture is the linear algebra aspect of differential
forms.

(ii) Fields to be integrated along geometries. Differential forms are to be paired

with test geometries. For example, flux and stress is to be evaluated on a piece
of surface. Calculus operators are adjoint operators of geometric operations
across the form-geometry pairing. This picture has been integrated in the
computation tools of Discrete Exterior Calculus.

(iii) Codimensional geometry. A point cloud represents a probability measure.

Similarly, a curve cloud, surface cloud, etc. are differential forms. Calculus
operators directly translate to geometric operations on these codimensional
geometries. This picture is the Geometric Measure Theory aspect of differential
forms.
Exterior calculus operations including the wedge product, interior product, exterior
derivative, and pullback are associated to geometric operations of intersection,
extrusion, boundary, and deformation, respectively. The course demonstrates that
many vector identities are examples of self-evident topological statements. The
differentiation of these calculus operations are also covered. The course also covers
Lie derivatives and their applications in continuum mechanics.

Part II: Graphics Topics in Exterior Calculus

While short applications are mentioned along the previous part, the second part of
the course focuses on a few applied topics.
(i) Chapter 3: Geometric optimizations. Optimization problems involving geome-
try is a crucial bit of computer graphics research. We discuss different geometry
representations and how they affect the downstream computations of geometric
properties like boundary and topological changes. We also provide a geometric
picture of the theory of optimizations, where readers will find out about the
important of choosing the correct metric for their gradient descent method.
 (ii) Chapter 4: Continuum Mechanics. Hyperelasticity models are the fundamental
physical models for soft body simulations. These physical models usually
involve many matrix fields representing the stress and strain. The course
decodes these matrices and describe them using the appropriate differential
forms. The formulation makes constitutive models more transparent.

(iii) Chapter 5: Fluid dynamics. The inviscid incompressible fluid flow is governed
by the Euler equation. By applying exterior calculus, we elucidate an alternative
formulation for the Euler equation that uses the velocity covector field as the
primary variable. Using this reformulation, one can obtain deeper insights in
the flow of fluids and come up with better algorithm for their simulation.

Speaker Biographies
Stephanie Wang – https://stephaniewang.page/
Dr. Stephanie Wang is a postdoctoral researcher in Computer Science and Engi-
neering at UCSD. She received her PhD in Mathematics from UCLA in 2020 with
a dissertation on simulating various materials using the Material Point Method.
Her research interests lie in in the intersection of mathematical analysis, scientific
computing, and computer graphics. She works on problems arising from physical
simulations and geometry processing by analyzing the mathematical models and
improving the numerical methods. In her spare time, she enjoys walking for many
hours in the mountains.

Mohammad Sina Nabizadeh – https://sinabiz.github.io/

Mohammad Sina Nabizadeh is a PhD student in Computer Science and Engineering
at UCSD. There, he is advised by Professors Albert Chern and Ravi Ramamoorthi. In
2020, he received his bachelors in Computer Engineering at the University of Waterloo,
Canada, where he was advised by Professor Christopher Batty at the Computational
Motion Group. His interests lie in many different fields of computer graphics research
such as geometric fluid dynamics, geometry processing, and rendering.

Albert Chern – https://cseweb.ucsd.edu/~alchern/

Prof. Albert Chern is an assistant professor in Computer Science and Engineering at
UCSD. He received his PhD in Applied and Computational Mathematics at Caltech
in 2017. He worked as a postdoctoral researcher in Mathematics at TU Berlin from
2017 to 2020 and began his assistant professorship at UCSD in 2020. Prof. Chern’s
research interests lie in the interplay among differential geometry, numerical partial
differential equations, and their applications in geometry processing and physics
simulations in computer graphics.

Acknowledgements
This work was funded by NSF CAREER Award 2239062 and UC San Diego Center
for Visual Computing. Additional support was provided by the WiGRAPH Rising
Stars Program, SideFX Software, and Activision Blizzard.
 Contents

I Fundamentals
1 Vectors and Covectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.1 Dual Pairing 15
1.1.1 Geometric picture of vectors and covectors . . . . . . . . . . . . . . . . . . . . . . . 16
1.1.2 Annihilator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.1.3 Adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.2 Endomorphisms and Bilinear Forms 18
1.2.1 Quadratic form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.3 Vector–Covector Conversion 1: Dual Basis 19
1.4 Vector–Covector Conversion 2: Metric 21
1.5 Vector Fields and Covector Fields 24
1.6 Differential of a Function 25
1.6.1 Conversion 1: to partial derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.6.2 Conversion 2: to gradient vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6.3 Pushforward of vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6.4 Pullback of covectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.7 Manifolds 29

2 Exterior Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1 From Multivariable Calculus to Exterior Calculus 31
2.1.1 Line and surface integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.1.2 Changing variables for multivariable integrals . . . . . . . . . . . . . . . . . . . . . . 33
2.1.3 Green, Gauss and Kelvin–Stokes theorems . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2 An Overview of the Language 35
2.3 Differential Forms 37
2.3.1 Integration and pullback . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.2 Wedge product and interior product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3.3 Exterior derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.4 Differential forms as codimensional geometries . . . . . . . . . . . . . . . . . . . . . 47
2.4 Vector Fields as Derivations 49
2.4.1 Algebra and derivations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4.2 Lie algebra for vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.4.3 Practical calculation of Lie bracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.4 Integrability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.5 Covariant Derivatives and Vector-Valued Forms 54
2.6 Lie Derivative 56
2.6.1 General definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.6.2 Lie derivative for functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.6.3 Lie derivative for vector fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.6.4 Lie derivative for differential forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.6.5 Useful identities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.6.6 Lie derivatives in the 3D Cartesian space . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.7 Pairings, Inner Products and Hodge Stars 64

II Graphics Topics in Exterior Calculus

3 Geometric Optimizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.1 Geometry representations 69
3.1.1 Lagrangian representations: manifold embedding . . . . . . . . . . . . . . . . . 69
3.1.2 Integration pairing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.1.3 Exterior calculus on combinatorial surfaces and curves (DEC) . . . . . . . . 73
3.1.4 Eulerian representations: geometric measure theory . . . . . . . . . . . . . . . . 74
3.2 A Geometric view of optimization problems 78
3.2.1 Unconstrained problems on (finite-dimensional) manifolds . . . . . . . . . . . 79
3.2.2 Unconstrained problems on functions and differential forms . . . . . . . . . . 79
3.2.3 Optimization with equality constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2.4 Karush–Kuhn–Tucker conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.3 Computational methods for optimization problems 85
3.3.1 Direct and iterative methods for unconstrained problems . . . . . . . . . . . . 85

4 Continuum Mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.1 Deformation map 90
4.2 Elasticity overview 91
4.3 Tensor Algebra 93
4.3.1 Tensor product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3.2 Tensor power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.3.3 Pullback bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.3.4 Type algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.4 Postulates of Elasticity 99
4.4.1 Deformation gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4.2 Pullback metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.4.3 Strain energy density function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5 Derivation of Elasticity 102
4.5.1 Summary of the derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.6 Cauchy stress tensor 105
4.7 More on the strain energy density function 107
4.7.1 Strains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.7.2 Isotropic models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.8 Conclusion 113

5 Fluid Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

5.1 Euler equations 116
5.1.1 Vector-based Euler equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.1.2 Covector Euler equations as a constrained minimization . . . . . . . . . . . . 119
5.2 Advantages of Covector Euler Equation 120
5.3 Conclusion 125

III Appendix
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
I
Fundamentals

1 Vectors and Covectors . . . . . . . . . . 15

1.1 Dual Pairing
1.2 Endomorphisms and Bilinear Forms
1.3 Vector–Covector Conversion 1: Dual Basis
1.4 Vector–Covector Conversion 2: Metric
1.5 Vector Fields and Covector Fields
1.6 Differential of a Function
1.7 Manifolds

2 Exterior Calculus . . . . . . . . . . . . . . . . . 31
2.1 From Multivariable Calculus to Exterior Calculus
2.2 An Overview of the Language
2.3 Differential Forms
2.4 Vector Fields as Derivations
2.5 Covariant Derivatives and Vector-Valued
Forms
2.6 Lie Derivative
2.7 Pairings, Inner Products and Hodge Stars
 1. Vectors and Covectors

This chapter is a review of Linear Algebra, but with emphasis on drawing distinction
between vectors and covectors. We state a few theorems from linear algebra along
the exposition. Most linear algebra textbooks would state the theorems for matrices,
and hence inherently basis dependent. We reprise the theorems using the language
of vectors and covectors. We skip the proofs of most of these theorems. Readers
can easily reconstruct the proofs by following the matrix counterpart from any
matrix-based linear algebra textbook.

1.1 Dual Pairing

The dual pairing is a recurring operation throughout linear algebra. Linear combi-
nations of vectors a1⃗v1 + a2⃗v2 + a3⃗v3 are pairing between a list of coefficients and a
list of vectors. Matrix multiplications pair row vectors and column vectors:
 
⊺  | | |  ⊺
r1 c1 r⊺1 c2 r⊺1 c3
 
r1
| |

| |

c1 c2 c3 = ⊺ . (1.1)
r⊺2 r2 c1 r⊺2 c2 r⊺2 c3
 
| | |

In a linear system of equations, for example

    
1 kg 1 kg x $/kg $3
= , (1.2)
2 kg 3 kg y /kg
$ $8

we see pairings between arrays quantities (rows with unit of kilograms) and an array
of prices (x and y) that produce scalars in costs (right-hand side). From these
examples, we observe that the pairings are between different types of objects. In
particular the pairings are not inner product. It does not make sense to talk about
“cosine of angles” between two vectors of different units!
 16 Chapter 1. Vectors and Covectors

⃗v ⃗v

α
=
α 2 α
= =
⃗0 α 1 ⃗0 α 1
= =
0 0
α
=
−1

Figure 1.1 Covector-vector pairing (left) and a simplified representation of a

covector using only a single slab (right). A covector α is a linear scalar
function that can be visualized as a series of equidistant codimension-
one planes (hyperplanes) representing its level sets. A vector is
represented by an arrow based at the origin ⃗0. The dual pairing
of covector α and vector ⃗v is computed by counting the number of
piercings of the arrow (⃗v ) through the hyperplanes (α).

To distinguish the two types of objects in the linear pairing, we call one of the
object vector and the other one covector. A covector is to be paired with a vector.
Definition 1.1 A covector is a scalar linear function of vectors. The evaluation,
also known as the dual pairing, of a covector α and a vector ⃗v is denoted by

α(⃗v ) = ⟨α|⃗v ⟩ = ⟨⃗v |α⟩. (1.3)

Definition 1.2 — Dual space. Let V be a real vector space. The space of all
linear
covectors on V is called the dual space V ∗ = {α : V −−−→ R}.

Theorem 1.1 For a finite dimensional space V , we have V ∗∗ ∼

= V .a In infinite
dimensional cases, we may have V ∗∗ ⊇ V .
a
This isomorphism is canonical, given by ev : V → V ∗∗ , ev(⃗v ) := (α 7→ ⟨α|⃗v ⟩). In particular,
this is a stronger isomorphism than just having the same dimension. For example, when V is a
finite dimensional space, one can say that V ∼
= V ∗ because their dimensions are the same. But
∗
there is no canonical map between V and V .

Throughout the note, we will postulate that V ∗∗ = V .

1.1.1 Geometric picture of vectors and covectors

The standard geometric picture of a vector ⃗v ∈ V is an arrow based at the origin
⃗0 ∈ V . The geometric postulates for a vector space V is that it is an affine space
with a distinguished point called the origin ⃗0. An affine space is a space equipped
with the notion of parallelism, or a projective space with a distinguished hyperplane
 1.1 Dual Pairing 17

⃗v

⃗u + ⃗v α β

⃗0

γ
⃗u
a⃗u

Figure 1.2 Addition of vectors (left) and covectors (right).

at infinity, or equivalently a Euclidean space without metric. The vector addition

⃗u + ⃗v is constructed by drawing a parallelogram formed by ⃗u , ⃗v and connecting the
origin to its opposite vertex (see Figure 1.2, left). The scaling of a vector a⃗v can be
constructed using similar triangles and diagonal bisection in parallelograms.
The geometric picture of a covector α ∈ V ∗ is a foliation of equidistant parallel
hyperplanes in V , each of which represents a level set {α = k } ⊂ V of the scalar
linear function α. The dual paring ⟨α|⃗v ⟩ measures the number of planes of α pierced
through by the arrow ⃗v . See Figure 1.1, left.
Since we know how to geometrically construct equidistant parallel hyperplanes
(again via the parallel postulate) we only need to draw two special level sets: {α = 0}
and {α = 1} for illustrating α. Note that {α = 0} = ker α passes through the origin.
Hence each covector is a slab with one side leaning against the origin. See Figure 1.1,
right.
The covector addition α + β is constructed by first intersecting the two slabs,
which forms a parallelogram prism, followed by building a new slab whose far side
from the origin passes through the diagonal of the parallelogram that is not incident
to the origin (see Figure 1.2). The scaling of a covector aα is defined by scaling the
thickness of the slab by 1/a . In particular, the narrower the slab is, the more dense
the foliation of level sets of α is, and the larger the slope the linear function α has.

1.1.2 Annihilator
It is important to note that the dual pairing is different from an inner product. The
dual pairing pairs a vector and a covector, whereas the inner product measures how
aligned two vectors are. We do not need the inner product structure (a metric) to
perform dual pairing. In fact, many inner product related notions we have seen in a
first course in linear algebra are independent of metric. Here, we give an example:
the notion of “orthogonal complement” of a subspace will be metric-independent if
we treat it as a subspace in the dual space.
 18 Chapter 1. Vectors and Covectors

Definition 1.3 — Annihilator. Let U ⊂ V be a linear subspace of the vector space

V . The annihilator U ◦ ⊂ V ∗ of U is a linear subspace in the dual space defined
by

U ◦ = {α ∈ V ∗ | ⟨α|⃗u ⟩ = 0 for all ⃗u ∈ U } (1.4)

Theorem 1.2 U ◦◦ = U .

1.1.3 Adjoint
linear
A linear map A : U −−−→ V from a vector space U to another vector space V
induces a dual (adjoint) map from the dual target space V ∗ back to the dual domain
U ∗.
linear
Definition 1.4 The adjoint of a linear map A : U −−−→ V is the linear map
linear
A∗ : V ∗ −−−→ U ∗ defined by

⟨A∗ λ|⃗u ⟩ = ⟨λ|A⃗u ⟩ for all ⃗u ∈ U and λ ∈ V ∗ . (1.5)

Theorem 1.3 A∗∗ = A.

In case U , V are equipped with bases that allow a matrix representation A = (Aij )
for A, the adjoint A∗ is simply the transpose A⊺ .
linear
Definition 1.5 The four fundamental subspaces of a linear map A : U −−−→ V
are

• kernel: ker(A) := {⃗u ∈ U | A⃗u = 0} ⊂ U ,

• image: im(A) := {A⃗u | ⃗u ∈ U } ⊂ V ,

• cokernel: ker(A∗ ) := {λ ∈ V ∗ | A∗ λ = 0} ⊂ V ∗ ,

• coimage im(A∗ ) := {A∗ λ | λ ∈ V ∗ } ⊂ U ∗ .

Theorem 1.4 — Theorem of four fundamental subspaces.

ker(A)◦ = im(A∗ ), im(A)◦ = ker(A∗ ) (1.6)

1.2 Endomorphisms and Bilinear Forms

In the previous section, we distinguish the difference between vectors and covectors.
In this section, we continue to draw the distinction between them, and separate
linear linear
linear transformations of type V −−−→ V and V −−−→ V ∗ . While both types can
be represented by a square matrix, their geometric picture are vastly different.
linear
A linear map A : V −−−→ V that maps a vector space to itself is called an
endomorphism. For example, the identity map is an endomorphism. An eigenvalue
 1.3 Vector–Covector Conversion 1: Dual Basis 19

problem A⃗v = λ⃗v is formulated for endomorphisms A. Powers Ak , exponentials e A

and so on are defined only for endomorphisms.
linear
A linear map B : V −−−→ V ∗ that maps a vector space to its dual space is called
a bilinear form. For each ⃗v ∈ V , the resulting value B (⃗v ) ∈ V ∗ is a covector. Recall
that a covector is a scalar-valued linear function that can take another vector and
return a scalar. That is, a bilinear map is to be evaluated like B (⃗v )(⃗ w ) bi-linearly
dependent on the two input vectors ⃗v , w ⃗ ∈ V . In other words, a bilinear map can
bilinear
also be viewed as B : V × V −−−−→ R.
linear
Note that the adjoint of a bilinear form B : V −−−→ V ∗ would have the same
linear
type B ∗ : V −−−→ V ∗ (using V ∗∗ = V ). Only for bilinear forms we can compare B
with B ∗ as they are objects of the same type.
A bilinear form B is said to be symmetric or self-adjoint if B = B ∗ . Equivalently,
a bilinear form is symmetric if B (⃗u )(⃗v ) = B (⃗v )(⃗u ) for all ⃗u , ⃗v ∈ V . A bilinear form
is said to be skew-symmetric if B = −B ∗ ; i.e. B (⃗u )(⃗v ) = −B (⃗v )(⃗u ).

1.2.1 Quadratic form

A symmetric bilinear form is equivalent to a quadratic form. A quadratic form
quadratic
q : V −−−−−→ R is a function such that q(a⃗v ) = a 2 q(⃗v ) for all scalar a ∈ R. A
linear
bilinear form B : V −−−→ V ∗ gives rise to a quadratic form by q(⃗v ) := B (⃗v )(⃗v ).
quadratic
Conversely, a quadratic form q : V −−−−−→ R gives rise to a symmetric bilinear form
linear
B : V −−−→ V ∗ using the formula for expanding a square
q(⃗u + ⃗v ) = q(⃗u ) + q(⃗v ) + 2B (⃗u )(⃗v ). (1.7)
The symmetric bilinear form B (⃗u )(⃗v ) = 12 (q(⃗u + ⃗v ) − q(⃗u ) − q(⃗v )) is called the
polarization of q. Therefore, a symmetric bilinear form is fundamentally the same
as a quadratic form.

1.3 Vector–Covector Conversion 1: Dual Basis

Vectors and covectors are different. The former is an arrow based at the origin, and
the latter is a slab bounded by a pair of parallel hyperplanes, one of which passes
through the origin. Without additional structure, there is no method converting a
vector to a covector.
One scenario where we can convert vectors to covectors is when we have a full set
of basis vectors. Suppose V is an n-dimensional vector space. Suppose ⃗v1 , . . . , ⃗vn are
n linearly independent vectors. Then they form a basis. Geometrically, they span a
non-degenerate parallelepiped. This parallelepiped has faces giving rise to n pairs of
parallel hyperplanes, one of which passes through the origin. As such, there emerges
n covectors. This construction is canonical and unique. The resulting covectors is
called the dual basis for the dual space V ∗ . See Figure 1.3, left.
Conversely, given n linearly independent covectors, as a basis for V ∗ , we can
geometrically visualize them as n slabs which intersect into parallelepiped with a
corner being the origin. By reading off the edge vectors of the parallelepiped based
at the origin, we get the find the dual basis for V .
The following is the algebraic definition of the above geometric construction.
20 Chapter 1. Vectors and Covectors

Theorem 1.5 — Dual basis. Let ⃗v1 , . . . , ⃗vn ∈ V be a basis for a vector space V .
Then there exists a unique set of covectors α1 , . . . , αn ∈ V ∗ forming a basis for
V ∗ such that
(
1 i =j
⟨αi |⃗vj ⟩ = δij = (1.8)
0 i ̸= j .

If V = Rn and ⃗v1 , . . . , ⃗vn are column vectors v1 , . . . , vn , then the relationship

between v1 , . . . , vn and the dual basis as row vectors α1 , . . . , αn is a matrix inversion
   −1
α1 | |
|

|
.. (1.9)
 = v1 · · · vn  .
  
 .
αn | |
|

Note that we need a full set of basis vectors to produce a full set of basis covectors.
We cannot turn a single vector to a single covector without any additional structure.
■ Example 1.1 — Diagonalization. Rn = R ⊕ R ⊕ · · · ⊕ R is a special vector
space called Cartesian space. Each element is an n-tuple of numbers. A map
D : Rn → Rn is called a diagonal matrix if the map is just a channel-wise scaling

D : (r1 , . . . , rn ) 7→ (λ1 r1 , . . . , λn rn ). (1.10)

linear
A basis ⃗v1 , . . . , ⃗vn for a vector space V induces a map X : Rn −−−→ V by
linear
X ((r1 , . . . , rn )) := ni=1 ri ⃗vi . Note that the inverse map X −1 : V −−−→ Rn is given
P
by X −1 (⃗u ) = (⟨α1 |⃗u ⟩, . . . ⟨αn |⃗u ⟩) where α1 , . . . , αn ∈ V ∗ is the dual basis.
A diagonalization of an endomorphism A ∈ End(V ) is to express the endo-
morphism by

A = X ◦ D ◦ X −1 , (1.11)

which is achieved by solving the eigenvalue problem.

Definition 1.6 The trivial Cartesian metric on Rn and Rn∗ = Rn is the identity
map I.
linear
If V is equipped with a metric, then a basis X : Rn −−−→ V is orthonormal if
and only if X : (Rn , I) → (V , ♭) is an isometry (a map that preserves the metric).

isometry
Theorem 1.6 If X : (Rn , I) −−−−→ (V , ♭) is an orthonormal basis, then

X −1 = X ∗ ◦ ♭ (1.12)
■

■ Example 1.2 — Polar decomposition. Let U , V be two spaces equipped with

 1.4 Vector–Covector Conversion 2: Metric 21

α2 ⃗v ♭
α1
⃗v
⃗e2
⃗e1

♭(⃗x )(⃗x ) = 1
“unit sphere”

Figure 1.3 Converting covectors to vectors and vice versa using either the dual
basis (left) or a metric (right).

metric ♭U , ♭V . The polar decomposition of a linear map A : U → V is

isometry
A = RY , where R : U −−−−→ V and Y ∈ End(U ) (1.13)

such that ♭U ◦ Y is a symmetric positive definite bilinear form. ■

1.4 Vector–Covector Conversion 2: Metric

A metric or an inner product is an additional structure for a vector space. The
most common vector space equipped with a metric is the Euclidean space (Rn , ⟨·, ·⟩)
where the Euclidean inner product ⟨⃗u , ⃗v ⟩ measures the length and the angle between
vectors ⃗u and ⃗v .
linear
Definition 1.7 A metric is an injective symmetric bilinear form ♭ : V −−−→ V ∗
(injectivity means that ker(♭) = {0} or that ♭ is invertible). We usually denote
♭(⃗u )(⃗v ) = ⟨⃗u , ⃗v ⟩ and ♭(⃗u ) = ⃗u ♭ . The inverse ♭−1 of the bilinear form ♭ is denoted
linear
by ♯ = ♭−1 : V ∗ −−−→ V .
The symbols ♭, ♯ are called flat and sharp respectively, together they are called
musical isomorphisms.
Recall that a symmetric bilinear form is equivalent to a quadratic form. In the
case of a metric, we denote the quadratic form by | · |2 :
|⃗v |2 = ♭(⃗v )(⃗v ) = ⟨⃗v ♭ |⃗v ⟩ = ⟨⃗v , ⃗v ⟩. (1.14)
Conversely, we can derive the musical isomorphism ♭ from the quadratic form | · |2
by using the polarization identity if V is a R-vector space:
1
|⃗u + ⃗v |2 − |⃗u |2 − |⃗v |2 . (1.15)


♭(⃗u )(⃗v ) = ⟨⃗u , ⃗v ⟩ =
2
A metric | · |2 on V induces a metric on V ∗ given by
|α|2V ∗ := ⟨α|α♯ ⟩. (1.16)
22 Chapter 1. Vectors and Covectors

Geometrically, we can visualize a metric structure | · |2 on V by the “unit sphere”

Q := ⃗x ∈ V |⃗x |2 = 1 ⊂ V , (1.17)


which is a non-degenerate quadratic surface in the affine space centered at the origin.
In addition to the parallelism structure in the vector space, we have a “compass”
allowing us to define distances and angles. The relationship between the arrow ⃗v ∈ V
and the slab ⃗v ♭ ∈ V ∗ is that the tip ⃗v of the arrow and the {⃗v ♭ = 1} hyperplane
of the slab are in polarity with respect to the quadratic surface Q. See Figure 1.3,
right.
Definition 1.8 In projective geometry, the polar hyperplane h to a point x with
respect to a quadratic hypersurface Q is the hyperplane that passes through all
the points on Q at which the tangent plane to Q contains x . If x is inside Q
so that there is no tangent plane of Q containing x , then h is defined by the
collection of points (outside of Q) whose polar hyperplane contains x . The point
x is called the pole of the plane h with respect to Q if h is the polar of x .

If α = ⃗v ♭ , then the parallel hyperplanes of α are orthogonal to ⃗v with respect to

the metric ♭.
Musical isomorphisms generally give different vector–covector conversions from
the dual basis construction. In fact:
Theorem 1.7 Let ⃗v1 , . . . , ⃗vn ∈ V be a basis for a vector space V and let α1 , . . . , αn
be its dual basis. Let ♭ be a metric on V . Then αi = ⃗vi♭ if and only if ⃗v1 , . . . , ⃗vn is
an orthonormal basis with respect to the metric ♭.

Definition 1.9 A metric is called positive definite if |⃗v |2 > 0 for all ⃗v ̸= 0.

For most discussion we do not need to assume that a metric is positive definite.
■ Example 1.3 — Eigenvalue problems for symmetric matrices. Eigenvalue prob-
lems are posed for endomorphisms and symmetry are adjective only for bilinear
forms. This makes us wonder what is going on with the famous theorem stating
that “if a matrix is symmetric, then it has real eigenvalues and the eigenvectors
are orthogonal.” The answer is that there is a hidden metric that is often omitted.
Here is the restored version of the statement.
linear
Let B , C be two bilinear forms of type V −−−→ V ∗ . Then it is sensible to pose
the eigenvalue problem as

B⃗v = λC ⃗v . (1.18)
1.4 Vector–Covector Conversion 2: Metric 23

linear
Theorem 1.8 If ♭ is a positive definite metric on V , and B : V −−−→ V ∗ is a
symmetric bilinear form. Then there exists an orthonormal basis ⃗v1 , . . . , ⃗vn
with respect to ♭ and real numbers λ1 , . . . , λn ∈ R such that

B⃗vi = λi ♭⃗vi . (1.19)

⃗vi and ⃗vj are perpendicular under both ♭ and B : ♭(⃗vi )(⃗vj ) = B (⃗vi )(⃗vj ) = 0 for
i ̸= j .

Theorem 1.9 The result of Theorem 1.8 holds (except that the eigenvalues
λ1 , . . . , λn may be complex) for the more general normal bilinear forms. A
linear
bilinear form B : V −−−→ V ∗ is called normal if

B ∗ ◦ ♯ ◦ B = B ◦ ♯ ◦ B ∗. (1.20)
■

■ Example 1.4 — Rotation matrices. A matrix R is called a special orthogonal

matrix or a rotation matrix if R⊺ R = I. Rotation is supposed to be an endomor-
phism, as it maps a vector to a rotated vector in the same space. But then the
adjoint R⊺ would be an endomorphism on the dual space V ∗ . It does not make
sense to compose R⊺ with R. What is missing here is some metric dependency.
linear
Definition 1.10 Let V be a vector space with metric ♭ : V −−−→ V ∗ . An endo-
morphism R ∈ End(V ) is said to be orthogonal or unitary if the composition
of the following cyclic sequence of linear maps is the identity endomorphism

VO
R / V (1.21)
♯ ♭

V∗ o R∗
V∗

If det(R) = 1, then R is called a special unitary transform. Another way to

describe unitarity is that the pullback metric R ∗ ◦ ♭ ◦ R by R is the same as
the metric ♭ = R ∗ ◦ ♭ ◦ R. In other words, it is a linear isometry.

Definition 1.11 In general, for two spaces U , V both equipped with metric
linear
♭U , ♭V , a linear map A : U −−−→ V is an isometry if

A∗ ◦ ♭V ◦ A = ♭U . (1.22)

Equivalently,

⟨A⃗x , A⃗y ⟩V = ⟨⃗x , ⃗y ⟩U for all ⃗x , ⃗y ∈ U . (1.23)

■
 24 Chapter 1. Vectors and Covectors

■Example 1.5 — Singular value decomposition. Let U , V be two spaces equipped

with metric ♭U , ♭V . The singular value decomposition (SVD) of a linear map
linear
A : U → V is to find an orthonormal basis EU : Rm −−−→ U and an orthonormal
linear
basis EV : Rn −−−→ V such that
 
σ1
A = EV ◦ 
 σ2  ◦ EU
 −1
(1.24)
.. |{z}
. EU∗ ◦♭
U

where σ1 ≥ σ2 ≥ · · · and the diagonal matrix may be truncated to a rectangular

matrix if m ̸= n. ■

1.5 Vector Fields and Covector Fields

A typical picture of a vector field is many arrows scattered over a domain, each
having a different based point. This collection of arrows represents a flow where each
arrow gives away the flow direction at the location of its base point. Note that the
arrows composing the vector field are based at different based points, which is in
contrast to vectors in a vector space which are all based at the same origin. Therefore,
to describe vector fields sensibly, we must consider a collection of vector spaces Vp
indexed by the base point p in the domain. A vector field ⃗v is an assignment of a
vector ⃗vp ∈ Vp for every point p in the domain.
With the above picture in mind, let us introduce the standard mathematical
notations for vector fields.
Let M ⊂ Rn be a region in an n-dimensional space representing a domain. For
each p ∈ M , define a vector space Tp M given by a copy of Rn . We call Tp M the
tangent space to the domain M at p. The elements of Tp M are called tangent
vectors at p.
The elements of Tp M are associated with velocities at which a particle at p can
travel. For each smooth parameterized curve γ : (−ϵ, ϵ) → M that passes through
p at t = 0, that is γ(0) = p, the instantaneous velocity dγ dt
at t = 0 is an element
of Tp M ; that is, γ̇(0) ∈ Tp M . Conversely, for each tangent vector ⃗vp ∈ Tp M , there
exists (non-uniquely) a parameterized curve γ : (−ϵ, ϵ) → M such that γ(0) = p and
γ̇(0) = ⃗vp .
The collection TM = {(p, ⃗vp ) | p ∈ M , ⃗vp ∈ Tp M } = p∈M Tp M of all tangent
F
vectors at various base points p is called the tangent bundle over M . There is a
projection operator π : TM → M , π : ⃗vp 7→ p. That is, π takes in a tangent vector
⃗vp ∈ TM and returns the information p ∈ M of which base point ⃗vp has. The
tangent space Tp M at p is the same as the preimage π −1 ({p}) of the point p via the
projection. The preimage π −1 ({p}) of a point via the projection of a bundle is often
called a fiber of the bundle.
A vector field is an assignment of an element in each fiber of the tangent bundle.
Such an object is called a section of the bundle.
 1.6 Differential of a Function 25

Definition 1.12 A section ⃗v of a bundle π : TM → M is a map ⃗v : M → TM such

that π ◦ ⃗v = idM . That is, π(⃗v (p)) = p for all p ∈ M .

The formal definition given above is equivalent to saying that at any given point
p ∈ M , the map ⃗v (p) is a vector in the tangent space Tp M of the point p, and hence
π(⃗v (p)) = p.
The space of all sections of the tangent bundle (i.e. the space of all vector fields
defined over M ) is denoted by Γ(TM ). So, a vector field would be instanced as
⃗v ∈ Γ(TM ).
A covector field α ∈ Γ(T ∗ M ) is a section of the cotangent bundle T ∗ M over M .
The cotangent bundle is defined such that each of its fiber Tp∗ M is the dual space of
the tangent space Tp M at the same base point.

