Quantum Mechanics and Geometry

(preliminary draft updated Dec 2024)

(Si Li)

Tsinghua University
 1

V (x)

x
a b
ˆ br
m
−i dx
a 2(V (x) − E)
| {z }
imaginary time ( ) for quantum tunneling

Thanks very much for your support of this note! Comments and suggestions are greatly
appreciated. You can contact me at [email protected]. The draft will be updated on
my homepage:

https://sili-math.github.io/

1
See Section 2.7.3

1
Contents

Preface 6

Part I Basic Principles 8

Chapter 1 Hilbert Space Formalism 9

1.1 State Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.1 Classical State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.2 Quantum State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.3 Schrödinger Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 Classical Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.2 Quantum Observables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.3 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3 Measurement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.1 Born Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.3.2 Collapse of the State Vector . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3.3 The Double Slit Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4 Uncertainty Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.1 Commutator and Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.2 Some Mathematical Subtleties . . . . . . . . . . . . . . . . . . . . . . . . 22
1.5 Wave Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.5.1 Hilbert Space of Quantum Particle . . . . . . . . . . . . . . . . . . . . . . 24
1.5.2 Wave Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5.3 Position and Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5.4 Hamiltonian Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.5.5 Stationary States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.6 Free Particle: Example of Continuous Spectrum . . . . . . . . . . . . . . . . . . . 29
1.6.1 Wave Packet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.6.2 Group Velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.7 Harmonic Oscillator: Example of Discrete Spectrum . . . . . . . . . . . . . . . . 32
1.7.1 Ladder Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
1.7.2 Ground State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

2
 1.7.3 Excited States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.8 Square Well: Example of Mixed Spectrum . . . . . . . . . . . . . . . . . . . . . . 37
1.8.1 Matching Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.8.2 Bound States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
1.8.3 Scattering States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
1.9 Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.9.1 Wave Packet Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.9.2 S-matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.9.3 Unitarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
1.9.4 Time Reversal Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
1.10 WKB Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.10.1 Approximation Scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.10.2 Turning Points and Airy Functions . . . . . . . . . . . . . . . . . . . . . . 55
1.10.3 Connection Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.10.4 Semi-classical Quantization Rule . . . . . . . . . . . . . . . . . . . . . . . 58
1.10.5 Quantum Tunneling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
1.11 Quantum Kepler Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.11.1 Angular Momentum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.11.2 Enhanced Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.11.3 Representations of so(3) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1.11.4 Energy Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
1.11.5 Hydrogen Atom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Chapter 2 Path Integral Formalism 71

2.1 Path Integral: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.1.1 Quantum Evolution and Feynman Kernel . . . . . . . . . . . . . . . . . . 71
2.1.2 Position and Momentum Representation . . . . . . . . . . . . . . . . . . . 73
2.2 Path Integral via Time Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.2.1 Free Particle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.2.2 Infinitesimal Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.2.3 Composition Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.2.4 Path Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
2.2.5 Imaginary Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.3 Gaussian Path Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.3.1 Gaussian Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.3.2 Zeta Function Regularization . . . . . . . . . . . . . . . . . . . . . . . . . 81
2.4 Harmonic Oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.4.1 Integral Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.4.2 Partition Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
2.5 Asymptotic Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

3
 2.5.1 Laplace’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
2.5.2 Method of Steepest Descent . . . . . . . . . . . . . . . . . . . . . . . . . . 89
2.5.3 Morse Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
2.5.4 Stokes Phenomenon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.6 Semi-classical Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
2.6.1 Semi-classical Feynman Kernel . . . . . . . . . . . . . . . . . . . . . . . . 99
2.6.2 Jacobi Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
2.6.3 Time-slicing Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2.7 Green’s Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
2.7.1 Green’s Function with Fixed Energy . . . . . . . . . . . . . . . . . . . . . 108
2.7.2 Semi-classical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
2.7.3 WKB via Path Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

Part II Geometric Perspectives 129

Chapter 3 Phase Space Geometry 130

3.1 Symplectic Geometry of Phase Space . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.1.1 Symplectic Vector Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
3.1.2 Lagrangian Grassmannian . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
3.1.3 Maslov Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.1.4 Symplectic Manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
3.2 Semi-classical Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.2.1 Semi-classical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
3.2.2 Half-Density . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.2.3 Maslov Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
3.3 Heisenberg Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.3.1 Heisenberg Lie Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
3.3.2 Heisenberg Group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3.3.3 Schrödinger Representation . . . . . . . . . . . . . . . . . . . . . . . . . . 153
3.3.4 Bargmann-Fock Representation . . . . . . . . . . . . . . . . . . . . . . . . 155
3.4 Weyl Wigner Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.4.1 Weyl Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.4.2 Moyal Product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
3.4.3 Wigner Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
3.5 Geometric Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
3.5.1 Dirac Quantization Principle . . . . . . . . . . . . . . . . . . . . . . . . . 167
3.5.2 The Groenewold-Van Hove Theorem . . . . . . . . . . . . . . . . . . . . . 169
3.5.3 Prequantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
3.5.4 Polarization and Quantum States . . . . . . . . . . . . . . . . . . . . . . . 175

4
 3.5.5 Quantum Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
3.6 Path Integral in Phase Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
3.6.1 Wick’s Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
3.6.2 Feynman Graph Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . 187
3.6.3 Weyl Quantization Revisited . . . . . . . . . . . . . . . . . . . . . . . . . 191
3.6.4 S 1 -Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
3.6.5 Hochschild Homology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

Chapter 4 Deformation Quantization 206

4.1 Deformation Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
4.1.1 Formal Deformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
4.1.2 Poisson Algebra as Classical Limit . . . . . . . . . . . . . . . . . . . . . . 208
4.2 Geometric Approach on Symplectic Manifolds . . . . . . . . . . . . . . . . . . . . 210
4.2.1 Weyl Bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
4.2.2 Symplectic Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
4.2.3 Fedosov’s Abelian Connection . . . . . . . . . . . . . . . . . . . . . . . . . 215
4.2.4 Symbol Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
4.2.5 Globalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
4.3 Poisson Manifold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
4.3.1 Polyvector Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222
4.3.2 Poisson Manifold and Poisson Cohomology . . . . . . . . . . . . . . . . . 223
4.3.3 Kontsevich’s Star Product . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
4.4 Trace Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
4.4.1 Differential Graded Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . 230
4.4.2 Quantum HKR Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
4.4.3 Twisting by Fedosov’s Abelian Connection . . . . . . . . . . . . . . . . . . 236
4.4.4 Trace Map and Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239

Bibliography 242

5
 Preface

In April 2021, Qiuzhen College ( ) was newly established at Tsinghua University

under the leadership of Professor Shing-Tung Yau. It homes the distinguished elite mathematics
program in China starting in 2021: the “Yau Mathematical Sciences Leaders Program” (
). This program puts strong emphasis on basic sciences related to
mathematics in a broad sense. Though majored in mathematics, students in this program are re-
quired to study fundamental theoretical physics such as classical mechanics, electromagnetism,
quantum mechanics, and statistical mechanics, in order to understand global perspectives of
theoretical sciences. It is an exciting challenge both for students and for instructors.

This preliminary note is intended for the course “Quantum Mechanics” that I lectured at
Tsinghua University in 2023 and 2024. It is a continuation of [36, 37] in this series. The first
part of the note is physics-oriented and aims to elucidate the fundamentals of quantum mechan-
ics. Chapter 1 explains the foundational principles of traditional Hilbert space and operator
approach to quantum mechanics. This formalism establishes a framework for describing the
evolution of quantum systems in terms of unitary transformations and measurements. Chap-
ter 2 explains the path integral approach to quantum mechanics, which characterizes quantum
dynamics of particles in terms of probabilistic paths. This formalism provides a powerful tool
for calculating transition amplitudes and understanding quantum phenomena, and has been
widely generalized and developed within modern quantum field theory. These two equivalent
formalisms are deeply intertwined and illustrate different faces of the exciting development of
quantum mechanics.

The second part of the note is mathematics-oriented with emphasis on modern geometric
and algebraic formulations of quantization. Chapter 3 explains the phase-space formalism of
quantum mechanics as well as the theory of geometric quantization and perturbative approach
to path integrals in phase space. Chapter 4 is focused on the algebraic theory of deformation
quantization which deals with the operator of quantum observables as a formal deformation
of classical observables. In this theory, the Planck’s constant h̄ is treated as a formal variable
parametrizing a family of quantum algebras.

Two parts of the note are closed related and we explain the bridge along the way. Some
useful resources that we consulted are listed at the end of this note.

6
7
 Part I

Basic Principles

8
 Chapter 1 Hilbert Space Formalism

In this chapter, we explain the foundational principles of Hilbert space and operator ap-
proach to quantum mechanics. This formalism establishes a framework for describing the evo-
lution of quantum systems in terms of unitary transformations and measurements.

1.1 State Space

1.1.1 Classical State

One fundamental difference between classical mechanics and quantum mechanics is how a
state of the system is described. Recall that in classical mechanics, the state of a particle is
usually described by a point in the phase space M , whose geometry is described by a symplectic

manifold. For most cases of our interest, M is parametrized by xi , pi , where

xi = (generalized) position,
p = (generalized) momentum conjugate to xi .
i

The time evolution of the state is described by a trajectory parametrized by the time:

xi (t), pi (t) : state at time t.

In the Hamiltonian formalism, this classical evolution is described by Hamilton’s equations

 i

 dx ∂H
 = ,
dt ∂pi


 dpi = − ∂H .
dt ∂xi
Here H is a function on the phase space, called the Hamiltonian function. The system of
Hamilton’s equations determines the evolution of the state.

1.1.2 Quantum State

In quantum mechanics, the state of the particle is represented by a vector in a Hilbert space
V. We will always assume that the Hilbert space V is separable.

Definition 1.1.1. A Hilbert space is a C-linear vector space equipped with a positive definite
Hermitian inner product such that the space is complete with respect to the induced norm.

9
 One can also talk about R-linear Hilbert spaces, but we will focus on the C-linear case. As
we will see, the angular phase factor will play an important role in quantum mechanics.
Let V be a Hilbert space. The Hermitian inner product is a sesquilinear pairing

h−|−i : V × V −→ C

which is conjugate linear in the first argument and linear in the second:

hλψ|ϕi = λhψ|ϕi
∀λ ∈ C, ψ, ϕ ∈ V.
hψ|λϕi = λhψ|ϕi

The Hermitian property says

hψ|ϕi∗ = hϕ|ψi ∀ψ, ϕ ∈ V.

The positive definite property says


hψ|ψi ≥ 0 ∀ψ ∈ V
hψ|ψi = 0 if and only if ψ = 0.

The induced norm on V is given by

p
kψk := hψ|ψi, ψ ∈ V.

Being a Hilbert space, V has to be complete with respect to the norm k·k.

Example 1.1.2. V = Cn . A vector u ∈ V is represented by

 
z1
.
u= .
 . , zi ∈ C.
zn

We can define the standard Hermitian inner product by

X
hu|vi = z i wi
i
 T  T
for u = z1 · · · zn , v = w1 · · · wn . Then Cn is a finite dimensional Hilbert space.

Example 1.1.3. V = L2 (R). For any f, g ∈ L2 (R), define

ˆ
hf |gi = f (x)g(x)dx.
R

This defines an infinite dimensional Hilbert space.

Remark 1.1.4. We will use Dirac’s bra-ket notation and write a vector ψ ∈ V as a “ket”

|ψi ∈ V.

10
A “bra” hϕ| for ϕ ∈ V represents the linear form

hϕ| : V −→ C
|ψi 7−→ hϕ|ψi.

In other words, the “bra” is related to the “ket” in terms of Hermitian conjugate.
For example of V = Cn , given
   
z1 w1
.  . 
u= .
 . , v= . 
 .  ∈ V,
zn wn

we can represent  
w1
 . 
1 “ket”: |vi = 
⃝ . 
 . 
w
 n 
⃝
2 “bra”: hu| = z 1 · · · zn .
Then the Hermitian inner product is
 
w1
   
.. 
hu|vi = z 1 · · · zn · 
 . .
wn

So what is the essential change for the description of state from classical to quantum?

Classical State Quantum State

Set Vector space

The algebraic structure on the space of states has been enhanced from a set to a vector
space! This in particular implies that we can take a linear superposition of two quantum states
|ψ1 i, |ψ2 i ∈ V to form another quantum state

λ1 |ψ1 i + λ2 |ψ2 i ∈V

where λ1 , λ2 ∈ C. This linear operation is not allowed on the set of classical states. As we will
see later in this note, many interesting quantum phenomena come out of this linear structure.
To get a first impression, consider a quantum system described by a two dimensional Hilbert
space spanned by
|↑i, |↓i.

For example, this could be a spin system, where |↑i indicates a state with “spin up”, and
|↓i indicates a state with “spin down”. Then we can take a combination to obtain a state
1 1
√ |↑i + √ |↓i.
2 2
This is a new state. But what is its spin? You should keep this simple question in mind along
the way. It is as good as the state |↑i or the state |↓i.

11
 As another example for illustration, let us consider how to describe the state for two
particles. In the classical case, if the set of states for particle i is Mi , i = 1, 2, then the state
space for the total is
M1 × M2 .

In the quantum case, if the state space for particle i is Vi , then the naive set-theoretical
product V1 × V2 does not work. This is simply because V1 × V2 is not a linear vector space.
Instead, we have to consider its linearization and arrive at the tensor product

V1 ⊗ V2 .

A main difference between M1 × M2 and V1 ⊗ V2 is that a point in M1 × M2 is of the form

m1 × m2 where m1 ∈ M1 , m2 ∈ M2

while a vector in V1 ⊗ V2 could be of the form

X
k
ψ i ⊗ ϕi where ψi ∈ V1 , ϕi ∈ V2 .
i=1

For example, this tensor structure is the origin of quantum entanglement.

1.1.3 Schrödinger Equation

We next discuss the law governing the time evolution of quantum states.

Definition 1.1.5. Let A be a linear operator on the Hilbert space V with a dense domain
Dom(A) ⊂ V. The adjoint operator A∗ of A is defined by

A∗ : Dom(A∗ ) −→ V

where the domain of A∗ is

n D E o
Dom(A∗ ) = ψ ∈ V ∃ψ̃ ∈ V : hψ|Aϕi = ψ̃ ϕ , ∀ϕ ∈ Dom(A) .

Here, for ψ ∈ Dom(A∗ ) and ψ̃ as above,

A∗ ψ = ψ̃.

Symbolically, the defining relation for the adjoint is

hA∗ ψ|ϕi = hψ|Aϕi

for ϕ ∈ Dom(A), ψ ∈ Dom(A∗ ).

Definition 1.1.6. If A = A∗ , then A is called self-adjoint.

12
Remark 1.1.7. When V is finite dimensional and A is represented by a matrix, A∗ is given by the
conjugate transpose. There a self-adjoint operator is the same as a Hermitian matrix. When V
is infinite dimensional, which is of our interest in quantum mechanics, self-adjoint operators are
much more delicate and we have to work with densely defined domain. Nevertheless, interested
readers can refer to standard context in functional analysis to clarify related statements.

In quantum mechanics, the time evolution of quantum state |ψ(t)i obeys the

d ^ |ψ(t)i.
Schrödinger equation : ih̄ |ψ(t)i = H
dt
^ is a self-adjoint operator, called the Hamiltonian operator. The positive number h̄ is
Here H
called the Planck constant, which is one of the fundamental constants in physics.
Assume H ^ is time-independent, then the Schrödinger equation can be formally solved by

^
|ψ(t)i = e−i H t/h̄ |ψ(0)i.

^
The operator e−i H t/h̄ (which can be defined using functional calculus) is a one-parameter
family of unitary operators on V, known as the time-evolution operators. Conversely, a strongly
continuous one-parameter unitary groups is generated infinitesimally by a self-adjoint operator.
This is known as Stone’s Theorem.

1.2 Observables

1.2.1 Classical Observables

In classical mechanics, observables are represented by functions on the phase space M

Obscl = O(M ).

Here O(M ) means (appropriate) space of functions on M . We will not specify the class of
functions to avoid technical discussions unnecessary for our purpose. You can safely work with
smooth functions here.
Since M is a symplectic manifold, the space of functions O(M ) on it naturally carries a
structure of Poisson bracket

{−, −} : O(M ) × O(M ) −→ O(M ).

It is a bilinear map and satisfies the following properties

• Skew-symmetry: {f, g} = −{g, f }.

• Leibniz rule: {f, gh} = {f, g}h + g{f, h}.

• Jacobi identity: {{f, g}, h} + {{g, h}, f } + {{h, f }, g} = 0.

13
 
For example, assume the phase space M is parametrized by xi and their conjugates {pi },
such that the symplectic structure takes the form
X
dxi ∧ dpi .
i

Then a classical observable is a function f (x, p) of xi , pi . The Poisson bracket takes the form
X  ∂f ∂g ∂f ∂g

{f, g} = − .
∂xi ∂pi ∂pi ∂xi
i

1.2.2 Quantum Observables

In quantum mechanics, quantum observables consist of self-adjoint linear operators on the

Hilbert space V of the system. We have seen one such quantum observable, the Hamiltonian
^ which governs the dynamical evolution of the system via the Schrödinger equation.
H,
Let Obsq denote the space of quantum observables. Let A, B be two self-adjoint operators

A = A∗ , B = B∗.

We consider their commutator

[A, B] := AB − BA.

Its adjoint is
[A, B]∗ = B ∗ A∗ − A∗ B ∗ = [B ∗ , A∗ ]
= [B, A] = −[A, B].
Therefore i[A, B] is again self-adjoint. The combination
i
[−, −]h̄ := − [−, −] : Obsq × Obsq −→ Obsq
h̄
defines the quantum bracket on quantum observables. Here the positive number h̄ is the Planck
constant. The quantum bracket is again skew-symmetric

[A, B]h̄ = −[B, A]h̄

and satisfies the Jacobi identity

[A, [B, C]h̄ ]h̄ + [B, [C, A]h̄ ]h̄ + [C, [A, B]h̄ ]h̄ = 0.

These properties follow from the associativity of the composition of operators.

The algebraic relation between classical and quantum observables can be summarized by

Obscl Obsq

quantization
Poisson algebra Associative algebra

classical limit
(·, {−, −}) (∗h̄ , [−, −]h̄ )

14
The classical Poisson bracket {−, −} measures the leading order noncommutativity of quantum
operators. Strictly speaking, we have extended Obsq to contain linear operators that are not
self-adjoint, in order to perform compositions to form an associative algebra. The study of such
algebraic correspondence is the main subject of deformation quantization. We will discuss in
details about this in Chapter 4.

1.2.3 Expectation

Let V be the Hilbert space of a quantum system. Let O be a quantum observable, which is
a self-adjoint operator on V. Assume the system is at the state |ψi ∈ V. The expectation value
of O of the system is defined to be

hψ|O|ψi
hOiψ := .
hψ|ψi

Here O|ψi = |Oψi is the action of O on the state |ψi . It would be convenient to normalize the
state such that
p
kψk = hψ|ψi = 1.

In many contexts, states of quantum system are referred to normalized vectors in the
Hilbert space. Then in the normalized case

hOiψ = hψ|O|ψi.

Note that the expectation values of quantum observables are real numbers. In fact,

O∗ =O
hψ|O|ψi = hψ|Oψi = hO∗ ψ|ψi ===== hOψ|ψi = hψ|Oψi = hψ|O|ψi

=⇒ hψ|O|ψi ∈ R.

Here we have used the self-adjoint property of O. This implies that eigenvalues of O are real.
Assume |ψi is a normalized eigenvector of A with eigenvalue λ, then
By our discussion above, λ has to be a real number. As we will discuss soon later, eigen-
values are the outcomes of physical measurement for quantum observables.

1.3 Measurement

1.3.1 Born Rule

We have discussed two basic postulates of quantum mechanics:

• a quantum state is a vector in a Hilbert space V.

• a quantum observable is a self-adjoint operator on V.

Now we turn to the interpretation of measurement for a quantum observable in a given

quantum state. The first statement is

15
 • possible outcomes of quantum measurements are eigenvalues of quantum observables.

Let O be a quantum observable, which is a self-adjoint operator on V. As we have seen,

self-adjointness implies that eigenvalues of O are all real numbers. The above postulate says
that these real numbers are the possible outcomes of measurements of O.
Assume the system is now at a state |ψi ∈ V. How come to obtain an eigenvalue of O out
of the quantum state |ψi under a measurement?
Firstly, assume |ψi is an eigenvector of O with eigenvalue λ

O|ψi = λ|ψi.

Then it is natural to expect that the measurement of O in the state |ψi is λ. This is indeed the
case. However, in general |ψi may not be an eigenvector of O, What do we get? The answer is
probabilistic, known as the Born rule.
Let us first mention that two vectors in V, which are proportional by a nonzero complex
number, represent the same physical states. In other words, each physical state is represented by
a “ray” in the Hilbert space V. For example, the expectation value of the quantum observable
O in a state |ψi
hψ|O|ψi
hOiψ =
hψ|ψi
is the same if you replace |ψi by a|ψi for any complex number a ∈ C∗ . Therefore we usually
normalize the state and represent a physical state by a vector of length 1

hψ|ψi = 1.

We will call such vector of unit length a normalized state.

Let O be a given quantum observable. Assume |ψ1 i and |ψ2 i are eigenvectors

O|ψ1 i = λ1 |ψ1 i, O|ψ2 i = λ2 |ψ2 i

with different eigenvalues λ1 6= λ2 . Then these two vectors |ψ1 i and |ψ2 i must be orthogonal

hψ1 |ψ2 i = 0.

This follows from the self-adjointness of O

λ2 hψ1 |ψ2 i = hψ1 |Oψ2 i = hOψ1 |ψ2 i = λ1 hψ1 |ψ2 i

=⇒ hψ1 |ψ2 i = 0.

If |ψ1 i and |ψ2 i have the same eigenvalue λ1 = λ2 , then they may not be orthogonal.
Nevertheless, we can apply the Gram-Schmidt procedure to rearrange them to be orthogonal.
Let us assume that the quantum state |ψi of the system under measurement can be de-
composed in terms of orthonormal eigenstates of the quantum observable O
X
|ψi = cα |ψα i, cα ∈ C.
α

16
Here O|ψα i = λα |ψα i for eigenvalues λα , and |ψα i’s are normalized vectors and orthogonal with
each other for different α’s. Things could be complicated in general where some λα ’s could
be the same (so with degenerate eigenspace) or eigenvalue spectrum could be continuous (see
Section 1.8 for one example). Let us not worry about these issues at the moment.
Assume |ψi is normalized. The normalization condition hψ|ψi = 1 implies (using the
orthonormal property of |ψα i’s)
X
|cα |2 = 1.
α
One interprets this formula as probability.
Born rule: the result of measurement of quantum observable O in the quantum state |ψi is
probabilistic: the probability of the measurement with answer λ is
X
P (λ) = |cα |2 .
λα =λ

Using the orthonormal property, the complex number cα can be obtained by

cα = hψα |ψi

which is called the probability amplitude. Thus the Born rule can be stated as

Probability = |Amplitude|2

Remark 1.3.1. In the case of continuous spectrum, we need to invoke spectral theorem to give
probability measure.

1.3.2 Collapse of the State Vector

The experiment for measurement in quantum mechanics involves the composite systems
essentially. Said in another way, any measurement of a quantum system (with an external
apparatus) will disturb the state. In fact, it is even more mysterious. After measurement of a
quantum observable O, with outcome of an eigenvalue λ, the state will change into an eigenstate
of the corresponding eigenvalue λ. This phenomenon is called the

“collapse of the state vector”.

Since the state becomes an eigenvector after the measurement, if there is no other disturbance
of the system, then the further repeated measurement of O will always produce λ.
There are a collection of views about the meaning of quantum mechanics, including the
above Born rule and collapse of the state vector. It coins the term “Copenhagen interpretation”.
It is not entirely satisfactory, and leads to many confusions and debates in the history. Concern-
ing the collapse, there is also the “many-worlds interpretation” by Hugh Everett, which assumes
that the state vector does not collapse, but all possible outcomes of quantum measurements are
realized in some parallel worlds. We will not go further into these issues here.
Nevertheless, quantum mechanics works, and has been one of the most successful framework
in physics. Many people take practical viewpoints. Here is a famous quote by David Mermin:

17
“If I were forced to sum up in one sentence what the Copenhagen interpretation says to me, it
would be ‘Shut up and Calculate!’ ”
So we see that even though the evolution of the state is deterministic, captured by the
Schrödinger equation, the prediction for the measurement is probabilistic. Furthermore, the
measurement will change the state. Then how do we test quantum mechanics?
To test such probabilistic prediction, we need to prepare a quantum ensemble consisting
of a large number of particles in the same state. Then we do the same measurement for each
particle. The outcome of each measurement ends up with some eigenvalue, which may differ
in a repeated measurement for another particle in the same state. Then we can collect the
statistical result of these measurement to check with the probabilistic prediction.

1.3.3 The Double Slit Experiment

Consider the following experiment set-up: there is a source S on the left which is able to
emit particles in all directions; there is a wall in the middle with two small holes on it; there is
a screen on the right which is a detector that can record each particle when it arrives.

S
2

source wall detector

We are looking for the outcome distribution of particle counts on the detector.
In the classical picture, if we close the hole 2, we find a distribution of particle counts on
the detector that come from the hole 1 only. Let us call this P1 . Similarly, if we close the hole
1 and leave the hole 2 open, we find a distribution P2 that counts particles that come from the
hole 2. Now if we open both holes, we shall see the distribution on the detector as the sum of
the above two
P = P1 + P2 .

The resulting distribution P would look like

1 P
Classical Picture:
2

wall detector

18
 In the quantum case, we have a superposition of the above two situations. From the linear
relation
hφ|ψ1 + ψ2 i = hφ|ψ1 i + hφ|ψ2 i

we see that it is the probability amplitude that is summed over from each case

A = A1 + A2 .

As we have seen, the probability distribution of the measurement is the square of the
amplitude
P1 = |A1 |2 , P2 = |A2 |2 ,

P = |A|2 = |A1 + A2 |2 .

Since the amplitudes are complex numbers,

P 6= P1 + P2 ,

we will observe interference pattern like that for light waves.

P

1
Quantum Picture:
2

wall detector

Now this can be tested for a beam of quantum particles, say electrons. The outcome indeed
exhibits the interference pattern of the quantum picture.
This is surprising. One can try to set up another apparatus near the hole to tell us each
time which hole each particle passes through the wall. To measure this, you will necessarily
disturb the state of the particle. For example, you may try to achieve this by using photons to
detect the electron at the hole. Such measurement will collapse the state vector. As a result, you
will find the interference pattern disappears, and we arrive at the classical picture. Quantum
particles behave as both waves and particles! This is the concept of wave-particle duality.
The double slit experiment was originally designed to demonstrate the wave behavior of
light. It was de Broglie who made the brave postulation that all matters have wave properties.
This prediction was soon verified for electrons.

1.4 Uncertainty Principle

1.4.1 Commutator and Uncertainty

Measurement of a quantum observable O will produce an eigenvalue of O and collapses

the state into an eigenvector of O. It is natural to ask whether we can measure two quantum

19
observables simultaneously. For example, we would like to measure the position and the mo-
mentum. This is not a problem in classical mechanics. However, we will run into trouble in
quantum mechanics.
Let A, B be two quantum observables. Our postulate on measurement asks for common
eigenvectors of A and B under a simultaneous measurement of them. As we have learned in
linear algebra, common eigenvectors may not exist if A and B do not commute. For example,
assume the case when A and B satisfy

[A, B] = iI

where I is the identity operator. Assume |ψi is a common eigenvector of A and B

A|ψi = λ1 |ψi, B|ψi = λ2 |ψi.

Then
i|ψi = [A, B]|ψi = (AB − BA)|ψi = (λ1 λ2 − λ2 λ1 )|ψi = 0.

Thus the problem of simultaneous measurement is related to the commutator of quantum

observables. Let us first introduce the notion of uncertainty of a quantum observable in a state.

Definition 1.4.1. The uncertainty of O in a quantum state |ψi is defined to be

 2 
2
(∆ψ O) := O − hOiψ .
ψ

Explicitly, let |ψi be a normalized state. Since hOiψ ∈ R, we have

D    E
(∆ψ O)2 = O − hOiψ ψ O − hOiψ ψ .

In particular, ∆ψ O = 0 if and only if

O|ψi = hOiψ |ψi,

i.e. when |ψi is an eigenvector of O. In this case, the measurement of O is certain, given by the
corresponding eigenvalue. In general, the uncertainty measures the average fluctuation around
the expectation. In probability, the uncertainty is also called the standard derivation. Note
that we also have
D E
(∆ψ O)2 = O2 − 2 hOiψ O + hOi2ψ = O2 ψ
− hOi2ψ .
ψ

A precise form of the uncertain principle is the following statement.

Proposition 1.4.2. Consider two quantum observables A, B in a quantum state |ψi. Then
1 2
(∆ψ A)2 (∆ψ B)2 ≥ h[A, B]iψ .
4

20
Proof: If one of ∆ψ A or ∆ψ B is zero, say ∆ψ A = 0. Then ψ is an eigenstate of A, and

h[A, B]iψ = hAψ|Bψi − hBψ|Aψi = hAiψ (hψ|Bψi − hBψ|ψi) = 0.

Thus we only need to consider the case when both ∆ψ A and ∆ψ B are nonzero.
Let A = A − hAiψ I and B = B − hBiψ I. Then we have A ψ = B ψ
= 0 and
 
∆ψ A = ∆ψ A, ∆ψ B = ∆ψ B, A, B = [A, B].

Therefore we can assume hAiψ = hBiψ = 0 without lost of generality.

Let |ψi be normalized. Then

(∆ψ A)2 = A2 = hψ|A2 |ψi
ψ
(∆ B)2 = B 2 = hψ|B 2 |ψi
ψ ψ

Since A, B are self-adjoint, the following quantities

A2 ψ
, B2 ψ
, i h[A, B]iψ

are all real numbers. Let γ1 , γ2 be two arbitrary real numbers. Then

0 ≤ k(γ1 A + iγ2 B) ψk2

= h(γ1 A + iγ2 B) ψ|(γ1 A + iγ2 B) ψi
= hψ|(γ1 A − iγ2 B) (γ1 A + iγ2 B)|ψi
= (∆ψ A)2 γ12 + i h[A, B]iψ γ1 γ2 + (∆ψ B)2 γ22 .
Since this holds for arbitrary real values of γ1 , γ2 , we have
1 2
(∆ψ A)2 (∆ψ B)2 ≥ h[A, B]iψ .
4

Remark 1.4.3. From the above proof, it is not hard to see that the equality in the above
proposition holds if and only if there exists real numbers (γ1 , γ2 ) 6= (0, 0) such that |ψi is an
eigenvector of γ1 A + iγ2 B.
Proposition 1.4.2 states that if h[A, B]iψ 6= 0, then there is a lower bound on the product
of the uncertainties of A and B. For readers who are careful about linear operators on Hilbert
spaces, you may worry about the domain for the operators in Proposition 1.4.2. There is indeed
such mathematical subtleties. We will briefly comment shortly in the next subsection.

Example 1.4.4. As we will discuss extensively later, the position operator x̂ and momentum p̂
in quantum mechanics satisfy the following commutation relation

[x̂, p̂] = ih̄.

Then the uncertainty inequality implies

h̄
(∆ψ x̂) (∆ψ p̂) ≥ .
2
This is the celebrated Heisenberg uncertainty relation. It says that in any quantum system,
we can not make precise measurement for both the position and momentum simultaneously.

21
1.4.2 Some Mathematical Subtleties

In quantum mechanics, we will often encounter a pair of quantum observables satisfying

[A, B] = ih̄.

This is called the “canonical commutation relation”. They arise naturally from the quantum
observables associated to canonical conjugate quantities. The position x̂ and momentum p̂ is
such an example.
For canonical commutation relation to hold, the Hilbert space V is necessarily infinite
dimensional. In fact, assume dimC V < ∞, and A, B are two linear operators on V satisfying
the canonical commutation relation. Then

0 = Tr[A, B] = Tr(ih̄) = ih̄ dimC V.

This is a contradiction.
Let us now assume V is an infinite dimensional Hilbert space. We know that bounded and
unbounded linear operators behave very differently. For examples, bounded linear operators
can be defined on the whole V, while unbounded linear operators can only be defined on dense
subspaces of V. The domains for relevant operators become a subtle issue, and we have to live
^ which captures the dynamical evolution
with that. For example, the Hamiltonian operator H,
of quantum states, is usually an unbounded operator. The next proposition reveals another
essential appearance of unbounded operators.

Proposition 1.4.5. Assume A, B are two self-adjoint operators on V that satisfy the canonical
commutation relation
[A, B] = ih̄.

Then at least one of A, B is unbounded.

Proof: Let h̄ = 1 for simplicity. Assume both A and B are bounded, so their operator norms
kAk < ∞, kBk < ∞. Since B is self-adjoint, we have kB n k = kBkn .
Inductively, one can show

[A, B k ] = i kB k−1 , k ≥ 1.

Then
k kBkk−1 = i kB n−1 = [A, B k ] ≤ 2 kAk B k ,

which implies that

k
kAk kBk ≥ , ∀k ≥ 1.
2
This is a contradiction.

Let Dom(A) denote the domain of A. A mathematical precise statement of Proposition

1.4.2 is that
1 2
(∆ψ A)2 (∆ψ B)2 ≥ h[A, B]iψ
4

22
holds for a state ψ belonging to the common domain of AB and BA, i.e., for

ψ ∈ Dom(AB) ∩ Dom(BA).

This indeed could lead to some mathematical subtleties. Here is one example.

Example 1.4.6. Consider the Hilbert space

V = L2 ([0, 1]).

A state ψ ∈ V is a function ψ(x) on x ∈ [0, 1] which is square integrable.

We define the position operator x̂ by multiplication with x

(x̂ψ) (x) := xψ(x).

x̂ is a bounded operator since x ∈ [0, 1].

We define the momentum operator p̂ by the differential operator
d
p̂ := −ih̄ .
dx
It is clear that the canonical commutation relation holds

[x̂, p̂ ] = ih̄.

This implies that p̂ must be unbounded, and so only densely defined.

For example, we can define p̂ on continuously differentiable functions on [0, 1] that satisfy
the periodic boundary condition
ψ(0) = ψ(1).

We can verify that for two such functions, the adjoint property holds:
ˆ 1  
d
hψ1 |p̂ψ2 i = ψ 1 (x) −ih̄ ψ2 (x) dx
0 dx
ˆ 1
1 d
= −ih̄ψ 1 (x)ψ2 (x) 0 + −ih̄ ψ1 (x)ψ2 (x)dx
0 dx
= 0 + hp̂ψ1 |ψ2 i = hp̂ψ1 |ψ2 i.

One can further show that we can extend such defined p̂ to a self-adjoint operator.
Now consider the following normalized state

ψ(x) = e2πix .

This is an eigenstate of p̂ since p̂ψ = 2πh̄ψ. Thus ∆ψ p̂ = 0. It is also straight-forward to

compute ∆ψ x̂ = √1 . On the other hand,
12

h[x̂, p̂]iψ = ih̄ 6= 0.

In this case, we find

(∆ψ x̂)2 (∆ψ p̂)2 = 0

23
while
1 2 1
h[x̂, p̂]iψ = h̄2 .
4 4
The uncertainty inequality of Proposition 1.4.2 fails in this case. The issue is about the domain.
x̂ψ is no longer periodic and does not lie in the domain of p̂.

In practice, we will work with nice state vectors in appropriate domain for relevant examples
in physics, so that the uncertainty inequality does hold. We will not check this subtlety all the
time. Careful readers can keep this in mind.

1.5 Wave Function

We now focus on studying quantum particles in the space Rn . Main examples will be
focused on one-dimensional case (n = 1) and three dimensional case (n = 3). We will use

x = (x1 , x2 , · · · , xn )

to parametrize positions in Rn .

1.5.1 Hilbert Space of Quantum Particle

The relevant Hilbert space that realizes the scalar quantum particle in Rn is

V = L2 (Rn ).

An element f ∈ V is a complex valued measurable function that is square integrable

ˆ
dn x |f (x)|2 < +∞.
Rn

Here dn x := dx1 dx2 · · · dxn is the standard measure on Rn . The Hermitian inner product is
ˆ
hg|f i := dn x g(x)f (x).
Rn

In particular, the norm of f is sˆ

kf k = dn x |f (x)|2 .
Rn

Thus a function f ∈ L2 (Rn ) will be also called normalizable. Otherwise, a function which
is not square integrable will be called non-normalizable.
As we have discussed before, physical states are represented by normalizable functions.
Nevertheless, non-normalizable functions will also play an important role. For example, given
any p = (p1 , · · · , pn ) ∈ Rn , the following function
X
eip·x/h̄ , where p · x = pi x i
i

is non-normalizable since
2
eip·x/h̄ =1 everywhere.

24
However, any normalizable function f can be expressed as a superposition
ˆ
1
f (x) = dn p eip·x/h̄ fˆ(p).
(2πh̄)n/2 Rn

This is the celebrated Fourier transform. The corresponding Fourier modes fˆ(p) can be also
obtained via the inverse Fourier transform
ˆ
1
fˆ(p) = dn x e−ip·x/h̄ f (x).
(2πh̄)n/2 Rn

1.5.2 Wave Function

Quantum mechanics describes a moving particle by a path (parametrized by the time t) in

the Hilbert space. Explicitly, this is realized by a complex valued function

Ψ(x, t)

such that Ψ(x, t) is normalizable at any fixed time t. This is called the wave function of the
quantum particle. The evolution of the wave function obeys the Schrödinger equation
∂Ψ ^ Ψ.
ih̄ =H
∂t

Here H^ is the Hamiltonian operator, which is a self-adjoint operator on the Hilbert space. We
^ shortly.
will discuss the form of H
First, we observe that the Schrödinger equation implies the invariance of norm under time
evolution. In fact,    
∂ ∂ ∂
ih̄ hΨ|Ψi = −ih̄ Ψ Ψ + Ψ ih̄ Ψ
∂t ∂t ∂t
D E D E
= −H^ΨΨ + ΨH ^Ψ

=0 ^
(using self-adjointness of H)
i.e., hΨ|Ψi does not depend on time t. Therefore we will assume the wave function is normalized
by hΨ|Ψi = 1 at any time, i.e., ˆ
dn x |Ψ(x, t)|2 = 1.
Rn
In the beginning, Schrödinger interpreted the wave function Ψ(x, t) as representing a parti-
cle that could spread out and disintegrate. There the magnitude of |Ψ(x, t)|2 would represent the
fraction density of the particle to be found at position x and time t. However, this turns out to
be inconsistent with experiments. Born figured out the solution and proposed the probabilistic
interpretation:
ˆ
dn x |Ψ(x, t)|2 = probability of finding the particle in the region V ⊂ Rn at time t.
V

This is precisely the Born rule. The magnitude of |Ψ(x, t)|2 is the probability density at x ∈ Rn .
The total probability over the whole space Rn is 1, as promised by the normalization condition.
Thus this probability interpretation is compatible with the quantum dynamics.

25
1.5.3 Position and Momentum

The quantum operator x̂i associated to the i-th position is defined to be multiplying by xi


x̂i f (x) := xi f (x).

In order to understand the quantum operator p̂j associated to the corresponding conjugate
momentum, we recall that conjugate variables are related by Fourier transform. If we go to the
momentum space
f (x) ←→ fˆ(p)

then the effect of p̂j should correspond to multiplying by pj in the Fourier dual

(p̂j f ) (x) ←→ pj fˆ(p)

Implementing this relation into the Fourier transform

ˆ
1
f (x) = dn p eip·x/h̄ fˆ(p)
(2πh̄)n/2 Rn

we find
∂
(p̂j f ) (x) = −ih̄ f (x).
∂xj
In other words, p̂j is represented by the differential operator

∂
p̂j = −ih̄ .
∂xj
It is now clear that the position operators and momentum operators satisfy the canonical
commutation relation
h i
x̂k , p̂j = ih̄δjk .

Here δjk is the Kronecker delta symbol. In particular, Heisenberg uncertainty relation holds

h̄
∆x̂k ∆p̂k ≥ .
2

f (x) fˆ(p)

∆x
∆p

1.5.4 Hamiltonian Operator

We consider a particle of mass m moving in the potential V (x, t). Classically, the dynamics
is described by the Hamiltonian function H (x, p) in the phase space

p2
H = + V (x, t).
2m
p2
Here 2m is the kinetic energy, and V is the potential energy which is real valued.

26
 In quantum mechanics, the Hamiltonian function will be quantized to a self-adjoint Hamil-
tonian operator. The natural candidate is to replace

xi → x̂i , pi → p̂i .

In this way we find the Hamiltonian operator

2
^ = − h̄ ∇2 + V (x, t),
H
2m
P
where ∇2 = ∂ ∂
i ∂xi ∂xi is the Laplacian operator.
We will mainly study the Hamiltonian operator of the above form in this chapter. The
corresponding Schrödinger equation now takes the explicit form
 
∂ h̄2 2
ih̄ Ψ(x, t) = − ∇ + V (x, t) Ψ(x, t).
∂t 2m
^ is a differential operator also leads to a local
The fact that the Hamiltonian operator H
form of the conservation of probability. Let

ρ(x, t) = |Ψ(x, t)|2

denote the probability density. As we have seen, the Schrödinger equation implies the conser-
vation of total probability ˆ
d
dn x ρ(x, t) = 0.
dt
This conservation law can be promoted to a standard local form
∂
ρ + ∇ · ~j = 0
∂t
for a current vector field ~j. Indeed, using the Schrödinger equation and reality of V ,
∂ ∂Ψ ∂Ψ
ρ= Ψ+Ψ
∂t ∂t ∂t
i ^  i ^ 
= HΨ Ψ − Ψ H Ψ
h̄ h̄
 
i h̄2
h̄2
= − ∇ Ψ Ψ + V ΨΨ +
2
Ψ∇ Ψ − V ΨΨ
2
h̄ 2m 2m
ih̄ 

=− ∇ · ∇Ψ Ψ − Ψ∇Ψ
2m



=⇒ ~j = ih̄ ∇Ψ Ψ − Ψ∇Ψ = h̄ Im Ψ∇Ψ .
2m m
Here ∇ = (∂x1 , · · · , ∂xn ) is the gradient operator. Thus


~j = h̄ Im Ψ∇Ψ
m
which is called the probability current. For any fixed region V ⊂ Rn , we have
ˆ ˆ ˆ ˆ
d n
d x ρ(x, t) = d x ∂t ρ = −
n
d x∇ · j = −
n ~ d~σ · ~j.
dt V V V ∂V

Here d~σ is the vector surface element on ∂V . This says that the probability of finding the
particle inside V changes by the flow of the probability current out of the boundary ∂V .

27
1.5.5 Stationary States

We consider the quantum particle moving in a potential V that only depends on the
^ is
position x ∈ Rn but not on the time: V = V (x). In this case, the Hamiltonian operator H
time-independent
2
^ = − h̄ ∇2 + V (x).
H
2m
We look for special solutions of the Schrödinger equation of the form

Ψ(x, t) = e−iωt ψ(x).

Inserting this expression into the Schrödinger equation, we find

^ ψ(x) = Eψ(x)
H (*)

^ Based on our
where E = h̄ω. In other words, ψ(x) is an eigenstate of the Hamiltonian H.
discussion on measurement, the eigenvalue E is naturally understood as the energy of the state
^ is self-adjoint, E
ψ. Equation (*) is called the time-independent Schrödinger equation. Since H
must be real.
Solutions of the form

Ψ(x, t) = e−iEt/h̄ ψ(x) with ^ ψ = Eψ

H

are also called stationary states. The probability density

|Ψ(x, t)|2 = |ψ(x)|2

^ at any time
does not depend on time t. These states are eigenstates of the Hamiltonian H

^ Ψ(x, t) = EΨ(x, t).

H

^ in the state Ψ is
In particular, the expectation of H
D E
^
H =E
Ψ

^ in the state Ψ vanishes

and the uncertainty of H

^ = 0.
∆Ψ H

Once we have found all the stationary states ψα with energy Eα , then a general solution
of the Schrödinger equation can be constructed as a superposition
X
cα e−iEα t/h̄ ψα (x).
α

Thus it is fundamental to understand stationary states first, i.e., to study the time-independent
Schrödinger equation. In general, H^ could have discrete spectrum and continuous spectrum.
P
For the continuous spectrum, the above sum α has to be replaced by an appropriate integral.
Both the discrete spectrum and the continuous spectrum have specific physical meanings.
We will illustrate by a few examples of different spectrum types in subsequence sections.

28
 1.6 Free Particle: Example of Continuous Spectrum

1.6.1 Wave Packet

Let us consider a free quantum particle where the potential V = 0. The Hamiltonian is
2
^ = − h̄ ∇2 .
H
2m
The free Schrödinger equation
∂ h̄2 2
ih̄ ψ(x, t) = − ∇ ψ(x, t)
∂t 2m
admits plane wave solutions by
ψk (x, t) = ei(k·x−ω(k)t) .

Here the wave vector k and the angular frequency ω are determined by the momentum p and
the energy E by
p = h̄k, E = h̄ω.

These are called the de Broglie relations, which work for general matter waves.
Using H^ = − h̄2 ∇2 , we find E = p2 , or
2m 2m

h̄k2
ω(k) = .
2m
This gives the dispersion relation in this case.
However, the plane wave solution Ψk is non-normalizable

|Ψk |2 = 1 at any x and t.

The integral of |Ψk |2 over space will be infinity. So Ψk does not give a physical state.
Nevertheless, a general solution can be obtained as a wave packet in terms of superposition
of plane waves ˆ
1
Ψ(x, t) = dn p ψ̂0 (p)ei(p·x−E(p)t)/h̄ .
(2πh̄)n/2
Clearly, let ψ0 (x) denote the wave function at t = 0

ψ0 (x) = Ψ(x, 0).

Then ψ̂0 (p) is the Fourier transform of ψ0 (x) to the momentum space. This gives the explicit
solution of the wave function from specified initial condition at t = 0.

Example 1.6.1 (Gaussian Packet). Consider the one-dimensional case n = 1. Let us give the
initial wave function at t = 0 by
e−x
2 /4δ 2 h̄
ip0 x/h̄
ψ0 (x) = e , δ > 0.
(2πδ 2 h̄)1/4
The coefficient is chosen such that the normalization condition holds
ˆ ˆ
1
dx e−x /2δ h̄ = 1.
2 2
dx |ψ0 (x)| = √
2
R 2πh̄δ R
2 2
e−x /2δ h̄
The magnitude of |ψ0 (x)|2 = √
2πh̄δ
is Gaussian.

29
 √
δ h̄

The parameter δ is related to the uncertainty of x̂

ˆ 1/2
2
√
∆ψ0 x̂ = dx x |ψ0 (x)|
2
= δ h̄.
R

The Fourier transform of ψ0 (x) is

ˆ
ψ̂0 (p) = dx e−ipx/h̄ ψ0 (x)
R
ˆ
e−x
2 /4δ 2 h̄
−i(p−p0 )x/h̄
= dx e
R (2πh̄δ 2 )1/4
e−(p−p0 )
2 δ 2 /h̄

= .
((2πh̄)−1 (2δ)−2 )1/4

The uncertainty of the momentum operator p̂ can be computed via ψ̂0

√
h̄
∆ψ0 p̂ = .
2δ
In this case, we find the Heisenberg uncertainty
h̄
∆ψ0 x̂ · ∆ψ0 p̂ = .
2
The wave function at any time t is therefore solved by
ˆ
1 2
Ψ(x, t) = dp ψ̂0 (p)ei(px−p t/2m)/h̄
2πh̄ R
ˆ
e−(p−p0 ) δ /h̄
2 2
1 2
= dp ei(px−p t/2m)/h̄
2πh̄ R ((2πh̄)−1 (2δ)−2 ) 1/4
!
p 0 t
2
1 ip0
( p0 t
x− 2m ) exp − x− m

= √
1/2 e
h̄
it
.
it 4 δ 2 + 2m h̄
2πh̄ δ + 2mδ

1.6.2 Group Velocity

We would like to understand how a localized wave packet move in the space. Let us start
with a general one-dimensional wave packet of the form
ˆ
1
Ψ(x, t) = √ dk ϕ̂0 (k)ei(kx−ω(k)t) .
2π
Here we have used the wave vector k instead of the momentum p = h̄k for the Fourier transform.
ω(k) is a function of k, describing the dispersion relation.
Assume ϕ̂0 (k) is concentrated near k = k0 , so only when k ∼ k0 is important.

30
 ϕ̂0

k0

We ask for which value of x and t such that the magnitude of the wave packet Ψ(x, t) takes
the largest value. For k ∼ k0 , we can approximate

ω(k) ' ω(k0 ) + ω ′ (k0 )(k − k0 ).

Substituting this into the wave packet, we have

ˆ
1
Ψ(x, t) = √ dk ϕ̂0 (k)ei(kx−ω(k)t)
2π ˆ
1 −iω(k0 )t+ik0 ω′ (k0 )t ′
'√ e dk ϕ̂0 (k)eik(x−ω (k0 )t)
2π
′
= e−iω(k0 )t+ik0 ω (k0 )t Ψ(x − ω ′ (k0 )t, 0).

Assume the peak of the magnitude of the initial wave function Ψ(x, 0) is at x = x0 .
The above calculation shows that the peak of the magnitude of Ψ(x, t) at time t appears
approximately at
x = ω ′ (k0 )t + x0 .

In other words, the peak of the wave packet moves at the velocity ω ′ (k0 ). This is called the
group velocity
dω
group velocity = .
dk k=k0
It describes the approximate speed at which the wave packet propagates.
In contrast, there is another notion called the phase velocity defined by
ω
phase velocity = .
k k=k0

It describes the speed at which the pure plane wave ei(k0 x−ω(k0 )t) propagates.
For linear dispersion relation

ω(k) = αk where α = const,

the group velocity and the phase velocity coincide, both equal to α. Electromagnetic waves in
the vacuum are such examples.
In general, group velocity is different from the phase velocity. Let us consider the example
of one-dimensional free quantum particles. The de Brogile relations

p = h̄k, E = h̄ω

and E = p2 /2m lead to the dispersion relation

h̄k 2
ω(k) = .
2m

31
 For a wave packet whose momentum is concentrated at p0 = h̄k0 , the group velocity is

dω h̄k0 p0
= = .
dk k=k0 m m

This is the expected velocity for a free particle with momentum p0 and mass m. As a comparison,
the phase velocity is
ω h̄k0 p0
= = .
k k=k0 2m 2m
In the Gaussian packet example 1.6.1,
!
p 0 t
2
1 ip0
( p0 t
x− 2m ) exp − x− m

Ψ(x, t) = √
1/2 e
h̄
it
.
it 4 δ 2 + 2m h̄
2πh̄ δ + 2mδ

The momentum mode ψ̂0 (p) is concentrated around p0 for small h̄. We see clearly that the peak
of Ψ(x, t) travels at the group velocity p0 /m.

1.7 Harmonic Oscillator: Example of Discrete Spectrum

In this section we study the exactly solvable example of harmonic oscillator, which is one
of the most important model in quantum physics.
For simplicity, we focus on the one-dimensional case. The classical Hamiltonian is

p2 1
H = + kx2
2m 2
where m is the particle’s mass, k > 0 is a constant. The potential V = 21 kx2 is quadratic in x
and time-independent.
Classically, the particle’s motion obeys the Hamilton’s equations


ẋ = ∂H = p

∂p m


ṗ = − ∂H = −kx
∂x
The equation of motion in x is
mẍ = −kx.

The force that is applied to the particle is governed by Hooke’s law F = −kx.
F = −kx

32
 The equation of motion is solved by

x(t) = A cos ωt + B sin ωt,

where A, B are constants, and r

k
ω=
m
is the angular frequency of oscillation.
Quantum mechanically, the Hamiltonian operator is
2 2
^ = − h̄ d + 1 kx2 .
H
2m dx2 2
It suffices to solve the time-independent Schrödinger equation for a stationary ψ(x)

^ ψ(x) = Eψ(x).
H

The corresponding wave function will then be given by ψ(x, t) = ψ(x)e−iEt/h̄ .

1.7.1 Ladder Operators

Remarkably, the time-independent Schrödinger equation for harmonic oscillator can be

exactly solved by a simple algebraic method. Let us write
2 2 2
^ = p̂ + k x̂2 = p̂ + mω x̂2 ,
H
2m 2 2m 2
where x̂ is the position operator, and p̂ = −ih̄ dx
d
is the momentum operator. x̂ and p̂ satisfy
the canonical commutation relation
[x̂, p̂] = ih̄.

We can rewrite the Hamiltonian operator as

^ = 1 p̂2 + m2 ω 2 x̂2
H
2m
1
= [(mωx̂ − ip̂) (mωx̂ + ip̂) − imω (x̂p̂ − p̂x̂)]
2m
1 1
= (mωx̂ − ip̂) (mωx̂ + ip̂) + h̄ω
2m 2 
1 1
= h̄ω (mωx̂ − ip̂) (mωx̂ + ip̂) +
2mh̄ω 2
 
1
= h̄ω a† a +
2

where we have introduced two operators

1 1
a† = √ (mωx̂ − ip̂) , a= √ (mωx̂ + ip̂) .
2mh̄ω 2mh̄ω
They are called ladder operators. The reason for the names will be clear soon. As the symbol
suggests, the two operators a† and a are adjoint of each other.

33
 Firstly we observe that the ladder operator satisfy the commutation relation
h i
a, a† = 1.

This leads to the following commutation relations

^ a† ] = h̄ωa† ,
[H, ^ a] = −h̄ωa.
[H,

The key is the following statement.

