This action might not be possible to undo. Are you sure you want to continue?

**Lecture Notes for Quantum Mechanics IV
**

Riccardo Rattazzi

May 25, 2009

2

Contents

1 The path integral formalism 5

1.1 Introducing the path integrals . . . . . . . . . . . . . . . . . . . . 5

1.1.1 The double slit experiment . . . . . . . . . . . . . . . . . 5

1.1.2 An intuitive approach to the path integral formalism . . . 6

1.1.3 The path integral formulation . . . . . . . . . . . . . . . . 8

1.1.4 From the Schr¨oedinger approach to the path integral . . . 12

1.2 The properties of the path integrals . . . . . . . . . . . . . . . . 14

1.2.1 Path integrals and state evolution . . . . . . . . . . . . . 14

1.2.2 The path integral computation for a free particle . . . . . 17

1.3 Path integrals as determinants . . . . . . . . . . . . . . . . . . . 19

1.3.1 Gaussian integrals . . . . . . . . . . . . . . . . . . . . . . 19

1.3.2 Gaussian Path Integrals . . . . . . . . . . . . . . . . . . . 20

1.3.3 O(ℏ) corrections to Gaussian approximation . . . . . . . . 22

1.3.4 Quadratic lagrangians and the harmonic oscillator . . . . 23

1.4 Operator matrix elements . . . . . . . . . . . . . . . . . . . . . . 27

1.4.1 The time-ordered product of operators . . . . . . . . . . . 27

2 Functional and Euclidean methods 31

2.1 Functional method . . . . . . . . . . . . . . . . . . . . . . . . . . 31

2.2 Euclidean Path Integral . . . . . . . . . . . . . . . . . . . . . . . 32

2.2.1 Statistical mechanics . . . . . . . . . . . . . . . . . . . . . 34

2.3 Perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . 35

2.3.1 Euclidean n-point correlators . . . . . . . . . . . . . . . . 35

2.3.2 Thermal n-point correlators . . . . . . . . . . . . . . . . . 36

2.3.3 Euclidean correlators by functional derivatives . . . . . . 38

2.3.4 Computing K

0

E

[J] and Z

0

[J] . . . . . . . . . . . . . . . . 39

2.3.5 Free n-point correlators . . . . . . . . . . . . . . . . . . . 41

2.3.6 The anharmonic oscillator and Feynman diagrams . . . . 43

3 The semiclassical approximation 49

3.1 The semiclassical propagator . . . . . . . . . . . . . . . . . . . . 50

3.1.1 VanVleck-Pauli-Morette formula . . . . . . . . . . . . . . 54

3.1.2 Mathematical Appendix 1 . . . . . . . . . . . . . . . . . . 56

3.1.3 Mathematical Appendix 2 . . . . . . . . . . . . . . . . . . 56

3.2 The ﬁxed energy propagator . . . . . . . . . . . . . . . . . . . . . 57

3.2.1 General properties of the ﬁxed energy propagator . . . . . 57

3.2.2 Semiclassical computation of K(E) . . . . . . . . . . . . . 61

3.2.3 Two applications: reﬂection and tunneling through a barrier 64

3

4 CONTENTS

3.2.4 On the phase of the prefactor of K(x

f

, t

f

; x

i

, t

i

) . . . . . 68

3.2.5 On the phase of the prefactor of K(E; x

f

, x

i

) . . . . . . . 72

4 Instantons 75

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.2 Instantons in the double well potential . . . . . . . . . . . . . . . 77

4.2.1 The multi-instanton amplitude . . . . . . . . . . . . . . . 81

4.2.2 Summing up the instanton contributions: results . . . . . 84

4.2.3 Cleaning up details: the zero-mode and the computation

of R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

5 Interaction with an external electromagnetic ﬁeld 91

5.1 Gauge freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

5.1.1 Classical physics . . . . . . . . . . . . . . . . . . . . . . . 91

5.1.2 Quantum physics . . . . . . . . . . . . . . . . . . . . . . . 93

5.2 Particle in a constant magnetic ﬁeld . . . . . . . . . . . . . . . . 95

5.2.1 An exercise on translations . . . . . . . . . . . . . . . . . 96

5.2.2 Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.2.3 Landau levels . . . . . . . . . . . . . . . . . . . . . . . . . 99

5.3 The Aharonov-Bohm eﬀect . . . . . . . . . . . . . . . . . . . . . 101

5.3.1 Energy levels . . . . . . . . . . . . . . . . . . . . . . . . . 105

5.4 Dirac’s magnetic monopole . . . . . . . . . . . . . . . . . . . . . 106

5.4.1 On coupling constants, and why monopoles are diﬀerent

than ordinary charged particles . . . . . . . . . . . . . . . 107

Chapter 1

The path integral formalism

1.1 Introducing the path integrals

1.1.1 The double slit experiment

One of the important experiments that show the fundamental diﬀerence between

Quantum and Classical Mechanics is the double slit experiment. It is interesting

with respect to the path integral formalism because it leads to a conceptual

motivation for introducing it.

Consider a source S of approximately monoenergetic particles, electrons for

instance, placed at position A. The ﬂux of electrons is measured on a screen

C facing the source. Imagine now placing a third screen in between the others,

with two slits on it, which can be opened or closed (see ﬁgure 1.1). When the

ﬁrst slit is open and the second closed we measure a ﬂux F

1

, when the ﬁrst slit

is closed and the second open we measure a ﬂux F

2

and when both slits are

open we measure a ﬂux F.

A B C

1

2

S

Figure 1.1: The double slit experiment: an electron source is located somewhere

on A, and a detector is located on the screen C. A screen B with two slits 1 and

2 that can be open or closed is placed in between, so that the electrons have to

pass through 1 or 2 to go from A to C. We measure the electron ﬂux on the

screen C.

In classical physics, the ﬂuxes at C in the three cases are expected to satisfy

the relation F = F

1

+ F

2

. In reality, one ﬁnds in general F = F

1

+ F

2

+ F

int

,

and the structure of F

int

precisely corresponds to the interference between two

5

6 CHAPTER 1. THE PATH INTEGRAL FORMALISM

waves passing respectively through 1 and 2:

F = [Φ

1

+ Φ

2

[

2

= [Φ

1

[

2

. ¸¸ .

F1

+[Φ

2

[

2

. ¸¸ .

F2

+2 Re (Φ

∗

1

Φ

2

)

. ¸¸ .

Fint

(1.1)

How can we interpret this result? What is the electron?

More precisely: Does the wave behaviour imply that the electron is a delo-

calized object? That the electron is passing through both slits?

Actually not. When detected, the electron is point-like, and more remark-

ably, if we try to detect where it went through, we ﬁnd that it either goes

through 1 or 2 (this is done by setting detectors at 1 and 2 and considering a

very weak ﬂux, in order to make the probability of coinciding detections at 1

and 2 arbitrarily small).

According to the Copenhagen interpretation, F should be interpreted as a

probability density. In practice this means: compute the amplitude Φ as if

dealing with waves, and interpret the intensity [Φ[

2

as a probability density for

a point-like particle position.

How is the particle/wave duality not contradictory? The answer is in the

indetermination principle. In the case at hand: when we try to detect which

alternative route the electron took, we also destroy interference. Thus, another

formulation of the indetermination principle is: Any determination of the al-

ternative taken by a process capable of following more than one alternative

destroys the interference between the alternatives.

Resuming:

• What adds up is the amplitude Φ and not the probability density itself.

• The diﬀerence between classical and quantum composition of probabilities

is given by the interference between classically distinct trajectories.

In the standard approach to Quantum Mechanics, the probability ampli-

tude is determined by the Schr¨odinger equation. The “Schr¨ odinger” viewpoint

somehow emphasizes the wave properties of the particles (electrons, photons,

. . . ). On the other hand, we know that particles, while they are described by

“a probability wave”, are indeed point-like discrete entities. As an example, in

the double slit experiment, one can measure where the electron went through

(at the price of destroying quantum interference).

This course will present an alternative, but fully equivalent, method to com-

pute the probability amplitude. In this method, the role of the trajectory of a

point-like particle will be formally “resurrected”, but in a way which is com-

patible with the indetermination principle. This is the path integral approach

to Quantum Mechanics. How can one illustrate the basic idea underlying this

approach?

1.1.2 An intuitive approach to the path integral formalism

In the double slit experiment, we get two interfering alternatives for the path of

electrons from A to C. The idea behind the path integral approach to Quantum

Mechanics is to take the implications of the double slit experiment to its extreme

consequences. One can imagine adding extra screens and drilling more and more

1.1. INTRODUCING THE PATH INTEGRALS 7

holes through them, generalizing the result of the double slit experiment by the

superposition principle. This is the procedure illustrated by Feynman in his

book “Quantum Mechanics and Path Integrals”.

Schematically:

• With two slits: we know that Φ = Φ

1

+ Φ

2

• If we open a third slit, the superposition principle still applies: Φ = Φ

1

+

Φ

2

+ Φ

3

• Imagine then adding an intermediate screen D with N holes at positions

x

1

D

, x

2

D

, . . . , x

N

D

(see ﬁgure 1.2). The possible trajectories are now labelled

by x

i

D

and α = 1, 2, 3, that is by the slit they went through at D and by

the slit they went through at B.

A B C D

1

2

3

1

2

N −1

N

Figure 1.2: The multi-slit experiment: We have a screen with N slits and a

screen with three slits placed between the source and the detector.

Applying the superposition principle:

Φ =

N

i=1

α=1,2,3

Φ

_

x

i

D

, α

_

Nothing stops us from taking the ideal limit where N → ∞ and the holes

ﬁll all of D. The sum

i

becomes now an integral over x

D

.

Φ =

α=1,2,3

_

dx

D

Φ(x

D

, α)

D is then a purely ﬁctious device! We can go on and further reﬁne our

trajectories by adding more and more ﬁctious screens D

1

,D

2

,. . . ,D

M

Φ =

α=1,2,3

_

dx

D1

dx

D2

dx

DM

Φ(x

D1

, x

D2

, . . . , x

DM

; α)

In the limit in which D

i

, D

i+1

become inﬁnitesimally close, we have speciﬁed

all possible paths x(y) (see ﬁgure 1.3 for a schematic representation, where the

screen B has been dropped in order to consider the simpler case of propagation

in empty space).

8 CHAPTER 1. THE PATH INTEGRAL FORMALISM

y

i

y

1

y

2

y

3

y

f

x(y

i

) x(y

1

)

x(y

2

)

x(y

3

)

x(y

f

)

x

y

Figure 1.3: The paths from (x

i

, y

i

) to (x

f

, y

f

) are labelled with the functions

x(y) that satisfy x(y

i

) = x

i

and x(y

f

) = x

f

.

In fact, to be more precise, also the way the paths are covered through time is

expected to matter (t →(x(y), y(t))). We then arrive at a formal representation

of the probability amplitude as a sum over all possible trajectories:

Φ =

All

trajectories

{x(t),y(t)}

Φ(¦x¦) (1.2)

How is this formula made sense of? How is it normalized?

1.1.3 The path integral formulation

Based on the previous section, we will start anew to formulate quantum me-

chanics. We have two guidelines:

1. We want to describe the motion from position x

i

at time t

i

to position x

f

at time t

f

with a quantum probability amplitude K(x

f

, t

f

; x

i

, t

i

) given by

K(x

f

, t

f

; x

i

, t

i

) =

All

trajectories

Φ(¦γ¦)

where ¦γ¦ is the set of all trajectories satisfying x(t

i

) = x

i

, x(t

f

) = x

f

.

2. We want classical trajectories to describe the motion in the formal limit

ℏ → 0. In other words, we want classical physics to be resurrected in the

ℏ →0 limit.

Remarks:

• ℏ has the dimensionality [Energy] [Time]. That is also the dimen-

sionality of the action S which describes the classical trajectories via the

principle of least action.

1.1. INTRODUCING THE PATH INTEGRALS 9

• One can associate a value of S[γ] to each trajectory. The classical trajec-

tories are given by the stationary points of S[γ] (δS[γ] = 0).

It is thus natural to guess: Φ[γ] = f(S[γ]/ℏ), with f such that the classical

trajectory is selected in the formal limit ℏ →0. The speciﬁc choice

Φ[γ] = e

i

S[γ]

ℏ

(1.3)

implying

K(x

f

, t

f

; x

i

, t

i

) =

{γ}

e

i

S[γ]

ℏ

(1.4)

seems promising for two reasons:

1. The guideline (2.) is heuristically seen to hold. In a macroscopic, classical,

situation the gradient δS/δγ is for most trajectories much greater than ℏ.

Around such trajectories the phase e

iS/ℏ

oscillates extremely rapidly and

the sum over neighbouring trajectories will tend to cancel.

1

(see ﬁgure

1.4).

x

i

x

f

γ

1

γ

2

Figure 1.4: The contributions from the two neighbouring trajectories γ

1

and γ

2

will tend to cancel if their action is big.

x

i

x

f

γ

cl

Figure 1.5: The contributions from the neighbouring trajectories of the classical

trajectory will dominate.

On the other hand, at a classical trajectory γ

cl

the action S[γ] is stationary.

Therefore in the neighbourhood of γ

cl

, S varies very little, so that all

trajectories in a tube centered around γ

cl

add up coherently (see ﬁgure

1.5) in the sum over trajectories. More precisely: the tube of trajectories

1

By analogy, think of the integral

R

+∞

−∞

dxe

if(x)

where f(x) ≡ ax plays the role of S/ℏ:

the integral vanishes whenever the derivative of the exponent f

′

= a is non-zero

10 CHAPTER 1. THE PATH INTEGRAL FORMALISM

in question consists of those for which [S −S

cl

[ ≤ ℏ and deﬁnes the extent

to which the classical trajectory is well deﬁned. We cannot expect to deﬁne

our classical action to better than ∼ ℏ. However, in normal macroscopic

situations S

cl

≫ℏ. In the exact limit ℏ →0, this eﬀect becomes dramatic

and only the classical trajectory survives.

Once again, a simple one dimensional analogy is provided by the inte-

gral

_

+∞

−∞

dxe

ix

2

/h

, which is dominated by the region x

2

<

∼

h around the

stationary point x = 0.

2. Eq. (1.3) leads to a crucial composition property. Indeed the action for

a path γ

12

obtained by joining two subsequent paths γ

1

and γ

2

, like in

ﬁg. 1.6, satisﬁes the simple additive relation S[γ

12

] = S[γ

1

]+S[γ

2

]. Thanks

to eq. (1.3) the additivity of S translates into a factorization property for

the amplitude: Φ[γ

12

] = Φ[γ

1

]Φ[γ

2

] which in turn leads to a composition

property of K, which we shall now prove.

Consider indeed the amplitudes for three consecutive times t

i

< t

int

< t

f

.

The amplitude K(x

f

, t

f

; x

i

, t

i

) should be obtainable by evolving in two

steps: ﬁrst from t

i

to t

int

(by K(y, t

int

; x

i

, t

i

) for any y ), and second from

t

int

to t

f

(by K(x

f

, t

f

; y, t

int

). Now, obviously each path with x(t

i

) = x

i

and x(t

f

) = x

f

can be obtained by joining two paths γ

1

(y = x(t

int

)) from

t

i

to t

int

and γ

2

(y = x(t

int

)) from t

int

to t

f

(see ﬁgure 1.6).

t

i

t

int

t

f

x

i

y

x

f

γ

1

γ

2

x

t

Figure 1.6: The path from (x

i

, t

i

) to (x

f

, t

f

) can be obtained by summing the

paths γ

1

from (x

i

, t

i

) to (y, t

int

) with γ

2

from (y, t

int

) to (x

f

, t

f

).

1.1. INTRODUCING THE PATH INTEGRALS 11

Thanks to eq. (1.3) we can then write:

_

dyK(x

f

, t

f

; y, t

int

)K(y, t

int

; x

i

, t

i

)

=

γ1,γ2

_

dye

i

ℏ

(S[γ1(y)]+S[γ2(y)])

=

γ(y)=γ2(y)◦γ1(y)

_

dye

i

ℏ

S[γ(y)]

= K(x

f

, t

f

; x

i

, t

i

) . (1.5)

Notice that the above composition rule is satisﬁed in Quantum Mechanics

as easily seen in the usual formalism. It is the quantum analogue of the

classical composition of probabilities

P

1→2

=

α

P

1→α

P

α→2

. (1.6)

In quantum mechanics what is composed is not probability P itself but

the amplitude K which is related to probability by P = [K[

2

.

It is instructive to appreciate what would go wrong if we modiﬁed the choice

in eq. (1.3). For instance the alternative choice Φ = e

−S[γ]/ℏ

satisﬁes the compo-

sition property but does not in general select the classical trajectories for ℏ →0.

This alternative choice would select the minima of S but the classical trajec-

tories represent in general only saddle points of S in function space. Another

alternative Φ = e

i(S[γ]/ℏ)

2

, would perhaps work out in selecting the classical

trajectories for ℏ →0, but it would not even closely reproduce the composition

property we know works in Quantum Mechanics. (More badly: this particular

choice, if S were to represent the action for a system of two particles, would

imply that the amplitudes for each individual particle do not factorize even in

the limit in which they are very far apart and non-interacting!).

One interesting aspect of quantization is that a fundamental unit of action

(ℏ) is introduced. In classical physics the overall value of the action in unphys-

ical: if

S(q, ˙ q) → S

λ

≡ λS(q, ˙ q) (1.7)

the only eﬀect is to multiply all the equations of motion by λ, so that the

solutions remain the same (the stationary points of S and S

λ

coincide).

Quantum Mechanics sets a natural unit of measure for S. Depending on the

size of S, the system will behave diﬀerently:

• large S → classical regime

• small S → quantum regime

In the large S limit, we expect the trajectories close to γ

cl

to dominate K

(semiclassical limit). We expect to have: K(x

f

; x

i

) ∼ (smooth function) e

i

S

cl

ℏ

.

In the S ∼ ℏ limit, all trajectories are comparably important: we must sum

them up in a consistent way; this is not an easy mathematical task.

12 CHAPTER 1. THE PATH INTEGRAL FORMALISM

Due to the mathematical diﬃculty, rather than going on with Feynman’s

construction and show that it leads to the same results of Schr¨odinger equation,

we will follow the opposite route which is easier: we will derive the path integral

formula from Schr¨odinger’s operator approach.

1.1.4 From the Schr¨oedinger approach to the path integral

Consider the transition amplitude:

K(x

f

, t

f

; x

i

, t

i

) ≡ ¸x

f

[ e

−

iH(t

f

−t

i

)

ℏ

[x

i

) = ¸x

f

[ e

−

iHt

ℏ

[x

i

) (1.8)

where we used time translation invariance to set: t

i

= 0, t

f

−t

i

= t.

To write it in the form of a path integral, we divide t into N inﬁnitesimal

steps and consider the amplitude for each inﬁnitesimal step (see ﬁgure 1.7). We

label the intermediate times t

k

= kǫ by the integer k = 0, . . . , N. Notice that

t

0

= 0 and t

N

= t.

0

ǫ 2ǫ (N −1)ǫ

t = Nǫ

Figure 1.7: The interval between 0 and t is divided into N steps.

We can use the completeness relation

_

[x) ¸x[ dx = 1 at each step nǫ:

¸x

f

[ e

−

iHt

ℏ

[x

i

) =

_

N−1

k=1

dx

k

¸x

f

[ e

−

iHǫ

ℏ

[x

N−1

) ¸x

N−1

[ e

−

iHǫ

ℏ

[x

N−2

)

¸x

2

[ e

−

iHǫ

ℏ

[x

1

) ¸x

1

[ e

−

iHǫ

ℏ

[x

i

) (1.9)

Consider the quantity: ¸x

′

[ e

−iHǫ/ℏ

[x). Using

_

[p) ¸p[ dp = 1:

¸x

′

[ e

−

iHǫ

ℏ

[x) =

_

dp ¸x

′

[p) ¸p[ e

−

iHǫ

ℏ

[x) (1.10)

If we stick to the simple case H =

ˆ p

2

2m

+V (ˆ x), we can write:

¸p[ e

−i

ǫ

ℏ

h

ˆ p

2

2m

+V (ˆ x)

i

[x) = e

−i

ǫ

ℏ

h

p

2

2m

+V (x)

i

¸p[x) +O

_

ǫ

2

_

(1.11)

where the O(ǫ

2

) terms arise from the non-vanishing commutator between ˆ p

2

/(2m)

and V (ˆ x). We will now assume, which seems fully reasonable, that in the limit

ǫ → 0 these higher orer terms can be neglected. Later on we shall come back

on this issue and better motivate our neglect of these terms.

We then get:

¸x

′

[ e

−

iHǫ

ℏ

[x) ≃

_

dp

_

e

−i

ǫ

ℏ

h

p

2

2m

+V (x)

i

e

i

ℏ

p(x

′

−x)

2πℏ

_

(1.12)

We can deﬁne:

x

′

−x

ǫ

≡ ˙ x:

¸x

′

[ e

−

iHǫ

ℏ

[x) ≃

_

dp

2πℏ

e

−i

ǫ

ℏ

h

p

2

2m

+V (x)−p˙ x

i

(1.13)

1.1. INTRODUCING THE PATH INTEGRALS 13

By performing the change of variables ¯ p ≡ p − m˙ x the integral reduces to

simple gaussian integral for the variable ¯ p:

¸x

′

[ e

−

iHǫ

ℏ

[x) ≃

_

d¯ p

2πℏ

e

−i

ǫ

ℏ

h

¯ p

2

2m

+V (x)−

m˙ x

2

2

i

(1.14)

=

_

m

2πiℏǫ

. ¸¸ .

1

A

e

i

ǫ

ℏ

[

1

2

m˙ x

2

−V (x)]

=

1

A

e

iL(x, ˙ x)

ǫ

ℏ

(1.15)

At leading order in ǫ we can further identify L(x, ˙ x)ǫ with the action S(x

′

, x) =

_

ǫ

0

L(x, ˙ x)dt. By considering all the intervals we thus ﬁnally get:

¸x

f

[ e

−

iHt

ℏ

[x

i

) = lim

ǫ→0

_

N−1

k=1

dx

k

1

A

N

e

i

ℏ

P

N−1

l=0

S(x

l+1

,x

l

)

= lim

ǫ→0

1

A

_

N−1

k=1

dx

k

A

. ¸¸ .

R

Dγ

e

i

S(x

f

,x

i

)

ℏ

(1.16)

where

_

Dγ should be taken as a deﬁnition of the functional measure over the

space of the trajectories. We thus have got a path integral formula for the

transition amplitude:

K(x

f

, t; x

i

, 0) =

_

D[x(t)]e

i

S

cl

(x, ˙ x)

ℏ

(1.17)

We see that we (actually somebody else before us!) had guessed well the form

of the transition amplitude. The path integral approach to QM was developed

by Richard Feyman in his PhD Thesis in the mid 40’s, following a hint from

an earlier paper by Dirac. Dirac’s motivation was apparently to formulate QM

starting from the lagrangian rather than from the hamiltonian formulation of

classical mechanics.

Let us now come back to the neglected terms in eq. (1.11). To simplify the

notation let us denote the kinetic operator as T = −i

p

2

2mℏ

and the potential

U = −iV/ℏ; we can then write

¸p[ e

ǫ(T+U)

[x) = ¸p[ e

ǫT

e

−ǫT

e

ǫ(T+U)

e

−ǫU

e

ǫU

[x) = ¸p[ e

ǫT

e

−ǫ

2

C

e

ǫU

[x) (1.18)

= e

−i

ǫ

ℏ

h

p

2

2m

+V (x)

i

¸p[ e

−ǫ

2

C

[x) (1.19)

where C is given, by using the Campbell-Baker-Haussdorf formula twice, as a

series of commutators between T and U

C =

1

2

[T, U] +

ǫ

6

¦[T, [T, U]] + [U, [U, T]]¦ +. . . (1.20)

By iterating the basic commutation relation [ ˆ p, V (ˆ x)]] = −iV

′

(ˆ x) and expanding

the exponent in a series one can then write

¸p[ e

−ǫ

2

C

[x) = 1 +ǫ

n=∞

n=0

r=n

r=1

s=r

s=0

ǫ

n

p

r−s

P

n,s,r

(x) (1.21)

14 CHAPTER 1. THE PATH INTEGRAL FORMALISM

where P

n,s,r

(x) is a homogenous polynomial of degree n + 1 − r in V and its

derivatives, with each term involving exactly r + s derivatives. For instance

P

1,0,1

(x) = V

′

(x). We can now test our result under some simple assumption.

For instance, if the derivatives of V are all bounded, the only potential problem

to concentrate on in the ǫ → 0 limit is represented by the powers of p. This

is because the leading contribution to the p integral in eq. (1.15) comes from

the region p ∼ 1/

√

ǫ, showing that p diverges in the small ǫ limit. By using

p ∼ 1/

√

ǫ the right hand side of eq. (1.21) is ∼ 1 +O(ǫ

3/2

) so that, even taking

into account that there are N ∼ 1/ǫ such terms (one for each step), the ﬁnal

result is still convergent to 1

lim

ǫ→0

_

1 +aǫ

3

2

_1

ǫ

= 1 . (1.22)

Before proceeding with technical developments, it is worth assessing the

rˆole of the path integral (P.I.) in quantum mechanics. As it was hopefully

highlighted in the discussion above, the path integral formulation is conceptually

advantageous over the standard operatorial formulation of Quantum Mechanics,

in that the “good old” particle trajectories retain some rˆole. The P.I. is however

technically more involved. When working on simple quantum systems like the

hydrogen atom, no technical proﬁt is really given by path integrals. Nonetheless,

after overcoming a few technical diﬃculties, the path integral oﬀers a much

more direct viewpoint on the semiclassical limit. Similarly, for issues involving

topology like the origin of Bose and Fermi statistics, the Aharonov-Bohm eﬀect,

charge quantization in the presence of a magnetic monopole, etc. . . path integrals

oﬀer a much better viewpoint. Finally, for advanced issues like the quantization

of gauge theories and for eﬀects like instantons in quantum ﬁeld theory it would

be hard to think how to proceed without path integrals!...But that is for another

course.

1.2 The properties of the path integrals

1.2.1 Path integrals and state evolution

To get an estimate of the dependence of the amplitude K(x

f

, t

f

; x

i

, t

i

) on its

arguments, let us ﬁrst look at the properties of the solutions to the classical

equations of motion with boundary conditions x

c

(t

i

) = x

i

, x

c

(t

f

) = x

f

.

Let us compute ∂

t

f

S

cl

. Where S

cl

is deﬁned

S

cl

≡ S[x

c

] =

_

t

f

ti

L(x

c

, ˙ x

c

)dt (1.23)

with x

c

a solution, satisfying the Euler-Lagrange equation:

_

∂

t

∂L

∂ ˙ x

−

∂L

∂x

_

x=xc

= 0 . (1.24)

We can think of x

c

as a function

x

c

≡ f(x

i

, x

f

, t

i

, t

f

, t)

1.2. THE PROPERTIES OF THE PATH INTEGRALS 15

such that

f(x

i

, x

f

, t

i

, t

f

, t = t

i

) ≡ x

i

= const

f(x

i

, x

f

, t

i

, t

f

, t = t

f

) ≡ x

f

= const

Diﬀerentiating the last relation we deduce:

_

∂

t

f

x

c

+∂

t

x

c

¸

t=t

f

= 0

=⇒ ∂

t

f

x

c

¸

¸

t=t

f

= −˙ x

c

¸

¸

t=t

f

(1.25)

Similarly, diﬀerentiating the relation at t = t

i

we obtain

∂

t

f

x

c

¸

¸

t=ti

= 0 (1.26)

And the following properties are straightforward:

∂

x

f

x

c

¸

¸

t=t

f

= 1 (1.27)

∂

x

f

x

c

¸

¸

t=ti

= 0 (1.28)

Using these properties, we can now compute:

∂

t

f

S = L(x, ˙ x)

¸

¸

t=t

f

+

_

t

f

ti

dt

_

∂L

∂x

∂

t

f

x +

∂L

∂ ˙ x

∂

t

f

˙ x

_

= L(x, ˙ x)

¸

¸

t=t

f

+

_

t

f

ti

dt

_

∂

t

_

∂L

∂ ˙ x

∂

t

f

x

__

−

_

t

f

ti

dt

_

∂

t

∂L

∂ ˙ x

−

∂L

∂x

_

∂

t

f

x

(1.29)

All this evaluated at x = x

c

gives:

∂

t

f

S

cl

= L(x

c

, ˙ x

c

)

¸

¸

t=t

f

+

_

∂L

∂ ˙ x

¸

¸

¸

¸

x=xc

∂

t

f

x

c

_¸

¸

¸

¸

¸

t=t

f

t=ti

=

_

L(x

c

, ˙ x

c

) −

∂L

∂ ˙ x

¸

¸

¸

¸

x=xc

˙ x

c

_¸

¸

¸

¸

¸

t=t

f

= −E (1.30)

We recognize here the deﬁnition of the hamiltonian of the system, evaluated

at x = x

c

and t = t

f

.

We can also compute:

∂

x

f

S =

_

t

f

ti

dt

_

∂

t

_

∂L

∂ ˙ x

∂

x

f

x

__

−

_

t

f

ti

dt

_

∂

t

∂L

∂ ˙ x

−

∂L

∂x

_

∂

x

f

x (1.31)

Evaluated at x = x

c

, this gives:

∂

x

f

S

cl

(x

c

) =

_

∂L

∂ ˙ x

¸

¸

¸

¸

x=xc

∂

x

f

x

c

_¸

¸

¸

¸

¸

t=t

f

t=ti

=

∂L

∂ ˙ x

¸

¸

¸

¸

x=xc,t=t

f

= P (1.32)

We recognize here the deﬁnition of the momentum of the system, evaluated

at x = x

c

and t = t

f

.

16 CHAPTER 1. THE PATH INTEGRAL FORMALISM

At this point, we can compare these results with those of the path integral

computation in the limit ℏ →0.

K(x

f

, t

f

; x

i

, t

i

) =

_

D[x(t)]e

i

S

ℏ

(1.33)

In the limit ℏ we expect the path integral to be dominated by the classical

trajectory and thus to take the form (and we shall technically demostrate this

expectation later on)

K(x

f

, t

f

; x

i

, t

i

) = F(x

f

, t

f

, x

i

, t

i

)

. ¸¸ .

smooth function

e

i

S

cl

ℏ

(1.34)

where F is a smooth function in the ℏ →0 limit. We then have:

∂

t,x

e

i

S

cl

ℏ

∼

1

ℏ

large (1.35)

∂

t,x

F ∼ O(1) comparatively negligible (1.36)

so that we obtain

−iℏ∂

x

f

K = ∂

x

f

S

cl

K +O(ℏ) = P

cl

K +O(ℏ) (1.37)

iℏ∂

t

f

K = −∂

t

f

S

cl

K +O(ℏ) = E

cl

K +O(ℏ) (1.38)

Momentum and energy, when going to the path integral description of particle

propagation, become associated to respectively the wave number (∼ (i∂

x

ln K)

−1

)

and oscillation frequency (∼ (i∂

t

ln K)).

The last two equations make contact with usual quantum mechanical rela-

tions, if we interpret K has the wave function and −iℏ∂

x

and iℏ∂

t

as respectively

the momentum and energy operators

−iℏ∂

x

K ∼ P K (1.39)

iℏ∂

t

K ∼ E K (1.40)

The function K(x

′

, t

′

; x, t) = ¸x

′

[ e

−

iH(t

′

−t)

ℏ

[x) is also known as the propaga-

tor, as it describes the probability amplitude for propagating a particle from x

to x

′

in a time t

′

−t.

From the point of view of the Schr¨odinger picture, K(x

′

, t; x, 0) represents

the wave function at time t for a particle that was in the state [x) at time 0:

ψ(t, x) ≡ ¸x[ e

−

iHt

ℏ

[ψ

0

) ≡ ¸x[ψ(t)) (1.41)

By the superposition principle, K can then be used to write the most general

solution. Indeed:

1. K(x

′

, t

′

; x, t) solves the Schr¨odinger equation for t

′

, x

′

as variables:

iℏ∂

t

′ K =

_

dx

′′

H(x

′

, x

′′

)K(x

′′

, t

′

; x, t) (1.42)

where H(x

′

, x

′′

) = ¸x

′

[ H [x

′′

).

1.2. THE PROPERTIES OF THE PATH INTEGRALS 17

2.

lim

t

′

→t

K(x

′

, t

′

; x, t) = ¸x

′

[x) = δ(x

′

−x)

Thus, given Ψ

0

(x) the wave function at time t

0

, the solution at t > t

0

is:

Ψ(x, t) = ¸x[ e

−

iH(t−t

0

)

ℏ

[Ψ

0

)

=

_

¸x[ e

−

iH(t−t

0

)

ℏ

[y) ¸y[Ψ

0

) dy

=

_

K(x, t; y, t

0

)Ψ

0

(y)dy (1.43)

• Having K(x

′

, t

′

; x, t), the solution to the Schr¨odinger equation is found by

performing an integral.

• K(x

′

, t

′

; x, t) contains all the relevant information on the dynamics of the

system.

1.2.2 The path integral computation for a free particle

Let us compute the propagator of a free particle, decribed by the lagrangian

L(x, ˙ x) = m˙ x

2

/2, using Feyman’s time slicing procedure. Using the result of

section 1.1.4 we can write

K(x

f

, t

f

; x

i

, t

i

) = lim

ǫ→0

1

A

_

N−1

k=1

dx

k

A

e

iS

ℏ

(1.44)

where:

ǫ =

t

f

−t

i

N

S =

N−1

k=0

1

2

m

(x

k+1

−x

k

)

2

ǫ

x

0

= x

i

x

N

= x

f

Let us compute eq. (1.44) by integrating ﬁrst over x

1

. This coordinate only

appears only in the ﬁrst and the second time slice so we need just to concentrate

on the integral

_

∞

−∞

dx

1

e

i

m

2ǫℏ

[(x1−x0)

2

+(x2−x1)

2

]

=

_

∞

−∞

dx

1

e

i

m

2ǫℏ

h

2(x1−

1

2

(x0+x2))

2

+

1

2

(x2−x0)

2

i

=

_

2πiǫℏ

m

1

2

. ¸¸ .

A

√

1

2

e

i

m

2ǫℏ

1

2

(x2−x0)

2

(1.45)

Notice that x

1

has disappeared: we “integrated it out” in the physics jargon.

Putting this result back into eq. (1.44) we notice that we have an expression

18 CHAPTER 1. THE PATH INTEGRAL FORMALISM

similar to the original one but with N − 1 instead of N time slices. Moreover

the ﬁrst slice has now a width 2ǫ instead of ǫ: the factors of 1/2 in both the

exponent and prefactor of eq. (1.45) are checked to match this interpratation.

It is now easy to go on. Let us do the integral on x

2

to gain more conﬁdence.

The relevant terms are

_

∞

−∞

dx

2

e

i

m

2ǫℏ

[

1

2

(x2−x0)

2

+(x3−x2)

2

]

=

_

∞

−∞

dx

2

e

i

m

2ǫℏ

h

3

2

(x2−

1

3

x0−

2

3

x3)

2

+

1

3

(x3−x0)

2

i

=

_

2πiǫℏ

m

2

3

. ¸¸ .

A

√

2

3

e

i

m

2ǫℏ

1

3

(x3−x0)

2

(1.46)

It is now clear how to proceed by induction. At the n-th step the integral

over x

n

will be:

_

dx

n

e

i

m

2ǫℏ

[

1

n

(xn−x0)

2

+(xn+1−xn)

2

]

=

_

dx

n

e

i

m

2ǫℏ

h

n+1

n

(xn−(

1

n+1

x0+

n

n+1

xn+1))

2

+

1

n+1

(xn+1−x0)

2

i

=

_

2πiǫℏ

m

n

n + 1

. ¸¸ .

A

√

n

n+1

e

i

m

2ǫℏ

1

n+1

(xn+1−x0)

2

(1.47)

Thus, putting all together, we ﬁnd the expression of the propagator for a

free particle:

K(x

f

, t

f

; x

i

, t

i

) = lim

ǫ→0

1

A

_

1

2

2

3

. . .

N −1

N

e

i

m

2ℏNǫ

(x

f

−xi)

2

= lim

ǫ→0

1

A

√

N

e

i

m

2ℏNǫ

(x

f

−xi)

2

=

_

m

2πiℏ(t

f

−t

i

)

e

i

m

2ℏ

(x

f

−x

i

)

2

t

f

−t

i

(1.48)

where we used the fact that Nǫ = t

f

−t

i

.

We can check this result by computing directly the propagator of a one-

dimensional free particle using the standard Hilbert space representation of the

propagator:

K(x

f

, t; x

i

, 0) =

_

¸x

f

[p) ¸p[ e

−

iHt

ℏ

[x

i

) dp

=

1

2πℏ

_

e

i

p

ℏ

(x

f

−xi)

e

i

p

2

t

2mℏ

dp

=

1

2πℏ

_

e

−

it

2mℏ

“

p−m

x

f

−x

i

t

”

2

+i

m

2ℏ

(x

f

−x

i

)

2

t

dp

=

_

m

2πiℏt

e

i

m(x

f

−x

i

)

2

2ℏt

= F(t)e

i

S

cl

(x

f

,t;x

i

,0)

ℏ

(1.49)

1.3. PATH INTEGRALS AS DETERMINANTS 19

These results trivially generalize to the motion in higher dimensional spaces, as

for the free particle the result factorizes.

• As we already saw, we can perform the same integral “` a la Feynman” by

dividing t into N inﬁnitesimal steps.

• It is instructive interpret the result for the probability distribution:

dP

dx

= [K[

2

=

m

2πℏt

(1.50)

Let us interpret this result in terms of a ﬂux of particles that started at

x

i

= 0 at t = 0 with a distribution of momenta

dn(p) = f(p)dp (1.51)

and ﬁnd what f(p) is. A particle with momentum p at time t will have

travelled to x = (p/m)t. Thus the particles in the interval dp will be

at time t in the coordinate interval dx = (t/m)dp. Therefore we have

dn(x) = f(xm/t)(m/t)dx. Comparing to (1.50), we ﬁnd f = 1/(2πℏ) =

const. Thus, dn(p) = (1/2πℏ)dp ⇒dn/dp ∼ [Length]

−1

[Momentum]

−1

.

We recover the result that the wave function ψ(x, t) = K(x, t; 0, 0) satisﬁes

ψ(x, 0) = δ(x) which corresponds in momentum space to an exactly ﬂat

distribution of momenta, with normalization dn(p)/dp = (1/2πℏ).

1.3 Path integrals as determinants

We will now introduce a systematic small ℏ expansion for path integrals and

show that the computation of the “leading” term in the expansion corresponds

to the computation of the determinant of an operator. Before going to that,

let us remind ourselves some basic results about “Gaussian” integrals in several

variables.

1.3.1 Gaussian integrals

Let us give a look at gaussian integrals:

_

i

dx

i

e

−

P

i,j

λij xixj

(1.52)

• With one variable, and with λ real and positive we have:

_

dxe

−λx

2

=

_

1

λ

_

dye

−y

2

=

_

π

λ

(1.53)

• Consider now a Gaussian integral with an arbitrary number of real vari-

ables

_

i

dx

i

e

−

P

i,j

λijxixj

(1.54)

where λ

ij

is real.

20 CHAPTER 1. THE PATH INTEGRAL FORMALISM

We can rotate to the eigenbasis: x

i

= O

ij

˜ x

j

, with O

T

λO = diag (λ

1

, . . . , λ

n

),

O

T

O = OO

T

= 1. Then,

i

dx

i

= [det O[

i

d˜ x

i

=

i

d˜ x

i

. Thus:

_

i

dx

i

e

−

P

i,j

λijxixj

=

_

i

d˜ x

i

e

−

P

i

λi ˜ x

2

i

=

i

_

π

λ

i

=

¸

1

det(

1

π

ˆ

λ)

(1.55)

where again the result makes sense only when all eigenvalues λ

n

are posi-

tive.

• Along the same line we can consider integrals with imaginary exponents,

which we indeed already encountered before

I =

_

dxe

iλx

2

= e

i sgn(λ)

π

4

_

π

[λ[

(1.56)

This integral is performed by deforming the contour into the imaginary

plane as discussed in homework 1. Depending on the sign of λ we must

choose diﬀerent contours to ensure convergence at inﬁnity. Notice that

the phase depends on the sign of λ.

• For the case of an integral with several variables

I =

_

i

dx

i

e

i

P

j,k

λ

jk

xjx

k

(1.57)

the result is generalized to

I = e

i(n+−n−)

π

4

¸

1

[ det(

1

π

ˆ

λ)[

(1.58)

where n

+

and n

−

are respectively the number of positive and negative

eigenvalues of the matrix λ

jk

.

1.3.2 Gaussian Path Integrals

Let us now make contact with the path integral:

K(x

f

, t

f

; x

i

, t

i

) =

_

D[x(t)]e

i

S[x(t)]

ℏ

(1.59)

with

S[x(t)] =

_

t

f

ti

L(x(t), ˙ x(t))dt (1.60)

• To perform the integral, we can choose the integration variables that suit

us best.

• The simplest change of variables is to shift x(t) to “center” it around the

classical solution x

c

(t):

x(t) = x

c

(t) +y(t) ⇒D[x(t)] = D[y(t)]

(the Jacobian for this change of variables is obviously trivial, as we indi-

cated).

1.3. PATH INTEGRALS AS DETERMINANTS 21

• We can Taylor expand S[x

c

(t) + y(t)] in y(t) (note that y(t

i

) = y(t

f

) =

0, since the classical solution already satisﬁes the boundary conditions

x

c

(t

i

) = x

i

and x

c

(t

f

) = x

f

). By the usual deﬁnition of functional deriva-

tive, the Taylor expansion of a functional reads in general

S[x

c

(t) +y(t)] = S[x

c

] +

_

dt

1

δS

δx(t

1

)

¸

¸

¸

¸

x=xc

y(t

1

) (1.61)

+

1

2

_

dt

1

dt

2

δ

2

S

δx(t

1

)δx(t

2

)

¸

¸

¸

¸

x=xc

y(t

1

)y(t

2

) +O

_

y

3

_

with obvious generalization to all orders in y. In our case S =

_

dtL(x, ˙ x)

so that each term in the above general expansion can be written as a single

dt integral

_

dt

1

δS

δx(t

1

)

¸

¸

¸

¸

x=xc

y(t

1

) ≡

_

dt

_

∂L

∂x

¸

¸

¸

¸

x=xc

y +

∂L

∂ ˙ x

¸

¸

¸

¸

x=xc

˙ y

_

(1.62)

1

2

_

dt

1

dt

2

δ

2

S

δx(t

1

)δx(t

2

)

¸

¸

¸

¸

x=xc

y(t

1

)y(t

2

) ≡

_

dt

_

∂

2

L

∂x

2

¸

¸

¸

¸

x=xc

y

2

+ 2

∂

2

L

∂x∂ ˙ x

¸

¸

¸

¸

x=xc

y ˙ y +

∂

2

L

∂ ˙ x

2

¸

¸

¸

¸

x=xc

˙ y

2

_

(1.63)

and so on.

The linear term in the expansion vanishes by the equations of motion:

_

δS

δx

¸

¸

¸

¸

x=xc

y(t)dt =

_ _

∂L

∂x

−

d

dt

∂L

∂ ˙ x

_¸

¸

¸

¸

x=xc

y(t)dt +

∂L

∂ ˙ x

y(t)

¸

¸

¸

¸

t=t

f

t=ti

= 0

(1.64)

Thus, we obtain the result:

K(x

f

, t

f

; x

i

, t

i

) =

_

D[y(t)]e

i

ℏ

h

S[xc(t)]+

1

2

δ

2

S

δx

2

y

2

+O(y

3

)

i

(1.65)

where δ

2

S/δx

2

is just short hand for the second variation of S shown in eq. (1.63).

We can rewrite it rescaling our variables y =

√

ℏ ˜ y:

K(x

f

, t

f

; x

i

, t

i

) = ^ e

i

ℏ

S[xc(t)]

_

D[˜ y(t)]e

i

2

δ

2

S

δx

2

˜ y

2

+O(

√

ℏ ˜ y

3

)

(1.66)

where the overall constant ^ is just the Jacobian of the rescaling.

2

The

interpretation of the above expression is that the quantum propagation of a

particle can be decomposed into two pieces:

1. the classical trajectory x

c

(t), which gives rise to the exponent e

iS[xc]/ℏ

.

2

Given that N is a constant its role is only to ﬁx the right overall normalization of the prop-

agator, and does not matter in assessing the relative importance of each trajectory. Therefore

it does not play a crucial role in the following discussion.

22 CHAPTER 1. THE PATH INTEGRAL FORMALISM

2. the ﬂuctuation y(t) over which we must integrate; because of the quadratic

term in the exponent, the path integral is dominated by y ∼ O

_

√

ℏ

_

(that

is ˜ y ∼ O(1)) so that y(t) represents the quantum ﬂuctuation around the

classical trajectory.

In those physical situations where ℏ can be treated as a small quantity, one

can treat O(ℏ) in the exponent of the path integral as a perturbation

K(x

f

, t

f

; x

i

, t

i

) = ^ e

i

ℏ

S[xc(t)]

_

D[˜ y(t)]e

i

2

δ

2

S

δx

2

˜ y

2

[1 +O(ℏ)] = (1.67)

= F(x

f

, t

f

; x

i

, t

i

)e

i

ℏ

S[xc(t)]

[1 +O(ℏ)] (1.68)

Where the prefactor F is just the gaussian integral around the classical trajec-

tory. The semiclassical limit should thus correspond to the possibility to reduce

the path integral to a gaussian integral. We will study this in more detail later

on.

In the small ℏ limit, the “classical” part of the propagator, exp(iS[x

c

]/ℏ)

oscillates rapidly when the classical action changes, while the prefactor (F and

the terms in square brakets in eq. (1.68)) depend smoothly on ℏ in this limit.

As an example, we can compute the value of S

cl

for a simple macroscopic

system: a ball weighing one gram moves freely along one meter in one second.

We will use the following unity for energy: [erg] = [g] ([cm]/[s])

2

.

S

cl

=

1

2

m

(x

f

−x

i

)

2

t

f

− t

i

=

1

2

mv∆x = 5000 erg s

ℏ = 1.0546 10

−27

erg s

⇒S

cl

/ℏ ∼ 10

30

1.3.3 O(ℏ) corrections to Gaussian approximation

It is a good exercise to write the expression for the O(ℏ) correction to the

propagator K in eq. (1.68). In order to do so we must expand the action to

order ˜ y

4

around the classical solution (using the same short hand notation as

before for the higher order variation of S)

S[x

c

+

√

ℏ˜ y]

ℏ

=

S[x

c

]

ℏ

+

1

2

δ

2

S

δx

2

˜ y

2

+

√

ℏ

3!

δ

3

S

δx

3

˜ y

3

+

ℏ

4!

δ

4

S

δx

4

˜ y

4

+O(ℏ

3/2

) (1.69)

and then expand the exponent in the path integral to order ℏ. The leading ℏ

1/2

correction vanishes because the integrand is odd under ˜ y →−˜ y

_

D[˜ y(t)]

√

ℏ

3!

δ

3

S

δx

3

˜ y

3

e

i

2

δ

2

S

δx

2

˜ y

2

= 0 (1.70)

and for the O(ℏ) correction to K we ﬁnd that two terms contribute

∆K = iℏe

i

ℏ

S[xc(t)]

^

_

D[˜ y(t)]

_

1

4!

δ

4

S

δx

4

˜ y

4

+i

1

2

_

1

3!

δ

3

S

δx

3

˜ y

3

_

2

_

e

i

2

δ

2

S

δx

2

˜ y

2

(1.71)

1.3. PATH INTEGRALS AS DETERMINANTS 23

1.3.4 Quadratic lagrangians and the harmonic oscillator

Consider now a special case: the quadratic lagrangians, deﬁned by the property:

δ

(n)

S

δx

n

= 0 for n > 2 (1.72)

For these lagrangians, the gaussian integral coresponds to the exact result:

K(x

f

, t

f

; x

i

, t

i

) = e

i

S[xc]

ℏ

_

D[y]e

i

ℏ

1

2

δ

2

S

δx

2

y

2

= const e

i

S[xc]

ℏ

_

det

δ

2

S

δx

2

_

−

1

2

(1.73)

To make sense of the above we must deﬁne this strange “beast”: the de-

terminant of an operator. It is best illustrated with explicit examples, and

usually computed indiretly using some “trick”. This is a curious thing about

path integrals: one never really ends up by computing them directly.

Let us compute the propagator for the harmonic oscillator using the deter-

minant method. We have the lagrangian:

L(x, ˙ x) =

1

2

m˙ x

2

−

1

2

mω

2

x

2

(1.74)

It is left as an exercise to show that:

S

cl

=

_

t

f

ti

L(x

c

, ˙ x

c

)dt =

mω

2

_

_

x

2

f

+x

2

i

_

cot(ωT) −

2x

f

x

i

sin(ωT)

_

(1.75)

where T = t

f

−t

i

.

Then,

_

t

f

ti

L(x

c

+y, ˙ x

c

+ ˙ y)dt =

_

t

f

ti

1

2

m

_

˙ x

2

c

−ω

2

x

2

c

_

dt +

_

t

f

ti

1

2

m

_

˙ y

2

−ω

2

y

2

_

dt

(1.76)

And thus, the propagator is:

K(x

f

, t

f

; x

i

, t

i

) = e

i

S

cl

ℏ

_

D[y]e

i

ℏ

R t

f

t

i

m

2

( ˙ y

2

−ω

2

y

2

)dt

= e

i

S

cl

(x

f

,x

i

,t

f

−t

i

)

ℏ

J(t

f

−t

i

) (1.77)

This is because y(t) satisﬁes y(t

i

) = y(t

f

) = 0 and any reference to x

f

and

x

i

has disappeared from the integral over y. Indeed J is just the propagator

from x

i

= 0 to x

f

= 0:

K(x

f

, t

f

; x

i

, t

i

) = e

i

S

cl

(x

f

,x

i

,t

f

−t

i

)

ℏ

K(0, t

f

; 0, t

i

)

. ¸¸ .

J(t

f

−ti)

(1.78)

We already have got a non trivial information by simple manipulations. The

question now is how to compute J(t

f

−t

i

).

24 CHAPTER 1. THE PATH INTEGRAL FORMALISM

1. Trick number 1: use the composition property of the transition amplitude:

K(x

f

, t

f

; x

i

, t

i

) =

_

dxK(x

f

, t

f

; x, t)K(x, t; x

i

, t

i

)

Applying this to J(t

f

−t

i

), we get:

J(t

f

−t

i

) = K(0, t

f

; 0, t

i

) =

J(t

f

−t)J(t −t

i

)

_

e

i

ℏ

(S

cl

(0,x,t

f

−t)+S

cl

(x,0,t−ti))

dx (1.79)

We can put the expression of the classical action in the integral, and we

get (T = t

f

−t

i

, T

1

= t

f

−t, T

2

= t −t

i

):

J(T)

J(T

1

)J(T

2

)

=

_

dxe

i

mω

2ℏ

[x

2

(cot(ωT1)+cot(ωT2))]

=

_

dxe

iλx

2

=

_

iπ

λ

=

_

2iπℏ

mω

sin(ωT

1

) sin(ωT

2

)

sin(ωT)

_1

2

(1.80)

The general solution to the above equation is

J(T) =

_

mω

2πiℏ sin(ωT)

e

aT

(1.81)

with a an arbitrary constant. So this trick is not enough to fully ﬁx the

propagator, but it already tells us a good deal about it. We will now

compute J directly and ﬁnd a = 0. Notice that in the limit ω → 0, the

result goes back to that for a free particle.

2. Now, let us compute the same quantity using the determinant method.

What we want to compute is:

J(T) =

_

D[y]e

i

ℏ

R t

f

t

i

m

2

( ˙ y

2

−ω

2

y

2

)

(1.82)

with the boundary conditions:

y(t

i

) = y(t

f

) = 0 (1.83)

We can note the property:

i

_

t

f

ti

m

2ℏ

_

˙ y

2

−ω

2

y

2

_

dt = −i

_

t

f

ti

m

2ℏ

y

_

d

2

dt

2

+ω

2

_

y dt =

i

2

_

t

f

ti

y

ˆ

Oy dt

(1.84)

where

ˆ

O = −

m

ℏ

_

d

2

dt

2

+ω

2

_

.

Formally, we have to perform a gaussian integral, and thus the ﬁnal result

is proportional to

_

det

ˆ

O

_

−1/2

.

To make it more explicit, let us work in Fourier space for y:

y(t) =

n

a

n

y

n

(t) (1.85)

1.3. PATH INTEGRALS AS DETERMINANTS 25

where y

n

(t) are the othonormal eigenfunctions of

ˆ

O:

y

n

=

_

2

T

sin(

nπ

T

t) , n ∈ N

∗

(1.86)

ˆ

Oy

n

(t) = λ

n

y

n

(t) (1.87)

_

y

n

y

m

dt = δ

nm

(1.88)

The eigenvalues are

λ

n

=

m

ℏ

_

_

nπ

T

_

2

−ω

2

_

(1.89)

Notice that the number of negative eigenvalues is ﬁnite and given by n

−

=

int(

Tω

π

), that is the number of half periods of oscillation contained in T.

The y

n

form a complete basis of the Hilbert space L

2

([t

i

, t

f

]) mod y(t

i

) =

y(t

f

) = 0. Thus, the a

n

form a discrete set of integration variables, and

we can write:

D[y(t)] =

n

da

n

√

2πi

˜

N

.¸¸.

“Jacobian”

(1.90)

where the 1/

√

2πi factors are singled out for later convenience (and also

to mimick the 1/A factors in our deﬁnition of the measure in eq. (1.16)).

We get:

_

y

ˆ

Oy dt =

m,n

_

a

m

a

n

y

m

ˆ

Oy

n

dt =

n

λ

n

a

2

n

(1.91)

And thus,

J(T) =

˜

N

_

n

da

n

√

2πi

e

i

2

P

n

λna

2

n

=

ill deﬁned

¸ .. ¸

_

[

n

λ

n

[

_

−

1

2

˜

N

. ¸¸ .

well deﬁned

e

i(n+−n−)

π

4

e

i(n++n−)

π

4

=

_

[

n

λ

n

[

_

−

1

2

˜

N e

−in−

π

2

(1.92)

Notice that in the continuum limit ǫ →0 in the deﬁnition of the measure

eq. (1.16), we would have to consider all the Fourier modes with arbitrarily

large n. In this limit both the product of eigenvalues and the Jacobian

˜

N are ill deﬁned. Their product however, the only physically relevant

quantity, is well deﬁned. We should not get into any trouble if we avoid

computing these two quantities separately. The strategy is to work only

with ratios that are well deﬁned. That way we shall keep at large from

mathematical diﬃculties (and confusion!), as we will now show.

Notice that

ˆ

O depends on ω, while

˜

N obviously does not: the modes y

n

do not depend on ω

3

. For ω = 0, our computation must give the free

3

Indeed a stronger result for the independence of the Jacobian on the quadratic action

holds, as we shall discuss in section 3.1.

26 CHAPTER 1. THE PATH INTEGRAL FORMALISM

particle result, which we already know. So we have

J

ω

(T) =

˜

N

_

[ det

ˆ

O

ω

[

_

−

1

2

e

−in−

π

2

(1.93)

J

0

(T) =

˜

N

_

[ det

ˆ

O

0

[

_

−

1

2

(1.94)

We can thus consider

J

ω

(T)

J

0

(T)

e

in−

π

2

=

_

[

det

ˆ

O

0

det

ˆ

O

ω

[

_1

2

=

_

n

[λ

n

(ω = 0)[

n

[λ

n

(ω)

[

_1

2

=

_

[

n

λ

n

(0)

λ

n

(ω)

[

_1

2

(1.95)

The eigenvalues of the operator

ˆ

O

ω

= −

m

ℏ

_

d

2

dt

2

+ω

2

_

over the space of

functions y(t) have been given before:

λ

n

(0)

λ

n

(ω)

=

_

nπ

T

_

2

_

nπ

T

_

2

−ω

2

=

1

1 −

_

ωT

nπ

_

2

=

1

1 −

_

a

nπ

_

2

(1.96)

∞

n=1

λ

n

(0)

λ

n

(ω)

=

∞

n=1

1

1 −

_

a

nπ

_

2

=

a

sin a

(1.97)

By eq. (1.95) we get our ﬁnal result:

J

ω

(T) = J

0

(T)

¸

ωT

[ sin(ωT)[

e

−in−

π

2

=

_

mω

2πiℏ[ sin(ωT)[

e

−in−

π

2

(1.98)

So that by using eq. (1.75) the propagator is (assume for simplicity ωT < π)

K(x

f

, T; x

i

, 0) =

_

mω

2πiℏ sin(ωT)

e

i

mω

2ℏ

h

(x

2

f

+x

2

i

) cot(ωT)−

2x

f

x

i

sin(ωT)

i

(1.99)

At this point, we can make contact with the solution of the eigenvalue prob-

lem and recover the well know result for the energy levels of the harmonic

oscillator. Consider ¦Ψ

n

¦ a complete orthonormal basis of eigenvectors of the

hamiltonian such that H [Ψ

n

) = E

n

[Ψ

n

). The propagator can be written as:

K(x

f

, t

f

; x

i

, t

i

) =

m,n

¸x

f

[Ψ

n

) ¸Ψ

n

[ e

−

iH(t

f

−t

i

)

ℏ

[Ψ

m

) ¸Ψ

m

[x

i

) =

n

Ψ

∗

n

(x

i

)Ψ

n

(x

f

)e

−

iEn(t

f

−t

i

)

ℏ

(1.100)

Let us now deﬁne the partition function:

Z(t) =

_

K(x, T; x, 0)dx =

n

__

[Ψ

n

(x)[

2

dx

_

e

−

iEnT

ℏ

=

n

e

−

iEnT

ℏ

(1.101)

For the harmonic oscillator, using eq. (1.100) we ﬁnd:

_

K(x, T; x, 0)dx =

1

2i sin

_

ωT

2

_ = e

−i

ωT

2

1

1 −e

−2i

ωT

2

=

∞

n=0

e

−i(2n+1)

ωT

2

(1.102)

1.4. OPERATOR MATRIX ELEMENTS 27

⇒E

n

= ℏω

_

n +

1

2

_

(1.103)

1.4 Operator matrix elements

We will here derive the matrix elements of operators in the path integral for-

malism.

1.4.1 The time-ordered product of operators

Let us ﬁrst recall the basic deﬁnition of quantities in the Heisenberg picture.