1.6 Differential of a Function

One of the most important applications of distinguishing vectors and covectors is to
understand the meaning of taking derivatives of a function.
Let M ⊂ Rn be a region in an n-dimensional space representing a domain, on
which we will consider generic scalar-valued non-linear functions g : M → R.
Definition 1.13 The differential of a function g : M → R at a point p ∈ M in a
domain M is a covector dg|p based at p. This covector takes in a vector ⃗vp based
at p (representing a small displacement about p) and returns the rate of change
⟨dgp |⃗vp ⟩ of g in the direction of ⃗vp . Concretely, for each ⃗vp ∈ Tp M consider a
parameterized curve γ : (−ϵ, ϵ) → M with γ(0) = p and γ̇(0) = ⃗vp , and define
D E d
dg|p ⃗vp := g(γ(t)). (1.25)
dt t=0

The differential dg of a scalar function g : M → R is a covector field dg ∈ Γ(T ∗ M )

describing the “slope” of g at every point by the best fitting linear function on every
tangent space.

1.6.1 Conversion 1: to partial derivatives

A coordinate system on M is a set of scalar functions x1 : M → R, . . . , xn : M → R
such that the covectors dx1 , . . . , dxn form a basis for every cotangent space Tp∗ M .
Let ⃗e1 , . . . , ⃗en ∈ Γ(TM ) be the dual basis of dx1 , . . . , dxn .
Definition 1.14 — Partial derivatives. The partial derivative of g along the k -th
direction in a coordinate system (x1 , . . . , xn ) is defined by

∂g
:= ⟨dg|⃗ek ⟩. (1.26)
∂xk

Equivalently, ∂g
∂x1
∂g
, . . . , ∂xn
are the coefficients when expressing the covector dg in
 26 Chapter 1. Vectors and Covectors
terms of the covector basis from the coordinate system
∂g ∂g
dg = dx1 + · · · dxn . (1.27)
∂x1 ∂xn

∂g
R The partial derivative ∂x 1
depends not only on the coordinate function x1 but
it depends on the entire coordinate system x1 , . . . , xn . This is because the
construction relies on taking the dual basis. Intuitively, the partial derivative
depends on which other variables are fixed during the variation.

1.6.2 Conversion 2: to gradient vector

The gradient vector of a function g is the vector whose direction is the steep-
est ascending direction and whose magnitude is the slope along that direction.

Definition 1.15 — Gradient. Let M be a domain equipped with metric ♭ on every

tangent space. Then the gradient grad g of a function g : M → R is defined by

grad g := (dg)♯ . (1.28)

One may use the polarity geometric picture of the musical isomorphism about the
unit sphere to see why the gradient vector points in the steepest ascending direction.
Note that the gradient of a function is independent of the coordinate. It does
however depend on a choice of metric.

R In many expositions in optimization or machine learning, the gradient is

defined by the array of partial derivatives
 
? ∂g ∂g
grad g = ,..., . (1.29)
∂x1 ∂xn
This is not true unless the coordinate system is orthonormal and therefore it
is misleading. Note that partial derivatives arise from the conversion of dual
basis, and the gradient arises from the conversion using a metric.

1.6.3 Pushforward of vectors

linear
The idea of the differential of a function dgp : Tp M −−−→ R of a scalar function
g : M → R can be extended for general maps.
Let ϕ : M → N be a general nonlinear map from a space M to another space N .
Then the differential d ϕ of the map ϕ is a linear map between the corresponding
tangent spaces
linear
d ϕp : Tp M −−−→ Tϕ(p) N . (1.30)
Such a linear map transforms a rate of change ⃗vp ∈ Tp M at the input p ∈ M to a
rate of change of (d ϕp )(⃗v ) ∈ Tϕ(p) N at the output ϕ(p) ∈ N (Figure 1.4). That is,
d ϕp (⃗vp ) describes how sensitive the value of ϕ is depending on variation in the ⃗v
direction.
 1.6 Differential of a Function 27

Definition 1.16 — Pushforward. Let ϕ : M → N be a general nonlinear map. The

linear
linear map ϕ∗ = d ϕ : TM → TN , with (ϕ∗ )p : Tp M −−−→ Tϕ(p) N , is called the
pushforward map of ϕ. It is defined such that for every scalar function g : N → R,
the composited scalar function f := (g ◦ ϕ) : M → R has the differential given by
the compositions of linear maps

df = dg ◦ ϕ∗ . (1.31)

Theorem 1.10 — Jacobian matrix. Suppose we put a coordinate system x1 , . . . ,

xm : M → R on M and a coordinate system y1 , . . . , yn : N → R on N . Let
⃗a1 , . . . , ⃗am be the coordinate vectors as the dual basis of the covector basis
dx1 , . . . , dxm . Similarly, let ⃗e1 , . . . , ⃗en be the dual basis of dy1 , . . . , dyn . Now,
call ϕi := yi ◦ ϕ : M → R as the resulting coordinate value of the Pm mapping ϕ.
For each vector ⃗v on M , we can write it under the basis as ⃗P v = i=1 vi ⃗ai . The
resulting vector after being applied by pushforward is ϕ∗⃗v = nj=1 wj ⃗ej . Then,
  ∂ϕ1 ∂ϕ1  
···
 
w1 ∂x1 ∂xm v1
 ..   .. ..   .. 
 . = . .  .  (1.32)
∂ϕn ∂ϕn
wn ∂x1
··· ∂xm
vm
| {z }
F

where the partial derivatives are defined in the way of Section 1.6.1. The matrix
F ∈ Rn×m is called the Jacobian matrix or the deformation gradient of ϕ under
the coordinates (x1 , . . . , xm ) and (y1 , . . . , yn ).

Theorem 1.11 The sensitivity analysis for a composition ψ ◦ ϕ : M → L of maps

ϕ : M → N and ψ : N → L is simply the composition of their pushforwards:

(ψ ◦ ϕ)∗ = ψ∗ ϕ∗ , or d (ψ ◦ ϕ)p = d ψϕ(p) d ϕp . (1.33)

This rule is called the chain rule.

Evaluating the differential by composing the chain of differentials of subcomponents
is the basis for the forward mode auto-differentiation programs.

1.6.4 Pullback of covectors

Definition 1.17 — Pullback of covectors. The adjoint

linear
ϕ∗ : Tϕ(p)
∗
N −−−→ Tp∗ M (1.34)

linear
of the pushforward ϕ∗ : Tp M −−−→ Tϕ(p) N is called the pullback operator (for
covectors) via the mapping ϕ : M → N . For each covector (field) α on N , ϕ∗ α
becomes a covector (field) on M and is called the pullback covector through ϕ.
28 Chapter 1. Vectors and Covectors

ϕ∗

ϕ∗⃗v
p q ⃗v ϕ(q)

ϕ∗ α ϕ(p)
α
ϕ∗
M N

Figure 1.4 Given a map ϕ: M → N , a point p ∈ M is mapped to ϕ(p) ∈ N ,

a vector ⃗v ∈ Tq M is sent to ϕ∗⃗v ∈ Tϕ(q) N using the pushforward
∗ N is sent to a covector ϕ∗ α ∈ T ∗ M
operator, and a covector α ∈ Tϕ(p) p
using the pullback operator.

By definition of adjoint,

⟨ϕ∗ α|⃗v ⟩ = ⟨α|ϕ∗⃗v ⟩ for all α ∈ Tϕ(p)

∗
N and ⃗v ∈ Tp M . (1.35)

See Figure 1.4.

When written in coordinates, the pullback operator is the transpose of the
Jacobian matrix.
The word pullback can also be used for composition of function. Suppose g : N →
R is a function on N and suppose ϕ : M → N . Then f : M → R defined by
f = g ◦ ϕ =: ϕ∗ g is called the pullback of g via the mapping ϕ.
In summary, the overloaded concept of pullback are given as follows. When it is
acted on functions and on covector fields: for ϕ : M → N ,
linear
ϕ∗ : (N → R) −−−→ (M → R), ϕ∗ g = g ◦ ϕ (1.36)
func func
linear
ϕ∗ : Γ(T ∗ N ) −−−→ Γ(T ∗ M ), ( ϕ∗ α)|p = (d ϕ|p )∗ (α|ϕ(p) ). (1.37)
covec covec

Now, the chain rule can be summarized as:

Theorem 1.12 — Pullback and differential commute. The pullbacks ϕ∗ for func-
tions g and covectors dg satisfy

d ( ϕ∗ g) = ϕ∗ (dg) (1.38)
func covec

Theorem 1.13 Let ϕ : M → N and ψ : N → L. Then

(ψ ◦ ϕ)∗ = ϕ∗ ◦ ψ ∗ . (1.39)
 1.7 Manifolds 29

■ Example 1.6 — Back-propagation. Suppose we have the sensitivity ϕ∗ = d ϕ,

ψ∗ = d ψ of maps ϕ : M → N and ψ : N → L. Suppose g : L → R is some “cost
function.” Through ϕ and ψ we can pull the function back and formulate a cost
function f on the parameter space M defined by f = g ◦ ψ ◦ ϕ.
How do we compute the differential dfp at a particular set of parameters
p ∈ M?
In a forward-mode differentiation method, one would first say that dfp is, under
a coordinate system (x1 , . . . , xm ), given by df = ∂x
∂f
1
dx1 + · · · + ∂x∂fm dxm where ∂x
∂f
i
=
df (⃗ai ) where ⃗ai is the coordinate vector. To find each of these partial derivatives,
we expand df = (dg)(d ψ)(d ϕ), and evaluate df (⃗ai ) = (dg)(d ψ)(d ϕ)(⃗ai ) for each i .
These pushforward operations have to be performed for m numbers of time.
An alternative approach is the so-called reversed-mode differentiation (a.k.a.
back-propagation)

dfp = (ϕ∗p )(ψϕ(p)

∗
)(dgψ(ϕ(p)) ). (1.40)

In this case, we only need to pullback one covector (since g is scalar valued),
rather than pushing forward m vectors like in the forward mode differentiation. ■

1.7 Manifolds
The domain M discussed in Section 1.5 and onward were assumed to be a subset
of Rn . In general, the above above discussion also works seamlessly when M is a
manifold. What is a manifold? At high level it is a generalization of curves and
surfaces, which are particularly interested in computer graphics. Its precise definition
in mathematics is somewhat subtle.
Consider the following curve that almost touches itself somewhere. When per-

forming computations on this curve, we care about the relation between point A and
point B a lot more than the relation between point A and point C, despite point C
being closer to point A in the “ambient” space, which is the whiteboard where the
curve was drawn on. This leads to us defining a customary “neighboring relation”
over a geometry instead of using the absolute Euclidean distance. Such neighboring
relation is formally referred to as topology. Mathematicians define topological spaces
in the broadest sense, and later define manifolds as topological spaces that can afford
differentiation and integration, similar to a Euclidean space.
30 Chapter 1. Vectors and Covectors

a manifold not a manifold

a manifold not a manifold

In essence, a manifold is a space that can be locally parameterized by a Cartesian

space. Through pullback and pushforward, calculus that is built upon differentials
of functions automatically carries over from Cartesian space to manifolds. So, in the
remainder of the course note, we can either think of a domain as a manifold or a
subset of Rn ; the language is the same.
 2. Exterior Calculus

In this chapter, we assume that the reader has a firm understanding of multivariable
calculus, as well as the linear algebra that involves vectors and covectors (Chapter 1).
However, you might wonder, why then, do we need to explore exterior calculus if
multivariable calculus already provides a comprehensive mathematical framework?
Historically, exterior calculus has emerged as a tool primarily because it provides
powerful notations to elucidate the principles of differential and integral calculus on
manifolds, abstract entities that can represent general domains and curved geometries.
Under the framework, one carefully distinguishes coordinate-dependent multivariable
quantities, and directly works with their geometrically meaningful, coordinate-free
representations.
It turns out that exterior calculus is a powerful language not limited to the abstract
manifold study. It proves to be useful in the broader application of multivariable
calculus, even for problems defined on flat domains. Examples include solid mechanics,
fluid dynamics, and optimization problems. The coordinate-free formalism exposes
the underlying structures far better than piles of indices pointing to various coordinate
components. One may write down equations that relate geometric measurements,
say in a continuum, at a higher level that is closer to our physical and geometric
intuition. Consequently, this approach results in a clearer, more straightforward
pathway for mathematical reasoning for the phenomena and algebraic identities
within these differential equations.

2.1 From Multivariable Calculus to Exterior Calculus

When it comes to multivariable integration in the introductory calculus, there are two
important topics: change of variables (change of coordinates) formula, and integral
theorems namely the Green, Gauss, Stokes theorems. One of the motivations of
exterior calculus is to work in a framework that the change of variables are automatic,
 32 Chapter 2. Exterior Calculus

and all the Stokes-like theorems are unified. Once we have such a framework, we
have a more unified way for talking about, for example, derivatives of changes of
variables (e.g. derivatives with respect to change of integration domain), which are
crucial in applications such as continuum mechanics and differentiable rendering.
Here is a quickly recap of these topics in multivariable calculus. We follow one
of the very first written textbooks on these topics, Maxwell’s 1873 “a Treatise on
Electricity and Magnetism, Vol. 1, Preliminary.” At the dawn of vector calculus,
Maxwell has made several remarks indicating that the natural structure to unify
multivariable calculus is to use the skew-symmetric structure of differential forms.
What exterior calculus is built upon is the linear algebra of these skew-symmetric
forms.

2.1.1 Line and surface integral

An important operation is the integration of the component of a vector field projected
along a line or a curve. This is called the circulation of a vector field along a curve
(not necessarily closed).
Let γ = (x1 , x2 , x3 ) : [0, S ] → R3 be a curve, and v = (v1 , v2 , v3 ) : R3 → R3 a
vector field. Then the circulation of the curve is given by
ˆ S
C = ⟨v, γ ′ ⟩ ds which can be expressed as
0
ˆ S
dx1 dx2 dx3
= v1 ds + v2 ds + v3 ds or simply
0 ds ds ds
ˆ S ˆ
= ⟨v, d γ⟩ = v1 dx1 + v2 dx2 + v3 dx3
0 γ([0,S ])
´
Here v1 dx1 + v2 dx2 + v3 dx3 is called a 1-form. This quantity C with integral γ([0,S ])
´b
generally depends on the entire path γ([0, S ]), as opposed to an integral a depending
only on the two end points a = γ(0), b = γ(S ). However, when within a certain
region the one-form takes the form of

v1 dx1 + v2 dx2 + v3 dx3 = d Ψ for some potential function Ψ

that is, is an exact differential within that region, the value of C becomes

C = Ψ(b) − Ψ(a)

and is the same for any two paths between a and b, provided the path can be changed
into the other by continuous motion without passing out of this region.1
1
The necessary condition for the 1-form v1 dx1 + v2 dx2 + v3 dx3 to be an exact differential of a
potential is
 
∂v3 ∂v2 ∂v1 ∂v3 ∂v2 ∂v1
curl v = − , − , − = (0, 0, 0).
∂x2 ∂x3 ∂x3 ∂x1 ∂x1 ∂x2

However, there are cases in which the condition curl v = 0 is fulfilled throughout a domain, and
yet the line integral C from a to b may be different for two curves. This may be the case if the
domain is multiply connected, i.e. there exists closed loop that cannot deform and shrink to a
 2.1 From Multivariable Calculus to Exterior Calculus 33

Another important operation that involves a surface integral is the flux of a vector
field through a surface. Let Σ be a surface, dA the surface area element,2 and n the
unit normal to the surface drawn towards the positive side of the surface. The flux
of v over the surface is given and expressed by
¨ ¨ ¨ ¨
⟨v, n⟩ dA = v1 n1 dA + v2 n2 dA + v3 n3 dA
Σ ¨ Σ ¨
Σ Σ¨
= v1 dx2 dx3 + v2 dx3 dx1 + v3 dx1 dx2 .
Σ Σ Σ

The quantity v1 dx2 dx3 + v2 dx3 dx1 + v3 dx1 dx2 is called a 2-form.
Here dx1 dx2 is the area element projected to the x1 x2 -plane. The flux of a
vector field through a surface is the sum of its contribution from all three projected
components, each of which is the area integral of the vector component normal to the
respective plane. It is important to note that these area integrals are signed integral
inherited from the orientation of the space. The normal of the dx2 dx3 -plane is x1 ,
the normal of the dx3 dx1 -plane is x2 , etc. In particular,
˜ when writing
˜ flux integral in
this form, the ordering of differentials matter v1 dx ˝2 3dx = − v 1 dx 3 dx2 .
Similarly, in this context, a volumetric integral h dx1 dx2 dx3 should have the
ordering of the differential dx1 dx2 dx3 positively oriented.
More generally, it turns out that the right structure for integration that evaluate
circulations, fluxes, etc., is infinitesimally a skew symmetric product of differentials
dxi1 · · · dxik . Later, in order not to be confused with the unsigned integrals, we shall
use the notation dxi1 ∧ · · · ∧ dxik to emphasize the skew symmetry nature. We will
spend 2.3 to study the linear algebra of these forms.

2.1.2 Changing variables for multivariable integrals

Suppose D is a domain of in the Cartesian space Rn with coordinates labeled
x1 , . . . , xn . Let F : D → Rn be a map into another Cartesian space with coordinates
labeled y1 , . . . , yn . This map parametrizes the image F (D) (as a signed multiset)
with yi = Fi (x1 , . . . , xn ) for i = 1, . . . , n. Let h : F (D) → R be a scalar function.
Then the integral
ˆ
h(y1 , . . . , yn ) dy1 · · · dyn
F (D)

can be written in terms of an integral over (x1 , . . . , xn ) ∈ D as

ˆ ˆ
h(y1 , . . . , yn ) dy1 · · · dyn = (h ◦ F )(x1 , . . . , xn ) J dx1 · · · dxn ,
F (D) D

point (nontrivial cycles), and if the two paths from a and b follow the opposite segments of such
nontrivial loops. In this case, the one path cannot be transformed into the other. Two cycles are
said to be equivalent if they together form the boundary of a surface embedded in the space. In a
multiply connected domain, every cycle (up to equivalence) is a multiple, or a linear combination,
of finitely many basis cycles.
If curl v = 0 and the circulations along those basis cycles vanish, then v1 dx +v2 dx2 +v3 dx3 = d Ψ.
2
Here dA does not mean the derivative d of a function A.
 34 Chapter 2. Exterior Calculus

where J is the Jacobian determinant:

 
∂F1/∂x1 · · · ∂F1/∂xn

J = det  ... ..  = det dF (e ), . . . , dF (e ) .

 
 ..
. .  1 n
∂Fn/∂x1 ··· ∂Fn/∂xn

Here e1 , . . . , en is the canonical basis for the Cartesian space.

In exterior calculus, we will simply write
ˆ ˆ
ω= F ∗ ω,
F (D) D

i.e., the integral of ω over the image F (D) of a map F is the same as the integral
of the pullback differential F ∗ ω via the map over the domain D. The notion of
pullback in exterior calculus will be the abstraction of the routine procedure of
changing variables for the integrand.

2.1.3 Green, Gauss and Kelvin–Stokes theorems

Another important topic of multivariable calculus includes theorems about integrals
of derivatives, which are higher dimensional versions of the fundamental theorem of
calculus.

• The surface integral of the flux through a closed surface may be expressed as
an integral over the enclosed volume.

• The circulation taken around a closed curve may be expressed in terms of a

surface integral taken over a surface subtended by the curve.

Let v : R2 → R2 be a vector field in the 2-dimensional Cartesian space. Let

D ⊂ R2 be a region, and let ∂D be its boundary curve. Then by integrating the
derivatives in each variable and noting the integration orientation to assign an
appropriate sign, we obtain Green’s Theorem
¨   ˛
∂v1 ∂v2
+ dx1 dx2 = v1 dx2 − v2 dx1 .
D ∂x1 ∂x2 ∂D

In general, in n-dimension, ni=1 ∂vi/∂xi is called the divergence of v. Its integral

P
over an n-dimensional region equals to the total (outward) flux through the boundary
surface. Take 3D for example. Suppose u : R3 → R3 is a vector field in 3D (with
coordinates labeled with y), D ⊂ R3 is a 3D domain, and ∂D is its boundary surface
with normal vectors denoted by n : ∂D → R3 . Then we have Gauss’ Divergence
Theorem
˚ X 3 ‹
∂ui
dy1 dy2 dy3 = u1 dy2 dy3 + u2 dy3 dy1 + u3 dy1 dy2
D i=1 ∂yi ∂D
‹
= ⟨v, n⟩ dA.
∂D
 2.2 An Overview of the Language 35

Let us come back to Green’s Theorem. Suppose : D ⊂ R2 → R3 is a parametrized

surface, whose restriction to the boundary f : ∂D → R3 is a parametrized space curve.
Let ⃗u : R3 → R3 be a vector field in 3D. Then the circulation of u along f (∂D) is
˛ ˛ 3 ˛ 3 X
2
X X ∂fi
⟨u, d f ⟩ = ui dyi = ui dxj
f (∂D) f (∂D) i=1 ∂D i=1 j =1 ∂xj
3 ¨     
Green’s Thm.
X ∂ ∂fi ∂ ∂fi
= ui − ui dx1 dx2
i=1 D ∂x2 ∂x1 ∂x1 ∂x2
3 X 3 ¨  
X ∂ui ∂fi ∂fj ∂ui ∂fi ∂fj
= − dx1 dx2
i=1 j =1 D ∂yj ∂x1 ∂x2 ∂yj ∂x2 ∂x1
3 X 3 ¨  
X ∂ui ∂uj ∂fi ∂fj
= − dx1 dx2
i=1 j =1 D ∂y j ∂y i ∂x 1 ∂x 2
¨ ¨
= det (curl u, d f (e1 ), d f (e2 )) dx1 dx2 = ⟨curl u, n⟩ dA,
D f (D)

which is the total flux of curl u = (∂u2/∂y3 − ∂u3/∂y2 , ∂u3/∂y1 − ∂u1/∂y3 , ∂u1/∂y2 − ∂u2/∂y1 )
through the surface f (D). This is the Kelvin–Stokes Theorem.
In exterior calculus, vector calculus operators such as grad (∇), curl (∇×), div
(∇·) all coalesce into a single derivative d . Similarly, important theorems such as
the Fundamental Theorem of Calculus, or Gauss’ and Green’s Theorem all become
instances of one single general statement, the so-called Stokes’ theorem
ˆ ˆ
dω = ω,
M ∂M

i.e., the integral of the differential of ω over the domain M is the same of ω over the
boundary of the domain.

2.2 An Overview of the Language

The main objects that exterior calculus deals with are differential forms. Functions,
or scalar fields, take a given point of the domain, and return a value. Extending this
idea, differential forms are to be evaluated (integrated) over given curves, surfaces,
or volumes, etc.
Specifically, let M be a 3-dimensional domain. On M , a k -form is a differential
form that is to be integrated over a k -dimensional (oriented) submanifold:
• A 3-form is to be integrated over a volumetric region. For example, the
mass density ρ of physical matter should be regarded as a 3-form, denoted by
´ ∈ Ω (M ). Given a region U ⊂ M , the total mass within U is denoted by
3
ρ
U
ρ.

• A 2-form is to be integrated over an oriented surface. For example, a flux ω of

a flow should be regarded as a 2-form, denoted by´ ω ∈ Ω2 (M ). It describes
each total flux over a given surface Σ, denoted by Σ ω.
36 Chapter 2. Exterior Calculus

• A 1-form is to be line integrated over a path. For example, a force field η

should be regarded as a 1-form, denoted by η ∈ ´Ω1 (M ). The total work done
by the force along a given path Γ is denoted by Γ η.
• A 0-form is a function, or a scalar field. For example, the temperature u of
a material is a 0-form, denoted by u ∈ Ω0 (M ). It is to be evaluated at each
point.
We will encounter only 5 algebraic and differential operators for differential forms:
• Pullback operator. If we have a differential form α ∈ Ωk (W ), and a map
f : M → W , then we have a pullback differential form f ∗ α ∈ Ωk (M ). It has a
natural definition
´ that
´ its evaluation over a k -dimensional surface Σ ⊂ M is
given by Σ (f ∗ α) = f (Σ) α.

• Exterior derivative. Denoted by d , the exterior derivative is a differential

operator that sends a k -form α to a (k + 1)-form d α. The
´ evaluation
´ of d α
over a (k + 1)-dimensional surface Σ is designed so that Σ d α = ∂Σ α.
• Wedge product. It is an algebraic operator that produces higher-degree forms.
If α ∈ Ωk (M ) and β ∈ Ωℓ (M ) then their wedge product α ∧ β ∈ Ωk +ℓ (M ).
• Interior product. Provided a given direction, the interior product as an algebraic
operator lowers the degree of a form. Suppose X is a tangent vector of M , and
α is a k -form, then the interior product iX α is a (k − 1)-form.
• Hodge star. Denoted by ⋆, it is an algebraic operator that turns a k -form α
into an (n − k )-form ⋆α, where n = dim(M ). Among these 5 operators, this
Hodge star operator is the only one that requires a Riemannian metric for M .
Macroscopic and Microscopic View
A differential 1-form η is a quantity that is waiting to be integrated over a curve
γ. This is the “macroscopic” viewpoint for differential forms. Let us take a short
parametrized curve γ, which is so short that γ is well-described by a point x = γ(0),
a tangent vector v = γ ′ (0) at x as the velocity for the curve, and a small ε as the
length of the interval of the parameter for the curve. Abusing
´ the notation slightly,
we write γ = εv. Then with continuity in η, we expect εv η ∼ O(ε). The limit
ˆ
1
lim η =: η(v)
ε→0 ε εv

is a quantity depending linearly on the velocity vector v. We say η(v) is the evaluation
of the 1-form on a tangent vector v. This is the algebraic, or the “microscopic,”
viewpoint of differential forms.
Conversely, given a 1-form η as a scalar-valued linear function on tangent vectors,
we can understand its “macroscopic” counterpart as follows. The integral of η over a
curve γ can be obtained by first partitioning the parameter interval of γ and then
taking the sum of microscopic evaluation
ˆ X
η γ ′ (ti ) (ti+1 − ti ).


η := lim
γ partion
refines i
 2.3 Differential Forms 37

Similarly, a differential 2-form ω in a “macroscopic” view is a quantity to be

evaluated over an oriented surface. Its “microscopic” version is a function linearly
evaluating on an infinitesimal oriented parallelogram spanned by two vectors v1 , v2 .
More precisely, ω(v1 , v2 ) is bilinear (linear in each component if fixing the other)
and skew symmetric ω(v1 , v2 ) = −ω(v2 , v1 ). To assemble a microscopic 2-form into
its integral over a surface, partition the surface into infinitesimal parallelograms and
sum the values of the bilinear forms.
As mentioned in Section 2.1, it turns out that the right structure for integration
that evaluate circulations, fluxes, etc., is infinitesimally a skew symmetric multilinear
form. In Section 2.3 we study the linear algebra of these forms.
Starting from Section 2.4 we will expand on related concepts about vector fields.
This will lead up to an important operation called Lie derivative in Section 2.6. This
operation is essential in many applications such as continuum mechanics that is often
overlooked in multivariable calculus. Finally, we introduce Hodge star in Section 2.7.

2.3 Differential Forms

Algebraically, differential forms are skew symmetric multilinear form. With such
an object defined over a domain, it becomes a mathematical object that can be
integrated over a submanifold (Section 2.3.1). In Section 2.3.2, we introduce how
differential forms are made from covectors by the theory of extension. Using the
same extension technique, we extend the notion of “the differential of a function”
to general differential forms (Section 2.3.3). Finally, in Section 2.3.4, we give a
geometric interpretation of differential forms and the operations among them.
Definition 2.1 Let M be a manifold. A k -form ω is a skew symmetric k -linear form
field; that is, for each p ∈ M , we have a function linear in each of its k arguments
multilinear
ωp J·, . . . , ·K : Tp M × · · · × Tp M −−−−−−→ R (2.1)
| {z }
k

satisfying

ωp JX1 , . . . , Xi , . . . , Xj , . . . , Xk K = −ωp JX1 , . . . , Xj , . . . , Xi , . . . , Xk K. (2.2)

Such a skew-symmetric k -linear form at p is collectively

ωp ∈ k Tp∗ M . (2.3)
V

The space of k -forms is denoted by Ωk (M ) = Γ(

Vk
T ∗ M ).a
a
The motivation for the notation “∧k T ∗ M ” will become clear in Section 2.3.2.

Space Ω0 (M ) of 0-forms are just space of functions Ω0 (M ) = {f : M → R}. Space

Ω1 (M ) of 1-forms are the space of covector fields.
Vk For each p
in an n-dimensional manifold M , the dimension of the vector space
Tp M is nk . Every k -form would be 0 if k > n.
∗
 38 Chapter 2. Exterior Calculus

■ Example 2.1 In the 3D Cartesian space, each vector written in coordinates

a = (a1 , a2 , a3 )⊺ can be converted into a 1-form (a)1-form or a 2-form (a)2-form ,
whose evaluation on a vector u or a pair of vectors u, v are defined by

(a)1-form JuK = a1 u1 + a2 u2 + a3 u3 (2.4)

 
a1 u1 v1
(a)2-form Ju, vK = det a2 u2 v2  . (2.5)
a3 u3 v3

A scalar function a, which is a 0-form a = (a)0-form , can be converted into a 3-form

(a)3-form
 
u 1 v 1 w1
(a)3-form Ju, v, wK = a det u2 v2 w2  . (2.6)
u 3 v 3 w3
■

2.3.1 Integration and pullback

Differential forms can be treated as objects to-be-integrated along a k -dimensional

surface. Suppose S : Dk → M is a parametric k -dimensional surface defined over
a k -dimensional parameter space Dk with values in M . Recall that dS = S∗
is its pushforward map. Let θ1 , . . . , θk , θi : Dk → R, be an arbitrary coordinate
system. That is, we have d θ1 , . . . , d θk as basis covector field, whose dual basis
⃗e1 , . . . , ⃗ek ∈ Γ(T Dk ) is a coordinate vector field.
´
Definition 2.2 Let ω ∈ Ωk (M ) be a k -form on M . The integral S ω of ω over a
surface S described above is defined by
ˆ ¨ ˆ 
ω := ··· ωS (θ1 ,...,θk ) JS∗⃗e1 , . . . , S∗⃗ek K d θ1 · · · d θk . (2.7)
S Dk

This definition is independent of reparametrization of S .

■ Example 2.2 In the 3D Cartesian space, for a scalar function a and a vector
field a we have
ˆ
a = a(p), (2.8)
ˆ p
ˆ
(a)1-form = a · d l, (2.9)
ˆC ¨C
(a)2-form = a · ndA, (2.10)
ˆS ˚S

(a)3-form = a dV . (2.11)
V V

■
 2.3 Differential Forms 39

Definition 2.3 For each smooth map ϕ : M → N , we define the pullback operator

linear
ϕ∗ : Ωk (N ) −−−→ Ωk (M ) (2.12)

that for each ω ∈ Ωk (N ),

(ϕ∗ ω)p JX1 , . . . , Xk K := ωϕ(p) Jϕ∗ X1 , . . . , ϕ∗ Xk K. (2.13)

Theorem 2.1 — Pullback as change of integration variables.

ˆ ˆ
ω= ϕ∗ ω. (2.14)
ϕ(S ) S

■Example 2.3 Let ϕ : R3 → R3 be a smooth map between the 3D Cartesian space.