Proposition 1.7.1. If ψ solves the time-independent Schrödinger equation with energy E, then
a† ψ (or aψ) solves the time-independent Schrödinger equation with energy E + h̄ω (or E − h̄ω).

^ ψ = Eψ. Then
Proof: Assume ψ satisfies H
 
^ a† ψ = [H,
H ^ a† ]ψ + a† H
^ψ

= h̄ωa† ψ + Ea† ψ
= (E + h̄ω) a† ψ.

The calculation for aψ is similar.

For this reason, the ladder operators a† , a allow us to climb up and down in energy. We
also call a† the raising operator and a the lowering operator.

1.7.2 Ground State

The second crucial statement is the following.

Proposition 1.7.2. The energy E of a stationary state is nonnegative.

Proof: For the normalized stationary state ψ with energy E,

ˆ  
^ h̄2 d2 1 2
E = hψ| H |ψi = dx ψ(x) − + kx ψ(x)
R 2m dx2 2
ˆ 2
h̄2 d 1
= dx ψ(x) + kx2 |ψ(x)|2 ≥ 0.
R 2m dx 2

Here we have cheated a bit by assuming without proof that ψ is differentiable with appropriate
decay condition at space infinity. We leave it to more careful readers.

Now given a stationary state ψ with energy E, we can use the lowering operator a to lower
the energy. Since the energy of a nonzero state is nonnegative, am ψ = 0 for m sufficiently large.
Let E0 ≥ 0 be the smallest possible energy. States with the lowest energy are called ground
states. Let ψ0 be one ground state with energy E0 . We will soon see that ground state is unique
(up to normalization) in this case. For ψ0 being a ground state, we must have

aψ0 = 0

34
which is the same as the differential equation
 
d
h̄ + mωx ψ0 (x) = 0.
dx

Up to a normalization, there is a unique solution given by

 mω 1/4
e− 2h̄ x .
mω 2
ψ0 (x) =
πh̄
The coefficient is chosen such that
ˆ
dx |ψ0 (x)|2 = 1.
R

The corresponding energy is

 
^ † 1 1
H ψ0 = h̄ω a a + ψ0 = h̄ωψ0
2 2

1
=⇒ E0 = h̄ω.
2
Note that classical mechanically, the smallest possible energy for harmonic oscillator is
zero (for example, the energy for a static particle sitting at x = 0). Quantum mechanically, the
smallest energy is E0 = 12 h̄ω! This lift of ground state energy is purely a quantum effect.

1.7.3 Excited States

Starting from the ground state, we obtain higher energy states simply by applying the
raising operator
 
† n 1
ψn (x) = An (a ) ψ0 (x) with En = n+ h̄ω,
2

where An is the normalization constant. To calculate An , we use the fact that a† and a are
adjoint of each other. Therefore
D E
hψn |ψn i = |An |2 (a† )n ψ0 (a† )n ψ0 = |An |2 hψ0 |an (a† )n |ψ0 i.
 
Using aψ0 = 0 and a, a† = 1, we find

an (a† )n |ψ0 i = nan−1 (a† )n−1 |ψ0 i = · · · = n!|ψ0 i.

Thus
hψn |ψn i = |An |2 n!hψ0 |ψ0 i.

The normalization condition hψn |ψn i = 1 gives An = √1 .

n!
So

1
ψn = √ (a† )n ψ0 .
n!
Finally we show that ψn ’s are all the stationary states.

35
Proposition 1.7.3. Let ψ be a stationary state with energy E. Then ψ must be of the form


ψn (up to normalization) for some n and E = n + 21 h̄ω.

Proof: Since the energies are bounded from below, there exists n ≥ 0 such that

an ψ 6= 0, an+1 ψ = 0.

Then a(an ψ) = 0. By the uniqueness of the ground state, we must have

an ψ = αψ0 for some α 6= 0.

Comparing both sides, this readily shows

 
1
E − nh̄ω = E0 =⇒ E = En = n+ h̄ω.
2

Assume ψ and ψn are linearly independent. Applying the Gram-Schmidt orthogonalization,

we can assume that
hψn |ψi = 0.

On the other hand, using the fact that a, a† are adjoint of each other,
D E
hψn |ψi ∝ (a† )n ψ0 ψ = hψ0 |an ψi = αhψ0 |ψ0 i = α 6= 0.

This is a contradiction. So up to normalization, ψ is the same as ψn .

Thus we have found all stationary states. The states ψn for n > 0 are called excited states,
which are created from the ground states ψ0 by applying a† . Note that unlike the classical
picture, the energies in the quantum case are discrete. They are quantized! Moreover, these
states {ψn }n≥0 form an orthonormal basis of the Hilbert space L2 (R).
The ground state is explicitly given by the Gaussian function
 mω 1/4
e− 2h̄ x .
mω 2
ψ0 (x) =
πh̄
The excited states ψn can be explicitly expressed in terms of Hermite polynomials.
For convenience, let us redefine the variable by
r
mω
y= x.
h̄
Then the ground state is
 mω 1/4
ψ0 (x) = c0 e−y
2 /2
, c0 := .
πh̄
The ladder operators are
    
 † 1 d 1 d

a = √ mωx − h̄ =√ y−
2mh̄ω dx 2 dy
   

 1 d 1 d
a = √ mωx + h̄ =√ y+
2mh̄ω dx 2 dy

36
 We can rewrite the raising operator as the composition of three operators
 
† 1 ŷ2 /2 d
· e−ŷ /2
2
a =√ e · −
2 dy
2 /2 2 /2
where eŷ means the operator by multiplying the function ey . Then
1
ψn (x) = √ (a† )n ψ0
n!
 
c0 1 ŷ 2 /2 d n −ŷ2 /2 −y2 /2
=√ √
n e · − ·e e
n! 2 dy
   
c0 −y 2 /2 y2 d n −y2
=√ e e − e
n!2n/2 dy
 mω 1/4 1
Hn (y)e−y /2 .
2
= √
πh̄ n
2 n!
Here
2 dn  −y2 
Hn (y) := (−1)n ey e
dy n
are the Hermite polynomials. The first few looks like

H0 (y) = 1
H1 (y) = 2y
H2 (y) = 4y 2 − 2
H3 (y) = 8y 3 − 12y
H4 (y) = 16y 4 − 48y 2 + 12
..
.

1.8 Square Well: Example of Mixed Spectrum

We have seen an example where the Schrödinger operator H^ has a pure continuous spectrum
^ has a pure discrete spectrum (harmonic oscillator). We
(free particle) and an example where H
now discuss an example where both the continuous and the discrete spectrum are present. It is
about the finite square well potential of depth V0 > 0 and width 2a in dimension one

−V0 −a ≤ x ≤ a
V (x) =
0 |x| > a

−a a
x
V (x)
−V0

37
1.8.1 Matching Condition

We consider the time-independent Schrödinger equation

 
h̄2 d2
− + V (x) ψ(x) = Eψ(x).
2m dx2
⃝
1 In the region |x| > a, the equation becomes

h̄2 d2
− ψ(x) = Eψ,
2m dx2
which has two linearly independent solutions given by
√
e± −2mEx/h̄
.

The solution ψ can be rewritten as a linear combination

√ √
−2mEx/h̄
ψ = c1 e + c 2 e− −2mEx/h̄
.

The coefficients c1 , c2 are to be determined. The solution ψ is smooth in the region |x| > a.
⃝
2 In the region |x| < a, the equation becomes

h̄2 d2
− ψ = (E + V0 ) ψ.
2m dx2
Again, ψ can be expressed as a linear combination of the two independent solutions
√
e± −2m(E+V0 )x/h̄ .

ψ is also smooth in the region |x| < a.

By ⃝
1 ⃝,
2 we see that the only possible place where the smoothness of ψ fails is when
x = ±a. The natural boundary condition to be imposed at x = ±a is

ψ and ψ ′ are continuous at x = ±a. (*)

Otherwise, ψ ′′ will have a δ-function contribution at x = ±a, breaking the Schrödinger equation.
We will call this boundary condition (*) the “matching condition”.

1.8.2 Bound States

We first consider the case when the solution ψ is normalizable. Such energy eigenstate is
called a bound state.
In the region |x| > a, ψ(x) is a linear combination of
√
e± −2mEx/h̄
.

For such ψ to be normalizable, it is necessary to have E < 0 and such that

 √
αe −2mEx/h̄ x < −a
ψ(x) = √
βe− −2mEx/h̄ x>a

38
for some constants α, β.
On the other hand, the equation
 
h̄2 d2
− + V (x) ψ = Eψ
2m dx2

implies  
h̄2 d2
ψ− ψ = hψ|(E − V (x))ψi,
2m dx2
i.e., ˆ ˆ
h̄2
dx ψ ′ (x) .
2 2
dx (E − V (x))|ψ(x)| =
2m
To obtain a nontrivial solution ψ, it is necessary to have E > infx V (x) = −V0 . So the
bound state appears only for energy satisfying

−V0 < E < 0.

−a a
x
V (x) E
−V0

Assume this holds. Then for |x| < a inside the well, ψ is a linear combination of
p ! p !
2m(E + V0 )x 2m(E + V0 )x
cos and sin .
h̄ h̄

To simplify notations, let

√ p
−2mE 2m(E + V0 )
λ= , µ= .
h̄ h̄
A further simplification comes from the observation that the potential is an even function

V (x) = V (−x).

If ψ(x) is a solution, then ψ(−x) is also a solution. Therefore any solution can be written
as a sum of an even solution and an odd solution
1 1
ψ(x) = (ψ(x) + ψ(−x)) + (ψ(x) − ψ(−x)) .
2 2
Without lost of generality, we can assume ψ is either even or odd.
Let us first consider ψ being an even function. Then


 αe−λx x>a


ψ(x) = β cos µx |x| < a α, β are constants.



 λx
αe x < −a

39
 Now we apply the matching condition. We only need to consider x = a since ψ(x) is even.


 lim ψ(x) = lim ψ(x)
x→a− x→a+

 lim ψ ′ (x) = lim ψ ′ (x)
x→a− x→a+

leads to 
β cos µa = αe−λa
−βµ sin µa = −αλe−λa

Dividing these two equations, we get

λ = µ tan µa (**)
√ p
Recall λ = −2mE/h̄, µ = 2m(E + V0 )/h̄. The above relation gives the allowed ener-
gies. Given E satisfying relation (**), we can solve for ψ(x) which is unique up to normalization.
Thus it gives a unique physical state with the corresponding energy E.
To understand solutions to (**), let us redefine
 p

 2m(E + V0 )a
u = µa =
h̄
√

 2mV a
 u0 = 0
h̄
We can express λ, µ in terms of u, u0

 u

µ = a
r  p

 u0 2 u20 − u2
λ = − µ2 =
a a
Then equation (**) becomes
r 
u 2 0
− 1 = tan u.
u
q
Solutions are given by the intersections of the curve y = (u0 /u)2 − 1 with the curve
y = tan u for 0 < u < u0 .

40
 y
y = tan u

q
y = (u0 /u)2 − 1

−π/2 0 π/2 π 3π/2 2π u0 u

In particular, the figure shows that there are a finite number of intersections, i.e., a finite
number of allowed energies. The number depends on the value of u0 . For larger u0 , which
means wider and deeper well, we have more bound states. Nevertheless, it is clear that we have
at least one solution, no matter how small u0 is.
We next briefly discuss the case for ψ being an odd function. Then


 αe−λx x>a


ψ(x) = β sin µx |x| < a




−αeλx x < −a

The matching condition at x = a gives


β sin µa = αe−λa
βµ cos µa = −αλe−λa

from which we find required relation for E

µ cot µa = −λ.

Using the same variables u and u0 as above, this is

r 
u0 2
− 1 = − cot u.
u
We plot the corresponding curves for 0 < u < u0

41
 y

y = − cot u

q
y= (u0 /u)2 − 1

−π/2 0 π/2 π 3π/2 2π u0 u

Again, we find only finite number of intersections. The number of odd bound state is bigger
for larger u0 , i.e., for wider and deeper well. However, for u0 sufficiently small, say u0 < π/2,
there will be no odd bound state.
In summary, we have found finite number of energies −V0 < E0 < E1 < · · · < EN < 0
where each Ei has exactly one bound state ψi (x). The wave function ψi (x) is even/odd if i is
even/odd. The ground state ψ0 is even, and it always exists.

−a a

V (x) x
E2
E1
E0
−V0

ψ0 (x)

−a a

42
 ψ1 (x)

−a a

ψ2 (x)

−a a

Let us briefly discuss the limit case V0 → +∞. This corresponds to u0 → +∞.

y y

y = tan u y = − cot u

u u

even case odd case

In this deep well limit, the intersections approximately happen when u is an integer multiple
of π/2. It follows that
p a (n + 1)π
2m(En + V0 ) '
h̄ 2
(n + 1)2 π 2 h̄2
=⇒ En + V 0 ' , n = 0, 1, 2, · · ·
2m(2a)2
Surprisingly, we will find this formula reappearing for the resonant transmission in the scattering
problem as we discuss next.

1.8.3 Scattering States

We have seen that the square well admits a finite number of bounded states. The bounded
^ Unlike the harmonic
state energies correspond to the discrete spectrum of the Hamiltonian H.
oscillator where the bound states form a basis of the Hilbert space, the space of bound states
of square well is finite dimensional and so can not span the whole Hilbert space. There will

43
 ^ like in the free particle case. In fact, when x is far
also exist the continuous spectrum of H
away, the potential is zero and the particle behaves like a free particle there. This suggests non-
normalizable solutions of the time-independent Schrödinger equation which behave like plane
waves of free particle in the far away region. As we will see shortly, this is indeed the case. These
solutions are called scattering states. The reason for the name will be explained in Section 1.9.
The scattering states appear for E > 0.
E>0

−a a
x
V (x)
−V0

The time-independent Schrödinger equation

 
h̄2 d2
− + V (x) ψ = Eψ
2m dx2
can be solved in each region in the same way as we did for bound state.


 Aeikx + Be−ikx x < −a


ψ(x) = C sin µx + D cos µx −a < x < a



 ikx
F e + Ge−ikx x>a

Here A, B, C, D, F, G are constants, and

√ p
2mE 2m(E + V0 )
k= > 0, µ= > 0.
h̄ h̄
At x = ±a, we again impose the matching condition.
Since E > 0, the wave function ψ(x) is oscillating instead of decaying when x → ∞, hence
is non-normalizable. However, such solutions will be the building block for scattering process
as we will discuss in Section 1.9. Let us first illustrate the basic idea. Let

ψ(x, t) = ψ(x)e−iEt/h̄

be the corresponding solution of the time-dependent Schrödinger equation. Let us consider the
region x < −a where    
2 2
i kx− h̄k t −i kx+ h̄k t
ψ(x, t) = Ae 2m
+ Be 2m
.
 2

i kx− h̄k t h̄k
The first term Ae 2m
is a plane wave moving to the right at phase velocity 2m . The
 2

−i kx+ h̄k t
second term Be 2m
is a plane wave moving to the left at the same phase velocity.
A

44
 With the above interpretation, let us consider the stationary solution ψ(x) of the form


 ikx −ikx x < −a
Ae + Be

ψ(x) = C sin µx + D cos µx −a < x < a




F eikx x>a

This wave function represents the following process: A wave of amplitude A is incident from
the left at x = −∞, and meets the square well; then a wave of amplitude B is reflected back to
the left, while a wave of amplitude F is transmitted through the square well and moves to the
right at x = +∞.
A
F
B

A: incident wave amplitude

B: reflected wave amplitude
F : transmitted wave amplitude

We define
|B|2
reflection coefficient: R=
|A|2
|F |2
transmission coefficient: T =
|A|2
R represents the probability of reflection, and T represents the probability of transmission.
From this physical interpretation, we should expect

T + R = 1.

One way to see this is to use the local form of probability conservation. Recall the proba-
bility current in Section 1.5.4  
h̄ d
j(x) = Im ψ ψ .
m dx
Substituting into the above ψ, we find

 h̄k  2 
 |A| − |B|2 x < −a
j(x) = m

 h̄k |F |2 x>a
m
For stationary solutions, the probability density ρ is time-independent. The conservation
equation
∂ρ ∂
+ j(x) = 0
∂t ∂x

45
implies that j(x) must be x-independent: accumulation of probability can not happen at any
region of space. It follows that

|A|2 = |B|2 + |F |2 =⇒ T + R = 1.

Now we move on to solve ψ(x) via the matching condition


Ae−ika + Beika = −C sin µa + D cos µa
x = −a :
ik Ae−ika − Beika
= µ (C cos µa + D sin µa)

C sin µa + D cos µa = F eika
x= a:
µ (C cos µa − D sin µa) = ikF eika

These four equations uniquely determine the five constants A, B, C, D, F up to a total

normalization. After a laborious calculation, we find


 F e−2ika
 =
 A cos 2µa − i sin2kµ
2µa
(k 2 + µ2 )

 B sin 2µa 2


 =i µ − k2
F 2kµ
We can compute the reflection coefficient R, the transmission coefficient T and check T +
R = 1 as promised. Explicitly, the transmission probability is
|F |2 1
T = =  p .
|A|2 1+
V02
sin 2 2a
2m(E + V 0 )
4E(E+V0 ) h̄

The plot of T as a function of E looks like

T

Note that there are certain values of E making T = 1. In this case we have the full
transmission as no waves are reflected: the well becomes transparent! From the above formula
of T , the full transmission happens when
(n + 1)2 π 2 h̄2
En + V 0 = , n∈Z such that En > 0.
2m(2a)2
Surprisingly, En corresponds to the bound state energies of the infinite square well that we
find previously. For the energy En , the wavelength of ψ inside the well is
2π 2πh̄ 4a
=p = .
µ 2m(En + V0 ) n+1

46
So the well width 2a fits an integer number of half wavelength. This phenomenon is called
resonant transmission.
We can also consider wave incidents from the right, and look for solutions of the form


 Be−ikx x < −a


ψ(x) = C sin µx + D cos µx −a < x < a




F eikx + Ge−ikx x>a

F
B
G

The interpretation is similar. The reflection and transmission coefficients are

|F |2 |B|2
R= , T = .
|G|2 |G|2

In general, we could have waves incident form both sides



 Aeikx + Be−ikx x < −a


ψ(x) = C sin µx + D cos µx −a < x < a



 ikx
F e + Ge−ikx x>a

A F

B G

Then A, G represent incoming waves and B, F represent outgoing waves.

1.9 Scattering
In this section we discuss the basic idea of scattering process in the case of one-dimensional
particles. We explain how this is related to the continuous spectrum of the Schrödinger operator.

1.9.1 Wave Packet Scattering

In the study of scattering problem, we consider particles that come from far away and
scatter against some potential produced by localized interaction.

47
 incoming outgoing

We consider one-dimensional particles scattering in a compactly supported potential V (x)

V (x) = 0 for |x| ≥ R.

V (x)

Assume a particle comes from x = −∞. Quantum mechanically, such a particle is repre-
sented by a wave packet

incoming
k0

before the interaction

When the particle enters the region of V , it interacts with the potential. Afterwards, it will
be transmitted through the potential toward x = +∞, or be reflected back toward x = −∞,
with certain probability.

reflected transmitted

after the interaction

Explicitly, let us represent an incoming free particle by a localized wave packet

ˆ
1
ψin (x, t) = √ dk α(k)ei(kx−ω(k)t) , ω(k) = h̄k 2 /2m.
2π R

48
Here α(k) is nonzero only in a small neighborhood of k0 > 0, so the wave packet will travel
forward with group velocity h̄k0 /m. We assume
ˆ
dk |α(k)|2 = 1,
R

so that the wave function is normalized

ˆ
dx |ψin (x, t)|2 = 1.
R

To describe the scattering process, we observe that the Schrödinger equation

 
∂ h̄2 d2
ih̄ ψ = − + V (x) ψ
∂t 2m dx2
is linear in ψ. This linearity leads to the following strategy. We first look for solutions which are
plane waves far away. Then we take the superposition of these asymptotic plane wave solutions
with respect to the coefficient α(k) to obtain the physical solution for the scattering.
Precisely, let us consider the time-independent Schrödinger equation
 
h̄2 d2 (h̄k)2
− + V (x) ψ = Eψ, where E = > 0.
2m dx2 2m

In the region |x| >> 0, the potential V (x) = 0 and so the wave function is given by linear
combinations of e±ikx . We look for the solution ψk of the form

eikx + B(k)e−ikx x → −∞
ψk (x) =
C(k)eikx x → +∞

eikx
Ceikx

Be−ikx

Here B(k), C(k) are constants that depend on k, which are determined by solving the time-
independent Schrödinger equation. Such solution ψk is non-normalizable, but lies in the con-
(h̄k)2
tinuous spectrum with energy Ek = 2m .
Assume we have found ψk ’s. Then we obtain a solution of the Schrödinger equation by
ˆ
1 h̄k2
ψ(x, t) = √ dk α(k)ψk (x)e−i 2m t .
2π R
In the region x → −∞, we have

ψ(x, t) = ψin (x, t) + ψR (x, t),

where ψin (x, t) is our prepared incoming wave packet above, and
ˆ  2

1 −i kx+ h̄k t
ψR (x, t) = √ dk α(k)B(k)e 2m
2π

49
represents the reflected wave packet.
In the region x → +∞, we have
ˆ  2

1 i kx− h̄k t
ψ(x, t) = √ dk α(k)C(k)e 2m
2π
which represents the transmitted wave packet. Thus this wave function ψ(x, t) contains the
quantum information about the scattering of incoming particle ψin with the potential V (x).

1.9.2 S-matrix

The S-matrix, or the scattering matrix, is about the relation for particle states before and
after a scattering process. We illustrate the S-matrix in the one-dimensional scattering process.
We consider a localized one-dimensional potential V (x) which is compactly supported. As
we have discussed above, the scattering process is completely determined by solutions of the
Schrödinger equation which are plane waves at |x| → ±∞ outside the potential barrier. We
look for solutions of the time-independent Schrödinger equation
 
h̄2 d2 (h̄k)2
− + V (x) ψ(x) = Eψ(x), E = ,
2m dx2 2m

which have the asymptotic plane wave behavior by


Aeikx + Be−ikx x → −∞
ψ(x) =
Ceikx + De−ikx x → +∞

A C

B D

• The case D = 0 represents a scattering process for an incident wave of amplitude A coming
from the left. Then C is the amplitude for the transmitted wave and B is the amplitude
for the reflected wave.

• The case A = 0 represents a scattering process for an incident wave of amplitude D coming
from the right. Then B is the amplitude for the transmitted wave and C is the amplitude
for the reflected wave.

In general, we could have both left and right incident waves. Let us represent the amplitudes
of the incoming waves by a column vector
!
A
Ψin =
D

50
and represent the amplitudes of the outgoing waves by a column vector
!
B
Ψout = .
C

Since the Schrödinger equation is linear, Ψout is related to Ψin by a linear relation

Ψout = SΨin

or explicitly ! ! !
B S11 S12 A
= .
C S21 S22 D
The matrix for the transition !
S11 S12
S=
S21 S22
is called the S-matrix. The matrix entries Sij are functions of the wave vector k, and these
functions are completely determined by the localized potential V (x).
To see the meaning of the entries of S, consider setting D = 0 and we have
! !
A S11 A
S = .
0 S21 A

This says that S11 is the reflection amplitude and S21 is the transmission amplitude for incident
wave from the left. Similarly, setting A = 0
! !
0 S12 D
S = .
D S22 D

This says that S12 is the transmission amplitude and S22 is the reflection amplitude for incident
wave from the right.
If we take the absolute value square of the transmission and reflection amplitudes, we find
the corresponding transmission and reflection coefficients.

1.9.3 Unitarity

The S-matrix is in fact unitary. To see this, consider the probability current of ψ
 
h̄ d
j(x) = Im ψ ψ .
m dx

For the stationary state, the probability density ρ = |ψ|2 is time-independent. Then the local
conservation of probability (which follows from Schrödinger equation for ψ)

∂ρ ∂
+ j=0
∂t ∂x
d
implies dx j(x) = 0. So j(x) must be constant.

51
 From the asymptotic behavior of ψ(x), we have

 h̄k  2 
 |A| − |B|2 x → −∞
j(x) = m  

 h̄k |C|2 − |D|2 x → +∞
m
Thus
|A|2 − |B|2 = |C|2 − |D|2 =⇒ |A|2 + |D|2 = |B|2 + |C|2 .

In other words,
t t
Ψout Ψout = Ψin Ψin .

So the linear transformation S preserves the Hermitian inner product, i.e., S is unitary

S ∗ S = 1.
t
Here S ∗ := S .

1.9.4 Time Reversal Symmetry

Assume the potential V = V (x) is time-independent. Then the Schrödinger equation

∂ ^ψ
ih̄ ψ=H
∂t
has a time reversal symmetry: if ψ(x, t) is a solution, then ψ(x, −t) is also a solution with the
time direction reversed. For time-independent Schrödinger equation, if ψ(x) solves
 
h̄2 d2
− + V (x) ψ(x) = Eψ(x),
2m dx2
then ψ(x) also gives a solution.
Now given ψ(x) with

Aeikx + Be−ikx x → −∞
ψ(x) =
Ceikx + De−ikx x → +∞

the time-reversed solution ψ(x) satisfies


Beikx + Ae−ikx x → −∞
ψ(x) =
Deikx + Ce−ikx x → +∞
! !
B A
In this presentation, Ψin = , Ψout = . Therefore
C D
! ! ! !
A B B −1 A
=S =⇒ =S .
D C C D

It follows that
unitarity of S
S = S̄ −1 ========⇒ S = St.

So time reversal symmetry implies the S-matrix is symmetric.

52
 1.10 WKB Approximation
The WKB method, named after Gregor Wentzel, Hendrik Kramers, and Léon Brillouin,
provides approximate solutions for linear differential equations with spatially slow-varying coef-
ficients. In applications to quantum mechanics, this is also called semi-classical approximation.

1.10.1 Approximation Scheme

Consider the one-dimensional time-independent Schrödinger equation

 
h̄2 d2
− + V (x) ψ = Eψ.
2m dx2
We will analyze the solution ψ in three different regions.

⃝
1 Classically allowed region. This corresponds to positions where V (x) < E. Classically, the
energy of motion is
p2
E= + V (x).
2m
So the classical particle can only move in this region.
⃝
2 Classically forbidden region. This corresponds to positions where V (x) > E. Classical
particles can not enter this region. However, as we have seen in previous examples of harmonic
oscillator and square well bound states, quantum particles can penetrate into this region with
certain probability.
⃝
3 Turning points. This corresponds to positions where V (x) = E.

V (x)
turning point

x
classically forbidden classically allowed classically forbidden

The WKB approximation scheme looks for solutions of the stationary wave function of the
exponential form (notation clarification: S(x) here is not the S-matrix)
i
ψ(x) = e h̄ S(x) , S(x) ∈ C.

Plugging into the Schrödinger equation,

 
h̄2 d2 i
− 2
− (E − V (x)) e h̄ S(x) = 0
2m dx

53

2
=⇒ S ′ (x) − ih̄S ′′ (x) = 2m(E − V (x)).

Treating h̄ as very small, this non-linear equation can be solved in order of h̄ by setting
∞
X
S(x) = h̄n Sn (x).
n=0

Equating two sides of

∞
!2 ∞
X X
h̄n Sn′ (x) − ih̄ h̄n Sn′′ (x) = 2m(E − V (x)),
n=0 n=0

we find 

 S0′ (x)2 = 2m(E − V (x))







 2S0′ (x)S1′ (x) = iS0′′ (x)



 ..
.



 X
n−1

 2S ′
(x)S ′
(x) = iS ′′
(x) − Si′ (x)Sn−i
′
(x)

 0 n n−1




i=1

 ..
.

The semi-classical approximation looks for the solution up to order h̄1 and neglects terms
of order h̄2 or higher. Thus we look for

S(x) = S0 (x) + h̄S1 (x) + O(h̄2 ).

This can be solved by the above recursive relation

ˆ xp
S0 (x) = ± 2m(E − V (y)) dy

i p
S1 (x) = ln 2m(E − V (x)) + C1
2
where C1 is some constant. Therefore
 
i 2


ψ(x) = exp S0 (x) + h̄S1 (x) + O(h̄ )
h̄
i
' e h̄ S0 (x) eiS1 (x)
A ´x√
e± h̄
i
2m(E−V (y)) dy
= 1
(2m(E − V (x))) 4

where A is some constant. The two expressions

A ´x√
e± h̄
i
2m(E−V (y)) dy
ψ± = 1
(2m(E − V (x))) 4

are the basic forms of the WKB approximation.

Note that this WKB form will blow up at the turning points where E = V (x). This
suggests that the WKB approximation is not good near turning points. There need some
special treatment there and we will come back to this shortly.

54
 V (x)

WKB approximated solution. Amplitude blows up near the turning points.

In the classically allowed region E > V (x), let us write

h̄2 k(x)2
E − V (x) = with k(x) > 0.
2m
Then the WKB approximated solution takes the form (A, B are some constants)
A ´x B ´x
ψ(x) ' p ei k(y)dy + p e−i k(y)dy .
k(x) k(x)
The first term with coefficient A represents a wave moving to the right, and the second term
represents a wave moving to the left. This function is oscillating in this region.
In the classically forbidden region E < V (x), let us write
h̄2 λ(x)2
E − V (x) = − with λ(x) > 0.
2m
Then the WKB approximated solution takes the form
A ´x B ´x
ψ(x) ' p e λ(y)dy + p e− λ(y)dy .
λ(x) λ(x)
This function is essentially exponential growing or decaying in this region.

1.10.2 Turning Points and Airy Functions

Now we consider the region near a turning point x = a. We look for a suitable approxima-
tion near this turning point that connects the oscillating WKB approximation on one side and
the exponential WKB approximation on the other side
V (x)

x
a

55
 We can approximate V (x) locally around x = a by a linear function

V (x) ' E + (x − a)V ′ (a).

This leads to the approximate equation

h̄2 d2
− ψ + (x − a)V ′ (a)ψ = 0.
2m dx2
If we make a change of variable
 1/3
2mV ′ (a)
z= (x − a),
h̄2
then the above approximate equation becomes

d2 ψ
− zψ = 0.
dz 2
This is the Airy equation.
The Airy equation has two linearly independent solutions, denoted by Ai(z) and Bi(z).
They are called Airy functions and are given by
ˆ  3 
1 ∞ t
Ai(z) = cos + zt dt
π 0 3
ˆ ∞  3   3 
1 t t
Bi(z) = exp − + zt + sin + zt dt.
π 0 3 3

The Airy functions Ai(z), Bi(z) have the approximate asymptotic behavior
  
 1 1 − 14 2 3

 2 √π |z| exp − 3 |z| 2 z → +∞
Ai(z) '  

 1 − 1 2 3 π
 √ |z| 4 cos |z| 2 − z → −∞
π 3 4
  
 1 − 14 2 3
 √
 π |z| exp 3 |z| 2 z → +∞
Bi(z) '  

 1 2 3 π
− √ |z|− 4 sin
1
|z| 2 − z → −∞
π 3 4
The key observation is that these approximate behaviors connect precisely the WKB solu-
tions on two sides of the turning point.

1.10.3 Connection Formula

Let us use the Airy functions to derive the connection condition for WKB solutions near
the turning point x = a.

1 Let us first consider the case V ′ (a) > 0.

⃝

56
 V (x)

x
a

The linear approximation of V near x = a is

V (x) ' E + V ′ (a)(x − a).

In the classically allowed region x < a near the turning point, we have approximately
p  
2m(E − V (x)) 2mV ′ (a) 1/2 √
k(x) = ' a−x
h̄ h̄2
and ˆ  1/2 ˆ
x
2mV ′ (a) x√
2
k(y) dy ' a − y dy = − |z|3/2 .
a h̄2 a 3
In the classically forbidden region x > a near the turning point, we have approximately
p  
2m(V (x) − E) 2mV ′ (a) 1/2 √
λ(x) = ' x−a
h̄ h̄2
and ˆ  1/2 ˆ
x
2mV ′ (a) x√
2
λ(y) dy ' y − a dy = |z|3/2 .
a h̄2 a 3
Assume the wave function has the leading behavior by a decaying exponential in the clas-
sically forbidden region, then the WKB solution must be the form
 ˆ x 
A
ψ(x) ' p exp − λ(y) dy , x > a.
λ(x) a
´x

In other words, the growing exponential term √ 1 exp a λ(y) dy can not appear.
λ(x)
Remark 1.10.1. Note that if ψ(x) has a leading behavior by a growing exponential, we can not
exclude the possible appearance of the decaying exponential term since this is dominated by
the growing exponential term and hence invisible in the leading behavior.

Comparing with the leading behavior of the Airy functions, we find that it can only be
connected by Ai(z) and
ˆ a 
2A π
ψ(x) ' p cos k(y) dy − , x < a.
k(x) x 4

Similarly, if we find the approximate wave function

ˆ a 
B π
ψ(x) ' p sin k(y) dy − , x<a
k(x) x 4

57
to the left of the turning point, then the leading behavior of Bi(z) implies
ˆ x 
B
ψ(x) ' − p exp λ(y) dy .
λ(x) a

In summary, we have found the following connection condition for WKB solutions near the
turning point x = a with V ′ (a) > 0.

V (x)

x
a

ˆ a   ˆ x 
2A π A
p cos k(y) dy − ⇐= p exp − λ(y) dy
k(x) x 4 λ(x) a
ˆ a  ˆ x 
B π B
p sin k(y) dy − =⇒ −p exp λ(y) dy
k(x) x 4 λ(x) a

The arrow is the implication direction.

2 Let us then consider the case V ′ (a) < 0. The discussion is similar. We find the connection
⃝
condition for WKB solutions near the turning point.

V (x)

x
a

 ˆ a  ˆ x 
A 2A π
p exp − λ(y) dy =⇒ p cos k(y) dy −
λ(x) x k(x) a 4
ˆ a  ˆ x 
B B π
−p exp λ(y) dy ⇐= p sin k(y) dy −
λ(x) x k(x) a 4

1.10.4 Semi-classical Quantization Rule

Let us consider a potential V (x) such that

V (x) → ∞ as x → ±∞.

58
We look for a bound state with energy E. By our experience with the harmonic oscillator, we
expect that the allowed energies should be discrete.
Assume also that there exist exactly two turning points x = a and x = b with a < b.

V (x)

x
a b

For a bound state ψ(x) which is normalizable, ψ(x) should decay as x → ±∞ in the
forbidden region. The WKB approximation thus takes the form
  ˆ a 

p
A
 λ(x) exp − x λ(y) dy
 x<a
ψ(x) =  ˆ x 

 B

p exp − λ(y) dy x>b
λ(x) b

By the connection formula, both these two behaviors will determine the behavior in the
classically allowed region in between. Then the consistency condition gives
ˆ x ˆ b 
π π
k(y) dy − = − k(y) dy − + nπ
a 4 x 4
ˆ b  
1
=⇒ k(y) dy = π n + for n ∈ Z.
a 2
This is known as the Einstein-Brillouin-Keller (EBK) semi-classical quantization condition
(or Keller-Maslov quantization condition), which improves the Bohr-Sommerfeld quantization
condition via the Maslov index correction ( 12 here).

Example 1.10.2. As an illustration, we consider the Harmonic oscillator with the potential
1
V (x) = mω 2 x2 .
2

59
 V (x)

x
q q
− 2E
mω 2
2E
mω 2

The EBK semi-classical quantization condition asks

q r
√ ˆ 2E    
2m mω 2 1 1 1
q E − mω x dx = π n +
2 2 =⇒ E= n+ h̄ω.
h̄ − 2E 2 2 2
mω 2

These are precisely the allowed energies of harmonic oscillators that we found before.

1.10.5 Quantum Tunneling

Consider a particle in the potential V (x) with energy E. Classically, the particle can only
move in the region V (x) ≤ E. Quantum mechanically, the particle has a chance to pass through
a potential barrier which is classically forbidden. This phenomenon is called quantum tunneling.

A V (x)
C

B
E

x
a b

Consider the potential as above, with V (x) → 0 as x → ±∞. Consider the energy E > 0
which is smaller than the height of the potential.
We consider the scattering problem with a wave incident from the left. Quantum mechan-
ically, it has a chance to pass through the potential barrier. The tunnelling probability is the
transmitted coefficient T as we discussed in Section 1.8.3. We will give an approximate formula
for T via the WKB method.

60
 Let us represent the WKB solution for the transmitted wave by
 ˆ x 
C π
ψ(x) ' p exp i k(y) dy − i , x >> b
k(x) b 4
ˆ x  ˆ x 
C π iC π
=p cos k(y) dy − +p sin k(y) dy − .
k(x) b 4 k(x) b 4
By the connection formula, the second term will match to an exponential that grows as we move
to the left from x = b.
ˆ b 
iC
−p exp λ(y) dy a<x<b
λ(x) x
 ˆ x  ˆ b 
iC
=− p exp − λ(y) dy exp λ(y) dy .
λ(x) a a

This in turn will match to the left of x = a by

ˆ b  ˆ a 
2iC π
−p exp λ(y) dy cos k(y) dy − .
k(x) a x 4
Decomposing this into left-moving and right-moving waves, the corresponding component
for the right-moving wave is
ˆ b   ˆ a 
iC π
−p exp λ(y) dy exp −i k(y) dy + i .
k(x) a x 4
This leads to the approximate relation
ˆ b 
|A| = |C| exp λ(y) dy .
a

The WKB approximated transmission coefficient is

 ˆ b   ˆ 
|C|2 2 bp
T ' = exp −2 λ(y) dy = exp − 2m(V (x) − E) dx .
|A|2 a h̄ a
Remark 1.10.3. We will revisit this WKB formula of transmission coefficient via path integral
method in Section 2.7.3. There the quantum tunneling is realized by a path in imaginary time.

Example 1.10.4. Consider the potential of a square barrier


 V0 |x| < a
V (x) =
0 else

V0 V (x)

E
x
−a a

The WKB approximated transmission coefficient is

 ˆ   
2 ap 4a p
T ' exp − 2m(V (x) − E) dx = exp − 2m(V0 − E) .
h̄ −a h̄

61
 1.11 Quantum Kepler Problem
In this section we study quantum particles in R3 under a potential of the form
Z
V (~r ) = − , Z > 0 constant.
r
Here ~r = (x1 , x2 , x3 ) are linear coordinates on R3 and
q
r = x21 + x22 + x23

is the length of ~r . The corresponding force is

Z ~r
F~ = −∇V = − 2 .
r r
In our assumption for Z > 0, this force is attractive via the inverse square law. For example,
gravitational force and attractive electrostatic force are of this type. The problem with inverse
square law is usually called the Kepler problem.
Classically, the Kepler problem exhibits rich symmetry and is completely integrable. Quan-
tum mechanically, we will see that the bound state spectrum can be also exactly solved by
symmetry. As an application, this allows us to compute the Hydrogen atom spectrum.

1.11.1 Angular Momentum

The angular momentum of the classical particle motion in R3 is

J~ = ~r × p~

where p~ is the classical momentum. In components,



 J 1 = x 2 p3 − x 3 p 2
X 

Ji = ijk xj pk or explicitly J 2 = x 3 p1 − x 1 p 3


j,k 

J 3 = x 1 p2 − x 2 p 1

Here ijk is the Levi-Civita symbol with 123 = 1. Their Poisson bracket relations are
X
{Ji , Jj } = ijk Jk .
k
.
In the quantum case, the angular momentums become the self-adjoint operators


 Jˆ1 = x̂2 p̂3 − x̂3 p̂2


X
Jˆi = ijk x̂j p̂k or explicitly Jˆ2 = x̂3 p̂1 − x̂1 p̂3


j,k 
ˆ
J3 = x̂1 p̂2 − x̂2 p̂1

The canonical commutation relations

[x̂k , p̂j ] = ih̄δkj

62
imply
h i X
Jˆk , Jˆj = ih̄ kjm Jˆm
m
h i X
Jˆk , x̂j = ih̄ kjm x̂m
m
h i X
Jˆk , p̂j = ih̄ kjm p̂m
m

Geometrically, these operators act on the Hilbert space L2 (R3 ) with pk = −ih̄ ∂x∂ k . Then

 i ˆ ∂ ∂

 J = x − x

 h̄
1 2
∂x3
3
∂x2


i ∂ ∂
Jˆ2 = x3 − x1

 h̄ ∂x 1 ∂x 3



 i ∂ ∂
 h̄ J3 = x1 ∂x − x2 ∂x
 ˆ
2 1

They generate rotations in the x2 x3 −plane, x3 x1 −plane, x1 x2 −plane, respectively. Let

Jˆ2 := Jˆ · Jˆ = Jˆ12 + Jˆ22 + Jˆ32 .

Then it is direct to check that Jˆ2 commutes with Jˆ1 , Jˆ2 , Jˆ3
h i h i h i
Jˆ2 , Jˆ1 = Jˆ2 , Jˆ2 = Jˆ2 , Jˆ3 = 0.

Jˆ2 is called the Casimir element.

The Hamiltonian operator of the Kepler problem
2
^ = p̂ − Z
H
2m r
is clearly rotational invariant. Therefore
h i
^ Jˆk = 0,
H, k = 1, 2, 3.

This can be also checked directly.

We are interested in bound states, which are normalizable solutions of the time-independent
Schrödinger equation. The energy E of such bound state has to be negative in this case.

V (r)

63
 For E < 0, let us denote

BE = {stationary states of energy E} .

Our goal is to find the allowed bound state energy E with non-trivial energy eigenspace BE .
Since the angular momentum Jˆi ’s commute with H, ^ they actually act on the space BE .
Thus BE forms a representation of SO(3), the three dimensional rotations. This puts important
constraints on BE , but not enough to determine E since generators of these symmetries do not
^
involve the Hamiltonian H.
On the other hand, the Kepler problem of inverse square law has an enhanced symmetry

SO(3) −→ SO(4)

whose generators do involve the Hamiltonian. This will enable us to compute the bound state
energy spectrum. We discuss next this enhanced symmetry.

1.11.2 Enhanced Symmetry

Classical Laplace-Runge-Lenz Vector

Consider a particle of mass m in the potential V = − Zr . The classical Hamiltonian is

p2 Z
H = − .
2m r
In the Kepler problem of inverse square force law, there exist an additional conserved
quantity called the Laplace-Runge-Lenz vector. The classical Laplace-Runge-Lenz vector is

~ = p~ × J − ~r .
~
A
mZ r
~ = (A1 , A2 , A3 ) follows from the Poisson bracket relations
The classical conservation of A

{H , Ak } = 0, k = 1, 2, 3

which can be verified directly. This allows us to solve the classical motion as follows.
Consider the inner product
 
p~ × J~ · ~r (~r × p~ ) · J~ J2
A~ · ~r = −r = −r = − r.
mZ mZ mZ
~ · ~r = Ar cos θ, where A is the length of A
Let us write A ~ and θ is the angle between A
~ and ~r .
Since both A and J 2 are conserved and hence are constants of motion, we find

J2 1
r= .
mZ 1 + A cos θ
This immediately implies that the orbit in the classical Kepler problem must be an ellipse
(A < 1), parabola (A = 1) or hyperbola (A > 1).

64
 Classically, direct computation shows the following Poisson bracket relations
2 X
{Ak , Aj } = − kjl Al H .
mZ 2
l

Moreover, we have
~ · J~ = 0
A
~·A
~ =1+ 2H ~ 2
A2 = A J .
mZ 2
Assume the classical energy is negative H < 0. Then A < 1 and the orbits are ellipses.
These are the classical analogue of bound states. Let us redefine two conserved vectors
q q
mZ 2 ~ mZ 2 ~
~
J + 2|H | A J − 2|H
~
|A
I~ = , ~ =
K .
2 2
Then they satisfy the following Poisson bracket relations
X
{Ii , Ij } = ijl Il
l
X
{Ki , Kj } = ijl Kl
l

{Ii , Kj } = 0

Thus {Ii , Ki } form the Lie algebra so(3) ⊕ so(3), which is the same as the Lie algebra so(4).
~ K
Since I, ~ are conserved (they Poisson commute with H ), we conclude that the classical Kepler
problem has enhanced SO(4) symmetry.

Quantum Laplace-Runge-Lenz Vector

Now we extend the above discussion to the quantum case. Define the quantum Laplace-
Runge-Lenz vector by
1   ~r
 = p̂ × Jˆ − Jˆ × p̂ − .
2mZ r
In the quantum case
p̂ × Jˆ 6= −Jˆ × p̂

since entries of p̂ and Jˆ do not commute. In components, we have


 1 h ˆ   i x̂1

 Â = p̂ J − p̂ ˆ2 − Jˆ2 p̂3 − Jˆ3 p̂2 −
J


1
2mZ
2 3 3
r

 1 h ˆ   i x̂2
Â2 = p̂3 J1 − p̂1 Jˆ3 − Jˆ3 p̂1 − Jˆ1 p̂3 −

 2mZ r

 1 h ˆ   i

 x̂3
 3
 = p̂ J
1 2 − p̂ ˆ
J
2 1 − ˆ
J p̂
1 2 − Jˆ p̂
2 1 −
2mZ r
h i P
Using the commutation relations Jˆk , p̂j = ih̄ kjm p̂m , we find
m

Jˆ × p̂ = −p̂ × Jˆ + 2ih̄p̂.

65
Therefore we can equivalently write
1   ~r
 = p̂ × Jˆ − ih̄p̂ − .
mZ r
We also have
 · Jˆ = Jˆ ·  = 0
2H^  
 ·  = 1 + Jˆ · Jˆ + h̄2 .
mZ 2
This last quantum relation differs from the classical one by a quantum correction h̄2 .
In the quantum case, we have the commutation relations (Exercise. See also [30] for a
detailed presentation)
h i h i
^ Âi = H,
H, ^ Jˆi = 0
h i 2 X
Âk , Âj = −ih̄ ^
kjl Jˆl H
mZ 2
l
h i X
Jˆk , Âj = ih̄ kjl Âl
l
h i X
Jˆk , Jˆj = ih̄ kjl Jˆl
l

^ they preserve the eigenspace BE (E < 0)

Since Â, Jˆ commute with H,

Âi , Jˆi : BE −→ BE .

Restricting to this subspace BE , we have

h i 2 X
Âk , Âj = −ih̄ kjl EJl .
mZ 2
l

Now we can perform the same construction as in the classical case and redefine
q q
2
Jˆ + mz  ˆ − mz 2 Â
J
2|E| 2|E|
Iˆ = , K̂ = on the subspace BE .
2 2
Their commutation relations again obey the Lie algebra of so(3) ⊕ so(3)
h i X
Iˆj , Iˆm = ih̄ jml Iˆl
l
h i X
K̂j , K̂m = ih̄ jml K̂l
l
h i
Iˆj , K̂m = 0

So we have quantum so(4) symmetry. The corresponding Casimir elements satisfy

mZ 2 h̄2
Iˆ · Iˆ = K̂ · K̂ = − on the subspace BE .
8|E| 4

It is this relation that allows us to compute the bound state energies of quantum Kepler problem.

66
1.11.3 Representations of so(3)

We review some basic facts about representations of the Lie algebra of so(3)

so(3) = A : 3 × 3 real matrix AT = −A, Tr A = 0 .

This Lie algebra is three-dimensional with a basis by

     
0 0 0 0 0 1 0 −1 0
     
t1 = 0 0 −1 ,
 t2 = 
 0 0 0 ,
 t3 = 
1 0 0
.
0 1 0 −1 0 0 0 0 0

Their commutation relations are

X
[ti , tj ] = ijk tk .
k

To study the representation, it is convenient to redefine the following complexified basis



 L := it1 − t2

 +
L− := it1 + t2




L3 := it3

They satisfy the commutation relations

[L3 , L+ ] = L+
[L3 , L− ] = −L−
[L+ , L− ] = 2L3

Moreover, L3 is Hermitian L3 = L∗3 and L+ , L− are Hermitian adjoint of each other L∗+ = L− .
A representation of so(3) is a vector space V together with a Lie algebra morphism

ρ : so(3) −→ gl(V ).

We are interested in finite dimensional complex representations. Irreducible complex represen-

tations of so(3) are classified: for each non-negative half-integer l = 0, 12 , 1, 23 , · · · , there exists
precisely one isomorphic class of irreducible representation Vl of dimC Vl = 2l + 1.
Let ρl : so(3) → gl(Vl ) denote the corresponding representation. Then on each Vl , the
element L3 can be diagonalized by
 
−l
 
 −l + 1 
 
 
 −l + 2 
 
 .. 
ρl (L3 ) =  . .
 
 
 l−2 
 
 l−1 
 
l

67
 The commutation relation
[L3 , L± ] = ±L±

says that the action of L+ (L− ) will raise (lower) the eigenvalue of L3 by one. So the actions
of L3 , L± on the representation space Vl look like

L+ L+ L+ L+

L3 : −l −l + 1 −l + 2 ··· l−2 l−1 l

L− L− L− L−

Another way to distinguish these representations is to consider the Casimir element

1
Ĉ = t21 + t22 + t23 = −L23 − (L+ L− + L− L+ ) .
2

The Casimir element Ĉ commutes with all ti ’s, hence becomes a constant when it acts on an
irreducible representation. The crucial result is that on the irreducible representation Vl

ρl (Ĉ) = −l(l + 1).

We can generalize the above discussion to the representation of the Lie algebra so(4) =
so(3) ⊕ so(3). There the finite dimensional irreducible representations are classified by
1 3
V k ⊗ Vl , k, l = 0, , 1, , · · · .
2 2
One copy of so(3) acts on the Vk -factor via the representation ρk and acts on the Vl factor
as the identity. The other copy of so(3) acts on the Vl -factor via the representation ρl and acts
on the Vk -factor as the identity. There are two Casimir elements Ĉ1 , Ĉ2 corresponding to the
two copies of so(3). In the representation Vk ⊗ Vl , we have

Ĉ1 = −k(k + 1), Ĉ2 = −l(l + 1) on Vk ⊗ Vl .

1.11.4 Energy Spectrum

Now we apply the so(3)-representation theory to analyze the quantum Kepler problem. We
^ with energy E < 0. This corresponds
consider the eigenspace BE of the Hamiltonian operator H
to bound states. A general spectral theory implies that BE is finite dimensional.
We have operators
Iˆi , K̂i : BE −→ BE

acting on BE . Thus BE forms a complex representation of the Lie algebra so(4) = so(3) ⊕ so(3).
h i X
Iˆj , Iˆm = ih̄ jml Iˆl
l
h i X
K̂j , K̂m = ih̄ jml K̂l
l
h i
Iˆj , K̂m = 0

68
 Comparing with our conventions in Section 1.11.3, the Casimir element Ĉ1 for the so(3)-
copy of {Ii } and the Casimir element Ĉ2 for the so(3)-copy of {Ki } are

Iˆ2 K̂ 2
Ĉ1 = − , Ĉ2 = − .
h̄2 h̄2
The algebraic relation on BE

mZ 2 h̄2
Iˆ · Iˆ = K̂ · K̂ = −
8|E| 4

implies Ĉ1 = Ĉ2 on BE . Thus BE consists of copies of Vk ⊗ Vk for some k. Then

mZ 2
Ĉ1 = Ĉ2 = −k(k + 1) =⇒ |E| =
2 .
8 k + 12 h̄2

Let n = 2k + 1 which is a positive integer. Then the possible bound state energies are

mZ 2
En = − , n = 1, 2, · · · .
2h̄2 n2
It turns out (via some further analysis) that each En does appear in the discrete spectrum and
each Vk ⊗ Vk appears precisely once

BEn = V n−1 ⊗ V n−1 .

2 2

In particular, the dimension of En -eigenstates is

 2
dim BEn = dim V n−1 = n2 .
2

1.11.5 Hydrogen Atom

The Hydrogen atom consists of a proton and an electron in dimension three. This can be
viewed as a two-body quantum mechanical problem.
Let {~xp , p~p } denote the position and momentum of the proton, and {~xe , p~e } denote the
position and momentum of the electron. The corresponding quantum operators satisfy the
canonical commutation relations
h i
(x̂p )i , (p̂p )j = ih̄δij
h i
(x̂e )i , (p̂e )j = ih̄δij

[x̂p or p̂p , x̂e or p̂e ] = 0.

The quantum Hamiltonian of the Hydrogen atom is

p̂2p p̂2
^=
H + e + V (|~xe − ~xp |)
2mp 2me

where mp is the proton mass and me is the electron mass. V is the central Coulomb potential

e2
V (r) = − ,
r

69
where e is the elementary electric charge.
We can simplify this problem by introducing the center-of-mass coordinates. Precisely, let
us define the center-of-mass position and momentum operators by
me x̂e + mp x̂p
x̂c = , p̂c = p̂p + p̂e .
me + mp

Define the relative position and momentum operators by

mp p̂e − me p̂p
x̂R = x̂e − x̂p , p̂R = .
me + mp

Then we can check that they still satisfy the canonical commutation relation
h i
(x̂c )i , (p̂c )j = ih̄δij
h i
(x̂R )i , (p̂R )j = ih̄δij

[x̂c or p̂c , x̂R or p̂R ] = 0.

We can work with x̂c , p̂c , x̂R , p̂R instead. The Hamiltonian operator now becomes
2 2
^ = p̂c + p̂R + V (|~xR |)
H
2Mc 2MR
where
me mp
Mc = me + mp , MR = .
me + mp
We can solve the time-independent Schrödinger equation by using separation of variables

ψ(xc , xR ) = ψc (xc )ψR (xR )

where ψc (xc ) and ψR (xR ) solve separately

p̂2c
ψ c = Ec ψ c
2Mc
 
p̂2R
+ V (|~xR |) ψR = ER ψR .
2MR
The total energy is
E = Ec + ER .