Given an operator

ˆ

O in the Schroedinger picture, the time evolved Heisenberg

picture operator is

ˆ

O(t) ≡ e

iHt

ˆ

Oe

−iHt

. (1.104)

In particular for the position operator we have

ˆ x(t) ≡ e

iHt

ˆ xe

−iHt

. (1.105)

Given a time the independent position eigenstates [x), the vectors

[x, t) ≡ e

iHt

[x) (1.106)

represent the basis vectors in the Heisenberg picture (notice the “+” in the

exponent as opposed to the “−” in the evolution of the state vector in the

Schroedinger picture!), being eigenstates of ˆ x(t)

ˆ x(t) [x, t) = x[x, t) . (1.107)

We have the relation:

¸x

f

, t

f

[x

i

, t

i

) =

_

D[x(t)]e

i

S[x(t)]

ℏ

. (1.108)

By working out precisely the same algebra which we used in section

Consider now a function A(x). It deﬁnes an operator on the Hilbert space:

ˆ

A(ˆ x) ≡

_

[x) ¸x[ A(x)dx (1.109)

Using the factorization property of the path integral we can derive the fol-

lowing identity:

_

D[x]A(x(t

1

))e

i

S[x]

ℏ

=

_

D[x

a

]D[x

b

]dx

1

A(x

1

)e

i

S[xa]+S[x

b

]

ℏ

=

_

dx

1

¸x

f

[ e

−

iH(t

f

−t

1

)

ℏ

[x

1

) ¸x

1

[ e

−

iH(t

1

−t

i

)

ℏ

[x

i

) A(x

1

) =

¸x

f

[ e

−

iH(t

f

−t

1

)

ℏ ˆ

A(ˆ x)e

−

iH(t

1

−t

i

)

ℏ

[x

i

) =

¸x

f

[ e

−

iHt

f

ℏ ˆ

A(ˆ x(t

1

))e

iHt

i

ℏ

[x

i

) =

¸x

f

, t

f

[

ˆ

A(ˆ x(t

1

)) [x

i

, t

i

) (1.110)

28 CHAPTER 1. THE PATH INTEGRAL FORMALISM

Let us now consider two functions O

1

(x(t)) and O

2

(x(t)) of x(t) and let us

study the meaning of:

_

D[x]O

1

(x(t

1

))O

2

(x(t

2

))e

i

S[x]

ℏ

(1.111)

By the composition property, assuming that t

1

< t

2

, we can equal it to (see

ﬁgure 1.8):

_

D[x

a

]D[x

b

]D[x

c

]dx

1

dx

2

e

i

S[xa]+S[x

b

]+S[xc]

ℏ

O

2

(x

2

)O

1

(x

1

) (1.112)

a

b

c

t

i

t

1

t

2

t

f

Figure 1.8: The path from (x

i

, t

i

) to (x

f

, t

f

) is separated into three paths a, b

and c. We have to distinguish t

1

< t

2

from t

2

< t

1

.

This can be rewritten:

_

¸x

f

, t

f

[x

2

, t

2

) O

2

(x

2

) ¸x

2

, t

2

[x

1

, t

1

) O

1

(x

1

) ¸x

1

, t

1

[x

i

, t

i

) dx

2

dx

1

(1.113)

Using the equation (1.109), it can be rewritten:

¸x

f

[ e

−

iH(t

f

−t

2

)

ℏ ˆ

O

2

(ˆ x)e

−

iH(t

2

−t

1

)

ℏ ˆ

O

1

(ˆ x)e

−

iH(t

1

−t

i

)

ℏ

[x

i

) =

¸x

f

, t

f

[

ˆ

O

2

(t

2

)

ˆ

O

1

(t

1

) [x

i

, t

i

) (1.114)

Recall that this result is only valid for t

2

> t

1

.

Exercise: Check that for t

1

> t

2

, one gets instead:

¸x

f

, t

f

[

ˆ

O

1

(t

1

)

ˆ

O

2

(t

2

) [x

i

, t

i

) (1.115)

Thus, the ﬁnal result is:

_

D[x]O

1

(x(t

1

))O

2

(x(t

2

))e

i

S[x]

ℏ

= ¸x

f

, t

f

[ T[

ˆ

O

1

(t

1

)

ˆ

O

2

(t

2

)] [x

i

, t

i

) (1.116)

where the time-ordered product T[

ˆ

O

1

(t

1

)

ˆ

O

2

(t

2

)] is deﬁned as:

T[

ˆ

O

1

(t

1

)

ˆ

O

2

(t

2

)] = θ(t

2

−t

1

)

ˆ

O

2

(t

2

)

ˆ

O

1

(t

1

) +θ(t

1

−t

2

)

ˆ

O

1

(t

1

)

ˆ

O

2

(t

2

) (1.117)

where θ(t) is the Heaviside step function.

1.4. OPERATOR MATRIX ELEMENTS 29

Exercise: Consider a system with classical Lagrangian L = m

˙ x

2

2

−V (x). Treat-

ing V as a perturbation and using the path integral, derive the well-known

formula for the time evolution operator in the interaction picture

U(t) = e

−iH0t/ℏ

ˆ

T[e

−

i

ℏ

R

t

0

ˆ

V (t

′

,ˆ x)dt

′

] (1.118)

where H

0

=

ˆ p

2

2m

and

ˆ

V (t, ˆ x) = e

iH0t/ℏ

V (ˆ x)e

−iH0t/ℏ

.

30 CHAPTER 1. THE PATH INTEGRAL FORMALISM

Chapter 2

Functional and Euclidean

methods

In this chapter, we will introduce two methods:

• The functional method, which allows a representations of perturbation

theory in terms of Feynman diagrams.

• The Euclidean path integral, which is useful for, e.g., statistical mechanics

and semiclassical tunneling.

We will show a common application of both methods by ﬁnding the pertur-

bative expansion of the free energy of the anharmonic oscillator.

2.1 Functional method

Experimentally, to ﬁnd out the properties of a physical system (for example

an atom), we let it interact with an external perturbation (a source, e.g. an

electromagnetic wave), and study its response.

This basic empirical fact has its formal counterpart in the theoretical study

of (classical and) quantum systems.

To give an illustration, let us consider a system with Lagrangian L(x, ˙ x) and

let us make it interact with an arbitrary external forcing source:

L(x, ˙ x) →L(x, ˙ x) +J(t)x(t)

K(x

f

, t

f

; x

i

, t

i

[J) =

_

D[x]e

i

ℏ

R

(L(x, ˙ x)+J(t)x(t))dt

(2.1)

K has now become a functional of J(t). This functional contains important

physical information about the system (all information indeed) as exempliﬁed

by the fact that the functional derivatives of K give the expectation values of

ˆ x (and thus of all operators that are a function of x):

ℏδ

iδJ(t

1

)

ℏδ

iδJ(t

n

)

K(x

f

, t

f

; x

i

, t

i

[J)

¸

¸

¸

J=0

= ¸x

f

, t

f

[ T [ˆ x(t

1

), . . . , ˆ x(t

n

)] [x

i

, t

i

)

(2.2)

31

32 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

As an example, let us see how the eﬀects of a perturbation can be treated

with functional methods.

Let us look at the harmonic oscillator, with an anharmonic self-interaction

which is assumed to be small enough to be treated as a perturbation:

L

0

(x, ˙ x) =

m

2

˙ x

2

−

mω

2

2

x

2

, L

pert

(x, ˙ x) = −

λ

4!

x

4

(2.3)

The propagator K

0

[J] for the harmonic oscillator with a source J(t) is:

K

0

[J] =

_

D[x]e

i

ℏ

R

(L0+J(t)x(t))dt

(2.4)

Consider now the addition of the perturbation. By repeated use of the

functional derivatives with respect to J (see eq. (2.2)) the propagator can be

written as:

K[J] =

_

D[x]e

i

ℏ

R

(L0−

λ

4!

x

4

+J(t)x(t))dt

=

_

D[x]

n

1

n!

_ _

−iλ

ℏ4!

_

n

x

4

(t

1

) x

4

(t

n

)e

i

ℏ

R

(L0+J(t)x(t))dt

dt

1

dt

n

=

n

1

n!

_ _

−iλ

ℏ4!

_

n

ℏ

4

δ

4

δJ

4

(t

1

)

ℏ

4

δ

4

δJ

4

(t

n

)

K

0

[J]dt

1

dt

n

≡ exp

_

−

_

dt

iλℏ

3

4!

δ

4

δJ

4

(t)

_

K

0

[J] (2.5)

Thus, when we take the J → 0 limit, we get the propagator of the anhar-

monic hamiltonian in terms of functional derivatives of the harmonic propagator

in presence of an arbitrary source J.

2.2 Euclidean Path Integral

Instead of computing the transition amplitude ¸x

f

[ e

−

iHt

ℏ

[x

i

), we could have

computed the imaginary time evolution:

K

E

≡ ¸x

f

[ e

−

βH

ℏ

[x

i

) , β ∈ R

+

(2.6)

which corresponds to an imaginary time interval t = −iβ.

0

ǫ 2ǫ (N −1)ǫ

β = Nǫ

Figure 2.1: The interval between 0 and β is divided into N steps.

It is straightforward to repeat the same time slicing procedure we employed

in section 1.1.4 (see ﬁgure 2.1) and write the imaginary time amplitude as a

path integral. The result is:

K

E

(x

f

, x

i

; β) =

_

D[x]

E

e

−

1

ℏ

SE[x]

(2.7)

2.2. EUCLIDEAN PATH INTEGRAL 33

where D[x]

E

is a path integration measure we shall derive below and where

S

E

[x] =

_

β

0

_

m

2

_

dx

dτ

_

2

+V (x)

_

dτ ∼ βE (2.8)

Notice that S

E

is normally minimized on stationary points.

From a functional viewpoint K

E

is the continuation of K to the imaginary

time axis

K

E

(x

f

, x

i

; β) = K(x

f

, x

i

; −iβ) . (2.9)

Notice that more sloppily we could have also gotten K

E

by “analytic con-

tinuation” to imaginary time directly in the path integral:

K(t) =

_

D[x]e

i

ℏ

R

t

0

L(x, ˙ x)dt

′

(2.10)

Then, with the following replacements

t

′

= −iτ

dt

′

= −idτ

˙ x = i

dx

dτ

t = −iβ

we obviously get the same result. Notice that the proper time interval in the

relativistic description ds

2

= −dt

2

+dx

2

i

, corresponding to a Minkovsky metric

with signature −, +, +, +, gets mapped by going to imaginary time into a metric

with Euclidean signature ds

2

E

= dτ

2

+dx

2

i

. Hence the ‘Euclidean’ suﬃx.

Let us now compute explicitly by decomposing the imaginary time interval

[0, β] into N steps of equal length ǫ (see ﬁgure 2.1). As in section 1.1.4 we can

write:

K

E

(x

f

, x

i

; β) =

_

¸x

N

[ e

−

Hǫ

ℏ

[x

N−1

) ¸x

1

[ e

−

Hǫ

ℏ

[x

0

) dx

1

dx

N−1

(2.11)

where x

f

= x

N

and x

i

= x

0

.

We can compute, just as before:

¸x

′

[ e

−

Hǫ

ℏ

[x) =

_

¸x

′

[p) ¸p[ e

−

Hǫ

ℏ

[x) dp

≈

1

2πℏ

_

e

ip

ℏ

(x

′

−x)−

ǫ

ℏ

“

p

2

2m

+V (x)

”

dp

=

1

2πℏ

_

e

−

ǫ

2mℏ

(p−i

m

ǫ

(x

′

−x))

2

−

m

2ǫℏ

(x

′

−x)

2

−

ǫ

ℏ

V (x)

dp

=

_

m

2πǫℏ

e

−

1

ℏ

»

m

2

“

x

′

−x

ǫ

”

2

+V (x)

–

ǫ

=

_

m

2πǫℏ

e

−

1

ℏ

LE(x,x

′

)ǫ

≡

1

A

E

e

−

S

E

(x

′

,x)

ℏ

(2.12)

34 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

and then proceed to obtain the Euclidean analogue of eq. (1.16)

K

E

(x

f

, x

i

; β) = lim

ǫ→0

_

N−1

k=1

dx

k

1

A

N

E

e

−

ℏ

P

N−1

l=0

SE(x

l+1

,x

l

)

= lim

ǫ→0

1

A

E

_

N−1

k=1

dx

k

A

E

. ¸¸ .

R

D[x]E

e

−

S

E

(x

f

,x

i

)

ℏ

(2.13)

Synthetizing: to pass from the real time path integral formulation to the Eu-

clidean time formulation, one simply makes the change iǫ → ǫ in the factor A

deﬁning the measure of integration, and modiﬁes the integrand according to

iS = i

_

t

0

_

1

2

m˙ x

2

−V (x)

_

dt −→−S

E

= −

_

β

0

_

1

2

m˙ x

2

+V (x)

_

dτ

e

iS/ℏ

−→e

−SE/ℏ

(2.14)

2.2.1 Statistical mechanics

Recalling the result of the previous section we have

K

E

(x

f

, x

i

; β) =

n

Ψ

∗

n

(x

f

)Ψ

n

(x

i

)e

−βEn/ℏ

(2.15)

where ¦Ψ

n

¦ is an othonormal basis of eigenvectors of the Hamiltonian. We can

then deﬁne the partition function:

Z[β] =

_

dxK

E

(x, x; β) =

n

e

−βEn/ℏ

(2.16)

We recognize here the statistical mechanical deﬁnition of the thermal parti-

tion function for a system at temperature k

B

T = ℏ/β, where k

B

is the Boltz-

mann constant.

We can observe the equivalence:

Z[β] =

_

dxK

E

(x, x; β) =

_

{x(0)=x(β)}

D[x]e

−SE

(2.17)

Thus, the thermal partition function is equivalent to a functional integral

over a compact Euclidean time: τ ∈ [0, β] with boundaries identiﬁed. The

integration variables x(τ) are therefore periodic in τ with period β.

Let us now check that K(−iβ) leads to the right partition function for the

harmonic oscillator. Using equation (1.102), we get:

Z[β] =

_

K(x, x; −iβ)dx =

n

e

−βω(n+

1

2

)

(2.18)

which is indeed the expected result.

2.3. PERTURBATION THEORY 35

2.3 Perturbation theory

The purpose of this section is to illustrate the use of functional methods in the

euclidean path integral. We shall do so by focussing on the anharmonic oscilla-

tor and by computing in perturbation theory the following physically relevant

quantities

• the time ordered correlators on the ground state: ¸0[T[ˆ x(τ

1

) . . . ˆ x(τ

n

)][0)

• the free energy and the ground state energy.

2.3.1 Euclidean n-point correlators

Consider the euclidean path integral deﬁned between initial time τ

i

= −

β

2

and

ﬁnal time τ

f

= +

β

2

. As we did for the real time path integral, we can consider

the euclidean n-point correlator

_

x(

β

2

)=x

f

x(−

β

2

)=xi

D[x(τ)]

E

x(τ

1

) . . . x(τ

n

)e

−

S

E

[x(τ)]

ℏ

(2.19)

By working out precisely the same algebra of section 1.4, the euclidean correlator

is also shown to be equal to the operatorial expression

¸x

f

[e

−βH/2ℏ

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]e

−βH/2ℏ

[x

i

) (2.20)

where

ˆ x

E

(τ) = e

Hτ/ℏ

ˆ xe

−Hτ/ℏ

(2.21)

represents the continuation to imaginary time of the Heisenberg picture position

operator. Indeed, comparing to eq. (1.105) we have, ˆ x

E

(τ) = ˆ x(−iτ).

It is interesting to consider the limit of the euclidean path integral and

correlators when β →∞. This is most directly done by working in the operator

formulation. By using the complete set of energy eigenstatets the euclidean

amplitude can be written as

K

E

(x

f

,

β

2

; x

i

, −

β

2

) = ¸x

f

[e

−

βH

ℏ

[x

i

) = (2.22)

=

n

ψ

n

(x

f

)ψ

n

(x

i

)

∗

e

−βEn/ℏ

β→∞

= ψ

0

(x

f

)ψ

0

(x

i

)

∗

e

−βE0/ℏ

_

1 +O(e

−β(E1−E0)ℏ

)

_

where E

0

and E

1

are the energies of respectively the ground state and the ﬁrst

excited state. From the above we conclude that for β ≫ ℏ/(E

1

− E

0

) the

operator e

−βH/ℏ

acts like a projector on the ground state. We arrive at the

same conclusion also for the correlators

lim

β→∞

¸x

f

[e

−

βH

2ℏ

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]e

−

βH

2ℏ

[x

i

) (2.23)

=¸0[T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)][0) ψ

0

(x

f

)ψ

0

(x

i

)

∗

e

−

βE

0

ℏ

_

1 +O(e

−β(E1−E0)/ℏ

)

_

36 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

where by [0) we indicate the energy ground state (not the x = 0 position eigen-

state!). By taking the ratio of eq. (2.23) and eq. (2.22) we thus ﬁnd

lim

β→∞

¸x

f

[e

−

βH

2ℏ

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]e

−

βH

2ℏ

[x

i

)

¸x

f

[e

−

βH

ℏ

[x

i

)

= ¸0[T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)][0)

(2.24)

where the dependence on x

i

and x

f

has cancelled out. The time ordered cor-

relators on the ground state are also known as Feynman correlators. Using the

path integral fomulation we can suggestively write

¸0[T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)][0) = lim

β→∞

_

D[x(τ)]

E

x(τ

1

) . . . x(τ

n

)e

−

S

E

ℏ

_

D[x(τ)]

E

e

−

S

E

ℏ

(2.25)

Notice that the structure of above equation is reminiscent of thermal averages

in classical statistical mechanics, with the denominator playing the role of a

partition function. As we already said, in the β →∞ limit the above equation

is independent of x

i

, x

f

. For the sake simplicity we shall then focus on the case

x

i

= x

f

= 0 and deﬁne, at ﬁnite β, the normalized n-point correlator as

ℏ

n

2

G

D

(τ

1

, . . . , τ

n

) =

_

xi=x

f

=0

D[x(τ)]

E

x(τ

1

) . . . x(τ

n

)e

−

S

E

ℏ

_

xi=x

f

=0

D[x(τ)]

E

e

−

S

E

ℏ

(2.26)

where the suﬃx D indicates that this normalized n-point correlator was com-

puted by imposing Dirichlet boundary conditions x

i

= x

f

= 0 in the path

integral. In the deﬁnition of G

D

, ℏ

n/2

is factored out for later convenience.

The euclidan Feynman correlators are related to the Feynman correlators in

real time

¸0[T[ˆ x(t

1

) . . . ˆ x(t

n

)][0) (2.27)

by performing an analytic continuation τ

i

→ it

i

in such a way as to preserve

the time ordering. This is most simply done via a counter-clockwise rotation,

the Wick rotation, in the complex time plane as shown in ﬁg. 2.2. Indeed the

time ordered correlators, both for real and imaginary time, are by construction

analytic functions of the time coordinates over the connected domains where

none of the time coordinates (ex. τ

1

, τ

2

, . . . , τ

n

) coincide. Each such domain

corresponds to a give ordering of the n time coordinates. The non-analiticity

is purely due to the step functions that enforce time ordering and is localized

at the points where at least two coordinates coincide. Then by continuing the

time coodinates on trajectories that do not cross (like in the Wick rotation) one

remains within the domain of analyticity.

The real time Feynman correlators play an important role in relativistic

quantum ﬁeld theory. The n-point correlators are associated to the amplitudes

describing the scattering of n

I

initial particles into n

F

ﬁnal particles with n =

n

I

+n

F

.

2.3.2 Thermal n-point correlators

One can also deﬁne the correlators using the path integral with periodic bound-

ary conditions

_

x(−

β

2

)=x(

β

2

)

D[x(τ)]

E

x(τ

1

) . . . x(τ

n

)e

−

S

E

[x(τ)]

ℏ

. (2.28)

2.3. PERTURBATION THEORY 37

Figure 2.2: Wick rotation from imaginary to real time. The arrows on the

coordinate axes indicate the future direction, both for real and imaginary time

This object should expectedly be related to statistical mechanics. Indeed, in

the operator formulation, eq. (2.28) can be written as

_

¸x[e

−

βH

2ℏ

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]e

−

βH

2ℏ

[x)dx = Tr

_

e

−

βH

ℏ

[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]

_

(2.29)

Normalizing by the partition function Z = Tr[e

−βH/ℏ

] we obtain he thermal

average of the time ordered operator product T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]

¸

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]

_

β

=

1

Z(β)

Tr

_

e

−

βH

ℏ

T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)]

_

(2.30)

which in the path integral language corresponds to the normalized n-point ther-

mal correlator

(2.30) =

_

x(−

β

2

)=x(

β

2

)

D[x(τ)]

E

x(τ

1

) . . . x(τ

n

)e

−

S

E

ℏ

_

x(−

β

2

)=x(

β

2

)

D[x(τ)]

E

e

−

S

E

ℏ

≡ ℏ

n

2

G

P

(τ

1

, . . . , τ

n

)

(2.31)

where the have deﬁned G

P

in analogy with the Dirichlet n-point correlator of the

previous section. G

D

and G

P

only diﬀer by the choice of boundary conditions,

respectively Dirichlet for G

D

and periodic for G

P

, applied to the corresponding

path integrals.

Like with G

D

, by performing the Wick rotation τ

k

→it

k

we obtain

ℏ

n

2

G

P

(it

1

, . . . , it

n

) =

1

Z(β)

Tr

_

e

−

βH

ℏ

T[ˆ x(t

1

) . . . ˆ x(t

n

)]

_

, (2.32)

that is the thermal average of the time ordered product in real time.

38 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

Finally, notice that for β →∞ we have

lim

β→∞

e

−

βH

ℏ

Z(β)

= [0)¸0[ (2.33)

This is expected, as β → ∞ corresponds to a statistical ensemble at zero tem-

perature. Then at β →∞ both Dirichlet and periodic correlators coincide with

the Feynman correlator

lim

β→∞

ℏ

n

2

G

P

(τ

1

, . . . τ

n

) = lim

β→∞

ℏ

n

2

G

D

(τ

1

, . . . τ

n

) = ¸0[T[ˆ x

E

(τ

1

) . . . ˆ x

E

(τ

n

)][0) .

(2.34)

From the path integral viewpoint, the coincidence of G

D

and G

P

in the above

equation represents the expected fact that for ﬁnite τ

1

, . . . , τ

n

the boundary

conditions in eqs. (2.26,2.31) become irrelevant for β →∞

2.3.3 Euclidean correlators by functional derivatives

Consider the euclidean path integral for our system coupled to an external source

J(τ). For the case of Dirichlet boundary conditions x

i

= x

f

= 0 we deﬁne

K

E

[β, J] ≡ K

E

(0,

β

2

; 0, −

β

2

[J) =

_

xi=x

f

=0

D[x]

E

exp

_

−

1

ℏ

_ β

2

−

β

2

dτ (L

E

(x, ˙ x) −J(τ)x(τ))

_

(2.35)

and similarly for periodic boundary conditions

Z[β, J] ≡ K

E

(0,

β

2

; 0, −

β

2

[J) =

_

xi=x

f

D[x]

E

exp

_

−

1

ℏ

_ β

2

−

β

2

dτ (L

E

(x, ˙ x) −J(τ)x(τ))

_

. (2.36)

The correlators can then be expressed in terms of functional derivatives. Using

the deﬁnitions in eqs. (2.26,2.31) we have

ℏ

n

2

G

D

(τ

1

, . . . , τ

n

) =

1

K

E

[β, 0]

ℏδ

δJ(τ

1

)

. . .

ℏδ

δJ(τ

n

)

K

E

[β, J]

¸

¸

¸

J=0

(2.37)

ℏ

n

2

G

P

(τ

1

, . . . , τ

n

) =

1

Z[β, 0]

ℏδ

δJ(τ

1

)

. . .

ℏδ

δJ(τ

n

)

Z[β, J]

¸

¸

¸

J=0

(2.38)

Like in section 2.1 we deﬁne K

0

E

[β, J] and Z

0

[β, J] to be the path integrals

associated to the simple harmonic oscillator with euclidean lagrangian L

0

E

=

m˙ x

2

2

+

mω

2

x

2

2

. We want to consider the general anharmonic oscillator with

lagrangian

L

E

=

m˙ x

2

2

+

mω

2

x

2

2

+ ∆V (x) = L

0

E

+ ∆V (x) (2.39)

where ∆V can be treated as a small perturbation of the dynamics of L

0

E

. Again

just following the procedure shown in section 2.1, the path integrals for the

2.3. PERTURBATION THEORY 39

anharmonic oscillator can be written as

K

E

[β, J] = exp

_

−

1

ℏ

_

dτ∆V

_

ℏδ

δJ(τ)

_

_

K

0

E

[β, J] (2.40)

Z[β, J] = exp

_

−

1

ℏ

_

dτ∆V

_

ℏδ

δJ(τ)

_

_

Z

0

[β, J] (2.41)

Summarizing: to compute any interesting quantity (correlator or partition

function) we must

1. compute the unperturbed K

0

E

and Z

0

,

2. learn to take their derivatives with respect to J(τ) in an eﬃcient way.

The ﬁrst step amounts to a Gaussian integral. The second is basically a problem

in combinatorics, whose solution has a convenient diagrammatic representation:

the Feynman diagrams.

2.3.4 Computing K

0

E

[J] and Z

0

[J]

K

0

E

[β, J] and Z

0

[β, J] are just gaussian integrals with an external source (eu-

clidean forced harmonic oscillator). To compute them we can use the method

of Green’s functions.

First, let us recall the result for a gaussian integral with a linear source (in

the discrete ﬁnite case). We want to compute the integral:

I[J] =

_

n

k

dx

k

e

−

1

2

x

i

Oijx

j

+J

k

x

k

(2.42)

where O

ij

is a positive deﬁnite n n matrix and we sum over repeated indices.

The inverse G of the matrix O is deﬁned by:

G

ij

O

jk

= O

kj

G

ji

= δ

i

k

(2.43)

Such an inverse matrix exists, since O

ij

is assumed positive deﬁnite. We

then perform the change of variables ¯ x

i

= x

i

−G

ij

J

j

which obviously has trivial

Jacobian

k

dx

k

=

k

d¯ x

k

(2.44)

and such that

−

1

2

x

i

O

ij

x

j

+J

k

x

k

= −

1

2

¯ x

i

O

ij

¯ x

j

+

1

2

J

k

G

kl

J

l

(2.45)

Thus, by performing the d¯ x integral:

I[J] = I[0] exp

_

1

2

J

i

G

ij

J

j

_

(2.46)

We can easily transpose that discrete example to our continuum case. Con-

centrating ﬁrst on K

0

E

for the sake of clarity (but the discussion on Z

0

will be

analogous) we have

K

0

E

[β, J] =

_

D[x]

E

e

−

1

ℏ

R

h

m

2

x(τ)

_

−

d

2

dτ

2

+ω

2

_

x(τ)−J(τ)x(τ)

i

=

_

D[x]

E

e

−

1

ℏ

R

[

1

2

x(τ)

ˆ

Ox(τ)−J(τ)x(τ)]

. (2.47)

40 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

Notice that on the space of functions that vanish at the boundaries (that is at

τ = ±β/2) the diﬀerential operator

ˆ

O = m(−d

2

/dτ

2

+ ω

2

) is hermitian and

positive deﬁnite. As a matter of fact, in terms of its matrix elements between

time coordinate eigenstates ψ

τ1

(τ) ≡ δ(τ −τ

1

) and ψ

τ2

(τ) ≡ δ(τ −τ

2

)

O(τ

1

, τ

2

) ≡

_

dτψ

τ1

(τ)

ˆ

Oψ

τ2

(τ) = m

_

∂

τ1

∂

τ2

+ω

2

_

δ(τ

1

−τ

2

)

= m

_

−∂

2

τ1

+ω

2

_

δ(τ

1

−τ

2

) (2.48)

using integration by parts and repeatedly using x(±β/2) = 0, we can rewrite

the exponent in eq. (2.47) as

−

_

1

2

x(τ

1

)x(τ

2

)O(τ

1

, τ

2

)dτ

1

dτ

2

+

_

J(τ

1

)x(τ

1

) (2.49)

in full analogy with −

1

2

x

i

O

ij

x

j

+ J

i

x

i

in eq. (2.42) We emphasize here that

O(τ

1

, τ

2

) is a function (better a distribution) in the two variables τ

1

and τ

2

.

Like in the discrete case, it is convenient to rewrite the exponent using

the inverse of

ˆ

O, that is the Green’s function satisfying Dirichlet boundary

conditions G

D

(τ

1

.τ

2

)

ˆ

O

τ1

G

D

(τ

1

, τ

2

) ≡ m

_

−∂

2

τ1

+ω

2

_

G

D

(τ

1

, τ

2

) = δ(τ

1

−τ

2

) (2.50)

G

D

_

±

β

2

, τ

_

= G

D

_

τ, ±

β

2

_

= 0 . (2.51)

G

D

(τ

1

, τ

2

) is just the matrix form of

ˆ

O

−1

. Using the two equations above one

can also check that O G

D

≡

_

O(τ

1

, τ

3

)G

D

(τ

3

, τ

2

)dτ

3

= δ(τ

1

−τ

2

). The solution

of eqs. (2.50,2.51) is given by

G

D

(τ

1

, τ

2

) =

1

mω

sinh

_

ω

_

β

2

+τ

<

__

sinh

_

ω

_

β

2

−τ

>

__

sinh (ωβ)

(2.52)

where τ

>

= max(τ

1

, τ

2

) and τ

<

= min(τ

1

, τ

2

).

Notice the properties:

lim

ω→0

G

D

(τ

1

, τ

2

) =

1

m

_

β

2

−[τ

1

−τ

2

[ −2

τ

1

τ

2

β

_

(2.53)

lim

β→∞

G

D

(τ

1

, τ

2

) =

1

2mω

e

−ω|τ1−τ2|

(2.54)

Shifting the integration variable

x(τ) =

_

G

D

(τ, τ

′

)J(τ

′

)dτ

′

+y(τ) ≡ G J +y (2.55)

the exponent is rewritten as (we use to represent the convolution in τ)

1

2

x

ˆ

O x +J x =

1

2

(y

ˆ

O y +J G

D

ˆ

O G

D

ˆ

J + 2y J) −(J y +J G

D

J)

=

1

2

(y

ˆ

O y −J G

D

J) (2.56)

2.3. PERTURBATION THEORY 41

so that we ﬁnd

K

0

E

[β, J] = exp

_

1

2ℏ

_

dτ

1

dτ

2

J(τ

1

)G

D

(τ

1

, τ

2

)J(τ

2

)

__

D[y]

E

e

−

1

ℏ

S

0

E

[y]

(2.57)

= e

1

2ℏ

J·GD·J

K

0

E

[β, 0] = e

1

2ℏ

J·GD·J

_

mω

2π sinh(ωβ)

(2.58)

Following precisely the same steps we can compute the partition function

Z

0

[β, J]. The only diﬀerence is that in this case the functional integral is over

the space of periodic functions in the interval [−β/2, β/2]. The inverse of

ˆ

O is

now given by the Green’s function G

P

satisfying period boundary conditions:

G

P

(β/2, τ) = G

P

(−β/2, τ), G

P

(τ, β/2) = G

P

(τ, −β/2). Thus we ﬁnd

Z

0

[β, J] = e

1

2ℏ

J·GP·J

_

periodic

D[y]

E

e

−

1

ℏ

S

0

E

[y]

(2.59)

= e

1

2ℏ

J·GP·J

Z

0

[β, 0] =

e

1

2ℏ

J·GP·J

2 sinh

βω

2ℏ

(2.60)

As for G

P

one ﬁnds

G

P

(τ, τ

′

) =

1

2mω

cosh

_

ω(

β

2

−[τ − τ

′

[)

_

sinh

_

ω

β

2

_ (2.61)

Notice that for β →∞ with τ

1

and τ

2

ﬁxed

1

we have

lim

β→∞

G

P

= lim

β→∞

G

D

=

1

2mω

e

−ω|τ1−τ2|

(2.62)

showing that for β →∞ the boundary conditions expectedly do not matter.

2.3.5 Free n-point correlators

Applying the results of the previous section to eqs. (2.38) we obtain a useful

expression for the n-point correlators in the free case, that is the harmonic

oscillator. The crucial remark is that the J independent prefactors K

0

E

[β, 0]

and Z

0

[β, 0] drop out from the computation and so we simply get

ℏ

n

2

G

D,P

(τ

1

, . . . , τ

n

) = ℏ

n

2

_

k=n

k=1

δ

δJ(τ

k

)

_

e

1

2ℏ

J·GD,P ·J

¸

¸

¸

J=0

. (2.63)

Notice that via the rescaling J → J/

√

ℏ we have put in evidence a factor ℏ

n

2

.

As the formulae are the same for Dirichlet and period boundary conditions, we

shall drop the P and D in what follows. Working out eq. (2.63) explicitly we

ﬁnd

• for n odd: G(τ

1

, . . . , τ

n

) = 0. This simply follows because exp(

1

2

J G J)

is an even functional of J and we take the limit J → 0 after computing

the derivatives.

1

More precisely for |τ

1

−τ

2

| ≪β, as well as |τ

1,2

±β/2| ≪β for G

D

.

42 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

• for n even

G(τ

1

, . . . , τ

n

) =

p∈σn

G(τ

p(1)

, τ

p(2)

) G(τ

p(n−1)

, τ

p(n)

) . (2.64)

σ

n

is the group of all permutations of n elements where the permutations

diﬀering by the exchange of two elements in a pair or the exchange of a

pair are identiﬁed. This group contains

n!

(n/2)!2

(n/2)

≡ (n − 1)!! = (n −

1)(n −3) 3 1 elements.

Let us see how this works in the simplest cases.

Notice ﬁrst of all that the two point correlator (D or P) simply equals the

corresponding Green’s function (thus our notation was wisely chosen!)

G(τ

1

, τ

2

) =

δ

δJ(τ

1

)

δ

δJ(τ

2

)

e

1

2ℏ

J·GD,P·J

¸

¸

¸

J=0

(2.65)

=

δ

δJ(τ

1

)

_

J(τ)G(τ, τ

2

)dτe

1

2ℏ

J·GD,P ·J

¸

¸

¸

J=0

(2.66)

= G(τ

1

, τ

2

) (2.67)

To compute the correlator for higher n it is convenient to introduce the graphical

notation

A)

_

G(τ

1

, τ

n

)J(τ

n

)dτ

n

B) G(τ

1

, τ

2

)

≡

≡

τ

1

τ

1

τ

2

Each functional derivative acts in two possible ways:

• on the exponent, i.e. giving a factor like in A)

• on a factor of type A) brought down by a previous derivative, thus giving

a factor like in B)

Using the above graphical rules to perform the derivatives one easily sees that

after setting J →0 the terms that survive give eq. (2.64). For instance, the four

points correlator is given by the graph shown in ﬁgure 2.3

G(τ

1

, τ

2

, τ

3

, τ

4

) = G(τ

1

, τ

2

)G(τ

3

, τ

4

) +G(τ

1

, τ

3

)G(τ

2

, τ

4

)+

G(τ

1

, τ

4

)G(τ

2

, τ

3

) . (2.68)

The general result we have found is known in quantum ﬁeld theory as Wick’s

theorem. It is normally derived by using creation and destruction operators.

The path integral allows for a more direct, functional, derivation, which is what

we have just shown.

2.3. PERTURBATION THEORY 43

τ

1

τ

1

τ

1

τ

2

τ

2

τ

2

τ

3

τ

3

τ

3

τ

4

τ

4

τ

4

+ +

Figure 2.3: A graphical representation of the four points correlator.

2.3.6 The anharmonic oscillator and Feynman diagrams

Using the diagrammatic method outlined in the previous section, and working

in a perturbative expansion in ∆V , we can compute the n-point correlators, the

propagator and partition function as written in section 2.3.3. For illustrative

purposes, we shall focus on the simple example of the anharmonic oscillator with

an x

4

potential: ∆V =

λ4

4!

x

4

. We shall compute the leading O(λ

4

) correction to

the free energy. Additional examples, involving other perturbations and other

observables, are extensively discussed in the exercise sessions.

From statistical mechanics we have that the free energy F(β) of a system in

equilibrium at ﬁnite temperature is

F(β) = −

ℏ

β

ln Z[β] ↔ Z[β] = e

−

β

ℏ

F(β)

(2.69)

We also recall that for β →∞the partition function is dominated by the ground

state, so that in this limit the free energy coincides with the ground state energy

lim

β→∞

F(β) = E

0

(2.70)

Expanding eq. (2.41) at ﬁrst order in λ

4

we have (notice that we also rescaled

J →

√

ℏJ)

Z[β] = Z

0

[β]e

−

ℏλ

4

ℏ4!

R

dτ(

δ

δJ(τ)

)

4

e

1

2

J·GP·J

¸

¸

¸

J=0

(2.71)

= Z

0

[β]

_

1 −

ℏλ

4

4!

_

dτ

_

δ

δJ(τ)

_

4

+O

_

λ

2

4

_

_

e

1

2

J·GP·J

¸

¸

¸

J=0

(2.72)

We learned from Wick’s theorem that the free four points correlator factor-

izes as

δ

δJ(τ

1

)

δ

δJ(τ

4

)

e

1

2

J·G·J

= (G(τ

1

, τ

2

)G(τ

3

, τ

4

)+

G(τ

1

, τ

3

)G(τ

2

, τ

4

) +G(τ

1

, τ

4

)G(τ

2

, τ

3

)) (2.73)

And thus, in the case at hand, the linear term in the expansion of Z[β] is

−

ℏλ

4

4!

_

dτ

_

δ

δJ(τ)

_

4

e

1

2

J·GP·J

¸

¸

¸

J=0

= −

3ℏλ

4

4!

_ β

2

−

β

2

G

2

P

(τ, τ)dτ . (2.74)

This result is graphically represented by ﬁgure 2.4: it simply corresponds to

ﬁgure 2.3 in the limiting case in which the four points are taken to coincide,

44 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

τ

1

= τ

2

= τ

3

= τ

4

= τ. The coordinate τ is called the interaction point, and

the four lines associated to this point form what is called an interaction vertex.

Notice that the factor 3 precisely counts the number of diagrams in the Wick

expansion of G(τ

1

, τ

2

, τ

3

, τ

4

). This graphical representation of a term in the

perturbative expansion of Z[β] is called a Feynman diagram.

3G(τ, τ)

2

≡ τ

Figure 2.4: A graphical representation of the part of the propagator linear in

λ

4

.

Substituting in eq. (2.72) the explicit expression for G

P

and performing the

(trivial) integral we ﬁnd

Z[β] = Z

0

[β]

_

1 −

ℏλ

4

32m

2

ω

3

(ωβ) coth

2

(

ωβ

2

) +O

_

λ

2

4

_

_

(2.75)

so that the free energy is

F =

ℏω

2

_

1 +

1

ωβ

ln(1 −e

−ωβ

) +

ℏλ

4

32m

2

ω

3

coth

2

(

ωβ

2

) +O

_

λ

2

4

_

_

. (2.76)

Notice that the second term in the expression in brackets represents the thermal

contribution to the harmonic oscillator free energy, while the third term gives

the leading anharmonicity correction. By taking the limit β → ∞ we obtain

the zero point energy of the anharmonic oscillator at lowest order in λ

4

E

0

=

ℏω

2

_

1 +

ℏλ

4

32m

2

ω

3

+O

_

λ

2

4

_

_

. (2.77)

Some comments on the above results are in order. By eq. (2.77), the dimension-

less parameter controlling the perturbative expansion is

¯

λ =

ℏλ

4

m

2

ω

3

. (2.78)

This result could have been anticipated in various ways. By eq. (2.71) the

expansion parameter must be proportional to ℏλ

4

, which however has dimension

mass

2

/time

3

. With the parameters at hand in the harmonic oscillator,

¯

λ is the

only dimensionless parameter proportional to ℏλ

4

. Also, more directly: the wave

function of the harmonic oscillator ground state is localized within a distance

X =

_

¸x

2

) ∼

_

ℏ/mω, and within this distance λ

4

X

4

/(mω

2

X

2

) ∼

¯

λ, sets the

relative size of the perturbation.

While

¯

λ ≪ 1 is, as was to be expected, necessary to apply perturbation

theory, the steps that lead to eqs. (2.76,2.77) leave something to be desired.

This is because in eq. (2.75) the ﬁrst term in the expansion is ∼

¯

λωβ, and

becomes inﬁnite in the limit β →∞, suggesting that the higher order terms in

the expansion also blow up in this limit. How can we then make sense of our

truncation at lowest order? Our procedure to compute E

0

makes sense as long

as there exists a range of β where

2.3. PERTURBATION THEORY 45

1. ∼

¯

λωβ ≪1

2. the contribution of the excited states to Z(β) is negligible, that is for

e

−

β

ℏ

(E1−E0)

≪1 −→

β

ℏ

(E

1

−E

0

) ≫1 (2.79)

which in our case, at lowest order corresponds, to ωβ ≫1

The simultaneous satisfaction of 1) and 2) implies again

¯

λ ≪ 1, and thus our

manipulations were technically justiﬁed.

But there is more. As a matter of fact, the Hamiltonian formulation tells us

that for β →∞ the exact partition function is

Z[β]

β→∞

−→ e

−

β

ℏ

E0

= e

−

βω

2

(1+

¯

λ

32

+... )

(2.80)

implying that the higher order corrections proportional to

¯

λωβ in eq. (2.75) must

build up into the exponential function. In the technical jargon one says that they

must exponentiate. Thus even though eq. (2.75) is not a good approximation

for β → ∞, the structure of the growing terms is ﬁxed to give an exponential.

As a corollary −ln Z[β]/β is well behaved for β → ∞ and for this quantity we

can safely apply perturbation theory.

In view of the above discussion we are now motivated to develop a picture

on the structure of the higher order terms in the expansion of Z[β]. The study

of the O(

¯

λ

2

) contribution is already very suggestive of what happens. Let us

focus on it ﬁrst. Expanding the exponent in eq. (2.71), we ﬁnd at O(λ

2

4

)

1

2!

_

ℏλ

4

4!

_

2

_

dτ

1

dτ

2

_

δ

δJ(τ

1

)

_

4

_

δ

δJ(τ

2

)

_

4

e

1

2

J·GP ·J

¸

¸

¸

¸

J=0

∼

¸

x

4

(τ

1

)x

4

(τ

2

)

_

(2.81)

Thus, we have to sum over diagrams that contain two points, τ

1

and τ

2

, with

four lines ending at each point. We can write three diﬀerent diagrams, as can

be seen in ﬁgure 2.5.

A

C

B

τ

1

τ

1

τ

1

τ

2

τ

2

τ

2

Figure 2.5: The three diﬀerent diagrams contributing to the corrections to the

ﬁrst energy level of the anharmonic oscillator. Diagrams A and B (at the top-

left corner and at the bottom) are connected, while diagram C consists of two

disconnected pieces.

46 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

The second order term is then shown, by direct computation, to be equal to:

1

2!

_

−

ℏλ

4

4!

_

2

_

dτ

1

dτ

2

_

24G

4

P

(τ

1

, τ

2

)+

72G

P

(τ

1

, τ

1

)G

2

P

(τ

1

, τ

2

)G

P

(τ

2

, τ

2

) + 9G

2

P

(τ

1

, τ

1

)G

2

P

(τ

2

, τ

2

)

¸

(2.82)

where each of these three addenda corresponds to one of the diagrams in ﬁgure

2.5.

The numerical factors in front of each of the three factors represents the

multiplicity of the corresponding Feynman diagram. It is a useful exercise to

reproduce these numbers. Let us do it for diagram A. To count its multiplicity

we have to count all possible ways to connect the 4 lines coming out of the τ

1

ver-

tex with the 4 lines coming out of the τ

2

vertex. Let us pick one given line from

τ

1

: it can connect to τ

2

in 4 possible ways. After having connected this ﬁrst line,

let us pick one of the remaining 3 lines from τ

1

: it can connect to τ

2

in 3 possible

ways. After this we pick one of the remaining two lines, which can connect to τ

2

in 2 possible ways. The remaining line can connect in just 1 way. Therefore the

number of combinations is 4 3 2 1 = 24. In the same way one can repro-

duce the factors 72 and 9 associated respectively to diagram B and C. The total

number of diagrams is therefore 24+72+9 = 105. On the other hand, eq. (2.81)

tells us that the number of diagrams should be the same as for the correlator

among 8 x’s (more precisely ¸x(τ

1

)x(τ

1

)x(τ

1

)x(τ

1

)x(τ

2

)x(τ

2

)x(τ

2

)x(τ

2

))). In-

deed the combinatoric factor of section 2.3.5 for the 8-point correlator gives

8!/(4! 2

4

) = 105.

In diagrams A and B there exist at least one line (in fact more) connecting

τ

1

and τ

2

. These diagrams are therefore called connected. On the other hand

in diagram C no line exists connecting τ

1

and τ

2

: diagram C consists of two

disconnected copies of the lowest order diagram in Fig.2.4. This implies that the

two-dimensional integral corresponding to C factorizes into the square of one-

dimensional integral. Diagrams A and B correspond to genuine, non-trivial,

two-dimensional integrals.

Putting together the diagrams in Figs. 2.4 and 2.5 the partition function

expanded at second order can be written as

Z[β] = Z

0

[β]

_

1 −

_

ℏλ

4

8

_

G

2

P

(τ, τ)dτ

_

+

1

2

_

ℏλ

4

8

_

G

2

P

(τ, τ)dτ

_

2

+A +B +O(λ

3

4

)

_

(2.83)

where we do not have written the explicit expressions of the connected diagrams

A and B. The point of the above expression is that the disconnected diagram

C has the right coeﬃcient to provide the ﬁrst step in the exponentiation of the

contribution of order λ

4

. This pattern continues at higher orders and for all

diagrams: the disconnected diagrams just serve the purpose of exponentiating

the contribution of the connected ones

Z[β] = Z

0

[β] e

(

P

connected diagrams)

. (2.84)

For instance, at O(λ

3

4

), in addition to new connected diagrams, there will be

disconnected diagrams corresponding to three copies of Fig. 2.4 and also dia-

grams made of one copy of Fig. 2.4 with one copy of A (or B). Notice, as it

should be evident, that the connected diagrams appear in the above exponent

2.3. PERTURBATION THEORY 47

precisely with the same coeﬃcient they have when they appear for the ﬁrst time

in the expansion of Z[β].

The exponentiation in Z[β] implies that for the free energy we simply have

F = F

0

+

_

connected diagrams

_

. (2.85)

Important result: in order to compute the free energy, and thus also the

ground state energy, we only need to compute the connected diagrams.

One can easily check exponentiation of the O(λ

4

) correction in Fig. 2.4

exp

_

−

ℏλ

4

4!

_

dτ

_

δ

δJ(τ)

_

4

_

e

1

2

J·GP·J

¸

¸

¸

¸

J=0

=

n

1

n!

_

−

ℏλ

4

4!

_

n

_

dτ

1

dτ

n

_

δ

δJ(τ

1

)

_

4

_

δ

δJ(τ

n

)

_

4

exp

_

1

2

J G

P

J

_

=

n

1

n!

__

−

3ℏλ

4

ℏ4!

_

dτG

2

P

(τ, τ)

_

n

+ (other topologies)

_

=

exp

_

−

3ℏλ

4

4!

_

dτG

2

P

(τ, τ)

_

[1 + (other topologies)] (2.86)

The general proof of eq. (2.84) is an exercise in combinatorics. We shall not

present it, but leave it as a challenge to the student. Instead in what follows

we want to give a simple proof, valid strictly speaking only for β → ∞. It is a

reﬁnement of the discussion around eq. (2.80).

Notice ﬁrst of all that connected diagrams like in ﬁgure 2.6 give a contribu-

tion that grows linearly in β.

Figure 2.6: Two examples of connected diagrams that contribute to the propa-

gator of the anharmonic oscillator. For example, the diagram on the left corre-

sponds to

_

dτ

1

dτ

2

G

4

(τ

1

, τ

2

) ∼ β.

To prove that, notice that in the large β limit, G(τ

1

, τ

2

) tends to zero ex-

ponentially when τ

1

≫ τ

2

. Consider then the general structure of a diagram

involving n interaction points

∝ λ

n

4

_

dτ

1

. . . dτ

n

F(τ

1

, . . . , τ

n

) (2.87)

where F consists of a product of G(τ

i

, τ

j

). It is graphically obvious that, if a

diagram is connected, the associated function F is signiﬁcantly diﬀerent than

zero only in the region τ

1

≃ τ

2

≃ ≃ τ

n

where all time coordinates τ

j

are

comparable. By integrating ﬁrst on τ

2

, . . . τ

n

, the integral converges fast and we

just get a number c so that

_

dτ

1

dτ

2

. . . dτ

n

F(τ

1

, . . . , τ

n

) =

_

dτ

1

c = cβ (2.88)

48 CHAPTER 2. FUNCTIONAL AND EUCLIDEAN METHODS

therefore connected diagrams give a contribution linear in β. This argument

easily generalizes to diagrams that consist of 2 or more disconnected pieces. In

that case the integrand is non vanishing when the time coordinates are compa-

rable within each connected subdiagram. Therefore a diagram consisting of k

disconnected diagrams will produce a contribution growing like β

k

.

Thus for β →∞ we have

Z[β] = Z

0

[β]

_

1 +A

1

β + A

2

β

2

+. . .

_

(2.89)

where A

k

is a sum over all diagrams involving k connected subdiagrams. In

particular A

1

is the sum over the connected diagrams.

Now, since Z[β] ∼ e

−βE0

, the coeﬃcients of β

n

must be such as to exponen-

tiate A

1

β, i.e.

A

n

=

1

n!

(A

1

)

n

(2.90)

and we can write:

lim

β→∞

Z[β] = lim

β→∞

Z

0

[β]e

A1β

(2.91)

⇒E

0

=

ℏω

2

−ℏA

1

Once again we found that the ground state energy is given by the sum of all

connected diagrams.

Chapter 3

The semiclassical

approximation

Let us ask an intuitive question. When do we expect the wave nature of a

particle propagation to be well appoximated by the properties of the trajectory

of a classical particle?

Well, as for the propagation of light, when the wavelength λ is much smaller

than the characteristic physical size of the system, diﬀraction phenomena are

unimportant, and we can use geometric optics.

In the quantum case, we expect similar results: if the wavelength λ is short

enough with respect to the typical length over which the potential varies L, we

can expect to be able to form a wave packet of size between λ and L, whose

motion should be well appoximated by that of a classical particle.

We have two approximation methods to analytically treat systems that can-

not be solved exactly:

Perturbation theory: Let us assume we are dealing with a system whose

hamiltonian H can be written as H = H

0

+λH

p

where H

0

is exactly solv-

able and λH

p

can be treated as a small perturbation (that is [λ¸n[H

p

[m)[ ≪

[E

(0)

n

−E

(0)

m

[, where [n) and E

(0)

n

are respectively the eigenstates and eigen-

values of H

0

). In this case, we work in perturbation theory in λ and we

compute the eigenvalues of the hamiltonian E

n

as a series in λ:

E

n

= E

(0)

n

+λE

(1)

n

+λ

2

E

(2)

n

+. . .

and similarly we expand the eigenvectors H.

Semiclassical methods: We use these methods when a system is close to the

classical limit ℏ →0, but more generally also when dealing with tunneling

phenomena. These methods oﬀer no diagrammatic representation: when

small couplings are involved, semiclassical contributions are proportional

to e

−1/λ

2

and thus they oﬀer non-perturbative corrections to the system.

49

50 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

3.1 The semiclassical propagator

The goal of this section is to compute the leading contribution to the path

integral in the formal limit ℏ →0. In practice this will correspond to deriving a

general expression for the Gaussian, or semiclassical, approximation discussed

in section 1.3.2. Let us recall it. Starting from the propagator

K(x

f

, t

f

; x

i

, t

i

) =

_

D[x]e

i

S[x]

ℏ

(3.1)

we expand it around the classical solution x

c

: x = x

c

+y and get

K(x

f

, t

f

; x

i

, t

i

) = e

i

S[xc]

ℏ

_

D[y]e

i

ℏ

“

1

2

δ

2

S

δx

2

y

2

+

1

3!

δ

3

S

δx

3

y

3

+...

”

(3.2)

The rescaling y =

√

ℏ˜ y makes it evident that the expansion in powers of y truly

corresponds to an expansion in powers of ℏ

K(x

f

, t

f

; x

i

, t

i

) = ^ e

i

S[xc]

ℏ

_

D[˜ y]e

i

2

δ

2

S

δx

2

˜ y

2

+

√

ℏO(y

3

)

. (3.3)

Finally, dropping orders higher than the second in the Taylor expansion of the

action, we obtain the Gaussian, or semiclassical, approximation to the propa-

gator:

K

sc

(x

f

, t

f

; x

i

, t

i

) = ^ e

i

S[xc]

ℏ

_

D[˜ y]e

i

2

δ

2

S

δx

2

˜ y

2

(3.4)

Let us now consider the simple case where:

S[x] =

_ _

m˙ x

2

2

−V (x)

_

dt (3.5)

We can expand at quadratic order around x

c

:

m˙ x

2

2

→

m

˙

˜ y

2

2

V (x) →

1

2

V

′′

(x

c

)˜ y

2

where we used the notation V

′′

(x

c

) = ∂

2

x

V (x)[

x=xc

.

Thus, we must compute:

I[x

c

] = ^

_

D[˜ y]e

i

2

R

dt(m

˙

˜ y

2

−V

′′

(xc)˜ y

2

)

. (3.6)

This is a Gaussian path integral like the one we found for the harmonic os-

cillator in subsection 1.3.4. More precisely if we write V

′′

(x

c

(t)) ≡ Ω

2

(t), the

computation generalizes that of the ﬂuctuation determinant for the harmonic

oscillator, for which Ω

2

(t) = ω

2

is time independent. We can thus proceed along

the same lines of subsection 1.3.4 and ﬁrst perform the change of variables

˜ y(t) =

n

˜ a

n

y

n

(t) (3.7)

3.1. THE SEMICLASSICAL PROPAGATOR 51

where y

n

are the orthonormalized eigenmodes of the diﬀerential operator

_

−m

d

2

dt

2

−V

′′

(x

c

(t))

_

≡

_

−m

d

2

dt

2

−Ω

2

(t)

_

(3.8)

which controls the quadratic action. We have again the change of measure

1

^ D[˜ y(t)] =

˜

^

n

d˜ a

n

√

2πi

(3.9)

so that in the end the gaussian integral in eq. (3.6) gives

I[x

c

] =

˜

^

_

det

_

−m

d

2

dt

2

−V

′′

(x

c

)

__

−

1

2

=

˜

^D

−

1

2

(3.10)

One ﬁnal, important, remark before going to the computation, is that the Jaco-

bian factor

˜

^ is independent of V

′′

(x

c

(t)). Consider indeed choosing a diﬀerent

orthonormal basis ¦y

′

n

(t)¦ on the space of square integrable functions on the

time interval T = t

f

−t

i

˜ y(t) =

n

˜ a

n

y

n

(t) =

m

˜ a

′

m

y

′

m

(t) . (3.11)

We have

˜ a

n

= U

nm

˜ a

′

m

(3.12)

where U

mn

is an (inﬁnite) unitary matrix satisfying U

†

U = 1 and detU = 1 as

it follows from standard linear algebra given that the two sets ¦y

n

(t)¦, ¦y

′

n

(t)¦

are orthonormalized. We thus have

n

d˜ a

n

√

2πi

=

m

d˜ a

′

m

√

2πi

(3.13)

which proves that the Jacobian

˜

^ is independent of the orthonormal basis and,

a fortiori, is independent of whatever diﬀerential operator has that basis as its

eigenbasis.

Let us now proceed. The computation of determinants as in eq. (3.10) was

ﬁrst carried out by Gelfand and Yaglom. In our derivation we shall follow

a clever, but less formal, derivation due to Coleman. A more mathematical

derivation can be found in [8]. The result we are are going to prove is that

D

˜

^

2

= c ψ

0

(t

f

) (3.14)

where:

• ψ

0

(t) satisﬁes the diﬀerential equation

−

_

md

2

dt

2

+ Ω

2

(t)

_

ψ

0

(t) = 0 (3.15)

with the boundary conditions

ψ

0

(t

i

) = 0 ψ

′

0

(t

i

) = 1 . (3.16)

1

Notice that, just to confuse the reader(!), we use a diﬀerent simbol for the Jacobian with

respect to eq. (1.90). This is because we are now working with rescaled variables ˜ y and ˜ an

which absorbed a power of ℏ. This is to orient the reader on the origin of the factors, but

beware that these Jacobians are anyway ill deﬁned when ǫ →0!

52 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

• c is a constant that does not depend on Ω

2

(t), and thus we can compute

it in the free particle case, Ω

2

(t) = 0.

Consider now two diﬀerent time dependent frequencies: Ω

2

1

(t) and Ω

2

2

(t).

Consider also the solution of the “eigenvalue” equation, generalizing eq. (3.15):

−

_

md

2

dt

2

+ Ω

2

1

(t)

_

. ¸¸ .

ˆ

O1

ψ

(1)

λ

(t) = λψ

(1)

λ

(t) (3.17)

with boundary conditions

ψ

(1)

λ

(t

i

) = 0 ∂

t

ψ

(1)

λ

(t

i

) = 1 . (3.18)

• For all λ, these two conditions ﬁx the solution everywhere, and thus we

can compute ψ

(1)

λ

(t

f

).

• λ is an eigenvalue of

ˆ

O

1

if, and only if ψ

(1)

λ

(t

f

) = 0. Indeed, remember

that we are looking for the eigenvalues of

ˆ

O

1

over the space of functions

that vanish at t

i

and t

f

.

Consider then the same problem for

ˆ

O

2

≡ −

_

md

2

dt

2

+ Ω

2

2

(t)

_

and call ψ

(2)

λ

the solution of the diﬀerential equation

ˆ

O

2

ψ

(2)

λ

= λψ

(2)

λ

with the same boundary

conditions in 3.18.

We have then the following

Theorem:

det

_

ˆ

O

1

−λ

_

det

_

ˆ

O

2

−λ

_ =

ψ

(1)

λ

(t

f

)

ψ

(2)

λ

(t

f

)

(3.19)

We shall now prove the above by following Coleman, and, as it is customary

in physics, emphasizing ideas above mathematical rigour.

Let us deﬁne ﬁrst:

f(λ) =

ψ

(1)

λ

(t

f

)

ψ

(2)

λ

(t

f

)

(3.20)

f satisﬁes the following properties:

• f(λ) has obviously all its zeroes in correspondence with the eigenvalues of

ˆ

O

1

and all its poles in correspondence with eigenvalues of

ˆ

O

2

. All these

zeroes and poles are simple as we show in the mathematical appendix of

subsection 3.1.2.

• f(λ) → 1 in the limit λ → ∞, λ / ∈ R

+

. This property follows by simple

solution of the diﬀerential equation 3.17. Indeed, for λ → ∞, we can

neglect Ω

2

1,2

(t), and thus the solution satisfying the boundary conditions

3.18 is written as (see the mathematical appendix 3.1.3 for more detail)

ψ

(1,2)

λ

(t) =

1

2i

_

m

λ

_

e

i

√

λ

m

(t−ti)

−e

−i

√

λ

m

(t−ti)

_

+O

_

Ω

2

1,2

λ

_

(3.21)

3.1. THE SEMICLASSICAL PROPAGATOR 53

One of the two exponentials grows for λ / ∈ R

+

. Thus, there are no zeroes

at t = t

f

, and we can safely neglect Ω

2

1,2

for all t. On the other hand, for

λ/m = [nπ/(t

f

− t

i

)]

2

, the zeroth order solution (ﬁrst term in eq. (3.21))

vanishes at t = t

f

and therefore the Ω

2

1,2

perturbation cannot be treated

as small, showing that lim

λ→∞

f(λ) ,= 1 for λ ∈ R

+

.