Let F = ∇ϕ, denoted as a matrix Fij = ∂ϕ
∂xj
i
called the deformation gradient. Let
J = det(F). Then

ϕ∗ (a)0-form = (a ◦ ϕ)0-form , (2.15)

0-form
∗
ϕ (a)1-form = (F⊺ a ◦ ϕ)1-form , (2.16)
1-form
∗
ϕ (a)2-form = (J F−1 a ◦ ϕ)2-form , (2.17)
2-form
∗
ϕ (a)3-form = (Ja ◦ ϕ)3-form . (2.18)
3-form

2.3.2 Wedge product and interior product

In Definition 2.1, we described what a differential form is. At each point, a differential
form is a skew symmetric k -linear form. If we have such an object, we can talk about
its integration and the associated changes of variables (pullback) (Section 2.3.1).
But how do we obtain these skew symmetric k -linear form in the first place? Some
instances of differential forms are concretely given, but only under a 3D Cartesian
coordinate system (Example 2.1). Is there a way to systematically construct skew
symmetric k -linear forms?
Well, we do have a starting point. Covectors are 1-forms. They are canonically
defined by taking the dual space of a vector space. By the theory of extension
introduced by Grassmann (1862), we can extend the elementary covectors into k
forms through a new multiplication called wedge product. We can also extend the
elementary vector–covector dual pairing to pairing k -forms and k vectors, making
k -forms truly objects satisfying Definition 2.1. The extension of dual pairing to
k -forms is called the interior product.
In fact, using the same technique of extension, we can extend operators other
than the vector–covector dual pairing. For example, at the level between functions
and covector fields, we have the notion of taking differentials (the differential of a
function is a covector field). We can extend such a differentiation notion, and obtain
40 Chapter 2. Exterior Calculus

exterior derivative (Section 2.3.3).

The following Definitions 2.4, 2.5 and 2.6 collectively serve as an alternative ax-
iomatic definition for differential forms that can replace Definition 2.1.
Definition 2.4 — Wedge product. Let V be a vector space, and let V ∗ be its dual
space. Define
V•
(2.19)
V2 V3
V∗ = R⊕V∗ ⊕( V ∗) ⊕ ( V ∗) ⊕ · · ·

as the free algebraa generated by a new multiplication called wedge product

bilinear
(2.20)
Vk Vℓ Vk +ℓ
∧: V∗ × V ∗ −−−−→ V∗

that satisfies the following two rules:

• Associativity: (α ∧ β) ∧ γ = α ∧ (β ∧ γ); (2.21)

• Anti-commutativity: α ∧ α = 0 for 1-forms α ∈ V ∗ . (2.22)

Here, k V ∗ in (2.19) is the collection of all linear combinations of degree-k wedge

V
products of covectors.
a
The space of all polynomial expressions but without commutativity on multiplication.

Definition 2.5 — Interior product. For each vector X ∈ V in a vector space V ,

define a linear map
linear
iX : ∧k V ∗ −−−→ ∧k −1 V ∗ (2.23)

that satisfies

• Base definition: iX α = ⟨α|X ⟩ for all 1-forms α ∈ V ∗ ; (2.24)

• Nilpotent: iX iX = 0; (2.25)

• Graded Leibniz rule: iX (α ∧ β) = (iX α) ∧ β + (−1)deg(α) α ∧ (iX β). (2.26)

Definition 2.6 — Evaluation of k -forms. For ω ∈ ∧k V ∗ and X1 , . . . , Xk ∈ V , define

ωJX1 , . . . , Xk K := iXk · · · iX2 iX1 ω. (2.27)

In particular,
(iX ω)JY1 , . . . , Yk −1 K = ωJX , Y1 , . . . , Yk −1 K. (2.28)
The wedge product and interior product generalize to the pointwise wedge product
and interior product for the space of k -form fields Ωk (M ) := Γ(∧k T ∗ M )
bilinear
∧ : Ωk (M ) × Ωℓ (M ) −−−−→ Ωk +ℓ (M ), (α ∧ β)p := αp ∧ βp (2.29)
linear
iX : Ωk (M ) −−−→ Ωk −1 (M ), (iX ω)p := iXp ωp , (2.30)
2.3 Differential Forms 41

where X ∈ Γ(TM ) is a vector field.

Theorem 2.2 — Graded anti-commutativity.

α ∧ β = (−1)k ℓ β ∧ α (2.31)

when α is a k -form and β is an ℓ-form.

Proof. First, we have

α ∧ β = −β ∧ α when k = ℓ = 1 (2.32)

as a consequence of (2.22):

0 = (α + β) ∧ (α + β) = α ∧ α} +α ∧ β + β ∧ α + β ∧ β .
| {z (2.33)
| {z }
0 0

For general k -forms and ℓ-forms, write α, β as linear combinations of ∧ of 1-forms,

and reorder α ∧ β into β ∧ α by applying the swaps (2.32) k ℓ number of times.

Theorem 2.3 For α1 , . . . , αk ∈ Ω1 (M ),

 
⟨α1 |X1 ⟩ · · · ⟨α1 |Xk ⟩
(α1 ∧ · · · ∧ αk )JX1 , . . . , Xk K = det  ... .. ..
. (2.34)
 
. .
⟨αk |X1 ⟩ · · · ⟨αk |Xk ⟩

Proof. Using the definition (2.27) P

and repeatedly Q apply the Leibniz rule (2.26)
and (2.24), we obtain an expression σ∈Sk (−1)|σ| ki=1 ⟨αi |Xσ(i) ⟩, which is Leibniz’s
formula for the determinant of the matrix ⟨αi |Xj ⟩.
From Theorem 2.3, it is apparent that the k -forms generated by k forms are
precisely the ones described in Definition 2.1.
■ Example 2.4 In the 3D Cartesian space R3 with covector basis dx , dy, dz ∈
Ω1 (R3 ), the conversion (Example 2.1) from Cartesian vectors/scalars to forms are
explicitly given by

(a)0-form = a, (2.35)
(a)1-form = a1 dx + a2 dy + a3 dz , (2.36)
(a)2-form = a1 dy ∧ dz + a2 dz ∧ dx + a3 dx ∧ dy (2.37)
= i(a)vec (dx ∧ dy ∧ dz ), (2.38)
(a)3-form = a dx ∧ dy ∧ dz . (2.39)

One can check that this is consistent with Example 2.1 using Theorem 2.3. ■
42 Chapter 2. Exterior Calculus

■ Example 2.5 In the 3D Cartesian space,

(a)0-form ∧ (b)0-form = (ab)0-form , (2.40)

(a)0-form ∧ (b)1-form = (ab)1-form , (2.41)
(a)0-form ∧ (b)2-form = (ab)2-form , (2.42)
(a)0-form ∧ (b)3-form = (ab)3-form , (2.43)
(a)1-form ∧ (b)1-form = (a × b)2-form , (2.44)
(a)1-form ∧ (b)2-form = (a · b)3-form . (2.45)

■ Example 2.6 In the 3D Cartesian space,

iv (a)1-form = (a · v)0-form , (2.46)

iv (a)2-form = (a × v)1-form , (2.47)
iv (a)3-form = (av)2-form . (2.48)

■ Example 2.7 — Vector algebra identities. The following 3D vector identities

are special cases of the Leibniz rule (2.26) for the interior product written out in
3D vector algebra.
Theorem 2.4 — Vector triple product. a × (b × c) = (a · c)b − (a · b)c.

(2.47) (2.44) (2.26)

Proof. (a × (b × c))1-form = −ia (b × c)2-form = −ia ((b)1-form ∧ (c)1-form ) =
(2.46)
−ia (b)1-form (c)1-form + (b)1-form ia (c)1-form = (−(a · b)c + (a · c)b)1-form .

Theorem 2.5 — Cauchy–Binet identity. (a×b)·(c×d) = (a·c)(b·d)−(a·d)(b·c).

(2.47)
(2.45) (2.44)
Proof. [(a × b) · (c × d)]3-form = (a × b)1-form ∧ (c × d)2-form = −ia (b)2-form ∧
(2.26)
(c)1-form ∧ (d)1-form = (b)2-form ∧ ia (c)1-form ∧ (d)1-form − (b)2-form ∧ (c)1-form ∧
(2.46)
(2.45)
ia (d)1-form = (a · c)(b · d)3-form − (a · d)(b · c)3-form .

Theorem 2.6 — Vector quadruple product. (a × b) × (c × d) = det(a, c, d)b −

det(b, c, d)a.

(2.44) (2.47)
Proof. [(a × b) × (c × d)]2-form = (a × b)1-form ∧ (c × d)1-form = ia (b)2-form ∧
(2.47),(2.45)
(2.26) (2.6)
ic (d)2-form = ia [(b)2-form ∧ ic (d)2-form ] − (b)2-form ia ic (d)2-form = ia (b · (d ×
(2.48)
c))3-form − (b)2-form det(d, c, a) = (a)2-form det(b, d, c) − (b)2-form det(d, c, a).
■
2.3 Differential Forms 43

Interaction with pullbacks

The pullback operator Definition 2.3 preserves the wedge product and interior product.
This is because Definition 2.3 is consistent to the pullback operator for covectors
(Definition 1.17 and Definition 1.4), and the rest of the extension construction is
canonical (Definition 2.4–2.6).
Theorem 2.7 — Pullback distributes over wedge products.

ϕ∗ (α ∧ β) = (ϕ∗ α) ∧ (ϕ∗ β). (2.49)

Theorem 2.8 — Passing pullback through interior product.

ϕ∗ (iϕ∗ (X ) ω) = iX (ϕ∗ ω). (2.50)

■ Example 2.8 — Matrix vector identities. We can write Theorem 2.7 and The-
orem 2.8 in the 3D Cartesian space using the correspondence given by Exam-
ple 2.3.
Theorem 2.9 Let u, v ∈ R3 , F ∈ R3×3 , and J = det(F). Then

J F−1 (u × v) = (F⊺ u) × (F⊺ v). (2.51)

F⊺ (u × (Fv)) = (J F−1 u) × v. (2.52)

Proof. (2.51) is (2.49) when both α and β are 1-forms. (2.52) is (2.50) when ω is
a 2-form.
■

Integral picture for interior product

Differential k -forms can be evaluated by inserting k vectors (Definition 2.6). Differ-
ential k -forms cam also be integrated over k -dimensional surface (Section 2.3.1). In
the former algebraic picture, the interior product is the “insertion” operation (2.28).
For the latter integral picture, it turns out that the interior product is dual to the
“extrusion” operation, stated as follows.
Suppose S : Dk → M is a (k −1)-dimensional surface. Let X ∈ Γ(TM ) be a vector
field. Consider extruding the (k − 1)-dimensional surface S into a k -dimensional
surface extϵX S along X . Concretely, it is the solution to the initial value problem of
flowing along X :
(extϵX S ) : [0, ϵ] × Dk → M , (2.53)
∂(extϵX S )
(t, θ1 , . . . , θk −1 ) = X |(extϵX S )(t,θ1 ,...,θk −1 ) (2.54)
∂t
(extϵX S )(0, θ1 , . . . , θk −1 ) = S (θ1 , . . . , θk −1 ). (2.55)
This geometric operation of extrusion is illustrated in Figure 2.1. The following
theorem states that the interior product using a vector field X measures the rate
of change of measurement when the integration domain is an extrusion along the
vector field.
 44 Chapter 2. Exterior Calculus

extϵX S

Figure 2.1 The extrusion of a (k − 1)-dimensional surface S along a vector field

X (left) yields a k -dimensional surface (right).

Theorem 2.10 — Extrusion and interior product.

ˆ ˆ
d
ω= iX ω. (2.56)
dϵ ϵ=0 extϵX S S

2.3.3 Exterior derivative

Taking the differential of a function produces a covector field (Definition 1.6). Using
a similar extension rule as in Definition 2.5, we obtain:
linear
Definition 2.7 — Exterior derivative. Define d : Ωk (M ) −−−→ Ωk +1 (M ) satisfying

• Base case: df is the differential of f for any 0-form f ; (2.57)

• Nilpotent: d ◦ d = 0; (2.58)

• Graded Leibniz rule: d (α ∧ β) = (d α) ∧ β + (−1)deg(α) α ∧ (d β). (2.59)

■ Example 2.9 In the 3D Cartesian space,

d (f )0-form = (∇f )1-form , (2.60)

d (v)1-form = (∇ × v)2-form , (2.61)
d (v)2-form = (∇ · v)3-form . (2.62)

Next, we state two important theorems about the exterior derivative. The first is
that exterior derivative and pullback commute. This is a consequence of the same
theorem for the base case (Theorem 1.12), and that its extension to the general
exterior derivatives (Definition 2.7) is defined using its interaction with ∧, which is
conserved under pullback (Theorem 2.7).
Theorem 2.11 — Exterior derivative and pullback commute.

ϕ∗ (d α) = d (ϕ∗ α). (2.63)

2.3 Differential Forms 45

S ∂S

Figure 2.2 The boundary of a k -dimensional surface S (left) is a (k − 1)-

dimensional surface (right).

The second important theorem is the Stokes theorem. Similar to how the interior
product is dual to extrusion (Theorem 2.10), the exterior derivative is dual to the
boundary operation on the integration domain (Figure 2.2).
Theorem 2.12 — Stokes Theorem.
ˆ ˆ
dα = α. (2.64)
S ∂S

Proof. Let the k -dimensional surface S be parameterized by ϕ : Ω ⊂ Rk → M .

By Theorem 2.1 and Theorem 2.11, the Stokes theorem reduces to the divergence
theorem on the k -dimensional region Ω ⊂ Rk .

■ Example 2.10 — Nilpotence (2.58) in 3D.

(∇ × )(∇) = 0, (2.65)
(∇ · )(∇ × ) = 0. (2.66)

■ Example 2.11 — Leibniz rule (2.59) in 3D. Let f : R3 → R, u, v : R3 → R3 be

function and vector fields. The Leibniz rule (2.59) becomes the following vector
calculus identities. When it is applied to a 0-form and a 1-form, we get
Theorem 2.13 ∇ × (f u) = (∇f ) × u + f (∇ × u).
For 0-form and 2-form:
Theorem 2.14 ∇ · (f u) = (∇f ) · u + f (∇ · u).
For 1-form and 1-form:
Theorem 2.15 ∇ · (u × v) = (∇ × u) · v − u · (∇ × v).
■

■Example 2.12 — Theorem 2.11 in 3D. In the 3D Cartesian space, Theorem 2.11
becomes the following vector identities. Let ϕ : R3 → R3 , F = ∇ϕ, and J =
46 Chapter 2. Exterior Calculus
det(F). Recall Example 2.3 for the pullback formulas in 3D. Theorem 2.11 applied
to 0-forms yields
Theorem 2.16 — Chain rule revisited. Let f : R3 → R. Then,

F⊺ (∇f ◦ ϕ) = ∇(f ◦ ϕ) (2.67)

Theorem 2.11 applied to 1-forms yields
Theorem 2.17 — Change of variables for curl. Let u : R3 → R3 . Then,

J F−1 ((∇ × u) ◦ ϕ) = ∇ × (F⊺ u ◦ ϕ). (2.68)

Theorem 2.11 applied to 2-forms yields
Theorem 2.18 — Piola identity. Let u : R3 → R3 . Then,

J ((∇ · u) ◦ ϕ) = ∇ · (J F−1 u ◦ ϕ). (2.69)

This example really shows the power of Theorem 2.11. The simple concept of
commutativity Theorem 2.11 becomes rather opaque vector identities especially
for the Piola identity (Theorem 2.18) and its counterpart for curl (Theorem 2.17)
which is in fact too complicated to be discovered with a name in vector calculus.
The Piola identity also implies the following identity [Evans(1998), Sec 8.1].
Specialize u into a constant coordinate vector, and note that the cofactor matrix
of F is given by cof(F) = J F−⊺ .
Theorem 2.19 The cofactor matrix of the Jacobian matrix of any map has
divergence-free rows.
■

■Example 2.13 — Theorem 2.12 in 3D. In the 3D Cartesian space, Theorem 2.12
becomes the following theorems. Theorem 2.12 applied to 0-forms becomes

Theorem 2.20 — Fundamental theorem of calculus. Let f : R3 → R, and C

be a curve connecting a ∈ R3 to b ∈ R3 . Then,
ˆ
∇f · d l = f (b) − f (a). (2.70)
C
Theorem 2.12 applied to 1-forms becomes
Theorem 2.21 — Kelvin–Stokes curl theorem. Let u : R3 → R, and S be a
surface in R3 with normal vector n. Then,
¨ ˛
(∇ × u) · ndA = u · d l. (2.71)
S ∂S
Theorem 2.12 applied to 2-forms becomes
 2.3 Differential Forms 47

ˆ
amount of
α=
S intersections
α

Figure 2.3 The integration of a k -form α along a k -dimensional surface S mea-

sures the total amount of intersections between S and the codimension-
k geometries representing α.

Theorem 2.22 — Gauss’ divergence theorem. Let u : R3 → R, and V be a

volumetric region in R3 with boundary normal n. Then,
˚ ‹
(∇ · u)dV = u · ndA. (2.72)
V ∂V
■

2.3.4 Differential forms as codimensional geometries

In this last part of Section 2.3, we provide visual representations for differential
forms, wedge products, interior products, and exterior derivatives. These geometric
interpretations are given in the appendix of [Yin et al.(2023)].
Geometrically, a vector field X is an assignment of an infinitesimal “arrow”
Xp ∈ Tp M at every point p ∈ M , whose directions and magnitudes depict some
“instantaneous flow velocity” within the domain M . A k -form α ∈ Ωk (M ), on the
other hand, is a distribution of (n − k )-dimensional (codimension-k ) oriented planes
over M . See Figure 2.4. For example, in 3D, a 3-form is illustrated as a point
cloud, a 2-form is a line segment cloud, a 1-form is a plane field, and a 0-form is a
superposition of sublevel sets of the corresponding scalar function.

Integration is counting signed intersection

The orientations´ and densities of the codimension-k plane cloud are given so that
the integration S α over a test k -surface S is the total signed intersection between
S and the codimensional-k plane cloud (Figure 2.3).3

Pullback is taking preimage

The pullback ϕ∗ α of a k -form α ∈ Ωk (M ) has its codimension cloud given by the
preimage of the codimension cloud of α through ϕ.

3
For n = 2, 3, every k -form at each tangent space admits a distinguished (n − k )-subspace that
represents the orientation of the plane field. For n > 3 and 1 < k < n − 1, the codimension-k plane
fields are generally no longer described by a distinguished oriented subspace but a superposition of
many. This is related to the decomposability of forms (Definition 4.3).
48 Chapter 2. Exterior Calculus

0-form 1-form 2-form 3-form

Figure 2.4 Differential k -forms can be represented by clouds of codimension-k
geometries.

α X
α iX α
β

α∧β

Figure 2.5 Wedge products are intersections (left) and interior products are
extrusions (right).

Wedge product is taking intersection

The wedge product α ∧ β of a k -form α ∈ Ωk (M ) and an ℓ-form β ∈ Ωℓ (M ) is a
(k + ℓ)-form whose codimension cloud is given by the intersection of the codimension
clouds of α and β (Figure 2.5, left).

Interior products are extrusions

The interior product iX α of a k -form α ∈ Ωk with a vector field X ∈ Γ(TM ) is a
(k − 1)-form whose codimension cloud is given by the extrusion of the codimension
cloud of α along X (Figure 2.5, right.)

Exterior derivatives are boundary operators

The exterior derivative operator d : Ωk (M ) → Ωk +1 (M ) takes the boundary of the
codimensional geometric representation. For each α ∈ Ωk (M ), its exterior derivative
d α ∈ Ωk +1 (M ) has the (n −k −1) dimensional cloud elements given by the boundaries
of
´ the (n − ¸ k )-dimensional cloud elements of α (Figure 2.6). The Stokes Theorem
S
d α = ∂S α can be interpreted as the invariant of intersection number when
swapping the boundary operation. A k -form α ∈ Ωk (M ) is closed if d α = 0. For
a closed form, the boundaries of each instance of the codimensional cloud cancel
out with the neighboring boundaries. As a consequence, the codimensional cloud
of a closed form stitches together into global pieces of (n − k )-dimensional surfaces
foliating the space (Figure 2.7).
 2.4 Vector Fields as Derivations 49

d
−−−−→

α dα

Figure 2.6 Taking exterior derivative is taking boundary.

Non-closed k -form Closed k -form

Figure 2.7 A k -form is closed when the boundaries of its codimension-k geome-
tries cancel out.

2.4 Vector Fields as Derivations

Now, we will take a pause on studying the operations on differential forms, and
expand on the idea of directional derivatives of functions. Directional derivatives
are the starting point of multivariable calculus. It turns out that the collection of
directional derivative operations form a space that is naturally endowed with an
algebra that encodes important information about the direction fields themselves.
This algebra is called a Lie algebra. Such an algebra is used in mechanics and robotics
for distinguishing constrained systems (holonomic constraints or non-holonomic
constraints).

2.4.1 Algebra and derivations

Definition 2.8 — Algebra. A vector space A is called an algebra if it is equipped
bilinear
with a bilinear function × : A × A −−−−→ A.

■ Example 2.14 The space Rn×n of n-by-n matrices is an algebra, with matrix-
matrix multiplication as the bilinear function. ■

■ Example 2.15 The space Ω0 (M ) of functions is an algebra, with pointwise

multiplication as the bilinear function. ■

Note that the multiplication structure (bilinear function) for an algebra does not
need to be associative:
 50 Chapter 2. Exterior Calculus

■ Example 2.16 (R3 , ×) with cross product × is an algebra. Note that u×(v×w) ̸=
(u × v) × w in general. ■

linear
Definition 2.9 — Derivations. Let (A, ×) be an algebra. A linear map X : A −−−→
A is called a derivation if

X (a × b) = (Xa) × b + a × (Xb) a, b ∈ A (2.73)

The space of all derivations over an algebra is denoted by Der(A).

Definition 2.10 — View vector fields as derivations. Each vector field X ∈ Γ(TM )
is also viewed as an operator on functions X : Ω0 (M ) → Ω0 (M ) defined by

Xf := df JX K. (2.74)

Note that this functional operator is a derivation on functions

X (fg) = (Xf )g + f (Xg). (2.75)

In fact, a converse statement is true: every derivation on function is a vector field :

Γ(TM ) = Der(Ω0 (M )). (2.76)

This equivalence between vector fields and the algebraic definition of derivations
on functions motivates many differential geometry textbooks to adopt (2.76) as the
definition for vector fields.
The takeaway for this abstraction is that we can think of each vector field as the
directional derivative operator (2.74) that obeys the derivation property (2.75).

2.4.2 Lie algebra for vector fields

Definition 2.11 — Lie algebra. An algebra (A, [·, ·]) is called a Lie algebra if the
bilinear
bilinear function [·, ·] : A × A −−−−→ A satisfies skew symmetry

[a, b] = −[b, a] (2.77)

and the Jacobi identity

[a, [b, c]] + [b, [c, a]] + [c, [a, b]] = 0. (2.78)

Here are two elementary examples for Lie algebra.

■ Example 2.17 (R3 , ×) with the 3D cross product × is a Lie algebra. ■

■ Example 2.18 (Rn×n , [·, ·]) is a Lie algebra, where the bracket operation is the
matrix commutator [A, B ] = AB − BA. ■
 2.4 Vector Fields as Derivations 51

Theorem 2.23 The space of vector fields Γ(TM ) is a Lie algebra, with Lie bracket
[X , Y ] of vector fields X , Y defined by

[X , Y ]f := X (Yf ) − Y (Xf ) (2.79)

when it is operated on functions.

Proof. We only need to verify two properties. First, [X , Y ] := X ◦ Y − Y ◦ X is

still a derivation on functions, and hence [·, ·] is a valid algebra operation. Second,
[X , Y ] satisfies the Jacobi identity. Each of these properties is easy to check straight
from definitions.

R Space of tangent vector fields, or equivalently the space of derivations on

functions, is naturally equipped with a Lie algebra structure.

2.4.3 Practical calculation of Lie bracket

Suppose M is a subset of Rn with coordinate system x 1 , . . . , x n . Let X , Y be two
vector fields. As derivations, we represent them as directional differential operators
n n
X ∂ X ∂
X = Xi , Y = Yi (2.80)
i=1
∂x i i=i
∂x i

where each components X i , Y i of the vector fields are functions defined Pover M . Let
Z = [X , Y ]. Our goal is to determine its component Z as in Z = i=1 Z i ∂x∂ i in
i n

terms of the components of X i , Y i . Expanding the definition of Lie bracket,

   
i ∂ j ∂f i ∂ j ∂f
X
Zf = X (Yf ) − Y (Xf ) = X i
Y j
−Y i
X j
(2.81)
i,j
∂x ∂x ∂x ∂x
X  ∂Y j j


∂ 2f
i ∂X ∂f
= X i
− Y + X i
Y j
− Y i j
X . (2.82)
i,j
∂x i ∂x i ∂x j | {z ∂x i ∂x j}
=0

Therefore,
∂Y j i ∂X
j
Zj = Xi − Y . (2.83)
∂x i ∂x i
If “∇” denotes taking componentwise differential, we can write

Z = [X , Y ] = ∇X Y − ∇Y X . (2.84)

Note that each of the coordinate-componentwise differential ∇X Y or ∇Y X depends

on the choice of coordinate. But the difference ∇X Y − ∇Y X becomes [X , Y ] which
is coordinate independent.
In practice, to compute the Lie bracket of two vector fields, one can choose any
coordinate system (which doesn’t need to be orthonormal) and take the mutual
componentwise derivative (2.84).
 52 Chapter 2. Exterior Calculus

2.4.4 Integrability

Why do we care about the Lie algebra structure for vector fields? One of the most
important application of the Lie structure for vector fields is the integrability to
coordinate system.

Theorem 2.24 Suppose x1 , . . . , xn ∈ Ω0 (M ) be a coordinate system, which gives

rise to covector basis field dx1 , . . . , dxn ∈ Ω1 (M ). Let ⃗e1 , . . . , ⃗en ∈ Γ(TM ) be the
dual basis from dx1 , . . . , dxn ∈ Ω1 (M ), called coordinate vectors. Then [⃗ei , ⃗ej ] = 0.
Conversely, if a basis vector field X1 , . . . , Xn ∈ Γ(TM ) satisfies [Xi , Xj ] = 0 for
each pair of i , j = 1, . . . , n. Then they are locally the coordinate vectors of some
coordinate system.

Definition 2.12 — Distribution. A distribution A is an assignment of a linear

subspace Ap ⊂ Tp M of the tangent space at each point p ∈ M . We say a vector
field X ∈ Γ(TM ) belongs to the distribution, denoted by X ∈ Γ(A), if Xp ∈ Ap
for all p ∈ M .

A distribution can be seen as a plane field defined over the space M . Like Figure 2.7,
there may be two cases: a general plane field may not stitch together with its
neighbor, or it may stitch together and form a foliation of submanifolds. In the
former case, we say the distribution is non-integrable. In the latter case, we say that
the distribution is integrable.
Definition 2.13 A distribution A ⊂ TM is said to be integrable if the plane field
associated to A is locally the tangent plane of a family of submanifolds in M .

Theorem 2.25 — Frobenius theorem of integrability. A distribution A ⊂ TM is

integrable if and only if Γ(A) ⊂ Γ(TM ) is not only a vector subspace but also a
Lie subalgebra; that is, the Lie bracket of vector fields in the distribution stays in
the distribution: [X , Y ] ∈ Γ(A) for all X , Y ∈ Γ(A).

This notion of integrability plays a crucial role in characterizing robotics and

mechanical constructs. Such systems are parameterized through general coordinates
like joint angles. Often, these parameters are subject to limitations concerning their
infinitesimal mobility.
The range of feasible infinitesimal movements constitutes a distribution over the
tangent bundle of the parameter space. If this distribution proves to be integrable,
the constraint is referred to as a holonomic constraint. Otherwise, if the distribution
is non-integrable, it is termed a non-holonomic constraint.
A system with holonomic constraints can be simplified to a lower-dimensional
system, which is the submanifold integrated from the distribution. On the other
hand, a system with non-holonomic constraints allows for certain maneuvers within
the realm of feasible movements, yet these can lead to directions transversal to the
distribution. Essentially, non-holonomic constraints are velocity restrictions that do
not sufficiently confine the positions to lower-dimensional spaces.
2.4 Vector Fields as Derivations 53

■ Example 2.19 — Rolling ball. A rolling ball moving on a surface without slipping
presents an example of a nonholonomically constrained system. In this case, the
constraint arises from the no-slip condition, which dictates that the point of the
ball in contact with the surface has zero velocity relative to the surface. Under this
constraint, the 3D orientation of the ball have only 2 degrees of freedom among
the 3-dimensional space of all 3D rotations. However, moving in an infinitesimal
cycle using the 2 degrees of freedom can yield a net motion into the 3rd degree of
freedom.
Let M = SO(3) be the space of 3D rotations. Consider a basis of vector fields

XR = [e1 ×]R, YR = [e2 ×]R, ZR = [e3 ×]R ∈ TR SO(3) (2.85)

ha i h 0 −c b i
at each R ∈ SO(3) ⊂ R3×3 . Here, [ bc ×] = c 0 −a is the matrix represen-
−b a 0
tation of cross product. The rolling-without-slipping condition means that the
allowed motion is in the span of X , Y . Note that, using the matrix components as
our coordinate system, the directional derivative of the coordinate is ∇X R = XR ,
∇Y R = YR . In particular, ∇X (Y ) = [e2 ×]∇X R = [e2 ×][e1 ×]R. By a similar
computation, we get

[X , Y ]R = ∇X YR − ∇Y XR = [e1 ×][e2 ×]R − [e2 ×][e1 ×]R (2.86)

= [e1 × e2 ]R = [e3 ×]R = ZR . (2.87)

In particular, Lie bracket between vector fields within the distribution span(X , Y )
can leave the distribution. Therefore, we have a non-holonomically constrained
system. ■

■ Example 2.20 — Parallel parking. The configuration space of a four-wheel car

on a plane is M = {(x , y, θ, κ): x , y ∈ R, θ, κ ∈ [0, 2π)}, where (x , y) denotes the
location of the car on the plane, θ denotes the angle between the central axis of
the car and the x -axis, and κ denotes the angle at which the front wheels are
turned from the car’s central axis.