The equation for ψc says that the center of mass moves as a free particle of mass Mc . The
equation for ψR says that the relative motion between the proton and the electron is a quantum
Kepler problem. By our result in Section 1.11.4, the energy ER for bound states are quantized

MR e 4
ER,n = − , n = 1, 2, · · · .
2h̄2 n2
The number of bound states with energy ER,n is n2 .
This formula explains precisely the emission spectrum of atomic hydrogen which occurs
when an electron transits, or jumps, from a higher energy state to a lower energy state. The
observed spectral lines match with the energy difference between two energy levels as above.

70
 Chapter 2 Path Integral Formalism

In this chapter, we explain the path integral approach to quantum mechanics, which char-
acterizes quantum dynamics of particles in terms of probabilistic paths. This formalism provides
a powerful tool for calculating transition amplitudes and understanding quantum phenomena.
It has been widely generalized and developed within modern quantum field theory. The pre-
sentation in this chapter will focus on intuition and examples to elucidate the basic idea.

2.1 Path Integral: Introduction

2.1.1 Quantum Evolution and Feynman Kernel

We have discussed the state space of quantum mechanics by wave functions ψ (vectors in
a Hilbert space) and the law of quantum time evolution by the Schrödinger equation

∂ ^ ψ.
ih̄ ψ=H
∂t
^ is the Hamiltonian operator (also called Schrödinger operator), which is a differential
Here H
operator that quantizes the classical Hamiltonian function H .
We focus on time-independent H ^ in this chapter. Viewing H
^ as a self-adjoint operator on
the Hilbert space of states, the time evolution of states via Schrödinger equation is solved by

^ ′′ −t′ )/h̄
ψ(t′′ ) = e−i H(t ψ(t′ ) , t′ < t′′ .

Thus the time evolution in quantum mechanics is completely encoded in the one-parameter
^
family of unitary operators e−i H t/h̄ on the Hilbert space.
^
As we will see, the operator e−i H t/h̄ can be represented by an integral kernel. This means
that the evolution of the wave function ψ(x, t) can be expressed by the integral relation
ˆ
ψ(x′′ , t′′ ) = dx′ K(x′′ , t′′ ; x′ , t′ )ψ(x′ , t′ ).

This integral kernel K plays the major role in a different formulation of quantum mechanics:
the “path integral” approach.
In classical mechanics, the principle of least action plays a primary role. The classical
system is usually described by an action functional
ˆ
S [x(t)] = dt L (x(t), ẋ(t))

71
where L is called the Lagrangian. The trajectories of classical particles are stationary points
of the action S, which can be described by the Euler-Lagrange equation
 
∂L d ∂L
− = 0.
∂xi dt ∂ ẋi

This Lagrangian formulation of classical mechanics is related to the Hamiltonian formula-

tion by the Legendre transform

H (x, p) = p · ẋ − L (x, ẋ).

∂L
Here pi = ∂ ẋi is the Legendre transform from ẋ, which is called the conjugate momentum of xi .
Remarkably, the study of this integral kernel K leads directly to the Lagrangian formu-
lation! This was first observed by Dirac in his study of canonical transformation of conjugate
variables in quantum mechanics. This viewpoint was not essentially used until Feynman who
developed the complete story of the “path integral approach to quantum mechanics”. In this
story, the integral kernel K has the interpretation as an “integration” over the space of paths
ˆ x(t′′ )=x′′
′′ ′′ ′ ′ i
K(x , t ; x , t ) = [Dx(t)] e h̄ S[x(t)] .
x(t′ )=x′

classical trajectory
x′ x′′

Here [Dx(t)] is expected to be certain measure over the space of paths

x : [t′ , t′′ ] −→ Space

with endpoints x(t′ ) = x′ , x(t′′ ) = x′′ .

One essential feature is that all paths will contribute to the integral kernel K through the
action functional S. This expression provides a direct relation between classical and quantum
mechanics. In the classical limit when h̄ → 0, the method of stationary phase suggests that
the above path integral will have dominate contributions from the stationary paths, which are
precisely the classical trajectories! This clean and intuitive interpretation has been generalized
and applied to many quantum physics and now become standard in textbooks.
Unfortunately, the path space is very big and infinite dimensional. In many quantum me-
chanical cases of our interest at hand, this can be related to Markovian evolution and Brownian

72
motion, thus the Wiener’s measure is available. For general path integral in quantum field
theory, the rigorous mathematical construction of the corresponding measure is yet unknown.
This has been one of the major foundational challenges for modern quantum theory.
Nevertheless, the path integral approach offers a deep insight into many quantum problems.
Even without a general rigorous measure available, we can still do many concrete calculations in
physics. Actually, one major motivation of Feynman in developing the path integral formulation
is to apply this to study quantum electrodynamics. One reason for the calculation power of
path integral lies in the formalism itself. For the usual finite dimensional integral
ˆ
f

we almost never compute it by definition of Riemann integral or Lebesgue integral. Instead,

we usually compute it by symmetry and differential equations that can be derived from certain
formal and natural properties provided by the integration. This is usually the situation how
we manipulate path integrals in physics. Assuming some natural elementary properties of the
path integral that we borrow from the ordinary integral, we can do many concrete calculations.
The above story of path integral is also called the Feynman path integral. The integral
kernel K is usually called the Feynman kernel in the literature.

2.1.2 Position and Momentum Representation

We will mainly focus on the Hilbert space

L2 (Rn ).

A state ψ(x) ∈ L2 (Rn ) is a square integrable measurable function

ˆ
dn x |ψ(x)|2 < ∞.

The inner product is ˆ

hψ1 |ψ2 i = dn x ψ1 (x)ψ2 (x).

It would be convenient to introduce the eigenvector |x′ i of the position operator x̂ by

x̂i x′ = x′i x′ .

Strictly speaking, |x′ i does not lie in the Hilbert space and corresponds to the continuous
spectra of the self-adjoint operator x̂. The wave function of |x′ i is the δ-function

x′ ⇝ δ(x − x′ )

which is non-normalizable. This state lies in the space of tempered distributions. Nevertheless,
we can formally work with such states to simplify many presentations. For example, we can
treat all eigenvectors of the position operator as a “basis” with normalized inner product by

x′′ x′ = δ(x′′ − x′ ).

73
 Any state |ψi can be expanded in this basis by
ˆ
|ψi = dn x ψ(x)|xi.

Thus the wave function has the interpretation as the coefficients in terms of such basis of position
eigenvectors. Equivalently, we can write

ψ(x) = hx|ψi.

This can be justified by

ˆ ˆ ˆ
hx|ψi = hx| dn x′ ψ(x′ ) x′ = dn x′ ψ(x′ ) x x′ = dn x′ ψ(x′ )δ(x − x′ ) = ψ(x).

We can also rewrite the formula

ˆ ˆ
|ψi = d x hx|ψi|xi = dn x |xihx|ψi
n

as the completeness relation ˆ

dn x |xihx| = 1

where 1 represents the identity operator.

Similarly, we can introduce the eigenvectors |p′ i of the momentum operator p̂ by

p̂i p′ = p′i p′ .

In our convention, we will normalize them by

p′′ p′ = (2πh̄)n δ(p′′ − p′ ).

Thus the completeness relation reads

ˆ
1
dn p |pihp| = 1.
(2πh̄)n
The position and momentum eigenvectors are related by Fourier transform
ˆ
|pi = dn x eix·p/h̄ |xi

or equivalently
hx|pi = eix·p/h̄ .

Its complex conjugate gives

hp|xi = e−ix·p/h̄ .

A state |ψi can be either expanded by the position eigenvectors to get

ψ(x) = hx|ψi

or expanded by the momentum eigenvectors to get

ψ̂(p) = hp|ψi.

74
They are related by
ˆ ˆ
1 1
ψ(x) = hx|ψi = hx| dn p |pihp|ψi =
dn p eix·p/h̄ ψ̂(p)
(2πh̄)n (2πh̄)n
ˆ ˆ
ψ̂(p) = hp|ψi = hp| d x |xihx|ψi = dn x e−ix·p/h̄ ψ(x)
n

which are precisely the Fourier transform formula.

We can use the above representation to express the integral kernel K. Let |ψi be an initial
state at t = 0. Let |ψ, ti denote the state at time t hence |ψi = |ψ, 0i. Then

^
|ψ, ti = e−i H t/h̄ |ψi.

The corresponding wave function is

^
ψ(x, t) = hx|ψ, ti = hx|e−i H t/h̄ |ψi.

If we compare at two different times t′ and t′′ ,

^ ′′ /h̄ ^ ′′ ′ ^ ′
|ψi = x′′ e−i H(t −t )/h̄ e−i H t /h̄ |ψi
ψ(x′′ , t′′ ) = x′′ e−i H t
ˆ 
^ ′′ −t′ )/h̄
′′ −i H(t ^ ′
= x e n ′ ′ ′
d x x x e−i H t /h̄ |ψi
ˆ
^ ′′ ′
= dn x′ x′′ e−i H(t −t )/h̄ x′ ψ(x′ , t′ )

we find the following expression for the Feynman kernel K

^ ′′ −t′ )/h̄
K(x′′ , t′′ ; x′ , t′ ) = x′′ e−i H(t x′ .

In summary, we can view K(x′′ , t′′ ; x′ , t′ ) as the matrix entries of the evolution operator
^ ′′ −t′ )/h̄
e−i H(t represented in the basis of the position eigenvectors.

2.2 Path Integral via Time Slicing

2.2.1 Free Particle

We start the study of the Feynman kernel K from the example of the free particle. The
Hamiltonian is
2
^ 0 = p̂ ,
H m = mass.
2m
We denote the Feynman kernel of the free particle by K0

^ ′′ −t′ )/h̄
K0 (x′′ , t′′ ; x′ , t′ ) = x′′ e−i H0 (t x′ .

To describe this kernel, we can first compute

^ p̂2 t using p2 t p2 t
hp|e−i H0 t/h̄ |xi = hp|e−i 2mh̄ |xi ======== e−i 2mh̄ hp|xi = e−i 2mh̄ e−ix·p/h̄ .
⟨p|p̂i =pi ⟨p|

75
It follows that
ˆ
′′ ^ 0 (t′′ −t′ )/h̄
−i H ′ 1 ^ ′′ ′
x e x = n
dn p x′′ p hp|e−i H0 (t −t )/h̄ x′
(2πh̄)
ˆ
1 p2 (t′′ −t′ ) ′′ ′
−i 2mh̄
= n
d n
p e ei(x −x )·p/h̄ .
(2πh̄)
Using the Gaussian integral formula
ˆ r
−au2 +bu π b2
du e = e 4a ,
a
the above integral is (strictly speaking we need to do analytic continuation. See Section 2.5.2)
  n m(x′′ −x′ )2
′′ ′′ ′ ′ m 2 i
h̄ 2(t′′ −t′ )
K0 (x , t ; x , t ) = e .
2πh̄i(t′′ − t′ )
This gives an explicit formula for the integral kernel of the free particle.
Remark 2.2.1. Note that when the time is purely imaginary with

i(t′′ − t′ ) = τ > 0

and when m = h̄2 , the kernel K becomes

1 (x′′ −x′ )2
−
e 4τ ,
(4πτ )n/2
which is precisely the kernel for the heat operator eτ ∇ on Rn . This is the expected result since
2

^ 0 = − h̄2 ∇2 = −h̄∇2 and

in this case H 2m
^ 1 (x′′ −x′ )2
x′′ eτ ∇ x′ = x′′ e− H0 τ /h̄ x′ = −
2
e 4τ .
(4πτ )n/2
We will come back to the discussion of imaginary time in Section 2.2.5.

2.2.2 Infinitesimal Time

We next consider the Feynman kernel for the general Hamiltonian operator
2
^ = p̂ + V (x)
H
2m
with the evolution in an infinitesinally small amount of time δt
^
K(x′′ , t + δt; x′ , t) = x′′ e−i H δt/h̄ x′ .

Keeping the first order in δt, we have approximately

^ p̂2 δt V (x̂)δt
e−i H δt/h̄ ' e−i 2mh̄ e−i h̄ .

Using the previous result on the free Feynman kernel, we find

p̂2 δt V (x̂)δt
K(x′′ , t + δt; x′ , t) ' x′′ e−i 2mh̄ e−i h̄ x′
p̂2 δt V (x′ )δt
= x′′ e−i 2mh̄ x′ e−i h̄
 m  n m(x′′ −x′ )2 V (x′ )δt
ei 2h̄δt −i h̄
2
=
2πh̄iδt  
 m  n i m  x′′ −x′ 2 −V (x′ ) δt
2 h̄ 2 δt
= e .
2πh̄iδt

76
 Note that the Lagrangian of the classical system emerges naturally. In a small amount of
 ′′ ′ 2
x −x
time, the quantity m
2 δt is approximately the kinetic energy. Thus the expression
 2
m x′′ − x′
− V (x′ )
2 δt

approximates precisely the kinetic energy minus the potential energy, i.e., the Lagrangian.

2.2.3 Composition Law

The quantum evolution fulfills the semi-group property

^ ′′′ −t′ )/h̄ ^ ′′′ −t′′ )/h̄ ^ ′′ −t′ )/h̄

e−i H(t = e−i H(t e−i H(t , t′ < t′′ < t′′′ .

In terms of the integral kernel, this becomes

^ ′′′ −t′ )/h̄ ^ ′′′ ′′ ^ ′′ ′
x′′′ e−i H(t x′ = x′′′ e−i H(t −t )/h̄ e−i H(t −t )/h̄ x′
ˆ
^ ′′′ ′′ ^ ′′ ′
= dn x′′ x′′′ e−i H(t −t )/h̄ x′′ x′′ e−i H(t −t )/h̄ x′ ,

i.e., ˆ
K(x′′′ , t′′′ ; x′ , t′ ) = dn x′′ K(x′′′ , t′′′ ; x′′ , t′′ )K(x′′ , t′′ ; x′ , t′ ).

This is the composition law for the Feynman kernel K.

Geometrically, this composition law can be illustrated by

x′′
x′ x′′′

t′ t′′ t′′′

Thinking about K(x′′′ , t′′′ ; x′ , t′ ) as a transition amplitude from x′ at time t′ to x′′′ at time t′′′ ,
the composition law says that this transition amplitude is the same as summing over all the
transitions at the intermediate time t′′ for all possible x′′ .
We can further subdivide the time interval for t0 = t′ < t1 < t2 < · · · < tN −1 < tN = t′′ .

t0 t1 ··· tN −1 tN
1 1
=

t′ t′′

77
Then the same consideration leads to the composition law
ˆ ˆ ˆ
K(x′′ , t′′ ; x′ , t′ ) = dn x1 dn x2 · · · dn xN −1

K(x′′ , t′′ ; xN −1 , tN −1 )K(xN −1 , tN −1 ; xN −2 , tN −2 ) · · · K(x1 , t1 ; x′ , t′ ).

x1 xN −2

x2
xN −1
x0 = x′ ··· xN = x′′

t0 t1 t2 tN −2 tN −1 tN
1 1
=

=
t′ t′′

This is again interpreted as summing over all possible intermediate transitions at time t1 , t2 , · · · , tN −1 .

2.2.4 Path Integral

Now we can subdivide the time interval [t′ , t′′ ] into small intervals for sufficiently large N

t0 = t′ < t1 < t2 < · · · < tN −1 < tN = t′′ ,

where
t′′ − t′
tj = t′ + j, = .
N
The composition law gives
ˆ NY
−1 Y
N −1
K(x′′ , t′′ ; x′ , t′ ) = dn xj K(xj+1 , tj+1 ; xj , tj ), x0 = x′ , xN = x′′ .
j=1 j=0

Applying our result for the integral kernel over small time interval, this is approximated by
 x 2 
 m N n/2 ˆ NY
NP
−1
−1 j+1 −xj
i
h̄
m
2 ϵ
−V (xj ) ϵ
' n
d xj e j=0
.
2πh̄i
j=1

In the limit N → ∞ or  → 0, this integral is expected to reach the following form of

Feynman path integral
ˆ x(t′′ )=x′′ ´ t′′
′′ ′′ ′
K(x , t ; x , t ) = ′
[Dx(t)] e h̄
i
t′ ( m2 ẋ2 −V (x))dt
x(t′ )=x′

for some suitable measure on the space of paths going from x′ at time t′ to x′′ at time t′′ .

78
 ···
x′ x′′

t′ t′′

Note that ˆ t′′ m 

S [x(t)] = ẋ2 − V (x) dt
t′ 2
is precisely the classical action functional. So the integral kernel K can be also written as
ˆ x(t′′ )=x′′
′′ ′′ ′ ′ i
K(x , t ; x , t ) = [Dx(t)] e h̄ S[x(t)] .
x(t′ )=x′

We will illustrate how to analyze this path integral in a suitable sense in subsequence sections.
It light of this above formula, Feynman’s path integral can be understood as a Lagrangian
formulation of quantum mechanics, providing an alternative viewpoint compared to the tradi-
tional Hamiltonian operator approach.

2.2.5 Imaginary Time

A more convenient way to obtain a mathematically better behaved path integral is to make
an analytic continuation in time to

t = −iτ, τ ∈ R.

This analytic continuation is called Wick rotation. The corresponding path integral is called
the Euclidean path integral. We denote the integral kernel in imaginary time by
^ ′′ ′
KE (x′′ , τ ′′ ; x′ , τ ′ ) := K(x′′ , −iτ ′′ ; x′ , −iτ ′ ) = x′′ e− H(τ −τ )/h̄ x′
ˆ x(τ ′′ )=x′′ ´ ′′ m 2
1 τ
= [DE x(τ )] e− h̄ τ ′ ( 2 ẋ +V (x))dτ .
x(τ ′ )=x′

At this point we can use the conditional Wiener measure to define [DE x(τ )]. In the math-
ematical literature, this above path integral representation for the Euclidean Feynman path
integral is established as the Feynman-Kac formula.
We will also denote the Euclidean action by
ˆ  
m 2
SE [x(τ )] = ẋ + V (x) dτ.
2
The Euclidean path integral becomes
ˆ x(τ ′′ )=x′′
′′ ′′ ′ ′
[DE x(τ )] e− h̄ SE [x(τ )] .
1
KE (x , τ ; x , τ ) =
x(τ ′ )=x′

The physical meaning of imaginary time is that the Euclidean kernel

^
ρ(x′′ , x′ ; β) := x′′ e−β H x′

becomes the density matrix in statistical mechanics.

79
 2.3 Gaussian Path Integral

2.3.1 Gaussian Integral

Recall the Gaussian integral formula

ˆ r
−ax2 π
dx e = , a > 0.
R a
This Gaussian integral can be extended to the imaginary phase a = −iλ via analytic continua-
p π q iπ
tion (see Example 2.5.2). It picks up the branch of a = λ by

ˆ r r
iλx2 iπ iπ π
dx e = = e 4 sign λ , λ ∈ R − {0}.
R λ |λ|

This can be generalized to n-dimensional case as follows.

Proposition 2.3.1. Let A = (aij ) be a symmetric positive definite real matrix. Then
ˆ
π n/2
dn x e−x
t Ax
=√ . (*)
Rn det A
P
n
Here xt Ax = aij xi xj . More generally, we have
i,j=1

ˆ
π n/2 1 Jt A−1 J
dn x e−x
t Ax+Jt ·x
=√ e4 .
Rn det A
 
λ1
 ..  −1
Proof: Let A = P 
 . P
 where P ∈ SO(n). We consider the change of variables
λn

x = P y.

Since P ∈ SO(n), this change of variables has trivial Jacobian dn x = dn y. Moreover

 
λ1
 .. 
xt Ax = yt P t AP y = yt P −1 AP y = yt 
 .  y = λ1 y12 + λ2 y22 + · · · + λn yn2 .

λn

Therefore
ˆ ˆ P
n
n r
Y
− λi yi2 π π n/2
−xt Ax
n
d xe = n
d ye i=1 = =√ .
Rn Rn λi det A
i=1

The case with a linear term Jt · x follows by completing the square.

Remark 2.3.2. This result can be analytically continued to the imaginary case A = −iΛ and
ˆ
π n/2 π n/2
= e 4 (n+ −n− ) p
t iπ
dn x eix Λx = p
Rn det(−iΛ) | det Λ|

where n+ and n− are respectively the number of positive and negative eigenvalues of Λ.

80
2.3.2 Zeta Function Regularization

Let us apply the idea of Gaussian integral to path integrals. To illustrate the basic idea,
we start with the simplest example of one-dimensional free particle to compute the kernel
ˆ x(T )=x′′ ´
i T m 2
′′ ′
K0 (x , T ; x , 0) = [Dx(t)] e h̄ 0 ( 2 ẋ ) dt .
x(0)=x′

Let xcl (t) denote the classical trajectory from x′ at time t = 0 to x′′ at time t = T
t ′′
xcl (t) = x′ + (x − x′ ).
T
Any path x(t) with x(0) = x′ , x(T ) = x′′ can be written as

x(t) = xcl (t) + γ(t),

where the path γ(t) satisfies γ(0) = γ(T ) = 0. We can view γ(t) as the quantum fluctuations
around the classical trajectory xcl (t).

xcl (t)
x′′

x′
x(t) = xcl (t) + γ(t)

Since the classical trajectory xcl (t) is a stationary point of the action, the action functional
S [x] = S [xcl + γ] has no linear dependence in γ. Thus
ˆ T ˆ T ˆ T ˆ T
m 2 m 2 m 2 m 2
S [x] = (ẋcl + γ̇) dt = ẋcl dt + γ̇ dt = S [xcl ] + γ̇ dt.
0 2 0 2 0 2 0 2
This can be also checked directly. We can write the second term via integration by part as
ˆ T ˆ  2
m 2 m T d
γ̇ dt = − γ 2
γ dt.
0 2 2 0 dt
Thus the path integral becomes
ˆ γ(T )=0 ´T  d2

0 γ −
im
K(x′′ , T ; x′ , 0) = e h̄ S[xcl ]
i γ dt
[Dγ(t)] e 2h̄ dt2 .
γ(0)=0

Observe that the path integral in γ(t) becomes the form of Gaussian integral, though in the
infinite dimensional space of paths γ(t) with the Dirichlet boundary conditions γ(0) = γ(t) = 0.
The analogy with the finite dimensional Gaussian integral is

i ←→ t
xi ←→ γ(t)
X ˆ
←→ dt
i
ˆ Y ˆ
dxi ←→ [Dγ(t)]
i

81
 Let us denote the elliptic operator

d2
A=− .
dt2
Comparing with the finite dimensional Gaussian integral, we would expect a result of the form
ˆ γ(T )=0 ´
im T
 2

γ − d 2 γ dt
= N (det A)− 2 ,
1
[Dγ(t)] e 2h̄ 0 dt

γ(0)=0

where N is some normalization constant to be determined.

We need to give a meaning to the determinant of the operator A. We consider the eigenvalue
problem for A with Dirichlet boundary conditions

Aγm (t) = λm γm (t), γm (0) = γm (T ) = 0.

We know from the general theory of eigenvalue problem that

0 < λ 1 ≤ λ2 ≤ · · · ≤ λm ≤ · · · , λm → ∞ as m → +∞

and {γm (t)} form an orthonormal basis of the Hilbert space of square integrable functions γ(t)
with γ(0) = γ(T ) = 0. Then the naive definition of det A would be
∞
Y
?
det A == λm .
m=1

However, this naive product is divergent since λm → ∞ as m → ∞.

Fortunately, there is a way out using analytic continuation. Let us define the zeta function
associated to the elliptic operator A by
∞
X 1
ζA (s) := , s ∈ C.
λsm
m=1

It is known that the series for ζA (s) is well-defined for Re s sufficiently large, and can be
analytically continued to the origin s = 0. Thus

′
ζA (0) is well defined.

Intuitively, the derivative formula

∞
X
′ − ln λm
ζA (s) = , Re s  0
λsm
m=1

Q
∞
suggests that the naive product λm should be defined by the analytic continuation
m=1

∞
Y ′
‘‘ λm ” := e−ζA (0) .
m=1

Then we can define the functional determinant of the operator A by

′
det A := e−ζA (0) .

82
 Let us see how this works. The eigenfunctions and eigenvalues of A are easily found
  

 mπt
γm (t) = cm sin , cm some constant
T
  2

λm = πm .
T
Thus the zeta function ζA (s) is
∞ 
X   2s
T 2s T
ζA (s) = = ζ(2s),
πm π
m=1

where
∞
X 1
ζ(s) =
ms
m=1

is the Riemann zeta function. Using the known result

1 1
ζ(0) = − , ζ ′ (0) = − ln 2π,
2 2
we compute
   
′ T ′ T
ζA (0) = 2 ln ζ(0) + 2ζ (0) = − ln − ln 2π = − ln 2T.
π π

Thus
′
det A = e−ζA (0) = 2T.

Plugging this into the path integral, we arrive at

N i N im(x′′ −x′ )2
K0 (x′′ , T ; x, 0) = √ e h̄ S[xcl ] = √ e 2h̄T .
2T 2T
This is consistent with our previous result for free particle, with the normalization constant
 m 1
2
N= .
πih̄
This result suggests that for general elliptic operator A in dimension one, we have the
Gaussian path integral
ˆ γ(T ) ´T  m 1
(det A)− 2
im 2 1
γAγ dt
[Dγ(t)] e 2h̄ 0 =
γ(0) πih̄

where det A is defined via the analytic continuation through the zeta function ζA (s)
′
det A := e−ζA (0) .

This method is called zeta function regularization.

83
 2.4 Harmonic Oscillator

2.4.1 Integral Kernel

As an illustration of the path integral method, we revisit the example of one-dimensional

harmonic oscillator. The Hamiltonian operator is

^ = 1 p̂2 + m ω 2 x̂2 .
H
2m 2
We calculate the integral kernel
ˆ x(T )=x′′ ´T
′′ ′ ′′ ^ T /h̄
−i H ′ i
( m2 ẋ2 − m2 ω2 x2 )dt .
K(x , T ; x , 0) = x e x = [Dx(t)] e h̄ 0
x(0)=x′

The action functional is ˆ T m m 2 2

S [x(t)] = ẋ2 − ω x dt.
0 2 2
Let xcl (t) be the classical trajectory which satisfies the classical equation of motion

ẍcl (t) = −ω 2 xcl (t), xcl (0) = x′ , xcl (T ) = x′′

with specified boundary condition xcl (0) = x′ and xcl (T ) = x′′ . This is solved by

sin ω(T − t) ′ sin ωt ′′

xcl (t) = x + x .
sin ωT sin ωT
We can decompose any path x(t) with x(0) = x′ and x(T ) = x′′ by

x(t) = xcl (t) + γ(t),

where γ(t) is an arbitrary path with boundary condition

γ(0) = γ(T ) = 0.

The action becomes

S [x(t)] = S [xcl (t) + γ(t)]

ˆ T
m 2 m 2 2
= S [xcl (t)] + γ̇ − ω γ dt
0 2 2
ˆ T  
m d2
= S [xcl (t)] + γ − 2 − ω 2 γ dt
2 0 dt
ˆ T
m
= S [xcl (t)] + γAγ dt.
2 0
Here A is the elliptic operator
d2
A=− − ω2.
dt2
The path integral is Gaussian and therefore
ˆ γ(T )=0 ´T  m 1 i
e h̄ S[xcl (t)] (det A)− 2 .
1
K(x′′ , T ; x′ , 0) = e h̄ S[xcl (t)]
i im 2
γAγ dt
[Dγ(t)] e 2h̄ 0 =
γ(0)=0 πih̄

84
 We compute det A via the zeta function regularization as in Section 2.3.2. The eigenfunc-
tions and eigenvalues are
  
 mπt

 γ (t) = cm sin , cm some constant
 m T
"  2 #
  πm 2  πm 2

 − ω2 = 1−
ωT
 λm =
T T πm

Then naively we find

"  2 !# "   #
∞
Y  πm 2 ωT
∞ 
Y πm 2 Y
∞
ωT 2
det A = 1− = 1− .
T πm T πm
m=1 m=1 m=1

The first term is regularized by Riemann zeta function, which is computed in Section 2.3.2
∞ 
Y πm 2
= 2T.
T
m=1

The second term is convergent. In fact, using

∞ 
Y  z 2  sin z
1− = ,
mπ z
m=1

we have
∞
"  2 #
Y ωT sin ωT
1− = .
πm ωT
m=1

Therefore  
sin ωT 2 sin ωT
det A = (2T ) = .
ωT ω
The Feynman kernel K becomes

mω  1 i
′′ ′ 2
K(x , T ; x , 0) = e h̄ S[xcl (t)] .
2πih̄ sin ωT
The action of the classical trajectory is
ˆ
m T

S [xcl (t)] = ẋcl 2 − ω 2 xcl 2 dt
2 0
ˆ
m T m T

= xcl ẋcl − xcl ẍcl + ω 2 xcl 2 dt
2 0 2 0
m
= (xcl (T )ẋcl (T ) − xcl (0)ẋcl (0))
2    
m ω ′ cos ωT ′′ ′′ m cos ωT ′ ω
= − x +ω x x − −ω x + x x′
′′
2 sin ωT sin ωT 2 sin ωT sin ωT
mω  ′ 2

= (x ) + (x′′ )2 cos ωT − 2x′ x′′ .
2 sin ωT
We arrive at the final result for the integral kernel of the quantum harmonic oscillator
  1 imω h i
mω 2x′ x′′
′′ ′ 2 ((x′ )2 +(x′′ )2 ) cot ωT − sin
K(x , T ; x , 0) = e 2h̄ ωT
.
2πih̄ sin ωT

85
2.4.2 Partition Function

We would like to compare the Feynman kernel of the harmonic oscillator with previous
result on the energy spectrum in Section 1.7. The link is the partition function defined by

^
Tr e−β H .

We will compute this partition function in two different ways.

The first way is to compute the partition function through the energy spectrum
∞
X ∞
X e− 2 βh̄ω
1
^
−β H −βEn −β (n+ 12 )h̄ω 1
Tr e = e = e = −βh̄ω
= .
1−e 2 sinh (βh̄ω/2)
n=0 n=0

The second way of computing the partition function is to use the Feynman kernel. Let us
consider the imaginary time
T = −iτ

and the Euclidean integral kernel

^
KE (x′′ , τ ; x′ , 0) = K(x′′ , −iτ ; x′ , 0) = x′′ e− H τ /h̄ x′ .

The partition function as a trace can be also expressed by

ˆ ˆ ˆ
^ ^
Tr e−β H = dx hx|e−β H |xi = dx KE (x, βh̄; x, 0) = dx K(x, −iβh̄; x, 0).

Plugging our explicit result for the Feynman kernel,

ˆ  1
mω 2 2mω sinh2 (βωh̄/2) 2
^
−β H − x
Tr e = dx e h̄ sinh(βωh̄)

R 2πh̄ sinh(βωh̄)
 1
mω πh̄ sinh(βωh̄) 2
=
2πh̄ sinh(βωh̄) 2mω sinh2 (βωh̄/2)
1
= .
2 sinh(βωh̄/2)

This is the same result as the energy spectrum calculation in a nontrivial way. Physics works!

2.5 Asymptotic Method

In this section, we review some basic tools for asymptotic analysis of the oscillatory inte-
grals. This will help us tackle path integrals in later sections to obtain semi-classical results
in quantum mechanics. The subject is rather classical, but could run easily into very technical
discussion. Instead, we choose the intuitive approach and illustrate the basic idea via examples,
aiming at motivating our later path integral manipulations.

86
2.5.1 Laplace’s Method

We start with Laplace’s method for analyzing integrals of the form

ˆ b
e−λf (x) dx
a

which provides the leading asymptotic approximation as λ → +∞. For simplicity, we assume

• [a, b] is a finite interval. The discussion can be generalized to the case when a = −∞ or
b = +∞ (or both) under further mild assumption of f near the infinity endpoint.

• f is a twice continuously differentiable function on [a, b] with a unique global minimum

at an interior point x0 ∈ (a, b) and

f ′′ (x0 ) > 0.

Under these assumptions, Laplace’ method shows

´b
e−λf (x) dx
lim a q = 1. (*)
λ→∞ e−λf (x0 ) 2π
λf ′′ (x 0)

We will usually write this as an asymptotic approximation

ˆ b s
−λf (x) −λf (x0 ) 2π
e dx ' e ′′ (x )
as λ → +∞.
a λf 0

The idea of the approximation formula (*) via Laplace’s method is that in the limit λ →
+∞, the integral is dominated by ˆ x0 +ε
e−λf (x) dx
x0 −ε

in a small neighborhood of the global minimum x0 . If we do Taylor series expansion around x0

(note that f ′ (x0 ) = 0)

f ′′ (x0 )
f (x) = f (x0 ) + (x − x0 )2 + O((x − x0 )3 )
2
and perform a change of variable
y
x = x0 + √ ,
λ
then
ˆ ˆ √λε  
x0 +ε
−λf (x) e−λf (x0 ) f ′′ (x )
− 2 0 y 2 +O √1
e dx = √ √ e
λ dy
x0 −ε λ − λε
ˆ
e−λf (x0 ) +∞ − f ′′ (x0 ) y2
' √ e 2 dy as λ → +∞
λ −∞
s
−λf (x0 ) 2π
=e .
λf ′′ (x0 )

87
 f (x)

f (x0 ) δ

a x0 − ε x0 x0 + ε b x

Outside [x0 − ε, x0 + ε], e−λf (x) = e−λf (x0 ) O(e−λδ )

The proof of (*) is basically to realize the above idea via a careful analysis of the error.
This asymptotic formula can be generalized to the case when a = −∞ or b = +∞ (or both)
under some further mild assumption of f near the infinity endpoint.
We can also generalize the above discussion to
ˆ b
g(x)e−λf (x) dx
a

where g(x) is positive. Then

ˆ s
b
−λf (x) −λf (x0 ) 2π
g(x)e dx ' g(x0 )e as λ → +∞.
a λf ′′ (x0 )

Example 2.5.1. Consider the Γ-function

ˆ +∞
Γ(s) = dx xs−1 e−x .
0

We consider its asymptotic behavior as s → +∞. The above discussion generalizes to this case.
Let us rewrite it as
ˆ +∞
Γ(s) = (s − 1)s dx e−(x−ln x)λ , λ=s−1
0

thus f (x) = x − ln x in this case. The minimum is at the point

1
f ′ (x0 ) = 1 − =0 =⇒ x0 = 1.
x0
It follows that we have an asymptotic approximation
s
2π
Γ(s) ' (s − 1)s e−λf (x0 )
λf ′′ (x0 )
√
= 2π(s − 1)s−1/2 e1−s
r √  
2π s s 1 s 1−s
= √ s 1− e
s s−1 s
r  
2π s s
' as s → +∞.
s e
This is known as the Stirling’s formula.

88
 The above discussion can be generalized to the n-dimensional case
ˆ
e−λf (x) dn x, x = (x1 , · · · , xn ).
Γ

Assume f (x) has a unique global minimum x0 in the interior of the domain for integration, and
the Hessian matrix Hf = (∂xi ∂xj f ) is positive definite at x0

Hf (x0 ) > 0.

By a similar consideration, in the limit λ → +∞, the dominate contribution comes from
the near neighborhood of x0 and the leading approximation is given by the Gaussian integral
ˆ  n/2
−λf (x0 ) −λ 2π 1
e n
d xe 2
xt Hf (x0 )x
= p e−λf (x0 ) .
Rn λ det(Hf (x0 ))
Thus Laplace’s method leads to
ˆ  n/2
−λf (x) 2π g(x0 )
n
d x g(x)e ' p e−λf (x0 ) as λ → +∞.
λ det(Hf (x0 ))

2.5.2 Method of Steepest Descent

Laplace’s method can be extended to the complex oscillatory integrals of the form
ˆ
I(λ) = dz g(z)e−λf (z)
C

where f (z) and g(z) are analytic functions of z. By Cauchy integral formula, this complex
integral is invariant under continuous deformations of C (with appropriate boundary condition
at infinity boundary of C). This will allow us to perform analytic continuation and asymptotic
analysis by deforming the contour suitably.

Example 2.5.2. Consider the following integral

ˆ
2
eix dx.
C

It can be viewed as an analytic continuation of the standard Gaussian integral

ˆ
e−λx dx,
2
λ > 0.
R

The analytic continuation from λ > 0 to λ = −i can be realized by choosing the contour:
y

Cθ
θ > 0 small
θ
x

89
Then ˆ
e−λz dz
2
I(λ) =
Cθ

is convergent for Re λ ≥ 0, Im λ ≤ 0, λ 6= 0. Thus I(λ) gives the analytic continuation from

λ ∈ R>0 to λ ∈ iR<0 . In particular
ˆ ˆ ˆ
π √
+∞
e−u du = e 4 i π.
π
iz 2 iz 2 i 2
I(−i) = e dz = e dz = e 4
Cθ Cπ −∞
4

Similarly, if we deform the contour clock-wise, then we obtain an analytic continuation

from λ > 0 to λ = i ˆ
π √
e−iz dz = e− 4 i π.
2

C− π
4

This explains our discussion on Gaussian integrals in Section 2.3.1.

In general, the method of steepest descent is to deform the contour C into a curve along
which e−f (z) decays fastest. To see how such a contour looks like, let us write

f (z) = u(x, y) + iv(x, y), z = x + iy.

So u = Re f and v = Im f . The steepest descent curve should follow the gradient of u since

|e−f | = e−u .

Since f (z) is analytic, u and v satisfy the Cauchy-Riemann equations

∂x u = ∂y v, ∂y u = −∂x v.

It follows that
∇u · ∇v = ∂x u∂x v + ∂y u∂y v = 0.

In other words, ∇v is perpendicular to the gradient direction of u, thus the steepest descent
curve will lie on a level set of v. This motivates the following strategy: we deform the contour
C into a contour C ′ such that
1 Im f is constant along C ′
⃝
2 C ′ passes through one or more points where
⃝

f ′ (z) = 0.

These are called saddle points. They are also the critical points of Re f along C ′ . Then
ˆ ˆ
I(λ) = g(z)e−λf (z) dz = e−iλ Im f g(z)e−λ Re f dz
C′ C′

and we can apply Laplace’s method.

Let us assume there is one non-degenerate saddle point z0 on the contour C ′ which is a
global minimum on C ′ . The condition of non-degenerate saddle point says

f ′ (z0 ) = 0, f ′′ (z0 ) 6= 0.

90
 Then a local computation via Gaussian integral gives
ˆ s
2π
g(z)e−λf (z) dz ' g(z0 )e−λf (z0 ) as λ → +∞.
C′ λf ′′ (z0 )

There are similar results for n-dimensional complex integrals. The branch for the square root
from the Gaussian integral is determined by the analytic continuation as in the above example.

Example 2.5.3. Consider the Gaussian integral

ˆ
2
eiλz dz, λ > 0.
C

Then f (z) = −iz 2 and

Re f = 2xy, Im f = y 2 − x2 .

The steepest descent contour is

y

C π4

2.5.3 Morse Flow

Concretely, curves of steepest descent can be constructed via Morse theory. We follow the
presentation [49] to illustrate the basic idea in our case. Consider the following flow equation
 dz

 = −f ′ (z)
du

 lim z(u) = saddle
u→−∞

where u ∈ R is the real parameter of the flow. Along the flow we have

df (z)
= −|f ′ (z)|2 .
du
So Im f is constant and Re f is decreasing along the flow, leading to a steepest descent curve.

Example 2.5.4. Let us consider an example

ˆ
αz 2
eλf (z) , f (z) = , α ∈ C∗ , λ > 0.
C 2

91
The saddle point is at z = 0. The curve C of steepest descend satisfies

Im f = 0, and Re f → −∞ along ∂C.

Such curve C can be constructed by solving the flow equation

 dz

 = −f ′ (z) = −αz
du

 lim z(u) = 0
u→−∞

Let α = Aeiθ , A > 0. There are two solutions pointing toward opposite directions
i
z(u) = ±eAu e 2 (π−θ) .

Here is the figure for the corresponding solutions (arrow indicates the flow direction)

π
2 − θ
2

The solutions flow along the direction with angle as in the figure
π θ π 1
− = − Arg(f ′′ (0)).
2 2 2 2
With appropriate orientations chosen, these two flows combine to form the curve C.

π
2 − θ
2

92
 In general, suppose we are to describe the curve of steepest descend for the integral
ˆ
eλf (z)
C

passing a saddle point z0 . Then the tangent line of C at z0 is along the direction with angle
π 1
− Arg(f ′′ (z0 )).
2 2

z0
π
2 − 12 Arg(f ′′ (z0 ))

2.5.4 Stokes Phenomenon

In applications of asymptotic method, we will often encounter cases when there are several
saddle points and we need to sum them all. However, as we vary parameters of the model, the
sum of the asymptotic expansions may exhibit discontinuous jump. This is known as the Stokes
phenomenon. Such jump phenomenon actually displays important physical behaviors. We will
explain the basic idea of Stokes phenomenon through a concrete example, the Airy integral.

Airy Function

We consider the following Airy integral along the real line

ˆ +∞  z3  ˆ +∞   3 
iλ −z z
I(λ) := e 3 dz = 2 cos λ −z dz
−∞ 0 3
for λ ∈ R>0 . This integral is convergent and is related to the standard Airy function Ai
ˆ  3 
1 +∞ z
Ai(x) = cos + xz dz
π 0 3
by
 2
I(λ) = 2πλ− 3 Ai −λ 3 .
1

We are interested in the asymptotic behavior of I(λ) as λ → +∞. As we discussed before,

this can be analyzed by the method of steepest descent. To apply this method, the first step is
to deform the contour C = R into contours of steepest descent. Let
 3 
z
f (z) = iλ −z .
3

93
 It has two saddle points

f ′ (z) = 0 =⇒ z = p± = ±1.

A curve of steepest descent is a contour which passes a saddle point and satisfies
⃝
1 Im f = constant
⃝
2 Re f → −∞ along the infinity endpoints of the curve.
Since
2
f (p± ) = ∓ iλ
3
we have two curves C± of steepest descent corresponding to the two saddle points p± .
To describe these curves, let us consider the flow equation
dz
= −f ′ (z) = iλ(z 2 − 1)
du
which can be written in real coordinates z = x + iy as

 dx
 = 2λxy
du (λ > 0)

 dy = λ(x2 − y 2 − 1)
du
We can draw the direction of the corresponding flow

x2 = y 2 + 1

−1 +1

This allows us to draw the curves C± as

94
 −1 +1

C− C+

It follows that for λ > 0

ˆ  3  ˆ 
z3
 ˆ 
z3

iλ z3 −z iλ −z iλ −z
I(λ) = e dz = e 3
dz + e 3
dz
R C− C+

i.e., the contour R is deformed to C− + C+ for steepest descent. Thus I(λ) has the asymptotic
behavior via the method of steepest descent
r    r   
π 2 iλ 1 π − 2 iλ 1
I(λ) ' e3 1+O + e 3 1+O as λ → +∞.
iλ λ −iλ λ

Stokes Ray

Now we consider the analytic continuation of the Airy integral

ˆ  3 
iλ z3 −z
e dz
C

as λ varies. In particular, we would like to analyze the asymptotic behavior in the limit

λ −→ +∞eiθ

as λ approaches ∞ in the direction of eiθ .

As λ varies, we need to deform the integration contour C accordingly so as to keep the
λ of steepest
integral convergent. Again, we can decompose C as a combination of curves C±
λ are associated to the two saddle points and described by
descent. The curves C±
⃝ λ passes the saddle point p = ±1
1 C± ±
⃝ λ
2 Im f = constant along C±
⃝
3 Re f → −∞ along the infinity endpoints of C±
λ

where  
z3
f = iλ −z .
3
Note that
2
Im (f (p± )) = ∓ Re λ.
3
So as long as Re λ 6= 0, curves C± do not intersect and approaches ∞ in different regions
λ

95
 P− P+

Re f = −∞ − 23 iλ

2
3 iλ

If we deform the contour C into

λ λ
C = n− C − + n+ C + ,

then the Airy integral becomes

ˆ ˆ ˆ
ef dz = n− ef dz + n+ ef dz.
C λ
C− λ
C+

λ will rotate with λ. As long as λ does not hit the locus

Now as we vary λ, the cycles C±
{Re λ = 0}, the curves C±
λ will vary continuously. However, when

Re λ = 0

so λ becomes pure imaginary,

Im(f (p− )) = Im(f (p+ ))

while
2
Re(f (p± )) = ± Im λ.
3
Thus one of the curve of steepest descent connects the two saddle points. The two rays

{λ ∈ iR>0 } ∪ {λ ∈ iR<0 }

in the λ-plane are called Stokes rays.

96
 Stokes rays
λ-plane

λ
C+
λ
C−
p+
p− z-plane

on the Stokes ray λ ∈ iR>0 : Re(f (p+ )) > Re(f (p− ))

λ will display a discrete transformation. The

When we cross the Stokes ray, the curves C±
following figure explains the crossing of one of the Stokes ray

λ
C− λ λ
C−
C− λ λ
C+ C+
λ
C+

across the Stokes ray

λ display a transformation as explained by the figure

The corresponding contours C±

C λ 7→ C λ
− −
C λ 7→ C λ ± C λ
+ + −

We apply the above result to the analytic continuation of

ˆ  3 
iλ z3 −z
e dz
C

as we vary λ and deform the contour C accordingly. We can decompose the contour C into a
λ
sum of curves of steepest descent C±

λ λ
C = n− C − + n+ C + .

97
 As a result, to keep the continuity of deformation of C along analytic continuation, the
numbers (n− , n+ ) will be locally constant away from the Stokes ray but display a jump

n+ 7→ n+
n 7→ n ± n
− − +

when we cross a Stokes ray.

Asymptotic Sum

Now we consider the asymptotic behavior of the Airy integral

ˆ  3 
iλ z3 −z
I(λ) = e dz as λ → +∞eiθ .
C
By the method of steepest descent, we first deform the contour C into a combination
λ λ
C = n− C − + n+ C + .

Then the method of steepest descent gives the leading asymptotic behavior
ˆ  3  ˆ  3 
iλ z3 −z iλ z3 −z
I(λ) = n− e dz + n+ e dz
λ
C− λ
C+
r r
π 2 iλ π − 2 iλ
' n− e 3 + n+ e 3 as λ → +∞eiθ .
iλ −iλ
Note that
´
⃝
1 if Im λ > 0, then λ dominates
´ C+
⃝
2 if Im λ < 0, then λ
C− dominates
The rays {Im λ = 0} on the real line separating the dominant asymptotic behaviors are some-
times called anti-Stokes rays

λ-plane
´
λ
C+ dominates

anti-Stokes ray
´
λ
C− dominates

This is compatible with our previous discussion on the Stokes jump. For example, consider
the case when we cross the Stokes ray λ ∈ R>0 .

λ-plane

´
λ
C+ dominates

98
 The asymptotic expansion
ˆ ˆ
I(λ) ' n− +n+
λ
C− λ
C+

will display a jump 

 n+ →
7 n+
n 7→ n ± n
− − +

This is possible since it does not alter the leading asymptotic behavior of I(λ), which is given
´
by the dominate term n+ C λ .
+

2.6 Semi-classical Approximation

The goal of this section is to apply the method of steepest descent to compute the asymp-
totic leading contribution to the Feynman kernel
ˆ x(t′′ )=x′′
′′ ′′ ′ ′ i
K(x , t ; x , t ) = [Dx(t)] e h̄ S[x(t)]
x(t′ )=x′

in the classical limit h̄ → 0. For simplicity, we focus on the one-dimensional case.

2.6.1 Semi-classical Feynman Kernel

The saddle point of S is the classical trajectory xcl (t). We write a general path x(t) by

x(t) = xcl (t) + γ(t)

where γ(t) satisfies the endpoint condition

γ(t′ ) = γ(t′′ ) = 0.

We can expand S [x(t)] around the classical trajectory xcl (t) and find
ˆ t′′   ˆ t′′  
m 2 m 2 1 ′′
S [x(t)] = ẋ − V (x) dt = S [xcl (t)] + γ̇ − V (xcl )γ dt + O(γ 3 ).
2
t ′ 2 t ′ 2 2
Thus the method of steepest descent leads to the following leading asymptotic contribution
ˆ γ(t′′ )=0 ´ ′′ m 2 1 ′′
i t
′′ ′′ ′ ′
[Dγ(t)] e h̄ t′ ( 2 γ̇ − 2 V (xcl )γ )dt
i 2
K(x , t ; x , t ) ' e h̄
S[xcl (t)]
as h̄ → 0.
γ(t′ )=0

This is called the semi-classical approximation.

Let us denote the semi-classical Feynman kernel by
ˆ γ(t′′ )=0 ´ ′′ m 2 1 ′′
i t
Ksc (x′′ , t′′ ; x′ , t′ ) := e h̄ S[xcl (t)] [Dγ(t)] e h̄ t′ ( 2 γ̇ − 2 V (xcl )γ )dt .
i 2

γ(t′ )=0

Apply our result on Gaussian path integral, the semi-classical Feynman kernel becomes
 m 1 i
Ksc (x′′ , t′′ ; x′ , t′ ) =
2
e h̄ S[xcl (t)]
πih̄ det A

99
where A is the elliptic operator

d2 1
A=− 2
− V ′′ (xcl (t)) .
dt m
The main goal of this subsection is to show the following semi-classical formula
 1
′′ ′′ ′ ′ m 2 i
Ksc (x , t ; x , t ) = e h̄ S[xcl (t)]
2πih̄ϕ0 (t′′ )

where ϕ0 (t) is the solution to the initial value problem

 

 d2 1 ′′
 − 2 − V (xcl (t)) ϕ0 (t) = 0
dt m


ϕ0 (t′ ) = 0, ϕ′0 (t′ ) = 1.

Determinant Computation

Let us denote
1 ′′
Θ(t) = V (xcl (t)) .
m
We present an intuitive computation of
 
d2
det A = det − 2 − Θ(t)
dt

due to Coleman [10].

The idea is to analyze the eigenvalue problem with initial condition at t = t′


Aϕλ (t) = λϕλ (t)

ϕ (t′ ) = 0,
λ ϕ′λ (t′ ) = 1.

For any λ, there exists a unique solution ϕλ (t) for the above initial value problem. The
key is to observe that
ϕλ (t′′ ) = 0

if and only if λ is an eigenvalue of A for the corresponding Dirichlet boundary value problem


Aϕλ (t) = λϕλ (t)

ϕ (t′ ) = ϕ (t′′ ) = 0.
λ λ

Now let us consider another operator

d2
à = − − Θ̃(t)
dt2
and similarly solve ϕ̃λ (t) for


Ãϕ̃λ (t) = λϕ̃λ (t)

ϕ̃ (t′ ) = 0,
λ ϕ̃′λ (t′ ) = 1.

100
Then we claim that
det(A − λ) ϕλ (t′′ )
= (*)
det(Ã − λ) ϕ̃λ (t′′ )
Intuitively this follows by “observing” that both sides are meromorphic functions of λ with
zeroes at eigenvalues of A and poles at eigenvalues of à (a careful analysis shows that they are
simple zeroes or poles). Let
det(A − λ) ϕλ (t′′ )
f (λ) = , g(λ) = .
det(Ã − λ) ϕ̃λ (t′′ )
f (λ)
Then the above consideration says that g(λ) is an entire function on C.
We next analyze the behavior of f (λ) and g(λ) as λ → ∞. Firstly, we have

lim f (λ) = 1.
λ→∞
λ∈R
/ +

Qualitatively this can be understood as follows. For Dirichlet boundary value problem, the
2
operator − dt
d
2 has eigenvectors

 
nπ
un (t) = sin (t − t′ )
t − t′
′′

 2 2
with eigenvalues nπ
t′′ −t′ . Therefore the shifted operator − dt
d
2 − λ has eigenvalues

 2
nπ
µn = − λ.
t − t′
′′

In the limit λ → ∞ for λ ∈

/ R+ , all eigenvalues |µn | → +∞. Thus Θ(t) is very small
2
comparing to the operator − dt
d
2 − λ in the limit λ → ∞, λ ∈
/ R+ . This small perturbation will
cause negligible effect in this limit. Therefore it is natural to expect

lim f (λ) = 1.
λ→∞
λ∈R
/ +

Secondly, we consider the limit

lim g(λ).
λ→∞
λ∈R
/ +

To analyze this limit, we first consider the inhomogeneous problem

 

 d2
 − 2 −λ u=f
dt

 ′
 u(t ) = 0, u′ (t′ ) = 0.

This is uniquely solved by

ˆ t √ 
1
u(t) = − √ sin λ(t − s) f (s) ds.
t′ λ
The corresponding Green’s operator is
ˆ t √ 
1
G(f )(t) := − √ sin λ(t − s) f (s) ds.
t′ λ

101
 Consider the original boundary value problem
 

 d2
 − 2 − λ − Θ(t) ϕλ (t) = 0
dt


 ϕλ (t′ ) = 0, ϕ′λ (t′ ) = 1.

We will write ϕλ (t) as

(0)
ϕλ (t) = ϕλ (t) + u(t)
(0)
where ϕλ (t) solves  

 d2 (0)
 − 2 − λ ϕλ = 0
dt


ϕ(0) (t′ ) = 0, ϕλ (t′ ) = 1.
(0)′
λ

(0)
Such ϕλ is explicitly found by

1 √ 
λ(t − t′ ) .
(0)
ϕλ (t) = √ sin
λ
Then the equation for ϕλ (t) becomes
   

 d2 (0)
 − 2 − λ u(t) = Θ(t) ϕλ (t) + u(t)
dt


u(t′ ) = 0, u′ (t′ ) = 0.