Consider now:

g(λ) =

det

_

ˆ

O

1

−λ

_

det

_

ˆ

O

2

−λ

_ =

det

_

−m∂

2

t

−λ −Ω

2

1

_

det (−m∂

2

t

−λ −Ω

2

2

)

(3.22)

• It has exactly the same simple poles and simple zeroes as f(λ).

• For λ →∞, λ / ∈ R

+

, we have g(λ) →1. Without aiming at mathematical

rigour (but a more thorough discussion is presented in Homework 8) this

result is qualitatively understood as follows. Notice that

ˆ

T ≡ −m∂

2

t

− λ

has eigenvalues λ

n

= (nπ/(t

f

− t

i

))

2

− λ. This result for λ

n

implies that

the norm [λ

min

[ of the eigenvalue of smallest norm λ

min

goes to inﬁnity

when λ →∞, λ / ∈ R

+

. In this situation the addition to this operator of Ω

2

1

or Ω

2

2

is a small perturbation with negligible eﬀects as λ →∞. Therefore

we expect g(λ) → 1 for λ → ∞, λ / ∈ R

+

. Obviously this reasoning does

not work for λ ∈ R

+

, as the function g has zero and poles for arbitrarily

large λ.

Putting everything together we have that the ratio

h(λ) =

f(λ)

g(λ)

is analytic everywhere (the singularity at the simple poles of f is neutralized

by simple zeroes of 1/g and conversely the singularity at the simple zeroes of

g is neutralized by simple zeroes of f) and tends to 1 as λ → ∞. By Cauchy

theorem the only such function is h(λ) ≡ 1, and that proves our theorem. More

precisely, applying Cauchy theorem to h(λ) −1 and integrating over a circle C

with radius R →∞ we have

h(λ) −1 =

1

2πi

_

C

h(z) −1

z −λ

dz →0 (3.23)

where in the last step we used that lim

|z|→∞

(h(z) −1) = 0.

Consider now Ω

2

1

(t) ≡ Ω

2

(t) and call ψ

(Ω)

0

the solution to the problem in

eqs. (3.17,3.18) at λ = 0. As a corollary of eq. (3.19) we have that

det

_

−

md

2

dt

2

−Ω

2

(t)

_

˜

^

2

1

ψ

(Ω)

0

(t

f

)

= c (3.24)

where c is independent of Ω(t), and where we have judiciously inserted the

normalization Jacobian (which, like c, is independent of Ω

2

) in order to deal only

with the physically interesting combination

˜

^/Det

1/2

. Provided we ﬁnd the

54 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

constant c, the solution to the diﬀerential equation problem in eqs. (3.17,3.18)

at λ = 0, gives us the determinant we were after! Finding the constant c is easy:

we just need to use eq. (3.24) for an operator whose determinat we already know.

The easiest option to do so is to consider the free particle case, where Ω(t) = 0.

In this case the determinat prefactor in K was computed in eq. (1.48)

1

˜

^

2

det

_

−

md

2

dt

2

_

=

2πi(t

f

−t

i

)ℏ

m

(3.25)

while the solution to eqs. (3.15,3.16) is

ψ

(0)

0

(t) = t −t

i

(3.26)

and thus:

c =

2πiℏ

m

(3.27)

The result for the gaussian integral in the general case is then

I[x

c

] =

¸

m

2πiℏψ

(Ω)

0

(t

f

)

(3.28)

3.1.1 VanVleck-Pauli-Morette formula

Let us now derive a useful implicit expression for ψ

(Ω)

0

(t

f

). The classical equa-

tion of motion tells us

m¨ x

c

+V

′

(x

c

) = 0 (3.29)

where the dot denotes derivation with respect to time. By diﬀerentiating it once

we get

m

˙

¨ x

c

+ V

′′

(x

c

) ˙ x

c

= 0 (3.30)

Thus, v

c

= ˙ x

c

satisﬁes the diﬀerential equation that we are trying to solve

in order to ﬁnd ψ

(Ω)

0

(t). However, it doesn’t satisfy in general the boundary

conditions ψ

0

(t

i

) = 0, ∂

t

ψ

0

(t

i

) = 1. Consider now the Wronskian:

W(t) = v

c

(t)

˙

ψ

0

(t) − ˙ v

c

(t)ψ

0

(t) (3.31)

where ψ

0

(t) = ψ

(Ω)

0

(t) is the function we are looking for.

Since ψ and v

c

satisfy the same diﬀerential equation, we have

˙

W(t) = 0, and

thus W(t) = W is a constant. We can then write:

v

2

c

d

dt

_

ψ

0

v

c

_

= v

2

c

_

˙

ψ

0

v

c

−

ψ

0

˙ v

c

v

2

c

_

= W = v

c

(t

i

)

⇒

d

dt

_

ψ

0

v

c

_

= v

c

(t

i

)

1

v

2

c

⇒ψ

0

(t) = v

c

(t)v

c

(t

i

)

_

t

ti

1

v

2

c

(t

′

)

dt

′

⇒ψ

0

(t

f

) = v

c

(t

f

)v

c

(t

i

)

_

t

f

ti

1

v

2

c

(t

′

)

dt

′

(3.32)

3.1. THE SEMICLASSICAL PROPAGATOR 55

Let us now write ψ

0

(t

f

) in terms of energy and potential. We have, from

the classical solution:

mv

2

c

(t)

2

= E−V (x

c

(t)), and dt =

dx

vc

. Writing v

f

= v

c

(t

f

)

and v

i

= v

c

(t

i

):

ψ

0

(t

f

) = v

f

v

i

_

x

f

xi

_

m

2(E −V (x))

_3

2

dx (3.33)

≡ v

f

v

i

_

x

f

xi

dx

v(x)

3

(3.34)

The prefactor is then:

I[x

c

] =

¸

1

2πiℏv

f

v

i

√

m

_

x

f

xi

dx

2(E−V (x))

(3.35)

Or in terms of velocities:

I[x

c

] =

¸

m

2πiℏv

f

v

i

_

x

f

xi

dx

v

3

(x)

(3.36)

Let us now express the result in terms of the classical action. We have:

S

c

=

_

t

f

ti

_

m˙ x

2

c

2

−V

_

dt =

_

t

f

ti

_

m˙ x

2

c

−E

_

dt =

_

x

f

xi

_

2m(E −V (x))dx −E(t

f

−t

i

) (3.37)

Thus:

∂S

c

∂x

f

= P

f

=

_

2m(E −V (x

f

)) (3.38)

∂

2

S

c

∂x

i

∂x

f

=

∂P

f

∂x

i

=

1

2

¸

2m

E −V (x

f

)

∂E

∂x

i

=

1

v

f

∂E

∂x

i

(3.39)

We can note that:

t

f

−t

i

=

_

t

f

ti

dt =

_

x

f

xi

dx

˙ x

=

_

x

f

xi

_

m

2(E −V (x))

dx (3.40)

0 =

∂(t

f

− t

i

)

∂x

i

= −

1

v

i

−

1

2

∂E

∂x

i

_

x

f

xi

_

m

2

dx

(E −V (x))

3

2

⇒

∂E

∂x

i

= −

m

v

i

_

x

f

xi

dx

v

3

(x)

(3.41)

Thus, the derivative of the action can be written as:

∂

2

S

c

∂x

i

∂x

f

= −

m

v

i

v

f

_

x

f

xi

dx

v

3

(x)

(3.42)

And we get in the end the Van Vleck-Pauli-Morette determinant:

I[x

c

] = N

¸

1

det

_

−m

d

2

dt

2

−V

′′

(x

c

)

_ =

¸

−

1

2πiℏ

∂

2

S

c

∂x

i

∂x

f

(3.43)

56 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

And thus we can write the ﬁnal form of the semiclassical propagator:

K

sc

(x

f

, t

f

; x

i

, t

i

) = e

i

Sc

ℏ

¸

−

1

2πiℏ

∂

2

S

c

∂x

i

∂x

f

(3.44)

3.1.2 Mathematical Appendix 1

Let us now prove that ψ

λ

(t

f

) only has simple zeroes. Consider the equation:

_

−

md

2

dt

2

+V

_

ψ

λ

(t) = λψ

λ

(t) (3.45)

with the boundary conditions ψ

λ

(t

i

) = 0, ∂

t

ψ

λ

(t

i

) = 1.

We will use the notation:

˙

ψ

λ

(t) ≡

d

dλ

ψ

λ

(t)

ψ

′

λ

(t) ≡

d

dt

ψ

λ

(t)

When we diﬀerentiate (3.45) with respect to λ, we get:

_

−

md

2

dt

2

+V

_

˙

ψ

λ

(t) = λ

˙

ψ

λ

(t) + ψ

λ

(t) (3.46)

Notice that since we impose the same boundary conditions for all λ, we have:

˙

ψ

λ

(t

i

) =

˙

ψ

′

λ

(t

i

) = 0 (3.47)

Consider then the “Wronskian”:

W = ψ

λ

˙

ψ

′

λ

−ψ

′

λ

˙

ψ

λ

(3.48)

W

′

= ψ

λ

˙

ψ

′′

λ

−ψ

′′

λ

˙

ψ

λ

= ψ

λ

1

m

_

(V −λ)

˙

ψ

λ

−ψ

λ

_

−

1

m

(V −λ)ψ

λ

˙

ψ

λ

= −

ψ

2

λ

m

0 (3.49)

On the other hand, we have the boundary conditions:

W(t

i

) = 0, W(t

f

) = ψ

λ

(t

f

)

˙

ψ

′

λ

(t

f

) −ψ

′

λ

(t

f

)

˙

ψ

λ

(t

f

) (3.50)

Thus, if both ψ

λ

(t

f

) and

˙

ψ

λ

(t

f

) vanish, the Wronskian is zero everywhere,

and so is its derivative. Thus, ψ

λ

(t) ≡ 0.

3.1.3 Mathematical Appendix 2

We want to write explicitly the solution to

(−∂

2

t

−λ −Ω

2

(t))ψ = 0 ψ(0) = 0 , ψ

′

(0) = 1 (3.51)

treating Ω

2

(t) as a perturbation. In order to carry on perturbation theory in Ω

2

we need the retarded Green’s function for the zeroth order equation G(t, t

′

) =

3.2. THE FIXED ENERGY PROPAGATOR 57

ˆ

T

−1

where ≡

ˆ

T ≡ −∂

2

t

+λ. By a simple computation analogous to the one used

to ﬁnd the euclidean Green’s function one ﬁnds (see section III of ref. [8] for

instance)

G(t, t

′

) = θ(t −t

′

)

1

2i

√

λ

_

e

−2i

√

λ(t−t

′

)

−e

2i

√

λ(t−t

′

)

_

. (3.52)

Also deﬁning the zeroth order solution

ψ

0

=

1

2i

√

λ

_

e

i

√

λ(t−ti)

−e

−i

√

λ(t−ti)

_

(3.53)

and writing ψ = ψ

0

+φ, eq. (3.51) becomes

ˆ

Tφ = Ω

2

(ψ

0

+φ) (3.54)

which by use of G =

ˆ

T

−1

and by employing standard functional operator no-

tation (by ◦ we indicate the product of functional operators), can be written

as

(1 −G◦ Ω

2

)φ = G◦ Ω

2

ψ

0

(3.55)

and solved iteratively in φ

φ = (1 −G◦ Ω

2

)

−1

G◦ Ω

2

ψ

0

=

∞

n=1

(G◦ Ω

2

)

n

ψ

0

(3.56)

and therefore

ψ =

∞

n=0

(G◦ Ω

2

)

n

ψ

0

. (3.57)

Notice that ψ deﬁned above automatically satisﬁes the boundary conditions

ψ(0) = 0, ψ

′

(0) = 1. For the lowest order correction we have

(G◦ Ω

2

)ψ

0

=

_

t

0

dt

′

1

2i

√

λ

_

e

−2i

√

λ(t−t

′

)

−e

2i

√

λ(t−t

′

)

_

Ω

2

(t

′

)ψ

0

(t

′

) (3.58)

which for λ / ∈ R

+

is ∼ O(Ω

2

/

√

λ)ψ

0

.

3.2 The ﬁxed energy propagator

3.2.1 General properties of the ﬁxed energy propagator

The ﬁxed-energy retarded propagator, sometimes simply Green’s function, is

deﬁned as the Fourier transform of the retarded propagator θ(t)K(x

f

, t; x

i

, 0)

K(E; x

f

, x

i

) ≡

_

∞

0

K(x

f

, t; x

i

, 0)e

iEt

ℏ

dt

=

_

∞

0

¸x

f

[ e

−

i

ˆ

Ht

ℏ

+

iEt

ℏ

[x

i

) dt

= lim

ǫ→0

_

∞

0

¸x

f

[ e

it

ℏ

(E−

ˆ

H+iǫ)

[x

i

) dt

= lim

ǫ→0

¸x

f

[

iℏ

E −

ˆ

H +iǫ

[x

i

) (3.59)

58 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

where, as usual when dealing with distributions, the factor e

−ǫt/ℏ

with ǫ →0

+

is introduced to insure converge of the integral at t → ∞. Notice that since

we integrate over t > 0, the exponent e

iEt/ℏ

ensures strong convergence of

the integral for any E in the upper complex plane, that is for ImE > 0. For

ImE > 0 the retarded propagator K(E; x

f

, x

i

) is thus an analytic function of

the complex variable E. This is an important property of retarded Green’s func-

tions: analyticity in the upper energy plane is the Fourier space manifestation

of causality, that is of retardation. For the advanced propagator we would have

instead analyticity in the lower complex E plane.

From the last equation above, the retarded propagator K(E; x

f

, x

i

) is just

the position space representation of the inverse of the Schr¨odinger operator,

ˆ

K =

iℏ

E−

ˆ

H+iǫ

. It thus satisﬁes the relation:

lim

ǫ→0

(E −

ˆ

H +iǫ)

ˆ

K = iℏ1 (3.60)

In coordinate representation, equation (3.60) becomes:

_

−

ℏ

2

2m

∂

2

x

+V (x) −E

_

K(E; x, y) = −iℏδ(x −y) . (3.61)

This equation, and the boundary conditions we discuss below, fully ﬁx K(E; x

f

, x

i

).

Let us discuss this in detail. Given two linearly independent solutions ψ

1

and

ψ

2

of the homogeneous Schr¨odinger’s equation at energy E

_

−

ℏ

2

2m

∂

2

x

+V (x) −E

_

ψ

1,2

= 0 (3.62)

the following function satisﬁes eq. (3.61)

f(E; x, y) =

2mi

ℏW

(θ(x −y)ψ

1

(x)ψ

2

(y) +θ(y −x)ψ

2

(x)ψ

1

(y)) (3.63)

where W = ψ

′

1

(x)ψ

2

(x) − ψ

′

2

(x)ψ

1

(x) is the wronskian; we recall that, by

eq. (3.62), W is constant over x. However in order for eq. (3.63) to coincide with

the retarded propagator, we must choose ψ

1

and ψ

2

satisfying the appropriate

boundary conditions. Which boundary conditions are required? The answer is

obtained by simple physical considerations. In order to illustrate that, let us

work on a case by case basis.

Let us focus ﬁrst on motion in a potential V (x) satisfying lim

|x|→∞

V (x) = 0,

which is relevant for barrier penetration problems and in the case of potential

wells of ﬁnite spacial width. In this situation any solution ψ of the Schroedinger

equation with positive energy behaves at [x[ → ∞ as a linear combination of

outgoing and ingoing plane waves

lim

x→+∞

ψ = a

+

e

−i

√

2mEx

+b

+

e

−i

√

2mEx

(3.64)

lim

x→−∞

ψ = a

−

e

−i

√

2mEx

+b

−

e

−i

√

2mEx

(3.65)

The solutions ψ

1,2

are selected by requiring K(E; x

f

, x

i

) to have the physically

correct behaviour at [x

f

[ → ∞. To ﬁnd out which behaviour is the physically

correct one, we must recall that the inverse Fourier transform, the propagator,

K(x

f

, t; x

i

, 0) =

_

∞

−∞

K(E; x

f

, x

i

)e

−iEt/ℏ

dE

2πℏ

(3.66)

3.2. THE FIXED ENERGY PROPAGATOR 59

represents the wave function Ψ(x

f

, t) at t > 0 for a state which at t = 0 was

a δ-function at some ﬁnite position x

i

, that is lim

t→0

Ψ(x

f

, t) = δ(x

f

− x

i

).

Then at any ﬁnite t > 0, for x

f

→+∞≫x

i

, the wave function Ψ(x

f

, t) should

correspond to a superposition of waves travelling out in the positive x direction,

ie with positive momentum and positive energy. This implies

lim

x

f

→+∞

K(E; x

f

, x

i

) ∝ e

+i

√

2mEx

f

(3.67)

Similarly, for x

f

→−∞, Ψ(x

f

, t) should be a superposition of waves travelling

out in the negative x direction, ie with negative momentum

lim

x

f

→−∞

K(E; x

f

, x

i

) ∝ e

−i

√

2mEx

f

(3.68)

This gives us all the constraints to fully ﬁx ψ

1,2

and write a general expression

for K(E; x

f

, x

i

) in the form 3.63. Before discussing that, notice that, analiticity

in the upper energy plane, tells us to how to continue

√

E from E > 0 to E < 0

2

. Then for E < 0 (or more precisely E = limǫ →0

+

(−[E[ + iǫ) the above

asymptotic behaviours become

lim

x

f

→+∞

K(E; x

f

, x

i

) ∝ e

−

√

2m|E|x

f

→0 (3.69)

lim

x

f

→−∞

K(E; x

f

, x

i

) ∝ e

+

√

2m|E|x

f

→0 (3.70)

corresponding to K representing an operator on the space of normalizable

wave functions. Indeed to ﬁnd the correct boundary conditions we could have

worked the other way around, and consider ﬁrst the asymptotic behaviour of

K(E, x

f

, x

i

) for E < 0. By eq. (3.61) there are two possibilities: e

+

√

2m|E|x

f

and e

−

√

2m|E|x

f

. The request that K be square integrable implies the choice

in eqs. (3.69, 3.70). Then by analytically continuing eqs. (3.69, 3.70) to E > 0

through a path in the ImE > 0 half-plane we get the outgoing wave boundary

conditions of eqs. (3.67, 3.68).

3

By eq. (3.63) the behaviour at x

f

→ +∞ and x

f

→ −∞ is controlled

respectively by ψ

1

and ψ

2

. Then eqs. (3.67, 3.68) are satisﬁed if

lim

x→+∞

ψ

1

(x) ∝ e

+i

√

2m|E|x

(3.71)

lim

x→−∞

ψ

2

(x) ∝ e

−i

√

2m|E|x

(3.72)

and provided we have such two solutions we can write the retarded propagator

as

K(E; x, y) =

2mi

ℏW

(θ(x −y)ψ

1

(x)ψ

2

(y) +θ(y −x)ψ

2

(x)ψ

1

(y)) (3.73)

Summarizing: retardation corresponds to outgoing waves at x →±∞.

2

This is because the cut of

√

E should be chosen at ImE < 0 to ensure analyticity of

K(E; x

f

, x

i

) at ImE > 0.

3

On the other hand, if we continue eq. (3.70) to E > 0 by a path in the ImE < 0 half-plane

we would get incoming instead of outgoing boundary conditions. This would correspond to

the advanced propagator, which is analytic in the lower energy half-plane.

60 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

We can generalize these considerations to other potentials. For instance

consider the case where lim

x→−∞

V (x) = 0 and lim

x→+∞

V (x) = +∞, so that

the region x → ∞ is not accessible at ﬁnite energy. It is easy to ﬁnd the right

boundary conditions for ψ

1,2

for this case too. In the case of ψ

2

the condition is

the same as before: outgoing plane wave. So we have again ψ

2

≡ ψ

−

. For ψ

1

the

two possible behaviours at x → +∞ by solving the Schroedinger equation are

exponentially growing or exponentially decreasing. We obviously must choose

the second one to make physical sense. Then whe have that ψ

1

is just the

stationary solution ψ

stat

of the Schroedinger equation with energy E: it belongs

to the physical Hilbert space. Notice on the other hand, that ψ

−

, an outgoing

wave at x → −∞, will always be, by current conservation, a combination of

growing and decreasing solution at x → ∞. Obviously ψ

−

does not belong to

the physical Hilbert space. But this does not concern the asymptotic behaviour

of K because of the θ step functions in eq. (3.63). In this case the propagator

is then

K(E; x, y) =

2mi

ℏW

(θ(x −y)ψ

stat

(x)ψ

−

(y) +θ(y −x)ψ

−

(x)ψ

stat

(y)) (3.74)

The last case to consider is the one where motion is fully limited: lim

|x|→+∞

V (x) =

+∞. In this case both ψ

1

and ψ

2

must correspond, at any real value of E, to

the exponentially dcreasing solutions at respectively x →+∞ and x →−∞.

To illustrate our results let us consider the simplest possibe case of a free

particle: V (x) ≡ 0. We have:

_

ψ

1

(x) = e

ikx

ψ

2

(x) = e

−ikx

(3.75)

where k =

√

2mE.

We then get the ﬁxed energy retarded propagator for the free particle:

K(E; x, y) =

m

ℏk

_

θ(x −y)e

ik(x−y)

+θ(y −x)e

ik(y−x)

_

=

m

ℏk

e

ik|x−y|

(3.76)

Which is precisely the expression derived in Homework 6, by inserting a com-

plete set of ﬁxed momentum states in eq. (3.59) and performing the momentum

integral using Cauchy theorem. Notice that we recover a result similar to the

correlator G

β

(τ

1

, τ

2

) in the limit β →∞.

Now, when we switch on the potential, we will have to ﬁnd the two following

solutions to the Schr¨odinger equation:

_

ψ

1

≡ ψ

+

(x) →A

+

e

ikx

x →+∞

ψ

2

≡ ψ

−

(x) →A

−

e

−ikx

x →−∞

(3.77)

We will then deﬁne the retarded propagator by:

K(E; x, y) =

2mi

ℏW

(θ(x −y)ψ

+

(x)ψ

−

(y) +θ(y −x)ψ

+

(y)ψ

−

(x)) (3.78)

3.2. THE FIXED ENERGY PROPAGATOR 61

Example: Barrier penetration

With the potential V (x) switched on, assuming that it has a barrier located

somewhere along the x axis, we need ψ

1,2

to correspond to the solutions to the

Schr¨odinger equation with the following asymptotic behaviour:

ψ

1

≡ ψ

+

(x) →

_

e

ikx

+B

+

e

−ikx

x →−∞

A

+

e

ikx

x →+∞

ψ

2

≡ ψ

−

(x) →

_

A

−

e

−ikx

x →−∞

e

−ikx

+B

−

e

ikx

x →+∞

where A

+

is the transmission coeﬃcient giving the transmission probability

p = [A

+

[

2

. Conservation of the probablity current moreover implies [A

+

[

2

+

[B

+

[

2

= 1. The physical interpretation of the solution ψ

+

is that it describes

an incoming wave ψ

in

= e

+ikx

travelling from −∞ which is scattered at the

barrier into an outgoing wave ψ

out

with ψ

out

= B

+

e

−ikx

for x → −∞ and and

ψ

out

= A

+

e

+ikx

for x → +∞. Similarly for motion in the opposite direction

described by ψ

−

. We can compute the Wronskian at x → −∞: W = 2A

−

ik.

This gives:

K(E; x, y) =

m

ℏk

1

A

−

_

θ(x −y)ψ

+

(x)ψ

−

(y) +θ(y −x)ψ

−

(x)ψ

+

(y)

¸

(3.79)

Now, in the limit x →∞, y →−∞, we ﬁnd:

K(E; x →∞, y →−∞) =

m

ℏk

A

+

e

ik(x−y)

=

A

+

v

e

ik(x−y)

(3.80)

where v = k/m is the velocity.

Thus, K directly gives the transmission coeﬃcient: the propagator (as it

should!) directly gives the amplitude for propagation through the barrier. Using

the above deﬁnitions, K(E; x, y) is written, in the limit x → ∞, y → −∞, in

an intuitive and concise way: K(E, x, y) →

1

v

ψ

out

(x)ψ

∗

in

(y).

Similarly, one can consider reﬂexion by a barrier by letting y → −∞, x →

−∞, and x > y. Far away, where V (x) →0, we have:

K(E; x, y) ∼

1

v

(ψ

in

(x)

. ¸¸ .

direct

+ψ

out

(x)

. ¸¸ .

reﬂected

)ψ

∗

in

(y)

=

1

v

_

e

ik(x−y)

+B

+

e

−ikx−iky

_

(3.81)

We see in the two contributions in the above equation the relics of the two

possible classical paths: a direct path from y to x, and a path from y to x where

the particle is reﬂected by the barrier, as is illustrated in ﬁgure 3.1. This will

become more transparent in the semiclassical path integral approach.

3.2.2 Semiclassical computation of K(E)

Let us now turn to the computation of K(E; x, y) in the semiclassical approx-

imation. This is expected to be a good approximation when we can neglect

62 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

y x

V

Figure 3.1: The two possible classical paths from y to x. The ﬁrst one goes

directly from y to x, whereas the second one is ﬁrst reﬂected by the potential

barrier before going to x.

derivatives higher than cubic in V (x). Using the semiclassical result in eq. (3.44)

we must perform a Fourier transform

K

sc

(E; x

f

, x

i

) =

_

∞

0

¸

−

1

2πiℏ

∂

2

S

c

∂x

i

∂x

f

e

i

Sc

ℏ

+

iEt

ℏ

dt (3.82)

where the classical action S

c

in a function of the initial and ﬁnal positions x

i

and x

f

, and the travel time t.

For ℏ → 0, or equivalently for E large enough, we expect a saddle point

computation of the integral over t to be a reliable approximation. So we pro-

ceed with such a an evaluation. Notice that in doing so we are introducing an

additional approximation in computing K(E; x

f

, x

i

) with respect to the gaus-

sian computation of K(x

f

, t; x

i

, 0). In Problem 1 of Homework 7 the condition

for the validity of these approximations is worked out.

Let us brieﬂy recall how the saddle point approximation works.

4

We have

to compute the following integral:

K =

_

∞

0

_

f(t) e

i

g(t)

ℏ

dt (3.83)

In the limit where ℏ → 0, this integral will be dominated by the region

of the stationary point t

∗

where g

′

(t) = 0. In this region, we can substitute

f(t) ≈ f(t

∗

) and g(t) ≈ g(t

∗

) +

1

2

g

′′

(t

∗

)(t − t

∗

)

2

. We will get the following

integral:

K =

_

f(t

∗

)e

i

g(t∗)

ℏ

_

∞

0

e

i

1

2

g

′′

(t∗)

ℏ

(t−t∗)

2

dt (3.84)

We will then be able to change the integration variable by δt ≡ t − t

∗

, and

extend the lower bound of the integral from −t

∗

to −∞, thus getting an easily

integrable Gaussian Integral.

Let us therefore look for the stationary point for the integral 3.82:

∂S

c

∂t

¸

¸

¸

¸

t=t∗

+E = 0 ⇒ (3.85)

−E

c

(t

∗

) +E = 0 ⇒ t

∗

= t(E) (3.86)

where E

c

(t) is the energy of the classical trajectory for the motion from x

i

to

x

f

in a time t, and t

∗

is the stationary point. Note that the the stationary point

4

Notice that we have already been using this concept throughout the course (in particular

in sections 1.3.2 and 3.1). So the explanation given here should truly have been at the

beginning! Better late than never.

3.2. THE FIXED ENERGY PROPAGATOR 63

approximation has selected a classical trajectory with energy E

c

= E, showing

that we are on the right track to derive a semiclassical approximation to the

propagator.

We have that the exponent at the stationary point t

∗

is given by:

S

c

+E

c

t

∗

=

_

t∗

0

(L +E) dt =

_

t∗

0

m˙ x

2

c

dt =

_

x

f

xi

p(x)dx (3.87)

where we used S

c

=

_

(m˙ xdx − Edt) and where p(x) =

_

2m(E −V (x)) is the

momentum of the classical trajectory with energy E.

Let us now ﬁnd the second order expansion of S

c

around t

∗

diﬀerentiating

the identity

∂Sc

∂t

= −E

c

(t)

∂

2

S

c

∂t

2

¸

¸

¸

¸

t=t∗

= −

∂E

c

∂t

(3.88)

To explicitly write the result we need to use some simple mechanical identi-

ties. By using ˙ x

c

= v(x) ≡

_

2(E

c

−V (x))/m,we have

t =

_

x

f

xi

dx

˙ x

c

=

_

x

f

xi

_

m

2(E

c

−V (x))

dx

⇒1 = −

∂E

c

∂t

_

x

f

xi

1

m

_

m

2(E −V (x))

_3

2

dx (3.89)

⇒−

∂E

∂t

=

1

1

m

_

x

f

xi

dx

v

3

= −v

i

v

f

∂

2

S

c

∂x

i

∂x

f

(3.90)

where in the second step we diﬀerentiated with respect to t, and then set t = t

∗

,

E

c

= E and where, again, v

f

= v(x

f

) and v

i

= v(x

i

).

The stationary phase integral is therefore:

_

dδt exp

_

i

2ℏ

_

∂

2

S

c

∂t

2

_

(δt)

2

_

=

¸

2πiℏ

∂

2

Sc

∂t

2

=

¸

2πiℏ

1

m

_

dx

v

3

(3.91)

Putting this into eq. (3.82), the ﬁnal result is:

K(E; x

f

, x

i

) =

¸

1

v(x

f

)v(x

i

)

exp

_

i

_

x

f

xi

p(x)

ℏ

dx

_

∼ ψ

WKB

(x

f

)ψ

∗

WKB

(x

i

)

(3.92)

where we used the result eq. (3.73) to deduce the form of the ﬁxed energy wave

functions corresponding to our approximation

ψ

WKB

(x) ∼

¸

1

v(x)

exp

_

i

_

x

x0

p(x

′

)

ℏ

dx

′

_

. (3.93)

This is the wave function which is derived in quantum mechanics by applying

the Wentzel-Kramers-Brillouin (WKB) approximation. Problem 1 of Home-

work 7 is devoted to the standard derivation of the WKB approximation in the

Schroedinger approach. That standard method also allows a direct study of the

64 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

domain of validity of the WKB approximation. The result is that one necessary

requirement for WKB to be valid is

ℏ

p

2

(x)

dp

dx

≡

dλ(x)

dx

≪1 (3.94)

corresponding to a physical situation where the De Broglie wavelength λ varies

by a relatively small amount over a distance of order λ. Of course, keeping the

classical parameters of the problem ﬁxed, this requirement would be eventually

satisﬁed by formally taking ℏ → 0. This limit, where the wavelength goes

to zero, is the quantum mechanics analogue of geometric optics limit for the

propagation of electromagnetic waves. In that limit the propagation of light

can be described by the propagation of particles along rays: the wavelength

being small diﬀraction phenomena are negligible.

The above equation can also be written as a constraint on the force F =

−V

′

(x) acting on the particle

ℏ

p

2

(x)

¸

¸

dp

dx

¸

¸

= ℏ

¸

¸

mF

p

3

¸

¸

≪1 . (3.95)

We thus see that at the points where the classical momentum is very small

the WKB approximation breaks down. In particular this is so at the turning

points of the classical motion where E = V and p = 0. Now, in all interesting

applications the wave function at the turning points is not directly needed,

which is reassuring. However in practically all interesting applications, in order

to compute the propagator or the wave function at points where the WKB

approximation applies one still has to deal with matching conditions close to

the turning points where the approximation breaks down. In the path integral

formulation the issue arises in that one has to consider in the semiclassical

computation of the propagator trajectories that go through turning points. The

standard method to overcome the breakdown of the WKB approximation at

the turning points is to deform the trajectory into the complex plane for the

coordinate x (and for time t as well) in such a way as to avoid the point where

v(x) = 0.

3.2.3 Two applications: reﬂection and tunneling through

a barrier

In this section we shall consider two basic applications of the WKB propagator

formalism. We shall not worry about the subtleties of reﬂection points, and

compute whatever we can compute. Sections 3.2.4 and 3.2.5 will delve with

more details into those issues.

Let us ﬁrst examine the process of reﬂection by a potential barrier. This

will be studied in detail in the exercises. Let us assume for that that we have

a potential V (x) that vanishes as x → −∞, and that has a barrier somewhere

along the x axis. When we put x

i

, x

f

→ −∞, x

f

> x

i

, the classical action

from x

i

to x

f

accepts two solutions: one going directly from x

i

to x

f

, and one

reﬂected ﬁrst by the barrier at a and then going to x

f

, as shown in ﬁgure 3.2.

Then, the Green’s function will be decomposed into the contributions of the

two classical paths: K(E) = K

1

+ K

2

. The contribution of the reﬂected path,

3.2. THE FIXED ENERGY PROPAGATOR 65

x

i

x

f

V

a

Figure 3.2: The two possible classical paths from x

i

to x

f

. The ﬁrst one goes

directly from x

i

to x

f

, whereas the second one is going ﬁrst from x

i

to a, and

then is reﬂected from a to x

f

.

K

2

, is interesting because it contains information about the phase shift induced

by the reﬂexion. K

1

is the same as without barrier.

It turns out that:

K

2

(E; x

f

, x

i

) =

¸

1

v(x

i

)v(x

f

)

e

−i

π

2

exp

_

i

ℏ

_

x

f

xi

p(x)dx

(2)

_

(3.96)

where

_

x

f

xi

p(x)dx

(2)

=

_

a

xi

[p(x)[ dx +

_

a

x

f

[p(x)[ dx is the integral over the re-

ﬂected path. [p(x)[ is deﬁned as

_

2m(E −V (x)) > 0.

The origin of the extra −i factor is that det

_

−m∂

2

t

−V

′′

(x)

_

< 0 over the

reﬂected trajectory, or more precisely this operator has one negative eigenvalue.

Now, if we assume x > y, x, y →−∞, and remember eq. (3.74) we can write:

K(E; x, y) = K

1

+K

2

= ψ

out

(y)ψ

stat

(x)

=

¸

1

v(x)v(y)

_

e

i

ℏ

R

x

y

|p(z)|dz

+e

−i

π

2

e

i

ℏ

(

R

a

y

|p(z)|dz+

R

a

x

|p(z)|dz)

_

=

¸

1

v(x)v(y)

e

i

ℏ

R

a

y

|p(z)|dz−i

π

4

_

e

−

i

ℏ

R

a

x

|p(z)|dz+i

π

4

+e

i

ℏ

R

a

x

|p(z)|dz−i

π

4

_

=

¸

1

v(y)

e

i

ℏ

R

a

y

|p(z)|dz−i

π

4

¸

1

v(x)

cos

_

1

ℏ

_

a

x

[p(z)[ dz −

π

4

_

(3.97)

Thus, we have found the physical eigenstate ψ

stat

of the Schr¨odinger prob-

lem:

ψ

out

(y) =

¸

1

v(y)

e

i

ℏ

R

a

y

|p(z)|dz−i

π

4

ψ

stat

(x) =

¸

1

v(x)

cos

_

1

ℏ

_

a

x

[p(z)[ dz −

π

4

_

(3.98)

Imagine now, that the motion is bounded also to the “far” left at b ≪ a.

Applying the same argument, to a wave that is reﬂected at b we have the

following result for ψ

stat

ψ

stat(b)

(x) =

1

_

v(x)

cos

_

1

ℏ

_

x

b

[p(z)[ dz −

π

4

_

(3.99)

Now, the two stationary solutions must coincide up to the sign, as they

solve the same problem (energy bound eigenstate in one-dimensional Quantum

66 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

Mechanics). We have:

ψ

a

(x) = cos

_

1

ℏ

_

a

x

[p(z)[ dz −

π

4

_

= cos (φ

a

(x))

ψ

b

(x) = cos

_

1

ℏ

_

x

b

[p(z)[ dz −

π

4

_

= cos (φ

b

(x)) (3.100)

Taking the derivatives, we ﬁnd:

∂

x

φ

a

= −∂

x

φ

b

⇒ φ

a

(x) +φ

b

(x) = θ (3.101)

where θ is independent of x.

We then get:

ψ

a

(x) = cos (φ

a

(x))

ψ

b

(x) = cos (θ −φ

a

(x)) = cos θ cos φ

a

−sin θ sinφ

a

⇒θ = mπ, m ∈ Z

⇒ψ

a

(x) = ±ψ

b

(x) (3.102)

This condition gives a constraint for the integral over the closed contour:

_

p(z)dz =

2

ℏ

_

a

b

[p(z)[ dz =

2

ℏ

_

x

b

[p(z)[ dz +

2

ℏ

_

a

x

[p(z)[ dz = 2π

_

m+

1

2

_

(3.103)

This is the famous Bohr-Sommerfeld semi-classical quantization condition.

Barrier penetration: semiclassical transmission coeﬃcient

Imagine now that the potential V (x) has a barrier so that a classical particle

with an energy E cannot go from −∞ to +∞ (see ﬁgure 3.3).

x

i a

b

x

f

E

V (x)

Figure 3.3: There is no classical trajectory going from x

i

to x

f

with an energy

E. The potential barrier is too high to let a particle go past it. Quantum

physics in the semi classical limit predict a non-vanishing probability for such a

particle to tunnel through the barrier from x

i

to x

f

.

However, we can deﬁne a semiclassical contribution to this process, which

can be seen in the WKB approximation. The equivalence between WKB and

Gaussian approximation requires therefore to ﬁnd the corresponding contribu-

tion in the path integral approach. Thus, we have to look for the stationary

3.2. THE FIXED ENERGY PROPAGATOR 67

point of the combined

_

D[x]dt integral:

K(E; x

f

, x

i

) =

_

∞

0

_

D[x]e

i

ℏ

(Sc+Et)

dt (3.104)

The result is that we can ﬁnd a stationary point provided t is complex!

Indeed, from the classical equation of motion, we ﬁnd that:

dx

dt

=

_

2(E −V (x))

m

(3.105)

Considering extending this equation at a < x < b, we have:

dx

dt

= +i

_

2 [E −V (x)[

m

(3.106)

where we chose the plus sign in order to ensure that the Green’s function remain

analytic.

Integrating over the full trajectory, we get:

t =

_

dt =

_

x

f

xi

_

m

2(E −V (x))

. ¸¸ .

v(x)

dx

=

_

a

xi

dx

v(x)

. ¸¸ .

real

−i

_

b

a

dx

[v(x)[

. ¸¸ .

imaginary

+

_

x

f

b

dx

v(x)

. ¸¸ .

real

(3.107)

The evolution of time in the complex plane along the trajectory is depicted

in ﬁgure 3.4.

Re(t)

Im(t)

−

_

b

a

dx

|v(x)|

Figure 3.4: The evolution of the time along the trajectory in the complex plane.

It acquires a negative imarinary part as the particle travels through the potential

barrier.

Now, regardless of the course of the complex value of t, we still have:

i

ℏ

(S

c

(t) +Et)

¸

¸

¸

¸

stat

=

i

ℏ

_

x

f

xi

_

2m(E −V (x)dx (3.108)

The Jacobian factor from the gaussian integration around the extremum of

68 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

t still compensates the determinant:

K(E; x

f

, x

i

) =

¸

1

v(x

f

)v(x

i

)

exp

_

i

ℏ

_

x

f

xi

_

2m(E − V (x)dx

_

=

¸

1

v(x

i

)

e

i

ℏ

R

a

x

i

|p(x)|dx

. ¸¸ .

ψ

∗

in

¸

1

v(x

f

)

e

i

ℏ

R x

f

b

|p(x)|dx

A+

¸ .. ¸

e

−

1

ℏ

R

b

a

|p(x)|dx

. ¸¸ .

ψout

(3.109)

where the transmission coeﬃcient is given by A

+

.

We can then derive the probability P of tunneling:

P = [A

+

[

2

= e

−

2

ℏ

R

b

a

|p(x)|dx

= e

−

2

ℏ

R

b

a

√

2m(V (x)−E)dx

(3.110)

3.2.4 On the phase of the prefactor of K(x

f

, t

f

; x

i

, t

i

)

We have so far been rather cavalier concerning the phase of our Gaussian in-

tegrals. However, as our previous discussion makes clear, when multiple tra-

jectories contribute to the amplitude it is essential to be in control of their

relative phases. In this subsection, we will give the basic idea for the case of

K(x

f

, t

f

; x

i

, t

i

). The next subsection is devoted to the phase of K(E; x

f

, x

i

),

which is very relevant to the discussion several applications like the computation

of energy levels in the WKB approximation (see Homework 7).

1) Getting the basic idea

In order to get an idea, let us go back to the determinant prefactor for the

harmonic oscillator which was computed in eq. (1.98). In that derivation we did

not pay any real attention to the phase factors that arose when some eigenvalues

became negative. Eq. (1.89) tells us that the number n

−

of negative eigenvalues

equals the integer part of ωT/π (T ≡ t

f

− t

i

). Thus for 0 < T < π/ω we have

n

−

= 0 so that the phase of the prefactor is the same as for the free particle

case. The speciﬁc phase of eq. (1.98) is only correct for this range of T. By

eq. (1.58), for n

−

,= 0 the phase of the harmonic oscillator prefactor relative

to that of the free particle is exp(−in

−

π

2

). Therefore the correct result for the

propagator prefactor is given by

J

ω

(T) = e

−in−

π

2

_

mω

2πiℏ[ sin(ωT)[

(3.111)

rather than by eq. (1.98). Thus when T increases by a half-period π/ω the

propagator aquires an extra phase factor = −i. With an extra full period 2π/ω

the propagator just ﬂips sign: (−i)

2

= −1. This result can also be stated as

1. For trajectories in which the velocity does not ﬂip sign there is no extra

phase factor

2. For trajectories where the velocity ﬂips sign n times the phase is either

e

inπ/2

or e

i(n−1)π/2

(depending on how many multiples of the half period

ﬁt into T)

3.2. THE FIXED ENERGY PROPAGATOR 69

3. For a trajectory corresponding to exactly k periods of motion the phase

factor is (−1)

k

This result is valid for any periodic motion not just for the harmonic oscillator.

There is a well developed mathematical theory, called Morse theory, underlying

the dependence of the number of negative eigenvalues n

−

on the properties

of the trajectory. n

−

is referred to as the Morse index. We are not going to

linger on this mathematical connection though. A simple paper with some more

details is for instance ref. [9].

2) An explicit non trivial example

5

x

i

x

f

x = 0

V (x) = θ(x)

mω

2

2

x

2

Figure 3.5: .

Let us see now a slightly less straightforward case by considering the motion

in the potential V (x) = θ(x)

1

2

mω

2

x

2

shown in ﬁg.3.5. The novelty with respect

to the harmonic oscillator is that the motion at x < 0 is not bounded, so

that motion is not periodic. Let us compute the ﬂuctuation determinant using

eqs. (3.15, 3.16). We also want to do so by focusing on a 1-dimensional family

of trajectories characterized by initial position x(0) = x

i

< 0 and velocity

˙ x(0) ≡ v > 0 at t = 0. The classical motion is characterized by 3 time, t,

intervals

1. For 0 ≤ t ≤ −

xi

v

≡ t

1

we have x(t) = x

i

+ vt ≤ 0: the particle moves in

the V (x) = 0 region with positive velocity.

2. For t

1

< t < t

1

+

π

ω

we have x(t) =

v

ω

sin[ω(t −t

1

)]: the particle moves in

the harmonic potential at x > 0.

3. For t ≥ t

1

+

π

ω

the trajectory is again “free”, x(t) = −v(t − t

1

+

π

ω

) < 0

but the velocity is now negative.

We want to compute the determinant function ψ

0

(t

f

) as t

f

is varied in the above

3 intervals. The operator controlling the quadratic ﬂuctuation is

−m

d

2

dt

2

− V

′′

(x(t)) = −m

d

2

dt

2

−mω

2

θ(t −t

1

)θ(t

1

+

π

ω

−t) (3.112)

where the product of step functions accounts for the fact that V ,= 0 only for

t

1

< t < t

1

+

π

ω

. The initial value problem in eqs. (3.15, 3.16) for the diﬀerential

5

This problem was proposed in one of the exercise sessions.

70 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

operator 3.112 is solved by patching the solution in the 3 time regions listed

above. In other words, ﬁrst the solution in region 1 is found. Second, the solution

in region 2 is found with initial conditions at t = t

1

dictated by continuity of ψ

0

and

˙

ψ

0

. Third this procedure is repeated to get the solution in region 3. One

ﬁnds respectively in region 1, 2 and 3

1. For 0 ≤ t ≤ −

xi

v

≡ t

1

the solution is ψ

0

(t) = t

2. For t

1

< t < t

1

+

π

ω

one has: ψ

0

(t) =

_

1

ω

2

+t

2

1

sin[ω(t − t

1

) + ϕ] where

0 ≤ ϕ ≤ π/2 is deﬁned by tanϕ = ωt

1

. Notice that ψ

0

(t

1

+ π/ω) =

−ψ

0

(t

1

) = −t

1

< 0.

3. For t ≥ t

1

+

π

ω

one has ﬁnally ψ

0

(t) = −t +

π

ω

< 0

Let us now increase t

f

starting from t

f

< t

1

. By the above result we have that

the determinant ψ

0

(t

f

) is positive and grows with t

f

in region 1. Indeed not

only is the determinant positive in region 1, but all eigenvalues are positive.

This is because for t

f

< t

1

eq. (3.112) reduces to the free particle quadratic

ﬂuctuation which is a positive deﬁnite diﬀerential operator, as we already know

very well. At some t

f

= t

∗

in region 2, t

1

< t

∗

< t

1

+ π/ω, the determinant

crosses zero once and becomes negative. In region 3, ψ

0

(t

f

) decreases and thus

stays negative without ever reaching zero again. The picture of what happens

is then obvious:

• For t

f

< t

∗

the eigenvalues are all positive 0 < λ

1

< λ

2

< λ

3

< . . .

• At t

f

= t

∗

the smallest eigenvalue λ

1

vanishes, and for any t

f

> t

∗

the

spectrum satisﬁes λ

1

< 0 < λ

2

< λ

3

< . . . : there is one and only one

negative eigenvalue.

For t

f

> t

∗

the gaussian integral prefactor is then (compare to eq. (3.28))

I[x

c

] = e

−i

π

2

_

m

2πiℏ[ψ

0

(t

f

)[

(3.113)

Another comment concerns the point at which the determinant ﬂips sign.

Notice that this happens at t

∗

strictly larger than the time t

r

= t

1

+

π

2ω

at

which reﬂection by the potential barrier happens. That is to say: ψ

0

> 0

before reﬂection. This is a general result that can be easily seen by using the

integral formula in eq. (3.32). From that equation is it obvious that ψ

0

> 0

for trajectories in which the velocity does not ﬂip sign. Let us show that a

stronger result holds, namely lim

t

f

→tr

ψ

0

(t

f

) > 0 when t

f

approaches t

r

from

below. For t →t

r

the velocity goes to zero linearly: v(t) ≃ a(−t +t

r

), with a a

positive coeﬃcient (so that the acceleration is ˙ v = −a < 0). Then the integral

in eq. (3.32) diverges but its prefator v

f

v

i

goes to zero so that the result is ﬁnite

lim

t

f

→tr

_

v(t

f

)v(t

i

)

_

t

f

ti

1

v

2

c

(t)

dt

_

= (3.114)

lim

t

f

→tr

v(t

i

)a(−t

f

+t

r

)

_

1

a

2

1

−t

f

+t

r

+ ﬁnite

_

=

v(t

i

)

a

> 0(3.115)

and therefore the determinant can only ﬂip sign after reﬂection.

3.2. THE FIXED ENERGY PROPAGATOR 71

re

t

R

−ǫ t

R

+ǫ

C

ǫ

t

R

Figure 3.6: .

3) The integral expression for ψ

0

and reﬂecting trajectories

One ﬁnal remark concerns the way to make sense of eq. (3.32) and eq. (3.34)

in the case of trajectories where v does change sign. These equations are ill

deﬁned in such cases: the factor v

f

v

i

is non-zero, but the integral diverges at

the points where v →0. This is because in deriving them we had to divide by v

somewhere. Notice however that no singularity should exist in ψ

0

is association

to the vanishing of the velocity. This is because ψ

0

is just given by the solution

of an ordinary diﬀerential equation with ﬁnite coeﬃcients. This suggest that

be should be able to ‘save’ eq. (3.32) and eq. (3.34). The way to do so is very

simple: we just need to slightly deform the trajectory into the complex (x, t)

plane in such a way as to avoid the singularities. When working with the time

integral eq. (3.32), by eq. (3.115) we see that the integrand has a double pole:

we just need to go around it as shown in ﬁg. 3.6. Notice also that it does not

make any diﬀerence to go around the pole from below, as in the ﬁgure, or from

above. This is because the singularity is a double pole and thus the integral over

a full oriented circle around it gives zero by Cauchy’s theorem. The integral on

the oriented half circle contour C

ǫ

is easily computed for ǫ →0 by parametrizing

t −t

r

= −ǫe

iθ

, with 0 ≤ θ ≤ π on C

ǫ

, so that dt = −iǫe

iθ

dθ. We have

lim

ǫ→0

_

Cǫ

dt

v

2

(t)

=

_

Cǫ

dt

a

2

(t −t

r

)

2

= (3.116)

1

ǫa

2

_

π

0

−ie

−iθ

dθ =

1

ǫa

2

[−1 −1] . (3.117)

This result precisely cancels the diverge of the integrals at t < t

r

−ǫ and t > t

r

+ǫ

so that we have in the end

ψ

0

= v

i

v

f

lim

ǫ→0

__

−

1

ǫa

2

+

_

tr−ǫ

ti

dt

v

2

(t)

_

+

_

−

1

ǫa

2

+

_

t

f

tr+ǫ

dt

v

2

(t)

__

(3.118)

We can also change variable t → x. The reﬂection point x

r

is deﬁned by

v(x

r

) =

_

2(E −V (x

r

))/m = 0, where E is the energy of the trajectory in

consideration. The equation of motion gives V

′

(x

r

)/m = a. Then we have that

t

r

±ǫ both correspond to x

r

−

a

2

ǫ

2

so that the above equation becomes

ψ

0

= v

i

v

f

__

−

1

ǫa

2

+

_

xr−

a

2

ǫ

2

xi

dx

[v

3

(x)[

_

+

_

−

1

ǫa

2

+

_

xr−

a

2

ǫ

2

x

f

dx

[v

3

(x)[

__

(3.119)

Using this formula, we can study the conditions on V (x) under which ψ

0

ﬂips

sign after reﬂection. Since for a reﬂecting trajectory v

i

v

f

< 0, we have that ψ

0

is negative if and only if the expression in curly bracket is positive. Now, if the

potential at x →−∞, i.e. in the classically accessible region, is ﬂat enough, then

72 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

it is obvious that, for x

i

or x

f

large enough and negative, the expression in curly

brakets is dominated by the positive integral in the region away from the turning

point. Thus we have ψ

0

< 0. However if lim

x→−∞

V (x) = −∞ fast enough,

then

_

dx/v

3

is dominated by the region close to the turning point. In this case

the expression in curly braket does not necessarily become positive for large

enough x

i

and x

f

and ψ

0

may never ﬂip sign. A simple and obvious example

where this is so is given by the linear potential V = cx. This is obvious because

in this case V

′′

= 0 and we have the free particle result ψ

0

(t

f

) = t

f

−t

i

> 0.

It is a good exercise to recover the result we found for ψ

0

in the potential in

ﬁg. 3.5 by using instead the integral representation of ψ

0

. We shall not do that

but just give some suggestions on how to try and do it. Using the space integral

representation eq. (3.34), one notices that v(x) has a cut in the complex x plane

along the x region accessible to classical motion. Physically this corresponds

to the velocity being double valued: v(x) = ±

_

2(E −V (x))/m. Then ψ

0

is

computed on a trajectory that goes around this cut...The willing student can

now go ahead and make a simple remark (on Cauchy integrals) that allows for

a direct computation.

3.2.5 On the phase of the prefactor of K(E; x

f

, x

i

)

Let us now focus on the phase of K(E; x

f

, x

i

), which is really what matters

in most applications. Again the delicate issue is how to deal with the turning

points. Two basic remarks are in order here.

The ﬁrst is that for x

i

or x

f

close to the turning points, the WKB approx-

imation to K(E; x

f

, x

i

) breaks down

6

because p(x) → 0 (see eq. (3.95)). The

saddle point approximation can however still be reliable if the trajectory can

be deformed into the complex C

2

plane for (t, x) in such a way that p(x) never

really crosses zero. In order for this procedure to makes sense two conditions

must be satisﬁed (see the discussion in the chapter on the WKB approximation

in Landau-Lifshitz [10])

• The potential V (x) is analytic in a region around the turning point. This

way, v(x) =

_

2(E −V (x))/m is also analytic in this region apart from

having a cut (due to the square root) in the physically accessible region

on the real x axis.

• The region of analyticity for V (x) overlaps with the region where the WKB

approximation applies, that is with the region where ℏ[mV

′

(x)/p(x)

3

[ ≪

1.

When the above two conditions are satisﬁed, a suitable deformation of the tra-

jectory can be chosen such that x goes around the turning points in such a

way that: 1) Physical quantities, like the action integral, are unaﬀected by the

deformation of the trajectory thanks to analyticity. 2) p(x) is never too small

6

We stress that the situation is diﬀerent with respect to the apparent singularity of ψ

0

(t

f

)

at the turning points which we encountered in the previous section. The gaussian approxima-

tion for K(x

f

, t

f

; x

i

, t

i

) does not really break down in the presence of turning points, witness

the fact that this approximation is exact for the harmonic oscillator. Instead the WKB com-

putation of K(E; x

f

, x

i

), as we already pointed out, requires an extra approximation, which

genuinely breaks down at the turning points.

3.2. THE FIXED ENERGY PROPAGATOR 73

and the WKB is manifestly valid.

7

We will explain with an example below how the deformation works. Notice,

for the moment, that by deforming the time line into the complex plane (like we

did for instance with barrier penetration), the trajectory x(t) which is deﬁned

to solve

m

d

2

x

dt

2

= −V

′

(x) , (3.120)

is also, in general, deformed into the complex x plane.

The second remark concerns the form of the prefactor itself. By continuing

the trajectory into the complex plane we can still make sense of the integral

_

dx/v

3

even when the velocity ﬂips sign somewhere. Then, as we already

know, all is left in the prefactor is the velocity at the initial and ﬁnal points

P(x

f

, x

i

) =

¸

−1

2πiℏ

∂

2

S

c

∂x

f

∂x

i

¸

2πiℏ

∂

2

Sc

∂t

2

=

¸

1

v(x

f

)v(x

i

)

. (3.121)

Since v(x

f

) changes sign at the turning point, the above prefactor gains a 1/

√

−1

factor precisely at reﬂection. The only question is that 1/

√

−1 being double

valued, we cannot yet tell, simply by staring at eq. (3.121), if the phase is +i or

−i. Also, just by staring at the above equation, it is not obvious how to proceed

for trajectories with multiple reﬂections. It turns out that the right answer is:

• At each reﬂection the prefactor gains a phase factor (−i) = e

−iπ/2

.

This result can be directly and easily checked for the example we discussed in

section 3.2.4, using the computations we have already made. There are two

contributions to the phase. One is the phase of the VanVleck-Pauli-Morette

prefactor in eq. (3.113). We computed this phase factor in section 3.2.4 and

found that it is e

−iπ/2

for x

f

and x

i

negative. The second comes from the phase

of the gaussian integral over t and is determined by the sign of ∂

2

S/∂t

2

. It is

easy to see that ∂

2

S/∂t

2

> 0. Indeed by eqs. (3.88) and (3.90) we have that

∂

2

S/∂t

2

is just the inverse of the expression in curly brackets in eq. (3.119). In

the case at hand that expression is clearly positive since, v

i

v

f

< 0 and since

ψ

0

(t

f

) < 0 by our direct solution of the diﬀerential equation. Since ∂

2

S/∂t

2

is

positive, like for a non-reﬂecting trajectory, the phase of the t integral is trivial

and the only non trivial contribution is that from the determinant, which equals

e

−iπ/2

. This is in agreement with the general result we stated above

There is one quick way to deduce in general that for v

i

v

f

< 0 the phase

in eq. (3.121) must be e

−iπ/2

. Without loss of generality and to simplify the

discussion let us consider a particle of mass m = 1, let us shift the deﬁnition of

the energy such that E = 0 for our trajectory, let us shift x so that the turning

point is x

r

= 0 and let us choose V (x) ≃ x in the vicinity of x = 0. The

equation of motion close to x = 0 is just

¨ x = −1 (3.122)

where again, without loss of gererality, we can focus on the solution x = −

1

2

t

2

with v ≡ ˙ x = −t. For real t the velocity ﬂips sign at t = 0. If we deform t into

7

Notice that the time contour shown in ﬁg.3.6 are an example of how to go around turning

points. To be manifestly in the WKB approximation ǫ should be large enough. However

because of analyticity the result of all integrals is independent of ǫ anyway!

74 CHAPTER 3. THE SEMICLASSICAL APPROXIMATION

t = 0

t = 0

−iǫ

iǫ

(a)

(b)

Figure 3.7: .

the complex plane we have two options to go around the point t = 0 where the

velocity vanishes. Either we go below like in ﬁg. 3.7a or above like in ﬁg. 3.7b.

These two options lead to diﬀerent (and conjugate) phase shifts for 1/

_

v(t

f

).

Which of the two deformations should we choose? It turns out deformation (a)

is acceptable while (b) is not. To understand why, we should go back and think

what it means to deform t into complex values in our original deﬁnition of the

path integral. It means that, rather than doing our simple (and real) slicing

dt = (t

f

− t

i

)/N, we are making some element dt of the sum

_

dt = t

f

− t

i

complex. As long as we write U(t

f

− t

i

) as a product of matrices this is not a

problem. A constraint arises however when we attempt to write this product of

matrices in terms of a path integral as done in section 1.1.4. This is because if

Re (−idt) > 0 in the exponent e

(−iHdt/ℏ)

, then the p integral in eq. (1.14) is not

convergent: it is growing at p →∞ (the ǫ interval of that equation corresponds

to our dt). This is just an inﬁnitesimal way of saying that the euclidean path

integral exists only for β > 0. Since our semiclassical approximation is based

on a path integral representation of the evolution amplitude, we should only

consider trajectories (real or complex) for which the path integral representation

makes sense. That is to say only trajectories such that Im(dt) ≤ 0. This the

case for trajectory (a) and not the case for trajectory (b). We are now just

left with computing the phase shift in 1/

_

v(t) = 1/

√

−t along trajectory (a).

When t goes from −∞ to +∞− iǫ the argument in

√

−t goes from +∞ to

−∞+ iǫ by passing in the upper half plane, therefore the phase of

√

−∞+iǫ

is e

iπ/2

. The resulting phase of the prefactor is then

1

_

v(−∞+iǫ)

= e

−iπ/2

1

_

[v(−∞)[

(3.123)

More formally we can compute the phase shift ∆ϕ by integrating over the t

contour

∆ϕ ≡ Im∆

_

ln(1/

√

v)

¸

= −Im

_

dv

2v

= −Im

_

+∞−iǫ

−∞

dt

2t

= −

1

2

π = −

π

2

(3.124)

Chapter 4

Instantons

4.1 Introduction

In Quantum Mechanics, perturbation theory can miss important features of a

physical system.

To illustrate this, let us give a quick look at a simple example: a one-

dimensional particle in a potential V (x) =

mω

2

x

2

2

+λx

3

.