θ
x

The car can move by hitting its acceleration and going forward in its current
direction θ and changing its direction by its current steering wheel angle κ, or the
 54 Chapter 2. Exterior Calculus
car can move by changing the angle of its steering wheel. These two actions each
corresponds to the following two generators:
   
cos θ 0
Xforward = κsin θ and Xturn = 00 . (2.88)
0 1

To see the effect of applying the following sequence of actions: move forward, steer
to the left, move backward, steer to the right, we can look at the Lie bracket of
these two generators:

[Xforward , Xturn ] = ∇Xforward Xturn − ∇Xturn Xforward

 0 0 − sin θ 0       
0 0 0 0 0 cos θ 0
= 00 00 cos0 θ 10 0 − 0 0 0 0 sinκ θ = 01 .
0 0 0 0 0 (2.89)
0 0 0 0 1 0 0 0 0 0 0

This means that the action sequence generates a spinning motion. We denote its
generator by  
0
Xspin = 01 . (2.90)
0

We can further analyze the effect of the following sequence of actions: spin to
the left, move forward, spin to the right, move backward by calculating the Lie
bracket

[Xspin , Xforward ] = ∇Xspin Xforward − ∇Xforward Xspin

    0 0 − sin θ 0    − sin θ 
0 0 0 0 cos θ 0
0
= 00000 0 0 sin
κ
θ − 00 00 cos0 θ 10 01 = cos0 θ . (2.91)
0 0 0 0 0 0 0 0 0 0 0

This means that the action sequence generates a sideway motion where the car
moves towards − sin θ, cos θ while its front points towards (cos θ, sin θ. We can
denote this generator as Xsideway . We’ve thus far obtained four generators of
motion:
       − sin θ 
cos θ 0 0
Xforward = κ , Xturn = 0 , Xspin = 1 , and Xsideway = cos0 θ .
sin θ 0 0 (2.92)
0 1 0 0

What we are seeing is that by introducing Lie bracket, the space of all possible
motion made from sequencing moving forward and turning the steering wheel
is actually enough to cover the entire tangent space Tp M of the configuration
manifold M , meaning the car is capable of traveling over the entire plane even
though it only has two motions. ■

2.5 Covariant Derivatives and Vector-Valued Forms

As we briefly remarked at (2.84), the componentwise differential “∇X Y ” of a vector
field Y along the X direction is coordinate dependent. What is a coordinate-
independent notion of ∇X Y ?
So far, we have only been looking at scalar-valued k -forms. (Componentwise
differentials in (2.84) is treating vectors as multi-component scalars.) Another
frequently occurring type of object is vector-valued k -forms. Recall a k -form is a
2.5 Covariant Derivatives and Vector-Valued Forms 55

section of the bundle

Vk ∗ T M . One may take a pointwise tensor product with the
Vk ∗
tangent bundle ( T M ) ⊗ TM , whose sections are both a k -form and a vector
field. These are vector-valued k -forms. For notation convenience we call
V 
k
Γ ( T M ) ⊗ TM = Ωk (M ; TM ).
∗
(2.93)

A vector field Y ∈ Γ(TM ) can also be seen as a vector-valued 0-form Y ∈ Ω0 (M ; TM ).

One may ask what is its directional derivative along a direction X ∈ Tp M at a point
p ∈ M.
The natural definition of the directional derivative of a vector would depend on a
metric.
Definition 2.14 Let (M , ♭) be a Riemannian manifold. A differential operator that
takes a vector field to a vector-valued 1-form

∇ : Ω0 (M ; TM ) → Ω1 (M ; TM ) (2.94)

is a Levi-Civita connection or a covariant derivative if

• Compatible with differential of scalar: ∇(fY ) = (df )Y +f ∇Y for f ∈ Ω0 (M )

and Y ∈ Ω0 (M ; TM ); (2.95)

• Compatible with metric: d ⟨X , Y ⟩ = ⟨∇X , Y ⟩ + ⟨X , ∇Y ⟩; (2.96)

• Compatible with Lie bracket (torsion free): ∇X Y − ∇Y X = [X , Y ]. (2.97)

Theorem 2.26 — Fundamental theorem of Riemannian geometry. The Levi-Civita

connection exists and is unique.
So, on a manifold with metric, we can just take differential of a vector field like
taking differential of a standard 0-form. The only difference in notation is that we
use ∇ instead of d .
Definition 2.15 — Exterior covariant derivative. Using the covariant derivative ∇
we can apply exterior derivative to vector-coefficient k -forms (i.e. vector-valued
k -forms). The operator

d ∇ : Ωk (M ; TM ) → Ωk +1 (M ; TM ) (2.98)

is characterized by that

• It is the same as ∇ when k = 0. (2.99)

• Graded Leibniz rule: d ∇ (α ∧ A) = (d α) ∧ A + (−1)deg(α) α ∧ d ∇ A for scalar-

valued form α and vector-valued form A. (2.100)

R Despite the similarity between d ∇ and d , there are some differences. For
example,
d∇ ◦ d∇ (2.101)
 56 Chapter 2. Exterior Calculus

is generally nonzero. In fact, d ∇ ◦ d ∇ = R∧ where R is a matrix-valued 2-form

called Riemann curvature tensor.

Definition 2.16 The identity is a vector valued 1-form

I ∈ Ω1 (M ; TM ), I JX K := X . (2.102)

The identity 1-form I is useful for expressing the flat of a vector

X ♭ = ⟨X , I ⟩. (2.103)

One checks that X ♭ JY K = ⟨X , I JY K⟩ = ⟨X , Y ⟩.

Theorem 2.27 The torsion free condition (2.97) for a derivative operator ∇ is
equivalent to

d ∇ I = 0. (2.104)

Proof. The exterior (covariant) derivative d on a 1-form α has an explicit formula of

(d α)JX , Y K = d (αJY K)JX K − d (αJX K)JY K − αJ[X , Y ]K, which we will give a proof
later in Theorem 2.36. Now, letting α be I , and X , Y be arbitrary vector fields, we
get

(d ∇ I )JX , Y K = ∇(I JY K)JX K − ∇(I JX K)JY K − I J[X , Y ]K (2.105)

= ∇X Y − ∇Y X − [X , Y ]. (2.106)

This expression vanishes if and only if the torsion free condition (2.97) holds.

2.6 Lie Derivative

Lie derivative is an extremely fundamental derivative in multivariable calculus, which
is unfortunately not mentioned in most of the course on multivariable calculus.
Lie derivative is defined by the rate of change of a time-dependent pullback
operator. Recall that the pullback operator is the just the unification of general
changes of coordinates/variables. The question about what the rate of change is
of a time-dependent change of coordinates/variables naturally arises. The unifying
concept for this rate of change is the Lie derivative, which is as important as how
changes of variables are important in calculus.
The use of Lie derivative is independently introduced, with different names, in
applied research fields when rates of change of changes of variables are unavoidable.
For example, its special cases include Reynolds transport theorem that is used in
differentiable rendering [Zhao et al.(2020b)]. Lie derivative also unifies convection
and stretching terms in continuum mechanics, which is the study of the differential
relations of a deformable body.
 2.6 Lie Derivative 57

2.6.1 General definition

Let A ∈ ⟨type⟩M be a field defined over M of a particular type ⟨type⟩. This space
⟨type⟩M can be functions (0-forms) Ω0 (M ), k -forms Ωk (M ), vector fields Γ(TM ), or
any other types that has a pullback operator

linear
φ∗ : ⟨type⟩M −−−→ ⟨type⟩M (2.107)
⟨type⟩

associated to any map φ : M → M . The Lie derivative is the differential operator

that measures the rate of change of the pullback φ∗t A of A when the time-dependent
⟨type⟩
map φt as a flow map has a known velocity field.
Definition 2.17 — Flow map generated by vector field. For each vector field
X ∈ Γ(TM ) on M , the flow map φt : M → M , t ∈ (−ϵ, ϵ), generated by X is the
solution to
∂φ
= X ◦ φ and φ0 = idM . (2.108)
∂t

Definition 2.18 — Lie derivative. Let X ∈ Γ(TM ) be a vector field on M and let
φt : M → M be the flow map generated by X (Definition 2.17). Define the Lie
derivative for ⟨type⟩

linear
LX : ⟨type⟩M −−−→ ⟨type⟩M (2.109)
⟨type⟩

by

∂
LX A := ( φ∗t A). (2.110)
⟨type⟩ ∂t t=0 ⟨type⟩

This notion of Lie derivative can be used in situations beyond taking derivative of
the infinitesimally flow map (2.108) generated around the identity map. Whenever
there is a time dependent map, the rate of change can of the pullback can be written
in terms of the Lie derivative.
Theorem 2.28 — Rate of change of pullback. Let M and W be two manifolds.
Let ϕt : M → W be a time-dependent map, and suppose there is ⃗u ∈ Γ(TW ) so
that
∂ϕt
= ⃗u ◦ ϕt . (2.111)
∂t
Let A ∈ ⟨type⟩W be some field defined on W . Then
! !
∂
ϕ∗t A = ϕ∗t L⃗u A . (2.112)
∂t ⟨type⟩ ⟨type⟩ ⟨type⟩
 58 Chapter 2. Exterior Calculus

Corollary 2.29 — Passing the time derivative through a pullback. Continue the
setup of Theorem 2.28. Suppose At ∈ ⟨type⟩W is a time-dependent field defined
on W . Then
! !
∂ ∂
ϕ∗t At = ϕ∗t At + L⃗u At . (2.113)
∂t ⟨type⟩ ⟨type⟩ ∂t ⟨type⟩

We call the operator ( ∂t

∂
+ L⃗u ) the Lie material derivative.

Definition 2.19 We say that a time-dependent field At ∈ ⟨type⟩W is Lie-advected

by ⃗u if
∂
A
∂t t
+ L⃗u At = 0; (2.114)
⟨type⟩

equivalently, for ∂ϕt

∂t
= ⃗u ◦ ϕt ,

ϕ∗t At = ϕ∗0 A0 is time independent. (2.115)

⟨type⟩ ⟨type⟩

2.6.2 Lie derivative for functions

The Lie derivative for functions is just the directional derivative.

Theorem 2.30 — Lie derivative for functions. For f ∈ Ω0 (M )

LX f = df JX K = Xf . (2.116)
0-form

Proof. Let φt be the flow map generated by X ∈ Γ(TM ) (Definition 2.17). Then
LX f = ∂t∂
|t=0 ( φ∗t f ) = ∂t
∂
|t=0 (f ◦ φt ) = df |φt J ∂φ
∂t
t
K|t=0 = df JX K.
0-form 0-form

2.6.3 Lie derivative for vector fields

The Lie derivative for vector fields is the Lie bracket (Theorem 2.23).

Theorem 2.31 — Lie derivative for vector fields. For X , Y ∈ Γ(TM )

LX Y = [X , Y ]. (2.117)
vec

To show Theorem 2.31, first note that the pullback of a vector field is given by

φ∗t Y = (d φt )−1 Y . (2.118)

vec

This pullback of vectors is natural, as we have

φ∗t (Yf ) = (φ∗t Y )( φ∗t f ). (2.119)

0-form vec 0-form
2.6 Lie Derivative 59

∇X Y = 0 LX Y = 0
vec

Y Y

X φt (p) X φt (p)

p p

Figure 2.8 The difference between the covariant derivative ∇X Y and the Lie
derivative LX Y = [X , Y ]. Vanishing covariant derivative implies
that Y is parallel transported by the path generated by X (left). The
dashed lines are geodesics constituting infinitesimal parallelograms for
the transportation (known as Schild’s ladder ), indicating the metric
dependency of ∇X . On the other hand, vanishing Lie derivative
implies that Y is the pushed forward by the flow map generated by
X (right), which is metric independent but requires the neighborhood
information of X instead of only along a path.

Proof of Theorem 2.31. Let φt be the flow map generated by X ∈ Γ(TM ), and let
f be an arbitrary function. Taking the time derivative of (2.119) yields

LX (Yf ) = (LX Y )f + Y ( LX f ) (2.120)

0-form vec 0-form
=⇒ X (Yf ) = (LX Y )f + Y (Xf ) (2.121)
vec
=⇒ (LX Y )f = X (Yf ) − Y (Xf ) = [X , Y ]f . (2.122)
vec

Comparison between covariant derivative ∇X Y and Lie derivative LX Y

We have learned two derivatives of a vector field Y ∈ Γ(TM ) along another vector
field X ∈ Γ(TM ). One is the covariant derivative ∇X Y (Definition 2.14), and the
other is the Lie derivative LX Y (Theorem 2.31). Figure 2.8 shows their differences.
vec
Consider the vector fields Y that satisfies ∇X Y = 0, and respectively LX Y = 0.
vec
The equation ∇X Y = 0 implies that Y is parallel transported along X , and
∇X Y measures how much Y deviates from being parallel when we take a step in
the X direction. This type of X -directional derivative of Y only depends on the
direction of X at a single point. But the result ∇X Y depends on a metric as is
required in Definition 2.14. In fact, since ∇X Y is a directional derivative, we have

∇X Y = LX∇ Y (2.123)
0-form

where Y is viewed as a vector-valued 0-form, and the Lie derivative is built by taking
the rate of change of the 0-form-typed pullback.
The equation LX Y = 0, on the other hand, implies that Y is pushed forward by
vec
the flow generated by X . The quantity LX Y measures how much Y deviates from
vec
 60 Chapter 2. Exterior Calculus

being purely pushed forward by the flow. The idea of pushforward by a flow, as well
as the Lie bracket, is independent of any metric. However, this type of “directional
derivative” of Y along a flow X requires the information of X in a neighborhood,
not only at one point. It is the neighborhood behavior of X that establishes an
infinitesimal pushforward (deformation gradient or Jacobian matrix) of the flow.

2.6.4 Lie derivative for differential forms

Finally, we give the formula for the Lie derivative for differential forms.

Theorem 2.32 — Cartan’s magic formula. For α ∈ Ωk (M )

LX α = iX d α + diX α. (2.124)
k -form

Lie derivatives for differential forms are particularly important, as they allow us to
take derivative on the integration domain:
Theorem 2.33 — Derivative of an integral. Let X be the vector field so that
ϕ̇ = X ◦ ϕ. Then,
ˆ ˆ  
d ∂
αt = αt + L X αt (2.125)
dt ϕt (S ) ϕt (S ) ∂t k -form

´ ´ ´ ´
Proof of Theorem 2.33. ∂
α =
ϕt (S ) t
∂
ϕ∗ α =
S t t
∂
(ϕ∗t αt ) = ϕ∗t ( ∂t
∂
α + LX α)
´ ∂
∂t ∂t S ∂t S
= ϕt (S ) ( ∂t α + LX α).

Now we give a simple but abstract proof for Cartan’s magic formula (2.124). This
proof still deserves geometric intuition, which we will provide afterward.

Proof of Theorem 2.32. The operator LX must satisfy the following 3 properties:
k -form

• Base definition: LX = LX when k = 0;

k -form 0-form

• Derivation: LX (α ∧ β) = (LX α) ∧ β + α ∧ (LX β);

form form form

• Commutativity with d : LX d α = d LX α.
form form

The “derivation” and “commutativity with d ” properties follow directly from the
analogous properties Theorem 2.7 and Theorem 2.11 for the pullback operator. Now,
notice that the operators on differential forms that satisfy these 3 properties must be
unique. This is because any differential form ω can be written as a linear combination
(withP0-form as coefficients) of wedge products of the d of some coordinate functions
ω = i1 <...<ik ωi1 ...ik dx i1 ∧ · · · ∧ dx ik , and the above 3 properties form a sufficient
set of rules to evaluate the Lie derivative of this expression. To complete the proof,
check that (2.124) satisfy all the 3 properties.
 2.6 Lie Derivative 61

boundary of extrusion of
extrusion boundary

d + +
φt (S ) = = + +
dt +
− + −
+ −

Figure 2.9 An intuition behind Cartan’s magic formula. The Lie derivative of a
differential form is dual to the rate of change of a moving integration
domain. This rate of change of a domain can be written as the
boundary of the extrusion plus the extrusion of the boundary.

Geometric intuition behind Cartan’s magic formula

The geometric intuition behind (2.124) is best appreciated when viewed under an
integral sign. To understand this, consider (2.125), but with α being independent of
t, and ϕt = φt representing the flow map generated by X around the identity map
as given by Definition 2.17:
ˆ ˆ
d
LX α = α. (2.126)
S dt φt (S )

In essence, the Lie derivative, through the integral pairing between an integration
domain and a differential form, acts as the dual operator to the rate of change of the
integration domain.
As illustrated in Figure 2.9, this rate of change of the integration domain can
be expressed as the sum of the boundary of the extrusion along the flow and the
extrusion of the boundary. Recall from Theorem 2.10 and Theorem 2.12 that the
operations of extrusion and taking boundary are dual to the interior product and
the exterior derivative, respectively. Thus, we arrive at Cartan’s magic formula.

2.6.5 Useful identities

We have explored different types of Lie derivatives in Sections 2.6.2–2.6.4. Here,
we collect a few identities by combining these various types of Lie derivatives.

Theorem 2.34 Let X , Y1 , . . . , Yk be vector fields and α be a k -form. Then

LX (αJY1 , . . . , Yk K) = (LX α)JY1 , . . . , Yk K + αJ[X , Y1 ], . . . , Yk K+

· · · + αJY1 , . . . , [X , Yk ]K (2.127)

Proof. This follows from taking the t-derivative of

φ∗t (αJY1 , . . . , Yk K) = ( φ∗t )Jφ∗t Y1 , . . . , φ∗t Yk K. (2.128)

0-form k -form vec vec

Similar to Theorem 2.34, we have:

 62 Chapter 2. Exterior Calculus

Theorem 2.35 Let X , Y be vector fields and α be a k -form. Then we have

LX (iY α) = i[X ,Y ] α + iY (LX α). (2.129)

The following identity is an explicit formula for the evaluation of the d of a 1-

form.
Theorem 2.36 For α ∈ Ω1 (M ), we have

(d α)JX , Y K = X ⟨α|Y ⟩ − Y ⟨α|X ⟩ − ⟨α|[X , Y ]⟩. (2.130)

(2.124) (2.129)
Proof. (d α)JX , Y K = iY iX d α = iY (LX α) − iY diX α = LX (iY α) − i[X ,Y ] α −
| {z }
Y ⟨α|X ⟩
Y ⟨α|X ⟩ = X ⟨α|Y ⟩ − ⟨α|[X , Y ]⟩ − Y ⟨α|X ⟩.

2.6.6 Lie derivatives in the 3D Cartesian space

■ Example 2.21 In the 3D Cartesian space, the Lie derivatives for differential
forms and vector fields correspond to the following expressions in vector calcu-
lus.
Theorem 2.37 — Lie derivatives in vector calculus.

Lu (a)0-form = (u · ∇a)0-form , (2.131)

0-form
⊺
Lu (a)1-form = (u · ∇a + (∇u) a)1-form , (2.132)
1-form

Lu (a)2-form = (u · ∇a − a · ∇u + (∇ · u)a)2-form , (2.133)

2-form

Lu (a)3-form = (u · ∇a + (∇ · u)a)3-form . (2.134)

3-form

Lu (a)vec = (u · ∇a − a · ∇u)vec . (2.135)

vec
The traditional material derivative along a vector field u defined in most
existing textbook is given by
D
Dt
:= ∂
∂t
+ u · ∇. (2.136)

The Lie material derivative, which measures the real rate of change of various
forms, deviates from the material derivative by additional terms:
∂
( ∂t D
+ Lu )(a)0-form = ( Dt a)0-form , (2.137)
0-form
∂
( ∂t D
+ Lu )(a)1-form = ( Dt a + (∇u)⊺ a)1-form , (2.138)
1-form
∂
( ∂t D
+ Lu )(a)2-form = ( Dt a − a · ∇u + (∇ · u)a)2-form , (2.139)
2-form
∂
( ∂t D
+ Lu )(a)3-form = ( Dt a + (∇ · u)a)3-form , (2.140)
3-form
∂
( ∂t D
+ Lu )(a)vec = ( Dt a − a · ∇u)vec . (2.141)
vec
2.6 Lie Derivative 63
Now, let us show Theorem 2.37.
Proof of (2.131). Eq. (2.131) follows directly from Theorem 2.30.
Proof of (2.132). Using (2.103), we write (a)1-form = ⃗a ♭ = ⟨⃗a , I ⟩, where ⃗a = (a)vec
and I is the identity vector-valued 1-form (Definition 2.16). For clarity we also
(2.124) (2.96)
write ⃗u = (u)vec . Now, Lu (a)1-form = L⃗u ⟨⃗a , I ⟩ = di⃗u ⟨⃗a , I ⟩ + i⃗u d ⟨⃗a , I ⟩ =
Thm. 2.27
∇ (2.96),(2.26)
d ⟨⃗a , ⃗u ⟩+i⃗u ⟨∇⃗a ∧I ⟩+i⃗u ⟨⃗a , d I ⟩ = ⟨∇⃗a , ⃗u ⟩+⟨⃗a , ∇⃗u ⟩+⟨∇⃗u ⃗a , I ⟩−⟨∇⃗a , ⃗u ⟩ =
⟨⃗a , ∇⃗u ⟩ + ⟨∇⃗u ⃗a , I ⟩ = ((∇u)⊺ a + ∇u a)1-form .
Proof of (2.135). Eq. (2.135) follows from Theorem 2.31 and (2.97).
Proof of (2.134). Let µ = (1)3-form be the 3D volume form. Then iu µ = (u)2-form
(2.124)
and Lu µ = diu µ = d (u)2-form = (∇·u)3-form . Now, Lu (a)3-form = Lu ((a)0-form µ)
= Lu (a)0-form µ + (a)0-form Lu µ = (∇u a + (∇ · u)a)3-form .
Proof of (2.133). Writing (a)2-form = i⃗a µ where ⃗a = (a)vec and µ is the 3D volume
(2.129) (2.134)
form, we have L⃗u (a)2-form = L⃗u (i⃗a µ) = i[⃗u ,⃗a ] µ + i⃗a (L⃗u µ) = i[⃗u ,⃗a ] µ + (∇ ·
(2.135)
u)i⃗a µ = (∇u a − ∇a u + (∇ · u)a)2-form .
In fact, during our proof of (2.133), we only used the fact that a 2-form is
the contraction of a vector and a 3-form, each of which has an established Lie
derivatives (2.134) and (2.135). What we may also explore is to expand the
left-hand side of (2.133) using Cartan’s formula (2.124):

Lu (a)2-form = diu (a)2-form + iu d (a)2-form (2.142)

= (∇ × (a × u) + (∇ · a)u)2-form . (2.143)

Combining this result with (2.133) yields the following vector calculus iden-
tity
Theorem 2.38 — Curl of cross product.

∇ × (a × u) = u · ∇a − a · ∇u + (∇ · u)a − (∇ · a)u. (2.144)

Similarly, combining Cartan’s magic formula for 1-forms with (2.132) gives us
the following vector calculus identity
Theorem 2.39 — Cross product of curl.

(∇ × a) × u = u · ∇a + ∇(a · u) − (∇u)⊺ a. (2.145)

■

■ Example 2.22 — Reynolds’ transport theorems. Consider (2.125) with S =

Ω ⊂ Rn being an n-dimensional region, and α = (f µ) being an n-form, where
f is some function and µ is the volume form in Rn . Let Ωt = ϕt (Ω) denote a
moving domain, and let u be the velocity of the moving boundary; more precisely,
 64 Chapter 2. Exterior Calculus

u is the velocity field written in coordinates so that ´ ϕ̇t = (u)vec´◦ ϕt . Then,

by Cartan’s magic formula and Stokes’ theorem, Ω L⃗u (f µ) = Ω di⃗u (f µ) =
¸ n-form

∂Ω
f (u) (n − 1)-form . Hence (2.125) can be written as the following vector calculus
identity:
Theorem 2.40 — Reynolds’ transport theorem. Let Ωt be a domain moving
with a velocity u. Then
ˆ ˆ ˛
d ∂f
f dV = dV + f u · ndA. (2.146)
dt Ωt Ωt ∂t ∂Ωt
The special case of the Reynolds’ transport theorem in n = 1 is:
Theorem 2.41 — Leibniz’s integral rule. In 1D calculus,
ˆ b(t) ˆ b(t)
d ∂f (t, x )
f (t, x )dx = dx + b ′ (t)f (t, b(t)) − a ′ (t)f (t, a(t)).
dt a(t) a(t) ∂t
(2.147)
Specializing (2.125) to the case of a surface in the 3D Cartesian space yields

Theorem 2.42 Let St be a surface in 3D moving at a velocity u. Let f be a

(time-dependent) vector field defined on the 3D space. Then,
¨ ¨   ˛
d ∂f
f · ndA = + (∇ · f )u · ndA + (f × u) · d l. (2.148)
dt St St ∂t ∂St

An alternative formula (using (2.133)) is

¨ ¨  
d ∂f
f · ndA = + u · ∇f − f · ∇u + (∇ · u)f · ndA (2.149)
dt St St ∂t
Specializing (2.125) to the case of a curve in the 3D Cartesian space yields

Theorem 2.43 Let Ct be a curve in 3D moving at a velocity u. Let f be a

(time-dependent) vector field defined on the 3D space. Then,
ˆ ˆ  
d ∂f
f · dl = + (∇ × f ) × u · d l + (f · u)|∂Ct . (2.150)
dt Ct Ct ∂t

Alternatively, using (2.132), we also have

ˆ ˆ  
d ∂f
f · dl = ⊺
+ u · ∇f + (∇u) f · d l. (2.151)
dt Ct Ct ∂t
■

2.7 Pairings, Inner Products and Hodge Stars

In this last section of the chapter, we introduce the Hodge star. The previous
operators we have introduced (∧, iX , d , Lie brackets, etc.) are built with canonical
2.7 Pairings, Inner Products and Hodge Stars 65

construction without depending on metric. In particular, they all have simple

interplay with the pullback operator. The Hodge star that we are studying here is
metric dependent. In particular, it does not have simple interplay with pullback or
other operators.
The Hodge star is defined with the notion of an L2 inner product. Let us first
clarify dual pairing and inner product.
Theorem 2.44 — Metric-independent pairing. The dual space of k -forms is iso-
morphic to (n − k )-forms.

∗
Ωk (M ) ∼ = Ωn−k (M ). (2.152)

This is done through the following dual pairing

bilinear
Ωk (M ) × Ωn−k (M ) −−−−→ R
ˆ
(α, β) 7→ α ∧ β. (2.153)
M

´ ˝
■ Example 2.23 In the 3D Cartesian space, M
(a)1-form ∧ (b)2-form = M
a · b dV .
■

Definition 2.20 — Inner product between forms. Suppose ♭ is an inner product

structure on a vector space V . This inner product structure defines a natural
volume form

µ ∈ ∧n V ∗ , µJX1 , . . . , Xn K = 1 if X1 , . . . , Xn are positively orthonormal.

(2.154)

The metric also defines a natural inner product on k -forms: for α, β ∈ ∧k V ∗

X
⟨α, β⟩ := αJXi1 , . . . , Xik K βJXi1 , . . . , Xik K (2.155)
1≤i1 <···<ik ≤n

using any orthonormal basis X1 , . . . , Xn .

Definition 2.21 — Hodge star. Suppose ♭ is an inner product structure on a vector

space V . Let µ be the associated volume form. Define
linear
⋆k : ∧k V ∗ −−−→ ∧n−k V ∗ (2.156)

such that

α ∧ ⋆β = ⟨α, β⟩µ. (2.157)

Definition 2.22 — L2 inner product. Suppose ♭ is an inner product structure on

66 Chapter 2. Exterior Calculus
Tp M for each p ∈ M . Define the L2 inner product for Ωk (M ) as
ˆ ˆ
⟪α, β⟫ := ⟨α, β⟩ µ = α ∧ ⋆β. (2.158)
M M

■ Example 2.24 In the 3D Cartesian space

⋆ (a)0-form = (a)3-form , (2.159)

⋆ (a)1-form = (a)2-form , (2.160)
⋆ (a)2-form = (a)1-form , (2.161)
⋆ (a)3-form = (a)0-form . (2.162)

Theorem 2.45 The Hodge star gives a duality relation between the wedge product
and the interior product

iX ⋆k α = (−1)k ⋆k +1 (X ♭ ∧ α). (2.163)

That is,
iv iv iv iv
∧0 V ∗ o ∧O 1 V ∗ o ∧O 2 V ∗ o ∧O 3 V ∗ o ··· ∧n−1 V ∗ o ∧O n V ∗ (2.164)
⋆0 ⋆1 ⋆2 ⋆3 ⋆n−1 ⋆n
 v♭ ∧  v♭ ∧  v♭ ∧   v♭ ∧ 
∧n V ∗ / ∧n−1 V ∗ / ∧n−2 V ∗ / ∧n−3 V ∗ ··· / ∧1 V ∗ / ∧0 V ∗
iv −iv iv (−1)n−1 iv

We conclude our exterior calculus chapter here. There are more objects one
can build with the Hodge star that we have not covered, such as the codifferential
δ = (−1)k ⋆−1 d ⋆ and Hodge Laplacian ∆ = d δ + δd .
Technically, exterior calculus with metric (i.e. with Hodge star mixed in) makes it
the so-called geometric calculus. In particular, one can replace the wedge product
with the geometric product formed by the geometric algebra (Clifford algebra) of the
given metric. One obtains the Dirac operator d + δ as the natural generalization of d
to geometric algebra. In this sense, just as exterior algebra is the geometric algebra
with 0 metric, the exterior calculus is the metric-independent geometric calculus.
That is why it reveals the important structures about the differential topological
aspect of multivariable calculus such as pullback, Lie derivative.
II
Graphics Topics in Exterior
Calculus

3 Geometric Optimizations . . . . . . . 69
3.1 Geometry representations
3.2 A Geometric view of optimization problems
3.3 Computational methods for optimization prob-
lems

4 Continuum Mechanics . . . . . . . . . . 89
4.1 Deformation map
4.2 Elasticity overview
4.3 Tensor Algebra
4.4 Postulates of Elasticity
4.5 Derivation of Elasticity
4.6 Cauchy stress tensor
4.7 More on the strain energy density function
4.8 Conclusion

5 Fluid Dynamics . . . . . . . . . . . . . . . . . . 115

5.1 Euler equations
5.2 Advantages of Covector Euler Equation
5.3 Conclusion
 3. Geometric Optimizations

In this chapter, we take a new, geometric look at the study of optimizations. This
includes optimization problems where the optimal solution in search is itself a
geometry (Section 3.1) and the geometric picture of the computation of optimization
problems (Section 3.2 and Section 3.3).

3.1 Geometry representations

In graphics, geometry can refer to a point distribution (a point cloud ), a curve,
a surface, or a volumetric object. They each have co-dimension 3, 2, 1, and 0,
and they correspond to the domains on which 0-form, 1-form, 2-form, 3-forms are
integrated.
Notation 3.1 Throughout this chapter, we use M to denote the material space
and W to denote the world space. M is the abstract space of the manifold, like
the index of particles of a point cloud, or the combinatorial structure of a mesh.
W is the world space where the manifold is embedded and in most case, just the 2-
or 3-dimensional Euclidean space. We always denotes the dimension of M as m
and the dimension of W as n.

3.1.1 Lagrangian representations: manifold embedding

Definition 3.1 An embedding is an injective smooth map f : M → W that
preserves the topological and the manifold structure. In the context of computer
graphics, this corresponds the positions of the vertices of a mesh / polycurve,
while the manifold M itself corresponds to the combinatorial structure of the mesh
/ polycurve.
70 Chapter 3. Geometric Optimizations
vertex 3
position (x3 , y3 )

vertex 1
position (x1 , y1 ) vertex 4
position (x4 , y4 )

vertex 2
position (x2 , y2 )

Manifold M Embedding f : M → W
Triangle id Vertices Vertex id Position
1 (1, 2, 3) 1 (x1 , y1 )
2 (3, 2, 4) 2 (x2 , y2 )
3 (x3 , y3 )
4 (x4 , y4 )

■ Example 3.1 — Function graph. Given a function graph z = g(x , y), we can
view it as an embedding

f:M → W
(x , y) 7→ (x , y, f (x , y)).