Using the above Green’s operator, this is equivalent to

  
(0)
u = G Θ ϕλ + u .

Let us rewrite this by

 
(0)
(1 − G ◦ Θ̂)u = G ◦ Θ̂ ϕλ ,

where Θ̂ is the operator

Θ̂(u)(t) = Θ(t)u(t).

Then we find the perturbative solution

  ∞
X  
−1 (0) (0)
u = (1 − G ◦ Θ̂) G ◦ Θ̂ ϕλ = (G ◦ Θ̂) n
ϕλ .
n=1

From this expression, we find that the correction u will have the asymptotic behavior
 
Θ (0)
u=O √ ϕλ
λ
(0)
which is small comparing to ϕλ in the limit λ → ∞, λ ∈
/ R. It follows that

ϕλ (t′′ )
(0)
lim g(λ) = lim (0)
= 1.
λ→∞
λ∈R
/ +
λ→∞
λ∈R
/ + ϕλ (t′′ )

102
 Combining the above two results, we find

f (λ)
lim = 1.
λ→∞ g(λ)
λ∈R
/ +

Since f (λ)/g(λ) is an entire function, it follows that f (λ)/g(λ) = 1. This shows (*).
Let us rewrite (*) as
det(A − λ) det(Ã − λ)
′′
= .
ϕλ (t ) ϕ̃λ (t′′ )
This implies
det A = cϕ0 (t′′ )

where c is a constant that does not depend on the potential V . The constant c can determined
2
by our result in the free case where à = − dt
d
2 . In the free case, we know from Section 2.3.2

 
d2
det(Ã) = det − 2 = 2(t′′ − t′ ).
dt

On the other hand, the differential equation


 d2
− ϕ̃0 (t) = 0
dt2


ϕ̃0 (t′ ) = 0, ϕ̃′0 (t′ ) = 1

is solved by
ϕ0 (t) = t − t′ .

We find
det Ã
c= = 2.
ϕ̃0 (t′′ )
Thus
det A = 2ϕ0 (t′′ ).

We have now arrived at the promised formula for the semi-classical Feynman kernel
 1
′′ ′′ ′ ′ m 2 i
Ksc (x , t ; x , t ) = e h̄ S[xcl (t)]
2πih̄ϕ0 (t′′ )

where ϕ0 (t) is the solution to the initial value problem

 

 d2 1 ′′
 − 2 − V (xcl (t)) ϕ0 (t) = 0
dt m


ϕ0 (t′ ) = 0, ϕ′0 (t′ ) = 1.

2.6.2 Jacobi Field

The function ϕ0 (t) appearing in the semi-classical Feynman kernel has a geometric inter-
pretation in terms of Jacobi field. We illustrate this connection together with a few applications.

103
 Consider all classical trajectories that start from x′ at the initial time t = t′ . They are
parametrized by the initial velocity v ′ at t = t′ . Let us denote xcl (t; v ′ ) for the classical trajectory
that solves the initial value problem for the equation of motion


mẍcl + V ′ (xcl ) = 0

x (t′ ; v ′ ) = 0,
cl ẋcl (t′ ; v ′ ) = v ′ .

Here ( ˙ ) means ∂( )
∂t . Thus xcl (t; v ′ ) gives a family of classical trajectories parametrized by v ′ .
We consider the variation of this family with respect to the parameter v ′
∂
J(t; v ′ ) = xcl (t; v ′ ).
∂v ′
This is called the Jacobi field. Differentiating the equation of motion mẍcl + V ′ (xcl ) = 0 with
respect to v ′ , we find that the Jacobi field satisfies

∂2
m J + V ′′ (xcl )J = 0
∂t2
which is called the Jacobi equation.
J

xcl
v′
x′

The initial condition gives

 
xcl (t′ ; v ′ ) = 0 J(t′ ; v ′ ) = 0
=⇒
ẋ (t′ ; v ′ ) = v ′ J(t
˙ ′; v′) = 1
cl

So J(t; v ′ ) satisfies the following initial value problem

 

 − ∂ 2 1 ′′
 − V (xcl ) J = 0
∂t2 m


J| ′ = 0, ˙ t=t′ = 1
J|
t=t

Thus the function ϕ0 (t) from the determinant in Section 2.6.1 is precisely the Jacobi field

ϕ0 = J.

Remark 2.6.1. The point t′′ at which J(t′′ ; v ′ ) = 0 is called a conjugate point. In presence of
a conjugate point, there is a family of classical trajectories that start from the same point at
t = t′ and end with the same point at t = t′′ .

104
 J

x′ x′′

Let us denote ˆ  
t′′
′′ ′′ ′ ′ 1
S cl (x , t ; x , t ) := mẋcl 2 − V (xcl ) dt
t′ 2
for the action on the classical trajectory x(t) that starts at x(t′ ) = x′ and ends at x(t′′ ) = x′′ .
We learn from classical mechanics that S cl (x′′ , t′′ ; x′ , t′ ) satisfies the Hamilton-Jacobi equation

 ∂S cl ∂S cl
 ′′ = p(t′′ ), = −p(t′ ) = −mv ′
∂x ∂x′

 ∂S cl = −E, ∂S cl
= E.
∂t′′ ∂t′
1 2
Here p(t) = mẋcl (t) is the conjugate momentum at time t and E = 2 mẋcl + V (xcl ) is the
energy along the trajectory xcl . It follows that
∂ 2 S cl ∂p(t′ ) ∂v ′ m 1 1 ∂ 2 S cl
′ ′′
= − ′′
= −m ′′
=− =⇒ = − .
∂x ∂x ∂x ∂x J(t′′ ) J(t′′ ) m ∂x′ ∂x′′
Thus the semi-classical contribution of the classical path xcl (t) can be also written as
 1  1   12
m 2 i
S m 2 i
S i ∂ 2 S cl i
e h̄ cl = e h̄ cl = e h̄ S cl .
2πih̄ϕ0 (t′′ ) 2πih̄J(t′′ ) 2πh̄ ∂x′ ∂x′′
In general, if we have several classical trajectories from x′ at t = t′ to x′′ at t = t′′ , then
the semi-classical Feynman kernel is the sum of contributions from all classical trajectories

X   12
′′ ′′ ′ ′ i ∂ 2 S cl i
Ksc (x , t ; x , t ) = e h̄ S cl
xcl 2πh̄ ∂x′ ∂x′′
xcl (t′ )=x′
xcl (t′′ )=x′′

where S cl (x′′ , t′′ ; x′ , t′ ) = S[xcl (t)] is the value of the classical action on the corresponding
classical trajectories. This is known as the Van Vleck-Pauli-Morette formula.

Example 2.6.2 (Free Particle). The classical trajectory is a straight line. Given initial point
(x′ , t′ ) and final point (x′′ , t′′ ), the unique classical trajectory is
x′ (t′′ − t) x′′ (t − t′ )
xcl (t) = + ′′ .
t′′ − t′ t − t′
Then ˆ t′′
′′ ′′ ′ ′ 1 m (x′′ − x′ )2
S cl (x , t ; x , t ) = mẋcl (t)2 dt =
t′ 2 2 t′′ − t′

105
 ∂ 2 S cl m
=⇒ = − ′′
∂x′ ∂x′′ t − t′
  1 m(x′′ −x′ )2
′′ ′′ ′ ′ m 2 i
h̄ 2(t′′ −t′ )
=⇒ Ksc (x , t ; x , t ) = e .
2πih̄(t′′ − t′ )
In the free case, the semi-classical approximation is exact, i.e., Ksc = K.

Example 2.6.3 (Harmonic Oscillator).

ˆ m
t′′
m 2 2
S[x(t)] = ẋ2 − ω x dt.
t′ 2 2
The equation of motion for the classical trajectory is

ẍ + ω 2 x = 0.

When ω(t′′ − t′ ) ∈
/ Zπ, there is a unique classical trajectory from x′ at t = t′ to x′′ at t = t′′ by

sin ω(t′′ − t) ′ sin ω(t − t′ ) ′′

xcl (t) = x + x .
sin ω(t′′ − t′ ) sin ω(t′′ − t′ )

Let
T := t′′ − t′

Then
ˆ m
′′ ′′ ′ ′
t′′
m 2 2
S cl (x , t ; x , t ) = ẋcl 2 −
ω xcl dt
t′ 2 2
mω  ′ 2

= (x ) + (x′′ )2 cos ωT − 2x′ x′′
2 sin ωT

∂ 2 S cl mω
=⇒ ′ ′′
=−
∂x ∂x sin ωT
  21  1 i
′′ ′′ ′ ′ i ∂ 2 S cl i mω 2
=⇒ Ksc (x , t ; x , t ) = e h̄ S cl = e h̄ S cl .
2πh̄ ∂x′ ∂x′′ 2πih̄ sin ωT
For the harmonic oscillator, the semi-classical approximation is also exact: Ksc = K.

2.6.3 Time-slicing Method

The semi-classical contribution to the Feynman kernel from a classical path xcl (t)
ˆ γ(t′′ )=0 ´ t′′  1 i
i
S[xcl (t)] i
( m2 γ̇ 2 − 21 V ′′ (xcl )γ 2 ) dt = m 2
e h̄ [Dγ(t)] e h̄ t′ e h̄ S[xcl (t)]
γ(t′ )=0 πih̄ det A

d2 1
where A = − 2
− V ′′ (xcl (t))
dt m
can be also understood from a heuristic computation via limit process of time-slicing for paths.

106
 ε

t0 t1 ··· tj tj+1 ··· tN −1 tN

1 1
=

=
t′ t′′

t′′ −t′
Let us subdivite the time interval [t′ , t′′ ] into N small intervals of width  = N . Let

V ′′ (xcl (tj ))
ωj2 :=
m
denote the value of V ′′ (xcl ) at the discrete point t = tj . By construction,
ˆ γ(t′′ )=0 ´ t′′
i
[Dγ(t)] e h̄ t′ ( m2 γ̇ 2 − 12 V ′′ (xcl )γ 2 ) dt
γ(t′ )=0
" #
x 2
 m  N ˆ NY i P
N −1 mωj2
−1 j+1 −xj
h̄
m
2 ε
− 2
x2j ε
2
= lim dxj e j=0 here x0 := 0, xN := 0
ε→0 2πih̄ε
j=1
 m  N ˆ NY
−1
im t
dxj e 2h̄ε x AN x
2
= lim here x = (x1 , · · · , xN −1 )
ε→0 2πih̄ε
j=1
 1
m 2
= lim .
ε→0 2πih̄ε det AN
Here AN is the (N − 1) × (N − 1) matrix
 
2 − ε2 ω12 −1 0
 .. 
 −1 2 − 2 ω22 . 
 
 .. .. .. 
AN =  . . . .

 
 .. 
 . 2 −  2 ωN
2
−2 −1 
0 −1 2 −  2 ωN
2
−1

Let us define u0 = ε and for 1 ≤ j ≤ N − 1

 
2 − ε2 ω12 −1 0
 .. 
 −1 2 − 2 ω22 . 
 
 . . .. 
uj = ε det 

.. .. . .

 
 .. 
 . 2 − 2 ωj−1
2 −1 
0 −1 2 − 2 ωj2

Then we have the recursive relation

uj+1 − 2uj + uj−1 2
+ ωj+1 uj = 0, 1 ≤ j ≤ N − 2.
ε2
In the continuum limit ε → 0, this becomes a differential equation for u(t) (uj = u(tj ))

ü(t) + ω 2 (t)u(t) = 0.

107
Here ω 2 (t) = V ′′ (xcl (t))/m. The initial condition
u1 − u0
u0 = ε, = 1 − ε2 ω12
ε
becomes the initial condition
u(t′ ) = 0, u′ (t′ ) = 1.

It follows that
lim ε det AN = u(t′′ )
ε→0

where u(t) solves 

 1
ü(t) + V ′′ (xcl (t))u(t) = 0
m

u(t′ ) = 0, u′ (t′ ) = 1.
This is the same result as we find before.

2.7 Green’s Function

2.7.1 Green’s Function with Fixed Energy

In the study of Schrödinger equation, it is useful to go to energy eigenstates and study the
stationary solutions with fixed energy. In the path integral formalism, we are thus led to define
the Green’s function G at fixed energy via the Fourier transform of the retarded Feynman kernel
θ(t)K(x′′ , t; x′ , 0). Here 
1 t≥0
θ(t) =
0 t<0
is the Heaviside step function. Precisely
ˆ
′′ ′ 1 +∞
G(x , x ; E) := dT ei(E+iε)T /h̄ θ(T )K(x′′ , T ; x′ , 0)
ih̄ −∞
ˆ
1 +∞
= dT ei(E+iε)T /h̄ K(x′′ , T ; x′ , 0)
ih̄ 0
ˆ
1 +∞ ^
= dT ei(E+iε)T /h̄ x′′ e−i H T /h̄ x′
ih̄ 0
1
= x′′ x′ .
^
E − H +iε
Here, as often used in distributions, a small positive ε > 0 has been introduced to ensure
convergence of the integral and we take ε → 0+ eventually.
The Green’s function G(x′′ , x′ ; E) is analytic in the region Im E > 0, reflecting the retar-
dation under Fourier transform. Knowing the Green’s function, we can recover the Feynman
kernel via the inverse Fourier transform
ˆ
i
K(x′′ , T ; x′ , 0) = dE e−iET /h̄ G(x′′ , x′ ; E).
2π R

108
Remark 2.7.1. As an illustration of the method ε → 0+ , the following distributional identity is
commonly used  
1 1
lim = ∓iπδ(x) + P.V. .
ε→0+ x ± iε x


Here δ(x) is the Dirac δ-function. P.V. x1 is the Cauchy principal value defining the distribution
 
1
P.V. : Cc∞ (R) −→ R
x
ˆ
f (x)
f 7−→ lim dx.
R−[−ε,ε] x
ε→0 +

We will simply drop the ε in formulae and keep in mind the meaning ε → 0+ . Thus
1
G(x′′ , x′ ; E) = x′′ x′ .
^
E−H
In functional analysis, the operator for z ∈ C\ spec(A), is called the resolvent of the
1
A−z ,
operator A. Thus G is precisely the resolvent integral kernel of the Hamiltonian H.^ It represents
^ and satisfies
the inverse of E − H
 
E−H ^ G(x′′ , x′ ; E) = x′′ x′ = δ(x′′ − x′ ).

^ is the discrete set {Ek } with orthonormal eigenstates {ψk }, then G

If the spectrum of H
can be written as a sum
X 1 X ψk (x′′ )ψ ∗ (x′ )
G(x′′ , x′ ; E) = x′′ |ψk i ψk x′ = k
.
^
E−H E − Ek
k k

Thus the energy eigenvalues {Ek } are detected by the poles of G. In general when both bound
and scattering states exist, we will have a spectral integral
ˆ
′′ ′ dp ψp (x′′ )ψp∗ (x′ )
G(x , x ; E) = .
R 2πh̄ E − E(p)
Example 2.7.2 (Free Particle). The Hamiltonian of the free particle is
2
^ 0 = p̂
H
2m
with the free Feynman kernel
 m  12 im(x′′ −x′ )2
K0 (x′′ , T ; x′ , 0) = e 2h̄T .
2πh̄iT
The energy spectrum is continuous. We can compute the free Green’s function by
1
G0 (x′′ , x′ ; E) = x′′ x′
E − ^
H
ˆ
1 1
= dp x′′ |pi p x′
2πh̄ R E − ^
H
ˆ
1 1 ′′ ′
= dp eip(x −x )/h̄ (ε → 0+ )
2πh̄ R E − p2 /2m + iε
ˆ
1 1 ′′ ′
= dp eip|x −x |/h̄ (ε → 0+ )
2πh̄ R E − p2 /2m + iε
√ √
We can compute the last integral by residue. The poles are located at p± = ± 2m E + iε. The
integral picks up a residue at p+

109
 p+

p−

′′ ′
!
i eip|x −x |/h̄
=⇒ G0 (x′′ , x′ ; E) = Resp=p+ (ε → 0+ )
h̄ E − p2 /2m + iε
r
1 m i√2mE|x′′ −x′ |/h̄
= e .
ih̄ 2E
h̄2 k2
In the physical region E > 0 with E = 2m (k > 0), we have
m ik|x′′ −x′ |
G0 (x′′ , x′ ; E) = e .
ih̄2 k
There is a standard way to construct Green’s function that we briefly recall for the case at
hand. Let ψ1 (x) and ψ2 (x) be two linearly independent solutions of
 
^ ψi (x) = 0
E−H

i.e.,  
h̄2 d2
− V (x) + E ψi (x) = 0.
2m dx2
Form the following function
2m
ξ(x, y; E) := (θ(x − y)ψ1 (x)ψ2 (y) + θ(y − x)ψ1 (y)ψ2 (x))
h̄2 W
where W = ψ1′ (x)ψ2 (x)−ψ2′ (x)ψ1 (x) is called the Wronskian which is a constant by the equation.
Using θ′ (x − y) = δ(x − y), it is direct to check that ξ(x, y; E) satisfies the equation
 2 2 
h̄ d
− V (x) + E ξ(x, y; E) = δ(x − y).
2m dx2

In order for ξ(x, y; E) = G(x, y; E) to hold, we need to choose ψ1 and ψ2 to satisfy appro-
priate boundary conditions. We illustrate this in the example of scattering problem.

Example 2.7.3. Let us consider the case

lim V (x) = 0
|x|→+∞

110
which is relevant for the scattering problem. Consider the inverse Fourier transform
ˆ
i
K(x′′ , T ; x′ , 0) = dE e−iET /h̄ G(x′′ , x′ ; E)
2π R
which can be viewed as a superposition of wavefunctions from different energies.
When x′′ → +∞, the Feynman kernel should behave like an outgoing plane waves with
positive momentum and energy. This tells
√
2mEx′′
G(x′′ , x′ ; E) ∝ e h̄ x′′ → +∞.
i
as

Similarly, behavior of an outgoing plane when x′′ → −∞ tells

√
2mEx′′
G(x′′ , x′ ; E) ∝ e− h̄ as x′′ → −∞.
i

This leads to the following asymptotic behavior for the solution ψ1 and ψ2
 √
ψ1 (x) ∝ e h̄i 2mEx , x → +∞
√
ψ (x) ∝ e− h̄ 2mEx , x → −∞
i
2

To illustrate this, consider the free particle when V = 0. The solutions

 2 2 
h̄ d
+ E ψi (x) = 0
2m dx2

for E > 0 with the expected boundary behavior is


ψ1 (x) = eikx √
2mE
, k= .
ψ (x) = e−ikx h̄
2

The Wronskian is W = ψ1′ ψ2 − ψ1 ψ2′ = 2ik. Then the Green’s function is

2m

G(x′′ , x′ ; E) =2 θ(x′′ − x′ )ψ1 (x′′ )ψ2 (x′ ) + θ(x′ − x′′ )ψ1 (x′ )ψ2 (x′′ )
h̄ 2ik 
m ′′ ′ ′ ′′
= 2 θ(x′′ − x′ )eik(x −x ) + θ(x′ − x′′ )eik(x −x )
ih̄ k
m ′′ ′
= 2 eik|x −x |
ih̄ k
which is precisely the formula we found above.
In general when a localized potential V (x) is turned on, we are in the situation of scattering
process. Consider the energy E > 0. Then the behavior

ψ1 (x) ∝ eikx , x → +∞

can be more precisely described by



eikx + B(k)e−ikx x → −∞
ψ1 (x) =

A(k)eikx x → +∞

111
 This solution represents an incoming plane wave from x = −∞ scattering through the po-
tential region. The coefficient B represents the amplitude of the reflected wave, and A represents
the amplitude of the transmitted wave.
Similarly, the behavior
ψ2 (x) ∝ e−ikx , x → −∞

can be precisely described by


A(k)e−ikx x → −∞
ψ2 (x) =
e−ikx + C(k)eikx x → +∞
This solution represents an incoming plane wave from x = +∞ scattering through the po-
tential region. The coefficient C represents the amplitude of the reflected wave, and A represents
the amplitude of the transmitted wave. Note that the transmission amplitudes from the left and
from the right are the same, due to time reversal symmetry.
We can compute the Wronskian from the behavior at x = ±∞ and find

W = ψ1′ (x)ψ2 (x) − ψ1 (x)ψ2′ (x) = 2ikA.

The Green’s function is thus given by

m  ′′ 
G(x′′ , x′ ; E) = 2 θ(x − x′ )ψ1 (x′′ )ψ2 (x′ ) + θ(x′ − x′′ )ψ1 (x′ )ψ2 (x′′ ) .
ikh̄ A
In the limit when x → −∞, x′′ → +∞, we find
′

m ik(x′′ −x′ )
G(x′′ , x′ ; E) =2 Ae , x′ → −∞, x′′ → +∞
ikh̄
which displays the information about the transmission amplitude. In Section 2.7.3, we will use
this formula to compute A and derive the WKB formula for quantum tunneling.

2.7.2 Semi-classical Analysis

Now we perform a semi-classical analysis of the Green’s function

ˆ
′′ ′ 1 ∞
G(x , x ; E) = dT eiET /h̄ K(x′′ , T ; x′ , 0)
ih̄ 0
in the asymptotic limit h̄ → 0. Recall the Feynman kernel
ˆ x(T )=x′′
′′ ′ i
K(x , T ; x , 0) = [Dx(t)] e h̄ S[x(t)] .
x(0)=x′

Combining the above two formulae, we find

ˆ ˆ
′′ ′ 1 ∞ i
G(x , x ; E) = dT [Dx(t)] e h̄ (ET +S[x(t)]) .
ih̄ 0
We consider the semi-classical approximation in the limit h̄ → 0 via the method of steepest
descent. The saddle points are given by (xcl , Ts ) such that

 δ


 δx (ET + S[x(t)]) =0
x(t)=xcl ,T =Ts

 ∂

 (ET + S[x(t)]) =0
∂T x(t)=xcl ,T =Ts

112
 The first equation is the same as
δS
=0
δx x=xcl ,T =Ts

i.e., xcl (t) is a classical trajectory from xcl (0) = x′ to xcl (Ts ) = x′′ . Plugging xcl into the second
equation and using the Hamilton-Jacobi equation, we find
∂S[xcl ]
E=− = E cl .
∂T
Here E cl is the energy of the classical trajectory xcl . Thus the saddle points are the set of pairs
(xcl , Ts ) where xcl is a classical trajectory from x′ to x′′ with energy E and travel time Ts .
Now let us compute the semi-classical contribution at a saddle point (xcl , Ts ). Let S cl denote
the action S[xcl ] on the classical trajectory xcl . Recall that we have the following semi-classical
approximation for the Feynman kernel
  12
′′ ′ i ∂ 2 S cl i
Ksc (x , T ; x , 0) = e h̄ S cl .
2πh̄ ∂x′′ ∂x′
This leads to the following semi-classical asymptotic behavior in the limit h̄ → 0

X   21  1
′′ ′ 1 i ∂ 2 S cl 2πh̄i 2 i
G(x , x ; E) ' e h̄ (ETs +S cl )
ih̄ 2πh̄ ∂x′′ ∂x′ ∂T2 S cl
(xcl ,Ts )
!1
1 X ∂ 2 S cl 2
∂x′′ ∂x′ i
= e h̄ (ETs +S cl ) .
h̄
(xcl ,Ts )
− ∂E ∂T
cl

This expression can be further simplified as follows. Recall from Section 2.6.2

∂ 2 S cl m
′′ ′
=−
∂x ∂x J(T )
where J is the Jacobi field solving
 2 

 d 1 ′′
 + V (xcl (t)) J(t) = 0
dt2 m


J(0) = 0, ˙
J(0) =1

along the trajectory. Let v cl (t) = ẋcl (t) denote the velocity along the trajectory. Differentiating
the trajectory equation
mẍcl (t) + V ′ (xcl (t)) = 0

with respect to t, we find  

d2 1 ′′
+ V (xcl (t)) v cl (t) = 0
dt2 m
i.e., v(t) solves the same equation as J.
Since v(t) is tangent to the trajectory while J(t) is along the normal direction, {v, J} form
two linearly independent solutions of the equation. The Wronskian

cl (t) − J(t)v̇ cl (t)

˙
W = J(t)v

113
is a constant by the equation. Evaluating W at t = 0, we find W = v cl (0). Then
  ˆ T
d J(t) W v cl (0) dt
= 2
= =⇒ J(T ) = v cl (0)v cl (T ) .
dt v cl (t) v cl (t) v cl (t)2 0 v cl (t)2

On the other hand,

ˆ T ˆ T ˆ x′′
dx dx
T = dt = = p
0 0 v cl (t) x′ 2(E cl − V (x))/m
ˆ x′′ ˆ ′′ ˆ
dT 1 dx 1 x dx 1 T dt
=⇒ =− =− =−
dE cl m x′ (2(E cl − V (x))/m)3/2 m x′ v cl 3 m 0 v cl (t)2
∂E cl m
=⇒ = −´ T .
∂T dt
0 v cl (t)2

It follows that
∂ 2 S cl − J(T
m
1 1
∂x′′ ∂x′ )
= = = .
− ∂E ∂T
cl ´T m
dt −v cl (0)v cl (T ) −ẋcl (0)ẋcl (T )
0 v cl (t)2

Let us also define W cl = E cl Ts + S cl . Then

ˆ Ts  
1
W cl = E cl Ts + mẋcl − V (xcl ) dt
2
0 2
ˆ Ts ˆ Ts
= E cl Ts + mẋcl 2 dt − E cl dt
0 0
ˆ Ts ˆ x′′
= mẋcl 2 dt = pdx.
0 x′
p
Here p = mẋ = 2m(E − V (x)) is the classical momentum. Thus we have arrived at the
following asymptotic semi-classical approximation
1 X
(−ẋcl (0)ẋcl (Ts ))− 2 e h̄ W cl
1
G(x′ , x′′ ; E) '
i

h̄
(xcl ,Ts )
1 X ´ ′′
i x
(−ẋcl (0)ẋcl (Ts ))− 2 e h̄ x′ pdx .
1
=
h̄
(xcl ,Ts )

Remark 2.7.4. It is worthwhile to emphasize one important point here. Since we have the
integral contour ˆ ∞
dT
0
for T , we are not summing over all possible classical trajectories (xcl , Ts ). In fact, to apply the
method of steepest descent, we have to deform the integral contour
ˆ ∞ ˆ
dT = dT
0 C

from Re≥0 to contour C which is a combination of curves of steepest descent. As we vary the
parameters x′ , x′′ , the sum of curves of steepest descent may display a discontinuous jump. This
is precisely the Stokes phenomenon.

114
Example 2.7.5 (Free Particle). Given energy E > 0, there exists a unique classical trajectory
from x′ to x′′ by
x′′ − x′
xcl (t) = x′ + t,
Ts
where r
m ′′
Ts = x − x′ .
2E
The semi-classical Green’s function is
1 ´ ′′
i x
(−ẋcl (0)ẋcl (Ts ))− 2 e h̄ x′ pcl dxcl
1
G(x′ , x′′ ; E) '
h̄
  1
1 2E − 2 i √2mE|x′′ −x′ |
= − e h̄
h̄ m
√
m ik|x′′ −x′ | 2mE
= 2 e , k= .
ih̄ k h̄
This is the same formula as we found before.

Example 2.7.6 (Linear Potential). We consider the example of a linear potential

V (x) = −λx, λ > 0.

This example plays an important role in deriving the connection formula for WKB approxima-
tion. We shall understand a different perspective of the connection formula via semi-classical
path integral in Section 2.7.3. We follow the presentation [9] to illustrate the basic idea first.
The equation of motion is
mẍ = λ.

Given T , there is a unique classical trajectory from x(0) = x′ to x(T ) = x′′ by

x′′ − x′ λT λ 2
xcl (t) = x′ + t− t+ t .
T 2m 2m
The corresponding action value is
ˆ T 
1 m ′′ λT ′′ λ2 T 3
S cl = 2
mẋcl + λxcl dt = (x − x′ )2 + (x + x′ ) − .
0 2 2T 2 24m

In the linear case, the semi-classical Feynman kernel is exact

  12 
′′ ′ i ∂ 2 S cl i m  12 i S cl
K(x , T ; x , 0) = e h̄ S cl = e h̄ .
2πh̄ ∂x′′ ∂x′ 2πih̄T

Let us now consider the saddle points along the T -integral for the Green’s function G(x′′ , x′ ; E).
Note that for a linear potential, a shift in energy is equivalent to a translation in x. Thus it is
enough to consider the case E = 0, so we have

W cl = S cl .

We assume E = 0 in the following discussions.

115
 The saddle point (xcl , Ts ) is located at the time Ts by

∂S cl
E=−
∂T T =Ts
m ′′ ′ 2 λ ′′ ′ λ2 Ts2
=⇒ (x − x ) − (x + x ) + =0
2Ts2 2 8m
 1
2m 2 √ ′′ √ ′
=⇒ Ts = (± x ± x ).
λ
There are four saddles in total, corresponding to all possible signs in the choice of square
root of x′ and x′′ . The value of the classical action is
2 1
 √ 3 √ 3
S cl = (2mλ) 2 ± x′′ ± x′
3
√ √
with signs of x′′ and x′ as that for Ts . The product of the initial and final velocities are
 ′′   ′′ 
x − x′ λTs x − x′ λTs
ẋcl (0)ẋcl (Ts ) = − +
Ts 2m Ts 2m
′′
(x − x )′ 2 λ 2
= − T2
2
Ts 4m2 s
λ λ2 2
= (x′′ + x′ ) − T
m 2m2 s
2λ √ √
= − (± x′′ )(± x′ ).
m
Thus the semi-classical contribution of the saddle (xcl , Ts ) to the Green’s function is
 1
1 m 2 i
√ √ e h̄ S cl .
h̄ 2λ(± x′′ )(± x′ )
Now let us identify which saddle points will contribute to the semi-classical Green’s func-
tion. The situation will depend on the locus of x′ and x′′ .

Case 0 < x′ < x′′

Both x′ and x′′ lie in the classically allowed region (x > 0).
V

x′ x′′ x

116
 All the saddle points have real values and we denote them by
 1 
2m 2 √ √ 
Ts(±±) = ± x′ ± x′′ .
λ

T -plane

(−−) (+−) (−+) (++)

Ts Ts Ts Ts

Let us consider curves of steepest descent in the T -plane. The complex oscillatory integral
is about the function (E = 0 here)
i i
e h̄ (ET +S cl ) = e h̄ S cl .

The curves of steepest descent in the T -plane are described by

• Im(iS cl ) = Re(S cl ) = Constant

• Re(iS cl ) = − Im(S cl ) → −∞ along boundary

• Pass through some saddle point.

The classical action S cl as a function of T is

m ′′ λT ′′ λ2 T 3
S cl (T ) = (x − x′ )2 + (x + x′ ) − .
2T 2 24m
(±±)
The corresponding values of S cl on Ts are
  2 1
 √ 3 √ 3
S cl Ts(±±) = (2mλ) 2 ± x′ ± x′′ .
3
  
(±±)
So all Re S cl Ts are different and the corresponding four curves C (±±) of steepest de-
scent do not intersect. They can be constructed by following the flow equation to u → +∞
dT
= −iS cl ′ (T )
du
i.e., " #
dT m ′′ λ λ 2T 2
= i − 2 (x − x′ )2 + (x′′ + x′ ) −
du 2T 2 8m
with the initial condition lim T = saddle point.
u→−∞

117
 Now let us apply the method in Section 2.5.3 to

f (T ) = iS cl (T ).

We have
λ2     
S cl ′ (T ) = − T − T (−−)
s T − T (+−)
s T − T (−+)
s T − T (++)
s .
8mT 2
At the four saddle points,
S cl ′′ > 0 S cl ′′ < 0 S cl ′′ > 0 S cl ′′ < 0

(−−) (+−) 0 (−+) (++) T

Ts Ts Ts Ts

So we can draw the tangent direction of steepest descent curve at each saddle point by

T -plane

(−−) (+−) (−+) (++)

Ts Ts Ts Ts

A bit further calculation shows that the integral contour {T ∈ R≥0 } is deformed into the
sum of the following two curves of steepest descent

(−+) (++)
Ts Ts

Case: 0 < x′ < x′′ .

Thus the semi-classical approximation of the Green’s function G(x′′ , x′ ; E) has contributions
from two saddle points
 1 
2m 2 √ √ 
Ts(−+) = − x′ + x′′
λ

118
  1  
2m 2 √ √
Ts(++) = x′ + x′′ .
λ

corresponds to a direct path from x′ to x′′ . The other saddle Ts

(−+) (++)
The saddle Ts
corresponds to a path that moves left from x′ , reflects at x = 0, then moves right to reach x′′ .

Case x′ < 0 < x′′

x′ lies in the classically forbidden region and x′′ lies in the classically allowed region.
V

x′ x′′ x

The four saddle points

 1 
2m 2 p √ 
Ts(±±) = ±i |x′ | ± x′′
λ

are all complex numbers.

T -plane

(+−) (++)
Ts Ts

(−−) (−+)
Ts Ts

λ2     
S cl ′ (T ) = − T − T (−−)
s T − T (+−)
s T − T (−+)
s T − T (++)
s .
8mT 2

119
 At the saddle points, we have
   −1
iS cl ′′ Ts(++) ∈ Ts(++) R>0
   −1
iS cl ′′ Ts(+−) ∈ − Ts(+−) R>0
   −1
iS cl ′′ Ts(−+) ∈ − Ts(−+) R>0
   −1
iS cl ′′ Ts(−−) ∈ Ts(−−) R>0

This allows us to draw the tangent directions of the steepest descent curves at saddles .

(+−) (++)
Ts Ts

(−−) (−+)
Ts Ts

The integral contour {T ∈ R≥0 } is deformed to the curve of steepest descend passing
 1 
2m 2 p √ 
Ts(−+) = −i |x′ | + x′′ .
λ
(−+)
Note that the saddle Ts has imaginary part. It corresponds to a unique classical path
from x′ to x′′ , but has to go through non-real times! This non-real time is due to the fact that
the path has to go through a portion of the classically forbidden region. Indeed
ˆ T ˆ T ˆ x′′
dx dx
T = dt = = p .
0 0 ẋ x′ 2m(E − V (x))

In the classical forbidden region, E < V (x) hence the integration will give rise to imaginary
contribution of the travel time.

Case x′ < x′′ < 0

Both x′ and x′′ lie in the forbidden region.

120
 V

x′ x′′ x

The four saddle points

 1  
2m 2 p p
Ts(±±) = ±i |x′ | ± i |x′′ |
λ

are all imaginary.

T -plane
(++)
Ts

(+−)
Ts

(−+)
Ts

(−−)
Ts

At the saddle points, we have

 
iS cl ′′ Ts(++) ∈ R>0
 
iS cl ′′ Ts(+−) ∈ R<0
 
iS cl ′′ Ts(−+) ∈ R>0
 
iS cl ′′ Ts(−−) ∈ R<0

This allows us to draw the tangent directions of steepest descent curves at saddles.

121
 (++)
Ts

(+−)
Ts

(−+)
Ts

(−−)
Ts

(−+)
The integral contour {T ∈ R≥0 } is deformed to two curves passing through Ts and
(−−) (−+)
Ts . The first curve goes from the origin, passes the saddle Ts , and then ends up with
(−−) (−−)
the saddle Ts . The second curve starts from the saddle Ts and goes to infinity. The
(−+)
contribution from the saddle Ts dominates.

2.7.3 WKB via Path Integral

Now we connect our discussion on the semi-classical approximation of path integrals to the
WKB formalism on the semi-classical approximation of wave functions.
The Green’s function
1
G(x′′ , x′ ; E) = x′′ x′
^
E−H
^ and satisfies
represents the inverse kernel of the operator E − H
 
^ x′′ G(x′′ , x′ ; E) = δ(x′′ − x′ ).
E−H

^ x′′ is the Hamiltonian operator expressed in the x′′ -coordinate.

Here H
In the previous subsection, we have shown the asymptotic semi-classical formula
1 X
(−ẋcl (0)ẋcl (Ts ))− 2 e h̄ W cl
1
G(x′′ , x′ ; E) '
i

h̄
(xcl ,Ts )

where the sum is over all classical trajectory xcl from x′ at time 0 to x′′ at time Ts , with
prescribed energy E. And

W cl = ETs + S cl = ETs + S[xcl ]

ˆ Ts  
1
= E + mẋcl − V (xcl ) dt
2
0 2
ˆ Ts ˆ x′′
= mẋcl 2 dt = p dx
0 x′

122
where
p
p = mẋ = 2m(E − V (x))

is the classical momentum. Thus semi-classically

 1
X
2
1 1 ´ ′′
 
i x
′′ ′
G(x , x ; E) ' −q q e h̄ x′ p dx .
h̄ 2(E−V (x′ )) 2(E−V (x′′ ))
(xcl ,Ts ) m m

We see that from the perspective of either x′ or x′′ , the semi-classical Green’s function
produces the structure of WKB approximation.

Quantum Tunneling

Let us frist apply the semi-classical Green’s function to the semi-classical computation of
the transmission coefficient for the barrier tunneling of the localized potential V (x).
V (x)

x
a b

As we have seen in Example 2.7.3, the Green’s function has the behavior
m ik(x′′ −x′ )
G(x′′ , x′ ; E) = 2 Ae , x′ → −∞, x′′ → +∞,
ikh̄
where A is the transmission amplitude.
On the other hand, we have a semi-classical asymptotic result
X 1 1
1
2
′′ ′ i
G(x , x ; E) ' e h̄ W cl .
h̄ −ẋ(0)ẋ(Ts )
(xcl ,Ts )

Note that in the limit region x′ → −∞, x′′ → +∞, the particle becomes free with velocity
h̄k
ẋ(0) = ẋ(Ts ) =
m
√
2mE
where k = h̄ .
To have a classical trajectory from x′ to x′′ along a time path from t = 0 to t = T , it
is necessary to go through a region of complex time! This is because both x′ and x′′ lie in
the classical allowed region, but the path has to go through the classically forbidden region in
between. To penetrate the barrier, the time has to be complex. In fact,
r
dx 2(E − V (x))
=
dt m

123
thus in the region E < V (x), t must go along the imaginary direction.
The total travel time for the trajectory is
ˆ Ts ˆ Ts ˆ x′′ r
dx m
Ts = dt = = dx
0 0
dx
dt x′ 2(E − V (x))
ˆ ar ˆ x′′ r ˆ br
m m m
= dx + dx −i dx .
x′ 2(E − V (x)) b 2(E − V (x)) a 2(V (x) − E)
| {z }
imaginary contribution of the complex time

For this classical trajectory in the complex time, we have

ˆ x′′ ˆ ap ˆ bp ˆ x′′p
W cl = p dx = 2m(E − V (x)) dx + i 2m(V (x) − E) dx + 2m(E − V (x)) dx.
x′ x′ a b

Thus the semi-classical Green’s function in the limit region x′ → −∞, x′′ → +∞ is
 √
′′ ′ m h̄i ´xa′ + ´bx′′ 2m(E−V (x)) dx − 1 ´ b √2m(V (x)−E) dx
G(x , x ; E) ' 2 e e h̄ a .
ih̄ k
Comparing with the result
m ik(x′′ −x′ )
G(x′′ , x′ ; E) = 2 Ae , x′ → −∞, x′′ → +∞,
ikh̄
we deduce the semi-classical transmission coefficient
´ √
2 b
T ' |A|2 = e− h̄ a 2m(V (x)−E) dx .

This is precisely the formula calculated from the WKB method in Section 1.10.5.

Connection Formula Revisited

Next we investigate the connection formula of WKB approximation near the turning point
where the potential is approximated by a linear potential, say

V (x) = −λx, λ > 0.

We assume the energy E = 0, so the turning point is x = 0.

Consider the Green’s function G(x′′ , x′ ; E = 0). Let us fix x′′ > 0 and vary x′ from the
region x′ > 0 to the region x′ < 0.
0 < x′ < x′′ x′ < 0 < x′′

V V

x′ x′′ x x′ x′′ x

124
 We would like to keep track of the semi-classical approximation along the deformation of
x′ . However, if we simply change x′ along the real axis, we will soon run into trouble when x′
hits the turning point 0. In fact when x′ = 0, the velocity at the turning point vanishes
r
2(E − V )
=0 at x = 0.
m
So the semi-classical approximation fails and we lose track of the asymptotic information.
Instead, we can consider the analytic continuation in the complex plane to get around the
turning point. For example, we will follow the change

x′ = reiθ , r>0

as θ varies from θ = 0 to θ = π.

x′ -plane

−r r

As we have seen in Example 2.7.6, the semi-classical asymptotic behavior of G(x′′ , x′ ; 0) is

contributed by the curves of steepest descent as in the figure

(−+) (++)
Ts Ts

when 0 < x′ < x′′

125
 (+−) (++)
Ts Ts

(−−) (−+)
Ts Ts

when x′ < 0 < x′′

Let us analyze what happens in between as we vary θ. Recall the four saddle points
 1 
2m 2 √ √ 
Ts(±±) = ± x′ ± x′′
λ
 1 
2m 2 i √ √ 
= ±e 2 θ r ± x′′ 0 < r < x′′ .
λ

The values of S cl at these saddle points are

  2 1
 √ 3 √ 3 2 1
 3 √ √ 3
3
S cl Ts(±±) = (2mλ) 2 ± x′ ± x′′ = (2mλ) 2 ±e 2 iθ r ± x′′ .
3 3
Recall the curves of steepest descent satisfy

• Im(iS cl ) = Re(Scl ) = constant

• Re(iS cl ) = − Im(S cl ) → −∞ along boundary

At the initial stage

0 < x′ < x′′
(−+) (++)
there are two contributing saddles Ts and Ts . Consider
  2 1
 3 √ √ 3
3
Scl Ts(−+) = (2mλ) 2 −e 2 iθ r + x′′
3
  2 1
 3 √ √ 3
3
Scl Ts(++) = (2mλ) 2 e 2 iθ r + x′′ .
3
As we vary θ, the saddle contributions to the asymptotic behavior may jump by Stokes
phenomenon, where one of the curve of steepest descent may hit another saddle point. The
Stokes phenomenon happens when
     
Re Scl Ts(−+) = Re Scl Ts(++)

that is when θ = π3 . We can draw the corresponding four saddles, the tangent directions of the
π
steepest descent at saddles, and the contributing curves of steepest descent at θ = 3

126
 (−+) (++)
Ts Ts

(−−) (−+)
Ts Ts

π (−+)
After we pass θ = 3, only the saddle Ts contributes to the semi-classical Green’s
function. The process can be illustrated by the following picture

(++)
(++) Ts
Ts

(−+) (++)
Ts Ts
(−+)
Ts
(−+)
Ts
π
θ=0 θ= 3 θ=π

This process explains how the connection formula works. At θ = 0, we have two saddle
(−+) (++)
contributions from Ts and Ts with phase factors
(−+) 1 √ 3 √ 3
i i 2
x′ + x′′ )
e h̄ S cl (Ts )
= e h̄ 3 (2mλ) 2 (−
(++) 1 √ 3 √ 3
i i 2
x′ + x′′ )
e h̄ S cl (Ts )
= e h̄ 3 (2mλ) 2 ( .

They contribute equally dominant to the semi-classical asymptotic.

π (−+)
At θ = π, after passing through θ = 3 by a Stokes jump, we have only Ts contributing
to the semi-classical asymptotic with phase factor
(−+) 1√ 3 1√ 3
x′′ −x′
e− h̄ 3 (2mλ) 2
i i 2 1 2
e h̄ S cl (Ts )
= e h̄ 3 (2mλ) 2 .

Similarly, let us consider semi-classical behavior of the Green’s function as we vary x′′

127
 x′ < x′′ < 0 x′ < 0 < x′′

V V

x′ x′′ x x′ x′′ x

(−+)
Ts

(−−)
Ts

when x′ < x′′ < 0 when x′ < 0 < x′′

When x′ < x′′ < 0, we have two saddle contributions from Ts
(−+) (−−)
and Ts with
(−+) 1 √ 3 √ 3
i 1 2
(2mλ) 2 (− −x′ + −x′′ )
e h̄ S cl (Ts )
=e h̄ 3

(−−) 1 √ 3 √ 3
i
S (T ) 1 2
−x′ − −x′′ )
e h̄ cl s = e h̄ 3 (2mλ) 2 (− .
(−+) (−−)
The saddle contribution from Ts dominates that from Ts .
(−−)
Another interesting phenomenon is that the steepest descent curve starting from Ts has
(−−)
only half of the full curve of steepest descent from Ts
(−−)
Ts

1
When we compute the semi-classical contribution, it gives 2 of the usual formula from the
Gaussian integral. Thus we write the saddle contribution in this case as
1
Ts(−+) + Ts(−−) .
2
As we vary x from the region x < 0 to the region x′′ > 0, these two steepest descent curves
′′ ′′

will deform into one steepest descent curve as illustrated.

A careful calculation shows that the above relations between semi-classical asymptotics on
two sides of the turning points precisely give rise to the WKB connection formula (see [9]).

128
 Part II

Geometric Perspectives

129
 Chapter 3 Phase Space Geometry

3.1 Symplectic Geometry of Phase Space

3.1.1 Symplectic Vector Space

Let V be a finite dimensional real vector space. Let

ω :V ×V →R
ω(u, v) = −ω(v, u), ∀u, v ∈ V

be a skew-symmetric bilinear form. We denote the null space of ω by

N = {u ∈ V | ω(u, v) = 0, ∀v ∈ V }.

We call ω non-degenerate or symplectic if its null space is trivial:

ω is non-degenerate ⇐⇒ N = {0}.

Definition 3.1.1. A symplectic vector space is a pair (V, ω) where ω is a non-degenerate skew-
symmetric bilinear form on V . In this case we say ω is a symplectic pairing.

Let c1 , · · · , cm be a basis of V . A skew-symmetric bilinear form ω is represented by a

skew-symmetric matrix in this basis

ωij = ω(ci , cj ), ωij = −ωji .

Then ω is non-degenerate if and only if the matrix (ωij ) is invertible, i.e.,

det(ωij ) 6= 0.

In the non-degenerate case, we can find a canonical representation of ω as follows:

Proposition 3.1.2. Let ω be a symplectic pairing on V . Then there exists a basis

{e1 , · · · , en , f1 , · · · , fn } of V such that

ω(ei , ej ) = ω(fi , fj ) = 0
∀i, j.
ω(e , f ) = δ = −ω(f , e )
i j ij j i

130
 In particular, dim V = 2n has to be even. In such a basis, ω is represented by the matrix
!
0 In
ωn := .
−In 0

Thus any symplectic vector space (V, ω) of dim = 2n is equivalent to the standard symplectic
space (R2n , ωn ) under the choice of a basis as in Proposition 3.1.2.

Definition 3.1.3. Let (V, ω) be a symplectic vector space. The symplectic group Sp(V, ω), or
simply Sp(V ), consists of invertible linear transformations ϕ : V → V such that

ω(ϕ(u), ϕ(v)) = ω(u, v), ∀u, v ∈ V.

When (V, ω) = (R2n , ωn ) is the standard symplectic space, we denote the corresponding
symplectic group by Sp(2n). Explicitly, let us represent ϕ ∈ Sp(2n) by
!
A B
ϕ= ,
C D

where A, B, C, D are n × n matrices. Then the condition of preserving the symplectic pairing is

ϕT ω n ϕ = ω n ,

i.e., ! ! ! !
AT CT 0 In A B 0 In
= .
BT DT −In 0 C D −In 0
That is,

! ! 
 AT C − C T A = 0


AT C − C T A AT D − C T B 0 In
= ⇐⇒ B T D − DT B = 0
BT C − DT A BT D − DT B −In 0 


 T
A D − C T B = In
!
a b
Example 3.1.4. In the case n = 1, we have ∈ Sp(2) if and only if ad − bc = 1. Thus
c d

Sp(2) = SL(2)

Definition 3.1.5. A complex structure on V is a real linear transformation J : V → V such that

J 2 = −1. J is said to be compatible with the symplectic structure ω (or called ω-compatible) if

ω(Ju, Jv) = ω(u, v), ∀u, v ∈ V

and
ω(u, Ju) > 0, ∀u 6= 0.

Let J be ω-compatible and define

gJ (−, −) := ω(−, J(−)).

131
 ⃝
1 gJ is symmetric:

gJ (u, v) = ω(u, Jv) = ω(Ju, J 2 v) = −ω(Ju, v) = ω(v, Ju) = gJ (v, u).

⃝
2 gJ is positive definite. This is by definition.

Thus gJ defines a positive definite inner product on V .

Example 3.1.6. Let {e1 , · · · , en , f1 , · · · , fn } be a basis of V such that


ω(ei , ej ) = ω(fi , fj ) = 0
.
ω(e , f ) = δ = −ω(f , e )
i j ij j i

Consider J : V → V defined in the above basis by

J(ei ) = fi , J(fi ) = −ei .

Then J defines a ω-compatible complex structure


gJ (ei , fj ) = 0
.
g (e , e ) = g (f , f ) = δ
J i j J i j ij

In particular, ω-compatible complex structure always exists.

Let J be a ω-compatible complex structure on V . Then V inherits a C-linear structure by

(a + ib) · u := au + bJ(u), ∀a + ib ∈ C.

There is an induced Hermitian structure on V

h−, −i : V × V −→ C

defined by
h−, −i := gJ (−, −) + iω(−, −).

⃝
1 Hermitian: for any u, v ∈ V ,

hJu, vi = gJ (Ju, v) + iω(Ju, v) = ω(Ju, Jv) − iω(u, Jv) = ω(u, v) − igJ (u, v) = −i hu, vi

hu, Jvi = gJ (u, Jv) + iω(u, Jv) = ω(u, J 2 v) + iω(u, Jv) = −ω(u, v) + igJ (u, v) = i hu, vi

⃝
2 Positivity: for any u ∈ V , u 6= 0,

hu, ui = gJ (u, u) > 0.

Given (V, ω, J), let

Sp(V ) = {ϕ ∈ GL(V ) | ω(ϕ(−), ϕ(−) = ω(−, −)}

GL(V, J) = {ϕ ∈ GL(V ) | ϕ(J(−)) = J(ϕ(−))}
O(V ) = {ϕ ∈ GL(V ) | gJ (ϕ(−), ϕ(−)) = gJ (−, −)}
U(V ) = {ϕ ∈ GL(V ) | hϕ(−), ϕ(−)i = h−, −i}

132
In terms of the standard symplectic space and the constructed J above, the above groups are

Sp(2n), GLn (C), O(2n), U(n).

The relation
gJ (−, −) = ω(−, J(−))

implies that if ϕ ∈ GL(V ) preserves any two of {ω, J, gJ }, then ϕ preserves the third and
therefore preserves the Hermitian structure. Thus we have proved

Proposition 3.1.7.
Sp(V ) ∩ GL(V, J) = U(V )
Sp(V ) ∩ O(V ) = U(V )
GL(V, J) ∩ O(V ) = U(V )

Example 3.1.8. Sp(2n) ∩ O(2n) = U(n) can be explicitly realized as follows. Let

A + iB ∈ U(n), A, B are real n × n matrices.

The unitary condition says


 AT A + B T B = I n
(AT − iB T )(A + iB) = In ⇐⇒ .
 AT B − B T A = 0

This implies !
A −B
∈ Sp(2n) ∩ O(2n).
B A
Thus Sp(2n) ∩ O(2n) = U(n) is realized explicitly by the map
!
A −B
A + iB 7−→ .
B A

Let
J = {ω-compatible complex structure}
P = {symmetric positive definite inner product}
The map J 7→ gJ defines
j : J ,−−→ P.

On the other hand, given g ∈ P, define K : V → V by the relation

ω(u, v) = g(Ku, v).

K may not lie in J. Nevertheless, K is invertible and skew-symmetric with respect to g

g(Ku, v) = ω(u, v) = −ω(v, u) = −g(Kv, u) = −g(u, Kv).

133
 Let K = RJ be the polar decomposition where R is symmetric positive definite and J is
orthogonal with respect to g. Since K is skew-adjoint (hence normal), RJ = JR. Then

K T = −K, RT = R =⇒ J T = −J =⇒ J 2 = −J T J = −1

i.e., J defines a complex structure on V . Moreover,

ω(Ju, Jv) = g(KJu, Jv) = g(JKu, Jv) = g(Ku, v) = ω(u, v)

ω(u, Ju) = g(Ku, Ju) = g(R(Ju), Ju) > 0 if u 6= 0

i.e., J is ω-compatible. Thus the map

r : P −→ J
g 7−→ J

defines a retraction of j : J → P.

Proposition 3.1.9. The space J of ω-compatible complex structures on (V, ω) is contractible.

Proof: We only need to show P is contractible. Given any g0 ∈ P, we define a homotopy

ϕt : P −→ P 0≤t≤1
g 7−→ tg + (1 − t)g0

Since ϕ1 = identity and ϕ0 = const map to g0 , this shows the contractibility of P.

Example 3.1.10. Consider (R2 , ω1 ). Let J be a ω1 -compatible complex structure. Preserving

ω1 says that J ∈ SL(2). J 2 = −1 implies further that J can written as
!
a −b
J = 1+a2 a ∈ R, b ∈ R − {0}
b −a

Positivity of ω1 (−, J(−)) says

! ! !
1+a2
0 1 a −b b −a
1+a2
=
−1 0 b −a −a b

is a symmetric positive definite matrix, that is b > 0. Therefore

( ! )
a −b
J= 1+a2
a ∈ R, b ∈ R>0
b −a

3.1.2 Lagrangian Grassmannian

Definition 3.1.11. Let (V, ω) be a symplectic vector space. A linear subspace L ⊂ V is called
isotropic if ω|L×L = 0, i.e.,
ω(u, v) = 0, ∀u, v ∈ L.