The point x = 0 is a classical minimum and it seems reasonable to study the

dynamics around this point upon quantization. For λ “small” enough, we can

apply perturbation theory and in principle compute the ground state energy to

any desired order in λ. In doing so, however, we would not realize, or get any

direct indication of one basic pathology of this ground state: its instability! Due

to the possibility of tunneling through the barrier, the ground state will be able

to decay down to the region where λx

3

is negative and dominates the harmonic

term (see ﬁgure 4.1).

V (x)

E

tunnel

Figure 4.1: The stability of the ground state localized around the local minimum

of the potential is broken by the possibility for the particle to tunnel through

the barrier to the part of the potential unbounded from below.

The probability of tunneling will be proportional to: P ∼ exp

_

−2

_

√

V m

ℏ

dx

_

∼

exp

_

−

m

3

ω

5

ℏλ

2

_

. It thus represents an eﬀect which vanishes, as λ →0, faster that

75

76 CHAPTER 4. INSTANTONS

any power of λ, implying that it cannot be seen at any ﬁnite order in perturba-

tion theory. This is what theorists normally call a non-perturbative eﬀect.

As a second example, let us look at the case of a particle in a double well

potential: V (x) = λ(x

2

− a

2

)

2

symmetric under reﬂection x → −x (see ﬁgure

4.2).

V (x)

−a

a

Figure 4.2: The double well potential has two relexion symmetric vacua, thus

leading to inexistant degenerate eigenstates in perturbation theory.

This potential has two classical ground states: x = ±a, which are equivalent

because of the reﬂection symmetry of the system. In perturbation theory, we

can ignore this doubling and expand around one of these minima, say x = a.

We introduce thus the new variable y = x − a. In this variable, V (y) is a

harmonic oscillator with extra cubic and quartic terms which we can treat as

perturbations as long as λ is “small”. We have:

V (y) = λ(y + 2a)

2

y

2

= 4λa

2

y

2

+ 4λay

3

+λy

4

(4.1)

≡

1

2

mω

2

y

2

+ω

√

2mλy

3

+λy

4

(4.2)

where in the last equation we traded a for the harmonic oscillator frequency ω.

For small λ, we can now compute the corrections to the lowest energy levels to

arbitrary orders in λ without having any direct evidence that there is another

minimum with its own localized levels!

The situation is here symmetric: a guy sitting at the other classical ground

state would compute in perturbation theory exactly the same levels

1

, but cor-

responding to a diﬀerent set of states where the wave function is localized near

x = −a instead of x = a. Thus, in perturbation theory, the energy levels, and

in particular the ground state, are all doubly degenerate to all orders.

However, from the exact quantum mechanical solution of the problem, we

know that for a unidimensional system, there is no degeneracy of the bound state

energy eigenvalues. Moreover, we know that the ground state wave function has

no nodes, and, in the case at hand, is therefore even under x →−x:

Ψ

0

(x) = Ψ

0

(−x)

1

Technically, this other guy expands the coordinate as x = −a −y

′

. Because of reﬂection

symmetry the potential in y

′

satisﬁes V (−a −y

′

) = V (a + y

′

): exactly the form in eq. (4.1).

Hence the levels, in perturbation theory, are the same.

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 77

On the other hand, the ﬁrst excited state has one node, and must therefore

be odd under reﬂection:

Ψ

1

(x) = −Ψ

1

(−x)

However, because of the doubling of all the levels that we just deduced with

our perturbative reasoning, the energy eigenvalues of these two lowest lying

states mus be degenate to all orders in perturbation theory. That is E

1

−E

0

= 0

to all orders in perturbation theory. The goal of this chapter is to show that

the leading contribution to E

1

−E

0

as λ →0 can be computed by semiclassical

methods. More precisely, via semiclassical trajectories in imaginary time called

instantons.

4.2 Instantons in the double well potential

What we want to study are the properties of the ground state(s) of the double

well potential: V (x) =

λ

4!

_

x

2

−a

2

_

2

. To do this, let us go back to the Euclidean

path integral.

We will compute ¸a[ e

−

βH

ℏ

[a) and ¸−a[ e

−

βH

ℏ

[a) for β →∞. If the barrier is

high and the the potential smooth, we can perform a semiclassical approximation

in Euclidean time.

To compute those two transition amplitudes, we must sum over all possible

stationary trajectories (i.e. with stationary action) in Euclidean time.

Remember that:

¸x

f

[ e

−

βH

ℏ

[x

i

) = K

E

(x

f

, x

i

; β) =

_

x

f

,

β

2

xi,−

β

2

D[x]e

−

S

E

[x]

ℏ

(4.3)

S

E

[x] =

_ β

2

−

β

2

dτ

_

m˙ x

2

2

+

λ

4!

_

x

2

−a

2

_

2

_

(4.4)

For each trajectory ¯ x, we can compute its contribution to K

E

in Gaussian

approximation:

K

E

=

˜

^

_

det

_

−m

d

2

dτ

2

+V

′′

(¯ x)

__

−

1

2

e

−

S

E

[¯ x]

ℏ

(1 +O(ℏ)) (4.5)

Notice that S

E

is positive deﬁnite. Now,

K

E

=

_

D[x]e

−

S

E

[x]

ℏ

(4.6)

is dominated by the local minima of S

E

[x] with given boundary conditions:

x

f

= x

_

β

2

_

x

i

= x

_

−

β

2

_

(4.7)

Imposing now that

δSE

δx

= 0 leads to the Euclidean equation of motion.

Notice that the solutions to this equation describe motion in the euclidean time

78 CHAPTER 4. INSTANTONS

V

E

(x)

−a

a

Figure 4.3: The Euclidean equation of motion corresponds to the real time

equation of motion with a reversed potential V

E

(x) = −V (x).

parameter τ of a classical system with reversed potential V

E

(x) = −V (x) (see

ﬁgure 4.3).

In the limit ℏ → 0, we therefore expect K

E

to be dominated by the local

minima of S

E

(if there exist more than one).

K

E

≈ K

1

+K

2

+K

3

+. . . (4.8)

Remark: we can make an analogy with the ordinary integrals. Suppose that

we have a function S(x), as depicted in ﬁgure 4.4.

S(x)

x

1 2 3

Figure 4.4: An ordinary function of one variable S(x), accepting three local

minima called 1, 2 and 3.

Then, in the limit ℏ →0, the integral

I =

_

∞

−∞

e

−

S(x)

ℏ

dx (4.9)

can be approximated by

I ≈ I

1

+I

2

+I

3

+. . .

I

i

= e

−

S(x

i

)

ℏ

¸

2πℏ

S

′′

(x

i

)

(4.10)

The anharmonic terms around the local minima give corrections to each

I

i

of relative size ℏ.

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 79

To compute the transition amplitudes ¸a[ e

−

βH

ℏ

[a), ¸−a[ e

−

βH

ℏ

[−a), ¸−a[ e

−

βH

ℏ

[a)

and ¸a[ e

−

βH

ℏ

[−a), we must study the classical trajectories in the upside down

potential V

E

(x) = −V (x), depicted in ﬁgure 4.3.

This potential accepts two types of solutions:

Trivial solutions: x(τ) ≡ ±a which will give the leading contribution to the

ﬁrst two amplitudes.

Less trivial solutions: the instanton, which starts at x = −a at τ = −∞and

rolls towards x = a at τ = ∞, and the anti-instanton, which rolls from a

to −a. These solutions are exact in the β →∞ limit.

The action on the instanton can be computed: as the initial velocity ˙ x(−∞)

is zero, then the Euclidean energy E

E

is zero, and we can write:

m˙ x

2

2

−V (x) = E

E

= 0 (4.11)

This allows us to write an expression for the Euclidean action:

m˙ x

2

2

= V (x) ⇒ ˙ x =

_

2V (x)

m

S

E

=

_

∞

−∞

m˙ x

2

dτ =

_

a

−a

m

_

2V (x)

m

dx =

_

a

−a

_

2mV (x)dx (4.12)

Thus, in the semiclassical limit, the transition amplitude will be proportional

to:

¸a[ e

−

βH

ℏ

[−a) ∼ e

−

S

E

ℏ

∼ e

−

R

a

−a

p(x)

ℏ

dx

≪1 (4.13)

We recover in this result the expected semiclassical tunneling coeﬃcient.

For the speciﬁc example of a quartic potential V (x) =

λ

4!

_

x

2

−a

2

_

2

, the

zero energy solution satisﬁes

˙ x = ±

ω

2a

_

x

2

−a

2

_

(4.14)

where we used the new variable ω

2

=

λa

2

3m

.

The solutions of these equations are the instanton and anti-instanton given

by

x(τ) = ±a tanh

_

ω

2

(τ −τ

0

)

_

(4.15)

The instanton (with the plus sign) is depicted in ﬁgure 4.5.

Remarks

• For the instanton solution, the value of the exponent

SE

ℏ

is given by:

S

E

ℏ

=

_

a

−a

_

2mV (x)

ℏ

dx =

2

3

_

λm

3

a

3

ℏ

=

2

3

3/2

√

λm

_

3m

λ

_

3/2

ω

3

ℏ

= 2

m

2

ω

3

ℏλ

=

2

¯

λ

(4.16)

where

¯

λ =

ℏλ

m

2

ω

3

is the dimensionless parameter that characterizes the

size of the perturbative corrections to the harmonic oscillator value for

80 CHAPTER 4. INSTANTONS

x(τ)

τ

τ

0

−a

a

Figure 4.5: The instanton solution in the double well potential, running from

−a to a.

the energy levels. We already encountered

¯

λ when we studied perturba-

tive corrections to the ground state energy of the harmonic oscillator with

a λx

4

perturbation. Since Feynman diagrams with k-loops contribute a

correction of order

¯

λ

k−1

, the coupling

¯

λ is often called a loop-expansion

parameter.

The part of the Euclidean propagator due to the instantons is therefore

proportional to e

−SE/ℏ

= e

−2/

¯

λ

. Thus, eﬀects associated with instantons

are non-perturbative in

¯

λ.

• At [τ −τ

0

[

1

ω

, we have that

¸

¸

x

2

(τ) −a

2

¸

¸

≈ e

−ω|τ−τ0|

≪ 1. Instantons

are thus well localized in Euclidean time: they exist for a short instant

∆τ ∼

1

ω

in Euclidean time (hence their name).

Thus, when β → ∞, up to exponentially small terms, we have x ≈ ±a

away from τ = τ

0

. This is a crucial property considering that we are

interested in the limit β → ∞: for β ≫

1

ω

the instanton solution looks

like a step function (see ﬁgure 4.6).

−a

a

τ

0

−

β

2

β

2

Figure 4.6: In the large β limit, the instanton solution looks like a step function.

After having jumped from −a to a, there is still plenty of time to bounce

back!

More technically, given the instanton solution x

I

(τ) = af(τ − τ

0

), where

[f(τ − τ

0

)[ = 1 −e

−ω|τ−τ0|

for [τ −τ

0

[ ≫

1

ω

, then

x

IA

(τ) = af(τ −τ

0

)f(τ

1

−τ) (4.17)

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 81

x

IA

(τ)

τ

τ

0

−a

a

τ

1

Figure 4.7: The instanton-anti-instanton quasisolution.

for τ

1

≫τ

0

is an approximate solution representing an instanton followed

by an anti-instanton. For such a conﬁguration, one can easily check that

δS

E

δx

¸

¸

¸

¸

x=xIA

∼ e

−ω|τ0−τ1|

. (4.18)

As long as we only need to consider situations where ω[τ

0

−τ

1

[ is so large

that the above exponential is smaller than any eﬀect we are trying to

compute, then we should by all means consider the conﬁguration x

IA

as

a stationary point of the action. And thus add the contribution of these

quasi-solutions to our semiclassical approximation to the path integral.

Indeed, we shall ﬁnd at the end that [τ

1

−τ

0

[ ∼ e

S

E

ℏ

. Thus, neglecting

that x

IA

is not an exact solution amounts to neglecting terms of order

O

_

e

−ω exp(SE/ℏ)

_

! These terms do not exactly vanish but are indeed very

small. Thus we must add the contribution of x

IA

to our saddle point

estimate of the path integral.

• By extending this argument, we are easily convinced that we must consider

all the approximate solutions with a number N of instantons + anti-

instantons. Notice that if we take n

I

instantons and n

A

anti-instantons,

we must have that n

I

−n

A

= 0 for a → a, and n

I

− n

A

= 1 for −a → a.

(what about −a →−a and a →−a ?)

An approximate solution with, e.g. three instantons at τ

1

, τ

3

and τ

5

and

two anti-instantons at τ

2

and τ

4

will be a good one provided τ

1

≪ τ

2

≪

τ

3

≪ τ

4

≪ τ

5

. We will check later whether this condition is satisﬁed for

the leading terms.

4.2.1 The multi-instanton amplitude

We will now compute our amplitudes by summing over all possible sequences of

widely separated instantons (and anti-instantons): this is the diluted instanton

gas approximation.

Each of these quasi-solutions contributes to the amplitude in two ways:

Exponent: the contribution to S

E

is localized at each instanton (or anti-

instanton) transition time. Indeed, away from these transition times, the

quasi-solution is close to ±a with a neglectible velocity. S

E

is therefore

zero in these regions.

82 CHAPTER 4. INSTANTONS

x(τ)

τ

τ

1

τ

2

τ

3

τ

4

τ

5

a

−a

Figure 4.8: A quasi-solution with three instantons localized at τ

1

, τ

3

and τ

5

,

and two anti-instantons localizd at τ

2

and τ

4

.

Each instanton (or anti-instanton) will give a factor S

I

in the action.

Due to the decomposition property of the action, the contribution of an

approximate solution with N instantons and anti-instantons will be:

S

N

= NS

I

⇒ e

−

S

N

ℏ

= e

−N

S

I

ℏ

(4.19)

Determinant prefactor: notice again that away from the very short instants

over which the transitions take place, the system sits at either one of its

classical vacua at x = ±a, around which V

′′

(x) =

mω

2

2

.

If we had V

′′

(¯ x) ≡

mω

2

2

throughout the interval

_

−

β

2

,

β

2

_

, then the pref-

actor would be the usual one for the harmonic oscillator:

˜

^

_

det

_

−m

d

2

dτ

2

+V

′′

(¯ x)

__

−

1

2

=

_

mω

2πℏ sinh(ωβ)

_1

2

≈

_

mω

πℏ

_1

2

e

−

ωβ

2

Given the factorization property of the path integral and, in particular, of

the Gaussian integral we are interested in, we expect that each transition

will correct the prefactor of the harmonic oscillator with a factor R.

We thus expect

_

¸

_

¸

_

˜

^

_

det

_

−m

d

2

dτ

2

+V

′′

(¯ x)

__

−

1

2

=

_

mω

πℏ

_1

2

e

−

ωβ

2

R

n

n = n

I

+n

A

(4.20)

for a quasi-solution with n instantons and anti-instantons.

Let us see how this argument can be made sharper, by focussing on the

simplest non-trivial case of an x

IA

solution. (The generalization to an

arbitrary sequence of I and A is straightforward.) To do so, we explicitly

write the Gaussian integral in factorized form by considering intermediate

times around instanton τ

±

1

= τ

1

±∆ and around anti-instanton τ

±

2

= τ

2

±∆

as shown in ﬁgure. We deﬁne the integration coordinates at intermediate

times as x(τ

±

i

) = x

±

i

, for i = 1, 2. The Euclidean propagator can thus be

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 83

x(τ)

τ

−

β

2

β

2

τ

−

1

τ

1

τ

+

1

τ

−

2

τ

2

τ

+

2

Figure 4.9: Intermediate times τ

±

i

= τ

i

± ∆ have been added to the instanton-

anti-instanton quasisolution, with [τ

i

−τ

j

[ ≫∆ ≫

1

ω

, so that at τ

±

i

, the solution

is almost in the vacuum.

factorized into the convolution of the propagator over ﬁve time intervals

(see ﬁgure)

K

E

(−a, −a;

β

2

, −

β

2

) =

_

dx

−

1

dx

+

1

dx

−

2

dx

+

2

K

E

(−a, x

+

2

;

β

2

, τ

+

2

)

. ¸¸ .

(1)

K

E

(x

+

2

, x

−

2

; τ

+

2

, τ

−

2

) K

E

(x

−

2

, x

+

1

; τ

−

2

, τ

+

1

)

. ¸¸ .

(2)

K

E

(x

+

1

, x

−

1

; τ

+

1

, τ

−

1

) K

E

(x

−

1

, −a; τ

−

1

, −

β

2

)

. ¸¸ .

(3)

(4.21)

Of course we work under the assumption [τ

2

−τ

1

[ ≫

1

ω

. The factorization

turns out to be useful if we also choose ∆ ≫

1

ω

. Indeed it is intuitively clear

that, for ω∆ ≫1, the amplitudes (1), (2) and (3) should be well approx-

imated (exponentially well indeed!) by harmonic oscillator amplitudes.

This is because, during the corresponding time intervals, the trajectory

sits very close to one or the other minimum, where the potential is well

approximated by the harmonic oscillator one. Moreover, we know that the

harmonic oscillator amplitude, for large euclidean time separation, is dom-

inated by the ground state (see section 2.3). We can thus approximate,

e.g. amplitude (2), by:

K

E

(x

−

2

, x

+

1

; τ

−

2

, τ

+

1

) =

¸

x

−

2

¸

¸

e

−

H

0

ℏ

(τ

−

2

−τ

+

1

)

¸

¸

x

+

1

_ _

1 +O(e

−ω∆

)

_

≈ ¸0

+

[ e

−

H

0

ℏ

(τ

−

2

−τ

+

1

)

[0

+

) Ψ

+

0

∗

(x

−

2

)Ψ

+

0

(x

+

1

)

= Ψ

+

0

∗

(x

−

2

)Ψ

+

0

(x

+

1

)e

−

ω

2

(τ

−

2

−τ

+

1

)

(4.22)

where H

0

represent the harmonic oscillator Hamiltonian around +a. Here

we also decomposed position eigenvectors [x) into eigenstates of the har-

monic oscillator Hamiltonian around +a: [x) =

n+

Ψ

+

n

(x) [n

+

), and we

kept only the ground state contribution.

By applying the same approximation to (1) and (3) as well, and expanding

into the similarly deﬁned harmonic oscillators eigenstates around −a, we

84 CHAPTER 4. INSTANTONS

can thus write the Euclidean propagator as

K

E

(−a, −a;

β

2

, −

β

2

) ≈

_

dx

−

1

dx

+

1

dx

−

2

dx

+

2

¸

¸

Ψ

−

0

(−a)

¸

¸

2

e

−

ωβ

2

Ψ

−

0

(x

+

2

) Ψ

+

0

∗

(x

−

2

)Ψ

+

0

(x

+

1

) Ψ

−

0

∗

(x

−

1

)e

ω

2

((τ

+

2

−τ

−

2

)+(τ

+

1

−τ

−

1

))

K

E

(x

+

2

, x

−

2

; τ

+

2

, τ

−

2

)K

E

(x

+

1

, x

−

1

; τ

+

1

, τ

−

1

)

=

_

mω

πℏ

_1

2

e

−

ωβ

2

R

2

(4.23)

where

R =

_

dx

−

1

dx

+

1

e

ω

2

(τ

+

1

−τ

−

1

)

Ψ

+

0

(x

+

1

) Ψ

−

0

∗

(x

−

1

)K

E

(x

+

1

, x

−

1

; τ

+

1

, τ

−

1

) (4.24)

R will be ﬁxed later on by studying more carefully the one-instanton

amplitude. This proves factorization as promised.

We are not done yet!

We must still sum over over the locations τ

1

, . . . , τ

n

of the instantons. Indeed,

as long as τ

1

≪τ

2

≪ ≪τ

n

, we still get an equally good approximate solution

by moving the positions τ

i

around. The summation over instanton and anti-

instanton location simply (and obviously) corresponds to the integral:

_ β

2

−

β

2

dτ

n

_

τn

−

β

2

dτ

n−1

_

τ2

−

β

2

dτ

1

=

β

n

n!

(4.25)

How do we normalize the dτ

i

integrals? This is not a problem for the mo-

ment: the normalization can be absorbed by the factor R (and this is not by

chance!).

Finally, one thing we have to be careful about is that instantons and anti-

instantons cannot be distributed arbitrarily: for example, starting at −a we

must ﬁrst encounter an instanton. For ¸−a[ e

−

βH

ℏ

[−a), all contributions will

have n

I

= n

A

, etc. . .

We can now put together all the pieces and compute the total semiclassical

amplitudes.

4.2.2 Summing up the instanton contributions: results

Let us compute ¸−a[ e

−

βH

ℏ

[−a) ﬁrst. The contributing solutions are depicted

in ﬁgure 4.10.

−a −a −a −a −a −a

+ + +

I I I A A A

Figure 4.10: The contributing solutions for the tranition from −a to −a are

solutions with any number of instanton-anti-instanton pairs.

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 85

Thus, the amplitude will be:

¸−a[ e

−

βH

ℏ

[−a) =

_

mω

πℏ

_1

2

e

−

βω

2

n=even

_

Re

−SI/ℏ

β

_

n

n!

[1 +O(ℏ)] (4.26)

Similarly, we can compute ¸a[ e

−

βH

ℏ

[−a):

¸a[ e

−

βH

ℏ

[−a) =

_

mω

πℏ

_1

2

e

−

βω

2

n=odd

_

Re

−SI/ℏ

β

_

n

n!

[1 +O(ℏ)] (4.27)

Performing the summation, we get:

¸±a[ e

−

βH

ℏ

[−a) =

_

mω

πℏ

_1

2

e

−

βω

2

1

2

_

e

Re

−S

I

/ℏ

β

∓e

−Re

−S

I

/ℏ

β

_

[1 +O(ℏ)]

(4.28)

Recalling that

¸x

f

[ e

−

βH

ℏ

[x

i

) =

n

Ψ

∗

n

(x

f

)Ψ

n

(x

i

)e

−

βEn

ℏ

(4.29)

and comparing to (4.28), we can read out (for β →∞) the energies of the lowest

energy states:

E

±

=

ℏω

2

±ℏRe

−

S

I

ℏ

(4.30)

Calling [+) and [−) the corresponding eigenstates, we can write

¸±a[ e

−

βH

ℏ

[−a) = Ψ

∗

+

(±a)Ψ

+

(−a)e

−

βE

+

ℏ

+ Ψ

∗

−

(±a)Ψ

−

(−a)e

−

βE

−

ℏ

(4.31)

Comparing to (4.28), and calling

ψ

0

(x) ≡

_

mω

πℏ

_1

4

e

−

mωx

2

2ℏ

(4.32)

the harmonic oscillator ground state wavefunction we get

Ψ

∗

−

(±a)Ψ

−

(−a) =

1

2

_

mω

πℏ

_1

2

=

_

1

√

2

ψ

0

(0)

_

2

Ψ

∗

+

(±a)Ψ

+

(−a) = ∓

1

2

_

mω

πℏ

_1

2

= ∓

_

1

√

2

ψ

0

(0)

_

2

⇒

_

Ψ

−

(x) is even Ψ

−

(a) = Ψ

−

(−a)

Ψ

+

(x) is odd Ψ

+

(a) = −Ψ

+

(−a)

(4.33)

as expected from the usual Schr¨odinger approach.

Basically,

Ψ

±

(x) =

1

√

2

(ψ

0

(x +a) ∓ψ

0

(x −a))

Let us now add a few comments about our results.

86 CHAPTER 4. INSTANTONS

1. Notice that in our computation, we have neglected the perturbative cor-

rections.

Technically speaking, equation (4.30) is “incorrect” because in it, we have

only retained the instanton correction which is of order

ℏe

−S/ℏ

∼ ℏe

−2/

¯

λ

(4.34)

and is smaller than the perturbative corrections (!) we previously com-

puted by Feynman diagrams:

∆E

n

∼ ℏω

¯

λ

n

(4.35)

The point, however, is that these perturbative correction aﬀect E

+

and

E

−

by exactly the same amount: they cancel in E

+

−E

−

.

A purist (as Coleman puts it) would only retain the instanton contribution

when writing the diﬀerence

E

+

−E

−

= 2ℏRe

−SI/ℏ

= 2ℏRe

−2/

¯

λ

(4.36)

One could go ahead and compute the perturbative corrections to ¸±a[ e

−

βH

ℏ

[−a)

with the same Feynman diagram techniques we used before; we would have

to compute the Green’s function around each solution.

For most of the time, the instanton solutions sit at either one or the other

minimum. The perturbative corrections are thus the same as those we

have computed before, plus corrections coming from the cubic anharmonic

term (see ﬁgure 4.11).

+ + +

Figure 4.11: The quantum corrections to the classical solutions. The ﬁrst di-

agram is a two-loop correction involving the λ

4

x

4

correction, the second one

is a two-loop correction involving the λ

3

x

3

correction, and the third one is a

three-loop correction.

For instance, ¸a[ e

−

βH

ℏ

[−a) becomes:

¸a[ e

−

βH

ℏ

[−a) =

_

mω

πℏ

_1

2

e

−

βE

0

(λ)

ℏ

n=odd

_

R(λ)e

−SI/ℏ

β

_

n

n!

E

0

(λ) =

ℏω

2

_

1 +

∞

n=1

c

n

¯

λ

n

_

R(λ) = R

_

1 +

∞

n=1

d

n

¯

λ

n

_

(4.37)

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 87

where c

n

are the loop coeﬃcient we have computed previously, and R is

the leading contribution to the prefactor we introduced before without

computing it explicitly.

Moral: we see the two classes of eﬀects at work:

• Non-perturbative eﬀects, determined by semiclassical methods. They

correspond to the non-trivial stationary points of the Euclidean ac-

tion.

• Perturbative eﬀects. They correspond to small quantum ﬂuctuations

around each stationary point of the action: they are the functional

integral analogue of the corrections to the saddle point approximation

of ordinary integrals. The perturbative corrections are conveniently

organized in a series of Feynman diagrams.

The leading O

_

¯

λ

0

_

term corresponds to evaluating the quantum ﬂuc-

tuation determinant

1

q

det(

ˆ

O)

.

- O

_

¯

λ

_

terms correspond to two-loops diagrams.

- O

_

¯

λ

2

_

terms correspond to three-loops diagrams.

- And so on. . .

2. We can now also check the validity of our dilute instanton approximation.

The exponential series

(Re

−S

I

/ℏ

β)

n

n!

is dominated by the terms with n ∼

Re

−SI/ℏ

β. Thus, the typical instanton density is: ρ ≈

n

β

= Re

−SI/ℏ

. It is

exponentially small! Their average separation in time is:

∆τ =

1

ρ

=

e

SI/ℏ

R

=

1

∆E

(4.38)

which is independent of β.

The sum over the instantons is truly necessary only to reproduce the

exponential behaviour of

_

e

−βH

ℏ

_

at β >

1

∆E

.

We can now go back and verify that the sequences of widely separated in-

stanton anti-instanton that are relevant to our result are stationary points

of the action up to eﬀects of order

δS

δx

¸

¸

¸

¸

x=xIA

∼ e

−ω|τ1−τ2|

≈ e

−[

ω

R

e

S

I

/ℏ

]

(4.39)

which is truly small! Much-much smaller than the e

−2/

¯

λ

eﬀects we have

been computing.

3. Another puzzle that one could have pointed out when we started our com-

putation concerned the relevance of multi-instanton trajectories. These

have an euclidean action = nS

I

which is bigger that the single instan-

ton one S

I

. The contribution of each such trajectory is thus weighted by

a much smaller exponential factor, and one may wonder if it makes any

sense to include them at all. For instance in ¸+a[e

−βH

[ −a) a IAI conﬁgu-

ration contributes with weight e

−3SI

, which is exponentially smaller than

the one instanton contribution e

−SI

. Now, at the end of our computation,

88 CHAPTER 4. INSTANTONS

we obviously know how to answer. Each multi-instanton contributes only

∝ e

−nSI

to the amplitude, however n-instantons have also a higher multi-

plicity ∝ R

n

β

n

/n! which for Rβ > e

−SI

compensates for the exponential

suppression. We can here draw a simple statistical mechanics analogy

(just an analogy!): the action exponent plays the role of Boltzmann factor

∝ e

−nSI

∼ e

−E/T

, the multiplicity ∝ R

n

β

n

/n plays the role of an en-

tropy factor e

S

, while the amplitude we are computing is the analogue of

the partition function e

−F/T

. Boltzmann suppressed conﬁgurations con-

tribute signiﬁcantly to the free energy at equilibrium if they have large

multiplicity.

4.2.3 Cleaning up details: the zero-mode and the compu-

tation of R

To wrap things up, we are left with explaining the origin of the dτ integral

on the instanton position and with computing R (the two things are obviously

related).

We have to study det

_

δ

2

S

δx

2

_¸

¸

¸

x=xI

, the ﬂuctuation determinant around the

one-instanton solution.

Surprise:

δ

2

S

δx

2

¸

¸

¸

¸

x=xI

= −m∂

2

τ

+V

′′

E

(x

I

) (4.40)

possesses one normalizable zero mode:

_

¸

_

¸

_

_

−m∂

2

τ

+V

′′

E

(x

I

)

_

y

0

(τ) = 0

_

∞

−∞

[y

0

(τ)[

2

dτ = 1

(4.41)

Indeed, we can compute from the equation of motion:

−m∂

2

τ

x

I

(τ) +V

′

E

(x

I

(τ)) = 0

⇒ −m∂

2

τ

˙ x

I

(τ) +V

′′

E

(x

I

(τ)) ˙ x

I

(τ) = 0 (4.42)

We can then compute the normalization of this zero mode:

_

˙ x

2

I

(τ)dτ =

1

m

_

m˙ x

2

I

(τ)dτ =

S

I

m

⇒ y

0

(τ) =

_

m

S

I

˙ x

I

(τ) (4.43)

The presence of Ψ

0

originates from the invariance of the action under time

translations: x

I

(τ + ∆) = x

∆

I

is a solution with same boundary conditions as

x

I

(τ), ∀∆ ∈ R, which implies that S(x

I

) = S(x

∆

I

). Thus,

δ

2

S

δx

2

= 0 along this

direction in the function space.

We can now expand the function x into modes around x

I

2

:

x(τ) = x

I

(τ) +

√

ℏ

n

y

n

(τ)˜ a

n

. ¸¸ .

y(τ)

(4.44)

2

We are using the same normalization of section 3.1

4.2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 89

Thus,

exp

_

−

1

2ℏ

δ

2

S

δx

2

y

2

_

= exp

_

−

1

2

n

λ

n

˜ a

2

n

_

= exp

_

−

1

2

λ

1

˜ a

2

1

−

1

2

λ

2

˜ a

2

2

+

_

(4.45)

as λ

0

= 0.

The integration over ˜ a

0

should be treated separately and with special care

(otherwise we would get a

1

λ0

nonsense).

Notice that d˜ a

0

is associated to a shift in the time position of the instanton:

x(τ) = x

I

(τ) +

_

ℏm

S

I

˙ x

I

(τ) ˜ a

0

+ ≈ x

I

_

τ + ˜ a

0

_

ℏm

S

I

_

+

x

I

(τ) = f(τ −τ

0

) ⇒

_

ℏm

S

I

d˜ a

0

= −dτ

0

(4.46)

Let us then choose an N such that the measure for the ﬂuctuation y(τ) is:

D[y] =

˜

^

n

d˜ a

n

√

2π

(4.47)

N is ﬁxed as usual by some simple reference system; for example the har-

monic oscillator:

S

E

=

m˙ y

2

2

+

mω

2

y

2

2

→ λ

h

n

(ω) = m

_

(

πn

β

)

2

+ω

2

_

(4.48)

˜

^

_

n

d˜ a

n

√

2π

e

−

1

2

P

n

λ

h

n

˜ a

2

n

=

˜

^

_

n

λ

h

n

_

−

1

2

=

_

mω

2πℏ sinh(βω)

=

˜

^

_

det (−m∂

2

τ

+mω

2

)

(4.49)

Let us now integrate in our case by separating out the integral on the zero

mode. We have

˜

^

_

n

d˜ a

n

√

2π

e

−

1

2

P

n

λn˜ a

2

n

=

_

dτ

0

_

S

I

mπℏ

˜

^

n≥1

d˜ a

n

√

2π

e

−

1

2

P

n

λn˜ a

2

n

=

_

dτ

0

_

S

I

mπℏ

˜

^

_

det (−m∂

2

τ

+mω

2

)

_

det (−m∂

2

τ

+mω

2

)

_

det

′

(−m∂

2

τ

+V

′′

(x

I

))

=

_

dτ

0

_

S

I

mπℏ

_

mω

πℏ

_1

2

e

−

βω

2

¸

det (−m∂

2

τ

+mω

2

)

det

′

(−m∂

2

τ

+V

′′

(x

I

))

(4.50)

where we multiplied and divided by the harmonic oscillator determinant, used

eq. (4.49) and took the βω ≫1 limit. By det

′

we indicate the determinant with

the zero mode removed: det

′

= λ

1

λ

2

λ

3

. . . . Now, by a straightforward but

tedious computation which we shall not do explicitly

3

one ﬁnds

_

S

I

mπℏ

¸

det (−m∂

2

τ

+mω

2

)

det

′

(−m∂

2

τ

+V

′′

(x

I

))

=

_

mω

πℏ

_1

2

˜ v ≡ R (4.51)

3

To be found in Coleman’s lectures and also proposed in one of the Homeworks to be

posted on line (next year...).

90 CHAPTER 4. INSTANTONS

where ˜ v is determined by the asymptotic behaviour of ˙ x

I

for [τ −τ

0

[ ≫

1

ω

:

˙ x

I

= ˜ ve

−ω|τ−τ0|

(4.52)

In the speciﬁc case of the

λ

4!

_

x

2

−a

2

_

2

potential, we have:

x(τ) = a

1 −e

−ω(τ−τ0)

1 +e

−ω(τ−τ0)

τ−τ0≫

1

ω

−→ a

_

1 −2e

−ω(τ−τ0)

_

⇒ ˙ x

τ−τ0≫

1

ω

−→ 2aωe

−ω|τ−τ0|

⇒ ˜ v = 2aω (4.53)

Thus, we can compute:

_

mω

πℏ

˜ v

2

_1

2

=

_

4ma

2

ω

3

πℏ

_

1

2

=

_

12m

2

ω

5

πℏλ

_

1

2

=

_

12

π

¯

λ

_1

2

ω (4.54)

We thus ﬁnally obtain the energy splitting between the two lowest energy

levels:

E

+

−E

−

= 2ℏ

_

mω

πℏ

_1

2

˜ ve

−

S

I

ℏ

(4.55)

Chapter 5

Interaction with an external

electromagnetic ﬁeld

We will now discuss the quantum mechanics of a particle interacting with an

external electromagnetic ﬁeld. We will see that the path integral approach oﬀers

a neat viewpoint on the issue of gauge invariance and on the role of topology.

5.1 Gauge freedom

5.1.1 Classical physics

The relativistically invariant action for a neutral spinless particle is

S = −mc

2

_

γ

dτ = −mc

2

_

γ

_

1 −

˙ x

2

c

2

dt (5.1)

where we have chosen the parametrization so that:

_

¸

¸

_

¸

¸

_

dτ

2

= (dx

0

)

2

−

1

c

2

(dx

i

)

2

=

_

(∂

σ

x

0

)

2

−

1

c

2

(∂

σ

x

i

)

2

_

dσ

2

σ = x

0

≡ t ⇒ dτ =

_

1 −

˙ x

2

c

2

dt

(5.2)

The simplest relativistic invariant interaction with the electromagnetic quadrivec-

tor ﬁeld A

µ

= (A

0

, −A

i

) is obtained by modifying eq. (5.1) to

S = −mc

2

_

γ

dτ −e

_

γ

A

µ

dx

µ

= −

_

γ

_

mc

2

_

1 −

˙ x

2

c

2

+eA

0

−eA

i

˙ x

i

_

dt (5.3)

The coeﬃcient e is the electric charge of the particle.

Notice that the gauge transformation A

µ

→ A

µ

+ ∂

µ

α does not aﬀect the

equations of motion since S only changes by boundary terms.

S →S −e (α(x

f

, t

f

) −α(x

i

, t

i

))

Consequently the classical equations of motion only depend on the electro-

magnetic ﬁelds F

µν

= ∂

µ

A

ν

− ∂

ν

A

µ

, which are local functions of A

µ

that are

not aﬀected by gauge transformations.

91

92CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

The classically relevant quantity is the electromagnetic ﬁeld F

µν

, but the

lagrangian involves a redundant quantity: A

µ

.

Let us consider now what happens in the Hamiltonian description of this

system.

From equation (5.3), we can ﬁnd the conjugate momenta:

p

i

=

∂L

∂ ˙ x

i

=

m˙ x

i

_

1 −

˙ x

2

c

2

+eA

i

(x, t) (5.4)

This leads to the hamiltonian:

H = p

i

˙ x

i

−L =

_

(p

i

−eA

i

)

2

c

2

+m

2

c

4

+eA

0

(5.5)

where A

0

= V is the electromagnetic potential.

Remark: both p

i

and H are not invariant under a change of gauge for A

µ

:

_

¸

¸

¸

_

¸

¸

¸

_

p

i

=

m˙ x

i

_

1 −

˙ x

2

c

2

+eA

i

(x, t) →p

′i

= p

i

−e∂

i

α

H =

_

(p

i

−eA

i

)

2

c

2

+m

2

c

4

+eA

0

→H

′

= H +e∂

0

α

(5.6)

(H, p

i

) transforms like e(A

0

, A

i

).

Notice that the kinetic momentum is gauge invariant (being just a function

of the velocity):

Π

i

≡

m˙ x

i

_

1 −

˙ x

2

c

2

= p

i

−eA

i

→

_

p

i

−e∂

i

α

_

−e

_

A

i

−∂

i

α

_

= Π

i

(5.7)

Now what is the transformation (5.6) in the Hamiltonian language? Deﬁn-

ing F(x, t) ≡ eα, we can write:

_

¸

_

¸

_

p

i

→p

′i

= p

i

−∂

i

F

x

i

→x

′i

= x

i

H(p, x) →H

′

(p

′

, x

′

) = H(p, x) +∂

0

F

(5.8)

It is a canonical transformation, in the sense that the following relations

are satisﬁed:

_

p

′i

, p

′j

_

=

∂p

′i

∂p

k

∂p

′j

∂x

k

−

∂p

′i

∂x

k

∂p

′j

∂p

k

= δ

ik

(−∂

kj

F) −(−∂

jk

F)δ

ik

= 0

_

p

′i

, x

′j

_

=

∂p

′i

∂p

k

∂x

′j

∂x

k

−

∂p

′i

∂x

k

∂x

′j

∂p

k

= δ

ik

δ

jk

= δ

ij

(5.9)

and, moreover, p

′i

and x

′i

satisfy the Hamilton equations for H

′

:

_

¸

¸

_

¸

¸

_

∂p

′i

∂t

=

∂H

′

∂x

′i

∂x

′i

∂t

=

∂H

′

∂p

′i

(5.10)

5.1. GAUGE FREEDOM 93

H and p

i

are not invariant under gauge transformations, but they are

covariant:

_

H, p

i

_

→

_

H, p

i

_

+e (∂

0

α, −∂

i

α) (5.11)

5.1.2 Quantum physics

Now, what happens to gauge invariance upon quantization? We will study the

path integral viewpoint ﬁrst, and the Hamiltonian viewpoint later.

In the path integral formalism, we can write very simply:

K(x

f

, t

f

; x

i

, t

i

) =

_

D[x]e

iS[x]

ℏ

→

_

D[x]e

iS[x]

ℏ

−

ie

ℏ

R x

f

,t

f

x

i

,t

i

dα

= e

−

ie

ℏ

α(x

f

,t

f

)

K(x

f

, t

f

; x

i

, t

i

)e

ie

ℏ

α(xi,ti)

(5.12)

We have the relation:

Ψ(x

f

, t

f

) =

_

K(x

f

, t

f

; x

i

, t

i

)Ψ(x

i

, t

i

)dx

i

(5.13)

The gauge covariance of this equation leads to:

Ψ(x, t) → e

−

ie

ℏ

α(x,t)

Ψ(x, t) (5.14)

This is what we should have expected!

Indeed, classically, the gauge transformation of the electromagnetic quadrivec-

tor ﬁeld is related to a canonical transformation of the Hamiltonian variables:

_

¸

¸

¸

¸

¸

_

¸

¸

¸

¸

¸

_

A

µ

gauge

transformation

−→ A

µ

+∂

µ

α

x

i

p

i

H

canonical

transformation

−→

_

¸

_

¸

_

x

i

p

i

−e∂

i

α

H +e∂

t

α

(5.15)

Quantum mechanically, we can make an analogy with this representation:

_

¸

_

¸

_

A

µ

gauge

transformation

−→ A

µ

+∂

µ

α

Ψ

canonical

transformation

−→ e

−

ie

ℏ

α

Ψ

(5.16)

Indeed, Ψ → UΨ with U

†

U = UU

†

= 1 is the most general canon-

ical transformation in the quantum description. Under the above rotation

of the state vectors, the matrix elements of any operator

ˆ

O transform like

¸Ψ

1

[

ˆ

O[Ψ

2

) → ¸Ψ

1

[U

†

ˆ

OU[Ψ

2

). As the only thing that matters are matrix el-

ements, very much like we do when going from the Schroedinger to the Heisen-

berg picture for time evolution, we could alternatively deﬁne the transformation

as one where state vectors are unaﬀected while operators transform according

to

ˆ

O → U

†

ˆ

OU. Canonical commutation relations are indeed preserved by the

that transformation:

[ ˆ p, ˆ x] = −iℏ1 →

_

U

†

ˆ pU, U

†

ˆ xU

¸

= U

†

[ ˆ p, ˆ x] U = −iℏU

†

U = −iℏ1. (5.17)

94CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

These results immediately tell us how to proceed in the Hamiltonian ap-

proach.

Let us quantize p

i

, x

j

→ ˆ p

i

, ˆ x

j

, with the canonical commutation relations:

_

ˆ p

i

, ˆ p

j

¸

=

_

ˆ x

i

, ˆ x

j

¸

= 0

_

ˆ p

i

, ˆ x

j

¸

= −iℏδ

ij

(5.18)

In the non-relativistic limit where

˙ x

c

≪ 1, the classical Hamiltonian takes

the form:

H = mc

2

+

1

2m

_

p

i

−eA

i

_

2

+eA

0

. (5.19)

Quantizing and neglecting the (now irrelevant) constant mc

2

we have the Hamil-

tonian operator

H(ˆ p, ˆ x) =

1

2m

_

ˆ p

i

−eA

i

(ˆ x)

_

2

+eA

0

(ˆ x) (5.20)

Let us now work in the coordinate representation, where

ˆ x

i

= x

i

ˆ p

i

= −iℏ

∂

∂x

i

(5.21)

The matrix elements of ˆ p

i

are not invariant under the gauge transformation

Ψ →e

−

ie

ℏ

α

Ψ:

¸Ψ

1

[ ˆ p

j

[Ψ

2

) =

_

d

3

xΨ

∗

1

(x)(−iℏ∂

j

)Ψ

2

(x)

→

_

d

3

xΨ

∗

1

e

ie

ℏ

α

(−iℏ∂

j

)e

−

ie

ℏ

α

Ψ

2

=

_

d

3

xΨ

∗

1

(−iℏ∂

j

−e∂

j

α)Ψ

2

(5.22)

However, the matrix elements of the kinetic momentum

ˆ

Π

j

= ˆ p

j

− e

ˆ

A

j

are

gauge invariant. Indeed we have

ˆ

Π

j

(A)Ψ(x) →

ˆ

Π

j

(A

′

)Ψ

′

(x)

=

_

−iℏ∂

j

−eA

j

+e∂

j

α

_

e

−

ie

ℏ

α

Ψ(x)

= e

−

ie

ℏ

α

_

−iℏ∂

j

−e∂

j

α −eA

j

+e∂

j

α

_

Ψ(x)

= e

−

ie

ℏ

α

ˆ

Π

j

(A)Ψ(x) (5.23)

and so the kinetic momentum

ˆ

Π

j

(A) is a covariant derivative (

ˆ

Π

j

(A)Ψ trans-

forms like Ψ). Thus, when taking matrix elements, the phase factor in eq. (5.23)

cancels out and we conclude that the matrix elements of

ˆ

Π

j

(A) are gauge in-

variant.

Now, what about the gauge covariance of the Hamiltonian and the Schr¨odinger

equation?

Under a gauge transformation the Hamiltonian transforms like:

H

α

−→H

α

=

1

2m

_

ˆ p

i

−e

ˆ

A

i

+e∂

i

α

_

2

+e

_

ˆ

A

0

+∂

t

α

_

(5.24)

Assume now that Ψ solves the Schr¨odinger equation with respect to H:

iℏ∂

t

Ψ = HΨ

5.2. PARTICLE IN A CONSTANT MAGNETIC FIELD 95

It is a good exercise to check that Ψ

α

= e

−

ie

ℏ

α

Ψ solves the Schr¨odinger

equation with respect to H

α

:

iℏ∂

t

Ψ

α

= H

α

Ψ

α

(5.25)

Thus, the Schr¨odinger equation is gauge covariant.

In the same way, we have that the probability density Ψ

∗

Ψ = [Ψ[

2

is gauge

invariant. Then what is the probability current?

It is expectedly given through the velocity operator:

ˆ v

j

=

ˆ

Π

j

m

=

1

m

_

−iℏ∂

j

−eA

j

_

(5.26)

J

j

(x) =

1

2

_

Ψ

∗

(x)ˆ v

j

Ψ(x) −

_

ˆ v

j

Ψ(x)

_

∗

Ψ(x)

_

=

−i

2m

_

Ψ

∗

_

ℏ∂

j

−ieA

j

_

Ψ−

__

ℏ∂

j

−ieA

j

_

Ψ

_

∗

Ψ

_

(5.27)

Indeed, by the Schr¨odinger equation, we can check that ∂

t

[Ψ[

2

+∂

j

J

j

= 0.

J

j

is also manifestly gauge invariant.

In order to prove current conservation let us deﬁne:

−→

D

i

= ℏ

−→

∂

i

−ieA

i

←−

D

i

= ℏ

←−

∂

i

+ieA

i

(5.28)

Using these deﬁnitions, we can write:

J

j

=

−i

2m

_

Ψ

∗

−→

D

j

Ψ−Ψ

∗

←−

D

j

Ψ

_

−→

∂

i

+

←−

∂

i

=

1

ℏ

_

−→

D

i

+

←−

D

i

_

−→

D

2

2m

=

D

2

2m

= eA

0

−H (5.29)

We can then compute:

∂

j

J

j

=

−i

2mℏ

_

Ψ

∗

_

−→

D

i

+

←−

D

i

__

−→

D

i

Ψ

_

−

_

Ψ

∗

←−

D

i

__

−→

D

i

+

←−

D

i

_

Ψ

_

=

−i

2mℏ

_

Ψ

∗

−→

D

i

−→

D

i

Ψ−Ψ

∗

←−

D

i

←−

D

i

Ψ

_

=

−i

2mℏ

_

Ψ

∗

_

D

2

Ψ

_

−

_

D

2

Ψ

_

∗

Ψ

_

=

i

ℏ

_

Ψ

∗

(H −eA

0

) Ψ−((H −eA

0

) Ψ)

∗

Ψ

_

= i

_

Ψ

∗

i∂

t

Ψ−(i∂

t

Ψ)

∗

Ψ

_

= −Ψ

∗

∂

t

Ψ−∂

t

Ψ

∗

Ψ = −∂

t

[Ψ[

2

(5.30)

5.2 Particle in a constant magnetic ﬁeld

Let us now study the system of a particle in a constant magnetic ﬁeld.

96CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

5.2.1 An exercise on translations

Let us ﬁrst study what happens with spatial translations in this particular case.

We take the magnetic ﬁeld along the z-axis:

B = (0, 0, B).

We can choose the electromagnetic quadrivector as:

A =

1

2

B ∧r ⇒ A

1

= −

1

2

By, A

2

=

1

2

Bx (5.31)

We then recover the magnetic ﬁeld by:

B

3

=

_

∇∧

A

_

3

= ǫ

321

∂

2

A

1

+ǫ

312

∂

1

A

2

= −∂

2

A

1

+∂

1

A

2

= B (5.32)

This possible gauge choice leads to the following Hamiltonian, in the coor-

dinate representation:

ˆ

H =

1

2m

_

ˆ

Π

i

(A)

_

2

=

1

2m

_

_

ˆ p

1

+

eB

2

y

_

2

+

_

ˆ p

2

−

eB

2

x

_

2

+ ˆ p

2

3

_

(5.33)

How are spatial translations realized?

• −iℏ∂

i

, the naive generators, do not commute with the Hamiltonian.

• Translations along the x and y axis are described by the unitary operator

U(a, b) such that U(a, b)Ψ(x, y) = Ψ(x + a, y + b). This operator can be

written as U(a, b) = exp

_

iaˆ p1

ℏ

+

ibˆ p2

ℏ

_

. Its action on operators is

ˆ

O(x +

a, y +b) = U(a, b)

ˆ

O(x, y)U

†

(a, b).

Let us now look at the translational invariance of the spectrum of the Hamil-

tonian. We have:

ˆ

H(x, y)Ψ

n

(x, y) = E

n

Ψ

n

(x, y)

U(a, b)

ˆ

H(x, y)Ψ

n

(x, y) = E

n

U(a, b)Ψ

n

(x, y)

U(a, b)

ˆ

H(x, y)U

†

(a, b)U(a, b)Ψ

n

(x, y) = E

n

U(a, b)Ψ

n

(x, y)

ˆ

H(x +a, y +b)Ψ

n

(x +a, y +b) = E

n

Ψ

n

(x +a, y +b) (5.34)

However,

ˆ

H(x +a, y +b) is related to

ˆ

H(x, y) by a gauge transformation:

_

¸

¸

_

¸

¸

_

¯

A

1

= −

1

2

B(y +b) = −

1

2

By +∂

x

_

−

Bbx

2

_

¯

A

2

=

1

2

B(x +a) =

1

2

Bx +∂

y

_

Bay

2

_

⇒

_

_

_

¯

A

µ

= A

µ

+∂

µ

α

α =

B

2

(ay −bx)

(5.35)

Thus, we can put these results together and write:

e

ie

ℏ

α

ˆ

H(x, y)e

−

ie

ℏ

α

Ψ

n

(x +a, y +b) = E

n

Ψ

n

(x +a, y +b) (5.36)

5.2. PARTICLE IN A CONSTANT MAGNETIC FIELD 97

We have then the following relation:

_

ˆ

H(x, y)

˜

Ψ

n

= E

n

˜

Ψ

n

˜

Ψ

n

= e

−

ie

ℏ

α

Ψ

n

(x +a, y +b) = e

i

eB

2ℏ

(bx−ay)

e

i

ℏ

(aˆ p1+bˆ p2)

Ψ

n

(x, y)

(5.37)

We have thus obtained the inﬁnitesimal generators of translations:

T

1

= ˆ p

1

−

eB

2

y

T

2

= ˆ p

2

+

eB

2

x (5.38)

Notice that these generators commute with Π

1

and Π

2

:

_

T

1

, Π

1

¸

=

_

ˆ p

1

−

eB

2

y, ˆ p

1

+

eB

2

y

_

= 0

_

T

1

, Π

2

¸

=

_

ˆ p

1

−

eB

2

y, ˆ p

2

−

eB

2

x

_

= −

eB

2

([ ˆ p

1

, x] + [y, ˆ p

2

]) = 0 (5.39)

However, T

1

and T

2

do not commute among themselves:

[T

1

, T

2

] =

_

ˆ p

1

−

eB

2

y, ˆ p

2

+

eB

2

x

_

=

eB

2

(−iℏ −iℏ) = −iℏeB (5.40)

• The group of translations is thus a projective realization of the classical

one.

• The commutator is related to the enclosed ﬂux.

• Previously, we chose one particular way to translate from (x, y) to (x +

a, y +b). Any other way would diﬀer by e

iγ

, with γ ∼ ℏeB.

5.2.2 Motion

We want to describe ﬁrst the classical motion of a particle in this constant

magnetic ﬁeld, oriented along the z-axis:

B = (0, 0, B). The classical equations

of motion give the following result:

• ¨ z = 0 ⇒z = z

0

+v

z

t.

• (x, y) describe circular orbits with frequency ω

c

=

eB

m

. If eB > 0, the

rotation is clockwise.

Notice: harmonicity: the frequency is independent of the radius of rota-

tion.

We get:

_

x = x

c

+Rcos ω

c

t

y = y

c

−Rsin ω

c

t

(5.41)

where (x

c

, y

c

) is the coordinate of the center of rotation. Its time evolution

is described by:

_

¸

¸

_

¸

¸

_

x

c

= x +

1

ω

c

˙ y = x +

1

ω

c

Π

y

m

y

c

= y −

1

ω

c

˙ x = y −

1

ω

c

Π

x

m

(5.42)

98CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

x

c

and y

c

are two constants of motion. What is the associated symmetry?

The answer is translations in (x, y).

Let us see this directly by working with the quantum Hamiltonian in some

gauge. Remember that we have seen that performing a gauge transformation

on the vector potential amounts to a phase rotation: Ψ →e

−

ie

ℏ

α

Ψ.

Let us pick Landau’s gauge which is also convenient to compute energy levels:

A

x

= −By A

y

= 0 (5.43)

The Hamiltonian is:

ˆ

H =

1

2m

_

(ˆ p

1

+eBˆ y)

2

+ ˆ p

2

2

+ ˆ p

2

3

_

(5.44)

The kinetic momenta are given by:

_

¸

¸

_

¸

¸

_

ˆ

Π

1

= ˆ p

1

+eBˆ y

ˆ

Π

2

= ˆ p

2

ˆ

Π

3

= ˆ p

3

(5.45)

• Notice that _

ˆ

Π

1

,

ˆ

Π

2

_

= −iℏeB ,= 0 (5.46)

The two velocity components in the plane perpendicular to

B do not

commute. Thus, they cannot be measured simultaneously with arbitrary

accuracy.

• The combinations

_

T

1

= ˆ p

1

T

2

= ˆ p

2

+eBx

(5.47)

commute with

ˆ

Π

1,2,3

. Thus, [T

i

, H] = 0, which implies that they are

conserved.

They indeed correspond to the coordinates of the center of the classical

motion. Comparing them to (5.42), we get:

_

T

1

= Π

1

−eBy = −eBy

c

T

2

= Π

2

+eBx = eBx

c

(5.48)

• Quantum mechanically, also T

1

and T

2

cannot be ﬁxed simultaneously as

[T

1

, T

2

] = −iℏeB (5.49)

• Given a solution of the Schr¨odinger equation Ψ(x, y, z), then

Ψ

a,b

(x, y, z) = e

i

ℏ

(aT1+bT2)

Ψ(x, y, z) (5.50)

is also a solution. What does this operation represent?

It is a space translation.

5.2. PARTICLE IN A CONSTANT MAGNETIC FIELD 99

This fact is obvious for e

i

aT

1

ℏ

= e

i

aˆ p

1

ℏ

= e

a∂x

.

For e

i

bT

2

ℏ

, we have:

e

i

bT

2

ℏ

Ψ(x, y, z) = e

i

eBx

ℏ

e

b∂y

Ψ(x, y, z) = e

i

eBx

ℏ

Ψ(x, y +b, z)

This is an ordinary translation composed with a gauge transformation.

• Notice that [T

1

, T

2

] ,= 0 implies that the translation group is non-commutative

in presence of a magnetic ﬁeld.

To be stressed and to conclude:

The pairs of canonically conjugated variables are (

ˆ

Π

1

,

ˆ

Π

2

), which are physical

velocities, and (T

1

, T

2

) which are the coordinates of the center of the orbit.

In a magnetic ﬁeld, the canonically conjugated variables are (velocity,velocity)

and (position,position), whereas in ordinary motion in a potential those are (ve-

locity,position).

5.2.3 Landau levels

In order to study the Landau quantization, let us continue working in the Lan-

dau gauge:

A

x

= −By A

y

= 0 (5.51)

• The rotation invariance around (x, y) = 0 is not manifest.

• The translational invariance in the x-direction is manifestly preserved,

whereas it is not manifest in the y-direction.

We have the Hamiltonian:

ˆ

H =

1

2m

_

(ˆ p

1

+eBˆ y)

2

+ ˆ p

2

2

+ ˆ p

2

3

_

(5.52)

As ˆ p

1

and ˆ p

3

commute with

ˆ

H, we can choose (p

3

, p

1

, E) as quantum num-

bers:

ˆ p

3

Ψ = p

3

Ψ

ˆ p

1

Ψ = p

1

Ψ

ˆ

HΨ = EΨ (5.53)

Such a state will satisfy:

Ψ(x, y, z) = e

i

ℏ

(p1x+p3z)

F(y) (5.54)

For such a state, the Hamiltonian is equivalent to:

ˆ

H =

1

2m

_

(p

1

+eBˆ y)

2

+ ˆ p

2

2

+p

2

3

_

(5.55)

It is the Hamiltonian of a harmonic oscillator in (ˆ p

2

, ˆ y) with frequency ω

c

=

eB

m

, which is the cyclotron frequency of the particle, and centered at y

c

= −

p1

eB

=

−

p1

mωc

:

ˆ

H =

1

2m

ˆ p

2

2

+

1

2

m

_

eB

m

_

2 _

ˆ y +

p

1

eB

_

2

+

p

2

3

2m

(5.56)

100CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

By quantizing this harmonic oscillator in (ˆ p

2

, ˆ y), we get the energy levels:

E(p

3

, p

1

, n) =

p

2

3

2m

+ℏω

c

_

n +

1

2

_

(5.57)

with the wave functions:

Ψ

p3,p1,n

(x, y, z) = e

−

i

ℏ

(p1x+p3z)

Ψ

n

_

y +

p

1

mω

c

_

Ψ

n

(y) =

_

mω

c

πℏ

_1

4 1

√

2

n

n!

H

n

_

y

_

mω

c

ℏ

_

e

−

mω

2

c

y

2

2ℏ

(5.58)

where Ψ

n

(y) is the ordinary wave function of a harmonic oscillator around y = 0

with frequency ω

c

, and H

n

(x) are the Hermite polynomials deﬁned as:

H

n

(x) = (−1)

n

e

x

2 d

n

dx

n

e

−x

2

(5.59)

• The energy is degenerate in p

1

∼ y

c

. This is related to translational

invariance.

• The result (5.57) is in accord with the classical property of the system: the

frequency of rotation does not depend on the radius: this is harmonicity.

• The shape of [Ψ

p3,p1,n=0

(x, y)[ is depicted in ﬁgure 5.1.

x

y

y

c

Figure 5.1: The shape of the norm of the wavefunction Ψ

p3,p1,n=0

(x, y). We see

that it is far from rotational invariant.