Here the material space M is the 2-dimensional parameter space of x and y, and
W is the 3-dimensional world space of x , y, and z . ■

Many classical geometric properties like curvature and second fundamental form
are defined using the embedding of a manifold. We give a few examples of them
below.
Definition 3.2 — Differential geometry of curves by embedding function. Let M
be an 1-dimensional manifold and f : M → W is an embedding of M into W = R3 .
Let p ∈ M be a point. Locally (in a neighborhood around p), we can parametrize
M as an interval (a, b) of the real line, and use the Cartesian coordinates x , y, z
for R3 and write the embedding f : (a, b) → R3 as f (t) = (x (t), y(t), z (t)) for
t ∈ (a, b). The unit tangent vector is given as

(x ′ (t), y ′ (t), z ′ (t))

T(t) = , (3.1)
|(x ′ (t), y ′ (t), z ′ (t))|ℓ2

where |(x , y, z )|ℓ2 = x 2 + y 2 + z 2 denotes the Euclidean norm. An intuitive

p

observation tells us that the faster the tangent direction changes, the curvier the
curve is. Notice that the curviness property only depends on the tangent direction,
not the derivative of the embedding function. We hence define the curvature as
3.1 Geometry representations 71

Figure 3.1 An abstract 1-dimensional manifold M (bottom left) that resembles

a circle can be embedded to any wiggly shape (top right) but we can
never embed it to something that has an open endpoints (top left) or
something knotted (bottom right).

the rate of change of the unit tangent:

κ(t) = Ṫ . (3.2)
ℓ2

R It is crucial that the readers see that the above definition of unit tangent and
curvature are both embedding dependent! Without an embedding function
f : M → W , the manifold still has its own (albeit abstract) tangent space
Tp M ∼ = R. An embedding f : M → W can convert the abstract tangent to an
R3 vector by the push-forward map
linear
df : Tp M ∼
= R −−−→ Tf (p) W ∼ = R3
 ′ 
x (t) (3.3)
linear  ′
1 7−−−→ y (t) .
z ′ (t)

R Another crucial insight is that the embedding function f : M → W has to

be smooth. The smoothness requirement means that the embedding must
preserve the manifold’s topology, including topological features like genus,
knottedness, or open endpoints. See Figure 3.1 and Figure 3.2. This is why
when a simulation artist uses meshed surfaces to model fracturing material (e.g.
tearing a piece of paper), their simulation pipeline cannot simply modify the
vertex positions but it sometimes have to modify the combinatorial structure
(often called remeshing) when the surface go through topological changes!
 72 Chapter 3. Geometric Optimizations

Figure 3.2 An abstract 2-dimensional manifold M (bottom left) that resembles

a sphere can be embedded to any wiggly shape (top right) but we
can never embed it to something that has boundary (top left) or
something with a nonzero genus (bottom right).

3.1.2 Integration pairing

Now that we’ve learned about the distinction between an abstract manifold M and
their embedding f : M → W into the world space, we can dive into the relation
f
between an embedded manifold M ,−
→ W and differential forms in the world space
Ωm (W ).
Definition 3.3 We define the integration pairing between an embedded k -dimensional
f
manifold M ,−
→W a
and a k -form α ∈ Ωk (W ) as the k -dimensional integral:
ˆ
⟨M |α⟩ := α. (3.4)
M

The integration pairing corresponds to function evaluation (k = 0), oriented line

integral (k = 1), oriented surface integral (k = 2), and the usual volumetric
integral (k = 3).
a
It is possible to consider a broader range of integration pairings: as long as the differential
forms are smooth, we can integrate them not only on smoothly embedded manifolds, but also on
rectifiable sets, which are patches of smoothly embedded manifolds joined together in a merely
continuous sense.

Theorem 3.1 The integration pairing is bilinear, meaning that we can take linear
 3.1 Geometry representations 73
combination in both the k -dimensional embedded manifolds and the k -forms: for
a1 , a2 ∈ R two real numbers, M1 , M2 ⊂ W two embedded k -dimensional manifolds,
α1 , α2 ∈ Ωk (W ) two k -forms,
ˆ ˆ
⟨a1 M1 + a2 M2 |α⟩ = a1 α + a2 α = a1 ⟨M1 |α⟩ + a2 ⟨M2 |α⟩, (3.5)
ˆM1 ˆM2
⟨M |a1 α1 + a2 α2 ⟩ = a1 α1 + a2 α2 = a1 ⟨M |α1 ⟩ + a2 ⟨M |α2 ⟩. (3.6)
M M

Theorem 3.2 By Stokes’ Theorem (Theorem 2.12), the exterior derivative d of

differential forms and the operation of taking boundary ∂ of the integration domain
are adjoint of each other under the integration pairing:
ˆ ˆ
Stokes
⟨S |d α⟩ = dα = α = ⟨∂S |α⟩. (3.7)
S ∂S

3.1.3 Exterior calculus on combinatorial surfaces and curves (DEC)

Definition 3.4 A tessellated (meshed) surface is a tuple (V , E , F ) of vertices V ,
edges E , and faces F . A vertex has dimension 0, so we called it a 0-cell. We call
an edge a 1-cell and a face a 2-cell. Together, we call (V , E , F ) a cell complex.
The boundary operator ∂ is defined by the following:

• Since, for example, a face can have multiple edges that form its boundary,
the boundary operator needs to map a face to a combination of multiple
edges.

• We denote the set of all weighted sum of faces as CF , weighted sum of edges
as CE , and weighted sum of vertices as CV . The orientation of an edge /
face corresponds to the sign of the weights eij = −eji , fijk = fjki = fkij =
−fjik = −fikj = −fkji . For reference, we referred to a weighted sum of cells
as a chain.

• We then define the boundary operator as the linear maps ∂1 , ∂2

∂ ∂
CV ←
−1
CE ←
−2
CF (3.8)

such that if eij , ejk , eki are the three edges forming the boundary of a triangu-
lar face fijk , then ∂2 (fijk ) = eij + ejk + eki , and if vi , vj are the two endpoints
forming the boundary of an edge eij , then ∂1 (eij ) = −vi + vj . Readers can
check that ∂1 ◦ ∂2 = 0 as the boundary of a face is a loop, which has no
endpoints.

Definition 3.5 With the above example in mind, we can define a generalized
m-dimensional cell complex as M = (M0 , · · · , Mm ) together with a boundary
 74 Chapter 3. Geometric Optimizations
operator and the chain complex:
∂ ∂ ∂
C0 ←
−1
C1 ←
−2
· · · ←m− Cm . (3.9)

This definition should cover poly-lines (1-dimensional cell complex), tesselated

(meshed) surfaces (2-dimensional cell complex), and volumetric meshes (e.g. tetra-
hedral mesh) (3-dimensional cell complex).

Since each Ck is a finite dimensional vector space, its dual space C k := (Ck )∗ is also
finite dimensional, too. Just like the dual of vectors is called a covector (linear maps
from vectors to real numbers), the dual of chains is called a cochain (linear maps
from chains to real numbers). In Theorem 3.2 we saw that the adjoint of boundary
operator ∂ is the exterior derivative d , thus gives rise to the cochain complex,
similar to (3.9):
∂ ∗ =d0 ∂ ∗ =d1 ∂ ∗ =dm−1
C 0 −−
1
−→ C 1 −−
2 m
−→ · · · −− −−−→ C m . (3.10)
The k -cells Mk = {c1 , · · · , cN } automatically form a basis of the space of weighted
sums of them, a.k.a. k -chains Ck . Each cell c ∈ Mk corresponds to a linear functional
on k -cochains by the integration pairing:
ˆ
k
c: α ∈ Ω (M ) 7→ α. (3.11)
c

We can therefore view the space of cochains C k as differential k -forms up to minor

difference that cannot be distinguished by integration on k -cells. More precisely,
suppose α, β ∈ Ωk (M ) are two differential k -forms such that their integral values are
the same on all k -cells, ˆ ˆ
α = β, for all c ∈ Ck . (3.12)
c c
For the purpose of any computation on the cell complex, α and β are virtually
indistinguishable. As a result, they can be represented as the same N real numbers:
´  ´
β ´c1 α

´c1
 β   α
 c2   c2 
 ..  =  ..  ∈ RN . (3.13)
 .   . 
´ ´
cN
β cN
α

In other words, α and β belongs to the same equivalence class, and the space of
cochains C k is exactly the collection of these equivalence classes. We can further
derive an entire theory about the exterior calculus of these “differential forms up
to minor differences that cannot be distinguished by the cell complex” (equivalence
classes). Another SIGGRAPH course [Crane(2018)] already covered a wide range
of exterior calculus operations on discrete surfaces and curves and we omit the
discussions here.

3.1.4 Eulerian representations: geometric measure theory

Eulerian representations, contrary to Lagrangian ones, seeks to represent geometries
contained in the world space W as functionals of differential forms.
3.1 Geometry representations 75

Recall that a Dirac-δ measure ρp of a point p ∈ W is an n-form (sometimes also

called a measure) where for any (volumetric) subset A ⊂ W ,
ˆ (
1, if p ∈ A
ρp = (3.14)
A 0, if p ∈
/ A.

The Dirac-δ n-form ρp induces a linear functional T via dual pairing:

linear
T : Ω0 (W ) −−−→ R
ˆ (3.15)
linear
f 7−−−→ f ρp .
W

Readers can verify that this functional T is linear in its input f ∈ Ω0 (W ) by verifying
that T (af + bg) = aT (f ) + bT (g) for a, b ∈ R and f , g ∈ Ω0 (W ).
We can approximate the Dirac-δ n-form by ρϵ defined as
ˆ
Vol(A ∩ Bp (ϵ))
ρϵ = . (3.16)
A Vol(Bp (ϵ))
Here Bp (ϵ) denotes the neighborhood ball centered at point p with radius ϵ > 0. The
induced linear functional on 0-forms is:
linear
Tϵ : Ω0 (W ) −−−→ R
´
fµ (3.17)
f 7−−−→ ´ p
linear B (ϵ)
.
Bp (ϵ)
µ

For continuous functions f , the above corresponds to the average of f on the ϵ-ball
Bp (ϵ) and it converges to the function value f (p) as ϵ → 0.1
The intuition behind Dirac-δ n-form is that each point p ∈ W corresponds to
a linear functional Tp that maps a 0-form f ∈ Ω0 (W ) to its evaluation f (p) at the
point p. A point p is of dimension 0. For geometries of other dimension, we can
similarly define a linear function on differential forms.
Definition 3.6 — Currents / Dirac-δ forms. One way to to represent a k -dimensional
geometry S ⊂ W is by the linear functional TS that maps differential k -forms
Ωk (W ) to their integral on S :
linear
TS : Ωk (W ) −−−→ R
ˆ (3.18)
linear
α 7−−−→ ⟨S |α⟩ = α.
S

We call the functional TS the k -current associated with S . In lieu of the dual
pairing between (n − k )-forms and k -forms:
ˆ
k n−k
α ∈ Ω (W ), β ∈ Ω (W ) 7→ α ∧ β, (3.19)
W

1
This is true for more general, integrable functions for almost all p ∈ W . Interested readers can
read about Lebesgue differentiation theorem.
76 Chapter 3. Geometric Optimizations
we can now view k -dimensional geometries as (n − k )-forms!

∗
{k -dimensional geomeries} ∼
= Ωk (W ) ∼ = Ωn−k (W ). (3.20)

We call the (n − k )-form δS ∈ Ωn−k (W ) satisfying the following equation the

Dirac-δ form associated with S :
ˆ ˆ
α ∧ δS = α. (3.21)
W S

We saw in Theorem 3.2 that the adjoint of the exterior derivative d is the
boundary operator ∂. In the context of currents / Dirac-δ forms, we have a similar
insight.

Theorem 3.3 Let S ⊂ W be a k -dimensional geometry and δS ∈ Ωn−k (W ) be the

associated Dirac-δ form. If we take an exact k -form d α ∈ Ωk (W ) for some smooth
α ∈ Ωk −1 (W ) as the input of the dual pairing, then
ˆ ˆ ˆ ˆ
Stokes
d α ∧ δS = dα = α= α ∧ δ∂S . (3.22)
W S ∂S W

We then apply Leibniz rule (2.59)

d (α ∧ δS ) = d α ∧ δS + α ∧ (−1)(k −1) d δS , (3.23)

| {z }
(n−1)-form

´ ´
Since
´ W
d (α ∧ δ S ) = ∂W
α ∧ δS = 0a , the left hand side of (3.22) turns into
W
α ∧ (−1) d δS . Since the equation holds true for all α ∈ Ωk −1 (W ), we conclude
k

that as an (n − k + 1)-form,

d δS = (−1)k δ∂S . (3.24)

a
In most cases, we can assume S is a bounded geometry, i.e. it doesn’t intersect the boundary
∂W the world space W ; therefore δS |∂M = 0. We will need to use testing smooth forms
α ∈ Ωkc −1 (M ) that vanish at the boundary ∂W / has compact support if S is not as well-
behaved.

R Recall that if we represented the geometry S ⊂ W as a manifold with an

embedding map, the process of taking the boundary ∂S would involve dealing
with the combinatorial structure representing the abstract manifold. It is when
we adopt the Eulerian view of geometry representations that we can use a
simple linear operation d to take the boundary of a geometry.

■ Example 3.2 — Surface area. Let W be the 3-dimensional Euclidean space.

Suppose S ⊂ W is a 2-dimensional geometry (surface) with boundary ∂S . The
surface area of S is the same as the surface integral of the unit normal vector field
3.1 Geometry representations 77
n: S → TW
ˆ ˆ
Area(S ) = 1dA = ⟨n, n⟩, (3.25)
S S

where ⟨·, ·⟩ denotes the Euclidean metric. At any given point p ∈ S , the unit
normal vector np ∈ Tp W is the vector that yields the largest flux, i.e.

⟨np , np ⟩ = max ⟨v, np ⟩. (3.26)

v∈Tp W ∼
=R3

Combining the above equations, we see that surface area is determined by the
maximal flux amongst all unit vector fields
ˆ
Area(S ) = max3
X♭ . (3.27)
X:W →R ,|X|≤1 S

Using the Hodge star and musical isomorphisms associated with Euclidean metric,
we can associate any vector field X: W → R3 with a 2-form α ∈ Ω2 (W ) such that
X = (⋆α)♯ . The constraint that X is a unit vector field is equivalent to

∥α∥max := max |αp | ≤ 1. (3.28)

p∈M

We now write everything using the Dirac-δ 1-form δS representing S :

ˆ
Area(S ) = max α (3.29)
α∈Ω2 (W ),∥α∥max ≤1 S
ˆ
= 2
max α ∧ δS (3.30)
α∈Ω (W ),∥α∥max ≤1 W
=: ∥δS ∥mass . (3.31)

The last two lines and is the same as saying ∥ · ∥mass is the dual norm of ∥ · ∥max .
By combining the boundary operator from Theorem 3.3 and the above deriva-
tion, we can now solve the classical Plateau problem of finding surface S of
minimal surface area subject to a fixed boundary curve Γ ⊂ W :

minimize Area(S )
S ⊂W : a surface
(3.32)
subject to ∂S = Γ.

The equivalent problem written in terms of Dirac-δ forms look like

minimize ∥η∥mass
η∈Ω1 (W ): a 1-form
(3.33)
subject to d η = δΓ .
■
 78 Chapter 3. Geometric Optimizations

Figure 3.3 Minimal surfaces computed using Eulerian representation as described

by Problem 3.33 [Wang and Chern(2021)]. Unlike Lagrangian rep-
resentation that requires tricky combinatorial effort in computing
surface boundary, problems formulated using Eulerian representations
only require a linear equation when assigning the boundary!

3.2 A Geometric view of optimization problems

Optimization is a type of math problems commonly found in engineering and science.
The most basic optimization problem was perhaps introduced in one’s first Calculus
class.
Problem 3.1 — Optimization on the real line. Given an objective function f : R → R,
find the optimal input x opt ∈ R such that f attains the minimal value at x opt , i.e.

f (x opt ) = min f (x ). (3.34)

x ∈R

We learned in Calculus that for differentiable f , a necessary condition that an input

x0 is optimal is that the first derivative of f attains zero:

f ′ (x0 ) = 0. (3.35)

The above equation is referred to as the first order optimality condition for the
optimization problem (3.34). If f is twice differentiable, then the first order optimality
condition together with the inequality

f ′′ (x0 ) ≥ 0 (3.36)

ensures that x0 is a local minimum, which means all nearby points x ∈ (x0 − ϵ, x0 + ϵ)
are less optimal than (or equally optimal to) x0 , i.e. f (x ) ≤ f (x0 ).
 3.2 A Geometric view of optimization problems 79

f (x0 )
f ′ (x0 ) = 0

x0 − ϵ x0 x0 + ϵ
In this section, we explore the geometric picture of optimization problems: re-
placing real line with finite-dimensional manifolds or even function spaces, adding
equality constraints, adding inequality constraints, the geometric meaning of La-
grange multipliers, and iterative numerical methods for solving optimization methods.
We focus on deriving the necessary condition for an optimal solution, similar to
(3.35), and leave the discussions of the sufficient condition for other textbooks, e.g.
[Boyd and Vandenberghe(2004)].

3.2.1 Unconstrained problems on (finite-dimensional) manifolds

Problem 3.2 M is an n-dimensional manifold and f : M → R is a function on M .
An unconstrained optimization problem on the manifold M is about finding a point
p ∈ M such that f (p) attains minimal value:

minimize f (p). (3.37)

p∈M

The locally optimal solution p ∈ M must satisfy that all its neighboring points
are less optimal than (or equally optimal to) p. This property can be captured by the
property that if γ: [−ϵ, ϵ] → M is a path in M passing through the point p = γ(0),
γ f
then the function F = f ◦ γ: [−ϵ, ϵ] → − R takes a local minimum when s = 0
− M →
and γ(s) = p. Apply the first order optimality condition (3.35), we get

d
0 = F ′ (0) = f ◦ γ(s)|s=0 = dfγ(0) Jγ̇(0)K = dfp Jv K, (3.38)
ds
where v ∈ Tp M = γ̇(0) is the tangent of the path γ at the point p. Since this has to
hold true for all path γ at all tangent direction v ∈ Tp M , we conclude that the first
order condition for a local optimal solution p is the same as the differential df of the
objective function f is zero as a 1-form at p:

dfp = 0. (3.39)

3.2.2 Unconstrained problems on functions and differential forms

We now take a look at unconstrained optimization problems on the space of functions
and differential forms Ωk (M ). When k = 0, the study of optimization problems
on the space of functions Ω0 (M ) is often called Calculus of Variations. One nice
property about differential forms is that they have a linear structure. To perturb an
input α ∈ Ωk (M ), one can simply pick another k -form α̊ ∈ Ωk (M ) and add it to α.
In Figure 3.5, we visualize this concept for 0-forms a.k.a. functions.
80 Chapter 3. Geometric Optimizations

u
u + ů

[ ]
a ů b

Figure 3.4 The function space V = Ω0 ([a, b]) is a vector space. With any function
u: [a, b] → R, we can always take another function ů: [a, b] → R and
add it to u to get a modified function u + ů: [a, b] → R. Effectively,
the tangent space at u is the entire function space Tu V = V.

Problem 3.3 M is an n-dimensional manifold and V = Ωk (M ) is the vector space of

all k -forms defined on M . Let E: V → R be a functional2 defined on V. We use the
notation E because a common origin of optimization problems on function spaces
is about minimizing an energy functional. An unconstrained optimization problem
on k -forms V = Ωk (M ) is about finding an optimal α ∈ V such that E[α] attains
minimal value:
minimize E(α). (3.40)
α∈V

The differential d E at an input α ∈ V is a covector d E|α ∈ V ∗ . We can understand

this infinite dimensional covector by pairing it with a tangent vector α̊ ∈ Tα V = V:
d
d Eα Jα̊K = d Eα α̊ := E(α + ϵα̊). (3.41)
dϵ ϵ=0

Each tangent vector α̊ ∈ Tα V = V acts on α by varying its value αp to αp + α̊p

at any point p ∈ M . Hence we call the tangent vector α̊ a variation. We use the
notation of α̊ to indicate that it is a variation in the object α. Likewise, we sometimes
write

E̊ = ⟨d Eα |α̊⟩ (3.42)

to denote the corresponding variation in E.

The locally optimal solution α ∈ V must satisfy that α + ϵα̊ is less optimal than
(or equally optimal to) α for small ϵ > 0. As a result, the function F (s) = E(α + sα̊)
must attain local minimum when s = 0 and F (0) = E(α). By (3.35),

d
0 = F ′ (0) = E(α + ϵα̊) = d Eα Jα̊K. (3.43)
dϵ ϵ=0

Since this has to hold true for all variations α̊ ∈ V, we conclude that the first order
condition for a local optimal solution α is the same as the differential d E of the
2
A functional is just another name for function with an emphasis that it may be a function over
a space of functions.
3.2 A Geometric view of optimization problems 81

energy functional E vanishes at α as a covector:

d E|α = 0 ∈ V ∗ . (3.44)

■ Example 3.3 Let V = {u : [a, b] → R} = Ω0 ([a, b]). Define E : V → R as

´b linear
E(u) = a
u 2 (x ) dx . The differential d E|u : V −−−→ R is
ˆ b
d
d Eu JůK = (u(x ) + ϵů(x ))2 dx (3.45)
d ϵ ϵ=0 a
ˆ b
= 2(u(x ) + ϵů(x ))ů(x ) dx (3.46)
a ϵ=0
ˆ b
= 2u(x )ů(x ) dx . (3.47)
a

Example 3.4 — Dirichlet energy. Let V = Ω0 ([a, b]). Define E : V → R as

´b
■

E(u) = a 21 u ′ (x )2 dx , which is often called the Dirichlet energy. We can use

linear
integration by parts to understand the differential d E|u : V −−−→ R,
ˆ b
d
d Eu JůK = (u ′ (x ) + ϵů ′ (x ))2 dx (3.48)
d ϵ ϵ=0 a
ˆ b
= u ′ (x )ů ′ (x ) dx (3.49)
a
ˆ b
IbyP ′ ′
= u (b)ů(b) − u (a)ů(a) − u ′′ (x )ů(x ) dx . (3.50)
a

The expression after the integration by parts appear more directly as an L2 linear
pairing (see Definition 2.22) with ů together with some finite dimensional dot
product pairing at the boundary. ■

∗
d E|α is the differential of E at α, which is a covector in V ∗ = Ωk (M ) ∼


= Ωn−k (M )
3
. It is not to be confused with the gradient of the functional grad E|α ∈ V = Ωk (M ).
It is also not to be confused with the differential of the input variable d α ∈ Ωk +1 (M ).
Let ♭V : V → V ∗ , or “uncurried” as ⟪·, ·⟫V = ♭V (·)(·), be some metric (inner product
structure) defined on V. Then we have a unique vector grad E|α := d E|♯αV defined
such that

⟪grad E|α , α̊⟫ = d E|α Jα̊K ∀ α̊ ∈ V. (3.51)

´b
■ Example 3.5 Let V = {u : [a, b] → R}. Define E : V → R as E(u) = u 2 (x ) dx ,
´b a
whose variation is known as d Eu JůK = a
2u(x )ů(x ) dx . If we define an inner

3
See Theorem 2.44.
82 Chapter 3. Geometric Optimizations
product structure by
ˆ b
⟪u, v ⟫ := u(x )v (x )m(x ) dx (3.52)
a

using some given positive function m : [a, b] → R>0 . Then the gradient of E is
given by

2u(x )
grad E|u (x ) = . (3.53)
m(x )
■

The domain for E does not need to be a vector space. One typical extension is
that A is an affine space parallel to V. That is, V is the tangent space of A at any
point α ∈ A. Then d E|α is a covector in V ∗ , and grad E|α is a vector in V. The first
order optimality condition for a minimization problem on an affine space A is the
same: the differential d E|α is zero as a covector in V ∗ .

u u(b) = cb
u(a) = ca
u + ů

ů(a) = 0
[ ]
a ů b ů(a) = 0

Figure 3.5 The function space with boundary conditions A = {u ∈

Ω0 ([a, b]), u(a) = ca , u(b) = cb } is an affine space. With any function
u ∈ A, we can always take another function that vanishes on the
boundary ů ∈ V = {ů: [a, b] → R,ů(a) = ů(b) = 0} and add it to u
to get a modified function u + ů: [a, b] → R. Effectively, the tangent
space at u is the space of functions with vanishing boundary values
Tu A = V.

■ Example 3.6 — Dirichlet energy with boundary conditions. Let A = {u : [a, b] →

R, u(a) = ca , u(b) = cb } for some constants ca , cb . Note that A is an affine space.

That is, the difference between any two elements in A form a vector space
V = {ů : [a, b] → R,ů(a) = ů(b) = 0} which is closed under additions and scalar
´b
multiplications. Let E : A → R be the Dirichlet energy E(u) = a 12 u ′ (x )2 dx for
u ∈ V. (3.50) states that
ˆ b
d Eu JůK = − u ′′ (x )ů(x ) dx . (3.54)
a
´b
Consider the m(x )-weighted inner product ⟪ů,v̊ ⟫ = a
ů(x )v̊ (x )m(x ) dx for
3.2 A Geometric view of optimization problems 83
ů,v̊ ∈ V and write
ˆ b
−u ′′ (x ) −u ′′
 
d Eu JůK = ů(x )m(x )dx = ⟪ ,ů⟫ . (3.55)
a m(x ) m}
| {z
grad Eu

The gradient grad Eu ∈ V as a function is

u ′′ (x )
grad Eu (x ) = − , x ∈ [a, b]. (3.56)
m(x )
■

■ Example 3.7 Let L : R × R → R, L = L(u1 , u2 ) be some expression. Let its

differential be denoted by
∂L ∂L
dL = du1 + du2 . (3.57)
∂u1 ∂u2
Now, consider A = {y : [a, b] → R | y(a) = ya , y(b) = yb } for some given ya , yb .
Note that A is an affine space with tangent vector space V = {ẙ : [a, b] →
R |ẙ(a) = ẙ(b) = 0}. Let E : A → R,
ˆ b
E(y) := L (y(x ), y ′ (x )) dx . (3.58)
a

Then its variation is

ˆ b 
∂L ∂L
d E|y JẙK = ′
(y(x ), y (x ))ẙ(x ) + ′ ′
(y(x ), y (x ))ẙ (x ) dx (3.59)
a ∂u1 ∂u2
ˆ b  
∂L d ∂L
= ′
(y(x ), y (x )) − ′
(y(x ), y (x )) ẙ(x ) dx . (3.60)
a ∂u1 dx ∂u2

A common shorthand notation is

ˆ b  
∂L d ∂L
d E|y JẙK = − ẙ dx . (3.61)
a ∂y dx ∂y ′
■

In a more general case, the domain of E is an infinite dimensional manifold M.

Its variation is a covector field d E ∈ Γ(T ∗ M) and its gradient with respect some
inner product structure is a vector field grad E ∈ Γ(T M).
Problem 3.4 M is an infinite dimensional manifold and E: M → R is a function
defined on M. The following defines an unconstrained optimization problem on M:

minimize E(y). (3.62)

y∈M

The necessary condition for a locally optimal solution y ∈ M is that variation of

 84 Chapter 3. Geometric Optimizations

the energy along any variation ẙ ∈ Ty M must vanish:

d E|y JẙK = 0, for all ẙ ∈ Ty M, (3.63)

or simply

d E|y = 0 ∈ Ty∗ M. (3.64)

Using any inner product, this vanishing variation condition becomes

grad E|y = 0. (3.65)

We call (3.65) the Euler–Lagrange equation. It is not a surprise that the optimality
condition for problems on finite dimensional manifold (Problem 3.2) and problems on
infinite dimensional manifold (Problem 3.4) look almost identical, but we usually use
Euler-Lagrange equation to refer to the metric-informed equation (3.65) for infinite
dimensional cases.
In the sections to follow, we always denote our optimization variable as x ∈ M
where M can be a finite- or infinite-dimensional manifold.

3.2.3 Optimization with equality constraints

An optimization with equality constraints takes the following general form. Let
M be a manifold, F : M → R be the objective function, and let G : M → U be a
constraint function for some vector space U. The optimization problem is stated as

minimize F(x )
x ∈M
(3.66)
subject to G(x ) = 0U .

Here 0U simply denotes the zero element in the vector space U. We assume both F
and G are smooth. The necessary condition for the optimal solution x is

d F|x Jx̊ K = 0 for all x̊ ∈ (d G|x )⊥ (3.67)

where (d G|x )⊥ denotes the annihilator subspace

(d G|x )⊥ = x̊ d G|x Jx̊ K = 0 . (3.68)



linear
Note that the type of d G|x is d G|x : Tx M −−−→ U. The condition implies that there
exists a Lagrange multiplier λ0 ∈ U ∗ such that G|x = 0 and
D E
d F|x + λ0 d G|x = 0 ∈ Tx M. (3.69)
U ∗ ×U

3.2.4 Karush–Kuhn–Tucker conditions

Let us look at optimization problems with both equality constraints and inequality
constraints. The energy function is still given as a function F : M → R on a general
manifold M. The equality constraints are given through a function G : M → U to a
vector space U. The inequality constraints are defined by a (proper) convex cone
 3.3 Computational methods for optimization problems 85

C ⊂ V in a vector space V and a function H : M → V. The optimization problem

takes the following form:

minimize

 F(y) subject to
y∈M
G(y) = 0U (3.70)


H(y) ∈ C.

The necessary stationary conditions for an optimal solution y0 ∈ M are the following
Karush–Kuhn–Tucker (KKT) conditions. There exist Lagrange multiplies λ0 ∈ U ∗ and
n o
µ0 ∈ C := µ ∈ V ⟨µ|v ⟩V ∗ ×V ≤ 0 for all v ∈ C ⊂ V ∗
◦ ∗
(3.71a)

(C ◦ is called the polar cone of C) such that G|y0 = 0, H|y0 ∈ C,

D E D E
d F|y0 + λ0 d G|y0 + µ0 d H|y0 = 0 ∈ Ty0 M, (3.71b)
U ∗ ×U V ∗ ×V

and
D E
µ0 H(y0 ) = 0. (3.71c)
V ∗ ×V

The last condition (3.71c) is called the complementary slackness condition. It

implies that if H(y0 ) is the interior of the cone C, then the Lagrange multiplier
term is deactivated µ0 = 0. Only when H(y0 ) lies on the boundary of the cone,
we see an associated “normal force” µ0 emerges to keep H(y0 ) in the interior of
the cone. Condition (3.71c) is the pullback of the corresponding complementary
slackness condition about the convex cone on V back to M via the map H. In (3.71b),
the 2nd and 3rd terms can also be interpreted as the pullbacks of the covectors
λ0 ∈ U ∗ , µ0 ∈ V ∗ via G and H respectively.

3.3 Computational methods for optimization problems

3.3.1 Direct and iterative methods for unconstrained problems
For an unconstrained problem on a (finite- or infinite-dimensional) manifold M

minimize F(x ), (3.72)

x ∈M

there are two ways to find the optimal solution. The first method is to solve the
optimality condition directly

d F|x = 0 ∈ Tx M. (3.73)
linear
The second method is to find a metric ♯ : T ∗ M −−−→ TM and solve the gradient
flow: starting with an initial guess x0 ∈ M, the path x : [0, ∞) → M satisfying the
following differential equation

ẋ (t) = −(d Fx (t) )♯ = − grad Fx (t) (3.74)

86 Chapter 3. Geometric Optimizations

will4 converge to the stable point x opt ∈ M with vanishing the gradient

(d F|x opt )♯ = grad F|x opt = 0. (3.75)

Numerically, this is often done via a forward Euler scheme, where a sequence of
intermediate states x (k ) is defined iteratively

x (0) = x0 (initial guess)

(3.76)
x (k +1) = x (k ) − ∆t (k ) (d F|x (k ) )♯ ,

where ∆t (k ) > 0 denotes the k -th time step, which can be a fixed value or determined
via line search. The iterative scheme (3.76) is often called the Gradient Descent
method.
■ Example 3.8 — An extended discussion on Dirichlet energy from Example 3.6.
Let A = {u : [a, b] → R, u(a) = ca , u(b) = cb } for some constants ca , cb . Consider
´b
the Dirichlet energy E(u) = a 12 u ′ (x )2 dx . We saw in Example 3.6 that
ˆ b ˆ b
d Eu JůK = ′ ′
u (x )ů (x ) dx = − u ′′ (x )ů(x ) dx , (3.77)
a a

and the gradient grad Eu associated with the m(x )-weighted inner product ⟪ů,v̊ ⟫L2 =
´b
a
ů(x )v̊ (x )m(x )dx is
u ′′ (x )
grad Eu (x ) = − . (3.78)
m(x )
Using the direct method, we can solve the optimality condition (also called
Euler-Lagrange equation) 
′′
u (x ) = 0

u(a) = ca (3.79)

u(b) = cb .