Proposition 3.1.12. Let L be an isotropic subspace of (V, ω). Then dim L ≤ 1

2 dim V.

134
Proof: The symplectic pairing ω induces a linear map

ω̃ : V −→ V ∗
v 7−→ ω(v, −)

which is an isomorphism by the non-degeneracy of ω. Let L ⊂ V be an isotropic subspace, then

ω̃ induces
ω̃L : L −→ (V /L)∗
u 7−→ ω(u, −)
Observe ω̃L is injective: if u ∈ L and ω̃L (u) = 0, then

ω(u, v) = 0 ∀v ∈ V =⇒ u = 0.

Therefore dim L ≤ dim (V /L), i.e.,

1
dim L ≤ dim V.
2

Definition 3.1.13. A linear subspace L of a symplectic vector space (V, ω) is called a La-
1
grangian subspace if L is isotropic and dim L = 2 dim V.

Example 3.1.14. Let {e1 , · · · , en , f1 , · · · , fn } be a basis of V such that


ω(ei , ej ) = ω(fi , fj ) = 0
ω(e , f ) = δ
i j ij

Then Span {e1 , · · · , en } and Span {f1 , · · · , fn } are Lagrangian subspaces of (V, ω).

Definition 3.1.15. Let N ⊂ V be a linear subspace. Define its ω-orthogonal complement by

N ⊥ := {u ∈ V | ω(u, n) = 0, ∀n ∈ N } .

It is clear that
N ⊂ V isotropic ⇐⇒ N ⊂ N ⊥.

Since ω is non-degenerate, we have

dim N ⊥ = dim V − dim N.

It follows by dimension counting that

L is Lagrangian subspace ⇐⇒ L = L⊥ .

The collection of all Lagrangian subspaces of (V, ω) will be denoted by L (V ), called the
Lagrangian Grassmannian of V . The symplectic group Sp(V ) acts naturally on L (V ):

Sp(V ) × L (V ) −→ L (V )
ϕ × L 7→ ϕ(L)

135
 Let J be a ω-compatible complex structure. It induces a Hermitian structure on V and

U(V ) = Sp(V ) ∩ O(V ).

Thus U(V ) acts on L (V ):

U(V ) × L (V ) −→ L (V ).

Let g(−, −) = ω(−, J(−)) be the corresponding symmetric positive-definite inner product.
Let L ∈ L (V ) be a Lagrangian subspace. Denote

L∗ := {u ∈ V | g(u, v) = 0, ∀v ∈ L}

to be the g-orthogonal complement of L. Then

V = L ⊕ L∗ .

By definition, we also have L∗ = J(L)⊥ . Since J is ω-compatible, J(L) is also a Lagrangian

subspace, thus J(L)⊥ = J(L). Therefore we have the g-orthogonal decomposition

V = L ⊕ J(L)

of V into a direct sum of two Lagrangian subspaces.

Proposition 3.1.16. U(V ) acts transitively on L (V ). Its stabilizer at L ∈ L (V ) is O(L), the

orthogonal group of L. Thus we can identify L (V ) as the homogeneous space

L (V ) = U(V )/ O(L).

Proof: For all L1 , L2 ∈ L (V ), let

ϕ : L1 −→ L2

be a g-orthogonal transformation. Then ϕ induces

φ⊕J◦φ◦J −1
ϕ̃ : L1 ⊕ JL1 L2 ⊕ JL2
==

==

V V

By construction, ϕ̃ is compatible with J and g, so

ϕ̃ ∈ U(V ) and ϕ̃(L1 ) = L2 .

This proves the transitivity of the U(V )-action.

Let ϕ̃ ∈ U(V ) and ϕ̃(L) = L. Then ϕ = ϕ̃|L ∈ O(L). Conversely, any ϕ ∈ O(L) leads to a
stabilizer ϕ̃ of L as above.

136
3.1.3 Maslov Index

Let (V, ω) be a 2n-dimensional symplectic vector space, with a ω-compatible complex struc-
ture J and a Lagrangian subspace L. By Proposition 3.1.16, we can identify

L (V ) ' U(V )/ O(L) ' U(n)/ O(n).

Let “det” be the determinant mapping

det : U(V ) −→ S 1 = {z ∈ C| |z| = 1}.

Restricting to the subgroup O(n), we have

det : O(n) −→ {±1} .

This leads to a well-defined map

det2 : L (V ) −→ S 1 .

−1
Let F = det2 (1) be the fiber over 1 ∈ S 1 . If [u] ∈ F is represented by u ∈ U(V ) with

det U = −1

then we can find O ∈ O(L) with det O = −1 such that

[u] = [uO] ∈ L (V ) and det(uO) = 1.

This implies that SU(V ) acts transitively on F with stabilizer at L by SO(L). Thus

F = SU(V )/ SO(L)

and we have a fibration

SU(n)/ SO(n) L (V )
det2

S1
SU(n) is simply connected. To see this, consider SU(n) acting on Cn by standard matrix
multiplication. It preserves the unit sphere S 2n−1 where SU(n) acts transitively. The stabilizer
is SU(n − 1). This gives a fibration (n ≥ 3)

SU(n − 1) SU(n)

S 2n−1

The homotopy exact sequence of this fibration gives

π1 (SU(n)) ' π1 (SU(n − 1)) ' · · · ' π1 (SU(2)) = π1 (S 3 ) = 1.

137
 Since SO(n) is connected, the homotopy exact sequence of the fibration

SO(n) SU(n)

SU(n)/ SO(n)

leads to
π1 (SU(n)/ SO(n)) = 1.

Back to the fibration

SU(n)/ SO(n) L (V )
det2

S1
we find that det2 induces an isomorphism
det2
π1 (L (V )) −−→ π1 (S 1 ) = Z.

This isomorphism does not depend on the choice of complex structure J (the collection of all
ω-compatible J’s is connected, actually contractible by Proposition 3.1.9) and does not depend
on the choice of the reference Lagrangian L (since U(n) is connected). Thus for any loop

γ : S 1 −→ L (V )

we can associate an integer m(γ) ∈ Z representing [γ] ∈ π1 (L (V )) ' Z. This integer is called
the Maslov index. Explicitly, m(γ) is the winding number of

det2 ◦ γ : S 1 −→ S 1 .

Example 3.1.17. Consider (V = R2 , ω). Any line L in R2 is a Lagrangian subspace. Thus

L (R2 ) = RP 1 = U(1)/ O(1) = S 1 /Z2 .

Let
γ : S 1 −→ L (R2 )
θ 7−→ eiθ
which winds around the unit circle.
R2

138
The composition det2 ◦ γ is
det2 ◦ γ : S 1 −→ S 1
eiθ 7−→ e2iθ
Thus the Maslov index of γ is
m(γ) = 2.

3.1.4 Symplectic Manifold

Definition 3.1.18. A symplectic manifold is a pair (M, ω) where M is a smooth manifold and
ω is a smooth 2-form on M such that

⃝
1 ω is closed: dω = 0

⃝
2 ω is non-degenerate: ∀p ∈ M ,

ω|p : Tp M × Tp M −→ R

defines a symplectic pairing on Tp M .


In local coordinates xi , ω can be written as

1X
ω= ωij (x)dxi ∧ dxj
2
i,j

where ωij (x) = −ωji (x). Condition dω = 0 becomes

∂i ωjk + ∂j ωki + ∂k ωij = 0, ∀i, j, k.

The non-degeneracy condition says

det (ωij (x)) 6= 0, ∀x.

Definition 3.1.19. Given a smooth function f on M , we define its associated Hamiltonian

vector field Vf ∈ Vect(M ) by the equation

ιVf ω = df.

Explicitly in local coordinates,

X ∂
Vf = ω ij ∂i f
∂xj
i,j

where ω ij is the inverse matrix of {ωij },i.e., satisfies
X
ω ik ωkj = δji .
k

In fact, X X X
ω ij ∂i f ι∂j ω = ω ij ∂i f ωjk dxk = ∂i f dxi = df.
i,j i,j,k i

139
Definition 3.1.20. The Poisson bracket of two smooth functions f, g on M is defined to be

{f, g} := ιVf ιVg ω.

Using ιVg ω = dg, this can be equivalently written as

{f, g} = ιVf (dg) = Vf (g).

In local coordinates,
X
{f, g} = ω ij ∂i f ∂j g.
i,j

Proposition 3.1.21. Hamiltonian vector fields preserve the symplectic form ω, i.e., LVf ω = 0.

Proof: Using Cartan’s formula, we have

LVf ω = dιVf + ιVf d ω = d(df ) + 0 = 0.

The symplectic form ω induces a volume form on M by

1 n
ω , n = dim M.
n!
As a corollary, Hamiltonian vector fields preserve the volume density (Liouville’s Theorem)
 
1 n
LVf ω = 0.
n!
Proposition 3.1.22. V{f,g} = [Vf , Vg ].

Proof: Using Cartan’s formula

ιV{f,g} ω = d {f, g} = dιVf ιVg ω = LVf ιVg ω − ιVf dιVg ω = LVf ιVg − ιVg LVf ω − ιVf ddg = ι[Vf ,Vg ] ω.

By the non-degeneracy of ω,
V{f,g} = [Vf , Vg ].

Proposition 3.1.23. (C ∞ (M ), {−, −}) defines a Lie algebra. The map

C ∞ (M ) −→ Vect(M )
f 7−→ Vf
is a Lie algebra homomorphism.

Proof: {−, −} is clearly skew-symmetric. We need to check Jacobi-identity:

{f, {g, h}} + {g, {h, f }} + {h, {f, g}} = 0.

In fact, using Proposition 3.1.22 and skew-symmetry of {−, −}

{{f, g} , h} = V{f,g} (h) = [Vf , Vg ](h) = Vf ({g, h}) − Vg ({f, h})

= {f, {g, h}} − {g, {f, h}} = − {{g, h} , f } − {{h, f } , g} .
The Lie algebra morphism follows from Proposition 3.1.22.

140
Example 3.1.24 (Cotangent Bundle). For any smooth manifold X, the total space of its
cotangent bundle M = T ∗ X is canonically a symplectic manifold.
 n o
Let q i be local coordinates on X. The frame ∂q∂ i defines naturally linear coordinates

{pi } along fibers of T ∗ X. Thus q i , pi form local coordinates on T ∗ X. Consider the 1-form
X
λ= pi dq i .
i

It is easy to check that this expression is independent of the choice of local coordinates and λ is
a globally defined 1-form on T ∗ X. λ is called the Liouville 1-form. The 2-form
X
ω = dλ = dpi ∧ dq i
i

defines a symplectic form on T ∗ X.

Definition 3.1.25. Let (M, ω) be a symplectic manifold. A submanifold L ⊂ M is called a

1
Lagrangian submanifold if dim L = 2 dim M and ω|L = 0.

This is equivalent to say that for all p ∈ L, the tangent space Tp L is a linear Lagrangian
subspace of (Tp M, ω|p ).
Let M = T ∗ X and ω be the canonical symplectic form. A submanifold L ⊂ M of dim L =
1
2 dim M is a Lagrangian submanifold if

ω|L = dλ|L = 0,

i.e., the restriction λ|L of the Liouville 1-form on L is a closed 1-form.

Let us consider a special case. Let S : X → T ∗ X be a section of the cotangent bundle.
Such a section is the same as specifying a 1-form α on X, and we denote it by

Sα : X −→ T ∗ X.

The image of Sα defines a submanifold

Lα := Sα (X) ⊂ T ∗ X.

One nice property of Sα is that we have tautologically

Sα∗ (λ) = α.

This immediately leads to the following

Proposition 3.1.26. Lα ⊂ T ∗ X is a Lagrangian submanifold if and only if α is a closed

1-form: dα = 0.

141
 T ∗X

Lα

Definition 3.1.27. A Lagrangian submanifold of the form Lα ⊂ T ∗ X associated to a closed 1-

form α is called a projectable Lagrangian submanifold of T ∗ X. Lα is called an exact Lagrangian
if α = df is an exact 1-form.

Example 3.1.28. For an exact Lagrangian Ldf , it is described by the equations

∂
pi = f (q).
∂q i
In this case, f is called a phase function of L. Note that the phase function is only determined
up to a shift by a constant.

Example 3.1.29. R2 = T ∗ R, ω = dp ∧ dq. Then any curve in R2 defines a Lagrangian

submanifold. Consider

L = p2 + q 2 = R2 ⊂ T ∗ R.
p

The restriction to L of the Liouville 1-form λ = pdq is closed but not exact since
ˆ
λ = ±πR2
L

where the sign ± depends on the orientation of L.

142
Definition 3.1.30. A diffeomorphism ϕ : M → M is called a symplectomorphism if ϕ preserves
the symplectic structure, i.e.,
ϕ∗ ω = ω.

Proposition 3.1.31. Let ϕ : M → M be a symplectomorphism and L ⊂ M be a Lagrangian

submanifold. Then ϕ(L) is also a Lagrangian submanifold.

Example 3.1.32. The flow φt : M → M generated by a Hamiltonian vector field Vf is a

symplectomorphism. In fact,
d ∗

(φt ω) = φ∗t LVf ω = 0 =⇒ φ∗t ω = ω, ∀t.
dt
Let us now specialize this example to the case

M = T ∗X and f : X → R.

Viewed as a function on M , the corresponding Hamiltonian vector field of f is

X ∂
Vf = ∂q i f .
∂pi
i

The symplectomorphism φt : T ∗ X → T ∗ X generated by Vf translates the fiber by tdf , thus

φ1 (Lα ) = Lα+df .

In particular, φ1 transforms the zero section to

φ1 (L0 ) = Ldf .

3.2 Semi-classical Quantization

3.2.1 Semi-classical Solution

We consider quantum particle moving in Rn in the potential V (x). The quantum Hamil-
tonian operator is
2
^ = − h̄ ∇2 + V (x)
H
2m
and the Schrödinger equation takes the form
∂ ^ ψ.
ih̄ ψ=H
∂t
We consider solutions of the form (called stationary states)

Ψ(x, t) = e−iEt/h̄ ψ(x)

Plugging into the Schrödinger equation, we find

^ ψ(x) = Eψ(x)
H (∗)

143
 ^ with eigenvalue E, which is interpreted as the energy. Equation
Thus ψ(x) is an eigenstate of H
(∗) is called the time-independent Schrödinger equation.
Let us consider a solution of (∗) by the form
i
ψ = e h̄ S(x) a(x, h̄).

S(x) is called the phase function. a(x, h̄) is the amplitude. Observe
  
h̄2 2 i
− ∇ + (V − E) ae h̄ S
2m
" ! #
(∇S)2 h̄2 2 ih̄
iS
= a + (V − E) − ∇ a− a∇ S + 2∇S · ∇a e h̄ .
2
2m 2m 2m

The idea of WKB method is to look for h̄-asymptotic solutions of the form
i
ψ = e h̄ S(x) (a0 (x) + a1 (x)h̄ + · · ·)

i.e., a(x, h̄) has the h̄-asymptotic series

∞
X
a(x, h̄) ∼ ak (x)h̄k , a0 (x) 6= 0.
k=0

Plugging this WKB ansatz into the above equation, we find



 (∇S)2

 +V =E

 2m

 a0 ∇ S + 2∇S · ∇a0 = 0
2



 ak ∇2 S + 2∇S · ∇ak = i∇2 ak−1 , k≥1

The approximate solution

i
e h̄ S(x) a0 (x)

is called the semi-classical approximation.

The leading order equation for S can be written as

H (x, p = ∇S) = E (∗∗)

p2
Here H (x, p) = 2m + V (x) is the classical Hamiltonian function. Equation (∗∗) is precisely the
Hamilton-Jacobi equation.
Geometrically, S defines an exact projectable Lagrangian submanifold

LdS ⊂ T ∗ Rn

and the Hamilton-Jacobi equation says that H is constant (= E) on LdS , i.e.

dH |LdS = 0.

In terms of the Hamiltonian vector field VH , this is

ιVH ω|LdS = 0

144
i.e.,
ω (VH (p), Tp LdS ) = 0, ∀p ∈ LdS .

Since LdS is a Lagrangian submanifold, this is equivalent to

VH (p) ∈ Tp LdS , ∀p ∈ LdS ,

i.e., the Hamiltonian vector field VH is tangent to LdS .

T ∗ Rn

VH

LdS

Thus the Hamilton-Jacobi equation leads to an exact projectable Lagrangian submanifold

LdS with phase function S such that the Hamiltonian vector field VH is tangent to LdS at every
point of LdS .
Now we consider the subleading order equation

a0 ∇2 S + 2∇S · ∇a0 = 0

which is called the homogeneous transport equation. It can be written as

∇ · a20 ∇S = 0.

This allows us to interpret it as

L∇S a20 |dn x| = 0.

145
Here |dn x| = |dx1 ∧ · · · ∧ dxn | is the canonical volume density on Rn . L∇S is the Lie derivative
with respect to the vector field
X ∂
∂i S .
∂xi
i

Geometrically, let π denote the diffeomorphism

π: LdS −→ Rn
(xi , pi = ∂i S) 7−→ (xi )

The Hamiltonian vector field VH is

X  ∂H ∂ ∂H ∂
 X
∂ 1 ∂

VH = − = ∂i V − pi .
∂xi ∂pi ∂pi ∂xi ∂pi m ∂xi
i i

Observe
1 X ∂
π∗ (VH |LdS ) = − ∂i S .
m ∂xi
i

Thus the homogeneous transport equation becomes

Lπ∗ (VH |L ) a20 |dn x| = 0.
dS

Since π : LdS → Rn is a diffeomorphism, we can equivalently describe this on LdS as

LVH |L π ∗ a20 |dn x| = 0.
dS

3.2.2 Half-Density

Recall that the density line bundle DensM on a manifold M is described by sections which

in local coordinates xi are of the form

ρx (x) |dn x|, ρx (x) smooth function.

 
Under coordinate transformation xi → y i , ρx and ρy in two coordinates are related by
 
∂y
ρx = ρy det .
∂x
   i
∂y ∂y
Here ∂x = ∂x j is the Jacobian matrix. In other words, the transition function of the
 
∂y
density line bundle is det ∂x . Invariantly, we can write it as

ρx |dn x| = ρy |dn y|.

This property allows us to integrate densities on manifold

ˆ
: Γ(M, DensM ) −→ R
M ˆ
ρ 7−→ ρ
M

146
Remark 3.2.1. If M is orientable, we can identify
^n
DensM ' T ∗ M, n = dim M

via a choice of orientation. Thus a section of DensM becomes an n-form. When M is unori-
entable, these two bundles are different.
Similarly, for any α > 0, we can define the α-density line bundle DensαM whose transition
function is given by   α
∂y
det .
∂x
Invariantly, we can write an α-density locally by

ρx |dn x|α .
 
The transition function says that under coordinate transformation xi → y i , we have

ρx |dn x|α = ρy |dn y|α .

Back to the semi-classical solution, it is better to identify a0 (x) as a half-density

a0 |dn x|1/2 .


Under the diffeomorphism π : LdS → Rn , it gives a half-density π ∗ a0 |dn x|1/2 on LdS . Then
the homogeneous transport equation can be written as
  
LVH |L π ∗ a0 |dn x|1/2 = 0.
dS

In summary, we have arrived at the following description: a semi-classical solution of the

time-independent Schrödinger equation is equivalent to the following geometric data:

⃝
1 an exact projectable Lagrangian submanifold L which lies in a level set of the Hamiltonian
function H (equivalently, VH is tangent to L).

⃝
2 a half-density Ω on L which is VH -invariant.

Given such data, we can write S for the phase function of L (defined up to a constant). Let

s : Rn −→ L
(xi ) 7−→ (xi , pi = ∂i S)
denote the diffeomorphism by the section map. Let

s∗ Ω = a0 (x) |dn x|1/2

be the corresponding half-density expressed on Rn . Then

i
e h̄ S(x) a0 (x)

gives a semi-classical solution.

There is an immediate advantage of such geometric description: L can be a general La-
grangian submanifold instead of an exact projectable one. For such L in a level set of H ,
the notion of VH -invariant half-density is still well-defined. This leads to the geometric semi-
classical state that we will discuss next.

147
3.2.3 Maslov Correction

Now we consider quantization on the Lagrangian submanifold L ⊂ T ∗ Rn which is not

necessarily projectable. The points of L where

π : L −→ Rn

fail to become local diffeomorphisms are called caustics.

Example 3.2.2. Consider the one dimensional Harmonic oscillator

p2 1
H = + kx2 .
2m 2
The level set H = E > 0 is a Lagrangian submanifold

L = {(x, p) | H (x, p) = E} .
p

L = {H = E}

u− u+ x

There are two caustics of L by r !

2E
u± = ± ,0 .
k

Caustics are not singularities of L. They actually reflect the way we choose the coordinate
x from the configuration space. In fact, if we consider the projection to the p-coordinate

πP : L −→ R
(x, p) 7−→ p

then πp is a local diffeomorphism around u± . It suggests that we could use appropriate co-
ordinates locally around the Lagrangian to quantize and then glue. This is precisely Maslov’s
technique.
Motivated by the projectable case, we define a geometric semi-classical state as follows.

Definition 3.2.3. A geometric semi-classical stationary state is a pair (L, Ω) where

1 L ⊂ T ∗ X is a Lagrangian submanifold which lies in a level set of the Hamiltonian H ,
⃝
⃝
2 Ω is a half-density on L which is invariant under the Hamiltonian vector field VH .

148
 Let us next explore how to implement the analogue of the semi-classical solution
i
e h̄ S a |dn x|1/2

on a geometric semi-classical state (L, Ω).

When L = LdS is exact projectable, the phase function S (viewed as a function of L) is
ˆ
S= λ

P
up to a constant. Here λ = pi dxi is the Liouville 1-form. This is the same as saying that
i

λ|L = dS.

For a general Lagrangian L, we have

λ|L is closed

but may not be exact. Nevertheless, we still wish to define the phase factor by
i
´z
λ
e h̄ z0 for z ∈ L.
´z
Here z0 ∈ L is some reference base point. The integral z0 λ is multi-valued and depends on the
integration path from z0 to z. Two paths from z0 to z will differ by a closed cycle γ, and the
´z
corresponding integrations z0 λ differs by
˛
λ.
γ
¸
Note that since λ is closed on L, the cycle integration γ λ only depends on the topological class
of γ in H1 (L).

z0

¸ i
´z
λ
γ λ ∈ 2πZ for all γ ∈ H1 (L), then e
1
Thus if h̄
h̄ z0 would be a well-defined function on L.
This suggests the following integrality condition (Bohr-Sommerfeld quantization condition)
˛
1
λ∈Z
2πh̄ γ

for all closed 1-cycle γ on L. This would be suffice when L is projectable. However, when
caustics appear, there is a further modification called the Maslov correction.

149
 The issue comes from the half-density piece of the semi-classical solution. In the projectable
case, L can be identified with Rn under the projection, and we can take the square root |dn x|1/2
uniformly on Rn and write
π∗ Ω = a(x) |dn x|1/2 .
i
Then e h̄ S a leads to the semi-classical solution. In general, we need to understand how to take
square root of a volume density consistently on L.
Let us assume L is oriented and we identify the density bundle with
^n
DensL ' T ∗L
Vn
which is the dual of T∗ L. At each point z ∈ L, the tangent plane

Tz L ⊂ Tz (T ∗ Rn ) = R2n

defines a Lagrangian linear subspace, i.e.,

Tz L ∈ L (R2n )

defines an element of the Lagrangian Grassmannian.

Recall we can identify
L (R2n ) ' U(n)/ O(n)

together with
det2
L (R2n ) −−→ S 1

which induces an isomorphism

π1 (L (Rn )) ' π1 (S 1 ).

Explicitly, let us choose the standard complex structure to identify

J
R2n ' Cn = Rn ⊕ JRn ,

where Rn = Span {e1 , · · · , en } and JRn = Span {f1 , · · · , fn }. Then {e1 , · · · , en } defines a C-
basis of Cn = R2n . For a Lagrangian subspace L0 , we choose an orthonormal basis {c1 , · · · , cn }
of L0 . Then  2
c1 ∧C · · · ∧ C cn
2
det (L0 ) = ∈ S1.
e1 ∧C · · · ∧ C en
Here ∧C is the C-linear wedge product on Cn . This allows us to define a map
 ^n 2
ϕ: L∨
0 − {0} −→ C∗

(α)2 7−→ re−iθ

where
r = (α(c1 , · · · , cn ))2 , eiθ = det2 (L0 ).

This map ϕ does not depend on the choice of the basis {c1 , · · · , cn } of L0 .

150
 Let us apply this construction to the Lagrangian submanifold L ⊂ T ∗ Rn . We obtain a map
 ^n 2
ϕ: T ∗ L − {zero section} −→ C∗ .

The map ϕ can be viewed as describing the geometric change of 2-densities on L relative to a
fixed reference 2-density on Rn . Let us write the S 1 -phase part of ϕ by
ϕ  ^n ∗  2
Θ= : T L − {zero section} −→ S 1 .
|ϕ|
Θ represents the geometric phase change for 2-densities along L.
Let us now consider the half-density Ω on L. To describe a semi-classical solution, we
need to compare Ω with the standard half-density on the configuration space Rn . To go around
caustics where a direct comparison with respect to the Jacobian blows up, we will use the
geometric comparison ϕ as above. By construction, we can think about ϕ as a continuous
parametrization of a 2-density relative to a reference one at a fixed point. For the half-density
Ω on L, Ω4 is a 2-density. Thus we can parametrize the half-density Ω by

ϕ(Ω4 )1/4 = Θ1/4 |ϕ(Ω4 )|1/4 .

However, Θ1/4 is multi-valued on L. Θ1/4 changes by

e− 4 (2πm(γ))
i

when it goes around a closed loop γ in L. Here m(γ) is the Maslov index.
Let us now combine the half-density with the phase function and consider the geometric
semi-classical wave
i
´z
e h̄ λ
Θ1/4 |ϕ(Ω4 )|1/4 .

Its multi-valueness is now reflected by the change along a closed 1-cycle γ via the phase factor
i
¸
λ− 4i (2πm(γ))
e h̄ γ .

Therefore the above geometric semi-classical wave is globally defined on L if and only if
˛
1 1
λ − m(γ) ∈ Z
2πh̄ γ 4

for all closed 1-cycle γ in L. This rule is the Keller-Maslov quantization condition. The factor
1
4 m(γ) is the Maslov correction. In physics literature, this is known as Einstein-Brillouin-Keller
(EBK) semi-classical quantization condition which improves the Bohr-Sommerfeld rule via the
Maslov index.

Example 3.2.4. Consider the one dimensional harmonic oscillator

p2 1
H = + kx2 .
2m 2
The level set
C = {H (x, p) = E}

151
is a Lagrangian submanifold. We have
˛ ˛ r
1 1 E k
λ= pdx = , ω= , m(C) = 2.
2π C 2π C ω m
The Keller-Maslov quantization condition reads
E 1
− ∈ Z.
h̄ω 2
This is compatible with the well-known spectrum
 
1
E = h̄ω n + , n = 0, 1, 2, · · · .
2

3.3 Heisenberg Group

3.3.1 Heisenberg Lie Algebra

Let (V, ω) be a symplectic vector space of dimension 2n. Consider the vector space of
dimension 2n + 1
hV := V ⊕ RK

where we use K for the basis of the new dimension. We can define a bracket on hV by

[u, v] = ω(u, v)K if u, v ∈ V

[−, K] = 0 i.e., K is a center element
It is easy to check that (hV , [−, −]) forms a Lie algebra. This is called the Heisenberg Lie
algebra. Explicitly, let {e1 , · · · , en , f1 , · · · , fn } be a symplectic basis of (V, ω). Then

[ei , fj ] = δij K
[ei , ej ] = [fi , fj ] = 0
[ei , K] = [fi , K] = 0
give the explicit bracket relations in the Heisenberg Lie algebra.
When V = R2n is the standard symplectic space, we denote

hn := hR2n .

Since any symplectic vector space (V, ω) of dimension 2n is symplectomorphic to the standard
R2n , we have a Lie algebra isomorphism

hV ' hn .

Let ϕ ∈ Sp(V ) be an element of the symplectic group. Since ϕ preserves the symplectic
pairing, it induces a Lie algebra isomorphism

ϕ: hV −→ hV
u + λK 7−→ ϕ(u) + λK
This induces a group homomorphism

Sp(V ) −→ Aut(hV ).

152
3.3.2 Heisenberg Group

The Heisenberg Group HV is the simply connected Lie group associated to the Lie algebra
hV . As a set, HV can be identified with hV via the exponential map
n o
HV = eu+λK .
u+λK∈hV

The multiplication law can be obtained via the Baker-Campbell-Hausdorff formula

1
eA eB = eA+B+ 2 [A,B]+higher brackets .

Since [hV , hV ] = RK is the center, all higher brackets vanish here. Thus
1
eu1 +λ1 K · eu2 +λ2 K = eu1 +u2 +(λ1 +λ2 + 2 ω(u1 ,u2 ))K u1 , u 2 ∈ V

When V = R2n is the standard symplectic space, we denote

Hn := HR2n .

For general symplectic vector space V of dimension 2n, a choice of symplectic basis gives a
group isomorphism
HV ' Hn .

There is also a group homomorphism

Sp(V ) −→ Aut(HV ).

Given ϕ ∈ Sp(V ), the induced group isomorphism is

HV −→ HV
eu+λK 7−→ eφ(u)+λK
This is compatible with the corresponding transformation on Heisenberg Lie algebra.

3.3.3 Schrödinger Representation

Let us consider V = R2n with standard symplectic basis {e1 , · · · , en , f1 , · · · , fn }. Let

S(Rn ) = {f (x) | f is a C-valued Schwartz function on Rn } .

S(Rn ) defines a representation of the Heisenberg Lie algebra hn by

i i i
ei 7−→
x̂ = xi
h̄ h̄
i ∂
fi 7−→ − p̂i = − i
h̄ ∂x
i
K 7−→
h̄
This can be exponentiated to a unitary representation of the Heisenberg group Hn . In fact,
let us identify vectors in hn as
r · e + s · f + λK.

Here r = (r1 , · · · rn ) ∈ Rn , s = (s1 , · · · sn ) ∈ Rn and

153
 e = (e1 , · · · en ), f = (f1 , · · · fn )

is the above symplectic basis.

Given a function ψ(x), we would like to define the transformed function

er·e+s·f+λK ψ.

To motivate how this is defined, we first use the group law to rewrite
1
er·e+s·f+λK = er·e+(λ− 2 r·s)K es·f .

By the corresponding Lie algebra action and the fact that ∂xi generates translation
 

es·f ψ (x) = e−s·∇ ψ (x) = ψ(x − s)
1 i 1
er·e+(λ− 2 r·s)K = e h̄ (r·x− 2 r·s+λ)

Therefore the expected transformation law is

  i 1
er·e+s·f+λK ψ (x) = e h̄ (r·x− 2 r·s+λ) ψ(x − s).

Definition 3.3.1. Given (r, s, λ) ∈ R2n+1 , we define

Th̄ (r, s, λ) : S(Rn ) −→ S(Rn )

by
i 1
(Th̄ (r, s, λ)ψ) (x) := e h̄ (r·x− 2 r·s+λ) ψ(x − s).

Proposition 3.3.2. The following composition law holds

 
1
Th̄ (r1 , s1 , λ1 )Th̄ (r2 , s2 , λ2 ) = Th̄ r1 + r2 , s1 + s2 , λ1 + λ2 + (r1 · s2 − r2 · s1 ) .
2

This follows from the above formal computation, and can be verified directly. Thus Th̄
defines a representation of the Heisenberg group Hn

er·e+s·f+λK ψ := Th̄ (r, s, λ)ψ.

It is clear that Th̄ preserves the Hermitian inner product. Thus we can extend Th̄ to

Th̄ (r, s, λ) : L2 (Rn ) −→ L2 (Rn ).

Th̄ defines a unitary representation of Hn on L2 (Rn ). Thus we have a group homomorphism

Th̄ : Hn → U(L2 (Rn ))

where U(L2 (Rn )) is the unitary group of L2 (Rn ). This is called the Schrödinger represen-
tation of Hn with parameter h̄. The parameter indicates the value of the central element
i
K 7−→ .
h̄

154
 The unitary representation Th̄ leads to a transformation of quantum operators via

AdTh̄ : Θ̂ 7−→ Th̄ Θ̂Th̄−1 .

For example,
AdTh̄ (r,s,λ) (x̂i ) = x̂i − si
AdTh̄ (r,s,λ) (p̂i ) = p̂i − ri

It turns out that the unitary representation Th̄ of the Heisenberg group Hn on L2 (Rn ) is
irreducible, i.e., the only closed subspaces of L2 (Rn ) which are invariant under all Hn actions
are {0} and L2 (Rn ) itself. Furthermore, the celebrated Stone-Von Neumann Theorem asserts
that this is essentially the only irreducible unitary representation of Hn on Hilbert spaces.

Theorem 3.3.3 (Stone-Von Neumann). Let ρ be any irreducible unitary representation of Hn

i
with center action by ρ(eK ) = e h̄ on a Hilbert space H. Then there exists a unitary operator
U : H → L2 (Rn ) such that
U ρU −1 = Th̄ .

Such U is uniquely determined up to a phase factor ξ with |ξ| = 1.

We refer to [21] for a proof of Theorem 3.3.3.

3.3.4 Bargmann-Fock Representation

There is another way to obtain the irreducible representation of the Heisenberg group in
terms of holomorphic functions. This is closely related to the Fock space construction.
Let us first define a Hermitian inner product between two functions on Cn by
ˆ
1
d2n z f (z)g(z)e−|z| /h̄ .
2
hf |gih̄ := n
(πh̄) Cn

Here f, g are not necessarily holomorphic, though we will restrict to holomorphic functions
below for Segal-Bargmann space. The normalization constant is chosen such that

h1|1ih̄ = 1.

The corresponding L2 -norm is denoted by

ˆ
1
d2n z |f (z)|2 e−|z|
2 /h̄
||f ||h̄2 = hf |f ih̄ = .
(πh̄)n Cn

Definition 3.3.4. The Segal-Bargmann space (or Bargmann-Fock space) is


Fn := f (z) holomorphic in Cn and ||f ||h̄2 < ∞ .

Proposition 3.3.5. Fn is a Hilbert space.

155
Proof: We need to show Fn is complete with respect to the norm || − ||h̄ . Let PR (z0 ) denote
the polydisk of radius R centered at z0 :

PR (z0 ) = z ∈ Cn | |z i − z0i | < R for i = 1, · · · , n .

For any holomorphic function f (z), we have the average property

ˆ
1
f (z0 ) = d2n z f (z).
(πR2 )n PR (z0 )

Using the Cauchy-Schwartz inequality,

ˆ
1
d2n z |f (z)|e−|z| /2h̄ e|z| /2h̄
2 2
|f (z0 )| ≤ 2 n
(πR ) PR (z0 )
ˆ !1 ˆ !1
2 2
1 |z|2 /h̄ 2 −|z|2 /h̄
≤ 2n
d ze d z |f (z)| e
2n
(πR2 )n PR (z0 ) PR (z0 )
ˆ !1
2
(πh̄)n |z| /h̄
2
≤ d 2n
z e ||f ||h̄ .
(πR2 )n PR (z0 )

This implies that for an open neighborhood V of z0 , we can choose a constant CV depending
on V such that
|f (z)| ≤ CV ||f ||h̄ , ∀z ∈ V

holds for any f ∈ Fn .

Now let {fn } be a Cauchy sequence in Fn . The above estimate shows that {fn } converges
locally uniformly to some limit function f , which must be holomorphic as well. So f ∈ Fn .
Therefore Fn is complete, i.e., a Hilbert space.

Proposition 3.3.6. Let f, g be polynomial functions in z. We have

zif g h̄
= hf |h̄∂z i gih̄ .

Proof: ˆ
1
d2n z z i f (z)g(z)e−|z| /h̄
i 2
zf g h̄
= n
(πh̄) Cn
ˆ  
1 ∂ −|z|2 /h̄
= d z f (z)g(z) −h̄ i e
2n
(πh̄)n Cn ∂z
ˆ
1
d2n z f (z) (h̄∂z i g(z)) e−|z| /h̄
2
=
(πh̄)n Cn
= hf |h̄∂z i gih̄ .

We define the raising and lowering operators a†i and ai by

∂
a†i = z i , ai = h̄ .
∂z i
They are adjoint of each other and satisfy

[ai , a†j ] = δij h̄.

156
 Given a multi-index I = {i1 , i2 , · · · , in }, we define the polynomial

zI
λI (z) := p
h̄|I| I!
where

z I := (z 1 )i1 (z 2 )i2 · · · (z n )in , |I| := i1 + i2 + · · · + in , I! := i1 !i2 ! · · · in !.

If we identify 1 as the vacuum state, then

(a† )i1 (a†n )in

λI = p 1 ··· p 1
h̄i1 i1 ! h̄in in !
are expressions of the excited states in the physics Fock space.

Proposition 3.3.7. {λI } form an orthonormal basis of Fn .

Proof: Orthogonality follows by the adjointness of a†i and ai . We can compute the norm
D E
z I z I h̄ = 1 (a†1 )i1 · · · (a†n )in ai11 · · · ainn 1
h̄
|I| |I|
= h̄ I!h1|1ih̄ = h̄ I!.

Thus {λI } form an orthonormal set. We next show they form a basis.
For all f ∈ Fn , we consider its Taylor series
X
f (z) = cI z I .
I

This series converges to f uniformly on compact subsets of Cn . Let χR denote the characteristic
function on the polydisk of radius R

1 if z ∈ PR (0)
χR (z) = .
0 otherwise

A direct computation using polar coordinate shows

z I χR z J χR h̄
= δIJ dI,R

where the constant dI,R increases with R and

lim dI,R = h̄|I| I!.

R→∞

On PR (0), the uniform convergence of the series and the above orthogonality imply
X
||f χR ||h̄2 = |cI |2 dI,R .
I

Taking R → ∞, we find
X
||f ||h̄2 = |cI |2 ||z I ||h̄2
I

This implies that {λI } form a basis.

157
Proposition 3.3.8. Let f ∈ Fn . Then

||z i f ||h̄2 = h̄||f ||h̄2 + ||h̄∂z i f ||h̄2

In particular, ai , a†i have the same domain on Fn .

Proof: The identity holds for f = λI , and in general by Parseval equation.

Next we consider representation of the Heisenberg group Hn on the Segal-Bargmann space

Fn . Recall the construction of raising operators and lowering operators in harmonic oscillator

a† = √1 (x̂ − ip̂)
2
a = √1 (x̂ + ip̂)
2
or 
x̂ = √1 (a + a† )
2
p̂ = √i (a† − a)
2

Let R2n be the standard symplectic space with symplectic basis {e1 , · · · , en , f1 , · · · , fn }.
Recall that in the Schrödinger representation, we have
i i
ei 7−→ x̂
h̄
i
fi 7−→ − p̂i
h̄
i
K 7−→
h̄
Comparing with the above formula of raising and lowering operators, this suggests to define the
representation ρ on Fn by  
i ∂
ei 7−→ √ z i + h̄ i
2h̄ ∂z
 
1 ∂
fi 7−→ √ z − h̄ i
i
2h̄ ∂z
i
K 7−→
h̄
Thus
1 1
r · e + s · f 7−→ √ (ir + s) · z + √ (ir − s) · ∇z .
2h̄ 2
Let us identify (r, s) with complex coordinates w ∈ Cn by
 
w = √1 (s − ir) r = √i (w − w)
2 or 2
w = √1 (s + ir) s = √1 (w + w)
2 2

Then we can write

1
w · z − w · ∇z .
r · e + s · f 7−→
h̄
This leads us to define the representation ρ of the Heisenberg group by
1
ρh̄ (w) = e h̄ w·z−w·∇z .

158
Using the Baker Campbell Hausdorff formula, these operators satisfy the composition law

ρh̄ (w1 )ρh̄ (w2 ) = ρh̄ (w1 + w2 )e 2h̄ (w1 ·w2 −w1 ·w2 )
1

= ρh̄ (w1 + w2 )e h̄ Im(w1 ·w2 ) .

i

In particular, we have
ρh̄ (w) = e h̄ w·z− 2h̄ |w| e−w·∇z
1 1 2

Thus we find the following explicit formula of the representation on Fn

(ρh̄ (w)f ) (z) := e h̄ w·z− 2h̄ |w| f (z − w).

1 1 2

This is the Bargmann-Fock representation of the Heisenberg group.

By the Stone-Von Newmann Theorem, there exists a unitary map

B : L2 (Rn ) −→ Fn

intertwining the Schrödinger representation Th̄ and the Bargmann-Fock representation ρh̄
B
L2 (Rn ) Fn
Th̄ ρh̄

B
L2 (Rn ) Fn

Explicitly, such B is found by

ˆ √
1
dn x e− 2h̄ (z·z−2
1
2z·x+x·x)
(Bψ) (z) = ψ(x).
(πh̄)n/4 Rn

To understand this formula, let us check at the Lie algebra level.

Th̄ i
r · e + s · f −→ r · x − s · ∇x
h̄
1 1
= − √ (w − w) · x − √ (w + w) · ∇x
2h̄ 2
   
1 1 h̄ 1 1
= w · √ x − √ ∇x − w · √ x + √ ∇x
h̄ 2 2 2h̄ 2
ρh̄ 1
r · e + s · f −→ w · z − w · ∇z
h̄
Comparing the above two infinitesimal transformations, B should intertwine
  
1 i h̄
z (Bψ) = B
i
√ x − √ ∂ xi ψ
2 2
  
1 i 1
∂z i (Bψ) = B √ x + √ ∂ xi ψ
2h̄ 2
It is not hard to check that the above defined B is designed to satisfy these two equations.

159
Let us also check the unitarity. Let ψ ∈ S(Rn ) be a Schwartz function. Then
ˆ ˆ √ √
1 − 2h̄
1
(z·z−2 2z·x+x·x+z̄·z̄−2 2z̄·y+y·y) −|z|2 /h̄
||Bψ||h̄2 = d n n
xd y ψ(x)ψ(y) d 2n
z e e
(πh̄)3n/2
√ ˆ ˆ
z=(r+is)/ 2 1
d xd y ψ(x)ψ(y) dn rdn s e− h̄ (r −r·(x+y)+is·(x−y)+ 2 (x·x+y·y))
1 2 1
n n
========= n 3n/2
2 (πh̄)
ˆ ˆ
1 − 4h̄
dn s e− h̄ s·(x−y)
1 i
n n (x·x−y·y)2
= n
d xd y ψ(x)ψ(y)e
(2πh̄)
ˆ
= dn xdn y ψ(x)ψ(y)e− 4h̄ (x·x−y·y) δ(x − y)
1 2

ˆ
= dn x ψ(x)ψ(x) = ||ψ||2L2 (Rn )

as expected. The unitary map B : L2 (Rn ) → Fn is called the Segal-Bargmann transform.

3.4 Weyl Wigner Transform

3.4.1 Weyl Quantization

Let f (y, ξ) ∈ Cc∞ (R2n ) be a smooth function on the phase space R2n with compact support.
We define the following operator on the state space
ˆ
1
Th̄ (f ) := dn ξdn y f (y, ξ)Th̄ (ξ, y)
(2πh̄)n R2n

where we denote the operator (see Definition 3.3.1)

Th̄ (ξ, y) := Th̄ (ξ, y, 0)

for simplicity. Explicitly, given a wave function ψ(x) ∈ S(Rn ), the operator Th̄ (f ) acts as
ˆ
1
Th̄ (f )(ψ)(x) = dn ξdn y f (y, ξ)(Th̄ (ξ, y)ψ)(x)
(2πh̄)n R2n
ˆ
1 i 1
= n
dn ξdn y f (y, ξ)e h̄ (ξ·x− 2 ξ·y) ψ(x − y)
(2πh̄) R2n
ˆ
y→x−y 1 i x+y
====== n
dn ξdn y f (x − y, ξ)e h̄ ξ·( 2 ) ψ(y)
(2πh̄) R2n
ˆ
1
= dn y Kf (x, y)ψ(y).
(2πh̄)n Rn

Here ˆ
i x+y
Kf (x, y) = dn ξ f (x − y, ξ)e h̄ ξ·( 2 ).
Rn
This can be rewritten as
  ˆ
1 1 i
Kf x + y, x − y = dn ξ f (y, ξ)e h̄ ξ·x
2 2 Rn

which is a partial Fourier transform of f in the variable ξ. Since the transformation

 
1 1
(x, y) −→ x + y, x − y
2 2

160
also preserves the measure, the transformation

f 7−→ Kf

preserves the L2 -norm.

Theorem 3.4.1. Th̄ defines a unitary map from L2 (R2n ) to the space of Hilbert-Schmidt oper-
ators on L2 (Rn ).

Proof: Kf is the kernel function of the operator Th̄ (f ). The theorem follows from the above
computation that f 7→ Kf is unitary.

Moreover, it is clear that for f ∈ L1 (R2n ), Th̄ (f ) defines a bounded operator on L2 (Rn ).
By the Schwartz Kernel Theorem, Th̄ further extends to a bijection

≃
Th̄ : S ′ (R2n ) −−→ L (S(Rn ), S ′ (Rn )).

Here S ′ (Rn ) is the space of tempered distributions and L (−, −) refers to continuous linear
maps.

Example 3.4.2. Let f (y, ξ) = δ(y − y0 , ξ − ξ0 ) be the Dirac δ-distribution at (ξ0 , y0 ). Then
1
Th̄ (f ) = Th̄ (ξ0 , y0 ).
(2πh̄)n

Example 3.4.3. Let f = 1 be the constant function. Then

ˆ  
n i
ξ· ( x+y
) n x+y
Kf = d ξ e h̄ 2 = (2πh̄) δ
Rn 2

Thus
Th̄ (f )(ψ)(x) = 2n ψ(−x)

is a reflection up to a rescaling.

Definition 3.4.4. Let f (x, p) ∈ S ′ (R2n ) be a tempered distribution on the phase space. We
define its Weyl transform to be
ˆ
1
Wh̄ (f ) := Th̄ (fbω ) = dn ydn ξ fbω (y, ξ)e h̄ (ξ·x̂−y·p̂)
i

(2πh̄)n R2n

where fbω is the symplectic Fourier transform (composition of the Fourier transform with J)
ˆ
b 1
dn xdn p e− h̄ (ξ·x−y·p) f (x, p).
i
fω (y, ξ) :=
(2πh̄)n R2n

Since Fourier transform extends to tempered distributions, fbω ∈ S ′ (R2n ). Hence Wh̄ (f )
defines a continuous linear map

Wh̄ (f ) : S(Rn ) −→ S ′ (Rn ).

161
 If f ∈ L2 (R2n ), then Wh̄ (f ) defines a Hilbert-Schmidt operator

Wh̄ (f ) : L2 (Rn ) −→ L2 (Rn ).

The operator Wh̄ (f ) is also called the Weyl quantization of the phase space function f .

Example 3.4.5. Consider the function

i
f (x, p) = e h̄ (a·x+b·p) , a, b ∈ Rn .

Its symplectic Fourier transform is

ˆ
b 1 i
fω (y, ξ) = n
dn xdn p e h̄ (x·(a−ξ)+p·(b+y)) = (2πh̄)n δ(y + b)δ(ξ − a).
(2πh̄)
Thus
i
Wh̄ (f ) = Th̄ ((2πh̄)n δ(y + b)δ(ξ − a)) = Th̄ (a, −b) = e h̄ (a·x̂+b·p̂)

i.e.
 i  i
Wh̄ e h̄ (a·x+b·p) = e h̄ (a·x̂+b·p̂) .

Expanding the above formula in powers of a and b, we find

Wh̄ (xI pK ) = Sym(x̂I p̂K ).

Here I, K are multi-indices and Sym(−) is the symmetrized order average.

For example, consider the n = 1 case. Then
1
Wh̄ (xp) = Sym(x̂p̂) = (x̂p̂ + p̂x̂)
2
1
Wh̄ (x2 p) = (x̂2 p̂ + x̂p̂x̂ + p̂x̂2 )
3
1
Wh̄ (x p ) = (x̂2 p̂2 + p̂2 x̂2 + x̂p̂x̂p̂ + x̂p̂2 x̂ + p̂x̂p̂x̂ + p̂x̂2 p̂)
2 2
6
This symmetrized ordering is also called the Weyl ordering.

3.4.2 Moyal Product

The Weyl transform identifies a function on the phase space with an operator on the state
space. Since composition of operators defines an associative product, it is natural to ask how
such product is reflected on phase space functions under the Weyl transform.
Recall from the composition law of the Heisenberg group, we have
 
i ξ1 ·y2 −ξ2 ·y1
Th̄ (ξ1 , y1 )Th̄ (ξ2 , y2 ) = Th̄ (ξ1 + ξ2 , y1 + y2 )e h̄ 2
.

Thus
ˆ ˆ
1
Th̄ (fb1 )Th̄ (fb2 ) = d n
ξ 1 d n
y b
f (y , ξ )T
1 1 1 1 h̄ 1 1 (ξ , y ) dn ξ2 dn y2 fb2 (y2 , ξ2 )Th̄ (ξ2 , y2 )
(2πh̄)2n R2n R 2n
ˆ
1
d ξ1 d y1 d ξ2 d y2 fb1 (y1 , ξ1 )fb2 (y2 , ξ2 )e 2h̄ (ξ1 ·y2 −ξ2 ·y1 ) Th̄ (ξ1 + ξ2 , y1 + y2 )
i
n n n n
=
(2πh̄)2n
ˆ  
1 n n b1 #fb2 (y, ξ)Th̄ (ξ, y)
= d ξd y f
(2πh̄)n

162
where

  ˆ
1
fb1 #fb2 (ξ, y) := dn ηdn z fb1 (ξ − η, y − z)fb2 (η, z)e 2h̄ ((ξ−η)·z−η·(y−z))
i

n
(2πh̄) R2n
ˆ
1
dn ηdn z fb1 (ξ − η, y − z)fb2 (η, z)e 2h̄ (ξ·z−η·y) .
i
=
(2πh̄)n R2n

Applying this to the Weyl transform, we find

Wh̄ (f )Wh̄ (g) = Th̄ (fbω )Th̄ (b

gω ) = Th̄ (fbω #b
gω ).

Thus we are led to define a product ∗h̄ on phase space functions by requiring

(f\
∗h̄ g)ω = fbω #b
gω

i.e.
ˆ  
1 bω #b i
(f ∗h̄ g)(x, p) = n
d n
ξd n
y f g ω (y, ξ)e h̄
(ξ·x−y·p)
(2πh̄)
ˆ ˆ
1
d ξd y dn ηdn z fbω (y − z, ξ − η)b
i i
n n
= 2n
gω (z, η)e 2h̄ (ξ·z−η·y) e h̄ (ξ·x−p·y)
(2πh̄)
ˆ
1
dn ξdn ydn ηdn z fbω (y, ξ)b
gω (z, η)e h̄ [(ξ+η)·(x+ 2 z)−(p+ 2 η)·(y+z)]
i 1 1
= 2n
(2πh̄)
ˆ
1
dn ξdn ydn ηdn z fbω (y, ξ)b
i i i
= 2n
gω (z, η)e h̄ (ξ·x−p·y) e h̄ (η·x−p·z) e 2h̄ (ξ·z−η·y) .
(2πh̄)

From the expression of the Fourier transform

ˆ
b 1
dn pdn xe− h̄ (ξ·x−p·y) f (x, p)
i
fω (y, ξ) = n
(2πh̄)
we have
\i f )
ξi fbω (y, ξ) = (−ih̄∂ x ω
\
yi fbω (y, ξ) = (ih̄∂ pi f ) ω

Thus we can formally write the above product as

 − →
← − ←− → − 
ih̄ ∂ ∂
− ∂p
∂ ∂
(f ∗h̄ g)(x, p) ∼ f (x, p)e 2 ∂xi ∂pi i ∂xi g(x, p).
←− −
→
∂ ∂
Here ∂(−) means applying the derivative to the function on the left, and ∂(−) means applying
the derivative to the function on the right. This formal expression becomes an exact formula
when f, g are polynomials.
The above defined ∗h̄ is called the Moyal product, or the Moyal-Weyl product. Thus
we have established
Wh̄ (f ∗h̄ g) = Wh̄ (f )Wh̄ (g).

163
3.4.3 Wigner Transform

Wigner transform and Wigner function

The Wigner transform is the inverse of the Weyl transform, thus takes an operator back to
a function on the phase space. Recall the Weyl transform

Wh̄ (f ) = Th̄ (fbω )

where fbω is the symplectic Fourier transform

ˆ
1
fbω (y, ξ) := dn xdn p e− h̄ (ξ·x−y·p) f (x, p).
i

n
(2πh̄) R2n

The operator Th̄ (fbω ) on L2 (Rn ) has a kernel 1

(2πh̄)n Kfbω (x, y) where
ˆ
dn ξ fbω (x − y, ξ)e h̄ ξ·(
i x+y
Kfbω (x, y) = 2 )
Rn

or equivalently   ˆ
1 1
dn ξ fbω (y, ξ)e h̄ ξ·x .
i
Kfbω x + y, x − y =
2 2 Rn
Via the inverse Fourier transform, we find
ˆ
1
dn ydn ξ e h̄ (ξ·x−y·p) fbω (y, ξ)
i
f (x, p) = n
(2πh̄) R2n
ˆ  
1 − h̄i y·p 1 1
= n
d ye Kfbω x + y, x − y .
(2πh̄)n Rn 2 2

This formula leads to the inverse map of the Weyl transform defined as follows.