Where is cylindrical symmetry gone? To make it manifest, we must su-

perimpose diﬀerent wavefunctions:

Ψ

p3,n2,n

(x, y, z) =

_

dp

1

Ψ

p3,p1,n

(x, y, z)f(p

1

, n

2

) (5.60)

• In Landau gauge, we have:

ˆ

Π

2

= ˆ p

2

ˆ

Π

1

= ˆ p

1

+eBˆ y

T

1

= ˆ p

1

T

2

= ˆ p

2

+eBˆ x (5.61)

We can check that:

_

ˆ

Π

1

, T

2

_

= eB([ ˆ p

1

, ˆ x] + [ˆ y, ˆ p

2

]) = 0 (5.62)

5.3. THE AHARONOV-BOHM EFFECT 101

We get:

mω

c

ˆ y =

ˆ

Π

1

−T

1

mω

c

ˆ x = T

2

−

ˆ

Π

2

(5.63)

We can thus conclude that in Landau gauge, for the states Ψ

p3,p1,n

, y

c

=

−

T1

mωc

is ﬁxed, whereas x

c

=

T2

mωc

is not, and is fully spread.

It is intructive to count the levels for the case where the motion is limited

to a region of sides L

x

, L

y

in the plane.

If the motion is limited to 0 x L

x

, then p

1

is quantized:

p

1

=

2πℏ

L

x

n

2

n

2

∈ Z (5.64)

Thus, the position y

c

is quantized:

y

c

=

p

1

eB

=

2πℏ

eBL

x

n

2

(5.65)

Note that it is the case whether or not the motion is limited in the y direction!

Now, if the motion is limited in the y direction as 0 y L

y

, then the

number of states of quantum numbers (n, p

3

) we can ﬁt in is:

N(n, p

3

) =

L

y

∆y

c

=

L

x

L

y

2πℏ

eB (5.66)

5.3 The Aharonov-Bohm eﬀect

The study of the vector potential in quantum mechanics opens a window on

deeper aspects of the quantum world, some of which have a full description only

in quantum ﬁeld theory:

1. The relevance of topology

2. The charge quantization

The Aharonov-Bohm eﬀect is one remarkable and yet simple example.

Consider an ideal inﬁnitely long and empty cylindrical shell such that parti-

cles do not have access to its interior. The propagation of particles in this setup is

characterized by a Schr¨odinger equation with boundary condition Ψ(x, y, z) = 0

on the surface of the cylinder. In the path integral approach, this corresponds

to summing only on paths that do not go through the interior of the cylinder.

Imagine now a modiﬁed arrangement where the cylindrical shell encloses a

uniform magnetic ﬁeld (see ﬁgure 5.2). This could, for instance, be realized by

placing an “inﬁnitely” long solenoid inside the shell.

Intuitively, we would think that physics (the propagation of charged parti-

cles) is unaﬀected outside the shell, as

B = 0 there. However, quantum me-

chanics tells that this conjecture is wrong!

To show this, we must consider the vector potential. Outside the shell,

B = 0, but

A ,= 0 as implied by Stokes’ theorem (see ﬁgure 5.3).

102CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

B ,= 0 B = 0

Figure 5.2: An inﬁnitely long cylinder has a constant magnetic ﬁeld inside it,

whereas outside it there is no magnetic ﬁeld.

γ

S = πr

2

0

B

γ = ∂σ

_

γ

A dx =

_

σ

B dσ = BS

Figure 5.3: An inﬁnitely long cylinder has a constant magnetic ﬁeld inside it,

whereas outside it there is no magnetic ﬁeld.

x

i

x

f

γ

1

γ

2

Figure 5.4: Two paths γ

1

and γ

2

from x

i

to x

f

that cannot be continuously

deformed into each other, due to the topological defect between them.

One possible gauge choice (the only one that manifestly respects cylindrical

symmetry) is A

φ

=

BS

2πr

, A

r

, A

z

= 0 in cylindrical coordinates, for r > r

0

.

5.3. THE AHARONOV-BOHM EFFECT 103

It is immediate to see that A

φ

has physical consequences by considering

quantum amplitudes in the path integral approach. Consider two paths γ

1

and

γ

2

like in ﬁgure 5.4.

• γ

1

cannot be continuously deformed into γ

2

.

• They are topologically inequivalent (they belong to diﬀerent homotopy

classes).

Consider the contribution of γ

1

and γ

2

to K(x

f

, x

i

; t):

γ

1

→e

iS

1

ℏ

= exp

_

iS

(1)

0

ℏ

−

ie

ℏ

_

γ1

A

i

dx

i

_

γ

2

→e

iS

2

ℏ

= exp

_

iS

(2)

0

ℏ

−

ie

ℏ

_

γ2

A

i

dx

i

_

(5.67)

where S

(i)

0

is the free particle action for γ

i

.

The phase diﬀerence induced by the vector potential is:

∆φ =

e

ℏ

__

γ2

A

i

dx

i

−

_

γ1

A

i

dx

i

_

=

e

ℏ

_

γ2−γ1

A

i

dx

i

= −

eBS

ℏ

≡ −

eΦ

ℏ

(5.68)

1. Apparently the presence of the solenoid is felt at the quantum level by

trajectories that are arbitrarily far from it!

2. Notice that the phase shift vanishes between paths that are topologically

equivalent, since they do not encircle any magnetic ﬂux.

The suprising relevance of the vector potential at the quantum level was ﬁrst

suggested by Aharonov and Bohm (1959) and later tested by Chambers (1960)

and more recently by Tonomura (1982).

The ideal “obvious” way to test this eﬀect is to perform the double slit ex-

periment in the presence of our inﬁnite and thin solenoid (in reality the solenoid

is replaced by a thin magnetized iron ﬁlament called a whisker), as in ﬁgure 5.5.

Two disjoint classes of trajectories contribute to the propagation amplitude

from the source S to the screen C:

K(x

f

, x

i

; t) =

_

D[x]

γ1

e

i

ℏ

“

S

(1)

0

−e

R

A

i

dx

i

”

+

_

D[x]

γ2

e

i

ℏ

“

S

(2)

0

−e

R

A

i

dx

i

”

(5.69)

By Stokes’ theorem,

_

γ1

A

i

dx

i

is the same for all trajectories in the class 1

(i.e. passing on the top of the solenoid). The same thing is true for the class 2.

Moreover, we have:

_

γ1

A

i

dx

i

−

_

γ2

A

i

dx

i

=

_

γ1−γ2

A

i

dx

i

= BS = Φ (5.70)

where Φ is the magnetic ﬂux through the solenoid.

Thus, the total contribution will be:

K(x

f

, x

i

; t) = e

−

ie

ℏ

R

γ

1

A

i

dx

i

_

φ

(0)

1

(x

f

) +e

ie

ℏ

Φ

φ

(0)

2

(x

f

)

_

= e

iα

_

φ

(0)

1

(x

f

) +e

ie

ℏ

Φ

φ

(0)

2

(x

f

)

_

(5.71)

104CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

x

i

S

γ

1

γ

2

solenoid

x

f

A

B C

Figure 5.5: The experimental setup to test the Aharonov-Bohm eﬀect. The two

paths γ

1

and γ

2

contribute to the propagation amplitude from the source S to

the screen C diﬀerently.

where φ

(0)

i

is the contribution of the class i in the absence of the solenoid.

The ﬁnal amplitude will be:

[K[

2

=

¸

¸

¸φ

(0)

1

¸

¸

¸

2

+

¸

¸

¸φ

(0)

2

¸

¸

¸

2

+ 2 Re

_

φ

(0)∗

1

e

ie

ℏ

Φ

φ

(0)

2

_

= F

1

+F

2

+F

int

(5.72)

This result tells us what the eﬀect of the solenoid is, no matter how compli-

cated φ

(0)

i

are. This is because its eﬀect is purely topological: the only property

of the trajectory that enters into the phase shift is its topology (whether it

belongs to the class 1 or 2).

In the case of the free propagation, we have (see ﬁgure 5.6):

φ

(0)∗

1

φ

(0)

2

= e

2

i

ℏ

ka

b

z

≡ e

2

i

ℏ

k

z

⇒ F

int

= 2 Re

_

e

i

„

2k

z

ℏ

+

eΦ

ℏ

«_

= 2 cos

_

2k

z

ℏ

+

eΦ

ℏ

_

(5.73)

Thus, as Φ is varied, the peaks of the interference pattern are shifted along

the z direction.

This eﬀect is periodic in the ﬂux Φ with period:

Φ

0

≡ ∆Φ =

2πℏ

e

= 4.135 10

−7

gauss cm

2

(5.74)

which represents a fundamental unit of magnetic ﬂux.

The eﬀect of the solenoid vanishes for Φ = nΦ

0

.

Moral:

1. The Aharonov-Bohm eﬀect shows that F

µν

underdescribes electromag-

netism; i.e. diﬀerent physical situations in a region may have the same

F

µν

in that region.

5.3. THE AHARONOV-BOHM EFFECT 105

a

−a

b

z

Figure 5.6: The experimental setup to test the Aharonov-Bohm eﬀect. The two

slits are located at a and −a along the vertical direction, the two screens are

separated by a distance b, and the position on the ﬁnal screen is described by

z.

2. The phase ϕ =

e

ℏ

_

A

i

dx

i

overdescribes electromagnetism; i.e. diﬀerent

phases in a region may describe the same physics. For example Φ

′

=

Φ +nΦ

0

.

3. What provides a complete description is the factor e

iϕ

= e

ie

ℏ

H

γ

Aµdx

µ

over

all possible closed paths in space-time, which after all is what enters in

the path integral.

Thus, electromagnetismis the gauge invariant manifestation of a non-integrable

phase factor, as stated by Yang and Wu (1975).

5.3.1 Energy levels

Let us now do a simple example on the energy levels of a charged particle

conﬁned on a ring of radius R around a solenoid (see ﬁgure 5.7).

solenoid

R

ring

Figure 5.7: A particle is conﬁned on a ring of radius R around a solenoid of

area σ went through by a magnetic ﬁeld of intensity B.

The vector potential describing this situation is such that:

_

σ

Bdσ =

_

∂σ

Adx (5.75)

106CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

The magnetic ﬂux is found to be:

Φ = Bσ = A

θ

(R)2πR (5.76)

Which ﬁxes the phase factor:

e

−

ie

ℏ

H

Adx

= e

−i

eΦ

ℏ

≡ e

−iϕ

(5.77)

We then ﬁnd the Hamiltonian for the particle:

H =

1

2m

(p

θ

−eA

θ

)

2

=

1

2m

_

−

iℏ∂

θ

R

−

eΦ

2πR

_

2

=

1

2m

ℏ

2

R

2

_

−i∂

θ

−

eΦ

2πℏ

_

2

=

1

2m

ℏ

2

R

2

_

−i∂

θ

−

ϕ

2π

_

2

(5.78)

The eigenstates of this Hamiltonian are:

Ψ

n

(θ) = e

inθ

n ∈ Z (5.79)

with energy levels:

E

n

=

1

2m

ℏ

2

R

2

_

n −

ϕ

2π

_

2

=

1

2m

ℏ

2

R

2

(n −α)

2

(5.80)

We can notice that when ϕ →ϕ+2πm, with m an integer, α →α+m, and

thus the spectrum is the same.

The conclusion is that what is physical is not ϕ itself, but the phase factor

e

iϕ

. Moreover, the physical angular momentum Rp

θ

= ℏ(n−α) is not quantized

in integer units of ℏ.

5.4 Dirac’s magnetic monopole

Let us now very brieﬂy and qualitatively discuss Dirac’s magnetic monopole.

Let us take our thin solenoid and cut it in two: the ﬂux will “expand” out.

The tip of the solenoid will act more or less as a point source of magnetic

ﬁeld: a magnetic monopole (see ﬁgure 5.8).

Of course

∇

B = 0 is still satisﬁed: the magnetic ﬂux is conserved as there

is the “incoming” ﬂux from the solenoid.

Notice that:

g

M

≡

_

σ

B dσ ⇒ B ≈

g

M

4πr

2

(5.81)

where σ is a sphere centered at the end of the solenoid, with the surface where

the solenoid crosses it excluded.

Assume now the ideal limit in which the solenoid is inﬁnitely thin.

If for all charged particles the charge e

j

the ﬂux Φ satisﬁes:

e

j

Φ

ℏ

= 2πn

j

(5.82)

there will be no way to tell the presence of the solenoid (also called Dirac’s

string). Our object will behave by all means as a point-like magnetic monopole!

Let us now turn the argument around:

5.4. DIRAC’S MAGNETIC MONOPOLE 107

B

B

Figure 5.8: The solenoid cut in two will act as a magnetic monopole. Maxwell’s

second law is still satisﬁed as the magnetic ﬂux is incoming from the solenoid.

Postulate: magnetic monopoles must exist in order to restore the symmetry

between magnetic charge and electric charge which is apparently broken

by Maxwell’s equations:

∇

E = ρ

E

∇

B = 0 (5.83)

Since

∇

**B = 0 follows not from dynamics but from kinematics: B
**

i

= ǫ

ijk

∂

j

A

k

,

a monopole state must involve an incoming ﬂux localized into some string;

but in order to preserve spherical symmetry the string should be unde-

tectable. Thus, the Aharonov-Bohm phase should be a multiple of 2π for

all existing electrically charged particles:

e

j

g

M

ℏ

= 2πn

j

(5.84)

Thus, the existence of a single monopole in the universe implies the quanti-

zation of the electric charge!

5.4.1 On coupling constants, and why monopoles are dif-

ferent than ordinary charged particles

The strength of the electromagnetic interactions is set by the ﬁne structure

constant:

α =

e

2

ℏc

≈

1

137

≪1 (5.85)

The electromagnetic interactions are thus “weak”.

The quantization condition implies that the minimal magnetic charge of a

monopole, if it existed, is:

g

min

=

2πℏ

e

(5.86)

Thus, the magnetic coupling constant, α

M

, is deﬁned as:

α

M

=

g

2

min

c

ℏ

=

(2π)

2

ℏc

e

2

=

(2π)

2

α

≈ 137 (2π)

2

≫1 (5.87)

The magnetic interactions among monopoles are therefore very strong!

108CHAPTER5. INTERACTIONWITH AN EXTERNAL ELECTROMAGNETICFIELD

Bibliography

[1] “Quantum Mechanics and Path Integrals” , R.P. Feynman and A.R. Hibbs,

McGraw-Hill, 1965. ( Chapters 1, 2 and 3: Great reading!!!)

[2] “Path Integral Methods and Applications”, R. MacKenzie, arXiv:quant-

ph/0004090. (readable overview on the subject)

The following books may be of help in complementing the lecture

notes.

[3] “Modern Quantum Mechanics”, J.J. Sakurai, The Benjamin/Cummings

Publishing Company, 1985. (chapters 2.5 and 2.6: quick introduction to

path integral and nice extended discussion of electromagnetic interactions,

Aharonov-Bohm, monopoles)

[4] “Aspects of Symmetry”, S. Coleman, Cambridge University Press, 1985.

(Sections 1-2 and appendices 1-2 of Chapter 7: The uses of the instantons.

I followed Coleman very closely in my lectures on instantons)

[5] “Path Integrals in Quantum Mechanics, Statistics and Polymer Physics”,

Hagen Kleinert, World Scientiﬁc, 1995. Chapter 3, in particular sections

3.9, 3.10 and 3.11 give a good account of Feynman diagrams

[6] “Topics in Advanced Quantum mechanics”, Barry R. Holstein, Addison-

Wesley 1982. Chapters 1 and 5 (contains a lot of subjects, which is nice,

but sometimes he is too sloppy)

[7] “Techniques and applications of Path Integration”, L.S. Schulman, John

Wiley & Sons Inc., 1981 Comprehensive but somewhat formal. Chapters

12, 17 and 18 overlap with presentations in the course.

Other references

[8] H. Kleinert and A. Chervyakov, “Simple Explicit Formulas for Gaussian

Path Integrals with Time-Dependent Frequencies,” Phys. Lett. A 245

(1998) 345 [arXiv:quant-ph/9803016].

[9] P. A. Horvathy, “The Maslov correction in the semiclassical Feynman in-

tegral,” arXiv:quant-ph/0702236.

109

110 BIBLIOGRAPHY

[10] “Quantum Mechanics”, Landau-Lifshitz.

2

Contents

1 The path integral formalism 1.1 Introducing the path integrals . . . . . . . . . . . . . . . . . 1.1.1 The double slit experiment . . . . . . . . . . . . . . 1.1.2 An intuitive approach to the path integral formalism 1.1.3 The path integral formulation . . . . . . . . . . . . . 1.1.4 From the Schr¨edinger approach to the path integral o 1.2 The properties of the path integrals . . . . . . . . . . . . . 1.2.1 Path integrals and state evolution . . . . . . . . . . 1.2.2 The path integral computation for a free particle . . 1.3 Path integrals as determinants . . . . . . . . . . . . . . . . 1.3.1 Gaussian integrals . . . . . . . . . . . . . . . . . . . 1.3.2 Gaussian Path Integrals . . . . . . . . . . . . . . . . 1.3.3 O(ℏ) corrections to Gaussian approximation . . . . . 1.3.4 Quadratic lagrangians and the harmonic oscillator . 1.4 Operator matrix elements . . . . . . . . . . . . . . . . . . . 1.4.1 The time-ordered product of operators . . . . . . . . 2 Functional and Euclidean methods 2.1 Functional method . . . . . . . . . . . . . . . . . . . . . . 2.2 Euclidean Path Integral . . . . . . . . . . . . . . . . . . . 2.2.1 Statistical mechanics . . . . . . . . . . . . . . . . . 2.3 Perturbation theory . . . . . . . . . . . . . . . . . . . . . 2.3.1 Euclidean n-point correlators . . . . . . . . . . . . 2.3.2 Thermal n-point correlators . . . . . . . . . . . . . 2.3.3 Euclidean correlators by functional derivatives . . 0 2.3.4 Computing KE [J] and Z 0 [J] . . . . . . . . . . . . 2.3.5 Free n-point correlators . . . . . . . . . . . . . . . 2.3.6 The anharmonic oscillator and Feynman diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 5 6 8 12 14 14 17 19 19 20 22 23 27 27 31 31 32 34 35 35 36 38 39 41 43 49 50 54 56 56 57 57 61 64

3 The semiclassical approximation 3.1 The semiclassical propagator . . . . . . . . . . . . . . . . . . . . 3.1.1 VanVleck-Pauli-Morette formula . . . . . . . . . . . . . . 3.1.2 Mathematical Appendix 1 . . . . . . . . . . . . . . . . . . 3.1.3 Mathematical Appendix 2 . . . . . . . . . . . . . . . . . . 3.2 The ﬁxed energy propagator . . . . . . . . . . . . . . . . . . . . . 3.2.1 General properties of the ﬁxed energy propagator . . . . . 3.2.2 Semiclassical computation of K(E) . . . . . . . . . . . . . 3.2.3 Two applications: reﬂection and tunneling through a barrier 3

4 3.2.4 3.2.5

CONTENTS On the phase of the prefactor of K(xf , tf ; xi , ti ) . . . . . On the phase of the prefactor of K(E; xf , xi ) . . . . . . . 68 72

4 Instantons 75 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 4.2 Instantons in the double well potential . . . . . . . . . . . . . . . 77 4.2.1 The multi-instanton amplitude . . . . . . . . . . . . . . . 81 4.2.2 Summing up the instanton contributions: results . . . . . 84 4.2.3 Cleaning up details: the zero-mode and the computation of R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 5 Interaction with an external electromagnetic ﬁeld 5.1 Gauge freedom . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 Classical physics . . . . . . . . . . . . . . . . . . . . . . . 5.1.2 Quantum physics . . . . . . . . . . . . . . . . . . . . . . . 5.2 Particle in a constant magnetic ﬁeld . . . . . . . . . . . . . . . . 5.2.1 An exercise on translations . . . . . . . . . . . . . . . . . 5.2.2 Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.3 Landau levels . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 The Aharonov-Bohm eﬀect . . . . . . . . . . . . . . . . . . . . . 5.3.1 Energy levels . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Dirac’s magnetic monopole . . . . . . . . . . . . . . . . . . . . . 5.4.1 On coupling constants, and why monopoles are diﬀerent than ordinary charged particles . . . . . . . . . . . . . . . 91 91 91 93 95 96 97 99 101 105 106 107

Chapter 1

**The path integral formalism
**

1.1

1.1.1

**Introducing the path integrals
**

The double slit experiment

One of the important experiments that show the fundamental diﬀerence between Quantum and Classical Mechanics is the double slit experiment. It is interesting with respect to the path integral formalism because it leads to a conceptual motivation for introducing it. Consider a source S of approximately monoenergetic particles, electrons for instance, placed at position A. The ﬂux of electrons is measured on a screen C facing the source. Imagine now placing a third screen in between the others, with two slits on it, which can be opened or closed (see ﬁgure 1.1). When the ﬁrst slit is open and the second closed we measure a ﬂux F1 , when the ﬁrst slit is closed and the second open we measure a ﬂux F2 and when both slits are open we measure a ﬂux F .

1 S 2

A

B

C

Figure 1.1: The double slit experiment: an electron source is located somewhere on A, and a detector is located on the screen C. A screen B with two slits 1 and 2 that can be open or closed is placed in between, so that the electrons have to pass through 1 or 2 to go from A to C. We measure the electron ﬂux on the screen C. In classical physics, the ﬂuxes at C in the three cases are expected to satisfy the relation F = F1 + F2 . In reality, one ﬁnds in general F = F1 + F2 + Fint , and the structure of Fint precisely corresponds to the interference between two 5

6

CHAPTER 1. THE PATH INTEGRAL FORMALISM

**waves passing respectively through 1 and 2: F = |Φ1 + Φ2 | = |Φ1 | + |Φ2 | + 2 Re (Φ∗ Φ2 ) 1
**

F1 F2 Fint 2 2 2

(1.1)

How can we interpret this result? What is the electron? More precisely: Does the wave behaviour imply that the electron is a delocalized object? That the electron is passing through both slits? Actually not. When detected, the electron is point-like, and more remarkably, if we try to detect where it went through, we ﬁnd that it either goes through 1 or 2 (this is done by setting detectors at 1 and 2 and considering a very weak ﬂux, in order to make the probability of coinciding detections at 1 and 2 arbitrarily small). According to the Copenhagen interpretation, F should be interpreted as a probability density. In practice this means: compute the amplitude Φ as if 2 dealing with waves, and interpret the intensity |Φ| as a probability density for a point-like particle position. How is the particle/wave duality not contradictory? The answer is in the indetermination principle. In the case at hand: when we try to detect which alternative route the electron took, we also destroy interference. Thus, another formulation of the indetermination principle is: Any determination of the alternative taken by a process capable of following more than one alternative destroys the interference between the alternatives. Resuming: • What adds up is the amplitude Φ and not the probability density itself. • The diﬀerence between classical and quantum composition of probabilities is given by the interference between classically distinct trajectories. In the standard approach to Quantum Mechanics, the probability amplitude is determined by the Schr¨dinger equation. The “Schr¨dinger” viewpoint o o somehow emphasizes the wave properties of the particles (electrons, photons, . . . ). On the other hand, we know that particles, while they are described by “a probability wave”, are indeed point-like discrete entities. As an example, in the double slit experiment, one can measure where the electron went through (at the price of destroying quantum interference). This course will present an alternative, but fully equivalent, method to compute the probability amplitude. In this method, the role of the trajectory of a point-like particle will be formally “resurrected”, but in a way which is compatible with the indetermination principle. This is the path integral approach to Quantum Mechanics. How can one illustrate the basic idea underlying this approach?

1.1.2

An intuitive approach to the path integral formalism

In the double slit experiment, we get two interfering alternatives for the path of electrons from A to C. The idea behind the path integral approach to Quantum Mechanics is to take the implications of the double slit experiment to its extreme consequences. One can imagine adding extra screens and drilling more and more

. Di+1 become inﬁnitesimally close.1.2).. xDM . where the screen B has been dropped in order to consider the simpler case of propagation in empty space). INTRODUCING THE PATH INTEGRALS 7 holes through them. . α D Nothing stops us from taking the ideal limit where N → ∞ and the holes ﬁll all of D. . .D2 . . xN (see ﬁgure 1. . α) D is then a purely ﬁctious device! We can go on and further reﬁne our trajectories by adding more and more ﬁctious screens D1 . α) In the limit in which Di . we have speciﬁed all possible paths x(y) (see ﬁgure 1. . Schematically: • With two slits: we know that Φ = Φ1 + Φ2 • If we open a third slit. generalizing the result of the double slit experiment by the superposition principle. Applying the superposition principle: N Φ= i=1 α=1. x2 . . 1 2 1 2 N −1 N A D 3 B C Figure 1.DM Φ= α=1. Φ= α=1. 3. The possible trajectories are now labelled D D D by xi and α = 1.2. .1.3 Φ xi .3 dxD Φ (xD . The sum i becomes now an integral over xD . .2. . xD2 .2. 2.3 dxD1 dxD2 · · · dxDM Φ (xD1 . the superposition principle still applies: Φ = Φ1 + Φ 2 + Φ3 • Imagine then adding an intermediate screen D with N holes at positions x1 . .3 for a schematic representation. This is the procedure illustrated by Feynman in his book “Quantum Mechanics and Path Integrals”.2: The multi-slit experiment: We have a screen with N slits and a screen with three slits placed between the source and the detector. that is by the slit they went through at D and by D the slit they went through at B.

2. That is also the dimensionality of the action S which describes the classical trajectories via the principle of least action. we want classical physics to be resurrected in the ℏ → 0 limit. yi ) to (xf . We want to describe the motion from position xi at time ti to position xf at time tf with a quantum probability amplitude K(xf .3 The path integral formulation Based on the previous section. tf . ti ) given by K(xf . xi . We then arrive at a formal representation of the probability amplitude as a sum over all possible trajectories: Φ= All trajectories {x(t).1.8 x x(yi ) CHAPTER 1. also the way the paths are covered through time is expected to matter (t → (x(y). x(tf ) = xf . In other words. Remarks: • ℏ has the dimensionality [Energy] × [Time]. tf . We want classical trajectories to describe the motion in the formal limit ℏ → 0. In fact. yf ) are labelled with the functions x(y) that satisfy x(yi ) = xi and x(yf ) = xf . xi . we will start anew to formulate quantum mechanics. y(t))). We have two guidelines: 1. . to be more precise. ti ) = All trajectories Φ({γ}) where {γ} is the set of all trajectories satisfying x(ti ) = xi . THE PATH INTEGRAL FORMALISM x(y1 ) x(y2 ) x(y3 ) x(yf ) y yi y1 y2 y3 yf Figure 1.2) How is this formula made sense of? How is it normalized? 1.3: The paths from (xi .y(t)} Φ({x}) (1.

so that all trajectories in a tube centered around γcl add up coherently (see ﬁgure 1. think of the integral −∞ dxeif (x) where f (x) ≡ ax plays the role of S/ℏ: the integral vanishes whenever the derivative of the exponent f ′ = a is non-zero 1 By . S varies very little. with f such that the classical trajectory is selected in the formal limit ℏ → 0.4) seems promising for two reasons: 1. The guideline (2. INTRODUCING THE PATH INTEGRALS 9 • One can associate a value of S[γ] to each trajectory. γ1 xf γ2 xi Figure 1. Therefore in the neighbourhood of γcl .5) in the sum over trajectories. In a macroscopic. at a classical trajectory γcl the action S[γ] is stationary.3) ei S[γ] ℏ (1. 1 (see ﬁgure 1.1. xi . tf .1. More precisely: the tube of trajectories R +∞ analogy. It is thus natural to guess: Φ[γ] = f (S[γ]/ℏ). xi γcl xf Figure 1. situation the gradient δS/δγ is for most trajectories much greater than ℏ. The classical trajectories are given by the stationary points of S[γ] (δS[γ] = 0).4: The contributions from the two neighbouring trajectories γ1 and γ2 will tend to cancel if their action is big. classical. On the other hand. The speciﬁc choice Φ[γ] = ei implying K(xf . ti ) = {γ} S[γ] ℏ (1.5: The contributions from the neighbouring trajectories of the classical trajectory will dominate.) is heuristically seen to hold. Around such trajectories the phase eiS/ℏ oscillates extremely rapidly and the sum over neighbouring trajectories will tend to cancel.4).

Eq. THE PATH INTEGRAL FORMALISM in question consists of those for which |S − Scl | ≤ ℏ and deﬁnes the extent to which the classical trajectory is well deﬁned.6: The path from (xi .6. tf ) can be obtained by summing the paths γ1 from (xi . tint ). 1. tint . ti ) to (y. this eﬀect becomes dramatic and only the classical trajectory survives. We cannot expect to deﬁne our classical action to better than ∼ ℏ. . y. ti ) for any y ). ti ) to (xf . (1. ti ) should be obtainable by evolving in two steps: ﬁrst from ti to tint (by K(y. Once again. tf . However. In the exact limit ℏ → 0. and second from tint to tf (by K(xf . in normal macroscopic situations Scl ≫ ℏ.3) leads to a crucial composition property. which is dominated by the region x2 < h around the ∼ stationary point x = 0. obviously each path with x(ti ) = xi and x(tf ) = xf can be obtained by joining two paths γ1 (y = x(tint )) from ti to tint and γ2 (y = x(tint )) from tint to tf (see ﬁgure 1. tf . The amplitude K(xf . like in ﬁg. Thanks to eq. tint ) with γ2 from (y. Now. xi . x γ1 y xi γ2 xf t ti tint tf Figure 1.3) the additivity of S translates into a factorization property for the amplitude: Φ[γ12 ] = Φ[γ1 ]Φ[γ2 ] which in turn leads to a composition property of K.10 CHAPTER 1. 2. xi . tint ) to (xf . tf ).6). Consider indeed the amplitudes for three consecutive times ti < tint < tf . Indeed the action for a path γ12 obtained by joining two subsequent paths γ1 and γ2 . a simple one dimensional analogy is provided by the inte2 +∞ gral −∞ dxeix /h . satisﬁes the simple additive relation S[γ12 ] = S[γ1 ]+S[γ2 ]. which we shall now prove. (1.

One interesting aspect of quantization is that a fundamental unit of action (ℏ) is introduced. we expect the trajectories close to γcl to dominate K Scl (semiclassical limit). tf . We expect to have: K(xf .1. q) → Sλ ≡ λS(q. tint . y. xi ) ∼ (smooth function) · ei ℏ . but it would not even closely reproduce the composition property we know works in Quantum Mechanics. would imply that the amplitudes for each individual particle do not factorize even in the limit in which they are very far apart and non-interacting!). In classical physics the overall value of the action in unphysical: if S(q.3). if S were to represent the action for a system of two particles. (1. .1.7) the only eﬀect is to multiply all the equations of motion by λ. For instance the alternative choice Φ = e−S[γ]/ℏ satisﬁes the composition property but does not in general select the classical trajectories for ℏ → 0. would perhaps work out in selecting the classical trajectories for ℏ → 0. q) ˙ ˙ (1. this is not an easy mathematical task. INTRODUCING THE PATH INTEGRALS Thanks to eq.5) Notice that the above composition rule is satisﬁed in Quantum Mechanics as easily seen in the usual formalism. xi . Another 2 alternative Φ = ei(S[γ]/ℏ) . It is instructive to appreciate what would go wrong if we modiﬁed the choice in eq. tint )K(y. ti ) . the system will behave diﬀerently: • large S → classical regime • small S → quantum regime In the large S limit. (1. tf . (More badly: this particular choice. It is the quantum analogue of the classical composition of probabilities P1→2 = α P1→α Pα→2 . (1.3) we can then write: dyK(xf .γ2 11 dye ℏ (S[γ1 (y)]+S[γ2 (y)]) dye ℏ S[γ(y)] γ(y)=γ2 (y)◦γ1 (y) i i = = K(xf . Quantum Mechanics sets a natural unit of measure for S. so that the solutions remain the same (the stationary points of S and Sλ coincide). ti ) = γ1 . (1. xi . In the S ∼ ℏ limit. Depending on the size of S. This alternative choice would select the minima of S but the classical trajectories represent in general only saddle points of S in function space. all trajectories are comparably important: we must sum them up in a consistent way.6) In quantum mechanics what is composed is not probability P itself but the amplitude K which is related to probability by P = |K|2 .

13) .4 From the Schr¨edinger approach to the path integral o iH(tf −ti ) ℏ iHt ℏ Consider the transition amplitude: K(xf .12 CHAPTER 1. We label the intermediate times tk = kǫ by the integer k = 0. o 1.12) x′ −x ǫ x |e ′ |x ≃ ǫ dp −i ℏ e 2πℏ h p2 2m +V i (x)−px ˙ (1. tf − ti = t.7). we can write: x h p2 2m +V h p2 ˆ 2m +V (ˆ) x i |x = e (x) i p|x + O ǫ2 (1. We will now assume.1. rather than going on with Feynman’s construction and show that it leads to the same results of Schr¨dinger equation. o we will follow the opposite route which is easier: we will derive the path integral formula from Schr¨dinger’s operator approach. . We can use the completeness relation xf | e− iHt ℏ |x x| dx = 1 at each step nǫ: |xN −1 xN −1 | e− x2 | e− iHǫ ℏ iHǫ ℏ |xi = N −1 k=1 dxk xf | e− iHǫ ℏ |xN −2 · · · iHǫ ℏ |x1 x1 | e− |xi (1. tf . ti ) ≡ xf | e− |xi = xf | e− |xi (1.9) Consider the quantity: x′ | e−iHǫ/ℏ |x . Later on we shall come back on this issue and better motivate our neglect of these terms. xi . we divide t into N inﬁnitesimal steps and consider the amplitude for each inﬁnitesimal step (see ﬁgure 1. Using x′ | e− iHǫ ℏ |p p| dp = 1: iHǫ ℏ |x = dp x′ |p p| e− p2 ˆ 2m ǫ −i ℏ |x (1. We then get: x′ | e− We can deﬁne: iHǫ ℏ |x ≃ ≡ x: ˙ − iHǫ ℏ dp e ǫ −i ℏ h p2 2m +V (x) i · e ℏ p(x −x) 2πℏ i ′ (1. .8) where we used time translation invariance to set: ti = 0. THE PATH INTEGRAL FORMALISM Due to the mathematical diﬃculty. 0 ǫ 2ǫ (N − 1)ǫ t = Nǫ Figure 1. Notice that t0 = 0 and tN = t. which seems fully reasonable.11) where the O(ǫ2 ) terms arise from the non-vanishing commutator between p2 /(2m) ˆ and V (ˆ). To write it in the form of a path integral. that in the limit x ǫ → 0 these higher orer terms can be neglected.7: The interval between 0 and t is divided into N steps. . . N .10) If we stick to the simple case H = p| e ǫ −i ℏ + V (ˆ).

1. (1. t. 0) = D[x(t)]ei Scl (x. [T. . Dirac’s motivation was apparently to formulate QM starting from the lagrangian rather than from the hamiltonian formulation of classical mechanics. following a hint from an earlier paper by Dirac.x) ˙ ℏ (1. we can then write p| eǫ(T +U) |x = = p| eǫT e−ǫT eǫ(T +U) e−ǫU eǫU |x = p| eǫT e−ǫ e ǫ −i ℏ 2 C ǫU e h 2m +V (x) p2 i |x1. as a series of commutators between T and U 1 ǫ C = [T.1. U ] + {[T.14) (1. To simplify the p2 notation let us denote the kinetic operator as T = −i 2mℏ and the potential U = −iV /ℏ.16) = lim 1 ǫ→0 A R N −1 k=1 Dγ dxk i S(xf . xi .17) We see that we (actually somebody else before us!) had guessed well the form of the transition amplitude.19) p| e−ǫ 2 C |x where C is given. By considering all the intervals we thus ﬁnally get: ˙ 0 xf | e− iHt ℏ |xi = lim N −1 k=1 ǫ→0 dxk 1 i PN −1 S(xl+1 .xi ) ℏ e A where Dγ should be taken as a deﬁnition of the functional measure over the space of the trajectories.11).18) ( (1.s. .x) ℏ 2πiℏǫ A h i (1.21) . We thus have got a path integral formula for the transition amplitude: K(xf . V (ˆ)]] = −iV ′ (ˆ) and expanding p x x the exponent in a series one can then write p| e−ǫ 2 n=∞ r=n s=r C |x = 1 + ǫ ǫn pr−s Pn. x) = ˙ ǫ L(x.20) 2 6 By iterating the basic commutation relation [ˆ. U ]] + [U. INTRODUCING THE PATH INTEGRALS 13 By performing the change of variables p ≡ p − mx the integral reduces to ¯ ˙ simple gaussian integral for the variable p: ¯ x′ | e− iHǫ ℏ |x ≃ = p2 ¯ x ǫ d¯ −i ℏ 2m +V (x)− m2˙ 2 p e 2πℏ ǫ 1 m 1 ˙2 ˙ ǫ ei ℏ [ 2 mx −V (x)] = eiL(x. T ]]} + .r (x) n=0 r=1 s=0 (1. Let us now come back to the neglected terms in eq. x)ǫ with the action S(x′ . (1.15) 1 A At leading order in ǫ we can further identify L(x. The path integral approach to QM was developed by Richard Feyman in his PhD Thesis in the mid 40’s. by using the Campbell-Baker-Haussdorf formula twice.xl ) e ℏ l=0 AN (1. x)dt. [U.

r (x) is a homogenous polynomial of degree n + 1 − r in V and its derivatives. xc )dt ˙ (1. Finally. t) . it is worth assessing the rˆle of the path integral (P.1 The properties of the path integrals Path integrals and state evolution To get an estimate of the dependence of the amplitude K(xf . after overcoming a few technical diﬃculties. (1. showing that p diverges in the small ǫ limit.0. Let us compute ∂tf Scl . Nonetheless. if the derivatives of V are all bounded. no technical proﬁt is really given by path integrals. ti ) on its arguments. tf .. tf . . xf . is however o technically more involved. This is because the leading contribution to the p integral in eq. charge quantization in the presence of a magnetic monopole. ti . satisfying the Euler-Lagrange equation: ∂t ∂L ∂L − ∂x ˙ ∂x = 0. let us ﬁrst look at the properties of the solutions to the classical equations of motion with boundary conditions xc (ti ) = xi . (1.2 1. xi . Where Scl is deﬁned tf Scl ≡ S[xc ] = ti L(xc . When working on simple quantum systems like the hydrogen atom. For instance.s. even taking into account that there are N ∼ 1/ǫ such terms (one for each step). x=xc (1.) in quantum mechanics.21) is ∼ 1 + O(ǫ3/2 ) so that.I. (1. the only potential problem to concentrate on in the ǫ → 0 limit is represented by the powers of p. Similarly. 1.1 (x) = V ′ (x).2. The P. We can now test our result under some simple assumption. the ﬁnal result is still convergent to 1 ǫ→0 lim 1 + aǫ 2 3 1 ǫ = 1. xc (tf ) = xf .15) comes from √ the region p ∼ 1/ ǫ. etc. By using √ p ∼ 1/ ǫ the right hand side of eq.24) We can think of xc as a function xc ≡ f (xi . For instance P1. As it was hopefully o highlighted in the discussion above. the path integral formulation is conceptually advantageous over the standard operatorial formulation of Quantum Mechanics. the path integral oﬀers a much more direct viewpoint on the semiclassical limit.14 CHAPTER 1. path integrals oﬀer a much better viewpoint.I.22) Before proceeding with technical developments.But that is for another course. with each term involving exactly r + s derivatives. THE PATH INTEGRAL FORMALISM where Pn. for advanced issues like the quantization of gauge theories and for eﬀects like instantons in quantum ﬁeld theory it would be hard to think how to proceed without path integrals!. . for issues involving topology like the origin of Bose and Fermi statistics.23) with xc a solution.. in that the “good old” particle trajectories retain some rˆle. the Aharonov-Bohm eﬀect.

we can now compute: tf ∂tf S = L(x. evaluated at x = xc and t = tf . ti . ti .2.31) Evaluated at x = xc .27) (1. tf . xf .1.32) We recognize here the deﬁnition of the momentum of the system.26) And the following properties are straightforward: ∂xf xc ∂xf xc t=tf t=ti =1 =0 (1. evaluated at x = xc and t = tf .25) Similarly. xc ) ˙ = + ∂L ∂x ˙ ∂L ∂x ˙ t=tf t=tf ∂tf xc x=xc t=ti L(xc . this gives: ∂xf Scl (xc ) = ∂L ∂x ˙ t=tf ∂xf xc x=xc t=ti = ∂L ∂x ˙ =P x=xc .30) We recognize here the deﬁnition of the hamiltonian of the system. THE PROPERTIES OF THE PATH INTEGRALS such that f (xi . diﬀerentiating the relation at t = ti we obtain ∂tf xc t=ti =0 (1. We can also compute: tf ∂xf S = ti dt ∂t ∂L ∂x x ∂x f ˙ tf − dt ∂t ti ∂L ∂L − ∂x ˙ ∂x ∂xf x (1. xf . t = ti ) ≡ xi = const f (xi . tf . xc ) − ˙ xc ˙ x=xc t=tf = −E (1. . x) ˙ t=tf + ti tf dt ∂L ∂L ∂t x + ∂t x ˙ ∂x f ∂x f ˙ ∂L ∂t x ∂x f ˙ tf t=tf + ti dt ∂t − dt ∂t ti ∂L ∂L − ∂x ˙ ∂x ∂tf x (1. x) ˙ = L(x.28) Using these properties.t=tf (1.29) All this evaluated at x = xc gives: ∂tf Scl = L(xc . t = tf ) ≡ xf = const Diﬀerentiating the last relation we deduce: ∂tf xc + ∂t xc =⇒ ∂tf xc t=tf t=tf 15 =0 t=tf = −xc ˙ (1.

t.42) where H(x′ .16 CHAPTER 1.39) (1. x. ti ) = F (xf . t′ . From the point of view of the Schr¨dinger picture.x ei 1 ℏ ∂t. we can compare these results with those of the path integral computation in the limit ℏ → 0.36) so that we obtain −iℏ∂xf K = ∂xf Scl · K + O (ℏ) = Pcl K + O (ℏ) (1. x.41) By the superposition principle. x. ti ) ei smooth function Scl ℏ (1. . tf . x′ as variables: o iℏ∂t′ K = dx′′ H(x′ . t) = x′ | e− ℏ |x is also known as the propagator.33) In the limit ℏ we expect the path integral to be dominated by the classical trajectory and thus to take the form (and we shall technically demostrate this expectation later on) K(xf . THE PATH INTEGRAL FORMALISM At this point. as it describes the probability amplitude for propagating a particle from x to x′ in a time t′ − t.x F ∼ O (1) Scl ℏ ∼ large comparatively negligible (1. xi . x′′ )K(x′′ . t) solves the Schr¨dinger equation for t′ . K(x′ . x) ≡ x| e− iHt ℏ |ψ0 ≡ x|ψ(t) (1. when going to the path integral description of particle propagation. t′ . The last two equations make contact with usual quantum mechanical relations. x. We then have: ∂t. xi . tf .34) where F is a smooth function in the ℏ → 0 limit. if we interpret K has the wave function and −iℏ∂x and iℏ∂t as respectively the momentum and energy operators −iℏ∂x K ∼ P · K iH(t′ −t) (1. K(xf . K can then be used to write the most general solution. 0) represents o the wave function at time t for a particle that was in the state |x at time 0: ψ(t. x′′ ) = x′ | H |x′′ . become associated to respectively the wave number (∼ (i∂x ln K)−1 ) and oscillation frequency (∼ (i∂t ln K)). xi . Indeed: 1. ti ) = D[x(t)]ei ℏ S (1. K(x′ . tf . t′ .37) (1.40) iℏ∂t K ∼ E · K The function K(x′ . t) (1.38) iℏ∂tf K = −∂tf Scl · K + O (ℏ) = Ecl K + O (ℏ) Momentum and energy.35) (1.

1. THE PROPERTIES OF THE PATH INTEGRALS 2.43) x| e iH(t−t0 ) − ℏ K(x. t) = x′ |x = δ(x′ − x) Thus. the solution at t > t0 is: Ψ(x. t). ti ) = lim where: ǫ= S= tf − ti N N −1 k=0 1 ǫ→0 A N −1 k=1 dxk iS eℏ A (1. t) contains all the relevant information on the dynamics of the system. decribed by the lagrangian L(x. This coordinate only appears only in the ﬁrst and the second time slice so we need just to concentrate on the integral ∞ −∞ ∞ −∞ dx1 ei 2ǫℏ [(x1 −x0 ) m 2 +(x2 −x1 )2 ] = = dx1 e h i 2 m i 2ǫℏ 2(x1 − 1 (x0 +x2 )) + 1 (x2 −x0 )2 2 2 2πiǫℏ 1 i m 1 (x2 −x0 )2 e 2ǫℏ 2 m 2 √1 A 2 (1. 1.44) by integrating ﬁrst over x1 . t) = x| e− = = iH(t−t0 ) ℏ |Ψ0 |y y|Ψ0 dy (1. t′ . x) = mx2 /2. t0 )Ψ0 (y)dy • Having K(x′ . the solution to the Schr¨dinger equation is found by o performing an integral.2. • K(x′ . t′ →t 17 lim K(x′ . y. Putting this result back into eq.2. using Feyman’s time slicing procedure. Using the result of ˙ ˙ section 1. t′ . (1. (1.44) 1 (xk+1 − xk )2 m 2 ǫ x0 = xi xN = xf Let us compute eq. t. x.2 The path integral computation for a free particle Let us compute the propagator of a free particle. t′ .4 we can write K(xf .45) Notice that x1 has disappeared: we “integrated it out” in the physics jargon.1. x. x. tf . xi . given Ψ0 (x) the wave function at time t0 .44) we notice that we have an expression .

The relevant terms are ∞ −∞ ∞ −∞ dx2 ei 2ǫℏ [ 2 (x2 −x0 ) m 1 2 +(x3 −x2 )2 ] 2 1 + 3 (x3 −x0 )2 = = dx2 e m i 2ǫℏ h 3 2 ( x2 − 1 x0 − 2 x3 ) 3 3 i 2πiǫℏ 2 i m 1 (x3 −x0 )2 e 2ǫℏ 3 m 3 √2 A 3 (1.49) .0) ℏ = F (t)ei (1. we ﬁnd the expression of the propagator for a free particle: K(xf .. It is now easy to go on.46) It is now clear how to proceed by induction. putting all together.t.18 CHAPTER 1. xi .45) are checked to match this interpratation. THE PATH INTEGRAL FORMALISM similar to the original one but with N − 1 instead of N time slices.xi . Moreover the ﬁrst slice has now a width 2ǫ instead of ǫ: the factors of 1/2 in both the exponent and prefactor of eq.48) where we used the fact that N ǫ = tf − ti . xi . At the n-th step the integral over xn will be: dxn ei 2ǫℏ [ n (xn −x0 ) m 1 2 +(xn+1 −xn )2 ] i = = dxn e m i 2ǫℏ h n+1 n 1 n 1 (xn −( n+1 x0 + n+1 xn+1 ))2 + n+1 (xn+1 −x0 )2 2 1 m 2πiǫℏ n ei 2ǫℏ n+1 (xn+1 −x0 ) m n+1 √ n (1. ti ) = lim = 1 A 2 m N − 1 i m (xf −xi )2 12 1 . 0) = = xf |p p| e− p iHt ℏ |xi dp p2 t 1 2πℏ 1 = 2πℏ ei ℏ (xf −xi ) ei 2mℏ dp e “ ”2 x −x it m − 2mℏ p−m f t i +i 2ℏ (xf −xi )2 t dp = m i m(xf −xi )2 2ℏt e 2πiℏt Scl (xf . We can check this result by computing directly the propagator of a onedimensional free particle using the standard Hilbert space representation of the propagator: K(xf . (1. Let us do the integral on x2 to gain more conﬁdence.. e 2ℏN ǫ = lim √ ei 2ℏN ǫ (xf −xi ) ǫ→0 A N 23 N 2 ǫ→0 (x −x ) m im f i e 2ℏ tf −ti 2πiℏ(tf − ti ) (1.47) A n+1 Thus. t. tf .

Thus. 0) satisﬁes ψ(x. as for the free particle the result factorizes.3. and with λ real and positive we have: dxe−λx = 2 1 λ dye−y = 2 π λ (1.52) • With one variable.50). Before going to that. let us remind ourselves some basic results about “Gaussian” integrals in several variables.50) Let us interpret this result in terms of a ﬂux of particles that started at xi = 0 at t = 0 with a distribution of momenta dn(p) = f (p)dp (1. .54) dxi e− i.j λij xi xj i where λij is real. dn(p) = (1/2πℏ)dp ⇒ dn/dp ∼ [Length]−1 ×[Momentum]−1 .1 Gaussian integrals Let us give a look at gaussian integrals: dxi e− i P i.3 Path integrals as determinants We will now introduce a systematic small ℏ expansion for path integrals and show that the computation of the “leading” term in the expansion corresponds to the computation of the determinant of an operator. t) = K(x. we ﬁnd f = 1/(2πℏ) = const. • It is instructive interpret the result for the probability distribution: m dP = |K|2 = dx 2πℏt (1. t. Comparing to (1. PATH INTEGRALS AS DETERMINANTS 19 These results trivially generalize to the motion in higher dimensional spaces.53) • Consider now a Gaussian integral with an arbitrary number of real variables P (1. A particle with momentum p at time t will have travelled to x = (p/m)t. 1. Thus the particles in the interval dp will be at time t in the coordinate interval dx = (t/m)dp.j λij xi xj (1. 1.1.3. Therefore we have dn(x) = f (xm/t)(m/t)dx. 0) = δ(x) which corresponds in momentum space to an exactly ﬂat distribution of momenta. • As we already saw. we can perform the same integral “` la Feynman” by a dividing t into N inﬁnitesimal steps.51) and ﬁnd what f (p) is. 0. We recover the result that the wave function ψ(x. with normalization dn(p)/dp = (1/2πℏ).

which we indeed already encountered before I= dxeiλx = ei sgn(λ) 4 2 π π |λ| (1. 1.59) L(x(t). Depending on the sign of λ we must choose diﬀerent contours to ensure convergence at inﬁnity. Notice that the phase depends on the sign of λ. Then. . . i dxi = |det O| i d˜i = i d˜i . we can choose the integration variables that suit us best. ti ) = with S[x(t)] = ti tf (1. • The simplest change of variables is to shift x(t) to “center” it around the classical solution xc (t): x(t) = xc (t) + y(t) ⇒ D[x(t)] = D[y(t)] (the Jacobian for this change of variables is obviously trivial. • Along the same line we can consider integrals with imaginary exponents.20 CHAPTER 1. tf .56) This integral is performed by deforming the contour into the imaginary plane as discussed in homework 1.j λij xi xj = i d˜i e− x P i λi xi2 ˜ = π = λi 1 1ˆ det( π λ) (1. • For the case of an integral with several variables I= i dxi ei P j. ˜ x x OT O = OOT = 1.58) where n+ and n− are respectively the number of positive and negative eigenvalues of the matrix λjk .k λjk xj xk (1. xi .55) i where again the result makes sense only when all eigenvalues λn are positive. . . THE PATH INTEGRAL FORMALISM We can rotate to the eigenbasis: xi = Oij xj . Thus: dxi e− i P i.60) • To perform the integral. x(t))dt ˙ (1.2 Gaussian Path Integrals D[x(t)]ei S[x(t)] ℏ Let us now make contact with the path integral: K(xf . as we indicated). λn ).3.57) the result is generalized to I = ei(n+ −n− ) 4 π 1 1ˆ | det( π λ)| (1. with OT λO = diag (λ1 . .

xi .64) Thus. In our case S = dtL(x.66) where the overall constant N is just the Jacobian of the rescaling. which gives rise to the exponent eiS[xc ]/ℏ . tf . the classical trajectory xc (t). ti ) = N · e ℏ S[xc (t)] · i ˜ D[˜(t)]e 2 δx2 y y i δ2 S 2 √ ˜ +O ( ℏ y 3 ) (1. PATH INTEGRALS AS DETERMINANTS 21 • We can Taylor expand S[xc (t) + y(t)] in y(t) (note that y(ti ) = y(tf ) = 0. the Taylor expansion of a functional reads in general S[xc (t) + y(t)] = S[xc ] + + 1 2 dt1 δS δx(t1 ) y(t1 ) x=xc (1.1. The linear term in the expansion vanishes by the equations of motion: δS δx y(t)dt = x=xc d ∂L ∂L − ∂x dt ∂ x ˙ y(t)dt + x=xc ∂L y(t) ∂x ˙ t=tf =0 t=ti (1.65) where δ 2 S/δx2 is just short hand for the second variation of S shown in eq.63) and so on. 2 The interpretation of the above expression is that the quantum propagation of a particle can be decomposed into two pieces: 1. xi .61) dt1 dt2 δ2S δx(t1 )δx(t2 ) x=xc y(t1 )y(t2 ) + O y 3 with obvious generalization to all orders in y. 2 Given that N is a constant its role is only to ﬁx the right overall normalization of the propagator. √ We can rewrite it rescaling our variables y = ℏ y : ˜ K(xf . we obtain the result: K(xf . By the usual deﬁnition of functional derivative. Therefore it does not play a crucial role in the following discussion. and does not matter in assessing the relative importance of each trajectory.63). . x) ˙ so that each term in the above general expansion can be written as a single dt integral dt1 δS δx(t1 ) y(t1 ) ≡ dt ∂L ∂x y+ x=xc x=xc ∂L ∂x ˙ y ˙ x=xc (1.3. since the classical solution already satisﬁes the boundary conditions xc (ti ) = xi and xc (tf ) = xf ).62) 1 2 dt1 dt2 δ2S δx(t1 )δx(t2 ) dt ∂2L ∂x2 x=xc y(t1 )y(t2 ) ≡ ∂2L ∂x∂ x ˙ yy + ˙ x=xc y2 + 2 x=xc ∂2L ∂ x2 ˙ y2 ˙ x=xc (1. tf . (1. ti ) = D[y(t)]e ℏ i h S[xc (t)]+ 1 2 δ2 S 2 y +O δx2 (y3 ) i (1.

Scl = 1 1 (xf − xi )2 = mv∆x = 5000 erg · s m 2 tf − ti 2 ℏ = 1. The leading ℏ1/2 correction vanishes because the integrand is odd under y → −˜ ˜ y √ 3 i ℏ δ S 3 2 δ2 S y 2 (1. In order to do so we must expand the action to order y 4 around the classical solution (using the same short hand notation as ˜ before for the higher order variation of S) √ √ 3 y ℏδ S 3 S[xc + ℏ˜] ℏ δ4 S 4 S[xc ] 1 δ 2 S 2 y + ˜ y + ˜ y + O(ℏ3/2 ) (1. the ﬂuctuation y(t) over which we must integrate. The semiclassical limit should thus correspond to the possibility to reduce the path integral to a gaussian integral. ti )e ℏ S[xc (t)] [1 + O (ℏ)] (1. ti ) = = N · e ℏ S[xc (t)] · i i ˜ D[˜(t)]e 2 δx2 y [1 + O (ℏ)] = (1. the “classical” part of the propagator.68) Where the prefactor F is just the gaussian integral around the classical trajectory. tf . because of the quadratic √ ℏ (that term in the exponent. We will study this in more detail later on. THE PATH INTEGRAL FORMALISM 2. We will use the following unity for energy: [erg] = [g] × ([cm]/[s])2 . the path integral is dominated by y ∼ O is y ∼ O (1)) so that y(t) represents the quantum ﬂuctuation around the ˜ classical trajectory. while the prefactor (F and the terms in square brakets in eq. one can treat O(ℏ) in the exponent of the path integral as a perturbation K(xf . (1.68)) depend smoothly on ℏ in this limit. exp(iS[xc ]/ℏ) oscillates rapidly when the classical action changes.22 CHAPTER 1. As an example. In those physical situations where ℏ can be treated as a small quantity.3 O(ℏ) corrections to Gaussian approximation It is a good exercise to write the expression for the O(ℏ) correction to the propagator K in eq. tf . we can compute the value of Scl for a simple macroscopic system: a ball weighing one gram moves freely along one meter in one second.67) y i δ2 S 2 F (xf . xi .68). (1. xi .70) D[˜(t)] y y e δx2 ˜ = 0 ˜ 3 3! δx and for the O(ℏ) correction to K we ﬁnd that two terms contribute ∆K = iℏe i ℏ S[xc (t)] N D[˜(t)] y 1 δ4S 4 1 1 δ3S 3 y +i ˜ y ˜ 4! δx4 2 3! δx3 2 ˜ e 2 δx2 y i δ2 S 2 (1.0546 · 10−27 erg · s ⇒ Scl /ℏ ∼ 1030 1.69) ˜ = + 2 3 ℏ ℏ 2 δx 3! δx 4! δx4 and then expand the exponent in the path integral to order ℏ.71) . In the small ℏ limit.3.