Alternatively, one can use the gradient flow method, and solve the differential
equation for u: [a, b] × [0, ∞) → R
u ′′ (x ,t)
(
d
u(x , t) = − grad E u(·,t) =
dt m(x )
(3.80)
u(x , 0) = u0 (x ),

where u0 (x ) ∈ A is an arbitrary initial function.

Interestingly, (3.79) corresponds to solving a Laplace problem with specified
boundary value, while (3.80) corresponds to solving the heat equation with an
arbitrary initial value. Both of them gives a harmonic function with the specified
boundary value! ■

4
The convergence to a stable point often requires additional assumptions on the objective
function F such as convexity, superlinearity, etc.. Thankfully most objective functions in computer
graphics do possess these nice properties.
3.3 Computational methods for optimization problems 87

■ Example 3.9 — Newton’s method. Let M be an m-dimensional manifold and

f : M → R be an objective function and we want to solve the following optimization
problem:
minimize f (x ). (3.81)
x ∈M

Suppose M has a coordinate system (x1 , · · · , xm ): M → Rm that are not necessarily

orthonormal. Given a point x ∈ M and a tangent x̊ ∈ Tx M , the differential can
be written in terms of the coordinates:
˚
f = dfx Jx̊ K (3.82)
 
x̊1
∂f ∂f  .. 
= (x )x̊1 + · · · + (x )x̊m , where x̊ =  .  (3.83)
∂x1 ∂xm
x̊m

We need a metric ⟨·, ·⟩ to define the gradient flow. Write the metric as a matrix-
valued function µ = µij : M → Rm×m using the coordinates:
      
* x̊1 ẙ1 + µ11 (x ) · · · µ1m (x ) ẙ1
 ..   ..   . . .
..   ...  (3.84)

 .  ,  .  = x̊1 · · · x̊m  .. ..  

x̊m ẙm µm1 (x ) · · · µmm (x ) ẙm

We can use the usual matrix algebra and compute

 
iT x̊1
dfx Jx̊ K = ∂x1 (x ) · · · ∂xm (x )  ... 
h
∂f ∂f
(3.85)
 

x̊m
 −1  ∂f   
* µ11 (x ) · · · µ1m (x ) ∂x1
(x ) x̊1 +
=  ... .. ..   .. ,  ..  . (3.86)

. .   .   . 
∂f
µm1 (x ) · · · µmm (x ) ∂xm
(x ) x̊m
| {z }
gradµ f |x

Since the original optimization Problem 3.81 doesn’t concern to assign a metric
to the manifold M , we can choose a metric ourselves such that the matrix µ(x )
coincides with the Hessian of f :
  ∂2f ∂2f

(x ) · · · (x )

µ11 (x ) · · · µ1m (x ) 2 ∂x1 ∂xm
 .. . .  ∂x1 . . .. (3.87)

 . .. ..  = 
  .. .. .
 =: Hf (x ).

2 2
µm1 (x ) · · · µmm (x ) ∂ f
(x ) · · · ∂ f
(x )
∂xm ∂x1 ∂x 2 m
88 Chapter 3. Geometric Optimizations
Optimization methods based on the gradient flow of this Hessian metric are called
Newton’s method:  ∂f 
∂x1
(x )
ẋ = Hf (x )−1  ...  . (3.88)
 
∂f
∂xm
(x )
■
 4. Continuum Mechanics

Continuum mechanics is the study of the statics and dynamics of deformable bodies.
These deformable bodies can behave like elastic solid, Newtonian fluid, elastoplastic
solid, viscoplasticity (non-Newtonian) fluid, ferrofluid, or plasma. Which behavior
should a deformable body follow depends on a constitutive model. These models boil
down to describing how a potential energy is defined as a function of the deformation
of the body, and how a dissipation function depends on both the deformation and
the rate of deformation of the body. The equations of motion are derived from the
model by variational principle.

• (Hyper)elasticity The potential energy is only a function of the metric induced

by the deformation from the world metric. Examples include elastic bodies
such as jelly, rubber, etc. Examples also include most solids undergone small
deformations or rigid motions.

• Newtonian fluid The potential energy is only a function of the induced change
of volume by the deformation; the dissipation is only a function of the rate of
deformation.

• Elastoplasticity An elasticity setup but the system carries extra variables that
encode a reference metric for rest state, and the dissipation function depends
on the rate of change of the rest metric. Examples include solids undergone
irreversible plastic deformation, hardening, and fracturing.

• Viscoplasticity In addition to elastoplasticity setup, the fluid carries a con-

vected tensor field describing the statistics of the direction microscopic anisotropic
molecules. The dissipation also depends on the rate of change of the defor-
mation. Examples are fluids consisting long polymers such as starch water
mixture.
 90 Chapter 4. Continuum Mechanics

• Ferrofluid A Newtonian fluid that carries an additional convected field of mag-

netization, adding magnetic terms to the potential energy and the dissipation
function. Examples are ferromagnetic liquids.

• Magnetohydrodynamic fluids Magnetohydrodynamic fluids are fluids with

free electrons, i.e. they are electrically conductive fluids. In particular, such
a continuum is a Newtonian fluid that carries trapped magnetic field due to
the electric conductivity of the fluid. This adds electromagnetic term to the
potential energy. Examples include liquid metals and ionized gas (plasmas).

The advantage of employing exterior calculus in formulating continuum mechanical

equations are

• Strong type system for tensors: Instead of identifying tensors as matrices or

arrays of various sizes, a stronger type system is able to distinguish the type
of, for example, stress and strain, which have different geometric and physical
meanings, even though their matrix representations are the same size.

• More fundamental and simpler understanding of the relationship between

objects: The language sorts out that many objects in continuum mechanics
are related by the canonical pullback operator, which is an important insight
when we need to invoke, for example, the commutativity between pullback and
exterior derivative.

• Exterior calculus includes the Lie derivative, which is fundamental for doing
calculus on deformations. Without the vocabulary of Lie derivatives, differential
equations that describe conservation laws are broken into many obscure terms.

• Theorems such as symmetry of stress tensors, conservation of circulation in

inviscid fluid etc can be reasoned just by type checking.

In this chapter, we will only focus on hyperelasticity.

4.1 Deformation map

A universal postulate for a continuum mechanical system is that the state of the
system is a map representing the shape of a deformable body.
Postulate 4.1 — Continuum mechanical system. We have the following setups for
a continuum mechanical system.

• M : a manifold representing material domain or Lagrangian coordinate.

• W : a manifold representing world coordinate or Eulerian coordinate.

• The state of deformation is described by a (time-dependent) embedding

ϕ = ϕ(t) : M → W . We call it deformation map or flow map.

• M is equipped with a time-independent ρM ∈ Ωn (M ) representing mass.

 4.2 Elasticity overview 91
• W is equipped with a time-independent metric ♭W .
The material domain M is typically Rn but sometimes just an index set of particles.
When M = R3 coordinate system is denoted by (X , Y , Z ) or interchangeably
(X 1 , X 2 , X 3 ). The world domain W is typically Rn . When W = R3 coordinate
system is denoted by (x , y, z ) or interchangeably (x 1 , x 2 , x 3 ).
Using the above objects, we can define the kinetic energy of a moving ob-
ject.
Definition 4.1 — Kinetic energy. The kinetic energy of a deformation map is given
by
ˆ
1 2
K(ϕ, ϕ̇) := |ϕ̇| ρM . (4.1)
M 2 ♭W

In a conservative Newtonian mechanical system, we can derive the equations of

motion by the kinetic energy K(ϕ, ϕ̇) and the potential energy U(ϕ) of the system,
both of which are functions of our state variable ϕ (and its time derivative ϕ̇ for the
kinetic energy). Explicitly, the equation of motion is given by

∇ ∂K
= −d U. (4.2)
dt ∂ ϕ̇

This equation is the critical point of the action

ˆ T 
S(ϕ) = K(ϕ, ϕ̇) − U(ϕ) dt. (4.3)
0

over the space of paths ϕ : [0, T ] → (M → W ) on the space (M → W ) of deforma-

tions with fixed initial and final positions.
To describe elasticity, we only need a minimal assumption on the form of the
potential energy U (Section 4.4). The remaining derivation for the equation of motion
is to evaluate the differential d U with respect to ϕ. This calculation will lead to the
emergence of the notion of stress with delicate tensor types (Section 4.5).
Before carrying out this calculation in Section 4.5, we first give an overview of the
classical exposition of elasticity in Section 4.2, followed by a preliminary on tensor
algebra in Section 4.3.

4.2 Elasticity overview

In this section, we give a quick overview of the classical exposition of elasticity. This
exposition is based on how the system was discovered and explained, instead of how it
could have been derived by calculating d U. In this section, we will introduce several
many matrices in this section, some of which represent strain and others represent
stress of various kinds. Later in Section 4.5, we will re-introduce these tensors using
a principled argument, from which we discover a more delicate understanding of
these tensors.
When studying elasticity, we first introduce the concept of stress. Phenomeno-
logically, a solid body has internal force. In the early 19th century, Augustin-Louis
92 Chapter 4. Continuum Mechanics

Cauchy showed that the internal force T acting on an infinitesimal internal surface,
known as the traction force, is a linear function of the surface normal. So we can
write T = σn where n is the surface normal, and σ is the matrix representing this
linear function. This matrix σ is called the Cauchy stress.
The net force acting on an infinitesimal volume is the given by the sum of
traction force on the volume’s boundary. Hence, the net force density is given
by the divergence fnet = ∇ · σ of the Cauchy stress. When the solid is in static
equilibrium, one can show that σ must be a symmetric matrix. However, in fact,
σ is symmetric even under accelerated motion for a different reason (under the
assumption of hyperelasticity).
Hyperelasticity is the model for large deformation. When it comes to this large
deformation, one starts to introduce many more other alternative stress measures.
Only writing a few major ones, we have the Kirchhoff stress τ , the first Piola–Kirchhoff
stress P, the second Piola–Kirchhoff stress S, the nominal stress N = P⊺ . Their
conversions are given by
Equation
for σ τ P S
σ= σ 1
J
τ 1
J
PF⊺ J −1 FSF⊺
⊺
τ = Jσ τ PF FSF⊺
P= J σF−⊺ τ F−⊺ P FS
S= J F−1 σF−⊺ F τ F−⊺
−1
F−1 P S

where the deformation gradient F = ∇ϕ is Jacobian matrix of the deformation map

ϕ, and J = det(F). A stress–strain relation is given by some function S = S(C)
where C = F⊺ F is called the Cauchy–Green tensor.
Where do all these stresses come from? Also, they are all represented by square
matrices of the same size, which are inherently coordinate dependent and there is no
information of what type they are (are they endomorphism or bilinear form, or some
other type?) What are their geometric intuition?
In fact, we do see something familiar in this conversion table. Recall Example 2.3
that the pullback of a 2-form in 3D involves J F−1 . In particular, the relationship
between P and σ looks like a pullback for 2-forms. This conversion relation between
P and σ hints that, perhaps, both of them are 2-form. If we think about it, we realize
that of course! σ is a force-(co)vector valued 2-form. The Cauchy stress represents
the traction force of a given infinitesimal surface. Instead of representing the traction
force as σn using a normal vector, a more geometrically accurate description is
that the force is given by σJX , Y K on an infinitesimal parallelogram spanned by
vectors X , Y . In that case, the 1st Piola–Kirchhoff stress P is merely the change of
coordinate to the material coordinate

P = ϕ∗ σ. (4.4)
2-form

From this example, we see that the theory of elasticity would be much more natural
when written in exterior calculus.
What is more, all of the above stress tensors arise automatically and carry specific
differential form types when we try to derive d U, essentially making exterior calculus
unavoidable. Following this first-principle derivation, we do not need any prerequisite
 4.3 Tensor Algebra 93

phenomenological knowledge about the existence of internal traction force, Cauchy

stress, or any other alternative stresses. They are just the results of variational
calculus. Using representations other than exterior calculus remove some information
about the physical and geometric nature of these stress tensors.

4.3 Tensor Algebra

What are tensors? In geometry, vectors, covectors, endomorphisms, bilinear forms,
differential forms, etc., are all special cases of tensors. When these objects are
written in basis or coordinate, they become arrays of numbers. For example, vectors,
covectors and (n − 1)-forms are represented as Rn under basis; endomorphisms and
bilinear forms are both represented as square matrices Rn×n .
For convenience, many introductions of tensors classify tensors by their array
size. Tensors that are represented by 1D array Rn of numbers are called order-1
tensors; tensors that are represented by a matrix Rn×n of numbers are called order-2
tensors. More generally, order-k tensors are arrays whose entries can be querried
by k indices. However, just as the distinctions between vectors and covectors are
important, describing tensors by their array size does not give enough information of
their geometric meaning.
Here we define tensors with the starting point of vector spaces and vector bundles
as building blocks. The construction of tensors retain the information of the type of
these vector spaces (bundle).

4.3.1 Tensor product

The notion of tensor product allows us to combine two vector spaces and produce a
new vector space.
Definition 4.2 — Tensor product space. Let U , V be two vector spaces. Consider
a new vector space, denoted by U ⊗ V , and a non-degenerate bilinear function
called tensor product
bilinear
(·) ⊗ (·) : U × V −−−−→ U ⊗ V . (4.5)

By the bilinearity we have rules such as

(c⃗u ) ⊗ ⃗v = ⃗u ⊗ (c⃗v ) = c(⃗u ⊗ ⃗v ) for all ⃗u ∈ U , ⃗v ∈ V and c ∈ R. (4.6)

and by the non-degeneracy ⃗u1 ⊗ ⃗v and ⃗u2 ⊗ ⃗v are linearly independent in U ⊗ V

whenever ⃗u1 , ⃗u2 are linearly independent in U .
Concretely, let (⃗ei ∈ U )i∈I be a basis for U and let (⃗fj ∈ V )j ∈J be a basis for V .
Then the vector space U ⊗ V is the linear combination of ⃗hij := ⃗ei ⊗ ⃗fj as
Pbasis vectors
with indices running over all i ∈ I , j ∈ J . The tensor product of ⃗u = i∈I u i ⃗ei ∈ U
and ⃗v = j ∈J v j ⃗fj ∈ V is given by
P

XX
⃗u ⊗ ⃗v := u i v j ⃗hij . (4.7)
i∈I j ∈J
 94 Chapter 4. Continuum Mechanics

In particular,

dim(U ⊗ V ) = dim(U ) dim(V ), (4.8)

in contrast to dim(U × V ) = dim(U ) + dim(V ).

A general element A ∈ U ⊗ V takes the form
XX
A= Aij ⃗hij . (4.9)
i∈I j ∈J

Here A = (Aij )i∈I ,j ∈J is the matrix representation of the tensor A.

Definition 4.3 — Decomposable tensor. An element A in U ⊗ V is said to be
decomposable if there exist ⃗
u ∈ U and ⃗v ∈ V so that A = ⃗u ⊗ ⃗v . In a basis
representation, decomposability means Aij = u i v j , i.e. the matrix A = (Aij )ij can
be expressed as an outer product A = uv⊺ .
A generic element in U ⊗ V is not decomposable, unless either U or V is a one-
dimensional vector space.
linear
■ Example 4.1 The space of endomorphisms End(U ) := {A : U −−−→ U } on a
vector space U is the tensor product space End(U ) = U ∗ ⊗ U . An endomorphism
is a vector-valued covector, as it takes in a vector and spits out a vector. ■

linear bilinear
■ Example 4.2 The space of bilinear forms {B : U −−−→ U ∗ } = {B : U ×U −−−−→
R} on a vector space U is the tensor product space U ∗ ⊗ U ∗ . ■

Definition 4.4 — Tensor product bundle. Let E1 , E2 be two vector bundles defined
over a common manifold M . Then E1 ⊗ E2 is a new vector bundle over M where
each fiber (E1 ⊗ E2 )p is defined by the tensor product space (E1 )p ⊗ (E2 )p for each
p ∈ M.

4.3.2 Tensor power

A special case of constructing tensor product spaces is taking tensor product between
the same vector space. Let U be a vector space. Then we may take U ⊗ U =: ⊗2 U
and continue tensoring more copies of the same space ⊗3 U := U ⊗ U ⊗ U and so on.

Definition 4.5 — Tensor power space. We recursively define ⊗k U by ⊗0 U :=

R, ⊗1 U = U , and a non-degenerate bilinear function
bilinear
⊗ : ⊗k U × ⊗ℓ U −−−−→ ⊗k +ℓ U (4.10)

such that it is associative

(A ⊗ B ) ⊗ C = A ⊗ (B ⊗ C ) for all A ∈ ⊗k U , B ∈ ⊗ℓ U , C ∈ ⊗m U . (4.11)

4.3 Tensor Algebra 95

■ Example 4.3 Space of bilinear forms on U is the tensor space ⊗2 U ∗ . ■

Under a basis for U and its dual basis for U ∗ , bilinear forms on U is represented
by a square matrix with size n-by-n where n = dim(U ). Just as the consideration
of symmetric matrices and skew-symmetric matrices are important when studying
bilinear forms, we should also discuss the symmetric and skew-symmetric tensor
powers.
Using a similar formal definition we define the following special spaces of tensor
power spaces.

Definition 4.6 — Symmetric tensor power space. We recursively define ⊙k U by

⊙0 U := R, ⊙1 U = U , and a non-degenerate bilinear function
bilinear
⊙ : ⊙k U × ⊙ℓ U −−−−→ ⊙k +ℓ U (4.12)

satisfying

• associativity:

(A ⊙ B ) ⊙ C = A ⊙ (B ⊙ C ) for all A ∈ ⊙k U , B ∈ ⊙ℓ U , C ∈ ⊙m U .
(4.13)

• symmetry:

⃗u ⊙ ⃗v = ⃗v ⊙ ⃗u for all ⃗u , ⃗v ∈ U . (4.14)

The space ⊙k U is included in ⊗k U by the inclusion map for decomposable

symmetric tensors:

⊙k U ,→ ⊗k U
X
⃗u1 ⊙ · · · ⊙ ⃗uk 7→ ⃗uσ(1) ⊗ · · · ⊗ ⃗uσ(k ) (4.15)
σ∈Sk

where Sk is the permutation group on {1, . . . , k }.

Definition 4.7 — Skew-symmetric tensor power space (Exterior algebra revisited).

We recursively define ∧k U by ∧0 U := R, ∧1 U = U , and a non-degenerate bilinear
function
bilinear
∧ : ∧k U × ∧ℓ U −−−−→ ∧k +ℓ U (4.16)

satisfying

• associativity:

(A ∧ B ) ∧ C = A ∧ (B ∧ C ) for all A ∈ ∧k U , B ∈ ∧ℓ U , C ∈ ∧m U .
(4.17)
 96 Chapter 4. Continuum Mechanics
• skew-symmetry:

⃗u ∧ ⃗u = 0 for all ⃗u ∈ U . (4.18)

The space ∧k U is included in ⊗k U by the inclusion map for decomposable

symmetric tensors:

∧k U ,→ ⊗k U
X
⃗u1 ∧ · · · ∧ ⃗uk 7→ (−1)|σ|⃗uσ(1) ⊗ · · · ⊗ ⃗uσ(k ) (4.19)
σ∈Sk

where Sk is the permutation group on {1, . . . , k }, and (−1)|σ| is +1 (resp. −1)

when σ is an even (resp. odd) permutation.

4.3.3 Pullback bundle

Suppose we have a deformation map ϕ : M → W . What is the type of its temporal
derivative ϕ̇ = ∂t
∂
ϕ and its spatial derivative d ϕ? The object ϕ̇ is to be evaluated at
a point in M , but its value is a vector in TW . The object d ϕ can take a vector at a
point in M , but spits out a vector in TW . The subtlety here is that TW is not a
bundle over M .
To characterize the type of objects such as ϕ̇ and d ϕ, we need the concept of
pullback bundle.
Definition 4.8 — Pullback bundle. Let E be a vector bundle over W and let
ϕ : M → W be a map. Then there induces a vector bundle Eϕ over M where each
fiber (Eϕ )p , p ∈ M , is defined by

(Eϕ )p := Eϕ(p) . (4.20)

For example Tϕ W becomes a bundle over M ; the fibers are (Tϕ W )x = Tϕ(x ) W .
Using pullback bundle, we can characterize the tensor types of derivatives of ϕ:

ϕ̇ ∈ Γ(Tϕ W ), d ϕ ∈ Γ(T ∗ M ⊗ Tϕ W ). (4.21)

4.3.4 Type algebra

We will be using the following isomorphisms between tensor spaces.
Recall that the space of linear homomorphisms between two vector spaces U , V
is denoted by
linear
Hom(U ; V ) := {A : U −−−→ V }. (4.22)

Theorem 4.1 Hom(U ; V ) ∼

= U∗ ⊗ V.
4.3 Tensor Algebra 97

Theorem 4.2 (U ⊗ V )∗ ∼
= U ∗ ⊗ V ∗ . The dual pairing between U ∗ ⊗ V ∗ and U ⊗ V
is given by

⟨α ⊗ β|⃗u ⊗ ⃗v ⟩ := ⟨α|⃗u ⟩⟨β|⃗v ⟩ (4.23)

for α ∈ U ∗ , β ∈ V ∗ , ⃗u ∈ U , and ⃗v ∈ V .
Combining Theorem 4.1 and Theorem 4.2 gives:
Theorem 4.3 Hom(U ; V )∗ ∼
= Hom(U ∗ ; V ∗ ). The dual pairing between a linear
operator B ∈ Hom(U ∗ ; V ∗ ) and A ∈ Hom(U ; V ) is given by

⟨B |A⟩ = tr(A∗ B ) = tr(B ∗ A) = tr(AB ∗ ) = tr(BA∗ ). (4.24)

Theorem 4.4 Let E be a vector bundle over M . Then Γ(E )∗ ∼

= Γ(E ∗ ⊗ ∧n T ∗ M ).
The dual pairing between a field S ∈ Γ(E ∗ ⊗ ∧n T ∗ M ) and a field X ∈ Γ(E ) is
given by
ˆ
⟪S |X ⟫ := ⟨S |X ⟩. (4.25)
M

Note that in (4.25) the pairing ⟨S |X ⟩ is a pointwise dual pairing between fibers of E ∗
and fibers of E , leaving ⟨S |X ⟩ ∈ Γ(∧n T ∗ M ) = Ωn (M ) that can still be integrated
over M . That is why the dual space of the space scalar functions are measures:
(Ω0 (M ))∗ ∼
= Ωn (M ). (4.26)
The dual space of the space of vector fields is the space of covector-valued measures.
Γ(TM )∗ ∼= Γ(T ∗ M ⊗ ∧n T ∗ M ). (4.27)

Theorem 4.5 Let U be an n-dimensional vector space. Then

U ⊗ ∧n U ∗ ∼
= ∧n−1 U ∗ . (4.28)

The isomorphism is given by

⃗u ⊗ µ 7→ i⃗u µ ⃗u ∈ U , µ ∈ ∧n U ∗ . (4.29)

Proof. One can check that (4.36) is non-degenerate, and that U ⊗ ∧n U ∗ and ∧n−1 U ∗
share the same dimension (note that dim(∧n U ∗ ) = 1). One can also construct an
explicit inverse map for (4.36). For each ω ∈ ∧n−1 U ∗ , pick a µ ∈ ∧n U ∗ and our goal
is to reconstruct ⃗u ∈ U so that i⃗u µ = ω. (Note that a different choice of µ will only
lead to rescaled ⃗u , and still resulting in the same i⃗u µ and ⃗u ⊗ µ using (4.6).) The
reconstruction goes as picking an arbitrary basis ⃗e1 , . . . , ⃗en for U and
n
X ωJ⃗e1 , . . . , ⃗ebk , . . . , ⃗en K
⃗u = u k ⃗ek , u k = (−1)k (4.30)
k =1
µJ⃗e1 , . . . , ⃗en K

where (·)
c means skipping over the term.
98 Chapter 4. Continuum Mechanics

linear
■ Example 4.4 — Cramer’s rule. Let A : U −−−→ U be a linear map. Given ⃗b ∈ U ,
find ⃗x ∈ U such that A⃗x = ⃗b.
Pick an arbitrary basis ⃗e1 , . . . , ⃗en for U and an arbitrary volume form µ ∈ ∧n U ∗ .
Note that det(A) = ( A∗ µ)/µ, which is independent of the choice of µ. We may
n-form
linear
assume µJ⃗e1 , . . . , ⃗en K = 1. Now, A : U −−−→ U induces a pullback on (n −1)-forms:
linear
A∗ : ∧n−1 U ∗ −−−→ ∧n−1 U ∗ . (4.31)
(n − 1)-form

Note that

A∗ (i⃗b µ) = iA−1⃗b A∗ µ = det(A)i⃗x µ. (4.32)

(n − 1)-form n-form

Hence, using (4.30) we can reconstruct ⃗x as x k ⃗ek with

Pn
i=1

1
xk = (−1)k (A∗ i⃗b µ)J⃗e1 , . . . , ⃗ebk , . . . , ⃗en K (4.33)
det(A)
1 ⃗b , . . . , ⃗an K
= µJ⃗a1 , . . . , |{z} (4.34)
det(A)
k -th
slot

where ⃗ai = A⃗ei . ■

The following is a more general version of Theorem 4.5.

Theorem 4.6 Let U be an n-dimensional vector space. Then

(⊗k U ) ⊗ ∧n U ∗ ∼
= ∧n−k U ∗ . (4.35)

The isomorphism is given by

(⃗u1 ⊗ · · · ⊗ ⃗uk ) ⊗ µ 7→ i⃗uk · · · i⃗u1 µ, ⃗u1 , . . . , ⃗uk ∈ U , µ ∈ ∧n U ∗ . (4.36)

Using Theorem 4.6 and 4.4 we can reason the dual space of the space of k -form
fields:

Ωk (M )∗ = Γ(∧k T ∗ M )∗ ∼
= Γ(∧k TM ⊗ ∧n T ∗ M ) ∼
= Γ(∧n−k T ∗ M ) = Ωn−k (M ).
(4.37)

Theorem 4.7 Ωk (M )∗ ∼
= Ωn−k (M ). The dual pairing between α ∈ Ωn−k (M ) and
β ∈ Ωk (M ) is given by
ˆ
⟪α|β⟫ = α ∧ β. (4.38)
M

See also Theorem 2.44.

 4.4 Postulates of Elasticity 99

4.4 Postulates of Elasticity

In addition to the general postulate Postulate 4.1 for continuum mechanics laid out in
Section 4.1, (hyper)elasticity has the following additional assumption.
Postulate 4.2 — Hyperelasticity. A hyperelasticity system is a conservative con-
tinuum ´mechanical system with the potential energy U(ϕ) taking the form of
U(ϕ) = M Ψ( ϕ∗ ♭W ), where ϕ∗ ♭W is the induced metric (a.k.a. pullback
⊙2 T ∗ W ⊙2 T ∗ W
(nonlinear)
metric) from ♭W by the deformation map ϕ, and Ψ : ⊙2 (T ∗ M ) −−−−−−→ ∧n T ∗ M
is a base-point preserving map. This function Ψ is called the strain energy density
function.
The equation of motion for a hyperelastic system is given by (4.2) (following
the least action principle)
∇
∂U
ρM ϕ̈ = − + fext (4.39)
∂ϕ

with possibly an external force fext ∈ Γ(∧n T ∗ M ⊗ ∧n Tϕ∗ W ).

The induced metric is the measurement of world-space distances and angles, as a

tensor defined on M . The elastic potential energy is the integral of some pointwise
evaluation of local energy Ψ, and this local energy is only a function of the local
induced metric.
We will derive −∂U
∂ϕ
in Section 4.5. We spend the remainder of this section to
explain objects such as the pullback metric and the deformation gradient.

4.4.1 Deformation gradient

Let ϕ : M → W be a deformation map. We define the deformation gradient as the
world-vector-valued 1-form on M :
Definition 4.9 — Deformation gradient.

F := d ϕ ∈ Γ(T ∗ M ⊗ Tϕ W ) ∼
= Γ(Hom(TM ; Tϕ W )). (4.40)

The deformation is the pushforward operator given from the flow map ϕ. Each
displacement as a vector in the material is mapped to a displacement as a tangent
vector in the world.
■ Example 4.5 Under a Cartesian coordinate in R3 , if we write
   1 
x ϕ (X , Y , Z )
y  = ϕ(X , Y , Z ) = ϕ2 (X , Y , Z ) (4.41)
z ϕ3 (X , Y , Z )
 100 Chapter 4. Continuum Mechanics
then the deformation gradient has a matrix representation given by
 1 1 1

∂ϕ ∂ϕ ∂ϕ
 ∂X2 ∂Y
∂ϕ2
∂Z
∂ϕ2 
F = d ϕ =  ∂ϕ
∂X ∂Y ∂Z 
(4.42)
∂ϕ3 ∂ϕ3 ∂ϕ3
∂X ∂Y ∂Z

4.4.2 Pullback metric

Recall that

♭W ∈ Γ(⊙2 T ∗ W ) ∼
= Γ(Homsym (TW ; T ∗ W )) (4.43)

denote the metric on the world space W . Now, with the deformation gradient F
and its adjoint, we find a diagram

TM
F / TW (4.44)
♭W

T ∗M o F∗
T ∗W

By composing these arrows, we obtain a symmetric bilinear form that can serve as a
metric on M .
Definition 4.10 — Pullback metric. Define

ϕ∗ : ⊙2 T ∗ W → ⊙2 T ∗ M , ϕ∗ ♭W := F ∗ ♭W F (4.45)
⊙2 T ∗ W ⊙2 T ∗ W

where F = d ϕ ∈ Γ(Hom(TM ; Tϕ W )) and ♭W ∈ Γ(Homsym (TW ; T ∗ W )) are

treated as linear maps like in (4.44). When viewing ϕ∗ ♭W as a bilinear form that
takes in two vectors, this definition is equivalent to
D E
⟨⃗a , ⃗b⟩ϕ∗ ♭W = d ϕJ⃗a K, d ϕJ⃗bK , ⃗a , ⃗b ∈ Tp M . (4.46)
♭W

We call

C := F ∗ ♭W F ∈ Γ(⊙2 T ∗ M ) (4.47)

the (right) Cauchy–Green tensor.