1
Definition 3.4.6. Let Θ be an operator on L2 (Rn ) with kernel function (2πh̄)n KΘ (x, y). Define
the Wigner transform Wig[Θ] as a function on the phase space by
ˆ  
1 − h̄i y·p 1 1
Wig[Θ](x, p) := n
d ye KΘ x + y, x − y .
(2πh̄)n Rn 2 2

It is clear from the above computation that

Wig[Wh̄ (f )] = f,

i.e., the Wigner transform is the inverse of the Weyl transform.

Example 3.4.7. Let Θ be the operator with kernel function

KΘ (x, y) = ψ1 (x)ψ2 (y).

Its Wigner transform is

ˆ    
1 − h̄i y·p 1 1
Wig [Θ] (x, p) = n
d ye ψ1 x + y ψ2 x − y .
(2πh̄)n Rn 2 2

164
 A particular case of great interest is when

KΘ (x, y) = ψ(x)ψ(y).

This operator Θ projects any state to a state proportional to |ψi. We denote its Wigner
transform by
Wig(ψ) := Wig [Θ]

or ˆ    
1 1 1
dn y e− h̄ y·p ψ x + y ψ x − y .
i
Wig(ψ)(x, p) =
(2πh̄)n Rn 2 2
Wig(ψ) is called the Wigner function, or the Wigner distribution, of the state ψ. Wigner
proposed it as a substitute for the the nonexistent joint probability distribution of momentum
and position in the quantum state ψ. Note that it does not make sense to speak of a joint
probability distribution for momentum and position due to uncertainty principle. Moreover,
Wig(ψ) is usually not a genuine probability density since it could take negative values (see the
example of harmonic oscillator below).

Proposition 3.4.8. ˆ
dn p Wig(ψ)(x, p) = |ψ(x)|2
ˆR
n

dn x Wig(ψ)(x, p) = |ψ̂(p)|2
Rn
´
dn x e− h̄ x·p ψ(x) is the Fourier transform of ψ.
i
1
Here ψ̂(p) = (2πh̄)n/2 Rn

Proof:
ˆ ˆ ˆ    
1 − h̄i y·p 1 1
n
d p Wig(ψ)(x, p) = n
d p n
d ye ψ x+ y ψ x− y
Rn (2πh̄)n Rn Rn 2 2
ˆ    
1 1
= dn y δ(y)ψ x + y ψ x − y
R n 2 2
= ψ(x)ψ(x).
Similarly,
ˆ ˆ ˆ    
1 − h̄i y·p 1 1
n
d x Wig(ψ)(x, p) = n
d x n
d ye ψ x+ y ψ x− y
Rn (2πh̄)n Rn Rn 2 2
ˆ ˆ
x→(x+y)/2 1
dn y e− h̄ (x−y)·p ψ(x)ψ(y)
i
========= dn x
y→x−y (2πh̄)n Rn R n
ˆ ˆ
1 − h̄i x·p i
= n
n
d xe ψ(x) dn y e h̄ y·p ψ(y)
(2πh̄) Rn Rn

=ψ(p)ψ(p).

This proposition says the integration of the Wigner function over the momentum space /
position space produces the probability distribution of the quantum state in the position space
/ momentum space. In particular, we have
ˆ
dn xdn p Wig(ψ)(x, p) = ||ψ||2L2 .
R2n

165
Assume ||ψ||L2 = 1 is normalized, then Wig(ψ) is also called the Wigner quasi-probability
distribution.

Proposition 3.4.9. Assume ||ψ||L2 = 1 is normalized. Then

1 1
− ≤ Wig(ψ) ≤ .
(πh̄)n (πh̄)n

Proof:
ˆ    
1 1 1
| Wig(ψ)(x, p)| ≤ dn y ψ x + y ψ x− y
(2πh̄)n Rn 2 2
ˆ   2ˆ   2 ! 12
1 1 1
≤ dn y ψ x + y dn y ψ x − y
(2πh̄)n Rn 2 Rn 2
1 1
= n
||ψ||L2 = .
(πh̄) (πh̄)n

As a corollary, let Ω ⊂ R2n be any region in the phase space. Assume ||ψ||L2 = 1 is
normalized. Then ˆ
Vol(Ω)
dn xdn p Wig(ψ)(x, p) ≤ .
Ω (πh̄)n
On the other hand, we have
ˆ
dn xdn p Wig(ψ)(x, p) = ||ψ||L2 = 1.
R2n

Comparing the above two expressions, we find that it is impossible for Wig(ψ) to concen-
trate in some region Ω with Vol(Ω) ≤ (πh̄)n . This is a reflection of the uncertainty principle.

Example 3.4.10. Consider the one-dimensional harmonic oscillator

p2 mω 2 2
H = + x .
2m 2


The energy levels are En = h̄ω n + 1
2 , n = 0, 1, 2, · · · , with the normalized eigenfunctions by

1 x x2
ψn (x) = p √ Hn e− 2a2
2n n!a π a
q
h̄
where a = mω and Hn are Hermite polynomials.
Let us consider the following linear superposition (coherent state)
∞
X |z|2 zn
ψz (x) = e− 2 √ ψn (x)
n=0 n!
√ !
1 |z|2 + z 2 2zx x2
= p √ exp − + − 2 .
a π 2 a 2a

166
 Let us consider the analogue of the Wigner function construction

Wz1 ,z2 (x, p)

ˆ 
1 i y  y
= dy e h̄ yp ψz1 x + ψ z2 x −
2πh̄ 2 2
ˆ    2 
1 i |z1 | + |z2 |2
2 z + z 22
= √ dy e h̄ yp exp − exp − 1
(2πh̄) a π 2 2
√ √ !  
2z1 (x + y/2) 2 z 2 (x − y/2) (x + y/2)2 (x − y/2)2
exp + exp − −
a a 2a2 2a2
 2     
1 x a 2 p2 |z1 |2 + |z2 |2 x/a − ipa/h̄ x/a + ipa/h̄
= exp − 2 − 2 exp − − z 1 z2 exp 2z2 √ + 2z 1 √ .
πh̄ a h̄ 2 2 2
(z1 z 2 )n
We can expand into power series in z1 , z2 and collect terms proportional to n! to find
 2  n    j
(−1)n x a2 p2 X n (−1)j 2j x2 a2 p2
Wig(ψn )(x, p) = exp − 2 − 2 + 2
πh̄ a h̄ j j! a2 h̄
j=0
(−1)n −u x 2 a 2 p2
= e Ln (2u), u= + 2
πh̄ a2 h̄
where Ln are the Laguerre polynomials

et dn −t n
Ln (t) = (e t ).
n! dtn
We list a first few terms of Laguerre polynomials

L0 =1
L1 (t) = −t + 1
1
L2 (t) = (t2 − 4t + 2)
2
1
L3 (t) = (−t3 + 9t2 − 18t + 6)
6
1
L4 (t) = (t4 − 16t3 + 72t2 − 96t + 24)
24
..
.

3.5 Geometric Quantization

3.5.1 Dirac Quantization Principle

Classical mechanics is rooted in the symplectic geometry of the phase space manifold M .
The space Obscl := C ∞ (M ) of classical observables carries a structure of Poisson bracket {−, −}.
We say {f1 , · · · , fm } is a complete set of classical observables if for g ∈ Obscl ,

{g, f1 } = · · · = {g, fm } = 0 =⇒ g = constant.


For example, on M = R2n , the set of position and momentum functions x1 , · · · , xn , p1 , · · · , pn
is complete.

167
 In quantum mechanics, we have a Hilbert space H representing the space of quantum states.
The space Obsq of quantum observables is an appropriate collection of linear operators on H
that carries the natural structure of operator commutator

[O1 , O2 ] = O1 O2 − O2 O1 .

Similarly, we say quantum observables {O1 , · · · , Om } is a complete set of quantum observables

if for O ∈ Obsq ,

[O, O1 ] = · · · = [O, Om ] = 0 =⇒ O = c1 is a multiple of the identity.

In Dirac’s Principle, quantization amounts to an assignment

Q : Obscl −→ Obsq
f 7−→ fˆ

of operators fˆ on some Hilbert space to classical observables f . This assignment has to satisfy

Q1: Q is R-linear

Q2: 1̂ = Q(1) is the identity operator

Q3: if f is real, then fˆ is self-adjoint

h i
\
Q4: fˆ, ĝ = ih̄{f, g}
n o
Q5: if {f1 , · · · , fm } is a complete set of classical observables, then fˆ1 , · · · , fˆm is a complete
set of quantum observables.

Q1,Q2,Q3 are natural requirements. Q4 is Dirac’s observation that the operator commu-
tator should be the quantum counterpart of the classical Poisson bracket. Q5 can be viewed as
an irreducibility condition. The Hilbert space H is in general infinite dimensional. But it could
be finite dimensional as well (such as describing internal symmetries).
Unfortunately, it is in general not possible to satisfy both Q4 and Q5. This is illustrated
by a “no-go” type result for quantization that we will discuss in Section 3.5.2.
Before that, let us first take a look at Weyl quantization on Schrödinger representation.
Let f (x, p) be a polynomial. Its Weyl quantization is the operator

Wh̄ (f ) = Sym(f (x̂i , p̂i ))

where
∂
x̂i = xi p̂i = −ih̄
∂xi
and Sym(−) is the symmetrized order average. We have

Wh̄ (f )Wh̄ (g) = Wh̄ (f ∗h̄ g)

168
where f ∗h̄ g is the Moyal product
 ←
− →− ←− → − 
ih̄ ∂ ∂
− ∂p
∂ ∂
(f ∗h̄ g)(x, p) = f (x, p)e 2 ∂xi ∂pi i ∂xi g(x, p).

In particular,
[Wh̄ (f ), Wh̄ (g)] = Wh̄ ([f, g]∗ )

where
[f, g]∗ := f ∗h̄ g − g ∗h̄ f

is the commutator of Moyal product. We compute

←− −→ ←
− −→ !3
(ih̄)3 ∂ ∂ ∂ ∂
[f, g]∗ = ih̄ {f, g} + f − g + ··· .
24 ∂xi ∂pi ∂pi ∂xi

Thus in the Schrödinger representation, the Weyl quantization violates Q4.

On the other hand, the above computation implies

[f, g]∗ = ih̄ {f, g}

if f, g are polynomials of deg ≤ 2. We will see next that deg = 4 leads to the obstruction.

3.5.2 The Groenewold-Van Hove Theorem

The Groenewold-Van Hove Theorem established several “No-Go” type results for quanti-
zation. We describe the simplest version in the following.
Consider quantization on R2n . Let

P ≤k := f ∈ R[xi , pi ]| deg f ≤ k .

The Schrödinger representation assigns elements of P ≤1 to operators

1 7−→ 1̂ = 1
xi 7−→ x̂i = xi
∂
pi 7−→ p̂i = −ih̄
∂xi
and
\
[x̂i , p̂j ] = ih̄δij = ih̄{x i , p }.
j

Theorem 3.5.1 (Groenewold). There is no R-linear map

Q : P ≤4 −→ self-adjoint operators on L2 (Rn )

satisfying

(1) Q(1) = 1, Q(xi ) = x̂i , Q(pi ) = p̂i

(2) [Q(f ), Q(g)] = ih̄Q({f, g}), ∀f, g ∈ P ≤4 .

169
Proof: We prove the case for n = 1. The proof for the case n > 1 is similar. Assume such Q
exists. Recall the Weyl quantization Wh̄ . By assumption (1)

Q(f ) = Wh̄ (f ) for f ∈ P ≤1 .

Observe the Poisson bracket satisfies


P ≤2 , P ≤1 ⊂ P ≤1

and
[Wh̄ (f ), Wh̄ (g)] = ih̄Wh̄ ({f, g}) for f ∈ P ≤2 , g ∈ P ≤1 .

Therefore
\
[Wh̄ (f ), ĝ] = ih̄{f, g} = [Q(f ), ĝ] for f ∈ P ≤2 , g ∈ P ≤1 .

Thus
[Q(f ) − Wh̄ (f ), x̂] = [Q(f ) − Wh̄ (f ), p̂] = 0.

It follows that the exponentiated one-parameter family of unitary operators eiα(Q(f )−Wh̄ (f ))
on L2 (Rn ) commute with the Heisenberg group action and give rise to automorphisms of the
Schrödinger representation. By Stone-von Neumann Theorem (Theorem 3.3.3), such automor-
phisms must be scalar multiplication so

Q(f ) − Wh̄ (f ) = const 1 for f ∈ P ≤2 .

1

The above constants on the right are actually zero. In fact, using x2 = 2 x2 , xp ,

1
Q(x2 ) = [Q(x2 ), Q(xp)]
2ih̄
1
= [Wh̄ (x2 ) + c1 , Wh̄ (xp) + c2 ] c1 , c2 are some constants
2ih̄
1
= [Wh̄ (x2 ), Wh̄ (xp)]
2ih̄
= Wh̄ (x2 ).

Similarly we can show Q(p2 ) = Wh̄ (p2 ) and Q(xp) = Wh̄ (xp). Thus

Q(f ) = Wh̄ (f ) for f ∈ P ≤2 .

Next we observe that for f ∈ P ≤3 , g ∈ P ≤1 ,

[Wh̄ (f ), Wh̄ (g)] = ih̄Wh̄ ({f, g})

still holds. Thus for f ∈ P ≤3 , g ∈ P ≤1 ,

[Q(f ) − Wh̄ (f ), ĝ] = ih̄Q({f, g}) − ih̄Wh̄ ({f, g}) = 0

since {f, g} ∈ P ≤2 . By the same argument as above, we have

Q(f ) − Wh̄ (f ) = const 1 for f ∈ P ≤3 .

170
 Similarly, using
1 3 1
x3 =
x , xp p3 = xp, p3
3 3
1 3 2 1 2 3
x2 p = x ,p xp2 = x ,p
6 6
we can deduce that the above constants on the right are zero. Hence

Q(f ) = Wh̄ (f ) for f ∈ P ≤3 .

Now we come to degree 4 polynomials. Consider

1 3 3 1 2
x 2 p2 = x ,p = x p, xp2 .
9 3
By assumption (2), we would have
1 1
Q(x2 p2 ) = [Q(x3 ), Q(p3 )] = [Q(x2 p), Q(xp2 )].
9ih̄ 3ih̄
On the other hand, using
3
[x3 , p3 ]∗ = 9(ih̄)x2 p2 + (ih̄)3
2
1
[x p, xp ]∗ = 3(ih̄)x p − (ih̄)3
2 2 2 2
2
we have

1 1 1 1
[Q(x3 ), Q(p3 )] = [Wh̄ (x3 ), Wh̄ (p3 )] = Wh̄ ([x3 , p3 ]∗ ) = Wh̄ (x2 p2 ) − h̄2
9ih̄ 9ih̄ 9ih̄ 6
and
1 1 1 1
[Q(x2 p), Q(xp2 )] = [Wh̄ (x2 p), Wh̄ (xp2 )] = Wh̄ ([x2 p, xp2 ]∗ ) = Wh̄ (x2 p2 ) + h̄2
3ih̄ 3ih̄ 3ih̄ 6
1
6= 3 3
[Q(x ), Q(p )].
9ih̄
Contradiction.

3.5.3 Prequantization

Since it is unrealistic to construct quantization satisfying all the conditions Q1-Q5, we have
to relax either condition Q4 or Q5. There is a natural way to construct a quantization of a
symplectic manifold which satisfies conditions Q1-Q4. This is known as the prequantization.
Let (M, ω) be a symplectic manifold. Recall the construction of Hamiltonian vector fields

C ∞ (M ) −→ Vect(M )
f 7−→ Vf , ιVf ω = df

which is a Lie algebra homomorphism

[Vf , Vg ] = V{f,g} .

171
 If we think of Vf as a first order differential operator acting on C ∞ (M ), then the assignment

f 7−→ ih̄Vf

satisfies Q1, Q3, Q4 but not Q2 since it sends 1 7−→ 0.

To remedy this, let L → M be a complex line bundle on M with a connection ∇. Let F∇
be the curvature 2-form. For any vector fields X, Y , we have

[∇X , ∇Y ] − ∇[X,Y ] = F∇ (X, Y ).

Given a function f , we use ∇f to denote the covariant derivative with respect to the
Hamiltonian vector field Vf
∇f := ∇Vf .

Then the above curvature formula reads

[∇f , ∇g ] − ∇{f,g} = F∇ (∇f , ∇g ) = ιVg ιVf F∇ .

Here ιVf is the interior product (contraction) with the vector field Vf .
Consider the following assignment of operators on Γ(M, L)

Q(f ) := ih̄∇f + f.

We have
[Q(f ), Q(g)] =[ih̄∇f + f, ih̄∇g + g]


=(ih̄)2 ∇{f,g} + F∇ (Vf , Vg ) + 2ih̄ {f, g}


=(ih̄) ih̄∇{f,g} + {f, g} + (ih̄)2 F∇ (Vf , Vg ) + ih̄ {f, g}
=ih̄Q({f, g}) + (ih̄)2 F∇ (Vf , Vg ) + ih̄ {f, g}
=ih̄Q({f, g}) + (ih̄)2 F∇ (Vf , Vg ) + ih̄ιVf ιVg ω
Thus if the curvature 2-form is
i
F∇ = − ω
h̄
then the following relation holds

[Q(f ), Q(g)] = ih̄Q({f, g}).

This calculation motivates the following definition.

Definition 3.5.2. A prequantum line bundle on (M, ω) is a Hermitian line bundle (L , h−, −i)
with a compatible connection ∇ such that its curvature 2-form F∇ satisfies
i
F∇ = − ω.
h̄
Here we recall a Hermitian line bundle is a complex line bundle L with a smoothly varying
family of Hermitian inner products h−, −i on fibers of L . Our convention for Hermitian inner
product is that h−, −i is complex-conjugate linear in the first argument and complex linear in
the second. Thus for any local sections s, s1 , s2 ∈ Γ(U, L ) and a local function f ∈ C ∞ (U ),

172
 • hs1 , s2 i = hs2 , s1 i

• hf s1 , s2 i = f hs1 , s2 i

• hs1 , f s2 i = f hs1 , s2 i

• hs, si ≥ 0 and hs, si = 0 ⇐⇒ s = 0.

The connection ∇ is said to be compatible with the Hermitian inner product if for any real
vector field X on M and local sections s1 , s2 of L ,

X hs1 , s2 i = h∇X s1 , s2 i + hs1 , ∇X s2 i .

The Hermitian structure h−, −i induces a Hermitian inner product h−|−i between global
sections of L by ˆ
ωn
hs1 |s2 i := hs1 , s2 i .
M n!
We obtain a Hilbert space L2 (M, L )
by square integrable measurable sections of L . Thus
ˆ 1
ωn 2
||s|| := hs, si < ∞ for s ∈ L2 (M, L ).
M n!
Let (L , ω, h−, −i) be a prequantum line bundle on (M, ω). For a smooth function f , we
assign the following operator

Q(f ) = ih̄∇f + f : Γ(M, L ) −→ Γ(M, L ).

The curvature condition F∇ = − h̄i ω implies

[Q(f ), Q(g)] = ih̄Q({f, g})

so condition Q4 is satisfied. Also we have

Q(1) = 1

so condition Q2 is satisfied. Condition Q1 on R-linearity is obvious.

Let us examine condition Q3. Let s1 , s2 ∈ Γc (M, L ) be two sections of L with compact
support. Let f be a real smooth function on M . Compatibility of ∇ with h−, −i implies

∇f hs1 , s2 i = h∇f s1 , s2 i + hs1 , ∇f s2 i .

ωn
Integrating over M with respect to the volume form n! , we find

h∇f s1 |s2 i + hs1 |∇f s2 i

ˆ
ωn
= (h∇f s1 , s2 i + hs1 , ∇f s2 i)
n!
ˆM
ωn
= ∇f hs1 , s2 i
n!
ˆM

ωn
= LVf hs1 , s2 i
n!
M
ˆ  n
ω
=− hs1 , s2 i LVf = 0.
M n!

173
Here in the last step, we have used the Liouville Theorem on the invariance of the Liouville
volume form under Hamiltonian flows. This implies the self-adjoint relation

hQ(f )s1 |s2 i = hs1 |Q(f )s2 i.

Thus a prequantum line bundle leads to a quantization that satisfies conditions Q1-Q4.
The above discussion is based on a given prequantum line bundle. However, it is not clear
whether prequantum line bundles exist or not. There is indeed a topological obstruction for it.
The 2-form
i
F∇
2π
is closed, and its de Rham cohomology class
 
i
F∇ = c1 (L )
2π
represents the first Chern class of L. One remarkable feature of Chern class is the integrality:
ˆ
i
F∇ ∈ Z
2π Σ
1
for any closed surface Σ. Therefore 2πh̄ [ω] represents an integral cohomology class. The converse
1
is also true: if 2πh̄ [ω] is integral, then there exists a prequantum line bundle.

Definition 3.5.3. We say the symplectic manifold (M, ω) is quantizable (for a particular value
of h̄) if ˆ
1
ω∈Z
2πh̄ Σ
for every closed surface Σ in M .

Thus on a quantizable phase space, we can construct a quantization satisfying Q1-Q4. Note
that such a quantization is far from expectation in quantum mechanics. For example

Q(f ) = ih̄∇f + f

is always a first order differential operator for any function f . But we should expect higher
order differential operators, such as Schrödinger operator, in quantum mechanics. The problem
is condition Q5, i.e., the above constructed Hilbert space L2 (M, L ) is too big. We will consider
how to cut down the degrees of freedom in Section 3.5.4.

Example 3.5.4. Consider the total space of the cotangent bundle

M = T ∗ X, ω = dθ

where θ is the Liouville 1-form. Since ω is exact,

ˆ ˆ
ω= dθ = 0
Σ Σ
for any closed surface Σ ⊂ M . Thus cotangent bundles are always quantizable. In fact, we can
choose L to be the trivial line bundle and the connection by
i
∇=d− θ.
h̄

174
 
Example 3.5.5. M = S 2 = x2 + y 2 + z 2 = 1 ⊂ R3 .

ω = A(xdydz + ydzdx + zdxdy), A > 0.

We have ˆ ˆ
ω= dω = 4πA.
S2 x2 +y 2 +z 2 ≤1

Thus (M, ω) is quantizable ⇐⇒ A ∈ h̄2 Z.

3.5.4 Polarization and Quantum States

The problem with the Hilbert space L2 (M, L ) arising from a prequantum line bundle L
on a quantizable symplectic manifold M is that it is too big. For example on M = R2n , a
section ψ of L depends on both the position x and the momentum p, thus can not be the
candidate of a wave function in quantum mechanics. We need a way to eliminate “half” of the
variables. The notion of polarization serves for this purpose.
Let T C M = T M ⊗R C denote the complexification of the tangent bundle T M . Let

: T C M −→ T C M
X 7−→ X

denote the complex conjugate. The symplectic form ω extends C-linearly to a non-degenerate
pairing on T C M and we still denote it by ω.
Assume dim(M ) = 2n. For any z ∈ M , a C-linear subspace P ⊂ TzC M is called a
Lagrangian subspace if dimC P = n and

ω(X, Y ) = 0, ∀X, Y ∈ P.

Definition 3.5.6. A polarization of (M, ω) is a complex subbundle P ⊂ T C M such that

1 for any z ∈ M , Pz is a Lagrangian subspace of TzC M

⃝

⃝
2 integrability: for any open subset U ,

X, Y ∈ Γ(U, P ) =⇒ [X, Y ] ∈ Γ(U, P )

Here [X, Y ] is the Lie bracket of two vector fields.

⃝
3 dim(Pz ∩ P z ) is constant for z ∈ M .

Condition ⃝
3 is a technical assumption for the purpose of certain constructions.

We will be mainly interested in the following two types of polarizations:

a) Real Polarization: Pz = P z , ∀z ∈ M

b) Kähler Polarization: Pz ∩ P z = {0}, ∀z ∈ M

175
Definition 3.5.7. Let (L , ∇) be a prequantum line bundle on (M, ω) and P be a polarization.
A section s ∈ Γ(M, L ) is called P -polarized if

∇X s = 0

for any vector field X in P .

Let us check the consistency of the above condition. Let X, Y be two vector fields in P .
Assume ∇X s = ∇Y s = 0 holds, then

[∇X , ∇Y ]s = 0.

On the other hand,

[∇X , ∇Y ] =∇[X,Y ] + F∇ (X, Y )
i
=∇[X,Y ] −
ω(X, Y ).
h̄
Since P is a polarization, [X, Y ] is still a vector in P and ω(X, Y ) = 0. This shows the
consistency of the defining equation for polarized sections.
Thus we can define the quantum Hilbert space associated to a prequantum line bundle L
and a polarization P by

L2 (M, L )P := s ∈ L2 (M, L ) | s is P -polarized .

Example 3.5.8. Consider the phase space R2n with the following prequantum line bundle

• L is the trivial line bundle on R2n with Hermitian inner product

hf, gi = f g

• The connection ∇ is
i
∇=d−θ
h̄
P
where θ = pi dxi is the Liouville 1-form. The curvature of ∇
i

i i
F∇ = − dθ = − ω
h̄ h̄
satisfies the prequantization condition.

We consider the polarization P spanned by vector fields along momentum directions

 
∂
P = Span .
∂pi
For any function f (x, p),
 
∂f i ∂ ∂f
∇ ∂ f= − θ f= .
∂pi ∂pi h̄ ∂pi ∂pi
Thus f is P -polarized if and only if
f = f (x)

176
does not depend on p. This gives the expected result for wave functions.
If we consider another polarization P ′ spanned by vector fields along position directions
 
′ ∂
P = Span .
∂xi

Since  
∂f i ∂ ∂f i
∇ ∂ f= i
− θ f= − pi f,
∂x i ∂x h̄ ∂xi ∂x i h̄
f is P ′ -polarized if and only if f is of the form
i
f = e h̄ x·p g(p).

Example 3.5.9. Consider the phase space Cn with Kähler symplectic form

¯
ω = i∂ ∂K.

K is a real function called the Kähler potential. We choose the prequantum line bundle L to
be the trivial line bundle with Hermitian inner product

hf1 , f2 i = f 1 f2 e− h̄
K

and compatible connection

1
∇=d− ∂K.
h̄
Its curvature
1 1¯ 1 ¯ i
F∇ = − d(∂K) = − ∂∂K = ∂ ∂K =− ω
h̄ h̄ h̄ h̄
satisfies the prequantization condition.
We choose the Kähler polarization spanned by vector fields of (0, 1)-type
 
∂ ∂
P = Span ,··· , n .
∂z 1 ∂z

Since
∂f
∇ ∂ f=
∂z j
∂z j

f is P -polarized if and only if f = f (z) is holomorphic. The resulting quantum Hilbert space
consists of holomorphic functions f (z) such that
ˆ n
K ω
|f (z)|2 e− h̄ < ∞.
Cn n!

For K = |z|2 , we find precisely the Segal-Bargmann space.

Example 3.5.10. We generalize the above construction to a complex manifold M with a her-
mitian holomorphic line bundle L . Let ∇ be the compatible connection on L such that

∇ = ∇1,0 + ∇0,1 ¯
where ∇0,1 = ∂.

i
Assume the first Chern class 2π F∇ defines a symplectic form on M .

177
 We consider the phase space (M, ω) where

ω = ih̄F∇ .

By construction, (L , ∇) defines a prequantum line bundle. We choose the polarization

P = T 0,1 M

spanned by vector fields of type (0, 1). In local holomorphic coordinates z i , P is spanned by
n o
∂
∂z i
. Thus P -potential sections are holomorphic sections of L .
Assume M is compact. Then the space H 0 (M, L ) of holomorphic sections is finite dimen-
sional. Thus we obtain a finite dimensional quantum Hilbert space

H = H 0 (M, L ).

Such Hilbert space can be used to describe internal degrees of freedom in a quantum system.
As an example, we consider

M = CP 1 = lines in C2 .

It is covered by two holomorphic charts

z ∈ U = C, w∈V =C

which are glued along C∗ by identifying

1
z= .
w
We consider the holomorphic line bundle

L = OCP 1 (k), k ∈ Z>0

whose transition function from U to V is wk on the intersection U ∩ V . A holomorphic section

is given by a pair (f (z), g(w)) where

• f (z) is holomorphic on U

• g(w) is holomorphic on V

• f (z)wk = g(w) on U ∩ V

We find f (z) can only be a polynomial of deg ≤ k, and g(w) = f (w−1 )wk . Thus
n o
H 0 (CP 1 , L ) = SpanC 1, z, · · · , z k

which has dim = k + 1. It is precisely the spin k/2 representation of su(2).

We can also write down the symplectic form explicitly. In the holomorphic chart U ,
dz ∧ dz
ω = ikh̄ .
(1 + |z|2 )2

178
Example 3.5.11. We continue with our previous example but change a point of view. Let
(L0 , ∇0 ) be a fixed Hermitian holomorphic line bundle on the complex manifold M whose first
Chern class gives a symplectic form
i
ω= F∇ .
2π 0
Let us consider the k-th tensor L = L0⊗k with induced connection ∇ from ∇0 . Then

F∇ = kF∇0 = −i(2πk)ω.

Thus (L , ∇) defines a prequantum line bundle if we identify

1
h̄ =
2πk
In particular, the semi-classical limit h̄ → 0 corresponds to k → ∞.
Assume M is compact, so the quantum Hilbert space H 0 (M, L ) is finite dimensional. Let
X
χ(M, L ) := (−1)i dim H i (M, L )
i

be the Euler characteristic of L . By Hirzebruch Riemann Roch Theorem,

ˆ
χ(M, L ) = ch(L ) td(M )
M

where td(M ) is the Todd class of the tangent bundle of X. In the case above where L0 is a
positive line bundle, we have

H i (M, L0⊗k ) = 0 for i > 0 and k sufficiently large

Thus in the k → ∞ limit, we have the leading behavior for L = L0⊗k

ˆ ˆ
i
dim H (M, L ) =
0 F ∇
e 2π td(M ) = ekω td(M )
M
ˆ M
ωn
∼k n n
= k Volω (M )
M n!
1
where 2n = dim M and Volω (M ) is the Liouville volume of M . Under the identification h̄ = 2πk ,
we find the semi-classical behavior
1
dim H 0 (M, L ) ∼ Volω (M ).
(2πh̄)n
This leads to the semi-classical intuition that the number of quantum states is approximately
the number of cells in phase space measured in unite of h̄.

3.5.5 Quantum Operators

Let (M, ω) be a quantizable symplectic manifold, with a chosen prequantum line bundle
(L , ∇) and a polarization P ⊂ T C M . Next we consider how to quantize functions f ∈ C ∞ (M )
to operators on the quantum Hilbert space L2 (M, L )P .
There is a special class of functions that are naturally quantized. These are functions that
preserve the polarization P .

179
Definition 3.5.12. A function f is called P -preserving if its Hamiltonian vector field Vf satisfies

[Vf , P ] ⊂ P

i.e., for any vector field X in P , the Lie bracket [Vf , X] is still a vector field in P . The collection
of P -preserving functions will be denoted by C ∞ (M )P .

Proposition 3.5.13. Let ψ ∈ Γ(M, L ) be a P -polarized section of L , and f ∈ C ∞ (M )P be a

P -preserving function. Then Q(f )ψ is also a P -polarized section of L .

Proof: Recall the prequantization map

Q(f ) = ih̄∇f + f.

Let X be any vector field in P . Then

[∇X , Q(f )] =[∇X , ih̄∇f + f ]

=ih̄∇[X,Vf ] + ih̄F∇ (X, Vf ) + X(f )
=ih̄∇[X,Vf ] + ω(X, Vf ) + X(f )
=ih̄∇[X,Vf ] − ιX ιVf ω + X(f )
=ih̄∇[X,Vf ] − ιX df + X(f )
=ih̄∇[X,Vf ] .

By assumption, [X, Vf ] is still a vector field in P . Thus for a P -polarized section ψ,

∇X Q(f )ψ =[∇X , Q(f )]ψ since ∇X ψ = 0

=ih̄∇[X,Vf ] ψ
=0 since [X, Vf ] lies in P .

Therefore Q(f )ψ is also a P -polarized section.

Thus for a P -preserving function f , we can simply quantize it to the operator Q(f ) on the
quantum Hilbert space.
For functions that are not P -preserving, their quantization is more tricky. The basic idea
is that their Hamiltonian flows will change the polarization, hence will change a P -polarized
state to a state that is polarized with a different polarization. Then one can construct a
map relating quantum Hilbert spaces associated to different polarizations. This is the idea of
Blattner-Kostant-Sternberg (BKS) kernel construction. We will not go into further details here.

Example 3.5.14. Consider the phase space

X
R2n , ω= dpi ∧ dxi
i

with prequantum line bundle

i X
∇=d− θ, θ= pi dxi
h̄
i

180
and polarization  
∂
P = Span .
∂pi
Consider a function f (x, p) whose Hamiltonian vector field is
X ∂ ∂

Vf = ∂ xi f − ∂ pi f i .
∂pi ∂x
i

From   X
∂ ∂ X ∂
, Vf = (∂pk i ∂xi f ) − (∂pk ∂pi f ) i
∂pk ∂pi ∂x
i i

we see that Vf preserves P if and only if

∂pk ∂pi f = 0, ∀i, k.

 
∞ 2n
P
Thus C (R )P = a(x) + bi (x)pi consists of functions at most linear in pi . Such a
P i
function f = a(x) + bi (x)pi is quantized to
i
  X
i ∂
Q(f ) =ih̄ Vf − θ(Vf ) + f = −ih̄ bi (x) i + a(x)
h̄ ∂x
i

as an operator acting on wave functions ψ = ψ(x).

Example 3.5.15. Consider the phase space

Cn , ¯
ω = i∂ ∂K

where K is the Kähler potential. We choose the prequantum line bundle L to be the trivial line
bundle with Hermitian inner product by

hf1 , f2 i = f 1 f2 e− h̄
K

and compatible connection by

1
∇=d− ∂K.
h̄
Its curvature
1 1¯ 1 ¯ i
F∇ = − d(∂K) = − ∂∂K = ∂ ∂K =− ω
h̄ h̄ h̄ h̄
satisfies the prequantization condition. We choose the Kähler polarization
 
∂
P = Span
∂z i

so polarized sections of L are holomorphic functions.

Let us write the Kähler form as
X
ω= gjm dz j ∧ dz m
j,m

181
where gjm = i∂j ∂m K. Let g mj denote the inverse matrix of gjm , i.e.,
X
gim g mj = δi j .
m

Let f be a function on Cn . Its Hamiltonian vector field is given by

X ∂ ∂

Vf = g ∂m f j − g ∂j f m
mj mj
∂z ∂z
j,m

∂ ∂
where ∂m = ∂z m and ∂j = ∂z j
. Since
  X
∂ ∂ mj

∂
, V f = g ∂ m f modulo P
∂z i j,m
∂z i ∂z j

we see that f is P -preserving if and only if

X ∂  
mj ∂
g f = 0, ∀i, j.
∂z i ∂z m
m

Now we specialize to the case when

K = |z|2 , gjm = iδjm .

Then the condition for P -preserving becomes

∂ ∂
f = 0, ∀i, j
∂z i ∂z j
i.e., f is at most linear in z i ’s. If we further require real condition, then the only real P -
preserving functions are of the form

f = a + bk z k + bk z k + ckj z k z j ,

where a ∈ R, bk , ckj ∈ C and ckj = cjk . The Hamiltonian vector of f is

X ∂ ∂

Vf = −i ∂k f k − ∂k f k .
∂z ∂z
k

Thus f is quantized to
 
1
Q(f ) =ih̄ Vf − ∂K(Vf ) + f
h̄
X ∂ X ∂
=h̄ ∂k f k − zk k f + f
∂z ∂z
k k
X X  ∂
=a + bk z k + h̄ bk + cjk z j
∂z k
k k

as an operator acting on holomorphic functions.

182
 This makes holomorphic representation particularly useful for quantization of quadratic
Hamiltonians. For example, consider one-dimensional Harmonic oscillator with Hamiltonian

H = |z|2 on C.

The above Kähler polarization quantizes H to the operator

∂
Q(H ) = h̄z
∂z
on the Segal-Bargmann space. Note that

Q(H )z n = nh̄z n ,

thus eigenvalues of Q(H ) are non-negative integers. However, this is not the correct spectrum:
the Maslov correction is missing.
This is the same problem that we have encountered in our discussion on semi-classical
quantization. The way out is to modify our quantum Hilbert space by including half-densities.
In the Kähler quantization, this modification amounts to define a quantum state ψ to be a
√ √
section of L ⊗ KM where KM is the canonical line bundle and KM is a square root of it.
In the Harmonic oscillator example above, this modification amounts to change the state
√
z n −→ z n dz

1

∂
which will then have the correct eigenvalue n + 2 h̄ under Q(H ) = h̄z ∂z .

3.6 Path Integral in Phase Space

Classical Hamiltonian mechanics is captured by the following action in the phase space
ˆ ˆ


S = (pdq − H dt) = pi (t)q̇ i (t) − H (p(t), q(t)) dt.

Here {pi (t), q i (t)} represents a path of classical particle in the phase space and t is the time
variable. Classical trajectories are extremals of this action and can be obtained via functional
variation (with fixed boundary values of q i )
ˆ   
d ∂H ∂H i
δS = δpi q̇ + pi (δq ) −
i i
δpi + δq dt
dt ∂pi ∂q i
ˆ  
∂H ∂H i
= δpi q̇ − ṗi δq −
i i
δpi − δq dt
∂pi ∂q i
ˆ     
∂H ∂H
= q̇ −
i
δpi − ṗi + i
δq dt.
∂pi ∂q i
Requiring the extremal condition δS = 0, we obtain the Hamilton’s equations

 ∂H
 i
q̇ = ∂p
i

 ∂H
ṗi = − i
∂q

183
 Quantum mechanically, transition amplitudes and correlation functions are captured by
the form of path integral ˆ
i
[DqDp] e h̄ S O.

In this section, we will discuss perturbative approach to such path integral in the asymptotic
h̄ → 0 limit in terms of Feynman’s combinatorial formula of graph expansion.

3.6.1 Wick’s Theorem

Gaussian Integral

We start with the finite dimensional situation. Let

X
n
Q(x) = Qij xi xj
i,j=1

be a positive definite quadratic form, i.e., the matrix (Qij ) is positive definite. Consider the
following volume form on Rn

p Yn
dxi − 1 Q(x)
Ω= det Q √ e 2h̄ , h̄ > 0.
i=1
2πh̄
´ −2x
1 2 √
By the Gaussian integral R dx e = 2π, it is easy to find
ˆ
Ω = 1.
Rn

Thus Ω defines a probability density on Rn .

Consider for simplicity a polynomial function f (x) ∈ R[xi ]. Define the following expectation
value with respect to the probability density Ω:
´ ˆ
n Ωf (x)
R´
hf (x)ix := = Ωf (x).
Rn Ω Rn

The subscript x indicates the integration variable. This expectation value defines a map

h−ix : R[xi ] −→ R.

To compute this expectation map, let us consider the auxiliary integral

 P i 
x Ji
Z[J] := e i J = (J1 , · · · , Jn )
x
p ˆ Yn P
dxi − 2h̄
1
Q(x)+ xi Ji
= det Q √ e i

Rn i=1 2πh̄

Completing the square, we find

1 X 1 X h̄
Q(x) − xi Ji = Q(xi − h̄ (Q−1 )ij Jj ) − Q−1 (J).
2h̄ 2h̄ 2
i j

184
Here (Q−1 )ij is the inverse matrix of Qij
X
Qik (Q−1 )kj = δi j
k

and Q−1 (J) is the quadratic expression

X
Q−1 (J) = (Q−1 )ij Ji Jj .
i,j

Thus ˆ
h̄ −1 h̄ −1 (J)
Q (J)
Z[J] = e 2 Ω = e2Q .
Rn
Now for a polynomial f (x), Taylor series expansion at x = 0 gives
P ∂
xi
∂ai
f (x) = e i f (a) .
a=0

Using this formula, we find

 P ∂ 
xi
∂ai
hf (x)ix = ei f (a)
a=0 x
∂
 
J= ∂a ∂ h̄ −1 ∂
===== Z f (a) = e 2 Q ( ∂a ) f (a)
∂a a=0 a=0

P
where Q−1 ∂
∂a = (Q−1 )ij ∂a
∂ ∂
i ∂aj is a second-order differential operator.
i,j

Wick’s Theorem

The obtained formula

h̄ −1 ∂
hf (x)ix = e 2 Q ( ∂a ) f (a)
a=0

has a combinatorial interpretation as follows. Let

i1 , i2 , · · · , i2m ∈ {1, 2, · · · , n}

be 2m indices. We consider the monomial xi1 xi2 · · · xi2m of degree 2m obtained from the index
set. Let us compute its expectation from the above formula:
 m
h̄ −1 ∂ h̄ m
1 X ∂ ∂
xi1 xi2 · · · xi2m x = e 2 Q ( ∂a ) ai1 ai2 · · · ai2m =  (Q−1 )ij i j  ai1 ai2 · · · ai2m
a=0 m! 2 ∂a ∂a
i,j
(*)
To compute this value, let

P (2m) = set of partitions of {1, 2, · · · , 2m} into unordered m pairs.

An element σ ∈ P (2m) can be described by a permutation σ ∈ S2m such that

σ(1) < σ(2), σ(3) < σ(4), ··· , σ(2m − 1) < σ(2m)

185
and
σ(1) < σ(3) < · · · < σ(2m − 3) < σ(2m − 1).

Then we identify this permutation

" #
1 2 3 ··· 2m − 1 2m
σ(1) σ(2) σ(3) · · · σ(2m − 1) σ(2m)

with an element of P (2m) by pairing

σ(1)σ(2) σ(3)σ(4) ··· σ(2m − 1)σ(2m)

| {z } | {z } | {z }
pair pair pair

In the above formula (*), it is computed to become the following sum

X
xi1 xi2 · · · xi2m x = h̄m (Q−1 )σ(1)σ(2) (Q−1 )σ(3)σ(4) · · · (Q−1 )σ(2m−1)σ(2m) .
σ∈P (2m)

This is called the Wick’s Theorem.

We can also draw each sum as
i1 i2 i3 ··· i2m−1 i2m

For each edge connecting i and j, we assign the factor h̄(Q−1 )ij to it.

Example 3.6.1 (Two-point function). The two-point function

xi xj x
= h̄(Q−1 )ij

is given by the inverse matrix of Q. (Q−1 )ij is also called the “propagator”.
i j

Example 3.6.2. The following four-point function is computed by

D E
xi xj xk xl =ijkl + ijkl + ijkl
x
h i
=h̄2 (Q−1 )ij (Q−1 )kl + (Q−1 )ik (Q−1 )jl + (Q−1 )il (Q−1 )jk .

We can also add “background” by considering the shift x → x + a

ˆ
hf (x + a)ix = Ω f (x + a)
Rn

which is now a function of a. A similar argument shows

h̄ −1 ∂
hf (x + a)ix = e 2 Q ( ∂a ) f (a).

If f (x) = xi1 xi2 · · · xik as above, then Wick’s Theorem in this case is a sum over partial pairings

186
 i1 i2 i3 ··· ik−1 ik

We assign the propagator to the paired indices and assign ai to unpaired index i.

Example 3.6.3. Let f (x) = xi xj xk xl . Then

hf (x + a)ix =ijkl + ijkl + ijkl + ijkl + ijkl + ijkl + ijkl + ijkl + ijkl + ijkl
h
=ai aj ak al + h̄ (Q−1 )ij ak al + (Q−1 )ik aj al + (Q−1 )il aj ak
i
+ (Q−1 )jk ai al + (Q−1 )jl ai ak + (Q−1 )kl ai aj
h i
+ h̄2 (Q−1 )ij (Q−1 )kl + (Q−1 )ik (Q−1 )jl + (Q−1 )il (Q−1 )jk .

3.6.2 Feynman Graph Expansion

Now we consider integrals of the form

p ˆ Y n
dxi − 1 S(x)
det Q √ e h̄
Rn i=1 2πh̄

where
1 λ
S(x) = Q(x) − I(x).
2 3!
P i j
Here Q(x) = Qij x x is a positive quadratic as before, and
i,j
X
I(x) = Iijk xi xj xk
i,j,k

is a cubic polynomial, called the “interaction”. The constants Iijk parametrize the cubics. The
constant λ is called the “coupling constant”.
Since the cubic approaches both ±∞ and grows faster than quadratic, the above integral
is simply “divergent”. There are essentially two ways out to make sense of it:

⃝
1 Complexify xi to complex variables z i and change the integration contour

Rn ⊂ C n =⇒ Γ ⊂ Cn

to some other contour Γ such that the integration becomes convergent. “Airy integral” is
such an example. This method is usually referred to as the non-perturbative method.

⃝
2 Treat λ as a perturbative parameter and compute the asymptotic series. This method is
usually referred to as the perturbative method.

We will focus on the perturbative method ⃝2 here. Let us rewrite

p ˆ Y n ˆ D λ E
dxi − 1 S(x) λ
“ det Q √ e h̄ ”=“ Ω e 3!h̄ I(x) ” = “ e 3!h̄ I(x) ”.
Rn i=1 2πh̄ Rn x

187
 √ Q
n
√dx e− 2h̄ Q(x)
i 1
Here Ω = det Q 2πh̄
is the probability density as before. Now we can redefine
i=1
the above divergent integral as a power series by
X∞  m 
λm 1
I(x) .
m!h̄m 3! x
m=0

1

m
Here each term 3! I(x) x
has a well-defined value which can be computed by Wick’s The-
orem. This will lead to a combinatorial graph formula for this power series.

Definition 3.6.4. By a graph γ, we refer to the following data

• V (γ) = set of vertices

• HE(Γ) = set of half-edges

• iΓ : HE(Γ) → V (γ) incidence map

• E(Γ): a perfect matching on HE(Γ) into pairs of two elements. Each pair is called an
“edge”.

• For each v ∈ V (Γ), # i−1
Γ (v) is called the valency of v.

Example 3.6.5. Consider the following Θ graph

a a′
b b′
Γ: 1 2
c c′

Here V = {1, 2}, HE = {a, b, c, a′ , b′ , c′ }, E = {(a, a′ ), (b, b′ ), (c, c′ )}. The incidence map is

iΓ : HE −→ V
{a, b, c} 7−→ 1
{a′ , b′ , c′ } 7−→ 2

Each vertex 1 and 2 has valency 3.

Definition 3.6.6. A graph isomorphism between two graphs Γ and Γ′ is a pair of bijections

σV : V (Γ) −→ V (Γ′ )
σHE : HE(Γ) −→ HE(Γ′ )

which are compatible with incident maps

σHE
HE(Γ) HE(Γ′ )
iΓ i′Γ diagram commutes
σV
V (Γ) V (Γ′ )

and compatible with edges for any a, b ∈ HE(Γ),

(a, b) ∈ E(Γ) ⇐⇒ (σHE (a), σHE (b)) ∈ E(Γ′ ).

188
 An automorphism of Γ is an isomorphism σ : Γ → Γ. Denote

Aut(Γ) = Group of automorphisms of Γ.

Example 3.6.7. Consider

a a′
b b′
Γ: 1 2
c c′

Here HE = {a, b, c, a′ , b′ , c′ }, V = {1, 2}, E = {(a, a′ ), (b, b′ ), (c, c′ )}. The incidence map is
as above. Here are two examples of automorphisms of Γ

⃝
1 σV : (1, 2) −→ (2, 1)
σHE : (a, b, c, a′ , b′ , c′ ) −→ (a′ , b′ , c′ , a, b, c)

⃝
2 σV : (1, 2) −→ (1, 2)
σHE : (a, b, c, a′ , b′ , c′ ) −→ (b, c, a, b′ , c′ , a′ )

Similarly, we find Aut(Γ) = Z2 × S3 . Here Z2 corresponds to permuting the two vertices, and
S3 correponds to permuting the three edges.

Example 3.6.8. Consider

Γ:

We find Aut(Γ) = Z2 × Z2 × Z2 .

Now let us consider the computation of

hI(x)m ix = h I(x)I(x) · · · I(x) i x

| {z }
m

Each I(x) gives a vertex with 3 half-edges

i1 i2 im

1 2 ··· m

j1 k1 j2 k2 jm km

Vertex set V = {1, 2, · · · , m}

Half-edge set HE = {i1 , j1 , k1 , i2 , j2 , k2 , , · · · , im , jm , km , }
Incidence map i : {is , js , ks } −→ {s}

189
By Wick’s Theorem, hI(x)m ix is a sum over perfect matching E
X
hI(x)m ix = h̄|E(ΓE )| ωΓE .
E∈Y

Here Y = {set of perfect matchings of HE} and ΓE is the graph by assigning the pairing E to
the vertex set V and the half-edge set HE.
For each graph Γ, ωΓ is the number by assigning
i

⃝
1 each vertex j k =⇒ Iijk

⃝
2 each edge i j =⇒ (Q−1 )ij

⃝
3 sum over indices

Example 3.6.9. Consider

a a′
b b′
Γ: 1 2
c c′
X
ωΓ = Ii1 j1 k1 Ii2 j2 k2 (Q−1 )i1 i2 (Q−1 )j1 j2 (Q−1 )k1 k2
i1 ,j1 ,k1
i2 ,j2 ,k2

It is clear that

ωΓ = ωΓ′ if Γ is isomorphic to Γ′ .

All graphs here are trivalent (valency 3 for each vertex). The set of bijections

σV : V −→ V
σHE : HE −→ HE

that preserve the incidence map is

G= Sm × (S3 )m .
|{z} | {z }
permuting vertices permuting half-edges

We have a natural G-action on Y with orbits

Y /G = {isomorphic class of trivalent graphs consisting of m-vertices} .

Thus
1 1 X |G|
hI(x)m ix = ωΓ h̄|E(Γ)|
m!(3!) m |G| | Aut(Γ)|
Γ∈Y /G
X ωΓ
= h̄|E(Γ)|
| Aut(Γ)|
Γ: trivalent graph
with m vertices
We can also refine this formula by grouping disconnected graphs into connected isomorphic
ones. This leads to the following Feynman graph expansion formula (details left to the reader).

190
Proposition 3.6.10 (Feynman Graph Expansion Formula). We have the following combinato-
rial graph expansion formula for the asymptotic power series
D λ E X∞  m 
I(x) λm 1
e 3!h̄ := I(x)
x m!h̄m 3! x
m=0
X ωΓ
= h̄|E(Γ)|−|V (Γ)| λ|V (Γ)|
| Aut(Γ)|
Γ: trivalent
 
 X ωΓ 
= exp  λ|V (Γ)| h̄l(Γ)−1 .
| Aut(Γ)|
Γ: connected
trivalent

Here for a connected graph Γ,

l(Γ) = |E(Γ)| − |V (Γ)| + 1 = 1 − χ(Γ)

is called the loop # of Γ.

l(Γ) = 3 − 2 + 1 = 2 l(Γ) = 6 − 4 + 1 = 3

In general, if the interaction contains terms of all possible degrees

λ3 λ4 λm
I3 (x) + I4 (x) + · · · + Im (x) + · · ·
3! 4! m!
then the series expension will have all possible graphs, where each vertex of valency m is
associated the value Im similarly. The general Feynman graph expansion formula reads
!  
* 1 P λm +  
I (x)
m! m asymptotic  X Y ωΓ 
======== exp   m (Γ)|  l(Γ)−1 
h̄
e m≥3  λ|V h̄
m
| Aut(Γ)| 
x Γ: connected m≥3
graph

where Vm (Γ) is the set of vertices with valency m.

3.6.3 Weyl Quantization Revisited

Let us apply the combinatorial formula of Feynman graph to the quantum mechanical
situation. The relevant space consists of paths on the phase space parametrized by the time.
For simplicity, we consider the phase space R2 . The discussion generalizes easily to R2n .
We use γ(t) = (X(t), P(t)) to represent a particle trajectory on the phase space R2 . Thus
γ can be viewed as a map
γ : R −→ R2
t 7−→ (X(t), P(t))
Let us denote
E = Map(R, R2 )

191
 The action S can be viewed as a function on E (with a suitable domain which is not relevant
in our current discussion)
S : E −→ R
ˆ  
S[γ] = PẊ − H dt
We are interested in the following expectation
´ i
E [Dγ] e
h̄
S[γ]
O[γ]
hOiγ = ´ i
S[γ]
E [Dγ] e
h̄

where O is some appropriate function on E and is called an observable. Let us first consider the
free case without Hamiltonian H , i.e.,
ˆ
S[γ] = S0 [γ] = PẊ dt.

As we will see, the zero Hamiltonian case has a topological nature.