78) We already have got a non trivial information by simple manipulations.xi . ti ) = ei = ei Scl ℏ · D[y]e ℏ i R tf ti m 2 ˙ (y2 −ω2 y2 )dt Scl (xf . Let us compute the propagator for the harmonic oscillator using the determinant method. 0. xi . . ti ) J(tf −ti ) (1. xc )dt = ˙ mω 2 xf 2 + xi2 cot(ωT ) − 2xf xi sin(ωT ) (1. This is a curious thing about path integrals: one never really ends up by computing them directly.75) where T = tf − ti .tf −ti ) ℏ · K(0.73) To make sense of the above we must deﬁne this strange “beast”: the determinant of an operator. ti ) = ei S[xc ] ℏ D[y]e ℏ 2 δx2 y i S[xc ] ℏ i 1 δ2 S 2 = const · e δ2S det 2 δx −1 2 (1.74) It is left as an exercise to show that: tf Scl = ti L(xc .72) For these lagrangians. It is best illustrated with explicit examples. the propagator is: K(xf .76) And thus. Then. xi . tf . x) = ˙ 1 1 mx2 − mω 2 x2 ˙ 2 2 (1. tf ti tf L(xc + y. We have the lagrangian: L(x.xi .4 Quadratic lagrangians and the harmonic oscillator Consider now a special case: the quadratic lagrangians. xi . tf .1.tf −ti ) ℏ · J(tf − ti ) (1. PATH INTEGRALS AS DETERMINANTS 23 1. tf . ti ) = ei Scl (xf . xc + y)dt = ˙ ˙ ti 1 m xc2 − ω 2 xc2 dt + ˙ 2 tf ti 1 m y 2 − ω 2 y 2 dt ˙ 2 (1. deﬁned by the property: δ (n) S =0 δxn for n > 2 (1.3. the gaussian integral coresponds to the exact result: K(xf . tf .77) This is because y(t) satisﬁes y(ti ) = y(tf ) = 0 and any reference to xf and xi has disappeared from the integral over y. Indeed J is just the propagator from xi = 0 to xf = 0: K(xf .3. The question now is how to compute J(tf − ti ). and usually computed indiretly using some “trick”.

tf . tf . and we get (T = tf − ti . x. t. let us compute the same quantity using the determinant method.0.80) The general solution to the above equation is J(T ) = mω eaT 2πiℏ sin(ωT ) (1. THE PATH INTEGRAL FORMALISM 1.83) i ti m 2 y − ω 2 y 2 dt = −i ˙ 2ℏ d2 dt2 tf ti m y 2ℏ i d2 + ω 2 y dt = dt2 2 tf ti ˆ y Oy dt (1. ti ) = dx K(xf . xi . Now. and thus the ﬁnal result −1/2 ˆ is proportional to det O . 0. T1 = tf − t. we get: J(tf − ti ) = K(0. What we want to compute is: J(T ) = with the boundary conditions: y(ti ) = y(tf ) = 0 We can note the property: tf D[y]e ℏ i R tf ti m 2 ˙ (y2 −ω2 y2 ) (1. xi .84) ˆ where O = − m ℏ +ω 2 .82) (1. but it already tells us a good deal about it. let us work in Fourier space for y: y(t) = n an yn (t) (1. we have to perform a gaussian integral.t−ti )) dx i (1.tf −t)+Scl (x. ti ) = J(tf − t)J(t − ti ) e ℏ (Scl (0. tf . Notice that in the limit ω → 0.24 CHAPTER 1. Trick number 1: use the composition property of the transition amplitude: K(xf .79) We can put the expression of the classical action in the integral. We will now compute J directly and ﬁnd a = 0. Formally. So this trick is not enough to fully ﬁx the propagator. To make it more explicit.85) .x. 2. T2 = t − ti ): J(T ) = J(T1 )J(T2 ) 2 mω dxei 2ℏ [x (cot(ωT1 )+cot(ωT2 ))] = 1 2 dxeiλx = 2 2iπℏ sin(ωT1 ) sin(ωT2 ) iπ = λ mω sin(ωT ) (1. t)K(x.81) with a an arbitrary constant. the result goes back to that for a free particle. ti ) Applying this to J(tf − ti ).

n n And thus. our computation must give the free 3 Indeed a stronger result for the independence of the Jacobian on the quadratic action holds. while N obviously does not: the modes yn 3 do not depend on ω . as we shall discuss in section 3. We should not get into any trouble if we avoid computing these two quantities separately. Their product however. The strategy is to work only with ratios that are well deﬁned.89) Notice that the number of negative eigenvalues is ﬁnite and given by n− = ω int( Tπ ). The yn form a complete basis of the Hilbert space L2 ([ti .1. we would have to consider all the Fourier modes with arbitrarily large n. that is the number of half periods of oscillation contained in T .1. ˆ ˜ Notice that O depends on ω. the only physically relevant quantity. In this limit both the product of eigenvalues and the Jacobian ˜ N are ill deﬁned.90) N 2πi n “Jacobian” √ where the 1/ 2πi factors are singled out for later convenience (and also to mimick the 1/A factors in our deﬁnition of the measure in eq.88) − ω2 (1. .16)). tf ]) mod y(ti ) = y(tf ) = 0. ill deﬁned −1 2 J(T ) = ˜ N n dan i Pn λn an2 √ = e2 2πi 1 −2 | n λn | i(n+ −n− ) 4 ˜ e N i(n +n ) π e + − 4 π well deﬁned = | n λn | π ˜ N e−in− 2 (1.92) Notice that in the continuum limit ǫ → 0 in the deﬁnition of the measure eq. For ω = 0.87) (1.16). That way we shall keep at large from mathematical diﬃculties (and confusion!).86) ˆ Oyn (t) = λn yn (t) yn ym dt = δnm The eigenvalues are λn = m ℏ nπ T 2 (1. and we can write: da ˜ √ n · D[y(t)] = (1. as we will now show. is well deﬁned.91) m. the an form a discrete set of integration variables. n ∈ N∗ 25 (1. We get: ˆ ˆ y Oy dt = am an ym Oyn dt = λn an2 (1.3. (1. Thus. PATH INTEGRALS AS DETERMINANTS ˆ where yn (t) are the othonormal eigenfunctions of O: yn = nπ 2 sin( t) T T . (1.

Consider {Ψn } a complete orthonormal basis of eigenvectors of the hamiltonian such that H |Ψn = En |Ψn .95) we get our ﬁnal result: Jω (T ) = J0 (T ) · π ωT e−in− 2 = | sin(ωT )| π mω e−in− 2 (1.75) the propagator is (assume for simplicity ωT < π) K(xf . xi . xi . So we have ˜ ˆ Jω (T ) = N | det Oω | ˜ ˆ J0 (T ) = N | det O0 | We can thus consider Jω (T ) in− π 2 = e J0 (T ) ˆ det O0 | | ˆ det Oω 1 2 1 −2 1 −2 e−in− 2 π (1. 0)dx = n |Ψn (x)|2 dx e− iEn T ℏ = n e− iEn T ℏ (1. ti ) = m. x.100) Let us now deﬁne the partition function: Z(t) = K(x.101) For the harmonic oscillator.98) 2πiℏ| sin(ωT )| So that by using eq.96) (1. (1.26 CHAPTER 1.n xf |Ψn Ψn | e− iH(tf −ti ) ℏ |Ψm Ψm |xi = iEn (tf −ti ) ℏ Ψ∗ (xi )Ψn (xf )e− n n (1. T .100) we ﬁnd: K(x. THE PATH INTEGRAL FORMALISM particle result.97) 1 λn (0) = λ (ω) n=1 1 − a n=1 n 2 a sin a By eq. The propagator can be written as: K(xf .95) 1 2 ˆ The eigenvalues of the operator Oω = − m ℏ functions y(t) have been given before: λn (0) = λn (ω) ∞ nπ 2 T nπ 2 − ω2 T ∞ nπ + ω2 over the space of = 1 1− = ωT 2 nπ = 1 1− a 2 nπ (1. 0) = 2xf xi mω i mω x 2 +xi2 ) cot(ωT )− sin(ωT ) e 2ℏ ( f 2πiℏ sin(ωT ) h i (1. (1. x. using eq. we can make contact with the solution of the eigenvalue problem and recover the well know result for the energy levels of the harmonic oscillator. tf .94) = n |λn (ω = 0)| | n |λn (ω) d2 dt2 1 2 = | n λn (0) | λn (ω) (1.99) At this point. T . T . which we already know.102) .93) (1. (1. 0)dx = 1 2i sin ωT 2 = e−i ωT 2 1 1− ωT e−2i 2 = ∞ n=0 e−i(2n+1) ωT 2 (1.

OPERATOR MATRIX ELEMENTS ⇒ En = ℏω n + 1 2 27 (1. ˆ We have the relation: xf .110) ˆx xf . ˆ ˆ Given a time the independent position eigenstates |x . tf | A(ˆ(t1 )) |xi .4 Operator matrix elements We will here derive the matrix elements of operators in the path integral formalism.109) Using the factorization property of the path integral we can derive the following identity: D[x]A(x(t1 ))ei S[x] ℏ = S[xa ]+S[xb ] ℏ D[xa ]D[xb ]dx1 A(x1 )ei dx1 xf | e− xf | e− xf | e − iHtf ℏ iH(tf −t1 ) ℏ = |xi A(x1 ) = |x1 x1 | e− iH(t1 −ti ) ℏ iH(t1 −ti ) ℏ iH(tf −t1 ) ℏ ˆx A(ˆ)e− ˆx A(ˆ(t1 ))e iHti ℏ |xi = (1. ˆ Given an operator O in the Schroedinger picture.104) In particular for the position operator we have x(t) ≡ eiHt xe−iHt . t ≡ eiHt |x (1.107) .4.4. the vectors |x. being eigenstates of x(t) ˆ x(t) |x. (1.105) (1. ti = D[x(t)]ei S[x(t)] ℏ (1. the time evolved Heisenberg picture operator is ˆ ˆ O(t) ≡ eiHt Oe−iHt .1 The time-ordered product of operators Let us ﬁrst recall the basic deﬁnition of quantities in the Heisenberg picture. tf |xi . ti |xi = . t = x |x.103) 1. (1. t .106) represent the basis vectors in the Heisenberg picture (notice the “+” in the exponent as opposed to the “−” in the evolution of the state vector in the Schroedinger picture!). 1. It deﬁnes an operator on the Hilbert space: ˆx A(ˆ) ≡ |x x| A(x)dx (1.108) By working out precisely the same algebra which we used in section Consider now a function A(x).1.

t2 O2 (x2 ) x2 . t2 |x1 . ti xf | e− iH(tf −t2 ) ℏ iH(t2 −t1 ) ℏ (1. t1 |xi .115) ˆ ˆ = xf .116) ˆ ˆ where the time-ordered product T [O1 (t1 )O2 (t2 )] is deﬁned as: ˆ ˆ ˆ ˆ ˆ ˆ T [O1 (t1 )O2 (t2 )] = θ(t2 − t1 )O2 (t2 )O1 (t1 ) + θ(t1 − t2 )O1 (t1 )O2 (t2 ) where θ(t) is the Heaviside step function.114) Recall that this result is only valid for t2 > t1 .8: The path from (xi . the ﬁnal result is: D[x]O1 (x(t1 ))O2 (x(t2 ))ei S[x] ℏ (1. t1 O1 (x1 ) x1 . Exercise: Check that for t1 > t2 .113) ˆ x O1 (ˆ)e− iH(t1 −ti ) ℏ |xi = (1. ti dx2 dx1 Using the equation (1. tf |x2 . it can be rewritten: ˆ x O2 (ˆ)e− ˆ ˆ xf .111) By the composition property. tf | O2 (t2 )O1 (t1 ) |xi . one gets instead: ˆ ˆ xf .28 CHAPTER 1. ti (1.112) t2 c tf Figure 1.109). tf ) is separated into three paths a. b and c. THE PATH INTEGRAL FORMALISM Let us now consider two functions O1 (x(t)) and O2 (x(t)) of x(t) and let us study the meaning of: D[x]O1 (x(t1 ))O2 (x(t2 ))ei S[x] ℏ (1. tf | O1 (t1 )O2 (t2 ) |xi . ti ) to (xf . We have to distinguish t1 < t2 from t2 < t1 .117) . tf | T [O1 (t1 )O2 (t2 )] |xi . ti Thus. (1.8): D[xa ]D[xb ]D[xc ]dx1 dx2 ei t1 ti a b S[xa ]+S[xb ]+S[xc ] ℏ O2 (x2 )O1 (x1 ) (1. we can equal it to (see ﬁgure 1. This can be rewritten: xf . assuming that t1 < t2 .

x . OPERATOR MATRIX ELEMENTS 2 29 ˙ Exercise: Consider a system with classical Lagrangian L = m x −V (x). Treat2 ing V as a perturbation and using the path integral. x) = eiH0 t/ℏ V (ˆ)e−iH0 t/ℏ .1. derive the well-known formula for the time evolution operator in the interaction picture i ˆ U (t) = e−iH0 t/ℏ T [e− ℏ Rt 0 ˆ V (t′ .ˆ)dt′ x ] (1.4.118) where H0 = p2 ˆ 2m ˆ ˆ and V (t.

THE PATH INTEGRAL FORMALISM .30 CHAPTER 1.

We will show a common application of both methods by ﬁnding the perturbative expansion of the free energy of the anharmonic oscillator. xi . 2. . x(tn )] |xi . x) + J(t)x(t) ˙ ˙ K(xf . to ﬁnd out the properties of a physical system (for example an atom). tf . ti |J) iδJ(t1 ) iδJ(tn ) = xf . statistical mechanics and semiclassical tunneling. This basic empirical fact has its formal counterpart in the theoretical study of (classical and) quantum systems. and study its response. ti x ˆ (2. x) and ˙ let us make it interact with an arbitrary external forcing source: L(x. ti |J) = D[x]e ℏ i R (L(x. . tf . .1 Functional method Experimentally. we let it interact with an external perturbation (a source. which is useful for.Chapter 2 Functional and Euclidean methods In this chapter.. To give an illustration. e. which allows a representations of perturbation theory in terms of Feynman diagrams. we will introduce two methods: • The functional method. e. an electromagnetic wave).1) K has now become a functional of J(t). tf | T [ˆ(t1 ). • The Euclidean path integral.2) 31 J=0 .x)+J(t)x(t))dt ˙ (2. xi . x) → L(x. .g. This functional contains important physical information about the system (all information indeed) as exempliﬁed by the fact that the functional derivatives of K give the expectation values of x (and thus of all operators that are a function of x): ˆ ℏδ ℏδ ··· K(xf .g. let us consider a system with Lagrangian L(x.

(2.32 CHAPTER 2. xi .2)) the propagator can be written as: K[J] = = = n i λ 4 D[x]e ℏ (L0 − 4! x +J(t)x(t))dt R D[x] n 1 n! −iλ ℏ4! dt n −iλ ℏ4! n x4 (t1 ) · · · x4 (tn )e ℏ i R (L0 +J(t)x(t))dt dt1 · · · dtn 1 n! ℏ4 δ 4 ℏ4 δ 4 ··· 4 K0 [J]dt1 · · · dtn 4 (t ) δJ 1 δJ (tn ) K0 [J] (2.1: The interval between 0 and β is divided into N steps.7) . 2 2 Lpert (x. β) = D[x]E e− ℏ SE [x] 1 (2.1) and write the imaginary time amplitude as a path integral.2 Euclidean Path Integral iHt ℏ Instead of computing the transition amplitude xf | e− computed the imaginary time evolution: KE ≡ xf | e− 0 ǫ 2ǫ βH ℏ |xi . Figure 2. let us see how the eﬀects of a perturbation can be treated with functional methods.4) Consider now the addition of the perturbation. we could have (2. The result is: KE (xf . we get the propagator of the anharmonic hamiltonian in terms of functional derivatives of the harmonic propagator in presence of an arbitrary source J. By repeated use of the functional derivatives with respect to J (see eq. It is straightforward to repeat the same time slicing procedure we employed in section 1. Let us look at the harmonic oscillator. x) = ˙ m 2 mω 2 2 x − ˙ x .1.4 (see ﬁgure 2.6) |xi .5) ≡ exp − iλℏ3 δ 4 4! δJ 4 (t) Thus. FUNCTIONAL AND EUCLIDEAN METHODS As an example. when we take the J → 0 limit. with an anharmonic self-interaction which is assumed to be small enough to be treated as a perturbation: L0 (x. β ∈ R+ β = Nǫ (N − 1)ǫ which corresponds to an imaginary time interval t = −iβ. x) = − ˙ λ 4 x 4! (2.3) The propagator K0 [J] for the harmonic oscillator with a source J(t) is: K0 [J] = D[x]e ℏ i R (L0 +J(t)x(t))dt (2. 2.

1. (2.12) . EUCLIDEAN PATH INTEGRAL where D[x]E is a path integration measure we shall derive below and where β 33 SE [x] = 0 m 2 dx dτ 2 + V (x) dτ ∼ βE (2.4 we can write: KE (xf . xi .x) ℏ ≡ e e ℏ 2πǫℏ AE » – “ ”2 m x′ −x +V (x) ǫ 2 ǫ (2.2. gets mapped by going to imaginary time into a metric with Euclidean signature ds2 = dτ 2 + dx2 . β) = xN | e− Hǫ ℏ |xN −1 · · · x1 | e− Hǫ ℏ |x0 dx1 · · · dxN −1 (2. +. xi .x′ )ǫ 1 − SE (x′ .9) Notice that more sloppily we could have also gotten KE by “analytic continuation” to imaginary time directly in the path integral: K(t) = D[x]e ℏ i Rt 0 L(x. As in section 1. −iβ) . corresponding to a Minkovsky metric i with signature −. Hence the ‘Euclidean’ suﬃx. +. Notice that the proper time interval in the relativistic description ds2 = −dt2 + dx2 . just as before: x′ | e− Hǫ ℏ |x = ≈ x′ |p p| e− ip Hǫ ℏ |x dp “ p2 2m +V 1 2πℏ 1 = 2πℏ = = eℏ ǫ (x′ −x)− ℏ (x) ” dp dp e− 2mℏ (p−i ǫ (x −x)) ǫ m ′ 2 ǫ m − 2ǫℏ (x′ −x)2 − ℏ V (x) 1 m −ℏ e 2πǫℏ m − 1 LE (x. β) = K(xf . xi . From a functional viewpoint KE is the continuation of K to the imaginary time axis KE (xf .10) Then. β] into N steps of equal length ǫ (see ﬁgure 2.11) where xf = xN and xi = x0 .x)dt′ ˙ (2. i E Let us now compute explicitly by decomposing the imaginary time interval [0.8) Notice that SE is normally minimized on stationary points.2. We can compute. +.1). with the following replacements t′ = −iτ dt′ = −idτ dx x=i ˙ dτ t = −iβ we obviously get the same result.

FUNCTIONAL AND EUCLIDEAN METHODS and then proceed to obtain the Euclidean analogue of eq. x. where kB is the Boltzmann constant.17) Thus. −iβ)dx = e−βω(n+ 2 ) 1 (2.2. Let us now check that K(−iβ) leads to the right partition function for the harmonic oscillator. x. one simply makes the change iǫ → ǫ in the factor A deﬁning the measure of integration.18) n which is indeed the expected result.102).1 Statistical mechanics Recalling the result of the previous section we have KE (xf . β) = n e−βEn /ℏ (2.16) We recognize here the statistical mechanical deﬁnition of the thermal partition function for a system at temperature kB T = ℏ/β. xi . Using equation (1. β) = n Ψ∗ (xf )Ψn (xi )e−βEn /ℏ n (2. xi . We can observe the equivalence: Z[β] = dxKE (x. β] with boundaries identiﬁed.xi ) ℏ e AE (2. the thermal partition function is equivalent to a functional integral over a compact Euclidean time: τ ∈ [0.13) 1 ǫ→0 AE R Synthetizing: to pass from the real time path integral formulation to the Euclidean time formulation.xl ) e ℏ l=0 AN E dxk − SE (xf . β) = {x(0)=x(β)} D[x]e−SE (2. x. The integration variables x(τ ) are therefore periodic in τ with period β.14) 2. . (1. We can then deﬁne the partition function: Z[β] = dxKE (x.16) KE (xf .15) where {Ψn } is an othonormal basis of eigenvectors of the Hamiltonian. β) = lim = lim N −1 k=1 N −1 k=1 D[x]E ǫ→0 dxk 1 − PN −1 SE (xl+1 .34 CHAPTER 2. and modiﬁes the integrand according to t iS = i 0 1 mx2 − V (x) dt −→ −SE = − ˙ 2 eiS/ℏ −→ e−SE /ℏ β 0 1 mx2 + V (x) dτ ˙ 2 (2. we get: Z[β] = K(x.

3 Perturbation theory The purpose of this section is to illustrate the use of functional methods in the euclidean path integral. We shall do so by focussing on the anharmonic oscillator and by computing in perturbation theory the following physically relevant quantities • the time ordered correlators on the ground state: 0|T [ˆ(τ1 ) . xi . Indeed.23) 1 + O(e−β(E1 −E0 )/ℏ ) .22) ψn (xf )ψn (xi )∗ e−βEn /ℏ ψ0 (xf )ψ0 (xi )∗ e−βE0 /ℏ 1 + O(e−β(E1 −E0 )ℏ ) = where E0 and E1 are the energies of respectively the ground state and the ﬁrst excited state. − ) 2 2 = = n β→∞ (2. This is most directly done by working in the operator formulation. ˆ ˆ It is interesting to consider the limit of the euclidean path integral and correlators when β → ∞. .3.20) xf |e− βH ℏ |xi = (2. . . . (1. . x(τn )]|0 x ˆ • the free energy and the ground state energy.19) By working out precisely the same algebra of section 1. .21) represents the continuation to imaginary time of the Heisenberg picture position operator.2. .1 Euclidean n-point correlators Consider the euclidean path integral deﬁned between initial time τi = − β and 2 ﬁnal time τf = + β . xE (τ ) = x(−iτ ). x(τn )e− SE [x(τ )] ℏ (2. 2. From the above we conclude that for β ≫ ℏ/(E1 − E0 ) the operator e−βH/ℏ acts like a projector on the ground state. We arrive at the same conclusion also for the correlators β→∞ x ˆ lim xf |e− 2ℏ T [ˆE (τ1 ) . we can consider 2 the euclidean n-point correlator x( β )=xf 2 x(− β )=xi 2 D[x(τ )]E x(τ1 ) . comparing to eq. . By using the complete set of energy eigenstatets the euclidean amplitude can be written as KE (xf . xE (τn )]e− 2ℏ |xi = 0|T [ˆE (τ1 ) . As we did for the real time path integral.3.4. xE (τn )]|0 ψ0 (xf )ψ0 (xi )∗ e− x ˆ βE0 ℏ βH βH (2. PERTURBATION THEORY 35 2. . .105) we have. the euclidean correlator is also shown to be equal to the operatorial expression xf |e−βH/2ℏ T [ˆE (τ1 ) . xE (τn )]e−βH/2ℏ |xi x ˆ where xE (τ ) = eHτ /ℏ xe−Hτ /ℏ ˆ ˆ (2. β β .

. 2. Indeed the time ordered correlators. .3. The real time Feynman correlators play an important role in relativistic quantum ﬁeld theory. x(tn )]|0 x ˆ (2. . . . τ2 .2.36 CHAPTER 2. in the β → ∞ limit the above equation is independent of xi . The n-point correlators are associated to the amplitudes describing the scattering of nI initial particles into nF ﬁnal particles with n = nI + nF . The euclidan Feynman correlators are related to the Feynman correlators in real time 0|T [ˆ(t1 ) . . are by construction analytic functions of the time coordinates over the connected domains where none of the time coordinates (ex. . ℏn/2 is factored out for later convenience. For the sake simplicity we shall then focus on the case xi = xf = 0 and deﬁne. Then by continuing the time coodinates on trajectories that do not cross (like in the Wick rotation) one remains within the domain of analyticity.25) Notice that the structure of above equation is reminiscent of thermal averages in classical statistical mechanics. (2. . . xE (τn )]|0 = lim x ˆ D[x(τ )]E x(τ1 ) .24) where the dependence on xi and xf has cancelled out. (2.26) xi =xf =0 where the suﬃx D indicates that this normalized n-point correlator was computed by imposing Dirichlet boundary conditions xi = xf = 0 in the path integral. . τn ) coincide. x(τn )e− SE [x(τ )] ℏ . xE (τn )]|0 x ˆ β→∞ (2. in the complex time plane as shown in ﬁg. both for real and imaginary time. the Wick rotation. Using the path integral fomulation we can suggestively write 0|T [ˆE (τ1 ) . . . Each such domain corresponds to a give ordering of the n time coordinates. . the normalized n-point correlator as ℏ GD (τ1 . (2.22) we thus ﬁnd lim x ˆ xf |e− 2ℏ T [ˆE (τ1 ) . By taking the ratio of eq. . at ﬁnite β. .2 Thermal n-point correlators One can also deﬁne the correlators using the path integral with periodic boundary conditions x(− β )=x( β ) 2 2 D[x(τ )]E x(τ1 ) . The non-analiticity is purely due to the step functions that enforce time ordering and is localized at the points where at least two coordinates coincide. τ1 .23) and eq. with the denominator playing the role of a partition function. .27) by performing an analytic continuation τi → iti in such a way as to preserve the time ordering. x(τn )e− D[x(τ )]E e− SE ℏ SE ℏ |xi = 0|T [ˆE (τ1 ) . . As we already said. . xE (τn )]e− 2ℏ |xi xf |e− βH ℏ βH βH β→∞ (2. . x(τn )e− D[x(τ )]E e− SE ℏ SE ℏ (2. This is most simply done via a counter-clockwise rotation. The time ordered correlators on the ground state are also known as Feynman correlators. xf . FUNCTIONAL AND EUCLIDEAN METHODS where by |0 we indicate the energy ground state (not the x = 0 position eigenstate!). . .28) . 2. . In the deﬁnition of GD . τn ) = n 2 xi =xf =0 D[x(τ )]E x(τ1 ) .

(2. x(tn )] Z(β) x(− β )=x( β ) 2 2 ≡ ℏ 2 GP (τ1 . xE (τn )]e− 2ℏ |x dx = Tr e− βH βH βH ℏ [ˆE (τ1 ) . xE (τn )] β = βH 1 Tr e− ℏ T [ˆE (τ1 ) . by performing the Wick rotation τk → itk we obtain ℏ 2 GP (it1 .3. . Indeed. (2. . . The arrows on the coordinate axes indicate the future direction. . in the operator formulation. τn ) n . .31) where the have deﬁned GP in analogy with the Dirichlet n-point correlator of the previous section. . .2: Wick rotation from imaginary to real time. . . eq. . both for real and imaginary time This object should expectedly be related to statistical mechanics. . . xE (τn )] x ˆ Z(β) (2. . itn ) = n βH 1 x ˆ Tr e− ℏ T [ˆ(t1 ) . Like with GD . . xE (τn )] x ˆ (2. . xE (τn )] x ˆ x ˆ T [ˆE (τ1 ) .29) Normalizing by the partition function Z = Tr[e−βH/ℏ ] we obtain he thermal average of the time ordered operator product T [ˆE (τ1 ) . applied to the corresponding path integrals. x(τn )e− D[x(τ )]E e− SE ℏ SE ℏ (2. . GD and GP only diﬀer by the choice of boundary conditions.2. PERTURBATION THEORY 37 Figure 2.28) can be written as x ˆ x|e− 2ℏ T [ˆE (τ1 ) .32) that is the thermal average of the time ordered product in real time.30) = x(− β )=x( β ) 2 2 D[x(τ )]E x(τ1 ) . . respectively Dirichlet for GD and periodic for GP . . . . . . .30) which in the path integral language corresponds to the normalized n-point thermal correlator (2.

.2. x ˆ β→∞ n n (2.38 CHAPTER 2. β β . x) − J(τ )x(τ )) ˙ (2. . notice that for β → ∞ we have e− ℏ = |0 0| lim β→∞ Z(β) βH (2. xE (τn )]|0 . the coincidence of GD and GP in the above equation represents the expected fact that for ﬁnite τ1 . We want to consider the general anharmonic oscillator with 2 + 2 lagrangian mx2 ˙ mω 2 x2 LE = + + ∆V (x) = L0 + ∆V (x) (2. J] ≡ KE (0. 0] δJ(τ1 ) δJ(τn ) J=0 0 Like in section 2. . τn ) = 0|T [ˆE (τ1 ) .26. Using the deﬁnitions in eqs.33) This is expected. τn ) = ℏ 2 GP (τ1 . . . . 0] δJ(τ1 ) δJ(τn ) J=0 1 ℏδ ℏδ .35) and similarly for periodic boundary conditions Z[β. − |J) = 2 2 β 2 1 D[x]E exp − ℏ xi =xf =0 −β 2 dτ (LE (x. .. KE [β.39) E 2 2 where ∆V can be treated as a small perturbation of the dynamics of L0 .3 Euclidean correlators by functional derivatives Consider the euclidean path integral for our system coupled to an external source J(τ ). . Then at β → ∞ both Dirichlet and periodic correlators coincide with the Feynman correlator β→∞ lim ℏ 2 GP (τ1 . J] ≡ KE (0. . J] (2.26. . . Again E just following the procedure shown in section 2.2.. the path integrals for the .3. x) − J(τ )x(τ )) . .31) we have ℏ 2 GD (τ1 . 0. For the case of Dirichlet boundary conditions xi = xf = 0 we deﬁne KE [β. .36) ˙ The correlators can then be expressed in terms of functional derivatives. . FUNCTIONAL AND EUCLIDEAN METHODS Finally. τn ) = n n 1 ℏδ ℏδ . . J] (2. (2. 0. (2. .1. β β . . .38) Z[β. τn the boundary conditions in eqs.. Z[β.31) become irrelevant for β → ∞ 2. . as β → ∞ corresponds to a statistical ensemble at zero temperature.37) KE [β. τn ) = lim ℏ 2 GD (τ1 .. (2.1 we deﬁne KE [β. J] and Z 0 [β. − |J) = 2 2 D[x]E exp − 1 ℏ β 2 xi =xf −β 2 dτ (LE (x. J] to be the path integrals associated to the simple harmonic oscillator with euclidean lagrangian L0 = E mω 2 x2 mx 2 ˙ . .34) From the path integral viewpoint.

J] are just gaussian integrals with an external source (euclidean forced harmonic oscillator). let us recall the result for a gaussian integral with a linear source (in the discrete ﬁnite case). R (2. whose solution has a convenient diagrammatic representation: the Feynman diagrams.42) where Oij is a positive deﬁnite n × n matrix and we sum over repeated indices. J] = D[x]E e 1 −ℏ R hm 2 d x(τ ) − dτ 2 +ω 2 x(τ )−J(τ )x(τ ) 2 i = 1 1 ˆ D[x]E e− ℏ [ 2 x(τ )Ox(τ )−J(τ )x(τ )] .46) We can easily transpose that discrete example to our continuum case. learn to take their derivatives with respect to J(τ ) in an eﬃcient way. The ﬁrst step amounts to a Gaussian integral.40) (2. by performing the d¯ integral: x I[J] = I[0] exp 1 Ji Gij Jj 2 (2. 2. To compute them we can use the method of Green’s functions. The inverse G of the matrix O is deﬁned by: Gij Ojk = Okj Gji = δ ik (2.4 0 Computing KE [J] and Z 0 [J] 0 KE [β.3. 2. J] 39 (2. We then perform the change of variables xi = xi − Gij Jj which obviously has trivial ¯ Jacobian dxk = d¯k x (2.43) Such an inverse matrix exists. Con0 centrating ﬁrst on KE for the sake of clarity (but the discussion on Z 0 will be analogous) we have 0 KE [β. J] = exp − ℏ = exp − ℏδ δJ(τ ) ℏδ dτ ∆V δJ(τ ) dτ ∆V 0 KE [β.2. We want to compute the integral: n I[J] = k dxk e− 2 x 1 i Oij xj +Jk xk (2.3.44) k k and such that 1 1 1 ¯ ¯ − xi Oij xj + Jk xk = − xi Oij xj + Jk Gkl Jl 2 2 2 Thus. compute the unperturbed KE and Z 0 . PERTURBATION THEORY anharmonic oscillator can be written as KE [β.45) (2.47) . First. J] Summarizing: to compute any interesting quantity (correlator or partition function) we must 0 1. The second is basically a problem in combinatorics. since Oij is assumed positive deﬁnite.41) Z 0 [β. J] and Z 0 [β. J] 1 ℏ 1 Z[β.

(2.55) the exponent is rewritten as (we use · to represent the convolution in τ ) 1 ˆ x·O·x+J ·x 2 = = 1 ˆ ˆ ˆ (y · O · y + J · GD · O · GD · J + 2y · J) − (J · y + J · GD · J) 2 1 ˆ (y · O · y − J · GD · J) (2.50.47) as − 1 x(τ1 )x(τ2 )O(τ1 . in terms of its matrix elements between time coordinate eigenstates ψτ1 (τ ) ≡ δ(τ − τ1 ) and ψτ2 (τ ) ≡ δ(τ − τ2 ) O(τ1 . it is convenient to rewrite the exponent using ˆ the inverse of O.τ2 ) 2 ˆ Oτ1 GD (τ1 . τ3 )GD (τ3 . τ2 ) = mω β 2 + τ< sinh ω β 2 − τ> sinh (ωβ) (2. τ2 ) is a function (better a distribution) in the two variables τ1 and τ2 . τ2 ) = (2.51) β GD ± .52) where τ> = max(τ1 . Like in the discrete case.48) 2 = m −∂τ1 + ω 2 δ(τ1 − τ2 ) using integration by parts and repeatedly using x(±β/2) = 0. ± β 2 = 0. Using the two equations above one can also check that O ·GD ≡ O(τ1 . (2.50) (2. ˆ GD (τ1 . τ2 ) is just the matrix form of O−1 . τ2 )dτ1 dτ2 + 2 J(τ1 )x(τ1 ) (2. that is the Green’s function satisfying Dirichlet boundary conditions GD (τ1 . τ2 ) and τ< = min(τ1 .49) 1 in full analogy with − 2 xi Oij xj + Ji xi in eq. τ2 )dτ3 = δ(τ1 −τ2 ).56) 2 .54) Shifting the integration variable x(τ ) = GD (τ. τ2 ). τ 2 = GD τ.40 CHAPTER 2. τ2 ) = δ(τ1 − τ2 ) (2. FUNCTIONAL AND EUCLIDEAN METHODS Notice that on the space of functions that vanish at the boundaries (that is at ˆ τ = ±β/2) the diﬀerential operator O = m(−d2 /dτ 2 + ω 2 ) is hermitian and positive deﬁnite.53) (2. τ2 ) ≡ ˆ dτ ψτ1 (τ )Oψτ2 (τ ) = m ∂τ1 ∂τ2 + ω 2 δ(τ1 − τ2 ) (2. Notice the properties: τ1 τ2 1 β − |τ1 − τ2 | − 2 m 2 β 1 −ω|τ1 −τ2 | lim GD (τ1 .42) We emphasize here that O(τ1 . τ ′ )J(τ ′ )dτ ′ + y(τ ) ≡ G · J + y (2. τ2 ) ≡ m −∂τ1 + ω 2 GD (τ1 .2. The solution of eqs. τ2 ) = e β→∞ 2mω ω→0 lim GD (τ1 . As a matter of fact.51) is given by 1 sinh ω GD (τ1 . we can rewrite the exponent in eq. (2.

0] = e 2ℏ J·GD ·J mω 2π sinh(ωβ) Following precisely the same steps we can compute the partition function Z 0 [β. . PERTURBATION THEORY so that we ﬁnd 0 KE [β. The only diﬀerence is that in this case the functional integral is over ˆ the space of periodic functions in the interval [−β/2. we shall drop the P and D in what follows. . τn ) = ℏ 2 n n k=1 δ δJ(τk ) e 2ℏ J·GD. As the formulae are the same for Dirichlet and period boundary conditions. . . β/2) = GP (τ. The crucial remark is that the J independent prefactors KE [β. 0] drop out from the computation and so we simply get k=n ℏ 2 GD.60) = e 2ℏ J·GP ·J Z 0 [β.3. (2.63) √ n Notice that via the rescaling J → J/ ℏ we have put in evidence a factor ℏ 2 . (2. as well as |τ1. GP (τ.59) (2.P ·J 1 J=0 .61) Notice that for β → ∞ with τ1 and τ2 ﬁxed β→∞ 1 we have (2. J].2 ± β/2| ≪ β for GD .57) (2. . 1 More precisely for |τ1 − τ2 | ≪ β. τ2 )J(τ2 ) 1 D[y]E e− ℏ SE [y] (2. τ ) = GP (−β/2. Working out eq. −β/2). τn ) = 0. J] = e 2ℏ J·GP ·J periodic 1 1 D[y]E e− ℏ SE [y] e 2ℏ J·GP ·J 2 sinh βω 2ℏ 1 1 0 (2. . that is the harmonic 0 oscillator.2.3. . The inverse of O is now given by the Green’s function GP satisfying period boundary conditions: GP (β/2.5 Free n-point correlators Applying the results of the previous section to eqs. J] = exp 1 41 1 2ℏ dτ1 dτ2 J(τ1 )GD (τ1 .63) explicitly we ﬁnd 1 • for n odd: G(τ1 . β/2]. Thus we ﬁnd Z 0 [β. 2. τ ) = 2mω sinh ω β 2 ′ (2. τ ). .62) lim GP = lim GD = β→∞ 1 −ω|τ1 −τ2 | e 2mω showing that for β → ∞ the boundary conditions expectedly do not matter. 0] 0 and Z [β. (2.38) we obtain a useful expression for the n-point correlators in the free case. 0] = As for GP one ﬁnds β ′ 1 cosh ω( 2 − |τ − τ |) GP (τ. This simply follows because exp( 2 J · G · J) is an even functional of J and we take the limit J → 0 after computing the derivatives.P (τ1 .58) 1 0 0 = e 2ℏ J·GD ·J KE [β. .

τ3 )G(τ2 .66) (2. Let us see how this works in the simplest cases. Notice ﬁrst of all that the two point correlator (D or P) simply equals the corresponding Green’s function (thus our notation was wisely chosen!) G(τ1 .e. τ4 )+ G(τ1 . (2. .64) σn is the group of all permutations of n elements where the permutations diﬀering by the exchange of two elements in a pair or the exchange of a n! pair are identiﬁed. which is what we have just shown. . τ2 )dτ e 2ℏ J·GD. functional.64). giving a factor like in A) • on a factor of type A) brought down by a previous derivative. τ2 )G(τ3 . the four points correlator is given by the graph shown in ﬁgure 2. This group contains (n/2)!2(n/2) ≡ (n − 1)!! = (n − 1)(n − 3) · · · 3 · 1 elements.42 CHAPTER 2. τn )J(τn )dτn ≡ τ1 B) G(τ1 . τn ) = p∈σn G(τp(1) . . τ2 ) ≡ τ2 Each functional derivative acts in two possible ways: • on the exponent. (2. i. (2. τ2 . For instance. FUNCTIONAL AND EUCLIDEAN METHODS • for n even G(τ1 . . τ2 ) = 1 δ δ e 2ℏ J·GD.P ·J δJ(τ1 ) = G(τ1 . τp(2) ) · · · G(τp(n−1) . thus giving a factor like in B) Using the above graphical rules to perform the derivatives one easily sees that after setting J → 0 the terms that survive give eq. . τ2 ) (2.P ·J δJ(τ1 ) δJ(τ2 ) J=0 δ 1 = J(τ )G(τ. τp(n) ) . τ4 ) = G(τ1 . derivation. τ4 ) + G(τ1 .3 G(τ1 . τ3 . The path integral allows for a more direct. It is normally derived by using creation and destruction operators. τ4 )G(τ2 .67) J=0 To compute the correlator for higher n it is convenient to introduce the graphical notation τ1 A) G(τ1 . τ3 ) .68) The general result we have found is known in quantum ﬁeld theory as Wick’s theorem.65) (2.

73) And thus. and working in a perturbative expansion in ∆V . PERTURBATION THEORY τ1 τ2 + τ3 τ4 τ3 τ4 τ1 τ2 + τ3 τ4 τ1 τ2 43 Figure 2.6 The anharmonic oscillator and Feynman diagrams Using the diagrammatic method outlined in the previous section. We shall compute the leading O(λ4 ) correction to 4! the free energy. are extensively discussed in the exercise sessions.2.70) Expanding eq.74) This result is graphically represented by ﬁgure 2. in the case at hand. τ )dτ . τ2 )G(τ3 . involving other perturbations and other observables. the linear term in the expansion of Z[β] is − ℏλ4 4! dτ δ δJ(τ ) 4 e 2 J·GP ·J 1 J=0 =− 3ℏλ4 4! β 2 −β 2 G2 (τ.41) at ﬁrst order in λ4 we have (notice that we also rescaled √ J → ℏJ) Z[β] = Z 0 [β]e− ℏ4! = Z 0 [β] 1 − ℏλ4 R δ dτ ( δJ(τ ) ) 4 e 2 J·GP ·J δ δJ(τ ) 4 1 J=0 (2.69) We also recall that for β → ∞ the partition function is dominated by the ground state. τ3 )G(τ2 . the propagator and partition function as written in section 2.3. we shall focus on the simple example of the anharmonic oscillator with an x4 potential: ∆V = λ4 x4 .3: A graphical representation of the four points correlator.4: it simply corresponds to ﬁgure 2. For illustrative purposes. (2. P (2. we can compute the n-point correlators.3 in the limiting case in which the four points are taken to coincide. From statistical mechanics we have that the free energy F (β) of a system in equilibrium at ﬁnite temperature is F (β) = − ℏ ln Z[β] β ↔ Z[β] = e− ℏ F (β) β (2.3.72) We learned from Wick’s theorem that the free four points correlator factorizes as δ 1 δ ··· e 2 J·G·J = (G(τ1 . Additional examples.71) e 2 J·GP ·J 1 ℏλ4 4! dτ + O λ42 J=0 (2. . 2. τ4 ) + G(τ1 .3. τ4 )+ δJ(τ1 ) δJ(τ4 ) G(τ1 . τ4 )G(τ2 . τ3 )) (2. so that in this limit the free energy coincides with the ground state energy β→∞ lim F (β) = E0 (2.3.

76) ωβ ℏλ4 (ωβ) coth2 ( ) + O λ42 32m2 ω 3 2 (2. This graphical representation of a term in the perturbative expansion of Z[β] is called a Feynman diagram.77). and becomes inﬁnite in the limit β → ∞.44 CHAPTER 2. τ3 .78) This result could have been anticipated in various ways.77) Some comments on the above results are in order. sets the relative size of the perturbation.2. more directly: the wave function of the harmonic oscillator ground state is localized within a distance ¯ X= x2 ∼ ℏ/mω. Notice that the factor 3 precisely counts the number of diagrams in the Wick expansion of G(τ1 . The coordinate τ is called the interaction point. suggesting that the higher order terms in the expansion also blow up in this limit. (2. By taking the limit β → ∞ we obtain the zero point energy of the anharmonic oscillator at lowest order in λ4 E0 = ℏω 2 1+ ℏλ4 + O λ42 32m2 ω 3 .4: A graphical representation of the part of the propagator linear in λ4 . τ4 ). Substituting in eq.75) the ﬁrst term in the expansion is ∼ λωβ.75) Notice that the second term in the expression in brackets represents the thermal contribution to the harmonic oscillator free energy. and the four lines associated to this point form what is called an interaction vertex. ¯ This is because in eq. necessary to apply perturbation theory. τ )2 ≡ τ Figure 2. (2.72) the explicit expression for GP and performing the (trivial) integral we ﬁnd Z[β] = Z 0 [β] 1 − so that the free energy is F = ℏω 2 1+ 1 ωβ ℏλ4 coth2 ( ln(1 − e−ωβ ) + ) + O λ42 2ω3 ωβ 32m 2 . (2. By eq.76. (2. the steps that lead to eqs. (2. By eq. m ω (2.77) leave something to be desired. as was to be expected. Also. and within this distance λ4 X 4 /(mω 2 X 2 ) ∼ λ. τ2 . (2. (2. 3G(τ. λ is the only dimensionless parameter proportional to ℏλ4 . which however has dimension ¯ mass2 /time3 . With the parameters at hand in the harmonic oscillator. while the third term gives the leading anharmonicity correction. ¯ While λ ≪ 1 is. the dimensionless parameter controlling the perturbative expansion is ℏλ4 ¯ λ= 2 3.71) the expansion parameter must be proportional to ℏλ4 . How can we then make sense of our truncation at lowest order? Our procedure to compute E0 makes sense as long as there exists a range of β where . FUNCTIONAL AND EUCLIDEAN METHODS τ1 = τ2 = τ3 = τ4 = τ .

. In the technical jargon one says that they must exponentiate. PERTURBATION THEORY ¯ 1. (2. with four lines ending at each point. As a matter of fact.75) is not a good approximation for β → ∞.75) must build up into the exponential function. we have to sum over diagrams that contain two points. τ1 and τ2 .5. τ1 A τ2 τ1 τ2 C B τ1 τ2 Figure 2. that is for e− ℏ (E1 −E0 ) ≪ 1 β 45 −→ β (E1 − E0 ) ≫ 1 ℏ (2. Expanding the exponent in eq. at lowest order corresponds.. the structure of the growing terms is ﬁxed to give an exponential. ) (2. (2. But there is more. Let us focus on it ﬁrst. The study ¯ of the O(λ2 ) contribution is already very suggestive of what happens.71). as can be seen in ﬁgure 2. we ﬁnd at O(λ2 ) 4 1 2! ℏλ4 4! 2 dτ1 dτ2 δ δJ(τ1 ) 4 δ δJ(τ2 ) 4 e 2 J·GP ·J J=0 1 ∼ x4 (τ1 )x4 (τ2 ) (2. As a corollary − ln Z[β]/β is well behaved for β → ∞ and for this quantity we can safely apply perturbation theory.81) Thus. ∼ λωβ ≪ 1 2. Thus even though eq.79) which in our case. . while diagram C consists of two disconnected pieces.5: The three diﬀerent diagrams contributing to the corrections to the ﬁrst energy level of the anharmonic oscillator. We can write three diﬀerent diagrams. the contribution of the excited states to Z(β) is negligible. the Hamiltonian formulation tells us that for β → ∞ the exact partition function is Z[β] β→∞ −→ e− ℏ E0 = e− β βω 2 ¯ λ (1+ 32 +.2.80) ¯ implying that the higher order corrections proportional to λωβ in eq. Diagrams A and B (at the topleft corner and at the bottom) are connected. and thus our manipulations were technically justiﬁed. to ωβ ≫ 1 ¯ The simultaneous satisfaction of 1) and 2) implies again λ ≪ 1.3. (2. In view of the above discussion we are now motivated to develop a picture on the structure of the higher order terms in the expansion of Z[β].

τ )dτ P + A + B + O(λ3 ) 4 (2. in addition to new connected diagrams. τ2 ) + 9G2 (τ1 . Diagrams A and B correspond to genuine. that the connected diagrams appear in the above exponent .5 for the 8-point correlator gives 8!/(4! 24) = 105.5. by direct computation. In the same way one can reproduce the factors 72 and 9 associated respectively to diagram B and C. to be equal to: 1 2! − ℏλ4 4! 2 dτ1 dτ2 24G4 (τ1 .4. τ2 ) P P P (2. at O(λ3 ).81) tells us that the number of diagrams should be the same as for the correlator among 8 x’s (more precisely x(τ1 )x(τ1 )x(τ1 )x(τ1 )x(τ2 )x(τ2 )x(τ2 )x(τ2 ) ). The point of the above expression is that the disconnected diagram C has the right coeﬃcient to provide the ﬁrst step in the exponentiation of the contribution of order λ4 . τ1 )G2 (τ1 . there will be 4 disconnected diagrams corresponding to three copies of Fig. 2.83) where we do not have written the explicit expressions of the connected diagrams A and B. Therefore the number of combinations is 4 × 3 × 2 × 1 = 24. let us pick one of the remaining 3 lines from τ1 : it can connect to τ2 in 3 possible ways.4 and also diagrams made of one copy of Fig. Putting together the diagrams in Figs. two-dimensional integrals. Notice. This implies that the two-dimensional integral corresponding to C factorizes into the square of onedimensional integral. It is a useful exercise to reproduce these numbers. τ2 )+ P 72GP (τ1 . The remaining line can connect in just 1 way. In diagrams A and B there exist at least one line (in fact more) connecting τ1 and τ2 . τ )dτ P + 1 2 ℏλ4 8 2 G2 (τ.46 CHAPTER 2. On the other hand in diagram C no line exists connecting τ1 and τ2 : diagram C consists of two disconnected copies of the lowest order diagram in Fig. eq.82) where each of these three addenda corresponds to one of the diagrams in ﬁgure 2.5 the partition function expanded at second order can be written as Z[β] = Z0 [β] 1 − ℏλ4 8 G2 (τ. To count its multiplicity we have to count all possible ways to connect the 4 lines coming out of the τ1 vertex with the 4 lines coming out of the τ2 vertex. After this we pick one of the remaining two lines. (2. This pattern continues at higher orders and for all diagrams: the disconnected diagrams just serve the purpose of exponentiating the contribution of the connected ones Z[β] = Z0 [β] e( P connected diagrams) .3. (2. FUNCTIONAL AND EUCLIDEAN METHODS The second order term is then shown.84) For instance. 2.2.4 with one copy of A (or B). which can connect to τ2 in 2 possible ways. 2. After having connected this ﬁrst line. On the other hand. The total number of diagrams is therefore 24 + 72 + 9 = 105. The numerical factors in front of each of the three factors represents the multiplicity of the corresponding Feynman diagram. Let us do it for diagram A. τ1 )G2 (τ2 .4 and 2. as it should be evident. Let us pick one given line from τ1 : it can connect to τ2 in 4 possible ways. τ2 )GP (τ2 . non-trivial. These diagrams are therefore called connected. Indeed the combinatoric factor of section 2.

τn ) = dτ1 × c = cβ (2. . .80). the associated function F is signiﬁcantly diﬀerent than zero only in the region τ1 ≃ τ2 ≃ · · · ≃ τn where all time coordinates τj are comparable. and thus also the ground state energy. . (2. notice that in the large β limit. τ2 ) ∼ β. Consider then the general structure of a diagram involving n interaction points ∝ λn 4 dτ1 . if a diagram is connected. τn ) (2. One can easily check exponentiation of the O(λ4 ) correction in Fig. 2. the diagram on the left corresponds to dτ1 dτ2 G4 (τ1 .88) . For example. (2. τ ) [1 + (other topologies)] The general proof of eq. .2. It is graphically obvious that. To prove that.84) is an exercise in combinatorics.87) where F consists of a product of G(τi . Instead in what follows we want to give a simple proof. .85) Important result: in order to compute the free energy. valid strictly speaking only for β → ∞.86) exp − 3ℏλ4 4! 2 dτ GP (τ.3. .6: Two examples of connected diagrams that contribute to the propagator of the anharmonic oscillator. We shall not present it. we only need to compute the connected diagrams. Notice ﬁrst of all that connected diagrams like in ﬁgure 2. dτn F (τ1 . the integral converges fast and we just get a number c so that dτ1 dτ2 . The exponentiation in Z[β] implies that for the free energy we simply have F = F0 + connected diagrams . .6 give a contribution that grows linearly in β. (2. τ ) δ δJ(τ1 ) n ··· δ δJ(τn ) 4 exp 1 J · GP · J 2 = n − 3ℏλ4 ℏ4! + (other topologies) = (2. . τj ). PERTURBATION THEORY 47 precisely with the same coeﬃcient they have when they appear for the ﬁrst time in the expansion of Z[β]. τ2 ) tends to zero exponentially when τ1 ≫ τ2 . G(τ1 . dτn F (τ1 . .4 exp − 1 n! 1 n! ℏλ4 4! − ℏλ4 4! dτ n δ δJ(τ ) 4 e 2 J·GP ·J J=0 1 = 4 n dτ1 · · · dτn 2 dτ GP (τ. Figure 2. but leave it as a challenge to the student. . . . . τn . It is a reﬁnement of the discussion around eq. . . By integrating ﬁrst on τ2 .

Now. In particular A1 is the sum over the connected diagrams.91) ⇒ E0 = ℏω − ℏA1 2 Once again we found that the ground state energy is given by the sum of all connected diagrams.90) An = (A1 )n n! and we can write: β→∞ lim Z[β] = lim Z0 [β]eA1 β β→∞ (2. FUNCTIONAL AND EUCLIDEAN METHODS therefore connected diagrams give a contribution linear in β. . since Z[β] ∼ e−βE0 . In that case the integrand is non vanishing when the time coordinates are comparable within each connected subdiagram. .89) where Ak is a sum over all diagrams involving k connected subdiagrams. the coeﬃcients of β n must be such as to exponentiate A1 β. 1 (2. . Thus for β → ∞ we have Z[β] = Z0 [β] 1 + A1 β + A2 β 2 + .48 CHAPTER 2. i.e. Therefore a diagram consisting of k disconnected diagrams will produce a contribution growing like β k . (2. This argument easily generalizes to diagrams that consist of 2 or more disconnected pieces.

In the quantum case. diﬀraction phenomena are unimportant.Chapter 3 The semiclassical approximation Let us ask an intuitive question. In this case. . we can expect to be able to form a wave packet of size between λ and L. These methods oﬀer no diagrammatic representation: when small couplings are involved. as for the propagation of light. but more generally also when dealing with tunneling phenomena. and similarly we expand the eigenvectors H. We have two approximation methods to analytically treat systems that cannot be solved exactly: Perturbation theory: Let us assume we are dealing with a system whose hamiltonian H can be written as H = H0 + λHp where H0 is exactly solvable and λHp can be treated as a small perturbation (that is |λ n|Hp |m | ≪ (0) (0) (0) |En −Em |. semiclassical contributions are proportional 2 to e−1/λ and thus they oﬀer non-perturbative corrections to the system. Semiclassical methods: We use these methods when a system is close to the classical limit ℏ → 0. When do we expect the wave nature of a particle propagation to be well appoximated by the properties of the trajectory of a classical particle? Well. we expect similar results: if the wavelength λ is short enough with respect to the typical length over which the potential varies L. when the wavelength λ is much smaller than the characteristic physical size of the system. . 49 . and we can use geometric optics. whose motion should be well appoximated by that of a classical particle. where |n and En are respectively the eigenstates and eigenvalues of H0 ). we work in perturbation theory in λ and we compute the eigenvalues of the hamiltonian En as a series in λ: (0) (1) (2) En = En + λEn + λ2 En + .

(3. xi .3) Finally.5) We can expand at quadratic order around xc : ˙ my 2 ˜ mx2 ˙ → 2 2 1 ′′ y V (x) → V (xc )˜2 2 where we used the notation V ′′ (xc ) = ∂x 2 V (x)|x=xc . for which Ω2 (t) = ω 2 is time independent.2) √ The rescaling y = ℏ˜ makes it evident that the expansion in powers of y truly y corresponds to an expansion in powers of ℏ K(xf ..3. we obtain the Gaussian. approximation to the propagator: Ksc (xf . tf . tf . Thus. tf .3. (3. or semiclassical. More precisely if we write V ′′ (xc (t)) ≡ Ω2 (t). ” (3. Starting from the propagator K(xf . xi .50 CHAPTER 3. We can thus proceed along the same lines of subsection 1.4) Let us now consider the simple case where: S[x] = mx2 ˙ − V (x) dt 2 (3. the computation generalizes that of the ﬂuctuation determinant for the harmonic oscillator.4 and ﬁrst perform the change of variables y (t) = ˜ n an yn (t) ˜ (3. tf .3. ti ) = ei S[xc ] ℏ D[y]e ℏ i “ 1 δ3 S 1 δ2 S 2 2 δx2 y + 3! δx3 y 3 +.1 The semiclassical propagator The goal of this section is to compute the leading contribution to the path integral in the formal limit ℏ → 0. ti ) = D[x]ei S[x] ℏ (3. xi . In practice this will correspond to deriving a general expression for the Gaussian..2. dropping orders higher than the second in the Taylor expansion of the action. approximation discussed in section 1. ti ) = N · ei S[xc ] ℏ ˜ D[˜]e 2 δx2 y y i δ2 S 2 (3. THE SEMICLASSICAL APPROXIMATION 3.7) .6) This is a Gaussian path integral like the one we found for the harmonic oscillator in subsection 1. ti ) = N · ei S[xc ] ℏ ˜ D[˜]e 2 δx2 y y i δ2 S 2 √ + ℏO (y 3 ) . or semiclassical.4. we must compute: I[xc ] = N D[˜]e 2 y i R ˙ dt(my 2 −V ′′ (xc )˜2 ) ˜ y . xi . Let us recall it.1) we expand it around the classical solution xc : x = xc + y and get K(xf .

(1.11) We have an = Unm a′ ˜ ˜m † (3. THE SEMICLASSICAL PROPAGATOR where yn are the orthonormalized eigenmodes of the diﬀerential operator −m d2 d2 − V ′′ (xc (t)) ≡ −m 2 − Ω2 (t) 2 dt dt d˜n a √ 2πi 51 (3.13) n m ˜ which proves that the Jacobian N is independent of the orthonormal basis and. The result we are are going to prove is that D = c · ψ0 (tf ) ˜ N2 (3. (3. We thus have d˜ a √ n = 2πi d˜′ a √ m 2πi (3.8) which controls the quadratic action.10) was ﬁrst carried out by Gelfand and Yaglom. but less formal. but beware that these Jacobians are anyway ill deﬁned when ǫ → 0! . {yn (t)} are orthonormalized. we use a diﬀerent simbol for the Jacobian with respect to eq.1.90).6) gives ˜ I[xc ] = N det −m d2 − V ′′ (xc ) dt2 1 −2 ˜ = N D− 2 1 (3. is that the Jaco˜ bian factor N is independent of V ′′ (xc (t)).15) with the boundary conditions ψ0 (ti ) = 0 (3. just to confuse the reader(!). (3.14) where: • ψ0 (t) satisﬁes the diﬀerential equation − md2 + Ω2 (t) ψ0 (t) = 0 dt2 ′ ψ0 (ti ) = 1 . The computation of determinants as in eq. important.10) One ﬁnal. a fortiori. ˜m ′ (3.9) n so that in the end the gaussian integral in eq.16) 1 Notice that. (3.12) where Umn is an (inﬁnite) unitary matrix satisfying U U = 1 and detU = 1 as ′ it follows from standard linear algebra given that the two sets {yn (t)}.3. In our derivation we shall follow a clever. derivation due to Coleman. We have again the change of measure1 ˜ N · D[˜(t)] = N · y (3. Let us now proceed. Consider indeed choosing a diﬀerent ′ orthonormal basis {yn (t)} on the space of square integrable functions on the time interval T = tf − ti y (t) = ˜ n an yn (t) = ˜ m a′ ym (t) . is independent of whatever diﬀerential operator has that basis as its eigenbasis. This is to orient the reader on the origin of the factors. remark before going to the computation. A more mathematical derivation can be found in [8]. This is because we are now working with rescaled variables y and an ˜ ˜ which absorbed a power of ℏ.

2 λ (3. This property follows by simple / solution of the diﬀerential equation 3. (3. Consider now two diﬀerent time dependent frequencies: Ω2 (t) and Ω2 (t). emphasizing ideas above mathematical rigour.17. λ ∈ R+ . and thus the solution satisfying the boundary conditions 1.2 3. and only if ψλ (tf ) = 0. THE SEMICLASSICAL APPROXIMATION • c is a constant that does not depend on Ω2 (t). We have then the following Theorem: ˆ det O1 − λ ˆ det O2 − λ = ˆ Consider then the same problem for O2 ≡ − (2) md2 2 dt2 + Ω2 (t) and call ψλ (2) λψλ with the same boundary ψλ (tf ) ψλ (tf ) (2) (1) (3.18. these two conditions ﬁx the solution everywhere.1.2) (t) = 1 2i √λ m i√ λ (t−ti ) − e−i m (t−ti ) + O e m λ Ω2 1. and thus we can compute it in the free particle case.3 for more detail) ψλ (1.19) We shall now prove the above by following Coleman. Indeed.18 is written as (see the mathematical appendix 3. (1) ˆ • λ is an eigenvalue of O1 if. All these zeroes and poles are simple as we show in the mathematical appendix of subsection 3.18) • For all λ. for λ → ∞. (1) (3. ˆ (2) the solution of the diﬀerential equation O2 ψλ = conditions in 3. • f (λ) → 1 in the limit λ → ∞. 1 2 Consider also the solution of the “eigenvalue” equation.21) . as it is customary in physics. Ω2 (t) = 0. and. Indeed.1.20) (2) ψλ (tf ) f satisﬁes the following properties: • f (λ) has obviously all its zeroes in correspondence with the eigenvalues of ˆ ˆ O1 and all its poles in correspondence with eigenvalues of O2 . generalizing eq.15): − md2 (1) (1) + Ω2 (t) ψλ (t) = λψλ (t) 1 dt2 ˆ O1 (3. remember ˆ that we are looking for the eigenvalues of O1 over the space of functions that vanish at ti and tf .2.17) with boundary conditions ψλ (ti ) = 0 (1) ∂t ψλ (ti ) = 1 . Let us deﬁne ﬁrst: (1) ψ (tf ) f (λ) = λ (3. and thus we (1) can compute ψλ (tf ). we can neglect Ω2 (t).52 CHAPTER 3.

we have g(λ) → 1. (3.2 λ/m = [nπ/(tf − ti )]2 . λ ∈ R+ .18) at λ = 0. (3. By Cauchy theorem the only such function is h(λ) ≡ 1. is independent of Ω2 ) in order to deal only ˜ with the physically interesting combination N /Det1/2 . there are no zeroes / at t = tf .17. Consider now: g(λ) = ˆ det O1 − λ ˆ det O2 − λ = 2 det −m∂t − λ − Ω2 1 2 − λ − Ω2 ) det (−m∂t 2 (3. As a corollary of eq. In this situation the addition to this operator of Ω1 / 2 or Ω2 is a small perturbation with negligible eﬀects as λ → ∞. Putting everything together we have that the ratio h(λ) = f (λ) g(λ) is analytic everywhere (the singularity at the simple poles of f is neutralized by simple zeroes of 1/g and conversely the singularity at the simple zeroes of g is neutralized by simple zeroes of f ) and tends to 1 as λ → ∞.1.21)) vanishes at t = tf and therefore the Ω2 perturbation cannot be treated 1. THE SEMICLASSICAL PROPAGATOR 53 One of the two exponentials grows for λ ∈ R+ . On the other hand. (3.23) where in the last step we used that lim|z|→∞ (h(z) − 1) = 0.3. like c. for 1.3. More precisely. Notice that T ≡ −m∂t − λ 2 has eigenvalues λn = (nπ/(tf − ti )) − λ. Provided we ﬁnd the .19) we have that det − md − Ω2 (t) dt2 ˜ N2 2 (Ω) 1 (Ω) ψ0 (tf ) =c (3. Consider now Ω2 (t) ≡ Ω2 (t) and call ψ0 the solution to the problem in 1 eqs. the zeroth order solution (ﬁrst term in eq. Thus. This result for λn implies that the norm |λmin | of the eigenvalue of smallest norm λmin goes to inﬁnity 2 when λ → ∞. λ ∈ R+ . applying Cauchy theorem to h(λ) − 1 and integrating over a circle C with radius R → ∞ we have h(λ) − 1 = 1 2πi C h(z) − 1 dz → 0 z−λ (3. as the function g has zero and poles for arbitrarily large λ. and we can safely neglect Ω2 for all t. and where we have judiciously inserted the normalization Jacobian (which. Without aiming at mathematical / rigour (but a more thorough discussion is presented in Homework 8) this 2 ˆ result is qualitatively understood as follows. Therefore we expect g(λ) → 1 for λ → ∞. Obviously this reasoning does / not work for λ ∈ R+ . • For λ → ∞.24) where c is independent of Ω(t). showing that limλ→∞ f (λ) = 1 for λ ∈ R+ .22) • It has exactly the same simple poles and simple zeroes as f (λ). λ ∈ R+ .2 as small. and that proves our theorem.