TM F / TW (4.48)
C ♭W
o 
T ∗M o F∗
T ∗W

The physical meaning of the pullback metric is best understood by (4.46). The
pullback metric gives the following inner product structure: The inner product
between two vectors ⃗a , ⃗b in the material is evaluated by taking their world-space
 4.4 Postulates of Elasticity 101

inner product; the deformation gradient F = d ϕ first realizes these material vectors
as world vectors before taking the world-space inner product.
Note that the inner product structure encodes the notion of lengths and angles.
In particular, it encodes how much the material is stretched and sheared (after
comparing it with a fixed reference material metric).
Definition 4.11 — Cauchy–Green constructor. Define

C : Γ(T ∗ M ⊗ Tϕ W ) → Γ(⊙2 T ∗ M ) (4.49)

C(F ) := F ∗ ♭W F . (4.50)

■ Example 4.6 Continuing Example 4.5 in the Cartesian 3D space, where the
metric tensor for ♭W is represented by the identity matrix, the right Cauchy–Green
tensor is given by

C = F⊺ F. (4.51)

4.4.3 Strain energy density function

One postulates that the elastic potential energy is only a function of the pullback
metric C , which encodes all information of the local world-space measurements of
distance and angle. This local potential energy is the strain energy density function
of C . The strain energy density function is a base-point preserving map between
vector bundles:
(nonlinear)
Ψ : ⊙2 T ∗ M −−−−−−→ ∧n T ∗ M . (4.52)

It is the core material property. The value Ψ(C ) describes how much energy is stored
(as an n-form) for a given pullback metric C . The function Ψp : ⊙2 Tp∗ M → ∧n Tp∗ M
can depend on p ∈ M for an inhomogeneous material.
Definition 4.12 — Derivative of Ψ. For each point p ∈ M , the differential d (Ψp ) of
the function Ψp : ⊙2 Tp∗ M → ∧n Tp∗ M over the domain ⊙2 Tp∗ M is a ∧n Tp∗ M -valued
covector of ⊙2 Tp∗ M ; that is,

d (Ψp ) ∈ (⊙2 Tp∗ M )∗ ⊗ ∧n Tp∗ M ∼

= (⊙2 Tp M ) ⊗ ∧n Tp∗ M . (4.53)

We denote this differential as

∂Ψ ∂Ψ
∈ Γ (⊙2 TM ) ⊗ ∧n T ∗ M , (4.54)


:= d (Ψp ).
∂C ∂C p

The base-point-preserving function (as a function of C )

∂Ψ
: Γ(⊙2 T ∗ M ) → Γ (⊙2 TM ) ⊗ ∧n T ∗ M (4.55)


∂C
 102 Chapter 4. Continuum Mechanics
is called a stress–strain relation.
For a general hyperelastic model, we do not need to make any further assumption.
For whichever Ψ that is modeled, there is a derived ∂C
∂Ψ
. We will only need to know
∂Ψ
∂C
∈ Γ(⊙ TM ⊗ ∧ T M ) to derive d U (Section 4.5).
2 n ∗

More discussion about special models and physical designs for Ψ will be given in
Section 4.7.

4.5 Derivation of Elasticity

Our goal is to derive ∂ϕ
∂
U(ϕ) for (4.39) for hyperelastic systems (Postulate 4.2). The
´
“computation tree” for evaluating U(ϕ) = M Ψ(C ) is given by

Def. 4.11 Sec. 4.4.3 ´

∞
C (M ; W )
d / Γ(T ∗ M ⊗ Tϕ W )
C /
Γ(⊙ T M )2 ∗ Ψ /
Γ(∧ T M ) n ∗ M
/ R
ϕ F = dϕ C = C(F ) Ψ(C ) U(ϕ)
(4.56)

To compute the variation ∂U

∂ϕ
we apply the method of back propagation. The tangent
spaces of (4.56) and the differentials are
´
d∇ dC|F dΨ|C
Γ(Tϕ W ) / ∗
Γ(T M ⊗ Tϕ W ) / 2
Γ(⊙ T M ) ∗ / n
Γ(∧ T M ) ∗ M / R. (4.57)

Here, the variation of the first map d requires the Levi-Civita connection ∇W for
♭W on W :
∂
∂ϵ ϵ=0
ϕϵ = ϕ̊ ∈ Γ(Tϕ W ) =⇒ ∂
( d ϕ)
∂ϵ ϵ=0 0-form ϵ
= d ∇ ϕ̊. (4.58)
TW -valued
0-form

Using the duality theorems in Section 4.3.4, we find the dual spaces of the sequence
of vector spaces in (4.57) and connect them by the adjoint maps:
´
d∇ dC|F dΨ|C
Γ(Tϕ W ) / ∗
Γ(T M ⊗ Tϕ W ) / Γ(⊙ T M )2 ∗ / n
Γ(∧ T M )∗ M / R

´
(d ∇ )∗ (dCF )∗ (dΨC )∗ ( M )∗
Γ( ∧ T M o
n ∗
Γ( ∧ T M o
n−1 ∗
Γ( ⊙ TM 2 o Ω (M ) o
0
R∗
⊗ Tϕ∗ W ) ∗
⊗ Tϕ W ) n ∗
⊗ ∧ T M)
(4.59)

Take 1 ∈ R∗ as the seed and back-propagate it using these adjoint maps. The
intermediate variables during the back-propagation are known as the stress tensors
of various kinds.
The adjoint of integration distributes 1 into the constant function
´
1 = ( M )∗ (1) ∈ Ω0 (M ). (4.60)
4.5 Derivation of Elasticity 103

The next adjoint map (d Ψ)∗ yields

∂Ψ
= (d Ψ|C )∗ (1) ∈ Γ(⊙2 TM ⊗ ∧n T ∗ M ), (4.61)
∂C

which we assume can be looked up (Section 4.4.3) while further discussions about its
evaluation can be found in Section 4.7.
Definition 4.13 — 2nd Piola–Kirchhoff stress tensor. The 2nd Piola–Kirchhoff stress
tensor (PK2) S ∈ Γ(⊙2 TM ⊗ ∧n T ∗ M ) is defined by

S ∂Ψ
:= = (d Ψ|C )∗ (1). (4.62)
2 ∂C
The next adjoint map (d C)∗ gives us:
Definition 4.14 — 1st Piola–Kirchhoff stress tensor.
∂Ψ
P := (d C|F )∗ = (d C|F )∗ (d Ψ|C )∗ (1) ∈ Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W ). (4.63)
∂C
called the 1st Piola–Kirchhoff stress tensor (PK1).

Finally,

∂U
= (d ∇ )∗ P = (d ∇ )∗ (d C|F )∗ (d Ψ|C )∗ (1) ∈ Γ(∧n T ∗ M ⊗ Tϕ∗ W ). (4.64)
∂ϕ

To complete the calculation, we derive explicit formulas for the adjoints (d ∇ )∗ and
(d C|F )∗ .

Theorem 4.8 Represent the domain Γ(⊙2 TM ⊗ ∧n T ∗ M ) and the target space
Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W ) of (d C|F )∗ as spaces of linear maps Γ(Homsym (T ∗ M ; TM ) ⊗
∧n T ∗ M ) and Γ(Hom(T ∗ M ; Tϕ∗ W ) ⊗ ∧n T ∗ M ) respectively using Theorem 4.1 and
Theorem 4.5. Then the linear map
linear
d C|∗F : Γ(Homsym (T ∗ M ; TM ) ⊗ ∧n T ∗ M ) −−−→ Γ(Hom(T ∗ M ; Tϕ∗ W ) ⊗ ∧n T ∗ M )
(4.65)

is explicitly given by

d C|∗F JH K = 2♭W FH , H ∈ Γ(Homsym (T ∗ M ; TM ) ⊗ ∧n T ∗ M ), (4.66)

in terms of linear maps F ∈ Γ(Hom(TM ; Tϕ W )) and ♭W ∈ Γ(Homsym (Tϕ W ; Tϕ∗ W )).

Proof. Differentiating (4.50) (C(F ) = F ∗ ♭W F ) with respect to F , we obtain

d CF JF̊ K = F̊ ∗ ♭W F + F ∗ ♭W F̊ , F̊ ∈ Γ(Hom(TM ; Tϕ W )). (4.67)

Since this is a pointwise linear operation, we may focus on its pointwise adjoint as in
 104 Chapter 4. Continuum Mechanics

the following diagram (with Γ(· · · ) and ⊗ ∧n T ∗ M dropped):

F̊ dC|F F̊ ∗ ♭W F +F ∗ ♭W F̊
Hom(TM ; Tϕ W ) / Homsym (TM ; T ∗ M ) (4.68)

dC|∗F
Hom(T ∗ M ; Tϕ∗ W ) o Homsym (T ∗ M ; TM )
H

Using the dual pairing between linear maps (Theorem 4.3) the adjoint d CF∗ must
satisfy
def’n of
 adjoint
 and (4.67)
∗ ∗
tr (d C|F JH K) F̊ = tr(H ∗ (F̊ ∗ ♭W F + F ∗ ♭W F̊ )) (4.69)
= tr(H ∗ F̊ ∗ ♭W F ) + tr(H ∗ F ∗ ♭W F̊ ) (4.70)
= 2 tr(H ∗ F ∗ ♭W F̊ ) (4.71)

where the last step uses invariance of trace under cyclic permutations (tr(ABCD) =
tr(BCDA)) and adjoints tr(A) = tr(A∗ ), and that H ∗ = H for H ∈ Homsym (TM ∗ ; TM ).
Therefore, d C|∗F JH K = 2♭W FH .

Theorem 4.9 The adjoint of d∇ is −d ∇ up to a boundary integral.

TW -valued T ∗ W -valued
0-form (n − 1)-form

d∇
Γ(Tϕ W )
0-form
/ Γ(T ∗ M ⊗ Tϕ W ) (4.72)

(d ∇ )∗
Γ(∧n T ∗ M ⊗ Tϕ∗ W ) o Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W )
= −d ∇
(n − 1)-form

Proof. For each ϕ̊ ∈ Γ(Tϕ W ) and P ∈ Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W ), we have

ˆ ˆ ˆ
d ⟨ϕ̊|P ⟩ = ⟨ d ∇
ϕ̊ ∧ P ⟩ + ⟨ϕ̊| ∗ d ∇ P ⟩ . (4.73)
(n − 1)-form TW -valued T W -valued
| M ¸ {z } | M 0-form {z } |
M (n − 1)-form

⟪d ∇ ϕ̊|P ⟫
= ⟨ϕ̊|P ⟩
{z }
∂M
⟪ϕ̊|d ∇ P ⟫

Therefore, up to a boundary term, the adjoint of d∇ is −d ∇ .

TW -valued T ∗ W -valued
0-form (n − 1)-form

4.5.1 Summary of the derivation

The derivation for the differential of the potential energy U(ϕ) follows the same logic
as a reversed mode automatic differentiation. We first perform a forward evaluation,
involving several intermediate spaces. Then we perform a back-propagation along
their dual spaces. The intermediate variables are the Piola–Kirchhoff stress tensors.
 4.6 Cauchy stress tensor 105

The following diagram shows the forward nonlinear evaluations (upper row) and the
linear back-propagation (lower row).

ϕ F = dϕ C = F ∗ ♭W F Ψ(C ) U(ϕ)
flow deformation Cauchy–Green energy total
map gradient tensor density ´ energy
C ∞ (M ; W )
d / ∗
Γ(T M ⊗ Tϕ W )
C / Γ(⊙
2 ∗
T M)
Ψ / n ∗
Γ(∧ T M )
M
/ R

´ ∗
∇ ∗ ∗ ∗ ( M)
n
Γ( ∧ T M ∗ o (d ) Γ( ∧ n−1
T M ∗ o (dCF ) 2
Γ( ⊙ TM o (dΨC ) Ω 0
(M ) o R∗
⊗ Tϕ∗ W ) ∗
⊗ Tϕ W ) ⊗ ∧n T ∗ M )

−d ∇ P P = ♭W FS S
2
= ∂C∂Ψ
1 1
(n − 1)-form 1st Piola– constant seed
2nd Piola– “one” function
Kirchhoff stress Kirchhoff stress
(minus)
force
(4.74)

The equation of motion for the flow map is given by substituting the result of the
derivative computation into (4.39)
∇
ρM ϕ̈ = d∇ P + fext . (4.75)
(n − 1)-form

■ Example 4.7 — Summary in 3D. We can translate the tensors in (4.74) into
matrices under the 3D Cartesian coordinate, continuing Example 4.5 (F = ∇ϕ)
and Example 4.6 (C = F⊺ F).
The 1st (P) and 2nd (S) Piola–Kirchhoff stress tensors are 3-by-3 matrices
given by
 
∂Ψ ∂Ψ(C)
S=2 Sij = 2 , P = FS. (4.76)
∂C ∂Cij

The elastic force is given by taking the divergence of the 1st Piola–Kirchhoff stress
tensor on the 2nd index:
!
∂U X ∂
fM = − =∇·P fi = Pij . (4.77)
∂ϕ j
∂X j

Here fM is interpreted as force density per material volume. ■

4.6 Cauchy stress tensor

The 1st Piola–Kirchhoff stress tensor P ∈ Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W ) is a world-covector-
valued (n − 1)-form. It is to be evaluated on a codimension-1 infinitesimal plane,
and it returns the force as a covector in the world space.

The 1st Piola–Kirchhoff stress tensor assigns a world force covector on

each codimension-1 plane in the material.
106 Chapter 4. Continuum Mechanics

By a change of coordinate of the same object, we have the Cauchy stress tensor.

The Cauchy stress tensor assigns a world force covector on each codimension-
1 plane in the world.

Definition 4.15 — Cauchy stress tensor. Given a 1st Piola–Kirchhoff stress tensor
P ∈ Γ(∧n−1 T ∗ M ⊗ Tϕ∗ W ), the Cauchy stress tensor is a covector-valued (n − 1)-
form

σ ∈ Γ(∧n−1 T ∗ W ⊗ T ∗ W ) (4.78)

(on the image of ϕ) defined such that

P= ϕ∗ σ. (4.79)
(n − 1)-form

■ Example 4.8 In a 3D Cartesian space, the relationship between the Cauchy

stress tensor σ and the 1st Piola–Kirchhoff stress tensor P is given by

P = J σF−⊺ , J = det(F). (4.80)

This is a direct application that in 3D

ϕ∗ (w)2-form = (J F−1 w)2-form ◦ ϕ. (4.81)

The multiplication by F−1 in (4.80) is from the right with a transpose because it
is the 2nd index of P and σ that corresponds to being a 2-form. ■

Theorem 4.10 Let f ∈ Γ(∧n T ∗ M ⊗ T ∗ W ) be the force covector measure defined

on the world so that the total elastic force d ∇ σ:

d∇ P = ϕ∗ f . (4.82)
(n − 1)-form n-form

Then

d∇ σ = f. (4.83)
(n − 1)-form

Proof. Exterior derivative and pullback commutes:

ϕ∗ d∇ σ= d∇ ϕ∗ σ= d∇ P = ϕ∗ f . (4.84)
n-form(n − 1)-form (n − 1)-form(n − 1)-form (n − 1)-form n-form

■ Example 4.9 In the 3D Cartesian setup, if fW represents the force density per
world-space volume and fM represents the force density per material-space volume.
 4.7 More on the strain energy density function 107
Then

fM = J fW ◦ ϕ, ∇ · σ = fW , ∇ · P = fM . (4.85)

Next, we show that the Cauchy stress tensor is symmetric. In contrast to physical
arguments about angular momentum conservation (many versions of which apply
only to W being Euclidean or under the assumption of equilibrium), we show the
symmetry by type checking.
First, consider a version of the Cauchy stress so that its force-covector is converted
into a vector using the inner product structure of the world:

σ̃ := ♯W σ ∈ Γ(∧n−1 T ∗ W ⊗ TW ) (4.86)
∼
= Γ(TW ⊗ TW ⊗ ∧n T ∗ W ). (4.87)

Note that in the 3D Cartesian representation σ̃ and σ share the same matrix
representation. Now, in the type of (4.87) it is sensible to ask whether σ̃ belong to
the symmetric sub-type
?
σ̃ ∈ Γ(TW ⊙ TW ⊗ ∧n T ∗ W ). (4.88)

Theorem 4.11 Under Postulate 4.2, we must have

σ̃ ∈ Γ(TW ⊙ TW ⊗ ∧n T ∗ W ). (4.89)

Proof. Regarding σ̃ as a homomorphism-valued measure

σ̃ ∈ Γ(Hom(T ∗ W ; TW ) ⊗ ∧n T ∗ W ) (4.90)

we have

ϕ∗ σ̃ = ( ϕ∗ σ̃)F ∗ = ♯W ( ϕ∗ σ)F ∗ (4.91)

n-form (n − 1)-form (n − 1)-form

= ♯W PF ∗ = ♯W ♭W FSF ∗ = FSF ∗ (4.92)

which is self-adjoint as S is self-adjoint. The pullback ϕ∗ on the n-form part does

n-form
not change the symmetry type of its homomorphism values.

4.7 More on the strain energy density function

In Section 4.4.3 we described that the strain energy density function is a pointwise
(base-point preserving) map from a Cauchy–Green tensor to an energy density
(nonlinear)
Ψ : ⊙2 T ∗ M −−−−−−→ ∧n T ∗ M . (4.93)

How do we model Ψ? We need more context of geometry and physics.

 108 Chapter 4. Continuum Mechanics

Postulate 4.3 — Existence of rest metric. There is a time-independent material

metric ♭M ∈ Γ(⊙2 T ∗ M ). Having ♭M , one can compare the pullback metric
C = F ∗ ♭W F with ♭M by their quotient

Ĉ := ♯M C = ♭−1
M C ∈ Γ(Hom(TM ; TM )) (4.94)

which is an endomorphism on TM (and by contrast C is a bilinear form):

TM F / TW TMO F / TW (4.95)
C ♭W ♭M Ĉ ♭W
o   
T ∗M o F∗
T ∗W T ∗M o F∗
T ∗W

The energy Ψ(C ) can be written as a scalar-valued function of Ĉ

Ψ(C ) = Ψ̂(Ĉ )ρM , Ψ̂ : End(TM ) → R, (4.96)

and Ψ̂ attains its minimum at Ĉ = idTM . The value of Ψ̂ is called the internal
energy or Helmholtz’s free energy.

The endomorphism Ĉ indicates how much itself deviates from the identity, as a
measurement of relative stretching. When the internal energy is only a function of
Ĉ and not on other “latent variables” such as entropy or temperature, then internal
energy is equivalent to Helmholtz’s free energy.

R In thermoelasticity, Ψ̂ can depend on entropy (isentropic material) or on

temperature (isothermal material). In the former case, Ψ̂ is the internal energy,
and in the latter case Ψ̂ is Helmholtz’s free energy.
In elastoplasticity, ♭M becomes variable as well. A changed rest metric ♭M
models a permanent plastic deformation.

4.7.1 Strains
A strain is an expression of Ĉ that measures its deviation from identity.

• Green–St. Venant strain EStV := 21 (Ĉ − I ).

p
• Biot strain EBiot := Ĉ − I .

• Hencky strain EHencky = Ĥ := 12 ln Ĉ .

• Almansi strain EAlmansi := 21 (I − Ĉ −1 ).

All these strains agree with each other to first order when Ĉ ≈ I .
In many scenario for large deformations, it is natural to consider Hencky’s
logarithmic strain as our main measurement for the deviation from identity. By a
 4.7 More on the strain energy density function 109

mere change of variable, one may also model the internal energy as a function of the
Hencky strain:

Ψ(C ) = Ψ̂(Ĉ )ρM = Ψ̂log (Ĥ )ρM , 0 ∈ argmin Ψ̂log (Ĥ ) (4.97)
Ĥ ∈End(TM )

■Example 4.10 — St. Venant–Kirchhoff model. In the St. Venant–Kirchhoff model,

one defines the internal energy as the following quadratic energy of some strain

λ̂
Ψ̂StVK (Ĉ ) = tr(E )2 + µ̂ tr(E 2 ). (4.98)
2

Here λ̂, µ̂ are the (per unit mass) Lamé constants.

The corresponding stress–strain relation is

∂Ψ(C ) ∂ Ψ̂(Ĉ )  
S =2 = ♯M ρM = λ̂ tr(E )♯M + 2µ̂♯M E ρM (4.99)
∂C ∂ Ĉ
= λ tr(E )♯M + 2µ♯M E (4.100)

where λ = λ̂ρM , µ = µ̂ρM are Lamé constants as n-forms.

In Cartesian coordinate,

S = λ tr(E)I + 2µE. (4.101)

4.7.2 Isotropic models

To design a more general Ψ̂(Ĉ ) beyond a quadratic function on some strain, one
takes advantage of another symmetry under the following isotropy assumption.
Postulate 4.4 — Isotropic material. We call a material isotropic if

Ψ̂(Ĉ ) = Ψ̂(R −1 ĈR) (4.102)

for all R ∈ End(TM ) so that R ∗ ♭M R = ♭M and det(R) = 1 (i.e. for rotation

transformations R with respect to ♭M ).

■ Example 4.11 (4.98) is isotropic. ■

Since ♭M Ĉ = C is self-adjoint, there exists a ♭M -orthonormal basis, or equivalently

a ♭M -rotation transformation R, that diagonalizes Ĉ . Therefore, for an isotropic
material (♭M -rotation invariance), Ψ̂(Ĉ ) depends only on the eigenvalues of Ĉ . The
linear
eigenvalues of Ĉ are also the squares of the singular values of F : TM −−−→ Tϕ W
using ♭M and ♭W as the inner product structures for the singular value decomposi-
tion.
110 Chapter 4. Continuum Mechanics

Theorem 4.12 A material is isotropic if and only if Ψ̂(Ĉ ) is a function of the

eigenvalues {λ1 , λ2 , . . . , λn } of Ĉ modulo permutation.

A general function of n variables {λ1 , . . . , λn } that is invariant under their permuta-

tion can be expressed as a function of the coefficients of the polynomial whose roots are
{λ1 , . . . , λn }. Therefore, Theorem 4.12 is equivalent to the following Theorem 4.13.
Definition 4.16 — Invariants of an endomorphism. The coefficients I1 , . . . , In of
the characteristic polynomial of Ĉ

PĈ (z ) = det(zI − Ĉ ) = z n − I1 z n−1 + I2 z n−2 − · · · + (−1)n In (4.103)

are called the principal invariants of Ĉ . The trace of the first n powers of Ĉ

Jk := tr(Ĉ k ), k = 1, . . . , n (4.104)

are called the main invariants of Ĉ . When the Hencky strain Ĥ = 12 ln(Ĉ ) (4.97) is
our main variable for material modeling, we also consider the principal invariants
and the main invariants of Ĥ . They are denoted by lower case letters:

det(zI − Ĥ ) =: z n − i1 z n−1 + i2 z n−2 − · · · + (−1)n in (4.105)

k
jk := tr(Ĥ ). (4.106)

Theorem 4.13 — Principal invariants. A material is isotropic if and only if Ψ̂(Ĉ )

can be written as

Ψ̂(Ĉ ) = w (I1 , . . . , In ) (4.107)

where I1 , . . . , In are the principal invariants of Ĉ , and w : Rn → R is some function.

R Similarly, isotropic materials can be modeled with Ψ̂log (Ĥ ) = w (i1 , . . . , in ).

The following example demonstrates that one can perform change of variables to
convert any functions of tr(Ĉ k ) (even when k > n) into an expression of the principal
invariants.
■ Example 4.12 In 3D, the 3 principal invariants of the Cauchy–Green endomor-
phism Ĉ are

I1 = λ1 + λ2 + λ3 = tr(Ĉ ) (4.108)
I2 = λ1 λ2 + λ1 λ3 + λ2 λ3 = 21 (tr(Ĉ )2 − tr(Ĉ 2 )) (4.109)
I3 = λ1 λ2 λ3 = det(Ĉ ). (4.110)
4.7 More on the strain energy density function 111
Conversely,

tr(Ĉ ) = I1 (4.111)
tr(Ĉ 2 ) = I12 − 2I2 (4.112)
3
tr(Ĉ ) = I13 − 3I1 I2 + 3I3 (4.113)

where the 3rd equation is obtained by taking trace on the Cayley–Hamilton

Theorem

Ĉ 3 − I1 Ĉ 2 + I2 Ĉ − I3 I = 0. (4.114)

Higher moments such as tr(Ĉ 4 ) can be obtained by repeated substitutions using

the Cayley–Hamilton equation: tr(Ĉ 4 ) = tr(Ĉ (Ĉ 3 )) = tr(Ĉ (I1 Ĉ 2 − I2 Ĉ + I3 I )).

tr(Ĉ 4 ) = I14 − 4I12 I2 + 4I1 I3 + 2I22 , (4.115)

tr(Ĉ 5 ) = I15 − 5I13 I2 + 5I12 I3 + 5I1 I22 − 5I2 I3 . (4.116)

In general,
k 
 
0 1 0 1
tr(Ĉ k ) = 1 0 0  0 (4.117)
 
0 1   I1  .
I3 −I2 I1 I12 − 2I2

Using power series, any expression tr(f (Ĉ )), for any analytic function f , can be
written as an analytic function of I1 , I2 , I3 . ■

A useful tip about taking the derivative of the determinant of a matrix with
respect to a matrix (endomorphism) is

∂ det(A)
= cof(A) = det(A)A∗−1 . (4.118)
∂A
Applying this identity to our characteristic polynomial we get

∂PĈ (z )
= −PĈ (z )(zI − Ĉ )∗−1 (4.119)
∂ Ĉ
and
∂Ik 1 dk ∂PĈ (z ) 1 dk  
= =− PĈ (z )(zI − Ĉ )∗−1 . (4.120)
∂ Ĉ k ! dz k z =0 ∂ Ĉ k ! dz k z =0

This allows us to compute the 2nd Piola–Kirchhoff stress for a general function
w (I1 , . . . , In ) of the invariants:
n
∂Ψ(C ) ∂ Ψ̂(Ĉ ) ∂ Ψ̂(Ĉ ) X ∂Ik ∂w
= ♯M ρM , = . (4.121)
∂C ∂ Ĉ ∂ Ĉ k =1
∂ Ĉ ∂Ik
112 Chapter 4. Continuum Mechanics

■ Example 4.13 — Neo-Hookean.

µ̂ λ̂ p
Ψ̂(Ĉ ) = w (I1 , . . . , In ) = (I1 − n − ln In ) + ( In − 1)2 . (4.122)
2 2
■

The following are special model one can consider in the framework of hyperelas-
ticity models. They are used in conformal mappings in geometry processing and
fluid simulations. Although they are usually not discussed together with elastic
solid simulation, we can describe them in our current discussion about the effects of
stress–strain energy designs.
■ Example 4.14 — Conformal energy. In 2D (n = 2), an energy that only penalizes
conformal distortion but not area distortion takes the form

Ψ̂log (Ĥ ) = w (i1 , i2 ) = w̃ i12 − 2i2 (4.123)

√ √
In terms of the singular values s1 = λ1 , s2 = λ2 of the deformation gradient F
with respect to ♭M and ♭W ,
 
s1
Q= i12 − 2i2 = ln (4.124)
s2

which is also known as the dilatation or the quasiconformal error. A quasicon-

Q
formal map is when the maximal conformal distortion, for example supM ( 1+e1−e Q
=
coth 2 ) over the domain M , is bounded or minimized.
−Q

A deformation is conformal or conformally constrained if Q = 0 everywhere.

When a deformation is conformal, F is locally a scaled isometry (scale rotation)
C = ϕ∗ ♭W = e 2u ♭M for some conformal factor e 2u . The conformal energy and the
conformal constraint will induce a “conformal stress” that restores conformality,
which are derived in [Soliman et al.(2021)]. Under the conformal constraint, the
flow map ϕ is still free to deform conformally, and the energy is insensitive to
those deformations. ■

■ Example 4.15 — Volumetric energy. In nD, an energy that only penalizes

volume change but on any other mode of deformation takes the form

Ψ̂(Ĉ ) = w (I1 , . . . , In ) = w (In ). (4.125)

Note that the nth principal invariant of Ĉ is In = J 2 where J = (ϕ∗ µW )/(µM )

where µW and µM are the volume forms with respect to the metrics ♭W and
♭M respectively. There are two notable consequences of this design of energy.
First, as we will derive in Chapter 5, the Cauchy stress tensor will always be a
scalar times the identity. This scalar is called the pressure. Second, the energy is
completely insensitive to any volume-preserving deformation; that is, the material
is unable to create any restoration force on other modes of deformation, such as
volume-preserving stretching and shearing. When a material behaves like this, we
 4.8 Conclusion 113
call it a fluid.
Inviscid barotropic compressible fluids are in fact exactly the same as a
hyperelastic body with the assumption of (4.125). If the energy (4.125) is stiff,
producing a volumetrically constrained system, then the flow map can only be a
volume preserving one. Note that the material becomes rigid in its volume, but
still free to shear and stretch. In that limit, we obtain an incompressible fluid.
The governing equation of motion becomes the incompressible Euler equation.
Perhaps a philosophical take-away is: The idea of fluid and the concept of
pressure can be derived from a minimalistic set of postulates, in contrast to
phenomenologically reasoned. ■

4.8 Conclusion
Exterior calculus naturally arises in continuum mechanics when we derive the equa-
tions of motion from first principle and keeping track of the types of objects. The
back-propagation method that we employed for the derivation demands the distinction
between vectors and their duals. The result is that we obtain stress tensors (1st
Piola–Kirchhoff stress or cauchy stress) as covector-valued (n − 1)-forms. This fits
very well with the original physical intuition of a stress tensor: there is an assignment
of force on every infinitesimal surface. The language of exterior calculus allows us to
accurately describe this delicate tensor type and distinguish it from other matrices
of the same array size. Moreover, changes of coordinates between the material space
and world space use only one unified concept: pullback. Using the commutativity
between pullbacks and exterior derivatives, we immediately obtain Theorem 4.10
without using any vector identity.
Overall, there is much less assumption and identities, and there is a more fine-
grained understanding of formulas in elasticity and symmetry of stress tensor revealed
by the exterior calculus language.
 5. Fluid Dynamics

In this chapter, we study the motion of fluids. Similar to the philosophy of Chapter 4,
we can characterize Newtonian fluids from a minimalistic set of postulates, and
subsequently derive and study the equations of motion based on first principles. Our
study will primarily concentrate on non-dissipative (inviscid) fluids. Hence, we only
need to look at the kinetic and potential energy of the physical system.
The following postulates follows Example 4.15.
Postulate 5.1 — Inviscid barotropic fluid. An inviscid barotropic fluid is an isotropic
hyperelastic body (Postulates 4.1–4.4) whose internal energy Ψ̂(Ĉ ) is only a function
of the volume change Ψ̂(Ĉ ) = w (det(Ĉ )).

In the limit that the internal energy in response to a volume change is infinitely stiff,
the deformable body becomes volumetrically rigid for every material element. In
that case, a barotropic fluid becomes incompressible.
Postulate 5.2 — Inviscid incompressible fluid. An inviscid incompressible fluid is a
volume-constrained (det(Ĉ ) is time-independent) barotropic fluid (Postulate 5.1).

R “Barotropic” means the compressible fluid has its internal energy, and conse-
quently the pressure, depend only on the change of fluid density (or the change
of specific volume). A more general compressible fluid, such as an isentropic
fluid governing gas dynamics, will have its internal energy depending on other
hidden variables such as entropy.

R For fluids, Postulate 4.3 can be simplified to that the material space has a rest
volume form µM , instead of a full-blown rest metric tensor ♭M .
 116 Chapter 5. Fluid Dynamics

We first derive the equations of motion for these systems, known as Euler equa-
tions. These classical equations describe the motion of fluid’s velocity as a vector-
valued 0-form. Then, we move on to looking at the same equations in their covector-
based variants, i.e. they describe the motion of fluid’s velocity as a 1-form. We can
arrive at the covector-based Euler equation either by changes of variables at the
PDE level, or from first principles. We show how this covector-based fluid equation
provides visuals and computational efficiency advantages for fluid simulations.

5.1 Euler equations

We follow the same principles established in the previous chapter and study our
dynamical system using its kinetic and potential energy, and its associated Euler-
Lagrange equation to prove the equations of motion.