The above integration is on the infinite-dimensional path space E, thus not well-defined
in the first space. In Chapter 2, we have treated path integral as a limit of finite dimensional
appproximation. When time is imaginary, this is related to Wienner measure integration.
In this section, we use an alternate perturbative approach. It turns out that we can borrow
the finite dimensional result and design the parallel combinatorial formula into the infinite-
dimensional setting. Let us see how this works in practice.
Let us first compare the finite-dimensional with the infinite-dimensional setting.
Finite Dim Infinite Dim
space Rn E = Map(R, R2 )
variable xi γ(t)
index i ∈ {1, 2, · · · , n} t∈R
P ´
sum R dt
i
P ´
action Q(x) = Qij x i xj S0 [γ] = PẊ dt
i,j
− 2h̄
1
Q(x) i
density e e h̄ S0 [γ]
In the finite-dimensional case, the two point function

xi xj x
= h̄(Q−1 )ij

plays the fundamental role as building block in computing general expectation values. It is
called the “propagator” and is given by the inverse matrix of the quadratic pairing Q.
The infinite-dimensional situation looks exactly the same. The free action
ˆ
S0 [γ] = PẊ dt

is quadratic in γ = (X, P). Let us write (ignore the boundary behavior so far)
ˆ
1  
S0 [γ] = PẊ − XṖ dt
2
ˆ " #" #
1 h i − dt
d
X
= X P d dt
2 Pdt

192
Thus " #" # !!
ˆ h i 0
i 1 1 − dt
d
X(t)
e S [γ]
h̄ 0 = exp − X(t) P(t) dt .
2h̄ i d
dt 0 P(t)
Comparing with the finite-dimensional situation
P
− 2h̄
1
Qij xi xj
e i,j

the analogue of the inverse matrix Q−1 is

" #−1
0 − dt
d
i d
dt 0

In analysis, the inverse of a differential operator is usually represented by an integral kernel

d
called the Green’s function. The Green’s function of dt is a function G(t1 , t2 ) satisfying

∂
G(t1 , t2 ) = δ(t1 − t2 ).
∂t1
Here δ(t1 − t2 ) can be viewed as the infinite-dimensional analogue of δij .
In terms of G(t1 , t2 ), we have
" #" # " #
0 − ∂t∂1 0 G(t1 , t2 ) δ(t1 − t2 ) 0
=
∂
∂t1 0 −G(t1 , t2 ) 0 0 δ(t1 − t2 )

The right hand side can be viewed as the (t1 , t2 )-entry of the infinite-dimensional identity matrix.
Thus the (t1 , t2 )-entry of the above inverse can be written as
" #−1 " #
0 − dt
d
0 iG(t1 , t2 )
i = .
d
dt 0 −iG(t1 , t2 ) 0
t1 ,t2

Then we could “define” the two-point functions in terms of this inverse by



 hX(t1 )P(t2 )iγ := ih̄G(t1 , t2 )


hP(t1 )X(t2 )iγ := −ih̄G(t1 , t2 )




hX(t1 )X(t2 )iγ = hP(t1 )P(t2 )iγ = 0

We are left to solve G(t1 , t2 ). The above expression asks for G(t1 , t2 ) such that

G(t1 , t2 ) = −G(t2 , t1 ).

d
Such Green’s function of dt is explicitly given by



1

 t1 > t 2
1 2
G(t1 , t2 ) = sgn(t1 − t2 ) = 0 t1 = t 2
2 



− 1 t1 < t 2
2

193
 G

1
2

t1 − t2
− 12

In fact, for any compactly supported test function f (t)

ˆ ˆ ˆ
′ 1 t2 ′ 1 ∞ 1 1
− dt1 G(t1 , t2 )f (t1 ) = dt1 f (t1 ) − dt1 f ′ (t1 ) = f (t2 ) − (−f (t2 )) = f (t2 ).
R 2 −∞ 2 t2 2 2

So
∂
G(t1 , t2 ) = δ(t1 − t2 )
∂t1
holds as a distributional equation.
In summary, we are interested in the following two-point function
´ i
[Dγ] e h̄ S0 [γ] X(t1 )P(t2 )
hX(t1 )P(t2 )iγ “ = ” ´ i
[Dγ] e h̄ S0 [γ]

Although we do not know a prior the precise information about the path integral measure, we
can still derive a reasonable result in comparison with the finite dimensional Gaussian integral
ih̄
hX(t1 )P(t2 )iγ := ih̄G(t1 , t2 ) = sgn(t1 − t2 ).
2
In the following, we use this formula to define our two-point function.
Note that this two-point function, or the propagator, does not depend on the precise value
of t1 , t2 , but only on the relative position



ih̄

 t1 > t 2
2
hX(t1 )P(t2 )iγ = 0 t1 = t2




− ih̄ t1 < t 2
2
This indicates a topological nature of this model, which is indeed the case.
Now let f (x, p) and g(x, p) be two polynomials . We plot them on the time line as
g f

t1 t2 t

We define a new function on the phase space by the following correlation

hf (x + X(t2 ), p + P(t2 ))g(x + X(t1 ), p + P(t1 ))iγ .

194
Here we perturb the variables x, p by quantum fluctuations X(t), P(t) at time t1 for g and at time
t2 for f . Equivalently, we can treat x and p as “background” shift. This correlation function
depends on the background variable {x, p} and defines a function on the phase space.
Let us apply Wick’s Theorem to compute this correlation. We represent each f and g as a
vertex whose valency is the polynomial degree

.. ..
. .

f g

Then Wick’s Theorem says

hf (x + X(t2 ), p + P(t2 ))g(x + X(t1 ), p + P(t1 ))iγ

x, p x, p
P ..
= x, p .
f g x, p
x, p

Here for each unpaired half-edge, we assign x or p as in the original vertex. For paired half-edges

X(t2 ) P(t1 )

=⇒ hX(t2 )P(t1 )iγ = i

2 h̄
f g

P(t2 ) X(t1 )

=⇒ hP(t2 )X(t1 )iγ = − 2i h̄

f g

For a graph with m propagators, there will be an automorphism group Sm by permuting

1
the edges, contributing m! in the Feynman graph expansion. Thus we find

hf (x + X(t2 ), p + P(t2 ))g(x + X(t1 ), p + P(t1 ))iγ

← → ← →
i
h̄ ∂ ∂
− ∂ ∂

g(x, p) = (f ∗h̄ g) (x, p)

2 ∂x ∂p ∂p ∂x
=f (x, p)e

which is precisely the Moyal product.

We can think about
g f

t1 t2 t

195
as inserting an operator g at time t1 , and then inserting another operator f at a later time t2 .
The total effect is the composition of two operators, which is represented quantum mechanically
by the Moayl product f ∗h̄ g.
This picture also gives a simple explanation of the associativity of the Moyal product ∗h̄ .
Given three functions f, g, h, we insert them in time order
h g f

t1 t2 t3 t

Consider the correlation

hf (x + X(t3 ), p + P(t3 ))g(x + X(t2 ), p + P(t2 ))h(x + X(t1 ), p + P(t1 ))iγ .

This value is invariant under local deformations of the time positions and only the order is
relevant. By our previous computation

(f ∗h̄ g)∗h̄ h = lim lim hf (x + X(t3 ), p + P(t3 ))g(x + X(t2 ), p + P(t2 ))h(x + X(t1 ), p + P(t1 ))iγ
t1 →t− +
2 t3 →t2

f ∗h̄ (g ∗h̄ h) = lim lim hf (x + X(t3 ), p + P(t3 ))g(x + X(t2 ), p + P(t2 ))h(x + X(t1 ), p + P(t1 ))iγ
−
t3 →t+
2 t1 →t2

The topological nature of the above correlation implies

(f ∗h̄ g) ∗h̄ h = f ∗h̄ (g ∗h̄ h) .

3.6.4 S 1 -Correlation

Let us now introduce topology and put our quantum mechanical model on S 1


γ(θ) = Xi (θ), Pi (θ) : S 1 −→ R2n .

Here θ ∈ S 1 is the angle variable and we identify

θ ∼θ+1

γ
R2n

S1

S 1 -Propagator

We consider the free action as before

ˆ
dXi
S0 [γ] = Pi (θ)Ẋi (θ) dθ, Ẋi (θ) = .
S1 dθ

196
We use Feynman graph expansion to define correlation function in this model. The propagator
is given by the two-point function
D ES 1
Xk (θ1 )Pj (θ2 ) = ih̄G(θ1 , θ2 )δjk
γ


d −1
where G(θ1 , θ2 ) is the Green’s function representing the integral kernel of the operator “ dθ ”.
On S 1 , we have a Fourier basis of functions
n o
e2πinθ .
n∈Z

The δ-function can be represented by

X
δ(θ1 , θ2 ) = e2πin(θ2 −θ1 ) .
n∈Z
d
Note that for n = 0, the constant function “1” can not be inverted by “ dθ ”. This is called
the zero mode, which is related to the nontrivial harmonics on S 1 . Let
X
b 1 , θ2 ) =
δ(θ e2πin(θ2 −θ1 )
n∈Z\{0}

where we have deleted the zero mode. The Green’s function G(θ1 , θ2 ) solves
∂ b 1 , θ2 )
G(θ1 , θ2 ) = δ(θ
∂θ1
X i 2πin(θ2 −θ1 )
=⇒ G(θ1 , θ2 ) = e = g(θ2 − θ1 )
2πn
n∈Z\{0}

where g(θ) is the following function on S 1

X i 2πinθ
g(θ) = e .
2πn
n∈Z\{0}

Proposition 3.6.11. Viewed as a periodic function on R with g(θ + 1) = g(θ), the value of
g(θ) on [0, 1) is

0 θ=0
g(θ) =
θ − 1 0 < θ < 1
2
Here we impose g(0) = 0 to ensure g(−θ) = −g(θ).
g

−1 − 12 0 1 1 θ
2

197
Proof: Let g(θ) be the periodic function as above. We can compute its Fourier expansion
ˆ 1  
−2πinθ 1
e θ− dθ ( for n 6= 0 )
0 2
  ˆ 1
i −2πinθ 1 1
= e θ− − e−2πinθ dθ
2πn 2 0 0
i
= .
2πn
´1

For n = 0, we have 0 θ − 12 dθ = 0. Thus

1 X i 2πinθ
θ− = e , 0<θ<1
2 2πn
n∈Z\{0}

as expected.

We can also check the distributional equation

∂ b 1 , θ2 ) = δ(θ1 , θ2 ) − 1
G(θ1 , θ2 ) = δ(θ
∂θ1

directly. For any smooth test function f (θ) on S 1 ,

ˆ ˆ θ2
′
− G(θ1 , θ2 )f (θ1 ) dθ1 = − g(θ2 − θ1 )f ′ (θ1 ) dθ1
S1 θ2 −1
ˆθ2  
1
=− θ2 − θ1 − f ′ (θ1 ) dθ1
θ2 −1 2
  θ2 ˆ θ2
1
= − θ2 − θ1 − f (θ1 ) − f (θ1 ) dθ1
2 θ2 −1 θ2 −1
ˆ
=f (θ2 ) − f (θ)dθ
S1
ˆ 1
= (δ(θ1 , θ2 ) − 1) f (θ1 ) dθ1
0

as expected.
Thus we have found the two-point function
D ES 1
Xk (θ1 )Pj (θ2 ) = ih̄G(θ1 , θ2 )δjk = ih̄g(θ2 − θ1 )δjk .
γ

Note that in the limit θ1 → θ2

D ES 1 ih̄ k
lim Xk (θ1 )Pj (θ2 ) =− δ
θ1 →θ2− γ 2 j
D ES 1 ih̄ k
lim Xk (θ1 )Pj (θ2 ) = δ
θ1 →θ2+ γ 2 j

Thus when θ1 and θ2 are close to each other, this two-point function becomes the two point
function on R. This is expected as the local geometry on S 1 is precisely R.

198
S 1 -Correlation

Let A = R[xi , pi ] denote the ring of polynomial functions on the phase space. Recall we
have a natural multiplication map

A ⊗ A ⊗ · · · ⊗ A −→ A
f0 ⊗ f1 ⊗ · · · ⊗ fm 7−→ f0 f1 · · · fm

As an application of the S 1 -quantum mechanical model, we construct a version of quantization

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 of the above multiplication map such that

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 = f0 f1 · · · fm + O(h̄).

Let

Cycm+1 (S 1 ) = (θ0 , θ1 , · · · , θm ) ∈ (S 1 )m+1 | θ0 , θ1 , · · · , θm are distinct in clockwise cyclic order .

θ0
θ1

θm

θ2

···

Using rotation invariance, we have

ˆ ˆ
1
dθ0 dθ1 · · · dθm = dθ1 · · · dθm = .
Cycm+1 (S 1 ) 1≥θ1 ≥···≥θm ≥0 m!

Definition 3.6.12. Let fi ∈ A. We define their S 1 -correlation

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 ∈ A[h̄]

by

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 (x, p)
ˆ
1
=m! dθ0 dθ1 · · · dθm hf0 (x + X(θ0 ), p + P(θ0 )) · · · fm (x + X(θm ), p + P(θm ))iSγ .
Cycm+1 (S 1 )

1
Here the correlation h−iSγ is defined via Wick’s Theorem and Feynman graphs. Explicitly,
1
hf0 (x + X(θ0 ), p + P(θ0 )) · · · fm (x + X(θm ), p + P(θm ))iSγ
P
m
∂ ∂
ih̄ G(θi ,θj )
∂x(i) ∂p(j)
:=e i,j=0
f0 (x(0) , p(0) ) · · · fm (x(m) , p(m) ) .
x(0) =···=x(m) =x
p(0) =···=p(m) =p

199
 It is clear that
ˆ
lim hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 = f0 f1 · · · fm m! dθ0 dθ1 · · · dθm = f0 f1 · · · fm
h̄→0 Cycm+1 (S 1 )

which corresponds to the term without propagators. Since fi ’s are polynomials, only finite
number of propagators can appear. Thus
⟨−⟩ 1
A ⊗ A ⊗ · · · ⊗ A −−−−
S
→ A[h̄].

Example 3.6.13. Given two polynomials f0 , f1 ∈ A = R[x, p],

ˆ ← → ← →
∂ ∂ ∂ ∂
ih̄G(θ0 ,θ1 ) ∂x +ih̄G(θ1 ,θ0 ) ∂p
hf0 ⊗ f1 iS 1 = dθ0 dθ1 f0 (x, p)e ∂p ∂x f1 (x, p)
Cyc2 (S 1 )
← ← →
ˆ 1
→
ih̄g(θ) ∂ ∂
∂x ∂p
− ∂ ∂
∂p ∂x
= dθ f0 (x, p)e f1 (x, p)
0
← ← →
ˆ 1
→
ih̄(θ−1/2) ∂ ∂
∂x ∂p
− ∂ ∂
∂p ∂x
= dθ f0 (x, p)e f1 (x, p)
0
← ← →
ˆ 1 →
2 ih̄θ ∂ ∂
∂x ∂p
− ∂p
∂ ∂
∂x
= dθ f0 (x, p)e f1 (x, p)
− 21
 2m
∞ ˆ
X 1 ← → ← →
2 (ih̄)2m∂ ∂ ∂ ∂ 
= θ2m dθ −f0 (x, p) 
f1 (x, p)
(2m)! ∂x ∂p ∂p ∂x
m=0 − 2
1

← → ← →
2m
X∞
(ih̄)2m ∂ ∂ ∂ ∂
= f0 (x, p)  −  f1 (x, p).
(2m + 1)22m (2m)! ∂x ∂p ∂p ∂x
m=0

For example,
hx ⊗ piS 1 = xp
h̄2
x 2 ⊗ p2 S1
= x 2 p2 −
6
Example 3.6.14. We use Wick’s Theorem to compute
ˆ
hx ⊗ xp ⊗ piS 1 =x p + 2!
2 2
dθ0 dθ1 dθ2 ih̄ (G(θ0 , θ1 ) + G(θ1 , θ2 ) + G(θ0 , θ2 )) xp
Cyc3 (S 1 )
ˆ
+ 2! dθ0 dθ1 dθ2 (ih̄)2 G(θ0 , θ1 )G(θ1 , θ2 )
Cyc3 (S 1 )
ˆ
2 2
=x p + 2xp(ih̄) dθ1 dθ2 (G(1, θ1 ) + G(θ1 , θ2 ) + G(1, θ2 ))
ˆ 1≥θ1 ≥θ2 ≥0
+ 2(ih̄)2 dθ1 dθ2 G(1, θ1 )G(θ1 , θ2 )
1≥θ1 ≥θ2 ≥0
ˆ      
1 1 1
2 2
=x p + 2(ih̄)xp dθ1 dθ2 θ1 − − (θ1 − θ2 ) − + θ2 −
1≥θ1 ≥θ2 ≥0 2 2 2
ˆ   
1 1
+ 2(ih̄)2 dθ1 dθ2 θ1 − − (θ1 − θ2 )
1≥θ1 ≥θ2 ≥0 2 2
ih̄
=x2 p2 + xp.
6

200
3.6.5 Hochschild Homology

As an application of the S 1 -correlation function (Definition 3.6.12), we explain how the

geometry of S 1 is related to the notion of Hochschild homology and present an explicit con-
struction of a quantum version of the Hochschild-Kostant-Rosenberg (HKR) map. We will apply
this quantum HKR map to construction trace map in deformation quantization in Section 4.4.

Hochschild Homology

Let A be an associative algebra over a base field k. We can define a chain complex
b b b b b
· · · → Cp (A) → Cp−1 (A) → · · · → C1 (A) → C0 (A)

where
Cp (A) = A⊗p+1 , here ⊗ = ⊗k .

The map b is defined by

b(a0 ⊗ a1 ⊗ · · · ⊗ ap ) =a0 a1 ⊗ a2 ⊗ · · · ⊗ ap − a0 ⊗ a1 a2 ⊗ · · · ⊗ ap
+ · · · + (−1)p−1 a0 ⊗ a1 ⊗ · · · ⊗ ap−1 ap + (−1)p ap a0 ⊗ a1 ⊗ · · · ⊗ ap−1 .

P
b = ±

··· ···

Check: b2 = 0.
b2 (a0 ⊗ a1 ⊗ · · · ⊗ ap )
=b (a0 a1 ⊗ a2 ⊗ · · · ⊗ ap − a0 ⊗ a1 a2 ⊗ · · · ⊗ ap ± · · ·)
=(a0 a1 )a2 ⊗ · · · ⊗ ap − a0 (a1 a2 ) ⊗ · · · ⊗ ap ± · · ·
=0 by associativity.
The chain complex (C• (A), b) is called the Hochschild chain complex and

HH• (A) := H• (C• (A), b)

is called the Hochschild homology.

Example 3.6.15.
b
C2 (A) −→ C1 (A) −→ C0 (A)
a0 ⊗ a1 7−→ a0 a1 − a1 a0
So HH0 (A) = A/[A, A] is the abelianization.

Theorem 3.6.16 (Hochschild-Kostant-Rosenberg (HKR)). Let A = k[y i ] be a polynomial ring

and Ω•A = k[y i , dy i ] be the algebraic differential forms. Then

HHp (A) = ΩpA .

201
Remark 3.6.17. HKR theorem says that on the standard commutative space k n

HH• = Differential forms.

In general for an associative but non-commutative algebra A, we can view

HH• = Non-commutative differential forms

which is a basic algebraic tool for doing calculus in the noncommutative world.
The HKR isomorphism is realized by

ρ : C• (A) −→ Ω•A
f0 ⊗ f1 ⊗ · · · ⊗ fp 7−→ f0 df1 ∧ df2 ∧ · · · ∧ dfp

Check: ρ(b(−)) = 0.

ρ(b(f0 ⊗ f1 ⊗ · · · ⊗ fp ))
=ρ (f0 f1 ⊗ f2 ⊗ · · · ⊗ fp − f0 ⊗ f1 f2 ⊗ · · · ⊗ fp ± · · ·)
=f0 f1 df2 ∧ · · · ∧ dfp − f0 d(f1 f2 ) ∧ df3 ∧ · · · ∧ dfp ± · · ·
+ (−1)p−1 f0 df1 ∧ · · · ∧ d(fp−1 fp ) + (−1)p fp f0 df1 ∧ · · · ∧ dfp−1
=0.

Thus ρ defines a map of chain complexes

ρ : (C• (A), b) −→ (Ω•A , 0).

A more precise description of HKR is that the HKR map ρ is a quasi-isomorphism, which
induces an isomorphism by passing to homology
≃
ρ : HH• (A) −→ Ω•A .

Quantum HKR

Let A = C[xi , pi ] be the ring of polynomial functions on the phase space R2n . We have a
canonical quantization of A to the associative Weyl algebra

Ah̄ := (A[h̄], ∗)

where ∗ is the Moyal product. We will describe a quantum version ρh̄ of the HKR map ρ that
intertwines the Hochschild chain complex C• (Ah̄ ) of the Weyl algebra Ah̄ .
Let f0 , f1 , · · · , fm ∈ A. Recall we have the S 1 -correlation as in Definition 3.6.12

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 ∈ A[h̄].

To simplify notations, let us denote

Of (θ) := f (x + X(θ), p + P(θ)), θ ∈ S1.

202
By construction
ˆ
1
hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1 = m! dθ0 dθ1 · · · dθm hOf0 (θ0 )Of1 (θ1 ) · · · Ofm (θm )iSγ .
Cycm+1 (S 1 )

Given f , its total differential is

X  ∂f ∂f

i
df = dx + dpi .
∂xi ∂pi
i

We denote
X 
Odf (θ) := O∂xi f (θ)dxi + O∂pi f (θ)dpi .
i

Definition 3.6.18. We define the quantum HKR map

ρh̄ : A⊗m+1 → Ωm
A [h̄]

by
ˆ
1
ρ (f0 ⊗ f1 ⊗ · · · ⊗ fm ) :=
h̄
dθ0 dθ1 · · · dθm hOf0 (θ0 )Odf1 (θ1 ) · · · Odfm (θm )iSγ
Cycm+1 (S 1 )

This can be formally written as

1
ρh̄ (f0 ⊗ f1 ⊗ · · · ⊗ fm ) = hf0 ⊗ df1 ⊗ · · · ⊗ dfm iS 1
m!
Using Wick’s Theorem, we have the explicit formula

ρh̄ (f0 ⊗ f1 ⊗ · · · ⊗ fm )
ˆ ih̄
P
m
G(θi ,θj ) ∂(i) ∂(j)
= dθ0 dθ1 · · · dθm e i,j=0 ∂x ∂p
f0 (x(0) , p(0) )df1 (x(1) , p(1) ) · · · dfm (x(m) , p(m) )
Cycm+1 (S 1 ) x(0) =···=x(m) =x
p(0) =···=p(m) =p

We next explore algebraic structures of ρh̄ . Let us consider the following integral
ˆ 
∂ 1
dθ0 dθ1 · · · dθm hOf0 (θ0 )Of1 (θ1 ) · · · Odfm (θm )iSγ
Cycm+1 (S 1 ) ∂θ1
∂ 1
− hOf0 (θ0 )Odf1 (θ1 )Of2 (θ2 ) · · · Odfm (θm )iSγ
∂θ2
+ ···

∂ S1
+(−1) Of0 (θ0 )Odf1 (θ1 ) · · · Odfm−1 (θm−1 )Ofm (θm ) γ
m−1
(†)
∂θm
´
We compute (†) in two ways. The first is to compute dθk ∂θ∂k (−) as a total derivative.
Due to the cyclic ordering
θk−1

θk

θk+1

203
we have ˆ ˆ θk−1 θk−1
∂ ∂
dθk (−) = dθk (−) = (−) .
∂θk θk+1 ∂θk θk+1

From Wick’s Theorem and the property of S 1 -propagator, we find the limit behavior
1 1
lim h· · · Of (θk−1 )Og (θk ) · · · iSγ = h· · · Of ∗g (θk−1 ) · · · iSγ
−
θk →θk−1

1 1
lim h· · · Of (θk )Og (θk+1 ) · · · iSγ = h· · · Of ∗g (θk+1 ) · · · iSγ
+
θk →θk+1

Therefore computing (†) in terms of total derivative leads to

1 1
hf0 ∗ f1 ⊗ df2 ⊗ · · · ⊗ dfm iS 1 − hf0 ⊗ f1 ∗ df2 ⊗ · · · ⊗ dfm iS 1
(m − 1)! (m − 1)!
1 1
− hf0 ⊗ df1 ∗ f2 ⊗ df3 ⊗ · · · ⊗ dfm iS 1 + hf0 ⊗ df1 ⊗ f2 ∗ df3 ⊗ · · · ⊗ dfm iS 1
(m − 1)! (m − 1)!
+···
1 1
+(−1)m−1 hf0 ⊗ df1 ⊗ · · · ⊗ dfm−1 ∗ fm iS 1 − (−1)m−1 hfm ∗ f0 ⊗ df1 ⊗ · · · ⊗ dfm−1 iS 1 .
(m − 1)! (m − 1)!

By the Moyal product formula, we have

d(f ∗ g) = df ∗ g + f ∗ dg.

Here
X
df ∗ g = (∂xi f ∗ g) dxi + (∂pi f ∗ g) dpi
i
X
f ∗ dg = (f ∗ ∂xi g) dxi + (f ∗ ∂pi g) dpi
i

Thus (†) becomes

1
hf0 ∗ f1 ⊗ df2 ⊗ · · · ⊗ dfm iS 1
(m − 1)!
1
− hf0 ⊗ d(f1 ∗ f2 ) ⊗ · · · ⊗ dfm iS 1
(m − 1)!
+ ···
1
+ (−1)m−1 hf0 ⊗ df1 ⊗ · · · ⊗ d(fm−1 ∗ fm )iS 1
(m − 1)!
1
+ (−1)m hfm ∗ f0 ⊗ df1 ⊗ · · · ⊗ dfm−1 iS 1
(m − 1)!
=ρh̄ (b(f0 ⊗ f1 ⊗ · · · ⊗ fm )).

Here b is the Hochschild differential with respect to the Moyal product.

1
The second way to approach (†) is to compute ∂
∂θk h· · · iSγ explicitly. Observe

∂ ∂
G(θk , θj ) = −1 = − G(θj , θk ) for θj 6= θk .
∂θk ∂θk

204
Thus
∂ S1
Of0 (θ0 )Odf1 (θ1 ) · · · Odfk−1 (θk−1 )Ofk (θk ) · · · Ofm (θm ) γ
∂θk
∂ D ES 1
=ih̄ i Of0 (θ0 )Odf1 (θ1 ) · · · Odfk−1 (θk−1 )O∂pi f (θk ) · · · Odfm (θm )
∂x γ
∂ D ES 1
− ih̄ Of0 (θ0 )Odf1 (θ1 ) · · · Odfk−1 (θk−1 )O∂xi f (θk ) · · · Odfm (θm ) .
∂pi γ

Introduce the following operator

X ∂ ∂
A −→ ΩA ι ∂ pi −
m−1
∆ : Ωm , ∆= i
ι∂ .
∂x ∂pi xi
i
P
Let ω −1 = ∂
∂xi
∧ ∂
∂pi be the Poisson bi-vector. Then ∆ is geometrically the Lie derivative
i

∆ = Lω−1 .

It is clear that ∆2 = 0. ∆ is the prototype of Batalin Vilkovisky (BV) operator in quantum

gauge theory. Then above computation leads to

(†) = ih̄∆ρh̄ (f0 ⊗ f1 ⊗ · · · ⊗ fm ).

Comparing the two computations, we find

ρh̄ (b(−)) = ih̄∆(ρh̄ (−)).

where b is the Hochschild differential with respect to the Moyal product.

We can extend ρh̄ h̄-linearly to

ρh̄ : C• (Ah̄ ) −→ Ω•A [h̄]

Then the above calculation shows that σ h̄ is a morphism of chain complexes

ρh̄ : (C• (Ah̄ ), b) −→ (Ω•A [h̄], ih̄∆).

Thus ρh̄ provides a quantization of ρ.

Remark 3.6.19. We can further localize h̄ and denote

A(h̄) := Ah̄ ⊗k[h̄] k[h̄, h̄−1 ].

The quantum HKR map ρh̄ is naturally extended to

ρh̄ : (C• (A(h̄) ), b) −→ (Ω•A [h̄, h̄−1 ], ih̄∆).

Both the Hochschild homology of A(h̄) and the homology of (Ω•A [h̄, h̄−1 ], ih̄∆) are concentrated
in degree 2n. Passing to homology, ρh̄ gives precisely the Feigin-Felder-Shoikhet trace formula
[20] on HH2n (A(h̄) ).

205
 Chapter 4 Deformation Quantization

In this chapter, we discuss the algebraic theory of deformation quantization which deals
with the operator of quantum observables as a formal deformation of classical observables. The
study of deformation quantization originated from the seminal work by Bayen-Flato-Frönsdal-
Lichnerowicz-Sternheimer [3] in 1978. In this theory, the Planck’s constant h̄ is treated as a
formal variable parametrizing a family of quantum algebras. One remarkable feature is that
deformation quantization exists on any Poisson manifold according to a theorem by Kontsevich
[34]. This greatly extends the scope of quantization beyond ordinary phase spaces.

4.1 Deformation Quantization

4.1.1 Formal Deformations

We fix a base field, the choice of which in specific case should be clear from the context.

Definition 4.1.1. Let V be a vector space. We use V[[λ]] to denote formal power series in λ
with coefficients in V ( )
∞
X
V[[λ]] := a i λi a i ∈ V
i=0

and use V((λ)) to denote formal Laurent series

(∞ )
X
V((λ)) := ai λi ai ∈ V, N ∈ Z .
i=N

When V = R is a ring, both R[[λ]] and R((λ)) inherit well-defined ring structure by
 
X  X  X X
a i λi bj λ j =  a i bj  λ k .
k i+j=k

Definition 4.1.2. Let (A, ·) be an associative algebra. A formal deformation of A is an asso-

ciative product ∗ on A[[λ]] such that

⃝
1 ∗ is λ-bilinear
(f (λ)a) ∗ b = a ∗ (f (λ)b) = f (λ)(a ∗ b)

for any a, b ∈ A[[λ]] and formal power series f (λ).

206
 ⃝
2 For any a, b ∈ A,
∞
X
a∗b=a·b+ λk µk (a, b)
k=1

where a · b is the product in A and µk : A × A → A are bilinear maps.

Condition ⃝
1 on λ-linearity implies that ∗ is completely captured by {µk } in Condition ⃝.
2
We can view ∗ as defining a family of associative products parametrized by λ such that

lim a ∗ b = a · b.
λ→0

Thus ∗ becomes a formal deformation of the associative product ·. It is called formal deformation
because we only consider formal power series in λ and ignore analytic properties (the analytic
property is in fact interesting and important, but irrelevant in the current discussion).

Example 4.1.3. Consider the ring C ∞ (R2n ) of smooth functions on the phase space R2n . Then
the Moyal product 
P ← → ← → 
λ ∂ ∂
− ∂ ∂
2 ∂xi ∂pi ∂pi ∂xi
f ∗ g = fe i g (λ = ih̄)

defines a formal deformation of C ∞ (R2n ). Here

  ← → k
 k X
← →
1 1 ∂ ∂ ∂ ∂
µk (f, g) = f  −  g.
k! 2 ∂xi ∂pi ∂pi ∂xi
i

Example 4.1.4. The above example extends to the formal power series ring R[[xi , pi ]]. The
same formula via Moyal product
 ← → ← → 
P
λ ∂
i
∂
− ∂p
∂ ∂
2 ∂x ∂p i i ∂xi
fe i g

defines a formal deformation of R[[xi , pi ]].

The requirement of associativity of ∗ gives a set of constraints on the bilinear maps {µk }.
For example at first-order in λ

(a ∗ b) ∗ c − a ∗ (b ∗ c)
=(a · b + µ1 (a, b)λ) ∗ c − a ∗ (b · c + µ1 (b, c)λ) + O(λ2 )
=(a · b) · c) + λ(µ1 (a · b, c) + µ1 (a, b) · c) − a · (b · c) − λ(µ1 (a, b · c) + a · µ1 (b, c)) + O(λ2 )
=λ(µ1 (a · b, c) + µ1 (a, b) · c − µ1 (a, b · c) − a · µ1 (b, c)) + O(λ2 ).

Here in the last step we have used the associativity of the undeformed product ·. Thus µ1 has
to satisfy the following equation

a · µ1 (b, c) − µ1 (a · b, c) + µ1 (a, b · c) − µ1 (a, b) · c = 0.

We remark that this is in fact a cocycle condition in Hochschild cohomology.

207
4.1.2 Poisson Algebra as Classical Limit

Definition 4.1.5. A Poisson algebra is a commutative algebra P with a bilinear map (called
the Poisson bracket)
{−, −} : P × P −→ P

satisfying the following properties: for any a, b, c ∈ P ,

⃝
1 Skew-symmetry: {a, b} = − {b, a}

⃝
2 Leibniz rule: {a, bc} = {a, b} c + b {a, c}

⃝
3 Jacobi identity: {a, {b, c}} + {b, {c, a}} + {c, {a, b}} = 0

Note that by skew-symmetry, Leibniz rule also holds for the first factor

{ab, c} = a {b, c} + {a, c} b.

Example 4.1.6. Let M be a symplectic manifold. Then

(C ∞ (M ), {−, −})

is a Poisson Algebra. Here {−, −} is the Poisson bracket associated to the symplectic structure.
In other words, classical observables form a Poisson algebra.

Let (A, ·) be a commutative algebra. Let (A[[λ]], ∗) be a formal deformation of A. Let

[a, b]∗ := a ∗ b − b ∗ a

denote the commutator with respect to the product ∗. Since ∗ is associative, [−, −]∗ satisfies
the Jacobi identity
[a, [b, c]∗ ]∗ + [b, [c, a]∗ ]∗ + [c, [a, b]∗ ]∗ = 0

(Exercise: Check this.) Therefore [−, −]∗ defines a Lie bracket on A[[λ]], which measures the
noncommutativity of ∗.
By the commutativity of the undeformed algebra (A, ·), we have for any a, b ∈ A

[a, b]∗ = λ(µ1 (a, b) − µ1 (b, a)) + O(λ2 ).

Let us define {−, −} : A × A −→ A by

{a, b} := µ1 (a, b) − µ1 (b, a)

or written as
1
{a, b} = lim (a ∗ b − b ∗ a)
λ→0 λ

which captures the first order noncommutativity of ∗.

Proposition 4.1.7. (A, ·, {−, −}) forms a Poisson algebra.

208
Proof: The skew-symmetry is obvious. The Leibniz rule follows from the λ → 0 limit of
1
([a, b ∗ c]∗ − [a, b]∗ ∗ c − b ∗ [a, c]∗ ) = 0.
λ
The Jacobi identity follows from the λ → 0 limit of
1
([a, [b, c]∗ ]∗ + [b, [c, a]∗ ]∗ + [c, [a, b]∗ ]∗ ) = 0.
λ2

We can summarize the above discussion as

λ→0
λ-family of Associative algebra −−−→ Poisson algebra.

In mechanical problems, we usually have a classical system with classical observables as a

Poisson algebra. The program of deformation quantization asks for constructing a quantized
algebra (A[[λ]], ∗) such that its classical limit is the prescribed Poisson algebra (λ = ih̄)
quantization
Poisson algebra Associative algebra
classical limit

Definition 4.1.8. A deformation quantization of a Poisson algebra (A, ·, {−, −}) is a formal
deformation (A[[λ]], ∗) such that
1
{−, −} = lim [−, −]∗ .
λ→0 λ
Given a Poisson algebra, its deformation quantization may not be unique.

Example 4.1.9. Let A = C ∞ (R2 ) with

∂f ∂g ∂f ∂g
{f, g} = − .
∂x ∂p ∂p ∂x
Let ∗ be the standard Moyal product
← → ← →
λ ∂ ∂
− ∂p
∂ ∂

f ∗ g = fe
2 ∂x ∂p ∂x
g.

For any α ∈ R, consider the operator

λ ∂ ∂
α ∂x
Nα := e 2 ∂p : A[[λ]] −→ A[[λ]].
λ ∂ ∂
is understood via its power series expansion. Nα is invertible and Nα−1 = N−α .
α ∂x
Here e 2 ∂p

Let us define
f ∗α g := Nα−1 ((Nα f ) ∗ (Nα g)) .

It is clear that (A[[λ]], ∗α ) defines a formal deformation for any α ∈ R. Explicitly,

 ← → ← →
λ ∂
(1−α) ∂x ∂
−(1+α) ∂ ∂

f ∗α g = f e
2 ∂p ∂p ∂x
g.

In particular,
1
{f, g} = lim (f ∗α g − g ∗α f )
λ→0 λ

209
holds for any α ∈ R. Thus (A[[λ]], ∗α ) is a deformation quantization of (A, ·, {−, −}) for any
α ∈ R. When α = 0, we get the Moyal product. When α = 1,
 ← →
−λ ∂p
∂ ∂

f ∗1 g = f e
∂x
g

is related to the normal ordering by placing the quantum operator p̂ = −λ ∂x

∂
to the right of x

x ∗1 p = xp
p ∗1 x = xp − λ

Thus ∗α=1 captures the composition of differential operators.

4.2 Geometric Approach on Symplectic Manifolds

Let (M, ω) be a symplectic manifold of dimension 2n. The classical observables on the
phase space M form a Poisson algebra (C ∞ (M ), {−, −}). DeWilde-Lecomte [13] obtained the
general existence of deformation quantization on symplectic manifolds via cohomological method
in 1983. Independently , Fedosov [18] presented another beautiful approach via differential
geometric method in 1985.
In this section we describe Fedosov’s geometric construction of deformation quantization on
symplectic manifolds. The basic idea is that locally (M, ω) is modelled on the standard symplec-
tic space R2n where Weyl quantization is available. Then we can glue such local quantizations
into a global one by a consistent parallel transport (i.e. by a flat connection).

4.2.1 Weyl Bundle

Formal Weyl Algebra

Let (V, ω) be a linear symplectic space of dimension 2n. Denote

∞
Y
b ) :=
O(V Symk (V ∗ ), V ∗ = HomR (V, R)
k=0

by the ring of formal power series on V . Let {e1 , · · · , e2n } be a linear basis of V and

ωij := ω(ei , ej ).

Let y 1 , · · · , y 2n be the dual basis of V ∗ , which can be viewed as a set of linear coordinates
on V . Then we can identify
b ) := R[[y i ]]
O(V

and express the symplectic form as

1 X
2n
ω= ωij dy i ∧ dy j .
2
i,j=1

210
 b ) is
The Poisson bracket on O(V
X
{f, g} = ω ij ∂yi f ∂yj g
i,j

where ω ij is the inverse matrix of ωij .

Definition 4.2.1. We define the formal Weyl algebra W(V ) associated to the linear symplectic
space (V, ω) by
 
b )[[λ]], ∗
W(V ) = O(V

b )[[λ]],
where for f, g ∈ O(V
P ← →
λ ∂ ∂
ω ij
2 ∂y i ∂y j
f ∗ g := f e i,j
g.

Here we have suppressed the dependence of W(V ) on the linear symplectic form ω for
simplicity. The above definition of ∗ does not depend on the choice of linear coordinates. It is
 
b ), {−, −} .
clear that W(V ) gives a deformation quantization of O(V

Weyl Bundle

Let (M, ω) be a 2n-dimensional symplectic manifold. For any p ∈ M ,

Tp M, ω|Tp M

defines a linear symplectic space.

Tp M

Let W(Tp M ) be the formal Weyl algebra associated to the linear symplectic space Tp M, ω|Tp M .
This can be viewed as a quantization localized near the point p ∈ M .

Definition 4.2.2. The Weyl bundle of (M, ω) is defined to be

∞
Y
W(M ) := Symk (T ∗ M )[[λ]].
k=0

The Weyl bundle has rank of infinity. At each p ∈ M , the fiber of W(M ) at p is

W(M )|p = W(Tp M ).

211
More precisely, a smooth section s of W(M ) is defined locally by
X
s(x, y, λ) = λk ak,i1 ···il (x)y i1 · · · y il
k≥0,l≥0
i1 ,··· ,il ≥0
 
where {ak,i1 ···il (x)} are smooth functions expressed in local coordinates xi of M , and y i
are induced fiber coordinates on T M .
For simplicity, we will just write W for the Weyl bundle W(M ) when the underlying man-
ifold M is clear from the context. We denote

Γ(U, W) = {smooth sections of W in the open subset U }

for smooth sections in U . The Moyal product on each fiber W(Tp M ) defines a fiberwise product
on W which we still denote by ∗. Explicitly, for any two smooth sections s1 , s2 of W, we have
P ← →
λ ∂ ∂
ω ij (x)
2 ∂y i ∂y j
(s1 ∗ s2 )(x, y, λ) = s1 (x, y, λ)e i,j
s2 (x, y, λ)

in local coordinates as above. Thus

{Γ(U, W} | U ⊂ M open subset}

form a sheaf of associative algebras on M .

4.2.2 Symplectic Connection

The space Γ(M, W) forms an associative algebra, but it is too large for deformation quan-
tization. Our next goal is to construct a flat connection (Fedosov’s abelian connection) on W
to cut down the degrees of freedom. Since W is a bundle built on tensors of the tangent bundle
T M , there is an induced connection on W from any connection on T M . We will start from
such a connection , and discuss how Fedosov trivializes the curvature in the next subsection.

Definition 4.2.3. Let (M, ω) be a symplectic manifold. A connection ∇ on the tangent bundle
T M is called a symplectic connection if

⃝
1 ∇ is torsion free
∇X Y − ∇Y X = [X, Y ], ∀X, Y ∈ Vect(M )

⃝
2 ∇ is compatible with ω
∇ω = 0.

Symplectic connection always exists. To see this, let us choose a metric g on M and let
∇g be the associated Levi-Civita connection. Then ∇g is torsion-free by construction, but may
not be compatible with ω. We look for

∇ = ∇g + α, α ∈ Ω1 (M, End(T M ))

212
such that ∇ becomes a symplectic connection. Let us write

∇X Y = ∇gX Y + α(X, Y ).

Torsion-free condition asks

∇X Y − ∇Y X − [X, Y ]
=∇gX Y + α(X, Y ) − ∇gY X − α(Y, X) − [X, Y ]
=α(X, Y ) − α(Y, X) = 0
so we need α to satisfy
α(X, Y ) = α(Y, X).
Compatibility with ω asks

Xω(Y, Z) =ω(∇X Y, Z) + ω(Y, ∇X Z)

which is equivalent to

ω(α(X, Y ), Z) − ω(α(X, Z), Y ) = Xω(Y, Z) − ω(∇gX Y, Z) − ω(Y, ∇gX Z).

Let ξ be the tensor defined by

ξ(X, Y, Z) = Xω(Y, Z) − ω(∇gX Y, Z) − ω(Y, ∇gX Z).

Skew-symmetry of ω implies
ξ(X, Y, Z) = −ξ(X, Z, Y ).
Torsion-freeness of ∇g and dω = 0 imply the total skew-symmetrization of ξ vanishes, i.e.

ξ(X, Y, Z) + ξ(Y, Z, X) + ξ(Z, X, Y ) = 0.

We look for α to satisfy


α(X, Y ) = α(Y, X)
ω(α(X, Y ), Z) − ω(α(X, Z), Y ) = ξ(X, Y, Z)

One such α can be found by solving the equation

1 1
ω(α(X, Y ), Z) = ξ(X, Y, Z) + ξ(Y, X, Z).
3 3
Non-degeneracy of ω guarantees this equation is uniquely solved. Such determined α clearly
satisfies α(X, Y ) = α(Y, X) and

ω(α(X, Y ), Z) − ω(α(X, Z), Y )

1
= (ξ(X, Y, Z) + ξ(Y, X, Z) − ξ(X, Z, Y ) − ξ(Z, X, Y ))
3
1
= (2ξ(X, Y, Z) − (ξ(Y, Z, X) + ξ(Z, X, Y )))
3
1
= (2ξ(X, Y, Z) + ξ(X, Y, Z))
3
=ξ(X, Y, Z)
as required.
Unlike the Riemannian case, symplectic connections are not unique.

213
Theorem 4.2.4. Symplectic connections on a symplectic manifold always exist. The set of
symplectic connections is an affine space modelled on Γ(M, Sym3 (T ∗ M )).

Proof: We have shown the existence of a symplectic connection above. Let ∇ and ∇′ be two
symplectic connections and they differ by

∇′X Y = ∇X Y + α(X, Y ).

Torsion-freeness requires α(X, Y ) = α(Y, X). Compatibility with ω asks ω(α(X, Y ), Z) =

ω(α(X, Z), Y ). Let ξ denote the tensor

ξ(X, Y, Z) := ω(α(X, Y ), Z).

Then the above two conditions become


ξ(X, Y, Z) = ξ(Y, X, Z)
ξ(X, Y, Z) = ξ(X, Z, Y )

which is equivalent to the total symmetry of ξ, i.e.

ξ ∈ Γ(M, Sym3 (T ∗ M )).

Definition 4.2.5. Let ∇ be a symplectic connection on (M, ω) and R∇ denote its curvature.


R∇ (X, Y )Z = ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] Z

for vector fields X, Y, Z on M . Define the symplectic curvature Rω∇ by the tensor

Rω∇ (X, Y, Z, T ) := ω(R∇ (X, Y )Z, T ).

Compatibility with ω implies

Rω∇ (X, Y, Z, T )


=ω( ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] Z, T )


= − ω(Z, ∇X ∇Y − ∇Y ∇X − ∇[X,Y ] T )
= − ω(Z, R∇ (X, Y )T )
=ω(R∇ (X, Y )T, Z)
=Rω∇ (X, Y, T, Z).

Thus Rω∇ is skew-symmetric in the first two arguments and symmetric in the last two

Rω∇ ∈ Γ(M, ∧2 (T ∗ M ) ⊗ Sym2 (T ∗ M )).

Since ∇ is torsion-free, Rω∇ also satisfies the First Bianchi identity

Rω∇ (X, Y, Z, T ) + Rω∇ (Y, Z, X, T ) + Rω∇ (Z, X, Y, T ) = 0.

214
4.2.3 Fedosov’s Abelian Connection

Let ∇ be a symplectic connection on (M, ω). It naturally induces a connection on all tensor
Symk (T ∗ M ) hence a connection on the Weyl bundle
∞
Y
W= Symk (T ∗ M )[[λ]].
k=0

Let xi be local coordinates on M and
1X
ω= ωij (x)dxi ∧ dxj , ωij = ω(∂xi , ∂xj ).
2
i,j

Let yi be the corresponding fiber coordinates on T M . We can represent the symplectic
curvature tensor
^
2
Rω∇ ∈ Γ(M, (T ∗ M ) ⊗ Sym2 (T ∗ M ))

in local coordinates as
1 X
Rω∇ = Rijkl (x)dxi ∧ dxj ⊗ y k y l
4
i,j,k,l

where Rijkl is skew-symmetric in the first two (ij)-index and symmetric in the last two (kl)-
index. The explicit identification is
1X
ω(∇2 (∂xk ), ∂xl ) = Rijkl dxi ∧ dxj .
2
i,j

Let us see how the curvature of ∇ on the Weyl bundle W looks like. In local coordinates
yi is dual to ∂xi , thus
!
X X
0=∇ 2
∂ xk ⊗ y k
= ∇2 (∂xk ) ⊗ y k + ∂xk ⊗ ∇2 y k .
k k

This implies
X X 1X
ω(∂xk , ∂xl )∇2 y k = − ω(∇2 (∂xk ), ∂xl ) ⊗ y k = − Rijkl dxi ∧ dxj ⊗ y k
2
k k i,j,k

i.e. X 1X
ωkl ∇2 y k = − Rijkl dxi ∧ dxj ⊗ y k
2
k i,j,k

1 X lm
=⇒ ∇2 y m =− ω Rijkl dxi ∧ dxj ⊗ y k .
2
i,j,k,l

We can write this in terms of the fiberwise Moyal product commutator as

 
X  
1 1 k l m 1 ∇ m
∇ y =
2 m
− Rijkl dx ∧ dx ⊗ y y , y
i j
= − Rω , y .
λ 4 λ ∗
i,j,k,l
∗

Using the Leibniz rule and Moyal product formula, we find for any section s of W
 
1 ∇
∇ s = − Rω , s
2
λ ∗

215
i.e. the curvature of the induced connection ∇ on the Weyl bundle is
 
1 ∇
∇ = − Rω , −
2
on W.
λ ∗

Another good property of this induced connection ∇ on W is its compatibility with the
fiberwise Moyal product: for any two sections s1 , s2 of W,

∇(s1 ∗ s2 ) = (∇s1 ) ∗ s2 + s1 ∗ ∇s2

i.e., the fiberwise Moyal product ∗ is ∇-parallel. This follows from the fact that ∗ is constructed
out of ω which is itself ∇-parallel.
On the other hand, we can modify the induced connection ∇ on W by terms which do not
come from a connection on T M . Specifically, for any 1-form valued in the Weyl bundle

γ ∈ Ω1 (M, W)

we can define a new connection D on W by

1
D =∇+ [γ, −]∗ .
λ
Such connection is also compatible with the fiberwise Moyal product ∗: for any two sections
s1 , s2 of W,
1
D(s1 ∗ s2 ) = ∇(s1 ∗ s2 ) + [γ, s1 ∗ s2 ]∗
λ
1 1
= ∇s1 ∗ s2 + s1 ∗ ∇s2 + [γ, s1 ]∗ ∗ s2 + s1 ∗ [γ, s2 ]∗
λ λ
= (Ds1 ) ∗ s2 + s1 ∗ Ds2 .

The upshot is that it is possible to find a γ such that D becomes a flat connection on W.
This is Fedosov’s key observation which will allow us to construct deformation quantization on
symplectic manifolds. To see this, let us first compute the curvature of D = ∇ + 1
λ [γ, −]∗

1 1
D 2 = ∇2 + [∇γ, −]∗ + 2 [γ, [γ, −]∗ ]∗
λ λ
1 1
= ∇ + [∇γ, −]∗ + 2 [[γ, γ]∗ , −]∗
2

 λ 2λ 
1 ∇ 1
= −Rω + ∇γ + [γ, γ]∗ , − .
λ 2λ ∗

Note that 1
2 [γ, γ]∗ = γ ∗ γ since γ is valued in 1-form. Thus we are looking for γ ∈ Ω1 (M, W)
such that −Rω∇ + ∇γ + 1
2λ [γ, γ]∗ lies in the center of W, i.e.

1
−Rω∇ + ∇γ + [γ, γ]∗ ∈ Ω2 (M )[[λ]].
2λ
Such a connection is called an abelian connection.
To construct an abelian connection γ, we first consider the operator

δ : W −→ Ω1 (M, W)

216
which is described in local coordinates by
X ∂
δ= dxi .
∂y i
i

Equivalently, if we denote
X
γ0 = ωij dxi y j ∈ Ω1 (M, W)
i

then
1
δ= [γ0 , −]∗ .
λ
Note that γ0 is ∇-parallel and
1 1 1 X 1X
∇γ0 + [γ0 , γ0 ]∗ = γ0 ∗ γ0 = ωik dxi ω kl ωjl dxj = − ωij dxi ∧ dxj = −ω
2λ λ 2 2
i,j,k,l i,j

which is the negative of the symplectic form.

We also consider the adjoint operator

δ ∗ : Ω1 (M, W) −→ W

defined by
X
δ∗ = y i ι ∂xi .
i

Both δ and δ∗ extend to forms of arbitrary degrees

δ : Ωp (M, W) −→ Ωp+1 (M, W)

δ ∗ : Ωp (M, W) −→ Ωp−1 (M, W)

and satisfy the following Hodge type relation

δδ ∗ + δ ∗ δ = p + q on Ωp (M, Symq (T ∗ M ))[[λ]].

In particular, δ is null-homotopic when p + q > 0. This immediately proves the following

Lemma 4.2.6. The cohomology of (Ω• (M, W), δ) is concentrated in degree 0 and

H • (Ω• (M, W), δ) = H 0 (Ω• (M, W), δ) = C ∞ (M )[[λ]].

We define a Z-grading “wt” of weight on the Weyl bundle W by assigning

wt(y i ) = 1, wt(λ) = 2.

It is clear that the fiberwise Moyal product preserves weights. Such weight defines a filtration

W = W 0 ⊃ W1 ⊃ · · · ⊃ W m ⊃ · · ·

where Wm consists of elements of weight ≥ m.

217
Theorem 4.2.7 (Fedosov). Given any sequence {ωk }k≥1 of closed 2-forms on M , there exists
γ ∈ Ω1 (M, W) of the form

γ = γ0 + γ + , γ + ∈ Ω1 (M, W3 )

such that
1
−Rω∇ + ∇γ + [γ, γ]∗ = ωλ .
2λ
P
Here ωλ = −ω + λk ωk . In particular, the connection D = ∇ + 1
λ [γ, −]∗ is an abelian
k≥1
connection whose curvature on W is zero.

Proof: Note that 1

λ [γ0 + γ + , −]∗ = δ + λ1 [γ + , −]∗ and ∇γ0 + 1
2λ [γ0 , γ0 ]∗ = −ω. Let us denote

1 + + X
Σ(γ + ) := δγ + + ∇γ + + [γ , γ ]∗ − Rω∇ − λk ωk .
2λ
k≥1

The required equation for γ is Σ(γ + ) = 0.

We decompose γ + into
γ + = γ3+ + γ4+ + · · ·
+ is homogeneous of weight m. Observe that both ∇ and [−, −] preserve the weights
where γm ∗

while δ decreases the weight by 1. We will solve Σ(γ + ) = 0 order by order in terms of weight.
The lowest order term in Σ(γ + ) has weight 2 and we need to solve

δγ3+ − Rω∇ − λω1 = 0.

Note that δ(Rω∇ + λω1 ) = δ(Rω∇ ) = 0 by the First Bianchi identity. By Lemma 4.2.6, we know
Rω∇ + λω1 must be δ-exact since it lies in cohomology degree 2. Hence γ3+ can be solved.
Assume we have found

+
γ≤m−1 = γ3+ + · · · + γm−1
+
, m≥4

such that
+
Σ(γ≤m−1 ) ∈ Ω2 (M, Wm−1 ).
+ such that the following holds
We look for γm

+
Σ(γ≤m ) ∈ Ω2 (M, Wm ).