28) 3.18) at λ = 0. Consider now the Wronskian: ˙ W (t) = vc (t)ψ0 (t) − vc (t)ψ0 (t) ˙ (Ω) (3. ∂t ψ0 (ti ) = 1. the solution to the diﬀerential equation problem in eqs.16) is ψ0 (t) = t − ti and thus: 2πiℏ m The result for the gaussian integral in the general case is then c= I[xc ] = m 2πiℏψ0 (tf ) (Ω) (0) = 2πi(tf − ti )ℏ m (3. vc = xc satisﬁes the diﬀerential equation that we are trying to solve ˙ (Ω) in order to ﬁnd ψ0 (t). it doesn’t satisfy in general the boundary conditions ψ0 (ti ) = 0. By diﬀerentiating it once we get ˙ mxc + V ′′ (xc )xc = 0 ¨ ˙ (3. we have W (t) = 0. The classical equation of motion tells us m¨c + V ′ (xc ) = 0 x (3.17.24) for an operator whose determinat we already know.54 CHAPTER 3.1. THE SEMICLASSICAL APPROXIMATION constant c.26) (3. In this case the determinat prefactor in K was computed in eq. We can then write: vc 2 ⇒ d dt d dt ψ0 vc ψ0 vc = vc 2 = vc (ti ) ˙ ψ0 ψ0 vc ˙ − 2 vc vc 1 vc 2 t ti = W = vc (ti ) ⇒ ψ0 (t) = vc (t)vc (ti ) ⇒ ψ0 (tf ) = vc (tf )vc (ti ) 1 dt′ vc 2 (t′ ) tf 1 vc 2 (t′ ) dt′ (3.3.15. and thus W (t) = W is a constant. (3.25) (3. gives us the determinant we were after! Finding the constant c is easy: we just need to use eq.31) where ψ0 (t) = ψ0 (t) is the function we are looking for.3.48) 1 md2 det − 2 ˜ dt N2 while the solution to eqs. (3. (3.30) Thus.27) (3. The easiest option to do so is to consider the free particle case. (1.1 VanVleck-Pauli-Morette formula (Ω) Let us now derive a useful implicit expression for ψ0 (tf ). However.29) where the dot denotes derivation with respect to time. where Ω(t) = 0. ˙ Since ψ and vc satisfy the same diﬀerential equation.32) ti .

40) (3.43) . 3 2 Writing vf = vc (tf ) ψ0 (tf ) = ≡ The prefactor is then: I[xc ] = Or in terms of velocities: vf vi xi xf vf vi xi m 2(E − V (x)) dx v(x)3 dx (3.34) √ 2πiℏvf vi m 1 xf dx xi 2(E−V (x)) (3. THE SEMICLASSICAL PROPAGATOR 55 Let us now write ψ0 (tf ) in terms of energy and potential. We have: tf Sc = ti mxc2 ˙ −V 2 tf dt = ti mxc2 − E dt = ˙ xf xi 2m(E − V (x))dx − E(tf − ti ) (3.36) Let us now express the result in terms of the classical action.41) Thus. from the classical solution: and vi = vc (ti ): mvc 2 (t) 2 = E − V (xc (t)).33) (3.38) (3. We have.37) Thus: ∂Sc = Pf = ∂xf 2m(E − V (xf )) ∂E 1 ∂E 2m = E − V (xf ) ∂xi vf ∂xi dx = x ˙ xf xi xf xi (3. and dt = xf dx vc .39) ∂ 2 Sc ∂Pf 1 = = ∂xi ∂xf ∂xi 2 We can note that: tf xf tf − ti = 0= dt = ti xi ∂(tf − ti ) 1 1 ∂E =− − ∂xi vi 2 ∂xi m ∂E =− ⇒ xf dx ∂xi vi xi v3 (x) m dx 2 (E − V (x)) 3 2 m dx 2(E − V (x)) (3.42) And we get in the end the Van Vleck-Pauli-Morette determinant: I[xc ] = N 1 det −m d2 dt2 − V ′′ (xc ) = − 1 ∂ 2 Sc 2πiℏ ∂xi ∂xf (3.3. the derivative of the action can be written as: ∂ 2 Sc m =− x ∂xi ∂xf vi vf x f i dx v 3 (x) (3.1.35) I[xc ] = m 2πiℏvf vi xf dx xi v 3 (x) (3.

we get: − md2 +V dt2 ˙ ˙ ψλ (t) = λψλ (t) + ψλ (t) (3.44) 3.1.47) (3.3 Mathematical Appendix 2 2 (−∂t − λ − Ω2 (t))ψ = 0 We want to write explicitly the solution to ψ(0) = 0 . the Wronskian is zero everywhere. and so is its derivative.46) Notice that since we impose the same boundary conditions for all λ.51) treating Ω2 (t) as a perturbation. ∂t ψλ (ti ) = 1. We will use the notation: d ˙ ψλ (t) ≡ ψλ (t) dλ d ′ ψλ (t) ≡ ψλ (t) dt When we diﬀerentiate (3. ψ ′ (0) = 1 (3. ti ) = ei Sc ℏ − ∂ 2 Sc 1 2πiℏ ∂xi ∂xf (3. we have: ˙ ˙′ ψλ (ti ) = ψλ (ti ) = 0 Consider then the “Wronskian”: ′ ˙ ˙′ W = ψλ ψλ − ψλ ψλ ˙ ˙ W ′ = ψλ ψ ′′ − ψ ′′ ψλ λ λ (3.2 Mathematical Appendix 1 md2 +V dt2 Let us now prove that ψλ (tf ) only has simple zeroes. 3. In order to carry on perturbation theory in Ω2 we need the retarded Green’s function for the zeroth order equation G(t.49) On the other hand. xi .48) 1 1 ˙ ˙ (V − λ)ψλ − ψλ − (V − λ)ψλ ψλ = ψλ m m ψ 2 =− λ 0 m (3.1. ′ ˙′ ˙ W (tf ) = ψλ (tf )ψλ (tf ) − ψλ (tf )ψλ (tf ) (3.50) ˙ Thus.45) with the boundary conditions ψλ (ti ) = 0. tf . if both ψλ (tf ) and ψλ (tf ) vanish. THE SEMICLASSICAL APPROXIMATION And thus we can write the ﬁnal form of the semiclassical propagator: Ksc (xf . ψλ (t) ≡ 0. t′ ) = .45) with respect to λ. Thus. we have the boundary conditions: W (ti ) = 0.56 CHAPTER 3. Consider the equation: − ψλ (t) = λψλ (t) (3.

t′ ) = θ(t − t′ ) √ e−2i λ(t−t ) − e2i λ(t−t ) .2 3. eq. THE FIXED ENERGY PROPAGATOR 57 2 ˆ ˆ T −1 where ≡ T ≡ −∂t + λ. t dt′ (3.57) Notice that ψ deﬁned above automatically satisﬁes the boundary conditions ψ(0) = 0.55) (1 − G ◦ Ω2 )φ = G ◦ Ω2 ψ0 and solved iteratively in φ ∞ φ = (1 − G ◦ Ω2 )−1 G ◦ Ω2 ψ0 = and therefore ψ= ∞ n=0 n=1 (G ◦ Ω2 )n ψ0 (3. sometimes simply Green’s function.2. For the lowest order correction we have (G ◦ Ω2 )ψ0 = which for λ ∈ R+ / √ √ ′ ′ 1 √ e−2i λ(t−t ) − e2i λ(t−t ) Ω2 (t′ )ψ0 (t′ ) 2i λ 0 √ is ∼ O(Ω2 / λ)ψ0 . t.3. xi . 0) K(E. ψ ′ (0) = 1. [8] for instance) √ √ ′ ′ 1 G(t.56) (G ◦ Ω2 )n ψ0 .53) and writing ψ = ψ0 + φ. (3. is deﬁned as the Fourier transform of the retarded propagator θ(t)K(xf . can be written as (3.51) becomes ˆ T φ = Ω2 (ψ0 + φ) (3. (3. xi ) ≡ = 0 ∞ 0 ∞ K(xf . 0)e xf | e− ∞ 0 ˆ iHt iEt ℏ + ℏ iEt ℏ dt |xi dt ˆ = lim ǫ→0 xf | e ℏ (E−H+iǫ) |xi dt iℏ ˆ E − H + iǫ |xi (3.2. xf . t. xi . 2i λ (3.54) ˆ which by use of G = T −1 and by employing standard functional operator notation (by ◦ we indicate the product of functional operators). By a simple computation analogous to the one used to ﬁnd the euclidean Green’s function one ﬁnds (see section III of ref.52) Also deﬁning the zeroth order solution ψ0 = √ √ 1 √ ei λ(t−ti ) − e−i λ(t−ti ) 2i λ (3.59) it = lim xf | ǫ→0 .58) 3.1 The ﬁxed energy propagator General properties of the ﬁxed energy propagator The ﬁxed-energy retarded propagator.

xf . xi ). as usual when dealing with distributions. xf . In this situation any solution ψ of the Schroedinger equation with positive energy behaves at |x| → ∞ as a linear combination of outgoing and ingoing plane waves x→+∞ lim ψ lim ψ = = a+ e−i a− e √ 2mEx + b+ e−i + b− e √ 2mEx (3. we must recall that the inverse Fourier transform.63) to coincide with the retarded propagator. Which boundary conditions are required? The answer is obtained by simple physical considerations. Let us discuss this in detail. 2m x (3.66) . (3. xi )e−iEt/ℏ dE 2πℏ (3.61) f (E. that is of retardation. xi ) is just the position space representation of the inverse of the Schr¨dinger operator. From the last equation above.63) ′ ′ where W = ψ1 (x)ψ2 (x) − ψ2 (x)ψ1 (x) is the wronskian. we recall that. o ˆ K = E−iℏ . xf . Given two linearly independent solutions ψ1 and ψ2 of the homogeneous Schr¨dinger’s equation at energy E o − ℏ2 2 ∂ + V (x) − E ψ1. THE SEMICLASSICAL APPROXIMATION where. xf . In order to illustrate that.65) x→−∞ √ −i 2mEx √ −i 2mEx The solutions ψ1. 0) = ∞ −∞ K(E. For the advanced propagator we would have instead analyticity in the lower complex E plane. x. K(xf .60) In coordinate representation.2 = 0 2m x (3. To ﬁnd out which behaviour is the physically correct one. t. which is relevant for barrier penetration problems and in the case of potential wells of ﬁnite spacial width.58 CHAPTER 3.62). y) = 2mi (θ(x − y)ψ1 (x)ψ2 (y) + θ(y − x)ψ2 (x)ψ1 (y)) ℏW (3. the retarded propagator K(E. y) = −iℏδ(x − y) . and the boundary conditions we discuss below. xi . we must choose ψ1 and ψ2 satisfying the appropriate boundary conditions. (3. (3. that is for ImE > 0.64) (3. However in order for eq. xi ) to have the physically correct behaviour at |xf | → ∞. the factor e−ǫt/ℏ with ǫ → 0+ is introduced to insure converge of the integral at t → ∞. xf . Notice that since we integrate over t > 0. Let us focus ﬁrst on motion in a potential V (x) satisfying lim|x|→∞ V (x) = 0. x.60) becomes: − ℏ2 2 ∂ + V (x) − E K(E. the exponent eiEt/ℏ ensures strong convergence of the integral for any E in the upper complex plane. the propagator.61) This equation.62) the following function satisﬁes eq. W is constant over x. This is an important property of retarded Green’s functions: analyticity in the upper energy plane is the Fourier space manifestation of causality. It thus satisﬁes the relation: ˆ H+iǫ ǫ→0 ˆ ˆ lim (E − H + iǫ)K = iℏ1 (3.2 are selected by requiring K(E. equation (3. by eq. For ImE > 0 the retarded propagator K(E. fully ﬁx K(E. xi ) is thus an analytic function of the complex variable E. let us work on a case by case basis.

2 This Summarizing: retardation corresponds to outgoing waves at x → ±∞. Then eqs. (3.3.69. t) should be a superposition of waves travelling out in the negative x direction. 3 By eq. Indeed to ﬁnd the correct boundary conditions we could have worked the other way around. (3. xi ) ∝ e−i √ 2mExf (3. The request that K be square integrable implies the choice in eqs.70) to E > 0 through a path in the Im E > 0 half-plane we get the outgoing wave boundary conditions of eqs. 3. Ψ(xf . xi ) ∝ e− 2m|E|xf → 0 (3.72) x→−∞ corresponding to K representing an operator on the space of normalizable wave functions. 3 On the other hand.68) This gives us all the constraints to fully ﬁx ψ1.68) are satisﬁed if √ (3. xi ) at Im E > 0. t) = δ(xf − xi ). if we continue eq. Then by analytically continuing eqs.61) there are two possibilities: e+ 2m|E|xf √ and provided we have such two solutions we can write the retarded propagator as K(E. (3. for xf → −∞.67. the wave function Ψ(xf .70).71) lim ψ1 (x) ∝ e+i 2m|E|x x→+∞ √ lim ψ2 (x) ∝ e−i 2m|E|x (3. (3. xf .68). and consider ﬁrst the asymptotic behaviour of √ K(E. This would correspond to the advanced propagator. (3. Then at any ﬁnite t > 0. y) = 2mi (θ(x − y)ψ1 (x)ψ2 (y) + θ(y − x)ψ2 (x)ψ1 (y)) ℏW (3. tells us to how to continue E from E > 0 to E < 0 2 . xi ) ∝ e+i √ 2mExf (3. By eq. Then for E < 0 (or more precisely E = lim ǫ → 0+ (−|E| + iǫ) the above asymptotic behaviours become √ lim K(E. xi ) ∝ e+ 2m|E|xf → 0 (3. 3.70) to E > 0 by a path in the Im E < 0 half-plane we would get incoming instead of outgoing boundary conditions. xf .63) the behaviour at xf → +∞ and xf → −∞ is controlled respectively by ψ1 and ψ2 . xf . for xf → +∞ ≫ xi . analiticity √ in the upper energy plane. x. xi ) for E < 0. xf . ie with negative momentum xf →−∞ lim K(E. ie with positive momentum and positive energy. (3. xf .2 and write a general expression for K(E. . (3. that is limt→0 Ψ(xf .67) Similarly. notice that. t) should correspond to a superposition of waves travelling out in the positive x direction.73) √ is because the cut of E should be chosen at Im E < 0 to ensure analyticity of K(E.63. THE FIXED ENERGY PROPAGATOR 59 represents the wave function Ψ(xf .69. This implies xf →+∞ lim K(E. xf . 3.70) xf →−∞ and e− 2m|E|xf . t) at t > 0 for a state which at t = 0 was a δ-function at some ﬁnite position xi . xi ) in the form 3. 3. which is analytic in the lower energy half-plane. Before discussing that.2.69) xf →+∞ √ lim K(E.67. xf .

It is easy to ﬁnd the right boundary conditions for ψ1. (3. that ψ− . Obviously ψ− does not belong to the physical Hilbert space. So we have again ψ2 ≡ ψ− . We obviously must choose the second one to make physical sense.59) and performing the momentum integral using Cauchy theorem. so that the region x → ∞ is not accessible at ﬁnite energy.76) Which is precisely the expression derived in Homework 6. x. Notice on the other hand. by inserting a complete set of ﬁxed momentum states in eq.78) . by current conservation. Now. THE SEMICLASSICAL APPROXIMATION We can generalize these considerations to other potentials.2 for this case too. will always be. τ2 ) in the limit β → ∞. x. we will have to ﬁnd the two following solutions to the Schr¨dinger equation: o ψ1 ≡ ψ+ (x) → A+ eikx ψ2 ≡ ψ− (x) → A− e−ikx x → +∞ x → −∞ (3.63). For instance consider the case where limx→−∞ V (x) = 0 and limx→+∞ V (x) = +∞. y) = 2mi (θ(x − y)ψ+ (x)ψ− (y) + θ(y − x)ψ+ (y)ψ− (x)) ℏW (3. In the case of ψ2 the condition is the same as before: outgoing plane wave.77) We will then deﬁne the retarded propagator by: K(E. an outgoing wave at x → −∞. Then whe have that ψ1 is just the stationary solution ψstat of the Schroedinger equation with energy E: it belongs to the physical Hilbert space. Notice that we recover a result similar to the correlator Gβ (τ1 . when we switch on the potential. at any real value of E. We have: ψ1 (x) = eikx ψ2 (x) = e−ikx (3.60 CHAPTER 3. y) = m θ(x − y)eik(x−y) + θ(y − x)eik(y−x) ℏk m ik|x−y| = e ℏk (3. We then get the ﬁxed energy retarded propagator for the free particle: K(E. x.74) The last case to consider is the one where motion is fully limited: lim|x|→+∞ V (x) = +∞. a combination of growing and decreasing solution at x → ∞. In this case both ψ1 and ψ2 must correspond. (3. To illustrate our results let us consider the simplest possibe case of a free particle: V (x) ≡ 0.75) √ where k = 2mE. But this does not concern the asymptotic behaviour of K because of the θ step functions in eq. to the exponentially dcreasing solutions at respectively x → +∞ and x → −∞. For ψ1 the two possible behaviours at x → +∞ by solving the Schroedinger equation are exponentially growing or exponentially decreasing. y) = 2mi (θ(x − y)ψstat (x)ψ− (y) + θ(y − x)ψ− (x)ψstat (y)) ℏW (3. In this case the propagator is then K(E.

THE FIXED ENERGY PROPAGATOR Example: Barrier penetration 61 With the potential V (x) switched on. 3.2. x. This will become more transparent in the semiclassical path integral approach. in 1 ∗ an intuitive and concise way: K(E. y) → v ψout (x)ψin (y).81) We see in the two contributions in the above equation the relics of the two possible classical paths: a direct path from y to x. x → ∞. Thus. as is illustrated in ﬁgure 3. and x > y. in the limit x → ∞. we ﬁnd: K(E. we have: K(E. Conservation of the probablity current moreover implies |A+ |2 + |B+ |2 = 1.2 to correspond to the solutions to the Schr¨dinger equation with the following asymptotic behaviour: o ψ1 ≡ ψ+ (x) → ψ2 ≡ ψ− (x) → eikx + B+ e−ikx A+ eikx A− e−ikx e−ikx + B− eikx x → −∞ x → +∞ x → −∞ x → +∞ where A+ is the transmission coeﬃcient giving the transmission probability p = |A+ |2 . Using the above deﬁnitions. This is expected to be a good approximation when we can neglect . Similarly for motion in the opposite direction described by ψ− . Similarly. where V (x) → 0. in the limit x → ∞. x. x. we need ψ1. y → −∞.80) where v = k/m is the velocity. x.79) Now.1. x. K directly gives the transmission coeﬃcient: the propagator (as it should!) directly gives the amplitude for propagation through the barrier. y) in the semiclassical approximation. y) ∼ 1 ∗ (ψin (x) + ψout (x))ψin (y) v direct reﬂected 1 ik(x−y) = e + B+ e−ikx−iky v (3.2 Semiclassical computation of K(E) Let us now turn to the computation of K(E. y → −∞. one can consider reﬂexion by a barrier by letting y → −∞. Far away. K(E. x → −∞. and a path from y to x where the particle is reﬂected by the barrier. This gives: K(E. y) is written. The physical interpretation of the solution ψ+ is that it describes an incoming wave ψin = e+ikx travelling from −∞ which is scattered at the barrier into an outgoing wave ψout with ψout = B+ e−ikx for x → −∞ and and ψout = A+ e+ikx for x → +∞.3. assuming that it has a barrier located somewhere along the x axis. y → −∞) = m A+ ik(x−y) A+ eik(x−y) = e ℏk v (3.2. y) = m 1 θ(x − y)ψ+ (x)ψ− (y) + θ(y − x)ψ− (x)ψ+ (y) ℏk A− (3. We can compute the Wronskian at x → −∞: W = 2A− ik.

this integral will be dominated by the region of the stationary point t∗ where g ′ (t) = 0.2 and 3. xi ) = 0 ∞ − Sc iEt ∂ 2 Sc 1 ei ℏ + ℏ dt 2πiℏ ∂xi ∂xf (3. So we proceed with such a an evaluation.1). t. derivatives higher than cubic in V (x). Notice that in doing so we are introducing an additional approximation in computing K(E. or equivalently for E large enough.62 CHAPTER 3.4 We have to compute the following integral: K= 0 ∞ f (t) ei g(t) ℏ dt (3. xi ) with respect to the gaussian computation of K(xf . xf . xf . and extend the lower bound of the integral from −t∗ to −∞. The ﬁrst one goes directly from y to x. (3.82: ∂Sc ∂t +E =0 t=t∗ ⇒ ⇒ t∗ = t(E) (3. and t∗ is the stationary point.84) 0 We will then be able to change the integration variable by δt ≡ t − t∗ . Let us brieﬂy recall how the saddle point approximation works. we expect a saddle point computation of the integral over t to be a reliable approximation.1: The two possible classical paths from y to x. In Problem 1 of Homework 7 the condition for the validity of these approximations is worked out. We will get the following integral: K= f (t∗ )ei g(t∗ ) ℏ ∞ ei 2 ′′ 1 g (t∗ ) (t−t∗ )2 ℏ dt (3. we can substitute 1 f (t) ≈ f (t∗ ) and g(t) ≈ g(t∗ ) + 2 g ′′ (t∗ )(t − t∗ )2 . Using the semiclassical result in eq. Note that the the stationary point 4 Notice that we have already been using this concept throughout the course (in particular in sections 1. 0).85) (3. So the explanation given here should truly have been at the beginning! Better late than never.83) In the limit where ℏ → 0. thus getting an easily integrable Gaussian Integral. In this region. Let us therefore look for the stationary point for the integral 3. . For ℏ → 0. THE SEMICLASSICAL APPROXIMATION V y x Figure 3.3.82) where the classical action Sc in a function of the initial and ﬁnal positions xi and xf . whereas the second one is ﬁrst reﬂected by the potential barrier before going to x. and the travel time t. xi .44) we must perform a Fourier transform Ksc (E.86) −Ec (t∗ ) + E = 0 where Ec (t) is the energy of the classical trajectory for the motion from xi to xf in a time t.

73) to deduce the form of the ﬁxed energy wave functions corresponding to our approximation ψW KB (x) ∼ 1 exp i v(x) x x0 p(x′ ) ′ dx ℏ . That standard method also allows a direct study of the .we have ˙ xf t= xi dx = xc ˙ xf xi xf xi m dx 2(Ec − V (x)) m 2(E − V (x)) ∂ 2 Sc ∂xi ∂xf 3 2 ∂Ec ⇒1=− ∂t ⇒− ∂E = ∂t 1 1 m 1 m dx (3. By using xc = v(x) ≡ 2(Ec − V (x))/m. showing that we are on the right track to derive a semiclassical approximation to the propagator.88) t=t∗ To explicitly write the result we need to use some simple mechanical identities. The stationary phase integral is therefore: dδt exp i 2ℏ ∂ 2 Sc ∂t2 (δt)2 = 2πiℏ ∂ 2 Sc ∂t2 = 2πiℏ 1 m dx v3 (3.2. (3. xf . THE FIXED ENERGY PROPAGATOR 63 approximation has selected a classical trajectory with energy Ec = E.87) where we used Sc = (mxdx − Edt) and where p(x) = 2m(E − V (x)) is the ˙ momentum of the classical trajectory with energy E.3. Ec = E and where. We have that the exponent at the stationary point t∗ is given by: t∗ t∗ Sc + Ec t∗ = (L + E) dt = 0 0 mxc2 dt = ˙ xf p(x)dx xi (3. Let us now ﬁnd the second order expansion of Sc around t∗ diﬀerentiating the identity ∂Sc = −Ec (t) ∂t ∂ 2 Sc ∂t2 =− ∂Ec ∂t (3. vf = v(xf ) and vi = v(xi ).93) This is the wave function which is derived in quantum mechanics by applying the Wentzel-Kramers-Brillouin (WKB) approximation.90) xf dx xi v 3 = −vi vf where in the second step we diﬀerentiated with respect to t. xi ) = 1 exp i v(xf )v(xi ) xf xi p(x) dx ℏ ∗ ∼ ψW KB (xf )ψW KB (xi ) (3. again.89) (3. Problem 1 of Homework 7 is devoted to the standard derivation of the WKB approximation in the Schroedinger approach. the ﬁnal result is: K(E. and then set t = t∗ . (3.82).92) where we used the result eq. (3.91) Putting this into eq.

94) corresponding to a physical situation where the De Broglie wavelength λ varies by a relatively small amount over a distance of order λ. Of course. Sections 3. In particular this is so at the turning points of the classical motion where E = V and p = 0. in order to compute the propagator or the wave function at points where the WKB approximation applies one still has to deal with matching conditions close to the turning points where the approximation breaks down. This will be studied in detail in the exercises. Let us assume for that that we have a potential V (x) that vanishes as x → −∞. and that has a barrier somewhere along the x axis. keeping the classical parameters of the problem ﬁxed.64 CHAPTER 3. 3. THE SEMICLASSICAL APPROXIMATION domain of validity of the WKB approximation. The above equation can also be written as a constraint on the force F = −V ′ (x) acting on the particle mF ℏ dp = ℏ 3 ≪ 1. as shown in ﬁgure 3. In the path integral formulation the issue arises in that one has to consider in the semiclassical computation of the propagator trajectories that go through turning points. which is reassuring. this requirement would be eventually satisﬁed by formally taking ℏ → 0. p2 (x) dx p (3. xf → −∞. is the quantum mechanics analogue of geometric optics limit for the propagation of electromagnetic waves. The standard method to overcome the breakdown of the WKB approximation at the turning points is to deform the trajectory into the complex plane for the coordinate x (and for time t as well) in such a way as to avoid the point where v(x) = 0. Then. and one reﬂected ﬁrst by the barrier at a and then going to xf . Now. in all interesting applications the wave function at the turning points is not directly needed. the classical action from xi to xf accepts two solutions: one going directly from xi to xf . However in practically all interesting applications.95) We thus see that at the points where the classical momentum is very small the WKB approximation breaks down. . where the wavelength goes to zero. We shall not worry about the subtleties of reﬂection points. This limit. In that limit the propagation of light can be described by the propagation of particles along rays: the wavelength being small diﬀraction phenomena are negligible.5 will delve with more details into those issues. The contribution of the reﬂected path. xf > xi .2.2.4 and 3.3 Two applications: reﬂection and tunneling through a barrier In this section we shall consider two basic applications of the WKB propagator formalism. Let us ﬁrst examine the process of reﬂection by a potential barrier. and compute whatever we can compute. the Green’s function will be decomposed into the contributions of the two classical paths: K(E) = K1 + K2 . The result is that one necessary requirement for WKB to be valid is dλ(x) ℏ dp ≡ ≪1 p2 (x) dx dx (3. When we put xi .2.2.

xi ) = where xf xi π 1 e−i 2 exp v(xi )v(xf ) i ℏ xf p(x)dx(2) xi (3. the two stationary solutions must coincide up to the sign. (3. is interesting because it contains information about the phase shift induced by the reﬂexion. xf . x.2: The two possible classical paths from xi to xf . or more precisely this operator has one negative eigenvalue. Applying the same argument. if we assume x > y.98) Imagine now.99) Now.74) we can write: K(E. THE FIXED ENERGY PROPAGATOR V 65 xi xf a Figure 3. whereas the second one is going ﬁrst from xi to a. It turns out that: K2 (E. and then is reﬂected from a to xf . Now. that the motion is bounded also to the “far” left at b ≪ a. and remember eq. as they solve the same problem (energy bound eigenstate in one-dimensional Quantum . K2 . The origin of the extra −i factor is that det −m∂t 2 − V ′′ (x) < 0 over the reﬂected trajectory. K1 is the same as without barrier.2. we have found the physical eigenstate ψstat of the Schr¨dinger probo lem: ψout (y) = ψstat (x) = a i 1 ℏ Ry |p(z)|dz−i π 4 e v(y) 1 cos v(x) 1 ℏ a x |p(z)| dz − π 4 (3. The ﬁrst one goes directly from xi to xf .96) p(x)dx(2) = a xi ﬂected path. y) = K1 + K2 = ψout (y)ψstat (x) = = = Ra Rx Ra i i π 1 e ℏ y |p(z)|dz + e−i 2 e ℏ ( y |p(z)|dz+ x |p(z)|dz) v(x)v(y) Ra Ra Ra π i i π i π 1 e ℏ y |p(z)|dz−i 4 e− ℏ x |p(z)|dz+i 4 + e ℏ x |p(z)|dz−i 4 v(x)v(y) a i 1 ℏ Ry |p(z)|dz−i π 4 · e v(y) |p(x)| dx + a xf |p(x)| dx is the integral over the re- 1 cos v(x) 1 ℏ a x |p(z)| dz − π 4 (3.3.97) Thus. y → −∞. to a wave that is reﬂected at b we have the following result for ψstat ψstat(b) (x) = 1 v(x) cos 1 ℏ x b |p(z)| dz − π 4 (3. |p(x)| is deﬁned as 2m(E − V (x)) > 0. x.

102) This condition gives a constraint for the integral over the closed contour: p(z)dz = 2 ℏ a b |p(z)| dz = 2 ℏ x b |p(z)| dz + 2 ℏ a x |p(z)| dz = 2π m + 1 2 (3. we ﬁnd: ∂x φa = −∂x φb where θ is independent of x. We then get: ψa (x) = cos (φa (x)) ψb (x) = cos (θ − φa (x)) = cos θ cos φa − sin θ sin φa ⇒ θ = mπ.100) Taking the derivatives. However. which can be seen in the WKB approximation. Thus.103) This is the famous Bohr-Sommerfeld semi-classical quantization condition. Quantum physics in the semi classical limit predict a non-vanishing probability for such a particle to tunnel through the barrier from xi to xf . m ∈ Z ⇒ ψa (x) = ±ψb (x) ⇒ φa (x) + φb (x) = θ (3. we have to look for the stationary . The equivalence between WKB and Gaussian approximation requires therefore to ﬁnd the corresponding contribution in the path integral approach. We have: ψa (x) = cos 1 ℏ 1 ψb (x) = cos ℏ π 4 x x π |p(z)| dz − 4 b |p(z)| dz − a = cos (φa (x)) = cos (φb (x)) (3. THE SEMICLASSICAL APPROXIMATION Mechanics). E V (x) xi a b xf Figure 3.3: There is no classical trajectory going from xi to xf with an energy E.101) (3. Barrier penetration: semiclassical transmission coeﬃcient Imagine now that the potential V (x) has a barrier so that a classical particle with an energy E cannot go from −∞ to +∞ (see ﬁgure 3. The potential barrier is too high to let a particle go past it. we can deﬁne a semiclassical contribution to this process.66 CHAPTER 3.3).

we ﬁnd that: dx = dt 2(E − V (x)) m (3. we still have: i (Sc (t) + Et) ℏ = stat i ℏ xf xi 2m(E − V (x)dx (3. xi ) = D[x]e ℏ (Sc +Et) dt i (3. we have: dx = +i dt 2 |E − V (x)| m (3.3.2.4.105) Considering extending this equation at a < x < b.104) The result is that we can ﬁnd a stationary point provided t is complex! Indeed. from the classical equation of motion. regardless of the course of the complex value of t. THE FIXED ENERGY PROPAGATOR point of the combined D[x]dt integral: ∞ 0 67 K(E.107) real imaginary real The evolution of time in the complex plane along the trajectory is depicted in ﬁgure 3. xf .108) The Jacobian factor from the gaussian integration around the extremum of .4: The evolution of the time along the trajectory in the complex plane. we get: xf t= dt = xi m dx 2(E − V (x)) v(x) b a a = xi dx −i v(x) dx + |v(x)| xf b dx v(x) (3. Im(t) Re(t) b dx a |v(x)| − Figure 3. Integrating over the full trajectory. Now. It acquires a negative imarinary part as the particle travels through the potential barrier.106) where we chose the plus sign in order to ensure that the Green’s function remain analytic.

However. when multiple trajectories contribute to the amplitude it is essential to be in control of their relative phases.110) 3. In this subsection. (1. Eq.98).2. For trajectories where the velocity ﬂips sign n times the phase is either einπ/2 or ei(n−1)π/2 (depending on how many multiples of the half period ﬁt into T ) . (1. The speciﬁc phase of eq. tf .58). We can then derive the probability P of tunneling: Rb Rb√ 2 2 P = |A+ |2 = e− ℏ a |p(x)|dx = e− ℏ a 2m(V (x)−E)dx (3. ti ) We have so far been rather cavalier concerning the phase of our Gaussian integrals. (1. In that derivation we did not pay any real attention to the phase factors that arose when some eigenvalues became negative.89) tells us that the number n− of negative eigenvalues equals the integer part of ωT /π (T ≡ tf − ti ). For trajectories in which the velocity does not ﬂip sign there is no extra phase factor 2. Thus when T increases by a half-period π/ω the propagator aquires an extra phase factor = −i. xi . (1. tf . as our previous discussion makes clear. xi . 1) Getting the basic idea In order to get an idea.98). for n− = 0 the phase of the harmonic oscillator prefactor relative to that of the free particle is exp(−in− π ).109) where the transmission coeﬃcient is given by A+ . Therefore the correct result for the 2 propagator prefactor is given by Jω (T ) = e−in− 2 π mω 2πiℏ| sin(ωT )| (3. By eq. This result can also be stated as 1. (1. let us go back to the determinant prefactor for the harmonic oscillator which was computed in eq.4 On the phase of the prefactor of K(xf . xf . which is very relevant to the discussion several applications like the computation of energy levels in the WKB approximation (see Homework 7).111) rather than by eq. xi ). ti ). With an extra full period 2π/ω the propagator just ﬂips sign: (−i)2 = −1.98) is only correct for this range of T . xi ) = 1 exp v(xf )v(xi ) 1 e v(xi ) i ℏ i ℏ xf xi 2m(E − V (x)dx A+ R xf b 1 |p(x)|dx − ℏ = Ra xi |p(x)|dx 1 i eℏ v(xf ) e Rb a |p(x)|dx ∗ ψin ψout (3. THE SEMICLASSICAL APPROXIMATION t still compensates the determinant: K(E.68 CHAPTER 3. xf . Thus for 0 < T < π/ω we have n− = 0 so that the phase of the prefactor is the same as for the free particle case. The next subsection is devoted to the phase of K(E. we will give the basic idea for the case of K(xf .

15. so that motion is not periodic.16) for the diﬀerential 5 This problem was proposed in one of the exercise sessions. The novelty with respect 2 to the harmonic oscillator is that the motion at x < 0 is not bounded. ˙ intervals 1. t.16). The classical motion is characterized by 3 time.5. There is a well developed mathematical theory. For t1 < t < t1 + ω we have x(t) = the harmonic potential at x > 0.3.15. <0 We want to compute the determinant function ψ0 (tf ) as tf is varied in the above 3 intervals. [9]. n− is referred to as the Morse index. We are not going to linger on this mathematical connection though. The initial value problem in eqs. 3. For t ≥ t1 + ω the trajectory is again “free”. 3. π 2. (3. Let us compute the ﬂuctuation determinant using eqs.2.5: . We also want to do so by focusing on a 1-dimensional family of trajectories characterized by initial position x(0) = xi < 0 and velocity x(0) ≡ v > 0 at t = 0.112) where the product of step functions accounts for the fact that V = 0 only for π t1 < t < t1 + ω . For a trajectory corresponding to exactly k periods of motion the phase factor is (−1)k This result is valid for any periodic motion not just for the harmonic oscillator. . 2) An explicit non trivial example 5 V (x) = θ(x) mω x2 2 2 xi xf x=0 Figure 3. v ω sin[ω(t − t1 )]: the particle moves in π ω) π 3. x(t) = −v(t − t1 + but the velocity is now negative. The operator controlling the quadratic ﬂuctuation is −m d2 d2 π − V ′′ (x(t)) = −m 2 − mω 2 θ(t − t1 )θ(t1 + − t) dt2 dt ω (3. underlying the dependence of the number of negative eigenvalues n− on the properties of the trajectory. called Morse theory. (3. For 0 ≤ t ≤ − xi ≡ t1 we have x(t) = xi + vt ≤ 0: the particle moves in v the V (x) = 0 region with positive velocity. THE FIXED ENERGY PROPAGATOR 69 3. Let us see now a slightly less straightforward case by considering the motion in the potential V (x) = θ(x) 1 mω 2 x2 shown in ﬁg.3. A simple paper with some more details is for instance ref.

For 0 ≤ t ≤ − xi ≡ t1 the solution is ψ0 (t) = t v 1 π 2.32). From that equation is it obvious that ψ0 > 0 for trajectories in which the velocity does not ﬂip sign. • At tf = t∗ the smallest eigenvalue λ1 vanishes. Third this procedure is repeated to get the solution in region 3. That is to say: ψ0 > 0 before reﬂection. . Second. The picture of what happens is then obvious: • For tf < t∗ the eigenvalues are all positive 0 < λ1 < λ2 < λ3 < .113) Another comment concerns the point at which the determinant ﬂips sign.112) reduces to the free particle quadratic ﬂuctuation which is a positive deﬁnite diﬀerential operator. as we already know very well.114) v(ti )a(−tf + tr ) 1 v(ti ) 1 + ﬁnite = > 0(3. One ﬁnds respectively in region 1. . (3. the determinant crosses zero once and becomes negative. In region 3. . 2 and 3 1.115) 2 −t + t a a f r and therefore the determinant can only ﬂip sign after reﬂection.112 is solved by patching the solution in the 3 time regions listed above. This is a general result that can be easily seen by using the integral formula in eq. : there is one and only one negative eigenvalue. t1 < t∗ < t1 + π/ω. π Notice that this happens at t∗ strictly larger than the time tr = t1 + 2ω at which reﬂection by the potential barrier happens. namely limtf →tr ψ0 (tf ) > 0 when tf approaches tr from below. ψ0 (tf ) decreases and thus stays negative without ever reaching zero again. This is because for tf < t1 eq. (3. For t → tr the velocity goes to zero linearly: v(t) ≃ a(−t + tr ). 3. For t ≥ t1 + π ω one has ﬁnally ψ0 (t) = −t + π ω <0 Let us now increase tf starting from tf < t1 . For t1 < t < t1 + ω one has: ψ0 (t) = ω2 + t2 sin[ω(t − t1 ) + ϕ] where 1 0 ≤ ϕ ≤ π/2 is deﬁned by tan ϕ = ωt1 .70 CHAPTER 3. . and for any tf > t∗ the spectrum satisﬁes λ1 < 0 < λ2 < λ3 < . Let us show that a stronger result holds. (3. By the above result we have that the determinant ψ0 (tf ) is positive and grows with tf in region 1.28)) I[xc ] = e−i 2 π m 2πiℏ|ψ0 (tf )| (3. the solution in region 2 is found with initial conditions at t = t1 dictated by continuity of ψ0 ˙ and ψ0 . (3. For tf > t∗ the gaussian integral prefactor is then (compare to eq. At some tf = t∗ in region 2. but all eigenvalues are positive. ﬁrst the solution in region 1 is found. Then the integral ˙ in eq. .32) diverges but its prefator vf vi goes to zero so that the result is ﬁnite tf limtf →tr limtf →tr v(tf )v(ti ) ti 1 dt vc 2 (t) = (3. Indeed not only is the determinant positive in region 1. In other words. Notice that ψ0 (t1 + π/ω) = −ψ0 (t1 ) = −t1 < 0. with a a positive coeﬃcient (so that the acceleration is v = −a < 0). THE SEMICLASSICAL APPROXIMATION operator 3.

We have ǫ→0 lim dt = a2 (t − tr )2 Cǫ Cǫ π 1 1 −ie−iθ dθ = 2 [−1 − 1] . The integral on the oriented half circle contour Cǫ is easily computed for ǫ → 0 by parametrizing t − tr = −ǫeiθ . (3.117) This result precisely cancels the diverge of the integrals at t < tr −ǫ and t > tr +ǫ so that we have in the end ψ0 = vi vf lim ǫ→0 − 1 + ǫa2 tr −ǫ ti dt 1 + − 2+ v 2 (t) ǫa tf tr +ǫ dt v 2 (t) (3. These equations are ill deﬁned in such cases: the factor vf vi is non-zero. by eq.re 3. Since for a reﬂecting trajectory vi vf < 0. This suggest that be should be able to ‘save’ eq. in the classically accessible region. i. Notice however that no singularity should exist in ψ0 is association to the vanishing of the velocity.6. (3. (3.119) Using this formula. The reﬂection point xr is deﬁned by v(xr ) = 2(E − V (xr ))/m = 0. The equation of motion gives V ′ (xr )/m = a. Then we have that tr ± ǫ both correspond to xr − a ǫ2 so that the above equation becomes 2 ψ0 = vi vf − 1 + ǫa2 xr − a ǫ2 2 xi (3. (3.34) in the case of trajectories where v does change sign.e. if the potential at x → −∞. or from above. as in the ﬁgure. This is because the singularity is a double pole and thus the integral over a full oriented circle around it gives zero by Cauchy’s theorem. The way to do so is very simple: we just need to slightly deform the trajectory into the complex (x.6: . we have that ψ0 is negative if and only if the expression in curly bracket is positive. but the integral diverges at the points where v → 0. 71 3) The integral expression for ψ0 and reﬂecting trajectories One ﬁnal remark concerns the way to make sense of eq. (3.116) (3. so that dt = −iǫeiθ dθ. is ﬂat enough.115) we see that the integrand has a double pole: we just need to go around it as shown in ﬁg. with 0 ≤ θ ≤ π on Cǫ .32) and eq.32).118) We can also change variable t → x.32) and eq. (3. When working with the time integral eq. Now. THE FIXED ENERGY PROPAGATOR t R − ǫ t R tR + ǫ Cǫ Figure 3. t) plane in such a way as to avoid the singularities. where E is the energy of the trajectory in consideration. This is because ψ0 is just given by the solution of an ordinary diﬀerential equation with ﬁnite coeﬃcients. we can study the conditions on V (x) under which ψ0 ﬂips sign after reﬂection.34). then 1 dx + − 2+ |v 3 (x)| ǫa xr − a ǫ2 2 xf dx |v 3 (x)| . This is because in deriving them we had to divide by v somewhere. Notice also that it does not make any diﬀerence to go around the pole from below. 3.2. 2 ǫa 0 ǫa dt = v 2 (t) (3.

.5 by using instead the integral representation of ψ0 . xf . xi ) Let us now focus on the phase of K(E. xi ). one notices that v(x) has a cut in the complex x plane along the x region accessible to classical motion. Then ψ0 is computed on a trajectory that goes around this cut. xf .2. xf . that is with the region where ℏ|mV ′ (x)/p(x)3 | ≪ 1. xi ). We shall not do that but just give some suggestions on how to try and do it. However if limx→−∞ V (x) = −∞ fast enough. as we already pointed out. In this case the expression in curly braket does not necessarily become positive for large enough xi and xf and ψ0 may never ﬂip sign. The saddle point approximation can however still be reliable if the trajectory can be deformed into the complex C2 plane for (t. for xi or xf large enough and negative. A simple and obvious example where this is so is given by the linear potential V = cx. like the action integral. 2) p(x) is never too small 6 We stress that the situation is diﬀerent with respect to the apparent singularity of ψ (t ) 0 f at the turning points which we encountered in the previous section. a suitable deformation of the trajectory can be chosen such that x goes around the turning points in such a way that: 1) Physical quantities. witness the fact that this approximation is exact for the harmonic oscillator. Instead the WKB computation of K(E. In order for this procedure to makes sense two conditions must be satisﬁed (see the discussion in the chapter on the WKB approximation in Landau-Lifshitz [10]) • The potential V (x) is analytic in a region around the turning point. xf . THE SEMICLASSICAL APPROXIMATION it is obvious that. xi . are unaﬀected by the deformation of the trajectory thanks to analyticity. (3. This way. 3. Using the space integral representation eq. xi ) breaks down6 because p(x) → 0 (see eq.72 CHAPTER 3. which genuinely breaks down at the turning points. (3. the WKB approximation to K(E. Thus we have ψ0 < 0. the expression in curly brakets is dominated by the positive integral in the region away from the turning point. Physically this corresponds to the velocity being double valued: v(x) = ± 2(E − V (x))/m.The willing student can now go ahead and make a simple remark (on Cauchy integrals) that allows for a direct computation. The gaussian approximation for K(xf . • The region of analyticity for V (x) overlaps with the region where the WKB approximation applies.95)). Again the delicate issue is how to deal with the turning points. It is a good exercise to recover the result we found for ψ0 in the potential in ﬁg. . The ﬁrst is that for xi or xf close to the turning points. requires an extra approximation. which is really what matters in most applications. When the above two conditions are satisﬁed.. 3. x) in such a way that p(x) never really crosses zero. v(x) = 2(E − V (x))/m is also analytic in this region apart from having a cut (due to the square root) in the physically accessible region on the real x axis.34). Two basic remarks are in order here. then dx/v 3 is dominated by the region close to the turning point. ti ) does not really break down in the presence of turning points. tf .5 On the phase of the prefactor of K(E. This is obvious because in this case V ′′ = 0 and we have the free particle result ψ0 (tf ) = tf − ti > 0.

in general. Also. It turns out that the right answer is: • At each reﬂection the prefactor gains a phase factor (−i) = e−iπ/2 . In the case at hand that expression is clearly positive since. Since ∂ 2 S/∂t2 is positive. Without loss of generality and to simplify the discussion let us consider a particle of mass m = 1. v(xf )v(xi ) (3. The second comes from the phase of the gaussian integral over t and is determined by the sign of ∂ 2 S/∂t2 . using the computations we have already made.122) 1 where again. This is in agreement with the general result we stated above There is one quick way to deduce in general that for vi vf < 0 the phase in eq. However because of analyticity the result of all integrals is independent of ǫ anyway! .4 and found that it is e−iπ/2 for xf and xi negative. (3. let us shift the deﬁnition of the energy such that E = 0 for our trajectory. By continuing the trajectory into the complex plane we can still make sense of the integral dx/v 3 even when the velocity ﬂips sign somewhere. (3. Notice. The only question is that 1/ −1 being double valued. without loss of gererality. It is easy to see that ∂ 2 S/∂t2 > 0. The second remark concerns the form of the prefactor itself.6 are an example of how to go around turning points. Indeed by eqs. all is left in the prefactor is the velocity at the initial and ﬁnal points P (xf . (3.2. One is the phase of the VanVleck-Pauli-Morette prefactor in eq. the phase of the t integral is trivial and the only non trivial contribution is that from the determinant.119). the above prefactor gains a 1/ −1 √ factor precisely at reﬂection. the trajectory x(t) which is deﬁned to solve d2 x (3.2. as we already know.3. simply by staring at eq. deformed into the complex x plane. vi vf < 0 and since ψ0 (tf ) < 0 by our direct solution of the diﬀerential equation.120) m 2 = −V ′ (x) . There are two contributions to the phase. let us shift x so that the turning point is xr = 0 and let us choose V (x) ≃ x in the vicinity of x = 0. we cannot yet tell. To be manifestly in the WKB approximation ǫ should be large enough. THE FIXED ENERGY PROPAGATOR 73 and the WKB is manifestly valid.7 We will explain with an example below how the deformation works.3. we can focus on the solution x = − 2 t2 with v ≡ x = −t. which equals e−iπ/2 .4. For real t the velocity ﬂips sign at t = 0.121). like for a non-reﬂecting trajectory.2. that by deforming the time line into the complex plane (like we did for instance with barrier penetration).90) we have that ∂ 2 S/∂t2 is just the inverse of the expression in curly brackets in eq.121) must be e−iπ/2 . it is not obvious how to proceed for trajectories with multiple reﬂections. for the moment. if the phase is +i or −i. If we deform t into ˙ 7 Notice that the time contour shown in ﬁg. (3. (3. just by staring at the above equation.88) and (3. Then.121) √ Since v(xf ) changes sign at the turning point. We computed this phase factor in section 3. dt is also. This result can be directly and easily checked for the example we discussed in section 3. xi ) = −1 ∂ 2 Sc · 2πiℏ ∂xf ∂xi 2πiℏ ∂ 2 Sc ∂t2 = 1 .113). The equation of motion close to x = 0 is just x = −1 ¨ (3.

we are making some element dt of the sum dt = tf − ti complex. This is just an inﬁnitesimal way of saying that the euclidean path integral exists only for β > 0.124) . We are now just √ left with computing the phase shift in 1/ v(t) = 1/ −t along trajectory (a). we should go back and think what it means to deform t into complex values in our original deﬁnition of the path integral. This is because if Re (−idt) > 0 in the exponent e(−iHdt/ℏ) . That is to say only trajectories such that Im (dt) ≤ 0. therefore the phase of −∞ + iǫ is eiπ/2 . then the p integral in eq. rather than doing our simple (and real) slicing dt = (tf − ti )/N . It means that. These two options lead to diﬀerent (and conjugate) phase shifts for 1/ v(tf ). (1.123) More formally we can compute the phase shift ∆ϕ by integrating over the t contour √ ∆ϕ ≡ Im ∆ ln(1/ v) = −Im dv = −Im 2v +∞−iǫ −∞ 1 π dt =− π=− 2t 2 2 (3. The resulting phase of the prefactor is then 1 v(−∞ + iǫ) = e−iπ/2 1 |v(−∞)| (3.7: . we should only consider trajectories (real or complex) for which the path integral representation makes sense.7a or above like in ﬁg. 3. THE SEMICLASSICAL APPROXIMATION t=0 −iǫ (a) (b) iǫ t=0 Figure 3. A constraint arises however when we attempt to write this product of matrices in terms of a path integral as done in section 1. As long as we write U (tf − ti ) as a product of matrices this is not a problem. To understand why. Which of the two deformations should we choose? It turns out deformation (a) is acceptable while (b) is not.7b.74 CHAPTER 3.4.1.14) is not convergent: it is growing at p → ∞ (the ǫ interval of that equation corresponds to our dt). √ When t goes from −∞ to +∞ − iǫ the argument in −t goes from +∞ to √ −∞ + iǫ by passing in the upper half plane. This the case for trajectory (a) and not the case for trajectory (b). Either we go below like in ﬁg. Since our semiclassical approximation is based on a path integral representation of the evolution amplitude. the complex plane we have two options to go around the point t = 0 where the velocity vanishes. 3.

perturbation theory can miss important features of a physical system.1: The stability of the ground state localized around the local minimum of the potential is broken by the possibility for the particle to tunnel through the barrier to the part of the potential unbounded from below. faster that 75 . In doing so.1). V (x) tunnel E Figure 4. To illustrate this.Chapter 4 Instantons 4. however.1 Introduction In Quantum Mechanics. √ Vm ℏ dx The probability of tunneling will be proportional to: P ∼ exp −2 exp 3 5 −m ω ℏλ2 ∼ . as λ → 0. or get any direct indication of one basic pathology of this ground state: its instability! Due to the possibility of tunneling through the barrier. For λ “small” enough. It thus represents an eﬀect which vanishes. we would not realize. we can apply perturbation theory and in principle compute the ground state energy to any desired order in λ. the ground state will be able to decay down to the region where λx3 is negative and dominates the harmonic term (see ﬁgure 4. let us give a quick look at a simple example: a one2 2 dimensional particle in a potential V (x) = mω2 x + λx3 . The point x = 0 is a classical minimum and it seems reasonable to study the dynamics around this point upon quantization.

76 CHAPTER 4. V (y) is a harmonic oscillator with extra cubic and quartic terms which we can treat as perturbations as long as λ is “small”. we know that the ground state wave function has no nodes. in the case at hand.2). INSTANTONS any power of λ.1).2: The double well potential has two relexion symmetric vacua. are all doubly degenerate to all orders. is therefore even under x → −x: Ψ0 (x) = Ψ0 (−x) 1 Technically.2) where in the last equation we traded a for the harmonic oscillator frequency ω. We introduce thus the new variable y = x − a. this other guy expands the coordinate as x = −a − y ′ . and. we can now compute the corrections to the lowest energy levels to arbitrary orders in λ without having any direct evidence that there is another minimum with its own localized levels! The situation is here symmetric: a guy sitting at the other classical ground state would compute in perturbation theory exactly the same levels1 . thus leading to inexistant degenerate eigenstates in perturbation theory. This potential has two classical ground states: x = ±a. We have: V (y) = λ (y + 2a)2 y 2 = 4λa2 y 2 + 4λay 3 + λy 4 √ 1 ≡ mω 2 y 2 + ω 2mλ y 3 + λy 4 2 (4. which are equivalent because of the reﬂection symmetry of the system. Thus. and in particular the ground state. implying that it cannot be seen at any ﬁnite order in perturbation theory. (4. However. are the same. For small λ. say x = a. Hence the levels. In this variable.1) (4. let us look at the case of a particle in a double well potential: V (x) = λ(x2 − a2 )2 symmetric under reﬂection x → −x (see ﬁgure 4. there is no degeneracy of the bound state energy eigenvalues. we can ignore this doubling and expand around one of these minima. As a second example. Because of reﬂection symmetry the potential in y ′ satisﬁes V (−a − y ′ ) = V (a + y ′ ): exactly the form in eq. In perturbation theory. we know that for a unidimensional system. the energy levels. V (x) −a a Figure 4. This is what theorists normally call a non-perturbative eﬀect. in perturbation theory. . but corresponding to a diﬀerent set of states where the wave function is localized near x = −a instead of x = a. from the exact quantum mechanical solution of the problem. in perturbation theory. Moreover.

xi . we must sum over all possible stationary trajectories (i. To do this.3) (4.2.4) SE [x] = −β 2 dτ mx2 ˙ λ 2 + x − a2 2 4! For each trajectory x. 4.6) is dominated by the local minima of SE [x] with given boundary conditions: xf = x β 2 β 2 (4. Now.2 Instantons in the double well potential What we want to study are the properties of the ground state(s) of the double 2 λ well potential: V (x) = 4! x2 − a2 . δx Notice that the solutions to this equation describe motion in the euclidean time .5) Notice that SE is positive deﬁnite. INSTANTONS IN THE DOUBLE WELL POTENTIAL 77 On the other hand. let us go back to the Euclidean path integral.e.7) xi = x − Imposing now that δSE = 0 leads to the Euclidean equation of motion. the energy eigenvalues of these two lowest lying states mus be degenate to all orders in perturbation theory.4. with stationary action) in Euclidean time. To compute those two transition amplitudes. β 2 xi . That is E1 − E0 = 0 to all orders in perturbation theory. we can perform a semiclassical approximation in Euclidean time. More precisely. because of the doubling of all the levels that we just deduced with our perturbative reasoning. Remember that: xf | e − βH ℏ |xi = KE (xf . via semiclassical trajectories in imaginary time called instantons. the ﬁrst excited state has one node.− β 2 D[x]e− 2 SE [x] ℏ (4. If the barrier is high and the the potential smooth. and must therefore be odd under reﬂection: Ψ1 (x) = −Ψ1 (−x) However. The goal of this chapter is to show that the leading contribution to E1 − E0 as λ → 0 can be computed by semiclassical methods. βH βH We will compute a| e− ℏ |a and −a| e− ℏ |a for β → ∞. we can compute its contribution to KE in Gaussian ¯ approximation: ˜ KE = N d2 det −m 2 + V ′′ (¯) x dτ −1 2 e− SE [x] ¯ ℏ (1 + O (ℏ)) (4. KE = D[x]e− SE [x] ℏ (4. β) = β 2 xf .

accepting three local minima called 1. .3). Then. Suppose that we have a function S(x). we therefore expect KE to be dominated by the local minima of SE (if there exist more than one). in the limit ℏ → 0. .9) −∞ 2πℏ S ′′ (xi ) (4. parameter τ of a classical system with reversed potential VE (x) = −V (x) (see ﬁgure 4. .3: The Euclidean equation of motion corresponds to the real time equation of motion with a reversed potential VE (x) = −V (x). Ii = e− S(xi ) ℏ ∞ e− S(x) ℏ dx (4. the integral I= can be approximated by I ≈ I1 + I2 + I3 + . KE ≈ K1 + K2 + K3 + .8) Remark: we can make an analogy with the ordinary integrals.78 VE (x) CHAPTER 4. S(x) 1 2 3 x Figure 4. .4. . 2 and 3. as depicted in ﬁgure 4. (4. INSTANTONS −a a Figure 4.10) The anharmonic terms around the local minima give corrections to each Ii of relative size ℏ.4: An ordinary function of one variable S(x). In the limit ℏ → 0.

which rolls from a to −a.5. the value of the exponent SE = ℏ = a −a SE ℏ is given by: 2mV (x) 2 dx = ℏ 3 3m λ 3/2 λm a3 3 ℏ (4.11) a 2mV (x)dx −a (4.15) 2 The instanton (with the plus sign) is depicted in ﬁgure 4. This potential accepts two types of solutions: Trivial solutions: x(τ ) ≡ ±a which will give the leading contribution to the ﬁrst two amplitudes. 2 λ For the speciﬁc example of a quartic potential V (x) = 4! x2 − a2 . in the semiclassical limit. −a| e− ℏ |a βH and a| e− ℏ |−a . the zero energy solution satisﬁes x=± ˙ ω x2 − a2 2a 2 (4. depicted in ﬁgure 4. These solutions are exact in the β → ∞ limit. Remarks • For the instanton solution.14) where we used the new variable ω 2 = λa . 3m The solutions of these equations are the instanton and anti-instanton given by ω x(τ ) = ±a tanh (τ − τ0 ) (4. which starts at x = −a at τ = −∞ and rolls towards x = a at τ = ∞. and we can write: mx2 ˙ − V (x) = EE = 0 2 This allows us to write an expression for the Euclidean action: mx2 ˙ = V (x) 2 2V (x) m a ∞ 2V (x) dx = m mx2 dτ = ˙ SE = m −a −∞ ⇒ x= ˙ βH ℏ SE ℏ (4.13) We recover in this result the expected semiclassical tunneling coeﬃcient. and the anti-instanton. −a| e− ℏ |−a . we must study the classical trajectories in the upside down potential VE (x) = −V (x). then the Euclidean energy EE is zero.4. Less trivial solutions: the instanton. The action on the instanton can be computed: as the initial velocity x(−∞) ˙ is zero.3.16) 3 2 √ λm 3/2 ω3 m2 ω 3 2 =2 = ¯ ℏ ℏλ λ ¯ where λ = mℏλ 3 is the dimensionless parameter that characterizes the 2ω size of the perturbative corrections to the harmonic oscillator value for . INSTANTONS IN THE DOUBLE WELL POTENTIAL βH βH 79 βH To compute the transition amplitudes a| e− ℏ |a .2. the transition amplitude will be proportional to: a| e− |−a ∼ e− ∼ e− Ra −a p(x) dx ℏ ≪1 (4.12) Thus.