5.1.1 Vector-based Euler equations

Following Postulate
´ 1 2 4.2, the kinetic energy for a flow map of a continuum´ is given
by K(ϕ̇) = M 2 |ϕ̇|♭W ρM , and together with the potential energy U(ϕ) = M Ψ(C ),
C = ϕ∗ ♭W = F ∗ ♭W F , we can find the equation of motion as in (4.2):

d ∂K ∂U
=− . (5.1)
dt ∂ ϕ̇ ∂ϕ

Using (4.75), and assuming no external forces, the equation simplifies to:
∇
ρM ϕ̈ = d∇ P (5.2)
(n − 1)-form

whre P = 2♭W F ∂C∂Ψ

. As demonstrated in Example 4.7, in 3D Cartesian space, the
equation is further simplified to:
∇
ρM ϕ̈ = ∇ · P, (5.3)

where P = FS, S = ∂C ∂Ψ
, F = ∇ϕ, and C = F⊺ F.
Now let us substitute the assumptions of Postulate 5.1 (Example 4.15) for fluids.
Specifically, the potential energy only depends on its top-dimensional principal
invariant, or equivalently Ψ̂(Ĉ ) = w (I3 ) = w (J 2 ) in 3D, where J = det(F). In the
∗
coordinate-free language, J = det(F ) = ϕµµMW where µW and µM are the volume
forms on W and M respectively.
Definition 5.1 We denote the mass form in the world coordinate as ρ ∈ Ωn (W ),
which given by the relation

ρM =: ϕ∗ ρ. (5.4)
n-form
5.1 Euler equations 117
Define the material-space and world-space mass density 0-forms by
ρM ρ
qM := ∈ Ω0 (M ), q := ∈ Ω0 (W ). (5.5)
µM µW
One can check that
qM
ϕ∗ q = q ◦ ϕ = . (5.6)
0-form J

Theorem 5.1 Under the assumption of Postulate 5.1, the Cauchy stress tensor
takes the form of a scalar times identity

σ = −pI ⊗ µW , p ∈ Ω0 (W ) (5.7)

We call p the pressure.

Proof. Using (4.118), the derivative of Ψ̂(Ĉ ) = w (I3 ) with respect to Ĉ is given by
∂ Ψ̂
∂ Ĉ
= w ′ (I3 )cof(Ĉ ) = w ′ (J 2 )J 2 Ĉ ∗−1 = w ′ (J 2 )J 2 ♭M C −1 . Then, by (4.121) we have
the 2nd Piola stress tensor given by
∂Ψ(C )
S =2 = 2♯M w ′ (J 2 )J 2 ♭M C −1 ρM (5.8)
∂C
= 2w ′ (J 2 )J 2 C −1 ρM (5.9)
= 2w ′ (J 2 )J 2 F −1 ♯W F ∗−1 ρM (5.10)

Transforming the 2nd Piola–Kirchhoff stress tensor to the 1st Piola–Kirchhoff and
subsequently Cauchy stress tensor, we get

P = ♭W FS = 2w ′ (J 2 )J 2 F ∗−1 ρM (5.11)
P = ϕ∗ σ =⇒ σ = 2w ′ (J 2 )J (I ⊗ ρ) (5.12)
2-form

We complete the proof by defining

p := −2w ′ (J 2 )Jq. (5.13)

Definition 5.2 — Fluid velocity. We define ⃗

u = u ∈ Γ(TW ) as the fluid velocity,
defined by

ϕ̇ = u ◦ ϕ = ϕ∗ u (5.14)
0-form

We use the notation u and ⃗u interchangeably, both of which represent the vector
type or the vector-valued 0-form type.
Also define

η := ♭W u ∈ Ω1 (W ) (5.15)
118 Chapter 5. Fluid Dynamics
as the fluid velocity covector.

Theorem 5.2 — Barotropic Euler equation. Under the general equation of motion
(5.2), the velocity field u satisfies

∂q
+ div(qu) = 0, (5.16)
∂t
∂u grad p
+ ∇u u = − , (5.17)
∂t q

where p is a function of q derived from (5.13).

Proof. The continuity equation (5.16) follows from ρM being time-independent in

terms of Lie derivative (2.134).
Noting that P = ϕ∗ σ and σ = −pI ⊗ µW , we can write the equation of motion
2-form
as:
∇
ρM ϕ̈ = d ∇ P = d ∇ ϕ∗ σ = ϕ∗ d ∇ σ, (5.18)
2-form 3-form
= ϕ∗ (− grad p ⊗ µW ) = −( ϕ∗ µW )( ϕ∗ grad p) (5.19)
3-form 3-form 0-form
= −J µM (grad p ◦ ϕ). (5.20)

Recall that ϕ̇ = u ◦ ϕ = ϕ∗ u. From Corollary 2.29, we can interpret the second

0-form
derivative of the flow map ϕ as the Lie material derivative of the velocity vector field
as a 0-form:
∇
∇ ∗   ∂u 
ϕ̈ = ϕ u = ϕ∗ + Lu∇ u . (5.21)
∂t 0-form 0-form ∂t 0-form

Further, note that J µM

ρM
= J
qM
= ϕ∗ 1
q
using (5.6). This allows us to now write the
0-form
Euler equations in the Eulerian coordinate:
∂u grad p
+ ∇u u = − . (5.22)
∂t q

R The formulation of the Euler equation through the above process describes
the conservation law on linear momentum.

For the remainder of this chapter, we assume that q = 1 to reduce the system
into an incompressible fluid system.
 5.1 Euler equations 119

Corollary 5.3 — Incompressible Euler equation. Under Postulate 5.2, the velocity
field u satisfies

(div u)µW = Lu µW = 0, (5.23)

n-form
∂u
+ ∇u u = −grad p, (5.24)
∂t

5.1.2 Covector Euler equations as a constrained minimization

One may re-derive the equations of motion in terms of the velocity covector field.
For this, we formulate the action minimization in terms of the Eulerian coordinate
energy as follows:
ˆ
1 2
minimize S (u) = |u| µW
u,ψ W 2 (5.25)
subject to ψ̇ + Lu ψ = 0 and Lu µ = 0.
0-form n-form

Here, ψ is the inverse flow map given by ψ ◦ ϕ = ϕ∗ ψ = id. The first constraint,
0-form
referred to as Lin’s Constraint, comes from the Lie material derivative for the 0-form
backward flow map ψ̇ + Lu ψ = 0, and the second constraint enforces incom-
0-form
pressibility for the fluid. Rewriting the action using Lagrange multipliers we get
to:
ˆ
1 2
S (u, ψ) = |u| µW − ⟨η0 |ψ̇ + Lu ψ⟩µW − ⟨λ| Lu µW ⟩, (5.26)
W 2

where λ ∈ Ω0 (W ) and η0 ∈ Γ(Tψ∗ M ) are the Lagrange multipliers for the incompress-
ibility and Lin’s constraint, respectively. By taking variations in terms of u and ψ
we can find the following equations of motion:

η = ψ ∗ η0 − d λ, (5.27)
1-form
∂η0
= 0. (5.28)
∂t

where η = u♭ . Using Theorem 2.28, we can take the time derivative of the above to
the covector-based Euler equations:

η̇ + Lu η = −d λ. (5.29)
1-form

R The covector formulation of the Euler equation describes a conservation law

on the circulation of velocity 1-form.
 120 Chapter 5. Fluid Dynamics

Consistency between (5.29) and (5.24)

We can check that (5.29) is consistent with (5.24) as follows. As shown in Exam-
ple 2.21, we can write the Lie material derivative of the velocity 1-form η = u♭ for
3D Cartesian space:

∂η  ∂u ♭
+ Lu η = + u · ∇u + (∇u)⊺ u . (5.30)
∂t 1-form ∂t

To get to the Euler equation for fluid dynamics, we can add a modified gradient
pressure term λ = p − 12 iu η = p − 12 |u|2 to both sides.

∂η  ∂u ♭
+ Lu η + d λ = + u · ∇u + (∇u)⊺ u + ∇λ (5.31)
∂t 1-form ∂t
 ∂u ♭
= + u · ∇u + ∇p = 0. (5.32)
∂t

Corollary 5.4 — Vorticity equation. By taking d on (5.29), we obtain

ω̇ + Lu ω = 0, (5.33)
2-form

where

ω = dη (5.34)

is called the vorticity 2-form.

5.2 Advantages of Covector Euler Equation

While the covector-based equation of motion is equivalent to the vector-based
equation, it can provide alternative advantages that are desirable for the animation
and simulation of fluid dynamics.
The common convention in fluid solvers is to freeze the velocity v ← u, and split
the Euler equations in two steps of advection and projection. These steps, however,
prove to induce additional numerical dissipation in the simulation which can hinder
the computational performance and visual fidelity of the results. We can see that
the covector-based formulation provides a formulation that can significantly improve
the visual fidelity of the results while reducing numerical dissipation.
Given a frozen background velocity field v, and splitting the Euler equations of
motion (5.17) and (5.29), we get to the following transport equations for covector-
based transport:

η̇ + Lv η = 0. (5.35)
1-form

and for vector-based transport equation:

u̇ + v · ∇u = 0. (5.36)
5.2 Advantages of Covector Euler Equation 121

Covector-based transport in vector calculus

From Example 2.21, we can transform (5.35) to a vector calculus equivalent of:

u̇ + v · ∇u = −(∇v) · u. (5.37)

This form reveals how the covector-based Euler equation involves a stretching term
that was not present in the vector-based transport (5.36).
This equation is also the basis for the impulse methods [Buttke(1993), Cortez(1995),
Feng et al.(2022)]. In the literature, (5.35) is treated as a non-trivial dynamical sys-
tem for each flowing particle Du Dt
= −(∇v) · u. The standard method for solving
this PDE is to further split the transport equation which brings about additional
instability due to the right-hand side stretching term.

Solving covector-based transport without inducing instability from the stretching

term
To remedy the problems associated with solving (5.37), we can instead study (5.35).
The solution to this PDE, as stated in Corollary 2.29, is given by:

ηt = e − Lv t η0 , (5.38)
= ψt∗ η0 . (5.39)

In 3D Cartisen domain (see Figure 5.1), this is simply

ut = (∇ψt )⊺ u ◦ ψt . (5.40)

Note the difference to the solution of vector-based Euler equation which is the
component-wise Lie-material derivative of a vector-valued 0-form u̇ + Lv u = 0
0-form
given by:

ut = u0 ◦ ψt . (5.41)

As opposed to the Impulse methods, the advection through the pullback operator
does not need the splitting of the stretching term, which allows for a more stable
transport equation.

Conservation of Circulation
Kelvin’s conservation¸ of circulation [Thomson(1868),
¸ Frisch and Villone(2014)] states
that the circulation C u · d l or equivalently C ηt , around every closed loop is con-
served. Since the solution to the covector-based transport is simply the pullback
of the 1-form field ηt = ψt∗ η0 , this conservation law follows naturally according to
Theorem 2.1. Concretely:
˛ ˛
∗
ψt η0 = η0 . (5.42)
C ψ(C )

As opposed to the vector-based formulation of Euler equations which results

in a conservation of linear momentum, the covector-based formulation leads to a
conservation of circulation.
122 Chapter 5. Fluid Dynamics
−∇p d ψ(x)⊺ u(ψ(x))
transported u(ψ(x))
velocity u(ψ(x))
dψ(x)⊺ u(ψ(x))
d ψ(x)
v x
u covector-
x transported
velocity
u ψ(x)

ψ(x)

(a) (b) (c)

Figure 5.1 (a) In a standard fluid solver, the velocity u is transported by the
flow v using an inverse flow map ψ. (b) This process can turn a
rotation motion into a divergent one which is subsequently damped
by the pressure projection. (c) Our advection method maintains the
vorticity by a multiplication with the transposed Jacobian of the
inverse flow map.

For simulating vortex-dominant phenomena, computations based on the con-

servation of circulation [Elcott et al.(2007), Nabizadeh et al.(2022)] are more advan-
tageous. The difference between the two motion laws (5.17) and (5.29) becomes
apparent in a flow dominated by vorticity. In (5.17), the pressure force is respon-
sible for the concentration of vorticity. Without the pressure, the inertial motion
(conservation of linear momentum) turns into a centrifugal force that makes vortices
disintegrate. In (5.29), the Lagrangian pressure plays no role in the persistence of
vorticity. The mechanism for the conservation of vorticity is entirely encoded in
the left-hand side of the equation. Such a property makes an algorithm based on
time-splitting into advection and pressure steps especially appealing.

Equivalence to a vortex method

The traditional advection-projection method introduces a splitting error that destroys
vorticity [Zhang et al.(2015), Zehnder et al.(2018)]. This phenomenon arises solely
from the advection step, since the projection step only modifies the velocity u with
a pure gradient which leads to no change in its curl.
Theorem 5.5 — Vector-based transport as a vortex method. Under the classical
advection equation ( ∂t
∂
+ v · ∇)u = 0, the vorticity w = ∇ × u evolves according
toa

∂
∂t
w + v · ∇w − w · ∇v = ⟨∇u × ∇v⟩. (5.43)

This modified vorticity equation deviates from the correct vorticity equation by a
term ⟨∇u × ∇v⟩.
a
In index notation ⟨∇u × ∇v⟩i = ϵijk ∂j u ℓ ∂k v ℓ .

Proof. The vector-based transport equation ( ∂t

∂
+ v · ∇)u = 0, is equivalent to the
following equation ∂t η + Lv η = ⟨∇v, u⟩ using the velocity covector η = u♭ . Taking
∂
5.2 Advantages of Covector Euler Equation 123

d on both sides of the equation yields

∂
∂t
+ Lv ω = d ⟨∇v, u⟩ = ⟨d| ∇{z
ω ∇v}, u⟩ − ⟨∇v ∧ ∇u⟩
= ⟨∇u ∧ ∇v⟩, =0

where the first identity is the Leibniz rule with metric d ⟨A ∧ B⟩ = ⟨d ∇ A ∧ B⟩ +

(−1)k ⟨A ∧ d ∇ B⟩, and the term d ∇ ∇v is zero since there is no curvature in flat space
(i.e. d ∇ d ∇ = 0). Converting 2-forms to vector fields in 3D using ω = iw µ, we obtain
∂
∂t
w + ∇v w − ∇w v = ⟨∇u × ∇v⟩.

Theorem 5.6 — Covector-based transport as a vortex method. The evolution of

w = ∇ × u that undergoes (5.37) is
∂
∂t
w + v · ∇w − w · ∇v = 0 (5.44)

which is the correct vorticity equation.

Proof. As shown in Corollary 2.29, Eq. (5.37) is equivalent to (5.35), i.e. ∂t

∂
η + Lv η =
0, given that η = u . Applying d to (5.35), and using the commutativity between
♭

d and Lv , we obtain ∂t
∂
ω + Lv ω = 0 for the vorticity 2-form ω = d η. In 3D, the
relationship between the 2-form ω = d u♭ and the vector field w = curl u is given by
ω = iw µ where µ ∈ Ω3 (M ) is the volume form. Using this relationship we obtain
∂ ∂
0 = ∂t ω + Lv ω = ∂t (iw µ) + Lv (iw µ)
= i∂w/∂t µ + i[v,w] µ + iw Lv µ (Lv µ = 0 since div v = 0)
| {z }
= i(∂w/∂t +[v,w]) µ, =0

where in the second line we used the identity of Lie derivative on contraction
Lv (iw α) = i[v,w] α + iw (Lv α). Therefore, ∂t
∂
w + [v, w] = 0. Finally, we arrive at
(5.44) by substituting [v, w] = ∇v w − ∇w v under Theorem 2.27.
By advancing u via the covector transportation (5.35) or (5.37), we implicitly solve
the vorticity equation (5.44), which is the modeling equation for vortex methods.
The ability to solve (5.44) at the velocity level without using the vorticity variable is
significant. Previous vortex methods which solve (5.44) have to include an expensive
integration that converts vorticity back to velocity.
Commutativity between covector transportation and pressure projection
In a traditional fluid solver, the splitting error between traditional advection and
the projection arises because two operations do not commute. Here we show that
advection and projection commute in CF. This property of CF fundamentally removes
the splitting error of these two operations.
Consider the equivalence classes of Ω1 (M ) where [ξ] = [η] whenever ξ −η = d φ for
some function φ. This is a natural abstraction for our discussion since two covector
fields are equivalent if and only if they share the same pressure projection result.
124 Chapter 5. Fluid Dynamics

The pressure projection can be understood as extracting the unique divergence-free

representative in each equivalence class [η] ∈ Ω1 (M )/ im(d ).
Now consider two covector fields ξ0 , η0 ∈ Ω1 (M ), divergence-free or not, and
transport them by the covector advection equation (5.38) to obtain ξt , ηt respectively.
Then

[ξ0 ] = [η0 ] if and only if [ξt ] = [ηt ]. (5.45)

To see this assertion, use (5.38) to express ξt − ηt as the pullback of ξ0 − η0 by the

inverse flow map

ξt − ηt = ψt∗ (ξ0 − η0 ) (5.46)

and apply Theorem 2.11 so that the pullback of an exact differential is still exact.
Therefore, whether one first projects then advects, advects then projects, or inserts a
projection (or reflection [Zehnder et al.(2018)]) at the halfway point of the advection,
one will obtain covector fields all in the same equivalence class and hence the same
divergence-free representative.
Note that the same argument does not apply to the traditional advections. The
transportation with a simple value look-up (5.41) generally turns an exact gradient
vector field into a non-gradient field.

Extending CF with a long-time characteristic mapping

The analysis in Section 5.2 implies that we may delay the pressure projection for a
longer time (rather than a time step) and just transport the velocity covector field η
using a long-time flow map.
Let ψt : Wt → M0 be the inverse flow map (Eulerian-to-Lagrangian map) which
is the Lagrangian marker carried by the history of the solution
∂
ψ
∂t t
+ ut · ∇ψt = 0, ψ0 = idM . (5.47)

Then the velocity covector field at the current time t is the pressure projection of

ηt = ψt∗ η0 . (5.48)

In other words, if we maintain a Lagrangian marker ψt we obtain the fluid state

through a single-step look up. This drastically reduces the amount of interpolation
in the advection-projection iteration.
Such a long-time method of characteristic mapping (MCM) is proposed by
[Tessendorf and Pelfrey(2011), Sato et al.(2018), Qu et al.(2019)]. We call the vari-
ant of the CF method based on (5.47) and (5.48) CF+MCM.

R CF+MCM is subtly different from the traditional MCM. The latter requires
an accumulation of the pressure gradient over time whereas the former does
not. In the traditional MCM [Qu et al.(2019)], one evaluates ut by (5.24)
integrating over time along particle trajectories:
´t
ut (x) = u0 (ψt (x)) + 0 (∇pτ )(ϕτ (ψt (x))) d τ. (5.49)
 5.3 Conclusion 125

In CF+MCM, the time integration of (5.29) yields

´t
ηt = ψt∗ η0 + ∗
0 (ϕτ ◦ ψt ) d λτ d τ
´  (5.50)
t
= ψt∗ η0 + d 0 (ϕτ ◦ ψt )∗ λτ d τ .

which is a comparable notation to (5.49) corresponds to

ut (x) = d ψt⊺ u0 (ψt (x))

´t 2
+ 0 d ψt⊺ d ϕ⊺τ (∇(pτ − |u2τ | ))(ϕτ (ψt (x))) d τ (5.51)
´ 2

= d ψt⊺ u0 (ψt (x)) + ∇ 0 (pτ − |u2τ | )(ϕτ (ψt (x))) d τ .
t

As omitted in (5.48), the second terms in (5.50) and (5.51) can be absorbed
in a single pressure projection as they are exact differentials. This is possible
since d (resp. ∇) in (5.50) (resp. (5.51)) can be pulled out of the integral by the
commutativity property (cf. Theorem 2.11) between d and pullback operators.
By contrast in (5.49), the ∇ in the pressure term cannot be pulled out of the
´t
time integral. The accumulated pressure 0 (∇pτ ) ◦ ϕτ ◦ ψt d τ is generally not
an exact gradient. Therefore, a traditional MCM must carefully record the
accumulated pressure. This procedure is entirely removed in CF+MCM.

5.3 Conclusion
One can write the equations of motion of fluid mechanics following the same process
established in the previous chapter. While this gets us to a vector-based Euler
equation following the conservation of linear momentum, with a modified procedure
of minimizing action in the Lagrangian frame, one can write the same equations of
motion in terms of velocity covector. These equation reveal multiple benefits when
it comes to simulating incompressible ideal fluids, since the formulation naturally
guarantees the preservation of circulation while involving a simple modification by
multiplying the looked up velocity with the inverse transpose of the flow map’s
jacobian (i.e. u ← (∇ψt )⊺ u ◦ ψt as opposed to u ← u ◦ ψt ). Additionally, this
formulation completely removes the splitting error between the pressure projection
and advection steps as it follows the commutativity between exterior derivative
and the pullback operator. The method emulates a vorticity-based Euler equation,
which is preferable in vortex-dominated flows, without incurring instability due to
the stretching term as the pullback operator is exact solution to the Lie-material
derivative for 1-form velocity field. This natural solution also allows for using long-
term method of characteristics mapping since the pullback operator remains the
exact solution regardless of the duration of the timesteps.
III
Appendix

Bibliography . . . . . . . . . . . . . . . . . . . . . 129

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
 Bibliography

[Boyd and Vandenberghe(2004)] Stephen P Boyd and Lieven Vandenberghe. 2004.

Convex optimization. Cambridge university press.
[Buttke(1993)] Tomas F Buttke. 1993. Velicity methods: Lagrangian numerical
methods which preserve the Hamiltonian structure of incompressible fluid flow.
In Vortex flows and related numerical methods. Springer, 39–57.
[Cortez(1995)] Ricardo Cortez. 1995. Impulse-based methods for fluid flow. https:
//doi.org/10.2172/87798
[Crane(2018)] Keenan Crane. 2018. Discrete differential geometry: An applied
introduction. Notices of the AMS, Communication (2018), 1153–1159.
[Crane et al.(2013)] Keenan Crane, Fernando De Goes, Mathieu Desbrun, and Peter
Schröder. 2013. Digital geometry processing with discrete exterior calculus. In
ACM SIGGRAPH 2013 Courses. 1–126.
[Elcott et al.(2007)] Sharif Elcott, Yiying Tong, Eva Kanso, Peter Schröder, and
Mathieu Desbrun. 2007. Stable, circulation-preserving, simplicial fluids. ACM
Transactions on Graphics (TOG) 26, 1 (2007), 4–es.
[Evans(1998)] Lawrence C Evans. 1998. Partial differential equations. American
Mathematical Society.
[Feng et al.(2022)] Fan Feng, Jinyuan Liu, Shiying Xiong, Shuqi Yang, Yaorui Zhang,
and Bo Zhu. 2022. Impulse Fluid Simulation. IEEE Transactions on Visualiza-
tion and Computer Graphics (2022).
[Flanders(1963)] Harley Flanders. 1963. Differential forms with applications to the
physical sciences by harley flanders. Elsevier.
130 BIBLIOGRAPHY

[Frankel(2011)] Theodore Frankel. 2011. The geometry of physics: an introduction.

Cambridge university press.

[Frisch and Villone(2014)] Uriel Frisch and Barbara Villone. 2014. Cauchy’s almost
forgotten Lagrangian formulation of the Euler equation for 3D incompressible
flow. The European Physical Journal H 39, 3 (2014), 325–351.

[Grinspun et al.(2006)] Eitan Grinspun, Mathieu Desbrun, Konrad Polthier, Peter

Schröder, and Ari Stern. 2006. Discrete differential geometry: an applied
introduction. ACM Siggraph Course 7, 1 (2006).

[Nabizadeh et al.(2022)] Mohammad Sina Nabizadeh, Stephanie Wang, Ravi Ra-

mamoorthi, and Albert Chern. 2022. Covector fluids. ACM Transactions on
Graphics (TOG) 41, 4 (2022), 1–16.

[Nicolet et al.(2021)] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. 2021. Large
steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG)
40, 6 (2021), 1–13.

[Qu et al.(2019)] Ziyin Qu, Xinxin Zhang, Ming Gao, Chenfanfu Jiang, and Baoquan
Chen. 2019. Efficient and conservative fluids using bidirectional mapping. ACM
Transactions on Graphics (TOG) 38, 4 (2019), 1–12.

[Sato et al.(2018)] Takahiro Sato, Christopher Batty, Takeo Igarashi, and Ryoichi
Ando. 2018. Spatially adaptive long-term semi-Lagrangian method for accurate
velocity advection. Computational Visual Media 4, 3 (2018), 223–230.

[Sellán et al.(2021)] Silvia Sellán, Jack Luong, Leticia Mattos Da Silva, Aravind
Ramakrishnan, Yuchuan Yang, and Alec Jacobson. 2021. Breaking Good:
Fracture Modes for Realtime Destruction. arXiv preprint arXiv:2111.05249
(2021).

[Soliman et al.(2021)] Yousuf Soliman, Albert Chern, Olga Diamanti, Felix Knöppel,
Ulrich Pinkall, and Peter Schröder. 2021. Constrained Willmore surfaces. ACM
Transactions on Graphics (TOG) 40, 4 (2021), 1–17.

[Tessendorf and Pelfrey(2011)] Jerry Tessendorf and Brandon Pelfrey. 2011. The
characteristic map for fast and efficient vfx fluid simulations. In Computer
Graphics International Workshop on VFX, Computer Animation, and Stereo
Movies. Ottawa, Canada.

[Thomson(1868)] William Thomson. 1868. On Vortex Motion. Earth and Environ-

mental Science Transactions of the Royal Society of Edinburgh 25, 1 (1868),
217–260.

[Wang and Chern(2021)] Stephanie Wang and Albert Chern. 2021. Computing
minimal surfaces with differential forms. ACM Transactions on Graphics (TOG)
40, 4 (2021), 1–14.
BIBLIOGRAPHY 131

[Yang et al.(2021)] Shuqi Yang, Shiying Xiong, Yaorui Zhang, Fan Feng, Jinyuan
Liu, and Bo Zhu. 2021. Clebsch gauge fluid. ACM Transactions on Graphics
(TOG) 40, 4 (2021), 1–11.

[Yin et al.(2023)] Hang Yin, Mohammad Sina Nabizadeh, Baichuan Wu, Stephanie
Wang, and Albert Chern. 2023. Fluid cohomology. ACM Transactions on
Graphics (TOG) 42, 4 (2023), 1–24.

[Zehnder et al.(2018)] Jonas Zehnder, Rahul Narain, and Bernhard Thomaszewski.

2018. An advection-reflection solver for detail-preserving fluid simulation. ACM
Transactions on Graphics (TOG) 37, 4 (2018), 1–8.

[Zhang et al.(2015)] Xinxin Zhang, Robert Bridson, and Chen Greif. 2015. Restoring
the missing vorticity in advection-projection fluid solvers. ACM Transactions
on Graphics (TOG) 34, 4 (2015), 1–8.

[Zhao et al.(2020a)] Shuang Zhao, Wenzel Jakob, and Tzu-Mao Li. 2020a. Physics-
based differentiable rendering: from theory to implementation. In ACM SIG-
GRAPH 2020 Courses. 1–30.

[Zhao et al.(2020b)] Shuang Zhao, Wenzel Jakob, and Tzu-Mao Li. 2020b. Physics-
based differentiable rendering: from theory to implementation. In ACM siggraph
2020 courses. 1–30.
 Index

1st Piola–Kirchhoff stress, 92, 103 Cycles, 33

2nd Piola–Kirchhoff stress, 92, 103
Decomposable tensor, 94
Action, 91 Deformation gradient, 27, 39, 92, 99
Adjoint, 18 Deformation map, 90
Algebra, 49 Derivation, 49, 50
Annihilator, 17 Diagonalization, 20
Differential, 25
Back-propagation, 29
Bilinear form, 18 Differential form, 35, 37
Dirac δ form, 75
Cartan’s magic formula, 60 Dirichlet energy, 81
Cauchy stress, 92, 105 Distribution, 52
Cauchy–Binet identity, 42 Divergence, 34
Cauchy–Green tensor, 100 Divergence theorem, 34, 47
Cayley–Hamilton theorem, 111 Dual basis, 19
Cell complex, 73 Dual norm, 77
Chain, 73 Dual pairing, 15
Chain rule, 27, 46 Dual space, 16
Circulation, 32, 35
Conformal energy, 112 Elasticity, 91
Continuum mechanics, 89 derivation, 102
Coordinate system, 25 postulate, 99, 108
Cotangent bundle, 25 Embedding, 69
Covariant derivative, 54, 59 Endomorphism, 18
Covector, 16 Euler equation, 116
Covector field, 24, 25 barotropic, 118
Cramer’s rule, 98 incompressible, 119
Current, 75 Euler–Lagrange equation, 84
INDEX 133

Eulerian coordinate, 90 Kelvin’s circulation theorem, 121

Eulerian representation, 74 Kelvin–Stokes theorem, 35, 46
Exterior calculus, 31 Kinetic energy, 91
discrete, 73 Kirchhoff stress, 92
Exterior covariant derivative, 55
Exterior derivative, 36, 44, 48 Lagrangian coordinate, 90
Extrusion, 43 Lagrangian representation, 69
Lamé constants, 109
Fiber, 24 Leibniz’s integral rule, 64
Flat, 21 Levi-Civita connection, 55
Flow map, 57, 90 Lie advection, 58
Fluid Lie algebra, 50
barotropic, 115 Lie bracket, 51
incompressible, 115 Lie derivative, 56, 57
Fluid dynamics, 115 differnetial form, 60
Fluid velocity, 117 function, 58
Flux, 33 vector field, 58
Four fundamental subspaces, 18 Lie material derivative, 58
Function graph, 70 Lin’s constraint, 119
Fundamental theorem of calculus, 46
Fundamental theorem of Riemannian ge- Main invariants, 110
ometry, 55 Manifold, 29
Material derivative, 62
Geometric measure theory, 74 Material space, 69, 90
Gradient, 26 Metric, 21
Gradient descent, 86 Musical isomorphism, 21
Green’s theorem, 34
Neo-Hookean model, 112
Helmholtz’s free energy, 108 Newton’s method, 87
Hodge star, 36, 64 Nominal stress, 92
Holonomic constraint, 52 Non-holonomic constraint, 52
Homomorphism
Orthonormal basis, 22
linear, 96
Partial derivative, 25
Identity 1-form, 56
Piola identity, 46
Impulse method, 121
Plateau problem, 77
Integrability
Polar, 22
distribution, 52
Polar cone, 85
Integration, 38, 47
Polar decomposition, 20
Integration pairing, 72
Polarization, 19, 21
Interior product, 36, 39, 48
Pole, 22
Internal energy, 108
Positive definite, 22
Isometry, 20, 23
Pressure, 113, 117
Isotropic material, 109
Principal invariants, 110
Jacobian, 27 Pullback, 27, 34, 36, 38, 47
covector, 27
Karush–Kuhn–Tucker conditions, 84 differential form, 34, 38, 47
134 INDEX

exterior derivative, 44
function, 28
interior product, 43
metric, 100
vector field, 58
wedge product, 43
Pullback bundle, 96
Pushforward, 26

Quadratic form, 19

Reynolds’ transport theorem, 63

Section, 25
Sharp, 21
Singular value decomposition, 24
St. Venant–Kirchhoff model, 109
Stokes theorem, 45
Strain, 108
Strain energy density function, 101, 107
Stress–strain relation, 102

Tangent bundle, 24
Tangent space, 24
Tangent vector, 24
Tensor, 93
Tensor power, 94
skew-symmetric, 95
symmetric, 95
Tensor product, 93
Tensor product bundle, 94
Traction force, 92

Unitary, 23

Vector, 16
Vector fields, 24
Vector quadruple product, 42
Vector triple product, 42
Vorticity, 120

Wedge product, 36, 39, 48, 95

World space, 69, 90