First observe that for any γ

  
1 ∇ 1
∇ + [γ, −]∗ −Rω + ∇γ + [γ, γ]∗ = 0
λ 2λ

which is essentially the Bianchi identity for the curvature. Let

+
Σ(γ≤m−1 ) = Om−1 modulo Ω2 (M, Wm ).

218
where Om−1 has weight m − 1. Apply the above Bianchi identity to γ0 + γ≤m−1
+
and use the
fact that ωk ’s are closed
1h + i
=⇒ ∇Σ(γ≤m−1
+ +
) + δΣ(γ≤m−1 )+ +
γ≤m−1 , Σ(γ≤m−1 ) = 0.
λ
Picking up the leading order in weight, we find

δOm−1 = 0.

+
On the other hand, the leading order in weight of Σ(γ≤m ) is

+
Σ(γ≤m +
) =Σ(γ≤m−1 +
) + δγm modulo Ω2 (M, Wm )
+
=Om−1 + δγm modulo Ω2 (M, Wm ).

Since δOm−1 = 0 and is a 2-form, by Lemma 4.2.6 we know Om−1 must be δ-exact. So we can
+ such that O
m−1 + δγm = 0, i.e., Σ(γ≤m ) ∈ Ω (M, Wm ) holds.
+ + 2
find γm
Thus we can recursively construct

γ = γ0 + γ3+ + γ4+ + · · · + γm
+
+ ···

such that
1
−Rω∇ + ∇γ + [γ, γ]∗ = ωh̄
2λ
holds. In particular, the curvature of D = ∇ + λ1 [γ, −]∗ is
 
1 ∇ 1 1
2
D = −Rω + ∇γ + [γ, γ]∗ , − = [ωh̄ , −] = 0
λ 2λ λ

since ωh̄ lies in the center of W.

4.2.4 Symbol Map

Let D = ∇ + λ1 [γ, −]∗ be Fedosov’s abelian connection which gives a flat connection on the
Weyl bundle W. It defines a de Rham chain complex

(Ω• (M, W), D)

with the associated cohomology

•
HD (M, W) := H • (Ω• (M, W), D).

• (M, W) is
Proposition 4.2.8. The cohomology HD

C ∞ (M )[[λ]] p=0
p
HD (M, W) =
0 p>0

219
Proof: Consider the following filtration induced by the weight

Fm Ωp (M, W) := Ωp (M, Wm−p ).

Here for k < 0, Wk = W0 = W. This gives a decreasing filtration

Ω• (M, W) = F0 Ω• (M, W) ⊃ F1 Ω• (M, W) ⊃ · · · ⊃ Fm Ω• (M, W) ⊃ · · ·

Fedosov’s abelian connection has the form

1 + 
D =δ+∇+ γ ,− ∗
λ
which clearly preserves Fm Ω• (M, W). The induced differential on the graded piece is
Fm Ω• (M, W)
δ : GrFm −→ GrFm , GrFm := .
Fm+1 Ω• (M, W)
Since W is built from formal power series, there is a natural vector space isomorphism
Y ≃
c F :=
Gr GrFm −−→ Ω• (M, W).
m≥0

It follows from Lemma 4.2.6 that

c F , δ) ' H • (Ω• (M, W), δ) = H 0 (Ω• (M, W), δ) = C ∞ (M )[[λ]].

H • (Gr

So the spectral sequence of the filtration degenerates at E1 -page and the proposition follows.

Let ΓfDlat (M, W) denote the space of flat sections of W

ΓfDlat (M, W) := {s ∈ Γ(M, W) | Ds = 0} .

Equivalently, we can identify

ΓfDlat (M, W) = HD
0
(M, W).

The above proposition says

D D
0 −→ ΓfDlat (M, W) −→ Ω0D (M, W) −→ Ω1D (M, W) −→ · · ·

is an exact sequence and

ΓfDlat (M, W) ' C ∞ (M )[[λ]].

This isomorphism can be described explicitly as follows.

Definition 4.2.9. Define the symbol map

σ : Γ(M, W) −→ C ∞ (M )[[λ]]
s(x, y, λ) 7→ s(x, 0, λ)

which sends all fiber variables y i to zero.

The above spectral sequence computation shows that

≃
σ : ΓfDlat (M, W) −→ C ∞ (M )[[λ]]

gives the required isomorphism.

220
4.2.5 Globalization

Fedosov’s abelian connection allows us to globalize local Weyl quantization to obtain a

deformation quantization on the whole manifold. We explain this construction in this subsection.
The connection D = ∇ + 1
λ [γ, −]∗ is compatible with the fiberwise Moyal product

D(s1 ∗ s2 ) = Ds1 ∗ s2 + s1 ∗ Ds2 .

Thus if s1 , s2 are flat sections, Ds1 = Ds2 = 0, then

D(s1 ∗ s2 ) = 0

so s1 ∗ s2 is also a flat section. Therefore

(ΓfDlat (M, W), ∗)

defines an associative algebra.

In terms of the isomorphism via symbol map

≃
σ : ΓfDlat (M, W) −→ C ∞ (M )[[λ]]

we have an induced associative product ∗D on C ∞ (M )[[λ]]

f ∗D g := σ(σ −1 (f ) ∗ σ −1 (g))

for f, g ∈ C ∞ (M )[[λ]]. In other words, the symbol map identifies two associative algebras
σ
∼
(ΓfDlat (M, W), ∗) −→ (C ∞ (M )[[λ]], ∗D )

Theorem 4.2.10 (Fedosov). (C ∞ (M )[[λ]], ∗D ) defines a deformation quantization on the sym-

plectic manifold (M, ω).

Proof: The induced product ∗D is clearly λ-linear and associative. Let f, g ∈ C ∞ (M ). By the
fiberwise Moyal product

f ∗D g = σ(σ −1 (f ) ∗ σ −1 (g)) = f g + O(λ)

so the limit λ → 0 of ∗D gives the classical product.

We are left to check the first order non-commutativity. Let
X
σ −1 (f (x)) = f (x) − fi (x)y i modulo W2 .
i

The flatness condition D(σ −1 (f )) = 0 implies

!
X ∂f
df (x) − δ i
fi (x)y = 0 =⇒ fi (x) = .
∂xi
i

221
Thus
1 1

[f, g]∗D = σ [σ −1 (f ), σ −1 (g)]∗
λ λ 
X −1 −1
∂σ (f ) ∂σ (g)
= σ ω ij (x) + O(λ)
∂y i ∂y j
i,j
X
= ω ij (x)fi (x)fj (x) + O(λ)
i,j

= {f, g} + O(λ)
as required.

4.3 Poisson Manifold

Functions on a symplectic manifold define a Poisson algebra. However, Poisson algebras
can arise in a more general setting on so-called Poisson manifolds. This includes a large class
of important examples, such as the dual g∗ of a Lie algebra g whose deformation quantization
leads to the universal enveloping algebra U (g).

4.3.1 Polyvector Fields

Definition 4.3.1. Let M be a smooth manifold. Denote

M
PV• (M ) = PVk (M ), PVk (M ) := Γ(M, ∧k T M ).
k

Elements of PVk (M ) are called k-polyvector fields. We denote |µ| = k for µ ∈ PVk (M ).
PV• (M ) is a graded commutative algebra under the wedge product

α ∧ β = (−1)|α||β| β ∧ α.

For k = 1, PV1 (M ) = Vect(M ) are vector fields on M . There is a Lie bracket on Vect(M )

[−, −] : Vect(M ) × Vect(M ) −→ Vect(M ).

This Lie bracket extends to polyvector fields by

[X1 ∧ · · · ∧ Xm , Y1 ∧ · · · ∧ Yn ]SN
X
:= (−1)i+j [Xi , Yj ] ∧ X1 ∧ · · · ∧ X bi ∧ · · · ∧ Xm ∧ Y1 ∧ · · · ∧ Ybj ∧ · · · ∧ Yn
i,j

and
X
[f, X1 ∧ · · · ∧ Xm ]SN := b i ∧ · · · ∧ Xm
(−1)i Xi (f )X1 ∧ · · · ∧ X
i

for vector fields Xi , Yj and function f . This bracket [−, −]SN is called the Schouten-Nijenhuis
bracket, also known as the Schouten bracket.

Proposition 4.3.2. The Schouten-Nijenhuis bracket satisfies the following

222
 • Skew-symmetry
[α, β]SN = −(−1)(|α|−1)(|β|−1) [β, α]SN

• Leibniz Rule
[α, β ∧ γ]SN = [α, β]SN ∧ γ + (−1)(|α|−1)|β| β ∧ [α, γ]SN

• Jacobi identity

[[α, β]SN , γ]SN = [α, [β, γ]SN ]SN − (−1)(|α|−1)(|β|−1) [β, [α, γ]SN ]SN

Proof: Direct computation.

The Jacobi identity can be also written in the symmetric form

(−1)(|α|−1)(|γ|−1) [α, [β, γ]SN ]SN +(−1)(|β|−1)(|α|−1) [β, [γ, α]SN ]SN +(−1)(|γ|−1)(|β|−1) [γ, [α, β]SN ]SN = 0.

4.3.2 Poisson Manifold and Poisson Cohomology

Poisson Manifold

Definition 4.3.3. A Poisson bivector on M is a 2-polyvector field P ∈ PV2 (M ) satisfying

[P, P ]SN = 0.

A pair (M, P ) where P is a Poisson bivector is called a Poisson manifold.

Let (M, P ) be a Poisson manifold. We can define a Poisson bracket on C ∞ (M )

{−, −}P : C ∞ (M ) × C ∞ (M ) −→ C ∞ (M )

by
{f, g}P := P (df ⊗ dg).

In local coordinates xi on M , let

1 X ij
P = P (x)∂xi ∧ ∂xj .
2
i,j

Then
X
{f, g}P = P ij (x)∂xi f ∧ ∂xj g.
i,j

Proposition 4.3.4. (C ∞ (M ), {−, −}P ) defines a Poisson algebra. The Jacobi identity of
{−, −}P is equivalent to [P, P ]SN = 0.

223
Proof: Let us check the Jacobi identity of {−, −}P .

{{f, g}P , h}P + cyclic permutations

X  
= P ij ∂i P kl ∂k f ∂l g ∂j h + cyclic permutations
X
= P ij ∂i P kl ∂k f ∂l g∂j h + cyclic permutations
 X 
1
= P ∂i P ∂k ∧ ∂l ∧ ∂j (df ⊗ dg ⊗ dh)
ij kl
2
= ([P, P ]SN ) (df ⊗ dg ⊗ dh) .

Example 4.3.5. Let (M, ω) be a symplectic manifold. In local coordinates

1
ω = ωij dxi ∧ dxj .
2
Let ω ij be the inverse of ωij . Then
1 X ij
ω −1 := ω ∂ xi ∧ ∂ xj
2
i,j

defines a Poisson bi-vector. The Poisson bracket {−, −}ω−1 is precisely the one on the symplectic
phase space. In this case

dω = 0 ⇐⇒ [ω −1 , ω −1 ]SN = 0.

Example 4.3.6 (Constant Poisson Structure). Let M = Rn and

1 X ij
P = P ∂ xi ∧ ∂ xj , P ij ∈ R.
2
i,j

Since P ij ’s are constants, [P, P ]SN = 0 holds. Such P is called a constant Poisson structure.

Example 4.3.7 (Linear Poisson Structure). Let M = Rn and

1 X ij X
P = P (x)∂xi ∧ ∂xj , P ij (x) = cij k
kx .
2
i,j k

Here ckij ∈ R are constants, thus P ij (x) is linear in x. The Poisson bracket reads
 X
xi , x j P
= cij k
kx .
k

It is clear from this expression that the Jacobi identity is equivalent to the condition that cij
k
defines the structure constant of a Lie algebra.
Intrinsically, we can identify M = g∗ as the linear dual of a Lie algebra g. The linear

coordinates xi can be viewed as a basis of g, under which the Lie bracket is expressed via the
structure constants cij
k . The linear Poisson structure is also called Lie-Poisson structure.

224
Poisson Cohomology

Let (M, P ) be a Poisson manifold. Define

dP : PV• (M ) −→ PV•+1 (M )
α 7−→ [P, α]SN
Using the Jacobi identity
[[P, P ]SN , −]SN = 2[P, [P, −]SN ]SN

and the Poisson condition [P, P ]SN = 0, we find

d2P = 0.

Thus (PV• (M ), dP ) defines a complex

d d d
0 −→ PV0 (M ) −→
P
PV1 (M ) −→
P
· · · −→
P
PVn (M ) −→ 0, n = dim M

which is called the Lichnerowicz complex.

Definition 4.3.8. The cohomology of the Lichnerowicz complex is called the Poisson cohomol-
ogy, denoted by
HPk (M ) := H k (PV• (M ), dP ).

Example 4.3.9. Consider an infinitesimal deformation of the Poisson structure

P −→ P + εQ, Q ∈ PV2 (M )

where ε is the infinitesimal parameter. The Poisson condition

[P + εQ, P + εQ]SN = 2ε[P, Q]SN = 0

says Q is dP -closed. Let X be a vector field. It generates a flow on M which changes

P −→ P + dP X.

So dP -exact Q corresponds to trivial deformations induced by diffeomorphisms. Thus

{first order deformations of P }
HP2 (M ) = .
{trivial deformations}
Example 4.3.10. Let (M, ω) be a symplectic manifold and P be the Poisson bi-vector P = ω −1 .
The non-degeneracy of ω induces a bundle isomorphism

ϕ : T M −→ T ∗ M
X 7−→ −ιX ω

In local coordinates ω = 21 ωij dxi ∧ dxj

X
ϕ(∂xi ) = − ωij dxj .
j

Extending to exterior tensors, ϕ induces an isomorphism

ϕ : PVk (M ) −→ Ωk (M )

between k-polyvector fields and differential k-forms. Observe that

225
 • for any function f ,  
X
ϕ(dP f ) = ϕ  ω ij ∂xj f ∂xi  = df
i,j

• for vector field ∂xk

 
1 X
ϕ(dP (∂xk )) =ϕ − ∂k ω ij ∂xi ∧ ∂xj 
2
i,j
1 X
=− ∂k ω ij ωim ωjl dxm ∧ dxl
2
i,j,m,l
1 X
= ∂k ωim ω ij ωjl dxm ∧ dxl
2
i,j,m,l
1X
= ∂k ωim dxm ∧ dxi
2
i,m
using 1X
===== − (∂i ωmk + ∂m ωki ) dxm ∧ dxi
dω=0 2
i,m
!
X
=−d ωki dx i

=dϕ(∂xk )

Using the Leibniz rule, we find ϕ identifies two complexes

φ
≃
(PV• (M ), dP ) −→ (Ω• (M ), d).

In particular, we have
φ
Hω2 −1 (M ) ' H 2 (M ).

This says that Poisson deformations of ω −1 are parametrized by H 2 (M ), i.e., are induced from
deformations of the symplectic form ω in H 2 (M ).

4.3.3 Kontsevich’s Star Product

Star Product

Definition 4.3.11. Let (M, P ) be a Poisson manifold. A star product (or a deformation
quantization) on (M, P ) is a formal deformation (C ∞ (M )[[λ]], ∗) of C ∞ (M )
∞
X
f ∗g =f ·g+ λk µk (f, g)
k=1

such that

• lim λ1 (f ∗ g − g ∗ f ) = {f, g}P

λ→0

• Each µk is a bi-differential operator.

226
 In other words, a star product on (M, P ) is a deformation quantization of the Poisson
algebra (C ∞ (M ), {−, −}P ) such that all µk ’s are “local”. We can also replace C ∞ (M ) by other
appropriate functions (such as polynomials/formal power series, etc) in specific context.

Example 4.3.12. Let (M, ω) be a symplectic manifold and P = ω −1 . Then Fedosov’s abelian
connection D leads to a star product ∗D on (M, ω −1 ). See Section 4.2.5.

Example 4.3.13. Let M = Rn and

1 X ij
P = P ∂xi ∧ ∂xj , P ij ∈ R
2
i,j

be a constant Poisson structure. When det(P ij ) = 0, P does not come from a symplectic
structure. Nevertheless, the same formula of Moyal product
λ P ←
− →
∂
−
∂
P ij
2 ∂xi ∂xj
f ∗ g := f e i,j
g

defines a star product on (Rn , P ).

Example 4.3.14. Let g be a Lie algebra and M = g∗ which is equipped with the Lie-Poisson
structure Pg . Let
U λ (g) := T (g)[λ]/ ∼
L
where T (g) = g⊗m is the tensor algebra of g and the quotient relation ∼ is
m≥0

x ⊗ y − y ⊗ x ∼ λ[x, y].

If we specialize to λ = 1, then U λ=1 (g) is the universal enveloping algebra of g. Let

M
Sym(g) := Symm (g) = lim U λ (g)
λ→0
m≥0

denote the ring of polynomial functions on g∗ . We can embed Sym(g) into U λ (g) via the
symmetrization map

ϕ : Sym(g)[λ] −→ U λ (g)
λl X
x1 x2 · · · xm λl 7−→ xσ(1) ⊗ · · · ⊗ xσ(m)
m!
σ∈Sm

which is a vector space isomorphism by the Poincaré-Birkhoff-Witt Theorem.

U λ (g) is an associative algebra with ⊗ as the product. This allows us to define a canonical
star product ∗g on Sym(g)[λ] by

(f1 ∗g f2 ) := ϕ−1 (ϕ(f1 ) ⊗ ϕ(f2 )).

For example, if X, Y ∈ g are linear, then

 
−1 −1 1 1 λ
X ∗g Y = ϕ (X ⊗ Y ) = ϕ (X ⊗ Y + Y ⊗ X) + (X ⊗ Y − Y ⊗ X) = XY + [X, Y ].
2 2 2

227
 X 2 ∗g Y =ϕ−1 (X ⊗ X ⊗ Y )

−1 1
=ϕ (X ⊗ X ⊗ Y + X ⊗ Y ⊗ X + Y ⊗ X ⊗ X)
3

2 1
+ X ⊗ (X ⊗ Y − Y ⊗ X) + ((X ⊗ Y − Y ⊗ X) ⊗ X)
3 3
λ 2
=X 2 Y + λX[X, Y ] + [X, [X, Y ]].
6
It is worthwhile to point out that the Moyal product formula does not work in this linear
λ2
Poisson case. The term 6 [X, [X, Y ]] gives a further correction in the above example. In general,
the Moyal product formula
λ P ←
−
∂
→
−
P ij (x) ∂ j
2 ∂xi ∂x
e i,j

fails to be associative when P ij (x) are not constant. This leads to a major difficulty in con-
structing star product on Poisson manifolds.

Example 4.3.15. Let M = R2 and P = xy∂x ∧ ∂y . Denote by

∂log x := x∂x , ∂log y := y∂y .

Then we can write P as

P = ∂log x ∧ ∂log y .

It follows that ← 
− →
− ←− →
−
λ
∂ log x ∂ log y − ∂ log y ∂ log x
f ∗ g := f e 2 g

defines a star product on (R2 , P ). In this star product

λ
x ∗ y = e 2 xy
y ∗ x = e− 2 xy
λ

In particular, we find
x ∗ y = eλ y ∗ x.

Kontsevich’s Star Product

In [34], Kontsevich figured out a remarkable explicit way to correct the Moral product
formula into an associative product using the quantum field theory of two-dimensional Poisson
σ-model (as explained later by Cattaneo-Felder [8]).

Theorem 4.3.16 (Kontsevich). Deformation quantization exists on any Poisson manifold.

We sketch the basic idea of Kontsevich’s construction. Consider the Poisson manifold
1 X ij
(Rn , P = P (x)∂xi ∧ ∂xj ).
2
i,j

Let Γr be a graph with r + 2 vertices and 2r (directed) edges drawn in the upper half-plane such
that two vertices are placed on the real line with only incoming edges and each of the other r
vertices is placed in the interior with two out-going edges from it.

228
 P P
P

f g f g

From each Γr , we can write down a bi-differential operator CΓr (f, g) by assigning P =
P
i ⊗
1 ∂ ∂
2 P ij (x) ∂x ∂xj
to each of the r vertices in the interior where we place the two derivations
i,j
∂ ∂
∂xi
and ∂xj
to its out-going edges. For example, the first graph in the above figure gives

1 X ij ∂f ∂g
P (x) i ⊗ j
2 ∂x ∂x
i,j

and the second graph gives

 
1 X ij X 1 ∂ km ∂2f ∂g
P P .
2 2 ∂xi ∂xk ∂xj ∂xm
i,j k,m

Then we can construct a star product of the form

X
µr (f, g) = ωΓr CΓr (f, g)
Γr

where the coefficients ωΓr is obtained from an appropriate Feynman graph integral over the
configuration space of the upper half-plane.
The Feynman graph integral arises from the correlation functions in the Poisson σ-model
ˆ ˆ
1
S[X, η] = i
ηi dx + P ij (x)ηi ∧ ηj
H 2 H

where X = (X i ) is a bosonic field describing mappings from the upper half-plane H to Rn

X : H −→ Rn

and η = (ηi ) is a fermionic field valued as a 1-form on H. The associativity of the star product
∗ follows from a version of Stokes Theorem.
The Moyal product formula corresponds to the graphs
P P P

···

f g f g f g

Kontsevich’s formula gives a manifest correction of the Moyal product by all possible graphs
with edges connecting two interior vertices.

229
 f g

4.4 Trace Map

We consider (M, ω) to be a symplectic manifold in this section. Let (C ∞ (M )[[λ]], ?) be a
deformation quantization. A λ-linear map

Tr : C ∞ (M )[[λ]] → R((λ))

is called a trace map if

Tr(f ? g) = Tr(g ? f ) or equivalently Tr([f, g]⋆ ) = 0

for any f, g ∈ C ∞ (M )[[λ]]. Since Tr is λ-linear, it suffices to assume f, g ∈ C ∞ (M ).

The star product ? describes the algebraic structure of quantized operators on the phase
space. If there would be a representation of the quantized operators, then the corresponding
trace on the representing Hilbert space would annihilate the operator commutators. A trace
map defined above precisely models this information without specifying the representation.
It turns out that on a symplectic manifold, there exists a unique nontrivial trace map (up
to a coefficient rescaling) on the deformation quantized algebra. In this section, we present an
explicit construction of this trace map in terms of the quantum Hochschild-Kostant-Rosenberg
(HKR) map in Section 3.6.5 and Fedosov’s abelian connection.

4.4.1 Differential Graded Algebra

Differential Graded Algebra (DGA)

Definition 4.4.1. A differential graded algebra (DGA) consists of a graded vector space
M
A= An
n∈Z

together with an associative product · and a linear map d : A → A (the differential) such that

⃝
1 the product · preserves the degree Am · An ⊆ Am+n

⃝
2 the differential d increases the degree by one: d : An → An+1 , and satisfies d2 = 0

⃝
3 graded Leibniz rule holds: for any a1 , a2 ∈ A

d(a1 · a2 ) = d(a1 ) · a2 + (−1)|a1 | a1 · d(a2 )

Here we write |a| for the degree of a, i.e., |a| = p for a ∈ Ap .

230
 For a graded algebra A, we will use [−, −] to denote the graded commutator

[a1 , a2 ] := a1 a2 − (−1)|a1 ||a2 | a2 a1 .

If [a1 , a2 ] = 0 for all a1 , a2 ∈ A, we say A is graded commutative.

Example 4.4.2. The de Rham complex (Ω• (M ), d) on a manifold M is a DGA. The differential
d is the de Rham differential. Ω• (M ) is graded commutative.

Example 4.4.3. Let (V, ω) be a linear symplectic space with linear coordinate {y i }. Let

b ) := R[[y i , dy i , λ]]
Ω(V

denote R[[λ]]-valued formal differential forms on V . Let dˆ denote the de Rham differential with
respect to variables {y i }
X ∂
dˆ = dy i
∂y i
i
b ) by
The Moyal product ? extends to a product on Ω(V

(fI (y, λ)dy I ) ? (gJ (y, λ)dy J ) := (fI (y, λ) ? gJ (y, λ))dy I ∧ dy J

b ), ?, d)
Here I, J are multi-indices. The triple (Ω(V ˆ defines a DGA, which will be used in Section
4.4.3 to construct the trace map.

Hochschild Chain Complex

In Section 3.6.5, we have discussed the Hochschild chain complex of an associative algebra
b b b b
· · · → Cp (A) → Cp−1 (A) → · · · → C1 (A) → C0 (A)

This generalizes naturally to differential graded algebras. We follow [46].

Let (A, d) be a DGA where d is the differential and

Cp (A) = A⊗p+1 .

The Hochschild differential b is modified to incorporate with the graded signs

P
k
X
p−1
1+ (|ai |+1)
b(a0 ⊗ a1 ⊗ · · · ⊗ ap ) := (−1) i=0 a0 ⊗ · · · ⊗ ak ak+1 ⊗ · · · ⊗ ap
k=0
P
p−1
|ap |+(|ap |+1) (|ai |+1)
+ (−1) i=0 ap a0 ⊗ a1 ⊗ · · · ⊗ ap−1 .

The differential d induces a differential on Hochschild chains (still called d) by

P
k−1
X
p
(|ai |+1)
d(a0 ⊗ a1 ⊗ · · · ⊗ ap ) := (−1) i=0 a0 ⊗ · · · ⊗ dak ⊗ · · · ⊗ ap .
k=0

It is straight-forward to check that

b2 = d2 = 0, bd + db = 0

231
 The chain complex C• (A) is double graded. One grading p is the Hochschild degree Cp (A).
There is another internal degree q induced from the tensor of the graded space of A
M
Cp (A)q = Aj 0 ⊗ Aj 1 ⊗ · · · ⊗ Aj p
j0 +···+jp =q

We have
b : Cp (A)q −→ Cp−1 (A)q
d : Cp (A)q −→ Cp (A)q+1
The Hochschild chain complex of the DGA (A, d) becomes the total complex of the double
complex C• (A) with differential d + δ.
It will be useful to introduce the following operations on Hochschild chains.

Definition 4.4.4. Given a ∈ A, we define

ada : A −→ A by ada (u) := [a, u] = au − (−1)|a||u| ua

and extend ada to Hochschild chains

ada : C• (A) −→ C• (A)

by
P
k−1
X
p
|a| (|ai |+1)
ada (a0 ⊗ a1 ⊗ · · · ⊗ ap ) = (−1) i=0 a0 ⊗ · · · ⊗ ada ak ⊗ · · · ⊗ ap .
k=0
For example, we have

ada (a0 ⊗ a1 ) = [a, a0 ] ⊗ a1 + (−1)|a|(|a0 |+1) a0 ⊗ [a, a1 ].

Definition 4.4.5. Given a ∈ A, we define

∧a : C• (A) −→ C• (A)

by
P
k
X
p
|a|+(|a|+1) (|ai |+1)
∧a (a0 ⊗ a1 ⊗ · · · ⊗ ap ) := (−1) i=0 a0 ⊗ · · · ⊗ ak ⊗ a ⊗ ak+1 ⊗ · · · ⊗ ap .
k=0
For example, we have

∧a (a0 ⊗ a1 ) = (−1)|a|+(|a|+1)(|a0 |+1) a0 ⊗ a ⊗ a1 + (−1)|a|+(|a|+1)(|a0 |+|a1 |+2) a0 ⊗ a1 ⊗ a.

Proposition 4.4.6. The following relations hold

ad[a,ã] = ada adã −(−1)|a||ã| adã ada
∧a ∧ã = (−1)(|a|+1)(|ã|+1) ∧ã ∧a
ada = b ∧a −(−1)|a|+1 ∧a b
 
∧ada ã = (−1)|a| ada ∧ã − (−1)|a|(|ã|+1) ∧ã ada

adda = d ada −(−1)|a| ada d

 
∧da = − d ∧a −(−1)|a|+1 ∧a d

Proof: Direct computation.

232
Twisting Construction

Let a ∈ A1 be a degree 1 element. Consider the twisting of the differential d by

da := d + ada .

Such da also satisfies the Leibniz rule

da (a1 · a2 ) = (da a1 ) · a2 + (−1)|a1 | a1 · da (a2 ).

By Proposition 4.4.6, we have

d2a = adda+ 1 [a,a] .

2

Definition 4.4.7. A Maurer-Cartan element of (A, d) is a degree 1 element a ∈ A1 satisfying

1
da + [a, a] = 0.
2
This is also called the Maurer-Cartan equation. Since |a| = 1, this is the same as

da + a2 = 0.

Thus if a is a Maurer-Cartan element, then (A, da ) defines a new DGA which will be called
the twist of (A, d) by a. The Hochschild chain complex of the twisted DGA is related to the
untwisted one by the following proposition.

Proposition 4.4.8. For any degree 1 element a ∈ A1 , we have

e−∧a (d + b)e∧a = d + b + ada − ∧da+a2 .

In particular if a is a Maurer-Cartan element, then the operator e∧a intertwines the Hochschild
chain complex of the twisted (A, da ) with the untwisted (A, d) in the following sense

(d + b)e∧a = e∧a (d + ada +b).

Proof: By Proposition 4.4.6,

[d, ∧a ] = −∧da
[b, ∧a ] = ada
[ada , ∧a ] = −∧ada (a) = −2∧a2

It follows that

1
e−∧a (d + b)e∧a = (d + b) + [d + b, ∧a ] + [[d + b, ∧a ] , ∧a ] = d + b + ada − ∧da − ∧a2 .
2

233
4.4.2 Quantum HKR Map

In Section 3.6.5, we have explained how S 1 -correlation on the phase space R2n leads to a
quantum HKR map. We reformulate this construction on the formal Weyl algebra and explain
how to glue it to the Weyl bundle on a symplectic manifold.

Local Case

Recall the S 1 -propagator in Section 3.6.4

X i 2πin(θ2 −θ1 )
G(θ1 , θ2 ) = g(θ2 − θ1 ) = e .
2πn
n∈Z\{0}

The periodic function g(θ) is explicitly given by


0 θ=0
g(θ) =
θ − 1 0<θ<1
2
Let (V, ω) be a linear symplectic space with
1X
ω= ωij dy i ∧ dy j .
2
i,j

The associated formal Weyl algebra is

b ) = (O(V )[[λ]], ∗) = (R[[y i , λ]], ∗).

W(V

The following definition is a reformulation of Definition 3.6.12 into the current setting.

Definition 4.4.9. We define the S 1 -correlation map

b ) ⊗ · · · ⊗ W(V
h−iS 1 : W(V b ) −→ W(V
b )

by

hf0 ⊗ f1 ⊗ · · · ⊗ fm iS 1
ˆ λ P
2
ω ij G(θα ,θβ ) ∂(α) ∂(β)
:= m! dθ0 dθ1 · · · dθm e i,j,α,β ∂y
i
∂y
j f0 (y(0) , λ) · · · fm (y(m) , λ) .
(α)
Cycm+1 (S 1 ) yi =y i

Here

Cycm+1 (S 1 ) = (θ0 , θ1 , · · · , θm ) ∈ (S 1 )m+1 | θ0 , θ1 , · · · , θm are distinct in clockwise cyclic order .

θ0
θ1

θm

θ2

···

234
 The following lemma will be useful when we discuss gluing the S 1 -correlation map on
general symplectic manifolds.
b )
Lemma 4.4.10. Assume h(y) is a constant or linear in y. Then for any fi ∈ W(V

X
m−1
hf0 ⊗ · · · ⊗ fi ⊗ h ⊗ fi+1 ⊗ · · · ⊗ fm iS 1 + hf0 ⊗ · · · ⊗ fm ⊗ hiS 1 = h hf0 ⊗ · · · ⊗ fm iS 1
i=0

Proof: This is clearly true when h = constant. Assume h is linear in y. Using the explicit
formula in Definition 4.4.9,
X
m−1
hf0 ⊗ · · · ⊗ fi ⊗ h ⊗ fi+1 ⊗ · · · ⊗ fm iS 1 + hf0 ⊗ · · · ⊗ fm ⊗ hiS 1
i=0
ˆ 1
= h hf0 ⊗ · · · ⊗ fm iS 1 + terms involving dθi G(θi , θj ).
0

The proposition follows by observing that

ˆ 1 ˆ 1  
1
dθi G(θi , θj ) = dθ θ − = 0.
0 0 2

Let us denote
b ) := R[[y i , dy i , λ]]
Ω(V

which represents R[[λ]]-valued formal differential forms on V . Let

X ∂
db = dy i
∂y i
i

denote the de Rham differential in {y i }’s.

Definition 4.4.11. We define the quantum HKR map

b ) ⊗ · · · ⊗ W(V
ρλ : W(V b ) −→ Ω(V
b )

by
1 D b 1 ⊗ · · · ⊗ df
bm
E
ρλ (f0 ⊗ f1 ⊗ · · · ⊗ fm ) = f0 ⊗ df .
m! S1
b = P ∂yi f dy i and
Here df
i
D E X
b 1 ⊗ · · · ⊗ df
f0 ⊗ df bm := f 0 ⊗ ∂ y i1 f 1 ⊗ · · · ⊗ ∂ y im f m dy i1 ∧ · · · ∧ dy im .
S1 S1
i1 ,··· ,im

Proposition 4.4.12. The quantum HKR map is a chain map

b )), b) −→ (Ω
ρλ : (C• (W(V b • (V ), λ∆).
P
Here ∆ = ω ij ∂y∂ i ι∂yj is the Lie derivative with respect to the Poisson bi-vector ω −1 .
i,j

Proof: The proof is the same as in Section 3.6.5.

235
Global Case

Let (M, ω) be a symplectic manifold and ∇ be a symplectic connection. Let W = W(M )

be the Weyl bundle and we define similarly the bundle of formal differential forms on T M
∞
Y
b = Ω(T
Ω b M ) := Symk (T ∗ M ) ⊗ ∧• T ∗ M [[λ]].
k=0

We also define the bundle C• (W) of Hochschild chains of W whose fiber over p ∈ M is the
b p M ), i.e.
Hochschild chains of W(T
∞
!⊗p+1
Y
Cp (W) := Symk (T ∗ M )[[λ]] .
k=0

The Hochschild differential b defines bundle maps

b : Cp (W) −→ Cp−1 (W)

and (C• (W), b) forms a chain complex of bundles.

The fiberwise quantum HKR map leads to a chain map of complexes of bundles

b • , λ∆).
ρλ : (C• (W), b) −→ (Ω

Here ∆ is the fiberwise BV operator, which in local coordinates is

X ∂
∆= ω ij (x) ι∂ .
∂y i yj
i,j

As bundle maps, all b, ∆ and ρλ are C ∞ (M )-linear.

b via the tensor
The symplectic connection ∇ also induces a connection (still called ∇) on Ω
construction. Since ω is ∇-parallel, both b and ∆ are ∇-parallel

∇(b) = ∇(∆) = 0.

Since ρλ is constructed out of ω and the S 1 -propagator G(θ1 , θ2 ), we know ρλ is also ∇-

parallel. Thus we have a commutative diagram

ρλ
Cp (W) bp
Ω
b λ∆

ρλ
Cp−1 (W) b p−1
Ω

where all maps in this diagram are ∇-parallel bundle maps.

4.4.3 Twisting by Fedosov’s Abelian Connection

Let D = ∇ + 1
λ [γ, −]∗ be a Fedosov’s abelian connection as in Theorem 4.2.7. It gives a
flat connection on W and extends to

D : Ω• (M, W) → Ω•+1 (M, W)

236
via the following Leibniz rule

D(α ⊗ s) = dα ⊗ s + (−1)|p| α ∧ ∇s, α ∈ Ωp (M ), s ∈ W.

The flatness of D implies D2 = 0 on Ω• (M, W). Thus we have a complex

(Ω• (M, W), D)

which is called the de Rham complex with values in W.

Since D is compatible with the fiberwise Moyal product, (Ω• (M, W), D) becomes a DGA
with Ωp (M, W) the degree p piece. We denote this DGA as

Ω(M, W) := (Ω• (M, W), D)

Thus we have the Hochschild chain complex as in Section 4.4.1

C• (Ω(M, W), D + b).

Let
db : Ω
b • −→ Ω
b •+1

denote the fiberwise de Rham differential. In local coordinate

X ∂
db = dy i .
∂y i
i

This db is a bundle map and compatible with the fiberwise Moyal product:

b 1 ∗ s2 ) = d(s
d(s b 1 ) ∗ s2 + s1 ∗ (ds
b 2 ), ∀s1 , s2 ∈ W.

We extend db to a C ∞ (M )-linear map

db : Ω• (M, Ω
b • ) −→ Ω• (M, Ω
b •+1 )

by
b ⊗ u) = (−1)p α ⊗ du,
d(α b b •.
α ∈ Ωp (M ), u ∈ Ω

Let
b • ) −→ Ω•+1 (M, Ω
∇ : Ω• (M, Ω b •)

be the map induced on tensors from the symplectic connection ∇. Let xi be local coordinates

on M and y i be the induced fiberwise coordinates on T M . Assume the connection ∇ acts as
X
∇y i = αji (x)y j ,
j

then  
X X
∇dy i = αji (x)dy j = −db b i.
αji (x)y j  = −d∇y
j j

237
A slight generalization of this computation shows

∇db + d∇
b =0 b • ).
on Ω• (M, Ω

which simply says db is ∇-parallel.

Similarly we extend the BV operator ∆ to

b • ) → Ω• (M, Ω
∆ : Ω• (M, Ω b •−1 )

by
∆(α ⊗ u) = (−1)|α| α ⊗ ∆U, b •.
α ∈ Ωp (M ), u ∈ Ω

Then ∆ is compatible with the symplectic connection ∇ in the sense

∇∆ + ∆∇ = 0 b • ).
on Ω• (M, Ω

b • to be valued in Ω• (M )
We also extend the quantum HKR map ρλ : C• (W) −→ Ω

b •)
ρλ : C• (Ω(M, W)) −→ Ω(M, Ω

by
P
m−1
k|αk |
ρ (α0 u0 ⊗ α1 u1 ⊗ · · · ⊗ αm um ) := (−1) k=1
λ
(α0 ∧ · · · ∧ αm ) ρλ (u0 ⊗ · · · ⊗ um ) .

Here αi ∈ Ω|αi | (M ) and ui ∈ W. The following diagram commutes

ρλ
C• (Ω(M, W)) b •)
Ω(M, Ω
b λ∆
ρλ
C• (Ω(M, W)) b •)
Ω(M, Ω

Now we are ready to incorporate Fedosov’s abelian connection D = ∇ + 1

λ [γ, −]∗ . Let us
consider the twisting of Hochschild chains by γ ∈ Ω1 (M, W)

e∧γ/λ : C• (Ω(M, W)) −→ C• (Ω(M, W)[λ−1 ]).

Here [λ−1 ] means localizing λ

(−)[λ−1 ] := (−) ⊗R[[λ]] R((λ)).

By the same computation as in Proposition 4.4.8

1
e−∧γ/λ (∇ + b)e∧γ/λ = ∇ + b + adγ/λ −∧∇(γ/λ)+γ/λ∗γ/λ = D + b − ∧ ∇ .
λ Rω +ωh̄
Here we have used ∇γ + 1
2λ [γ, γ]∗ = Rω∇ + ωh̄ in Theorem 4.2.7. Thus
 
1
e∧γ/λ (D + b) = ∇ + b + ∧Rω∇ +ωh̄ e∧γ/λ .
λ

238
Definition 4.4.13. We define the Fedosov-twisted quantum HKR map by

b • )[λ−1 ].
ρλD := ρλ ◦ e∧γ/λ : C• (Ω(M, W)) −→ Ω(M, Ω

The following proposition is a reformulation of the BV quantization result in [23] [28].

Proposition 4.4.14. The following diagram commutes

ρλ
C• (Ω(M, W)) D b • )[λ−1 ]
Ω(M, Ω
1 b ∇
D+b ∇+h̄∆+ λ dRω
ρλ
C• (Ω(M, W)) D b • )[λ−1 ]
Ω(M, Ω

Equivalently, ρλD gives a chain map of complexes

 
b • −1 1b ∇
ρD : (C• (Ω(M, W)), D + b) −→ Ω(M, Ω )[λ ], ∇ + h̄∆ + dRω .
λ
λ

Proof: Observe
b ∇ + ωh̄ ) = dR
d(R b ∇
ω ω

which is linear in y. It follows from the defintion of ρλ and Lemma 4.4.10 that

b ∇ · ρλ (−).
ρλ ◦ ∧Rω∇ +ωh̄ (−) = dR ω

Thus
ρλD ◦ (D + b) = ρλ ◦ e∧γ/λ ◦ (D + b)
 
1
= ρ ◦ ∇ + b + ∧Rω∇ +ωh̄ ◦ e∧γ/λ
λ
λ
 
1b ∇
= ∇ + h̄∆ + dR ◦ ρλ ◦ e∧γ/λ
λ ω
 
1b ∇
= ∇ + h̄∆ + dRω ◦ ρλD .
λ

4.4.4 Trace Map and Index

Definition 4.4.15. We define the bundle map (here dim M = 2n)

ˆ
b • )[λ−1 ] −→ Ω(M )((λ))
: Ω(M, Ω
Ber

by  n
ˆ
1  1X
s(x, y, λ) := ω ij (x)ι∂yi ι∂yj  s .
Ber n! 2
i,j y i =0
dy i =0
´
Remark 4.4.16. The map Ber is a version of Berezin integral on Grassmann variables.

239
 ´
Note that b 2n ). Let us denote
= 0 unless s ∈ Ω(M, Ω
Ber s

1X
ω̂ := ωij (x)dy i ∧ dy j
2
i,j

Then for any α ∈ Ω(M ), we have ˆ

ω̂
α = α.
Ber n!
´
Proposition 4.4.17. Ber is a chain map intertwining
ˆ  
b • −1 1b ∇
: Ω(M, Ω )[λ ], ∇ + λ∆ + dRω −→ (Ω(M )((λ)), d).
Ber λ

Here d is the de Rham differential on M .

b • ).
Proof: For any s ∈ Ω(M, Ω
ˆ ˆ
∆(s) = 0 and b ∇∧s=0
dR ω
Ber Ber
´
b ∇ contains y which goes to zero under
b 2n while dR
since ∆(s) can not lie in Ω ω Ber .
Let us assume s is of the form s = ω̂
α n! where α ∈ Ω(M ). Since ∇ is a sympletic connection,
∇ω̂ = 0. Thus ˆ   ˆ ˆ
ω̂ ω̂ ω̂
∇ α = dα = dα = d α
Ber n! Ber n! Ber n!
This proves the proposition.
´
Proposition 4.4.18. The composition of the Fedosov-twisted quantum HKR map with Ber
gives a chain map
ˆ
◦ρλD : (C• (Ω• (M, W)), D + b) −→ (Ω• (M )((λ)), d).
Ber

Proof: This follows from Proposition 4.4.14 and Proposition 4.4.17.

Recall that the space ΓfDlat (M, W) of flat sections of W defines a deformation quantization.
´
Restricting Ber ◦ρλD to ΓfDlat (M, W), we obtain a chain map
ˆ
◦ρλD : (C• (ΓfDlat (M, W)), b) −→ (Ω• (M )((λ)), d).
Ber

Definition 4.4.19. Define the trace map on the deformation quantized algebra ΓfDlat (M, W)

Tr : ΓfDlat (M, W) → R((λ))

by ˆ ˆ ˆ ˆ
Tr(s) := ρλD (s) = ρλ (e∧γ/λ (s)).
M Ber M Ber

Remark 4.4.20. This approach to trace in Definition 4.4.19 is equivalent to [23, Definition 2.41].

Theorem 4.4.21. The trace map Tr satisfies the following properties

240
 1 For f ∈ C ∞ (M ),
⃝ ˆ 
−1 (−1)n ωn
Tr σ (f ) = f + O(λ) .
n!λn M n!

⃝
2 For any s1 , s2 ∈ ΓfDlat (M, W) which are compactly supported on M ,

Tr(s1 ∗ s2 ) = Tr(s2 ∗ s1 ).

1 the leading λ-order contribution of Tr σ −1 (f ) comes from

Proof: For ⃝,
ˆ ˆ
ρλ (f ⊗ (γ0 /λ)⊗2n )
M Ber
ˆ ˆ * +
1 b ⊗ · · · ⊗ dγ
b0
= f ⊗ dγ
M Ber (2n)!λ
2n | 0 {z }
2n S1
ˆ ˆ  
1 b 0 )2n + O(λ)
= f (dγ
(2n)!λ2n M Ber
P ˆ 0 = − P ωij dxi dy j .
Here γ0 = ωij dxi y j ∈ Ω1 (M, W), and dγ
i,j i,j
Using a Darboux coordinate, it is straight-foward to compute
1 b 2n ω n ω̂
(dγ0 ) = (−1)n .
(2n)! n! n!

Then ⃝
1 follows from
ˆ ˆ ˆ ˆ n ˆ
1 b 2n n ω ω̂ n ωn
f (dγ0 ) = (−1) f = (−1) f .
(2n)! M Ber M Ber n! n! M n!

For ⃝,
2 by Proposition 4.4.18 we have

Tr(s1 ∗ s2 − s2 ∗ s1 )
= Tr(b(s1 ⊗ s2 ))
ˆ ˆ
= ρλD (b(s1 ⊗ s2 ))
M Ber
ˆ ˆ
= d ρλD (s1 ⊗ s2 ) = 0.
M Ber

One remarkable property of the trace map is

ˆ
Tr(1) = e−ωλ /λ Â(M )
M

Here ωλ appears in Theorem 4.2.7 which parametrizes the deformation quantization. This
formula is called the Algebraic Index Theorem which was established by Fedosov [19] and Nest-
Tsygan [40] as the algebraic analogue of Atiyah-Singer Index theorem. One geometric approach
to compute the algebraic index is to use the explicit trace map (Definition 4.4.19) and reduce it
to an exact semi-classical calculation in terms of Feynman diagrams. This is explained in [23]
[28], and Â(M ) appears as the one-loop contribution.

241
 Bibliography

[1] Arnol d, Vladimir I., Alexander B. Givental, Sergei P. Novikov. Symplectic Geometry. In:
Arnold, V.I., Novikov, S.P. (eds) Dynamical Systems IV. Encyclopaedia of Mathematical
Sciences, vol 4. Springer, Berlin, Heidelberg. (2001)

[2] Bates, Sean, and Alan Weinstein. Lectures on the Geometry of Quantization. Vol. 8. Amer-
ican Mathematical Soc., 1997.

[3] Bayen, François, Moshé Flato, Christian Fronsdal, André Lichnerowicz, and Daniel Stern-
heimer. Deformation theory and quantization. I. Deformations of symplectic structures.
Annals of Physics 111, no. 1 (1978): 61-110. Deformation theory and quantization. II.
Physical applications. Annals of Physics 111, no. 1 (1978): 111-151.

[4] Berezin, Felix Aleksandrovich. Feynman path integrals in a phase space. Soviet Physics
Uspekhi 23.11 (1980): 763.

[5] Bordemann, Martin. Deformation quantization: a survey. Journal of Physics: Conference

Series. Vol. 103. No. 1. IOP Publishing, 2008.

[6] Born, Max. Statistical interpretation of quantum mechanics. Science 122.3172 (1955): 675-
679.

[7] Brylinski, Jean-Luc. Loop spaces, characteristic classes and geometric quantization.
Springer Science & Business Media, 2009.

[8] Cattaneo, Alberto S., and Giovanni Felder. Poisson sigma models and deformation quan-
tization. Modern Physics Letters A 16.04n06 (2001): 179-189.

[9] Carlitz, Robert D., and Denis A. Nicole. Classical paths and quantum mechanics. Annals
of Physics 164.2 (1985): 411-462.

[10] Coleman, Sidney. Aspects of symmetry: selected Erice lectures. Cambridge University Press,
1988.

[11] Curtright, Thomas L., and Cosmas K. Zachos. Quantum mechanics in phase space. Asia
Pacific Physics Newsletter 1.01 (2012): 37-46.

[12] De Gosson, Maurice A. Symplectic geometry and quantum mechanics. Vol. 166. Springer
Science & Business Media, 2006.

242
[13] De Wilde, Marc, and Pierre BA Lecomte. Existence of star-products and of formal deforma-
tions of the Poisson Lie algebra of arbitrary symplectic manifolds. Letters in Mathematical
Physics 7 (1983): 487-496.

[14] Dirac, Paul Adrien Maurice. The principles of quantum mechanics. No. 27. Oxford Uni-
versity Press, 1981.

[15] Faddeev, Ljudvig D., and Oleg Aleksandrovich I�A�kubovskiĭ. Lectures on quantum mechan-
ics for mathematics students. Vol. 47. American Mathematical Soc., 2009.

[16] Feynman, Richard P., Albert R. Hibbs, and Daniel F. Styer. Quantum mechanics and path
integrals. Courier Corporation, 2010.

[17] Feynman, Richard Phillips, and Laurie M. Brown. Feynman’s thesis: a new approach to
quantum theory. World Scientific, 2005.

[18] Fedosov, Boris V. A simple geometrical construction of deformation quantization. Journal

of differential geometry 40.2 (1994): 213-238.

[19] Fedosov, Boris V. Deformation quantization and index theory. Mathematical topics 9
(1996).

[20] Feigin, Boris, Giovanni Felder, and Boris Shoikhet. Hochschild cohomology of the Weyl
algebra and traces in deformation quantization. Duke Mathematical Journal 127 (3) 487 -
517, (2005)

[21] Folland, Gerald B. Harmonic analysis in phase space. No. 122. Princeton university press,
1989.

[22] Gel’fand, Israel M., and Akiva M. Yaglom. Integration in functional spaces and its appli-
cations in quantum physics. Journal of Mathematical Physics 1.1 (1960): 48-69.

[23] Grady, Ryan E., Qin Li, and Si Li. Batalin Vilkovisky quantization and the algebraic index.
Advances in Mathematics 317 (2017): 575-639.

[24] Ginzburg, Victor. Lectures on noncommutative geometry. arXiv:math/0506603 (2005).

[25] Griffiths, David J., and Darrell F. Schroeter. Introduction to quantum mechanics. Cam-
bridge University Press, 2018.

[26] Grosche, Christian, Frank Steiner, and Frank Steiner. Handbook of Feynman path integrals.
Vol. 145. Berlin: Springer, 1998.

[27] Guillemin, Victor, and Shlomo Sternberg. Semi-classical analysis. Boston, MA: Interna-
tional Press, 2013.

243
[28] Gui, Zhengping, Si Li, and Kai Xu. Geometry of localized effective theories, exact semi-
classical approximation and the algebraic index. Communications in Mathematical Physics
382.1 (2021): 441-483.

[29] Hall, Brian C. Holomorphic methods in analysis and mathematical physics. Contemporary
Mathematics 260 (2000): 1-60.

[30] Hall, Brian C. Quantum theory for mathematicians. Springer Publication, 2013.

[31] Hislop, Peter D., and Israel Michael Sigal. Introduction to spectral theory: With applications
to Schrödinger operators. Vol. 113. Springer Science & Business Media, 2012.

[32] Holstein, Barry R., and Arthur R. Swift. Path integrals and the WKB approximation.
American Journal of Physics 50.9 (1982): 829-832.

[33] Kaiser, David. History: Shut up and calculate!. Nature 505.7482 (2014): 153-155.

[34] Kontsevich, Maxim. Deformation quantization of Poisson manifolds. Letters in Mathemat-

ical Physics 66 (2003): 157-216.

[35] Landau, Lev Davidovich, and Evgenii Mikhailovich Lifshitz. Quantum mechanics: non-
relativistic theory. Vol. 3. Elsevier, 2013.

[36] Li, Si. Classical Mechanics and Geometry. Boston, MA: International Press of Boston,
2023.

[37] Li, Si. Electromagnetism and Geometry. Boston, MA: International Press of Boston, 2023.

[38] Littlejohn, Robert G. The semiclassical evolution of wave packets. Physics reports 138.4-5
(1986): 193-291.

[39] Loday, Jean-Louis. Cyclic homology. Vol. 301. Springer Science & Business Media, 2013.

[40] Nest, Ryszard, and Boris Tsygan. Algebraic index theorem. Communications in Mathemat-
ical Physics 172, 223-262 (1995)

[41] Shankar, Ramamurti. Principles of quantum mechanics. Springer Science & Business Me-
dia, 2012.

[42] Sharan, Pankaj. Star-product representation of path integrals. Physical Review D 20.2
(1979): 414.

[43] Susskind, Leonard, and Art Friedman. Quantum mechanics: the theoretical minimum.
Basic Books, 2014.

[44] Takhtadzhian, Leon Armenovich. Quantum mechanics for mathematicians. Vol. 95. Amer-
ican Mathematical Soc., 2008.

244
[45] Teschl, Gerald. Mathematical methods in quantum mechanics. Vol. 157. American Mathe-
matical Soc., 2014.

[46] Tsygan, Boris. Cyclic homology. Cyclic homology in non-commutative geometry. Vol. 121,
73-113. Springer Science & Business Media (2003)

[47] Von Neumann, John. Mathematical foundations of quantum mechanics: New edition. Vol.
53. Princeton University Press, 2018.

[48] Weinberg, Steven. Lectures on quantum mechanics. Cambridge University Press, 2015.

[49] Witten, Edward. Analytic continuation of Chern-Simons theory. AMS/IP Stud. Adv. Math
50 (2011): 347.

[50] Woodhouse, Nicholas Michael John. Geometric quantization. Oxford university press, 1992.

[51] Zinn-Justin, Jean. Path integrals in quantum mechanics. OUP Oxford, 2004.

245