Since Feynman diagrams with k-loops contribute a ¯ ¯ correction of order λk−1 . up to exponentially small terms.6). a −β 2 τ0 β 2 −a Figure 4. the instanton solution looks like a step function. Instantons • At |τ − τ0 | ω .17) . then xIA (τ ) = af (τ − τ0 )f (τ1 − τ ) (4.6: In the large β limit. there is still plenty of time to bounce back! More technically. running from −a to a. Thus. we have that x (τ ) − a are thus well localized in Euclidean time: they exist for a short instant 1 ∆τ ∼ ω in Euclidean time (hence their name).5: The instanton solution in the double well potential. 1 2 2 ≈ e−ω|τ −τ0 | ≪ 1. The part of the Euclidean propagator due to the instantons is therefore ¯ proportional to e−SE /ℏ = e−2/λ . given the instanton solution xI (τ ) = af (τ − τ0 ). we have x ≈ ±a away from τ = τ0 . Thus. INSTANTONS τ Figure 4.80 x(τ ) a τ0 −a CHAPTER 4. ¯ the energy levels. This is a crucial property considering that we are 1 interested in the limit β → ∞: for β ≫ ω the instanton solution looks like a step function (see ﬁgure 4. We already encountered λ when we studied perturbative corrections to the ground state energy of the harmonic oscillator with a λx4 perturbation. when β → ∞. the coupling λ is often called a loop-expansion parameter. where 1 |f (τ − τ0 )| = 1 − e−ω|τ −τ0 | for |τ − τ0 | ≫ ω . eﬀects associated with instantons ¯ are non-perturbative in λ. After having jumped from −a to a.

three instantons at τ1 . one can easily check that δSE δx ∼ e−ω|τ0 −τ1 | .g. τ3 and τ5 and two anti-instantons at τ2 and τ4 will be a good one provided τ1 ≪ τ2 ≪ τ3 ≪ τ4 ≪ τ5 .7: The instanton-anti-instanton quasisolution. INSTANTONS IN THE DOUBLE WELL POTENTIAL xIA (τ ) a 81 τ0 −a τ1 τ Figure 4. For such a conﬁguration. Notice that if we take nI instantons and nA anti-instantons. (what about −a → −a and a → −a ?) An approximate solution with. the quasi-solution is close to ±a with a neglectible velocity. Thus.2. neglecting that xIA is not an exact solution amounts to neglecting terms of order O e−ω exp(SE /ℏ) ! These terms do not exactly vanish but are indeed very small. Each of these quasi-solutions contributes to the amplitude in two ways: Exponent: the contribution to SE is localized at each instanton (or antiinstanton) transition time. we must have that nI − nA = 0 for a → a. and nI − nA = 1 for −a → a. we are easily convinced that we must consider all the approximate solutions with a number N of instantons + antiinstantons. We will check later whether this condition is satisﬁed for the leading terms. then we should by all means consider the conﬁguration xIA as a stationary point of the action. e. Thus we must add the contribution of xIA to our saddle point estimate of the path integral.1 The multi-instanton amplitude We will now compute our amplitudes by summing over all possible sequences of widely separated instantons (and anti-instantons): this is the diluted instanton gas approximation. 4. for τ1 ≫ τ0 is an approximate solution representing an instanton followed by an anti-instanton. we shall ﬁnd at the end that |τ1 − τ0 | ∼ e ℏ .18) x=xIA As long as we only need to consider situations where ω|τ0 − τ1 | is so large that the above exponential is smaller than any eﬀect we are trying to compute.2. Indeed.4. SE is therefore zero in these regions. . (4. away from these transition times. • By extending this argument. SE Indeed. And thus add the contribution of these quasi-solutions to our semiclassical approximation to the path integral.

82 x(τ ) a CHAPTER 4.20) for a quasi-solution with n instantons and anti-instantons. around which V ′′ (x) = mω . Let us see how this argument can be made sharper. We deﬁne the integration coordinates at intermediate times as x(τi± ) = x± . The Euclidean propagator can thus be i .) To do so. β .19) Determinant prefactor: notice again that away from the very short instants over which the transitions take place. then the prefx 2 2 2 actor would be the usual one for the harmonic oscillator: ˜ N d2 det −m 2 + V ′′ (¯) x dτ −1 2 2 = mω 2πℏ sinh(ωβ) 1 2 ≈ mω πℏ 1 2 e− ωβ 2 Given the factorization property of the path integral and. 2 If we had V ′′ (¯) ≡ mω throughout the interval − β . Due to the decomposition property of the action. in particular. the contribution of an approximate solution with N instantons and anti-instantons will be: SN = N S I ⇒ e− SN ℏ = e−N SI ℏ (4. (The generalization to an arbitrary sequence of I and A is straightforward. We thus expect d2 ˜ x N det −m 2 + V ′′ (¯) dτ n = nI + nA −1 2 = mω πℏ 1 2 e− ωβ 2 Rn (4. the system sits at either one of its 2 classical vacua at x = ±a. we expect that each transition will correct the prefactor of the harmonic oscillator with a factor R. 2.8: A quasi-solution with three instantons localized at τ1 . for i = 1. we explicitly write the Gaussian integral in factorized form by considering intermediate ± ± times around instanton τ1 = τ1 ±∆ and around anti-instanton τ2 = τ2 ±∆ as shown in ﬁgure. Each instanton (or anti-instanton) will give a factor SI in the action. INSTANTONS τ1 τ2 τ3 τ4 τ5 τ −a Figure 4. and two anti-instantons localizd at τ2 and τ4 . τ3 and τ5 . by focussing on the simplest non-trivial case of an xIA solution. of the Gaussian integral we are interested in.

2. τ1 ) 2 1 = ≈ = x− e− 2 0 + | e− ∗ H0 ℏ − + (τ2 −τ1 ) x+ 1 ω 1 + O(e−ω∆ ) ∗ − + −τ1 ) H0 ℏ − + (τ2 −τ1 ) Ψ+ (x− )Ψ+ (x+ )e− 2 (τ2 0 2 0 1 |0+ Ψ+ (x− )Ψ+ (x+ ) 2 0 1 0 (4.22) where H0 represent the harmonic oscillator Hamiltonian around +a. with |τi − τj | ≫ ∆ ≫ ω . We can thus approximate. Here we also decomposed position eigenvectors |x into eigenstates of the harmonic oscillator Hamiltonian around +a: |x = n+ Ψ+ (x) |n+ .9: Intermediate times τi± = τi ± ∆ have been added to the instanton1 anti-instanton quasisolution. where the potential is well approximated by the harmonic oscillator one. so that at τi± . This is because. τ1 . − ) (4. and expanding into the similarly deﬁned harmonic oscillators eigenstates around −a. The factorization 1 turns out to be useful if we also choose ∆ ≫ ω . τ2 ) KE (x− . 1 1 2 2 (1) + − − + KE (x+ . amplitude (2).τ )× 2 2 β + − − KE (x+ . the amplitudes (1). x+ . Indeed it is intuitively clear that.4. x− . −a. the trajectory sits very close to one or the other minimum. x2 .21) 1 1 1 2 (3) 1 Of course we work under the assumption |τ2 − τ1 | ≫ ω . for ω∆ ≫ 1. β β . during the corresponding time intervals. x− .g. and we n kept only the ground state contribution. Moreover. x+ . τ2 . is dominated by the ground state (see section 2.3). (2) and (3) should be well approximated (exponentially well indeed!) by harmonic oscillator amplitudes. τ1 ) × 2 2 2 1 (2) β + . τ1 ) KE (x− . e. for large euclidean time separation. we know that the harmonic oscillator amplitude. −a. τ1 . τ2 . factorized into the convolution of the propagator over ﬁve time intervals (see ﬁgure) KE (−a. we . By applying the same approximation to (1) and (3) as well.− ) = 2 2 + dx− dx+ dx− dx+ KE (−a. the solution is almost in the vacuum. τ2 . by: − + KE (x− . INSTANTONS IN THE DOUBLE WELL POTENTIAL + τ1 83 x(τ ) − τ2 −β 2 − τ1 τ1 τ2 τ β 2 + τ2 Figure 4.

. . β β .84 CHAPTER 4.10: The contributing solutions for the tranition from −a to −a are solutions with any number of instanton-anti-instanton pairs. τ2 . We are not done yet! We must still sum over over the locations τ1 . τ2 )KE (x+ .10.25) How do we normalize the dτi integrals? This is not a problem for the moment: the normalization can be absorbed by the factor R (and this is not by chance!). x− . . + −a −a −a |−a ﬁrst. τ1 ) 2 2 1 1 = where R= mω πℏ 1 2 e− ωβ 2 R2 (4. .24) 0 1 0 1 1 1 1 1 ω + − ∗ R will be ﬁxed later on by studying more carefully the one-instanton amplitude. 4. −a. For −a| e− ℏ |−a .2 Summing up the instanton contributions: results βH ℏ Let us compute −a| e− in ﬁgure 4. starting at −a we βH must ﬁrst encounter an instanton. The summation over instanton and antiinstanton location simply (and obviously) corresponds to the integral: β 2 τn τ2 −β 2 dτn −β 2 dτn−1 · · · −β 2 dτ1 = βn n! (4.− ) ≈ 2 2 ∗ + dx− dx1 dx− dx+ Ψ− (−a) e− 0 1 2 2 ∗ ω + − 2 ωβ 2 + + Ψ− (x+ ) Ψ0 (x− )Ψ0 (x+ ) Ψ− (x− )e 2 ((τ2 −τ2 0 2 2 1 0 1 + − )+(τ1 −τ1 )) + − + − KE (x+ . etc. one thing we have to be careful about is that instantons and antiinstantons cannot be distributed arbitrarily: for example. INSTANTONS can thus write the Euclidean propagator as KE (−a. τ1 . Indeed. all contributions will have nI = nA . τn of the instantons.23) + − dx− dx+ e 2 (τ1 −τ1 ) Ψ+ (x+ ) Ψ− (x− )KE (x+ . Finally. The contributing solutions are depicted + I A −a −a I A I A −a + ··· Figure 4. x− . x− . as long as τ1 ≪ τ2 ≪ · · · ≪ τn . This proves factorization as promised. . we still get an equally good approximate solution by moving the positions τi around. τ1 . .2. We can now put together all the pieces and compute the total semiclassical amplitudes. τ1 ) (4. .

.28). we get: ±a| e− βH ℏ |−a = mω πℏ 1 2 e− βω 2 · −SI /ℏ 1 Re−SI /ℏ β β · [1 + O (ℏ)] e ∓ e−Re 2 (4.30) ± ℏRe− ℏ 2 Calling |+ and |− the corresponding eigenstates. we can write ±a| e− βH ℏ |−a = Ψ∗ (±a)Ψ+ (−a)e− + βE+ ℏ + Ψ∗ (±a)Ψ− (−a)e− − βE− ℏ (4. and calling ψ0 (x) ≡ mω πℏ 1 4 e− mωx2 2ℏ (4. 1 Ψ± (x) = √ (ψ0 (x + a) ∓ ψ0 (x − a)) 2 Let us now add a few comments about our results.2. we can read out (for β → ∞) the energies of the lowest energy states: SI ℏω E± = (4.28).33) as expected from the usual Schr¨dinger approach.27) Performing the summation.28) Recalling that xf | e− βH ℏ |xi = ∗ Ψn (xf )Ψn (xi )e− n βEn ℏ (4. INSTANTONS IN THE DOUBLE WELL POTENTIAL Thus. o Basically. we can compute a| e− a| e − βH ℏ βH ℏ |−a : Re−SI /ℏ β n! n mω |−a = πℏ 1 2 e − βω 2 · n=odd · [1 + O (ℏ)] (4.26) Similarly.4.29) and comparing to (4.32) the harmonic oscillator ground state wavefunction we get Ψ∗ (±a)Ψ− (−a) = − 1 2 1 2 mω πℏ mω πℏ 1 2 = 1 2 1 √ ψ0 (0) 2 2 Ψ∗ (±a)Ψ+ (−a) = ∓ + ⇒ =∓ 1 √ ψ0 (0) 2 2 Ψ− (x) is even Ψ− (a) = Ψ− (−a) Ψ+ (x) is odd Ψ+ (a) = −Ψ+ (−a) (4. the amplitude will be: −a| e− βH ℏ 85 |−a = mω πℏ 1 2 e− βω 2 · n=even Re−SI /ℏ β n! n · [1 + O (ℏ)] (4.31) Comparing to (4.

A purist (as Coleman puts it) would only retain the instanton contribution when writing the diﬀerence E+ − E− = 2ℏRe−SI /ℏ = 2ℏRe−2/λ ¯ (4. is that these perturbative correction aﬀect E+ and E− by exactly the same amount: they cancel in E+ − E− . however. and the third one is a three-loop correction. The perturbative corrections are thus the same as those we have computed before. the instanton solutions sit at either one or the other minimum. + + + ··· Figure 4. equation (4.37) R(λ) = R 1 + n=1 ∞ n=1 ¯ dn λn .11). we would have to compute the Green’s function around each solution. plus corrections coming from the cubic anharmonic term (see ﬁgure 4.30) is “incorrect” because in it. For instance.35) The point. INSTANTONS 1. we have neglected the perturbative corrections. we have only retained the instanton correction which is of order ℏe−S/ℏ ∼ ℏe−2/λ ¯ (4.34) and is smaller than the perturbative corrections (!) we previously computed by Feynman diagrams: ¯ ∆En ∼ ℏω λn (4. The ﬁrst diagram is a two-loop correction involving the λ4 x4 correction.11: The quantum corrections to the classical solutions. For most of the time. Notice that in our computation.86 CHAPTER 4. the second one is a two-loop correction involving the λ3 x3 correction.36) βH One could go ahead and compute the perturbative corrections to ±a| e− ℏ |−a with the same Feynman diagram techniques we used before. Technically speaking. a| e− a| e− βH ℏ βH ℏ |−a becomes: mω πℏ ℏω 2 1 2 |−a = e− βE0 (λ) ℏ n=odd R(λ)e−SI /ℏ β n! n E0 (λ) = 1+ ∞ ¯ cn λn (4.

38) The sum over the instantons is truly necessary only to reproduce the −βH 1 exponential behaviour of e ℏ at β > ∆E . For instance in +a|e−βH | − a a IAI conﬁguration contributes with weight e−3SI .4. We can now go back and verify that the sequences of widely separated instanton anti-instanton that are relevant to our result are stationary points of the action up to eﬀects of order δS δx ∼ e−ω|τ1 −τ2 | ≈ e−[ R e ω SI /ℏ ] ¯ (4. These have an euclidean action = nSI which is bigger that the single instanton one SI . determined by semiclassical methods. 3. ˆ det(O ) ¯ . Thus. which is exponentially smaller than the one instanton contribution e−SI . Moral: we see the two classes of eﬀects at work: • Non-perturbative eﬀects. . . The contribution of each such trajectory is thus weighted by a much smaller exponential factor.O λ2 terms correspond to three-loops diagrams. They correspond to the non-trivial stationary points of the Euclidean action. We can now also check the validity of our dilute instanton approximation. the typical instanton density is: ρ ≈ n = Re−SI /ℏ . ¯ The leading O λ0 term corresponds to evaluating the quantum ﬂuctuation determinant q 1 . • Perturbative eﬀects.And so on. They correspond to small quantum ﬂuctuations around each stationary point of the action: they are the functional integral analogue of the corrections to the saddle point approximation of ordinary integrals. The perturbative corrections are conveniently organized in a series of Feynman diagrams. and R is the leading contribution to the prefactor we introduced before without computing it explicitly. . and one may wonder if it makes any sense to include them at all. . eSI /ℏ 1 1 = = ρ R ∆E (4. 2. INSTANTONS IN THE DOUBLE WELL POTENTIAL 87 where cn are the loop coeﬃcient we have computed previously.2.O λ terms correspond to two-loops diagrams. Now. at the end of our computation.39) x=xIA which is truly small! Much-much smaller than the e−2/λ eﬀects we have been computing. (Re β) is dominated by the terms with n ∼ The exponential series n! Re−SI /ℏ β. ¯ . It is β exponentially small! Their average separation in time is: −SI /ℏ n ∆τ = which is independent of β. Another puzzle that one could have pointed out when we started our computation concerned the relevance of multi-instanton trajectories.

1 . 4. INSTANTONS we obviously know how to answer.3 Cleaning up details: the zero-mode and the computation of R To wrap things up.40) possesses one normalizable zero mode: ′′ −m∂τ 2 + VE (xI ) y0 (τ ) = 0 Indeed. Surprise: δ2 S δx2 x=xI . Each multi-instanton contributes only ∝ e−nSI to the amplitude. We can here draw a simple statistical mechanics analogy (just an analogy!): the action exponent plays the role of Boltzmann factor ∝ e−nSI ∼ e−E/T .42) We can then compute the normalization of this zero mode: xI 2 (τ )dτ = ˙ ⇒ y0 (τ ) = 1 m mxI 2 (τ )dτ = ˙ SI m (4.41) ⇒ (4. which implies that S(xI ) = S(x∆ ).88 CHAPTER 4.44) yn (τ )˜n a n y(τ ) 2 We are using the same normalization of section 3. we can compute from the equation of motion: ′′ − m∂τ 2 xI (τ ) + VE (xI (τ ))xI (τ ) = 0 ˙ ˙ ′ − m∂τ 2 xI (τ ) + VE (xI (τ )) = 0 ∞ −∞ |y0 (τ )| dτ = 1 2 (4. the ﬂuctuation determinant around the δ2S δx2 x=xI ′′ = −m∂τ 2 + VE (xI ) (4. δ S = 0 along this I δx2 direction in the function space. ∀∆ ∈ R. We can now expand the function x into modes around xI 2 : √ x(τ ) = xI (τ ) + ℏ (4. while the amplitude we are computing is the analogue of the partition function e−F/T .2. we are left with explaining the origin of the dτ integral on the instanton position and with computing R (the two things are obviously related).43) m xI (τ ) ˙ SI The presence of Ψ0 originates from the invariance of the action under time translations: xI (τ + ∆) = x∆ is a solution with same boundary conditions as I 2 xI (τ ). Thus. Boltzmann suppressed conﬁgurations contribute signiﬁcantly to the free energy at equilibrium if they have large multiplicity. however n-instantons have also a higher multiplicity ∝ Rn β n /n! which for Rβ > e−SI compensates for the exponential suppression. the multiplicity ∝ Rn β n /n plays the role of an entropy factor eS . We have to study det one-instanton solution.

48) + → λh (ω) = m ( )2 + ω 2 SE = n 2 2 β P h d˜n a 1 ˜ 2 ˜ √ e− 2 n λn an = N 2π −1 2 ˜ N λh n n = ˜ N det (−m∂τ 2 + mω 2 ) (4. Notice that d˜0 is associated to a shift in the time position of the instanton: a x(τ ) = xI (τ ) + ℏm xI (τ ) · a0 + · · · ≈ xI ˙ ˜ SI ⇒ τ + a0 ˜ ℏm SI + ··· (4.2.). Now..47) n N is ﬁxed as usual by some simple reference system. (4. .50) SI mω · mπℏ πℏ e− βω 2 · det (−m∂τ 2 + mω 2 ) det′ (−m∂τ 2 + V ′′ (xI )) where we multiplied and divided by the harmonic oscillator determinant.4. We have ˜ N = = P d˜n a 1 ˜ 2 √ e− 2 n λn an = 2π dτ0 n SI ˜ ·N mπℏ P 1 d˜n a ˜ 2 √ e− 2 n λn an 2π n≥1 dτ0 dτ0 SI · mπℏ ˜ N · det (−m∂τ 2 + mω 2 ) 1 2 det (−m∂τ 2 + mω 2 ) det′ (−m∂τ 2 + V ′′ (xI )) (4. By det′ we indicate the determinant with the zero mode removed: det′ = λ1 · λ2 · λ3 · . . INSTANTONS IN THE DOUBLE WELL POTENTIAL Thus. by a straightforward but tedious computation which we shall not do explicitly 3 one ﬁnds SI mπℏ det (−m∂τ 2 + mω 2 ) = det′ (−m∂τ 2 + V ′′ (xI )) mω πℏ 1 2 v≡R ˜ (4.49) n mω = 2πℏ sinh(βω) Let us now integrate in our case by separating out the integral on the zero mode.45) as λ0 = 0. The integration over a0 should be treated separately and with special care ˜ 1 (otherwise we would get a λ0 nonsense). for example the harmonic oscillator: mω 2 y 2 πn my 2 ˙ (4.. .46) xI (τ ) = f (τ − τ0 ) ℏm d˜0 = −dτ0 a SI d˜n a √ 2π Let us then choose an N such that the measure for the ﬂuctuation y(τ ) is: ˜ D[y] = N (4. exp − 1 δ2S 2 y 2ℏ δx2 = exp − 1 2 λn an2 ˜ n 89 1 1 ˜ ˜ = exp − λ1 a12 − λ2 a22 + · · · 2 2 (4.49) and took the βω ≫ 1 limit. .51) 3 To be found in Coleman’s lectures and also proposed in one of the Homeworks to be posted on line (next year. used eq.

we can compute: mω 2 v ˜ πℏ 1 2 = 4ma2 ω 3 πℏ 1 2 = 12m2 ω 5 πℏλ 1 2 = 12 ¯ πλ 1 2 ω (4.53) 1 − e−ω(τ −τ0) 1 + e−ω(τ −τ0) 1 τ −τ0 ≫ ω −→ x −→ ω 2aωe−ω|τ −τ0 | ˙ v = 2aω ˜ τ −τ0 ≫ 1 Thus.55) ve ℏ ˜ E+ − E− = 2ℏ πℏ .90 CHAPTER 4. INSTANTONS 1 ω: where v is determined by the asymptotic behaviour of xI for |τ − τ0 | ≫ ˜ ˙ xI = v e−ω|τ −τ0 | ˙ ˜ In the speciﬁc case of the x(τ ) = a ⇒ ⇒ λ 4! (4. we have: a 1 − 2e−ω(τ −τ0) (4.54) We thus ﬁnally obtain the energy splitting between the two lowest energy levels: 1 mω 2 − SI (4.52) x2 − a2 2 potential.

1 5. S → S − e (α(xf .3) ˙ c2 The coeﬃcient e is the electric charge of the particle. tf ) − α(xi .1 Gauge freedom Classical physics x2 ˙ dt c2 The relativistically invariant action for a neutral spinless particle is S = −mc2 dτ = −mc2 1− (5.1) γ γ The simplest relativistic invariant interaction with the electromagnetic quadrivector ﬁeld Aµ = (A0 .2) dτ − e γ Aµ dxµ = − mc2 γ 1− x2 ˙ + eA0 − eAi xi dt (5.1) to S = −mc2 γ where we have chosen the parametrization so that: dτ 2 = (dx0 )2 − 1 (dxi )2 = (∂σ x0 )2 − 1 (∂σ xi )2 dσ 2 c2 c2 ˙2 σ = x0 ≡ t ⇒ dτ = 1 − x dt c2 (5. which are local functions of Aµ that are not aﬀected by gauge transformations. We will see that the path integral approach oﬀers a neat viewpoint on the issue of gauge invariance and on the role of topology. ti )) Consequently the classical equations of motion only depend on the electromagnetic ﬁelds Fµν = ∂µ Aν − ∂ν Aµ . (5. Notice that the gauge transformation Aµ → Aµ + ∂µ α does not aﬀect the equations of motion since S only changes by boundary terms. 91 .Chapter 5 Interaction with an external electromagnetic ﬁeld We will now discuss the quantum mechanics of a particle interacting with an external electromagnetic ﬁeld. −Ai ) is obtained by modifying eq. 5.1.

moreover. but the lagrangian involves a redundant quantity: Aµ .9) k ∂xk ∂p ∂x ∂pk Now what is the transformation (5. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD The classically relevant quantity is the electromagnetic ﬁeld Fµν .10) . we can ﬁnd the conjugate momenta: pi = ∂L = ∂ xi ˙ mxi ˙ 1− x2 ˙ c2 + eAi (x.6) H = (pi − eAi )2 c2 + m2 c4 + eA → H ′ = H + e∂ α 0 0 (H. t) ≡ eα. Notice that the kinetic momentum is gauge invariant (being just a function of the velocity): Πi ≡ mxi ˙ 1− x2 ˙ c2 = pi − eAi → pi − e∂i α − e Ai − ∂i α = Πi (5. pi ) transforms like e(A0 .92CHAPTER 5. we can write: pi → p′i = pi − ∂i F (5. x) → H (p . Let us consider now what happens in the Hamiltonian description of this system. x ) = H(p. in the sense that the following relations are satisﬁed: p′i . x′j ∂p′i ∂p′j ∂p′i ∂p′j − k k = δik (−∂kj F ) − (−∂jk F )δik = 0 k ∂xk ∂p ∂x ∂p ′i ′j ∂p ∂x ∂p′i ∂x′j = − k = δik δjk = δij (5. p′i and x′i satisfy the Hamilton equations for H ′ : ′i ∂H ′ ∂p = ∂t ∂x′i ′i ∂x ∂H ′ = ∂t ∂p′i (5. t) (5.5) where A0 = V is the electromagnetic potential. Ai ).6) in the Hamiltonian language? Deﬁning F (x. x) + ∂0 F and.7) It is a canonical transformation. t) → p′i = pi − e∂i α x2 ˙ 1 − c2 (5.8) xi → x′i = xi ′ ′ ′ H(p.4) This leads to the hamiltonian: H = pi xi − L = ˙ (pi − eAi ) c2 + m2 c4 + eA0 2 (5. From equation (5. p′j = p′i . Remark: both pi and H are not invariant under a change of gauge for Aµ : ˙i i p = mx + eAi (x.3).

xi .12) We have the relation: Ψ(xf . but they are covariant: (5. the gauge transformation of the electromagnetic quadrivector ﬁeld is related to a canonical transformation of the Hamiltonian variables: gauge Aµ transformation Aµ + ∂µ α −→ i i x x (5. Ψ → U Ψ with U † U = U U † = 1 is the most general canonical transformation in the quantum description.1.17) . the matrix elements of any operator O transform like ˆ ˆ Ψ1 |O|Ψ2 → Ψ1 |U † OU |Ψ2 .ti ) ie ie (5.t) Ψ(x.16) Ψ −→ Ψ Indeed. ti )Ψ(xi . As the only thing that matters are matrix elements.ti iS[x] ℏ dα = e− ℏ α(xf .5. t) ie (5. classically. we can make an analogy with this representation: Aµ gauge transformation canonical transformation −→ Aµ + ∂µ α e − ie α ℏ (5. we can write very simply: K(xf . very much like we do when going from the Schroedinger to the Heisenberg picture for time evolution. −∂i α) 5. ti )e ℏ α(xi . tf .1.14) This is what we should have expected! Indeed. xi . what happens to gauge invariance upon quantization? We will study the path integral viewpoint ﬁrst. GAUGE FREEDOM 93 H and pi are not invariant under gauge transformations. tf . pi → H. and the Hamiltonian viewpoint later.15) canonical i transformation p −→ pi − e∂i α H + e∂t α H Quantum mechanically.2 Quantum physics Now. t) → e− ℏ α(x.13) The gauge covariance of this equation leads to: Ψ(x. tf . we could alternatively deﬁne the transformation as one where state vectors are unaﬀected while operators transform according ˆ ˆ to O → U † OU . pi + e (∂0 α. x] = −iℏ1 → p ˆ U † pU.11) H.tf ) K(xf . x] U = −iℏU † U = −iℏ1 . Canonical commutation relations are indeed preserved by the that transformation: [ˆ. Under the above rotation ˆ of the state vectors. U † xU = U † [ˆ. ti ) = → D[x]e iS[x] ie ℏ − ℏ D[x]e R xf . ˆ ˆ p ˆ (5. In the path integral formalism.tf xi . xi . ti )dxi (5. tf ) = K(xf .

18) ˙ In the non-relativistic limit where x ≪ 1. xj → pi .22) ˆ ˆ However. what about the gauge covariance of the Hamiltonian and the Schr¨dinger o equation? Under a gauge transformation the Hamiltonian transforms like: H −→ Hα = α 1 ˆ pi − eAi + e∂i α ˆ 2m 2 ˆ + e A0 + ∂t α (5. Now.20) Let us now work in the coordinate representation.94CHAPTER 5. xj = −iℏδij ˆ ˆ (5. when taking matrix elements.23) ˆ cancels out and we conclude that the matrix elements of Πj (A) are gauge invariant.23) ˆ ˆ and so the kinetic momentum Πj (A) is a covariant derivative (Πj (A)Ψ transforms like Ψ). the classical Hamiltonian takes c the form: 1 2 H = mc2 + pi − eAi + eA0 .24) Assume now that Ψ solves the Schr¨dinger equation with respect to H: o iℏ∂t Ψ = HΨ . Thus. Let us quantize pi . xj = 0 ˆ ˆ pi . where xi = xi ˆ pi = −iℏ ˆ ∂ ∂xi (5. Indeed we have ˆ ˆ Πj (A)Ψ(x) → Πj (A′ )Ψ′ (x) ie ˆ = e− ℏ α Πj (A)Ψ(x) = e− ℏ α −iℏ∂j − e∂j α − eAj + e∂j α Ψ(x) ie = −iℏ∂j − eAj + e∂j α e− ℏ α Ψ(x) ie (5. xj .21) The matrix elements of pi are not invariant under the gauge transformation ˆ ie Ψ → e− ℏ α Ψ: Ψ1 | pj |Ψ2 = ˆ → d3 xΨ∗ e ℏ α (−iℏ∂j )e− ℏ α Ψ2 = 1 ie ie d3 xΨ∗ (x)(−iℏ∂j )Ψ2 (x) 1 d3 xΨ∗ (−iℏ∂j − e∂j α)Ψ2 1 (5. pj = xi .19) 2m Quantizing and neglecting the (now irrelevant) constant mc2 we have the Hamiltonian operator H(ˆ. (5. the phase factor in eq. the matrix elements of the kinetic momentum Πj = pj − eAj are ˆ gauge invariant. with the canonical commutation relations: ˆ ˆ ˆ ˆ pi . INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD These results immediately tell us how to proceed in the Hamiltonian approach. x) = p ˆ 1 pi − eAi (ˆ) ˆ x 2m 2 + eA0 (ˆ) x (5. (5.

by the Schr¨dinger equation. we can write: Jj = ← − − → −i Ψ∗ D j Ψ − Ψ∗ D j Ψ 2m − → ← − → ← − 1 − ∂i+ ∂i= Di + Di ℏ −2 → D2 D = = eA0 − H 2m 2m − → ← − Di + Di Ψ (5. we can check that ∂t |Ψ| + ∂j J j = 0. In order to prove current conservation let us deﬁne: − → − → D i = ℏ ∂ i − ieAi ← − ← − D i = ℏ ∂ i + ieAi Using these deﬁnitions.26) J j (x) = 1 ∗ Ψ∗ (x)ˆj Ψ(x) − v j Ψ(x) Ψ(x) v ˆ 2 −i = Ψ∗ ℏ∂j − ieAj Ψ − ℏ∂j − ieAj Ψ 2m ∗ Ψ 2 (5.2 Particle in a constant magnetic ﬁeld Let us now study the system of a particle in a constant magnetic ﬁeld. Then what is the probability current? It is expectedly given through the velocity operator: vj = ˆ ˆ 1 Πj −iℏ∂j − eAj = m m (5.2. PARTICLE IN A CONSTANT MAGNETIC FIELD ie 95 o It is a good exercise to check that Ψα = e− ℏ α Ψ solves the Schr¨dinger equation with respect to Hα : iℏ∂t Ψα = Hα Ψα (5. the Schr¨dinger equation is gauge covariant. o J j is also manifestly gauge invariant.28) (5.29) We can then compute: ∂j J j = − → ← − −i − → ← − Ψ∗ D i + D i D i Ψ − Ψ∗ D i 2mℏ −i ←← − − − − →→ = Ψ∗ D i D i Ψ − Ψ∗ D i D i Ψ 2mℏ −i ∗ = Ψ∗ D 2 Ψ − D 2 Ψ Ψ 2mℏ i Ψ∗ (H − eA0 ) Ψ − ((H − eA0 ) Ψ)∗ Ψ = ℏ ∗ = i Ψ∗ i∂t Ψ − (i∂t Ψ) Ψ = −Ψ∗ ∂t Ψ − ∂t Ψ∗ Ψ = −∂t |Ψ| 2 (5.25) Thus.27) Indeed.5. o 2 In the same way.30) 5. we have that the probability density Ψ∗ Ψ = |Ψ| is gauge invariant. .

y + b) = En Ψn (x + a. We have: ˆ H(x.35) Thus.1 An exercise on translations Let us ﬁrst study what happens with spatial translations in this particular case. y) = En U (a. y)Ψn (x. A2 = Bx 2 2 (5. y) = Ψ(x + a. We take the magnetic ﬁeld along the z-axis: B = (0. b)H(x. the naive generators. B). y + b) = En Ψn (x + a. b)Ψn (x. y + b)Ψn (x + a.2. H(x + a. This operator can be p ˆ p ˆ ˆ written as U (a.31) We then recover the magnetic ﬁeld by: B3 = ∇ ∧ A 3 = ǫ321 ∂2 A1 + ǫ312 ∂1 A2 = −∂2 A1 + ∂1 A2 = B (5.32) This possible gauge choice leads to the following Hamiltonian. y + b). we can put these results together and write: ie ie ˆ e ℏ α H(x. y) ˆ U (a. b)Ψn (x. b)Ψn (x. y + b) = U (a. y)e− ℏ α Ψn (x + a. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD 5. y + b) (5. 0.36) . y + b) (5. We can choose the electromagnetic quadrivector as: A= 1 B ∧r 2 ⇒ 1 1 A1 = − By. b) = exp iaℏ 1 + ibℏ 2 . b)O(x. b)U (a. b)Ψ(x. b) such that U (a. Let us now look at the translational invariance of the spectrum of the Hamiltonian. in the coordinate representation: 1 ˆi ˆ H= Π (A) 2m 2 = 1 2m p1 + ˆ eB y 2 2 + p2 − ˆ eB x 2 2 + p32 ˆ (5. y) by a gauge transformation: A1 = − 1 B(y + b) = − 1 By + ∂x − Bbx ¯ 2 2 2 ¯ Bay 1 1 A2 = B(x + a) = Bx + ∂y 2 2 2 ¯ Aµ = Aµ + ∂µ α α = B (ay − bx) 2 ⇒ (5. • Translations along the x and y axis are described by the unitary operator U (a.34) ˆ ˆ However. y + b) is related to H(x.33) How are spatial translations realized? • −iℏ∂i . y)U † (a. do not commute with the Hamiltonian. Its action on operators is O(x + ˆ a. b). y) ˆ U (a.96CHAPTER 5. y) = En U (a. y)Ψn (x. b)H(x. y) ˆ H(x + a. y) = En Ψn (x. y)U † (a.

39) However.42) 1 y = y − x = y − 1 Πx c ˙ ωc ωc m . If eB > 0. Π2 eB y.37) We have thus obtained the inﬁnitesimal generators of translations: eB y 2 eB T 2 = p2 + ˆ x 2 T 1 = p1 − ˆ Notice that these generators commute with Π1 and Π2 : ˆ T1 . y + b). p2 ]) = 0 p ˆ 2 2 (5. • Previously. y) 97 (5. with γ ∼ ℏeB. we chose one particular way to translate from (x.2. • The commutator is related to the enclosed ﬂux. eB m. y) to (x + a. yc ) is the coordinate of the center of rotation. p2 − ˆ 2 eB y =0 2 eB eB x =− ([ˆ1 . Π1 = p1 − T1 .2.2 Motion We want to describe ﬁrst the classical motion of a particle in this constant magnetic ﬁeld. y + b) = ei 2ℏ (bx−ay) e ℏ (ap1 +bp2 ) Ψn (x. The classical equations of motion give the following result: • z = 0 ⇒ z = z0 + vz t.41) where (xc . y)Ψn = En Ψn eB i ie ˆ ˆ ˜ Ψn = e− ℏ α Ψn (x + a. Any other way would diﬀer by eiγ . p2 + ˆ x = (−iℏ − iℏ) = −iℏeB 2 2 2 (5. x] + [y. T2 ] = p1 − ˆ eB eB eB y. We get: x = xc + R cos ωc t y = yc − R sin ωc t (5. T1 and T2 do not commute among themselves: [T1 .40) • The group of translations is thus a projective realization of the classical one. oriented along the z-axis: B = (0. 5. the Notice: harmonicity: the frequency is independent of the radius of rotation. 0. ¨ • (x. B). PARTICLE IN A CONSTANT MAGNETIC FIELD We have then the following relation: ˜ ˆ ˜ H(x.38) (5. y) describe circular orbits with frequency ωc = rotation is clockwise.5. Its time evolution is described by: x = x + 1 y = x + 1 Πy c ˙ ωc ωc m (5. p1 + ˆ 2 eB = p1 − ˆ y.

42). y. Thus. What does this operation represent? It is a space translation.2. Let us pick Landau’s gauge which is also convenient to compute energy levels: Ax = −By The Hamiltonian is: 1 ˆ (ˆ1 + eB y)2 + p22 + p32 p ˆ ˆ ˆ H= 2m The kinetic momenta are given by: Π1 = p1 + eB y ˆ ˆ ˆ ˆ Π = p2 ˆ 2 ˆ Π3 = p3 ˆ • Notice that (5. Remember that we have seen that performing a gauge transformation ie on the vector potential amounts to a phase rotation: Ψ → e− ℏ α Ψ. then o Ψa. y. they cannot be measured simultaneously with arbitrary accuracy.b (x.50) .98CHAPTER 5. which implies that they are conserved.48) • Quantum mechanically.49) (5. • The combinations T 1 = p1 ˆ T2 = p2 + eBx ˆ (5. T2 ] = −iℏeB • Given a solution of the Schr¨dinger equation Ψ(x.45) ˆ ˆ Π1 . Let us see this directly by working with the quantum Hamiltonian in some gauge. Thus. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD xc and yc are two constants of motion.3 .43) (5. we get: T1 = Π1 − eBy = −eByc T2 = Π2 + eBx = eBxc (5. Π2 = −iℏeB = 0 (5. z) = e ℏ (aT1 +bT2 ) Ψ(x. Comparing them to (5. y). also T1 and T2 cannot be ﬁxed simultaneously as [T1 . i (5.44) Ay = 0 (5. They indeed correspond to the coordinates of the center of the classical motion. H] = 0. z) is also a solution. What is the associated symmetry? The answer is translations in (x.46) The two velocity components in the plane perpendicular to B do not commute. [Ti .47) ˆ commute with Π1. z). y.

2.56) . z) This is an ordinary translation composed with a gauge transformation.velocity) and (position. which p1 − mωc : i (5. we can choose (p3 . we have: bT2 ℏ ei Ψ(x. and (T1 . In a magnetic ﬁeld.2. whereas in ordinary motion in a potential those are (velocity. 5. T2 ) which are the coordinates of the center of the orbit. z) = ei Ψ(x. Π2 ). the canonically conjugated variables are (velocity. p1 . To be stressed and to conclude: ˆ ˆ The pairs of canonically conjugated variables are (Π1 . z) = e ℏ (p1 x+p3 z) F (y) For such a state. y) = 0 is not manifest.54) (5. z) = ei eBx ℏ eb∂y Ψ(x. y. whereas it is not manifest in the y-direction. eBx ℏ . which are physical velocities. • Notice that [T1 .3 Landau levels In order to study the Landau quantization.52) ˆ As p1 and p3 commute with H. T2 ] = 0 implies that the translation group is non-commutative in presence of a magnetic ﬁeld. PARTICLE IN A CONSTANT MAGNETIC FIELD This fact is obvious for ei For e bT i ℏ2 aT1 ℏ 99 = ei ap1 ˆ ℏ = ea∂x .55) It is the Hamiltonian of a harmonic oscillator in (ˆ2 . and centered at yc = − eB = 1 2 1 ˆ H= p + m ˆ 2m 2 2 eB m 2 y+ ˆ p1 eB 2 + p32 2m (5. the Hamiltonian is equivalent to: 1 2 ˆ H= (p1 + eB y ) + p22 + p32 ˆ ˆ 2m eB m .5.position). y + b. • The translational invariance in the x-direction is manifestly preserved. E) as quantum numˆ ˆ bers: p3 Ψ = p3 Ψ ˆ p1 Ψ = p1 Ψ ˆ ˆ HΨ = EΨ Such a state will satisfy: Ψ(x.position). y. y ) with frequency ωc = p ˆ p1 is the cyclotron frequency of the particle. We have the Hamiltonian: 1 ˆ (ˆ1 + eB y )2 + p22 + p32 p ˆ ˆ ˆ H= 2m (5.53) (5.51) • The rotation invariance around (x. y. let us continue working in the Landau gauge: Ax = −By Ay = 0 (5.

59) • The energy is degenerate in p1 ∼ yc .p1 .n2 .n (x.61) dp1 Ψp3 .n (x. y). n) = with the wave functions: Ψp3 . p1 . we get the energy levels: p ˆ E(p3 . • The result (5. p2 ]) = 0 (5. y ).p1 . We see that it is far from rotational invariant. T2 = eB ([ˆ1 . y. Where is cylindrical symmetry gone? To make it manifest.60) .58) where Ψn (y) is the ordinary wave function of a harmonic oscillator around y = 0 with frequency ωc .1.62) ˆ Π1 = p1 + eB y ˆ ˆ T2 = p2 + eB x ˆ ˆ (5. n2 ) (5. x] + [ˆ.1: The shape of the norm of the wavefunction Ψp3 . • The shape of |Ψp3 . y)| is depicted in ﬁgure 5. z)f (p1 .57) is in accord with the classical property of the system: the frequency of rotation does not depend on the radius: this is harmonicity.100CHAPTER 5.n (x. we have: ˆ Π2 = p2 ˆ T 1 = p1 ˆ We can check that: ˆ p ˆ y ˆ Π1 . z) = e− ℏ (p1 x+p3 z) Ψn y + Ψn (y) = mωc πℏ 1 4 i 1 p32 + ℏωc n + 2m 2 (5. x yc y Figure 5. y. y. z) = • In Landau gauge. we must superimpose diﬀerent wavefunctions: Ψp3 .p1 .n=0 (x. and Hn (x) are the Hermite polynomials deﬁned as: Hn (x) = (−1)n ex 2 dn −x2 e dxn (5.57) p1 mωc mωc ℏ e− mωc2 y2 2ℏ 1 √ Hn y 2n n! (5.n=0 (x. This is related to translational invariance. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD By quantizing this harmonic oscillator in (ˆ2 .p1 .

yc = T1 T2 − mωc is ﬁxed. this corresponds to summing only on paths that do not go through the interior of the cylinder. THE AHARONOV-BOHM EFFECT We get: mωc y = Π1 − T1 ˆ ˆ ˆ mωc x = T2 − Π2 ˆ 101 (5.3. if the motion is limited in the y direction as 0 y Ly . The charge quantization The Aharonov-Bohm eﬀect is one remarkable and yet simple example. we would think that physics (the propagation of charged particles) is unaﬀected outside the shell. Imagine now a modiﬁed arrangement where the cylindrical shell encloses a uniform magnetic ﬁeld (see ﬁgure 5. quantum mechanics tells that this conjecture is wrong! To show this. B = 0. If the motion is limited to 0 x Lx . Outside the shell.n . This could. Consider an ideal inﬁnitely long and empty cylindrical shell such that particles do not have access to its interior. some of which have a full description only in quantum ﬁeld theory: 1. The propagation of particles in this setup is characterized by a Schr¨dinger equation with boundary condition Ψ(x. Intuitively. then the number of states of quantum numbers (n. be realized by placing an “inﬁnitely” long solenoid inside the shell. whereas xc = mωc is not.5.3 The Aharonov-Bohm eﬀect The study of the vector potential in quantum mechanics opens a window on deeper aspects of the quantum world. It is intructive to count the levels for the case where the motion is limited to a region of sides Lx . as B = 0 there. but A = 0 as implied by Stokes’ theorem (see ﬁgure 5.2). then p1 is quantized: p1 = 2πℏ n2 Lx n2 ∈ Z (5. y.66) 5. for the states Ψp3 . the position yc is quantized: yc = 2πℏ p1 n2 = eB eBLx (5.3). . The relevance of topology 2. However. and is fully spread. p3 ) we can ﬁt in is: N (n. p3 ) = Lx Ly Ly = eB ∆yc 2πℏ (5. z) = 0 o on the surface of the cylinder. Ly in the plane. we must consider the vector potential.65) Note that it is the case whether or not the motion is limited in the y direction! Now.63) We can thus conclude that in Landau gauge. for instance. In the path integral approach.64) Thus.p1 .

whereas outside it there is no magnetic ﬁeld.102CHAPTER 5. Az = 0 in cylindrical coordinates. whereas outside it there is no magnetic ﬁeld. Ar . γ2 xi γ1 xf Figure 5. for r > r0 . One possible gauge choice (the only one that manifestly respects cylindrical BS symmetry) is Aφ = 2πr . .3: An inﬁnitely long cylinder has a constant magnetic ﬁeld inside it.4: Two paths γ1 and γ2 from xi to xf that cannot be continuously deformed into each other. due to the topological defect between them. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD B=0 B=0 Figure 5.2: An inﬁnitely long cylinder has a constant magnetic ﬁeld inside it. B γ S = πr0 2 γ = ∂σ γ A · dx = σ B · dσ = BS Figure 5.

5.3. THE AHARONOV-BOHM EFFECT

103

It is immediate to see that Aφ has physical consequences by considering quantum amplitudes in the path integral approach. Consider two paths γ1 and γ2 like in ﬁgure 5.4. • γ1 cannot be continuously deformed into γ2 . • They are topologically inequivalent (they belong to diﬀerent homotopy classes). Consider the contribution of γ1 and γ2 to K(xf , xi ; t): γ1 → e γ2 → e

(i)

iS1 ℏ

= exp = exp

iS0 ℏ iS0 ℏ

(1)

− −

ie ℏ ie ℏ

Ai dxi

γ1

iS2 ℏ

(2)

Ai dxi

γ2

(5.67)

where S0 is the free particle action for γi . The phase diﬀerence induced by the vector potential is: ∆φ = e ℏ Ai dxi − Ai dxi

γ1

=

γ2

e ℏ

γ2 −γ1

Ai dxi = −

eΦ eBS ≡− (5.68) ℏ ℏ

1. Apparently the presence of the solenoid is felt at the quantum level by trajectories that are arbitrarily far from it! 2. Notice that the phase shift vanishes between paths that are topologically equivalent, since they do not encircle any magnetic ﬂux. The suprising relevance of the vector potential at the quantum level was ﬁrst suggested by Aharonov and Bohm (1959) and later tested by Chambers (1960) and more recently by Tonomura (1982). The ideal “obvious” way to test this eﬀect is to perform the double slit experiment in the presence of our inﬁnite and thin solenoid (in reality the solenoid is replaced by a thin magnetized iron ﬁlament called a whisker), as in ﬁgure 5.5. Two disjoint classes of trajectories contribute to the propagation amplitude from the source S to the screen C: K(xf , xi ; t) = D[x]γ1 e

i ℏ

“ ” R (1) S0 −e Ai dxi

+

D[x]γ2 e

i ℏ

“ ” R (2) S0 −e Ai dxi

(5.69)

By Stokes’ theorem, γ1 Ai dxi is the same for all trajectories in the class 1 (i.e. passing on the top of the solenoid). The same thing is true for the class 2. Moreover, we have: Ai dxi − Ai dxi =

γ2 γ1 −γ2

Ai dxi = BS = Φ

(5.70)

γ1

**where Φ is the magnetic ﬂux through the solenoid. Thus, the total contribution will be: K(xf , xi ; t) = e
**

− ie ℏ R

γ1

Ai dxi

φ1 (xf ) + e ℏ Φ φ2 (xf )

ie

(0)

ie

(0)

= eiα φ1 (xf ) + e ℏ Φ φ2 (xf )

(0)

(0)

(5.71)

104CHAPTER 5. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD

xf γ1 xi S γ2 solenoid

A

B

C

Figure 5.5: The experimental setup to test the Aharonov-Bohm eﬀect. The two paths γ1 and γ2 contribute to the propagation amplitude from the source S to the screen C diﬀerently.

**where φi is the contribution of the class i in the absence of the solenoid. The ﬁnal amplitude will be: |K| = φ1
**

2 (0) 2

(0)

+ φ2

(0)

2

+ 2 Re φ1

(0)∗

e ℏ Φ φ2

ie

(0)

= F1 + F2 + Fint

(5.72)

This result tells us what the eﬀect of the solenoid is, no matter how compli(0) cated φi are. This is because its eﬀect is purely topological: the only property of the trajectory that enters into the phase shift is its topology (whether it belongs to the class 1 or 2). In the case of the free propagation, we have (see ﬁgure 5.6): φ1

(0)∗ (0) φ2

= e2 ℏ

i ka b z

≡ e2 ℏ k

i

z

⇒ Fint = 2 Re e

« „ 2k z eΦ i ℏ + ℏ

= 2 cos

2k z eΦ + ℏ ℏ

(5.73)

Thus, as Φ is varied, the peaks of the interference pattern are shifted along the z direction. This eﬀect is periodic in the ﬂux Φ with period: Φ0 ≡ ∆Φ = 2πℏ = 4.135 · 10−7 gauss × cm2 e (5.74)

which represents a fundamental unit of magnetic ﬂux. The eﬀect of the solenoid vanishes for Φ = nΦ0 . Moral: 1. The Aharonov-Bohm eﬀect shows that Fµν underdescribes electromagnetism; i.e. diﬀerent physical situations in a region may have the same Fµν in that region.

5.3. THE AHARONOV-BOHM EFFECT a z −a b

105

Figure 5.6: The experimental setup to test the Aharonov-Bohm eﬀect. The two slits are located at a and −a along the vertical direction, the two screens are separated by a distance b, and the position on the ﬁnal screen is described by z.

e 2. The phase ϕ = ℏ Ai dxi overdescribes electromagnetism; i.e. diﬀerent phases in a region may describe the same physics. For example Φ′ = Φ + nΦ0 .

3. What provides a complete description is the factor eiϕ = e ℏ γ Aµ dx over all possible closed paths in space-time, which after all is what enters in the path integral. Thus, electromagnetism is the gauge invariant manifestation of a non-integrable phase factor, as stated by Yang and Wu (1975).

ie

H

µ

5.3.1

Energy levels

Let us now do a simple example on the energy levels of a charged particle conﬁned on a ring of radius R around a solenoid (see ﬁgure 5.7).

solenoid

R

ring Figure 5.7: A particle is conﬁned on a ring of radius R around a solenoid of area σ went through by a magnetic ﬁeld of intensity B. The vector potential describing this situation is such that: Bdσ =

σ ∂σ

Adx

(5.75)

and thus the spectrum is the same. with m an integer.78) The eigenstates of this Hamiltonian are: Ψn (θ) = einθ with energy levels: En = ϕ 1 ℏ2 n− 2m R2 2π 2 n∈Z 1 ℏ2 2 (n − α) 2m R2 (5.106CHAPTER 5.8).81) gM ≡ B · dσ ⇒ B ≈ 4πr2 σ where σ is a sphere centered at the end of the solenoid. If for all charged particles the charge ej the ﬂux Φ satisﬁes: ej Φ = 2πnj ℏ (5. Of course ∇ · B = 0 is still satisﬁed: the magnetic ﬂux is conserved as there is the “incoming” ﬂux from the solenoid.82) there will be no way to tell the presence of the solenoid (also called Dirac’s string). Moreover.4 Dirac’s magnetic monopole Let us now very brieﬂy and qualitatively discuss Dirac’s magnetic monopole.80) We can notice that when ϕ → ϕ + 2πm. but the phase factor eiϕ .77) We then ﬁnd the Hamiltonian for the particle: H= 1 1 (pθ − eAθ )2 = 2m 2m = 1 ℏ2 2m R2 − iℏ∂θ eΦ − R 2πR eΦ 2πℏ 2 2 −i∂θ − = ϕ 1 ℏ2 −i∂θ − 2m R2 2π 2 (5. α → α + m. Notice that: gM (5. the physical angular momentum Rpθ = ℏ(n−α) is not quantized in integer units of ℏ. with the surface where the solenoid crosses it excluded. Let us take our thin solenoid and cut it in two: the ﬂux will “expand” out. Assume now the ideal limit in which the solenoid is inﬁnitely thin.79) = (5. Our object will behave by all means as a point-like magnetic monopole! Let us now turn the argument around: . The conclusion is that what is physical is not ϕ itself. The tip of the solenoid will act more or less as a point source of magnetic ﬁeld: a magnetic monopole (see ﬁgure 5. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD The magnetic ﬂux is found to be: Φ = Bσ = Aθ (R)2πR Which ﬁxes the phase factor: e− ℏ ie (5. 5.76) H Adx = e−i eΦ ℏ ≡ e−iϕ (5.

8: The solenoid cut in two will act as a magnetic monopole.4. Postulate: magnetic monopoles must exist in order to restore the symmetry between magnetic charge and electric charge which is apparently broken by Maxwell’s equations: ∇ · E = ρE ∇·B = 0 (5. a monopole state must involve an incoming ﬂux localized into some string.83) Since ∇· B = 0 follows not from dynamics but from kinematics: Bi = ǫijk ∂j Ak . the existence of a single monopole in the universe implies the quantization of the electric charge! 5. is: 2πℏ (5. and why monopoles are different than ordinary charged particles The strength of the electromagnetic interactions is set by the ﬁne structure constant: e2 1 α= ≈ ≪1 (5. Maxwell’s second law is still satisﬁed as the magnetic ﬂux is incoming from the solenoid. The quantization condition implies that the minimal magnetic charge of a monopole. αM .85) ℏc 137 The electromagnetic interactions are thus “weak”.84) Thus. if it existed.4. the Aharonov-Bohm phase should be a multiple of 2π for all existing electrically charged particles: ej gM = 2πnj ℏ (5. but in order to preserve spherical symmetry the string should be undetectable.1 On coupling constants. Thus. is deﬁned as: αM = (2π)2 (2π)2 ℏc gmin 2 c = = ≈ 137 · (2π)2 ≫ 1 2 ℏ e α (5. DIRAC’S MAGNETIC MONOPOLE B 107 B Figure 5. the magnetic coupling constant.5.86) gmin = e Thus.87) The magnetic interactions among monopoles are therefore very strong! .

108CHAPTER 5. INTERACTION WITH AN EXTERNAL ELECTROMAGNETIC FIELD .

2 and 3: Great reading!!!) [2] “Path Integral Methods and Applications”.5 and 2.10 and 3. 109 . A. MacKenzie. Hibbs. Chapters 1 and 5 (contains a lot of subjects. ( Chapters 1. J. Other references [8] H. AddisonWesley 1982. 1985. [3] “Modern Quantum Mechanics”. Sakurai. (Sections 1-2 and appendices 1-2 of Chapter 7: The uses of the instantons.9.P. “The Maslov correction in the semiclassical Feynman integral. Cambridge University Press. Feynman and A. I followed Coleman very closely in my lectures on instantons) [5] “Path Integrals in Quantum Mechanics. A 245 (1998) 345 [arXiv:quant-ph/9803016]. Coleman. which is nice. Chapters 12. Hagen Kleinert. [9] P. Kleinert and A.” Phys.” arXiv:quant-ph/0702236. 1981 Comprehensive but somewhat formal. Chervyakov. Aharonov-Bohm. McGraw-Hill. John Wiley & Sons Inc. Statistics and Polymer Physics”.S.J. Lett.6: quick introduction to path integral and nice extended discussion of electromagnetic interactions. arXiv:quantph/0004090. 1995. but sometimes he is too sloppy) [7] “Techniques and applications of Path Integration”.Bibliography [1] “Quantum Mechanics and Path Integrals” . R. L. 1965. Horvathy. 1985. Chapter 3. S. World Scientiﬁc. R. The Benjamin/Cummings Publishing Company. monopoles) [4] “Aspects of Symmetry”.11 give a good account of Feynman diagrams [6] “Topics in Advanced Quantum mechanics”. (chapters 2. in particular sections 3..R. Schulman. Barry R. 17 and 18 overlap with presentations in the course. “Simple Explicit Formulas for Gaussian Path Integrals with Time-Dependent Frequencies. (readable overview on the subject) The following books may be of help in complementing the lecture notes. 3. Holstein.

110 [10] “Quantum Mechanics”. BIBLIOGRAPHY . Landau-Lifshitz.

- An Introductory Lecture on Feynman Path Integrals
- Lectures Pi
- Path Integral
- Lectures Pi
- 311 DAS _Field
- art%3A10.1023%2FB%3AJOSS.0000019836.37663.d9.pdf
- ch1
- Ladder Operators and the Quantum Harmonic Oscillator
- 9781107009790_index
- 141 8.6 Lecture Notes
- Zeh-2002-Decoherence- Basic Concepts and Their Interpretation.9506020v3
- QMSE_01_Syllabus
- Wavefunction Collapse - Wikipedia, The Free Encyclopedia
- Why Probability in Quantum Mechanics is Given by the Wave Function Squared | Sean Carroll.pdf
- c7-6
- Integration
- Mecanica cuantica
- Supersymmetric Quantum Mechanics (wikipedia extract)
- Consistency Tests of Classical and Quantum Models for a Quantum Annealer
- Quantum Tunnelling
- Nelson Pinto-Neto and Diego Moraes Pantoja- Bohmian Quantization of the Big Rip
- Review Exam 1 2011
- Syll Sample Mstat PSB 2014
- UT Dallas Syllabus for math1326.502 05s taught by Kenneth Zamberlan (kenzam)
- M1.doc
- Handout Central
- Harald J. W. Muller-Kirsten Introduction to Quantum Mechanics Schrodinger Equation and Path Integral 2003.pdf
- Paradoxes
- Bode Phase Gain
- Fully Quantum Fluctuation Theorems
- Rattazzi-PathIntegral

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd