H.

Kleinert, PATH INTEGRALS

January 5, 2016 (/home/kleinert/kleinert/books/pathis/pthic2.tex)
A dancing shape, an image gay,
To haunt, to startle, and waylay
W. Wordsworth (1770–1840), Phantom of Delight

2
Path Integrals — Elementary Properties and
Simple Solutions

The operator formalism of quantum mechanics and quantum statistics may not
always lead to the most transparent understanding of quantum phenomena. There
exists another, equivalent formalism in which operators are avoided by the use of
infinite products of integrals, called path integrals. In contrast to the Schrödinger
equation, which is a differential equation determining the properties of a state at a
time from their knowledge at an infinitesimally earlier time, path integrals yield the
quantum-mechanical amplitudes in a global approach involving the properties of a
system at all times.

2.1 Path Integral Representation of Time Evolution

Amplitudes
The path integral approach to quantum mechanics was developed by Feynman1 in
1942. In its original form, it applies to a point particle moving in a Cartesian co-
ordinate system and yields the transition amplitudes of the time evolution operator
between the localized states of the particle (recall Section 1.7)

(xb tb |xa ta ) = hxb |Û(tb , ta )|xa i, tb > ta . (2.1)

For simplicity, we shall at first assume the space to be one-dimensional. The exten-
sion to D Cartesian dimensions will be given later. The introduction of curvilinear
coordinates will require a little more work. A further generalization to spaces with
a nontrivial geometry, in which curvature and torsion are present, will be described
in Chapters 10–11.

2.1.1 Sliced Time Evolution Amplitude

We shall be interested mainly in the causal or retarded time evolution amplitudes
[see Eq. (1.303)]. These contain all relevant quantum-mechanical information and
1
For the historical development, see Notes and References at the end of this chapter.

89
90 2 Path Integrals — Elementary Properties and Simple Solutions

possess, in addition, pleasant analytic properties in the complex energy plane [see
the remarks after Eq. (1.310)]. This is why we shall always assume, from now on,
the causal sequence of time arguments tb > ta .
Feynman realized that due to the fundamental composition law of the time evo-
lution operator (see Section 1.7), the amplitude (2.1) could be sliced into a large
number, say N + 1, of time evolution operators, each acting across an infinitesimal
time slice of thickness ǫ ≡ tn − tn−1 = (tb − ta )/(N + 1)> 0:

(xb tb |xa ta ) = hxb |Û (tb , tN )Û (tN , tN −1 ) · · · Û (tn , tn−1 ) · · · Û (t2 , t1 )Û (t1 , ta )|xa i. (2.2)

When inserting a complete set of states between each pair of Û’s,

Z ∞
dxn |xn ihxn | = 1, n = 1, . . . , N, (2.3)
−∞

the amplitude becomes a product of N-integrals

N Z
Y ∞  NY
+1
(xb tb |xa ta ) = dxn (xn tn |xn−1 tn−1 ), (2.4)
n=1 −∞ n=1

where we have set xb ≡ xN +1 , xa ≡ x0 , tb ≡ tN +1 , ta ≡ t0 . The symbol Π[· · ·] denotes

the product of the quantities within the brackets. The integrand is the product of
the amplitudes for the infinitesimal time intervals

(xn tn |xn−1 tn−1 ) = hxn |e−iǫĤ(tn )/h̄ |xn−1 i, (2.5)

with the Hamiltonian operator

Ĥ(t) ≡ H(p̂, x̂, t). (2.6)

The further development becomes simplest under the assumption that the Hamil-
tonian has the standard form, being the sum of a kinetic and a potential energy:

H(p, x, t) = T (p, t) + V (x, t). (2.7)

For a sufficiently small slice thickness, the time evolution operator

e−iǫĤ/h̄ = e−iǫ(T̂ +V̂ )/h̄ (2.8)

is factorizable as a consequence of the Baker-Campbell-Hausdorff formula (to be

proved in Appendix 2A)
2 X̂/h̄2
e−iǫ(T̂ +V̂ )/h̄ = e−iǫV̂ /h̄ e−iǫT̂ /h̄ e−iǫ , (2.9)

where the operator X̂ has the expansion

i ǫ 1 1
 
X̂ ≡ [V̂ , T̂ ] − [V̂ , [V̂ , T̂ ]] − [[V̂ , T̂ ], T̂ ] + O(ǫ2 ) . (2.10)
2 h̄ 6 3

H. Kleinert, PATH INTEGRALS

2.1 Path Integral Representation of Time Evolution Amplitudes 91

The omitted terms of order ǫ4 , ǫ5 , . . . contain higher commutators of V̂ and T̂ . If we

neglect, for the moment, the X̂-term which is suppressed by a factor ǫ2 , we calculate
for the local matrix elements of e−iǫĤ/h̄ the following simple expression:
Z ∞
hxn |e−iǫH(p̂,x̂,tn )/h̄ |xn−1 i ≈ dxhxn |e−iǫV (x̂,tn )/h̄ |xihx|e−iǫT (p̂,tn )/h̄ |xn−1 i
−∞
Z ∞
−iǫV (x̂,tn )/h̄
Z ∞ dpn ipn (x−xn−1 )/h̄ −iǫT (pn ,tn )/h̄
= dxhxn |e |xi e e . (2.11)
−∞ −∞ 2πh̄
Evaluating the local matrix elements,
hxn |e−iǫV (x̂,tn )/h̄ |xi = δ(xn − x)e−iǫV (xn ,tn )/h̄ , (2.12)
this becomes
hxn |e−iǫH(p̂,x̂,tn )/h̄ |xn−1 i ≈
∞ dpn
Z
exp {ipn (xn − xn−1 )/h̄ − iǫ[T (pn , tn ) + V (xn , tn )]/h̄} . (2.13)
−∞ 2πh̄
Inserting this back into (2.4), we obtain Feynman’s path integral formula, consisting
of the multiple integral
N Z  NY
+1 Z
" #
∞ ∞ dpn i N
Y  
(xb tb |xa ta ) ≈ dxn exp A , (2.14)
n=1 −∞ n=1 −∞ 2πh̄ h̄

where AN is the sum

N +1
AN =
X
[pn (xn − xn−1 ) − ǫH(pn , xn , tn )]. (2.15)
n=1

2.1.2 Zero-Hamiltonian Path Integral

Note that the path integral (2.14) with zero Hamiltonian produces the Hilbert space
structure of the theory via a chain of scalar products:
N Z  NY
+1 Z
" #
Y ∞ ∞ dpn i PN+1 pn (xn −xn−1 )/h̄
(xb tb |xa ta ) ≈ dxn e n=1 , (2.16)
n=1 −∞ n=1 −∞ 2πh̄
which is equal to
N Z
Y ∞  NY
+1 N Z
Y ∞  NY
+1
(xb tb |xa ta ) ≈ dxn hxn |xn−1 i = dxn δ(xn − xn−1 )
n=1 −∞ n=1 n=1 −∞ n=1
= δ(xb − xa ). (2.17)
whose continuum limit is
Dp i R dtp(t)ẋ(t)/h̄
Z Z
(xb tb |xa ta ) = Dx e = hxb |xa i = δ(xb − xa ). (2.18)
2πh̄
92 2 Path Integrals — Elementary Properties and Simple Solutions

In the operator expression (2.2), the right-hand side follows from the fact that
for zero Hamiltonian the time evolution operators Û(tn , tn−1 ) are all equal to unity.
At this point we make the important observation that a momentum variable
pn inside the product of momentuma integrations in the expression (2.16) can be
generated by a derivative p̂n ≡ −ih̄∂xn outside of it. In Subsection 2.1.4 we shall
go to the continuum limit of time slicing in which the slice thickness ǫ goes to zero.
In this limit, the discrete variables xn and pn become functions x(t) and p(t) of the
continuous time t, and the momenta pn become differential operators p(t) = −ih̄∂x(t) ,
satisfying the commutation relations with x(t):

[p̂(t), x(t)] = −ih̄. (2.19)

These are the canonical equal-time commutation relations of Heisenberg.

This observation forms the basis for deriving, from the path integral (2.14), the
Schrödinger equation for the time evolution amplitude.

2.1.3 Schrödinger Equation for Time Evolution Amplitude

Let us split from the product of integrals in (2.14) the final time slice as a factor, so
that we obtain the recursion relation
Z ∞
(xb tb |xa ta ) ≈ dxN (xb tb |xN tN ) (xN tN |xa ta ), (2.20)
−∞

where
dpb (i/h̄)[pb (xb −xN )−ǫH(pb ,xb ,tb )]
∞
Z
(xb tb |xN tN ) ≈e . (2.21)
−∞ 2πh̄

The momentum pb inside the integral can be generated by a differential operator

p̂b ≡ −ih̄∂xb outside of it. The same is true for any function of pb , so that the
Hamiltonian can be moved before the momentum integral yielding
dpb ipb (xb −xN )/h̄ −iǫH(−ih̄∂x ,xb ,tb )/h̄
∞
Z
−iǫH(−ih̄∂xb ,xb ,tb )/h̄
(xb tb |xN tN ) ≈ e e =e b δ(xb −xN ).
−∞ 2πh̄
(2.22)
Inserting this back into (2.20) we obtain

(xb tb |xa ta ) ≈ e−iǫH(−ih̄∂xb ,xb,tb )/h̄ (xb tb −ǫ|xa ta ), (2.23)

or
1 1 h −iǫH(−i∂x ,xb ,tb+ǫ)/h̄ i
[(xb tb +ǫ|xa ta ) − (xb tb |xa ta )] ≈ e b − 1 (xb tb |xa ta ). (2.24)
ǫ ǫ
In the limit ǫ → 0, this goes over into the differential equation for the time evolution
amplitude

ih̄∂tb (xb tb |xa ta ) = H(−ih̄∂xb , xb , tb )(xb tb |xa ta ), (2.25)

which is precisely the Schrödinger equation (1.301) of operator quantum mechanics.

H. Kleinert, PATH INTEGRALS

2.1 Path Integral Representation of Time Evolution Amplitudes 93

2.1.4 Convergence of of the Time-Sliced Evolution Amplitude

Some remarks are necessary concerning the convergence of the time-sliced expression
(2.14) to the quantum-mechanical amplitude in the continuum limit, where the
thickness of the time slices ǫ = (tb − ta )/(N + 1) → 0 goes to zero and the number
N of slices tends to ∞. This convergence can be proved for the standard kinetic
energy T = p2 /2M only if the potential V (x, t) is sufficiently smooth. For time-
independent potentials this is a consequence of the Trotter product formula which
reads
 N +1
e−i(tb −ta )Ĥ/h̄ = lim e−iǫV̂ /h̄ e−iǫT̂ /h̄ . (2.26)
N →∞

If T and V are c-numbers, this is trivially true. If they are operators, we use Eq. (2.9)
to rewrite the left-hand side of (2.26) as
 N +1  2 X̂/h̄2
N +1
e−i(tb −ta )Ĥ/h̄ ≡ e−iǫ(T̂ +V̂ )/h̄ ≡ e−iǫV̂ /h̄ e−iǫT̂ /h̄ e−iǫ .

The Trotter formula implies that the commutator term X̂ proportional to ǫ2 does
not contribute in the limit N → ∞. The mathematical conditions ensuring this
require functional analysis too technical to be presented here (for details, see the
literature quoted at the end of the chapter). For us it is sufficient to know that
the Trotter formula holds for operators which are bounded from below and that
for most physically interesting potentials, it cannot be used to derive Feynman’s
time-sliced path integral representation (2.14), even in systems where the formula
is known to be valid. In particular, the short-time amplitude may be different
from (2.13). Take, for example, an attractive Coulomb potential V (x) ∝ −1/|x|
for which the Trotter formula has been proved to be valid. Feynman’s time-sliced
formula, however, diverges even for two time slices. This will be discussed in detail in
Chapter 12. Similar problems will be found for other physically relevant potentials
such as V (x) ∝ l(l + D − 2)h̄2 /|x|2 (centrifugal barrier) and V (θ) ∝ m2 h̄2 /sin2 θ
(angular barrier near the poles of a sphere). In all these cases, the commutators
in the expansion (2.10) of X̂ become more and more singular. In fact, as we shall
see, the expansion does not even converge, even for an infinitesimally small ǫ. All
atomic systems contain such potentials and the Feynman formula (2.14) cannot be
used to calculate an approximation for the transition amplitude. A new path integral
formula has to be found. This will be done in Chapter 12. Fortunately, it is possible
to eventually reduce the more general formula via some transformations back to a
Feynman type formula with a bounded potential in an auxiliary space. Thus the
above derivation of Feynman’s formula for such potentials will be sufficient for the
further development in this book. After this it serves as an independent starting
point for all further quantum-mechanical calculations.
In the sequel, the symbol ≈ in all time-sliced formulas such as (2.14) will imply
that an equality emerges in the continuum limit N → ∞, ǫ → 0 unless the potential
94 2 Path Integrals — Elementary Properties and Simple Solutions

has singularities of the above type. In the action, the continuum limit is without
subtleties. The sum AN in (2.15) tends towards the integral
Z tb
A[p, x] = dt [p(t)ẋ(t) − H(p(t), x(t), t)] (2.27)
ta

under quite general circumstances. This expression is recognized as the classical

canonical action for the path x(t), p(t) in phase space. Since the position variables
xN +1 and x0 are fixed at their initial and final values xb and xa , the paths satisfy
the boundary condition x(tb ) = xb , x(ta ) = xa .
In the same limit, the product of infinitely many integrals in (2.14) will be called
a path integral . The limiting measure of integration is written as
N Z  NY
+1 Z
" #
∞ ∞ dpn x(tb )=xb Dp
Z Z
D′x
Y
lim dxn ≡ . (2.28)
N →∞
n=1 −∞ n=1 −∞ 2πh̄ x(ta )=xa 2πh̄
By definition, there is always one more pn -integral than xn -integrals in this product.
While x0 and xN +1 are held fixed and the xn -integrals are done for n = 1, . . . , N, each
pair (xn , xn−1 ) is accompanied by one pn -integral for n = 1, . . . , N +1. The situation
is recorded by the prime on the functional integral D ′x. With this definition, the
amplitude can be written in the short form
x(tb )=xb Dp iA[p,x]/h̄
Z Z
′
(xb tb |xa ta ) = Dx e . (2.29)
x(ta )=xa 2πh̄
The path integral has a simple intuitive interpretation: Integrating over all paths
corresponds to summing over all histories along which a physical system can possibly
evolve. The exponential eiA[p,x]/h̄ is the quantum analog of the Boltzmann factor
e−E/kB T in statistical mechanics. Instead of an exponential probability, a pure phase
factor is assigned to each possible history: The total amplitude for going from xa , ta
to xb , tb is obtained by adding up the phase factors for all these histories,
eiA[p,x]/h̄ ,
X
(xb tb |xa ta ) = (2.30)
all histories
(xa ,ta ) ❀ (xb ,tb )

where the sum comprises all paths in phase space with fixed endpoints xb , xa in
x-space.

2.1.5 Time Evolution Amplitude in Momentum Space

The above observed asymmetry in the functional integrals over x and p is a conse-
quence of keeping the endpoints fixed in position space. There exists the possibility
of proceeding in a conjugate way keeping the initial and final momenta pb and pa
fixed. The associated time evolution amplitude can be derived by going through
the same steps as before but working in the momentum space representation of the
Hilbert space, starting from the matrix elements of the time evolution operator
(pb tb |pa ta ) ≡ hpb |Û(tb , ta )|pa i. (2.31)

H. Kleinert, PATH INTEGRALS

2.1 Path Integral Representation of Time Evolution Amplitudes 95

The time slicing proceeds as in (2.2)–(2.4), with all x’s replaced by p’s, except in
the completeness relation (2.3) which we shall take as
Z ∞ dp
|pihp| = 1, (2.32)
−∞ 2πh̄
corresponding to the choice of the normalization of states [compare (1.186)]

hpb |pa i = 2πh̄δ(pb − pa ). (2.33)

In the resulting product of integrals, the integration measure has an opposite asym-
metry: there is now one more xn -integral than pn -integrals. The sliced path integral
reads
N N Z ∞
"Z #
∞ dpn Y
Y 
(pb tb |pa ta ) ≈ dxn
n=1 −∞ 2πh̄ n=0 −∞
N
( )
i X
× exp [−xn (pn+1 − pn ) − ǫH(pn , xn , tn )] . (2.34)
h̄ n=0

The relation between this and the x-space amplitude (2.14) is simple: By taking
in (2.14) the first and last integrals over p1 and pN +1 out of the product, renaming
them as pa and pb , and rearranging the sum N
P +1
n=1 pn (xn − xn−1 ) as follows

N
X +1
pn (xn − xn−1 ) = pN +1 (xN +1 − xN ) + pN (xN − xN −1 ) + . . .
n=1
. . . + p2 (x2 − x1 ) + p1 (x1 − x0 )
= pN +1 xN +1 − p1 x0
−(pN +1 − pN )xN − (pN − pN −1 )xN −1 − . . . − (p2 − p1 )x1
N
X
= pN +1 xN +1 − p1 x0 − (pn+1 − pn )xn , (2.35)
n=1

the remaining product of integrals looks as in Eq. (2.34), except that the lowest
index n is one unit larger than in the sum in Eq. (2.34). In the limit N → ∞ this
does not matter, and we obtain the Fourier transform
Z
dpb ipb xb /h̄ Z dpa −ipa xa /h̄
(xb tb |xa ta ) = e e (pb tb |pa ta ). (2.36)
2πh̄ 2πh̄
The inverse relation is
Z Z
(pb tb |pa ta ) = dxb e−ipb xb /h̄ dxa eipa xa /h̄ (xb tb |xa ta ). (2.37)

In the continuum limit, the amplitude (2.34) can be written as a path integral
p(tb )=pb D′p
Z Z
(pb tb |pa ta ) = DxeiĀ[p,x]/h̄, (2.38)
p(ta )=pa 2πh̄
96 2 Path Integrals — Elementary Properties and Simple Solutions

where
Z tb
Ā[p, x] = dt [−ṗ(t)x(t) − H(p(t), x(t), t)] = A[p, x] − pb xb + pa xa . (2.39)
ta

If the Hamiltonian is independent of x and t, the sliced path integral (2.34)

becomes trivial. Then the N + 1 integrals over xn (n = 0, . . . , N) can be done
yielding a product of δ-functions δ(pb − pN ) · · · δ(p1 − p0 ). As a conseqence, the
integrals over the N momenta pn (n = 1, . . . , N) are all squeezed to the initial
momentum pN = pN −1 = . . . = p1 = pa . A single a final δ-function 2πh̄δ(pb − pa )
remains, accompanied by the product of N + 1 factors N n=0 e
−iǫH(pa )/h̄
, which is
Q
−i(tb −ta )H(p)/h̄
equal to e . Hence we obtain:

(pb tb |pa ta ) = 2πh̄δ(pb − pa )e−i(tb −ta )H(p)/h̄ . (2.40)

Inserting this into Eq. (2.36), we find a simple Fourier integral for the time evolution
amlitude in x-space:

dp ip(xb −xa )/h̄−i(tb −ta )H(p)/h̄

Z
(xb tb |xa ta ) = e . (2.41)
2πh̄
Note that in (2.40) contains an equal sign rather than the ≈-sign since the right-hand
sign is the same for any number of time slices.

2.1.6 Quantum-Mechanical Partition Function

A path integral symmetric in p and x arises when considering the quantum-
mechanical partition function defined by the trace (recall Section 1.17)
 
ZQM (tb , ta ) = Tr e−i(tb −ta )Ĥ/h̄ . (2.42)

In the local basis, the trace becomes an integral over the amplitude
(xb tb |xa ta ) with xb = xa :
Z ∞
ZQM (tb , ta ) = dxa (xa tb |xa ta ). (2.43)
−∞

The additional trace integral over xN +1 ≡ x0 makes the path integral for ZQM
symmetric in pn and xn :
N Z  NY
+1 Z NY
+1
" # "ZZ #
∞ ∞ ∞ dpn ∞ dxn dpn
Z Y
dxN +1 dxn = . (2.44)
−∞ n=1 −∞ n=1 −∞ 2πh̄ n=1 −∞ 2πh̄

In the continuum limit, the right-hand side is written as

NY
+1
"ZZ #
∞ dxn dpn Dp
I Z
lim ≡ Dx , (2.45)
N →∞
n=1 −∞ 2πh̄ 2πh̄

H. Kleinert, PATH INTEGRALS

2.1 Path Integral Representation of Time Evolution Amplitudes 97

and the measures are related by

∞ x(tb )=xa Dp Dp
Z Z Z I Z
′
dxa Dx ≡ Dx . (2.46)
−∞ x(ta )=xa 2πh̄ 2πh̄
H
The symbol indicates the periodic boundary condition x(ta ) = x(tb ). In the
momentum representation we would have similarly
∞ dpa p(tb )=pa D′p Dp
Z Z Z I Z
Dx ≡ Dx , (2.47)
−∞ 2πh̄ p(ta )=pa 2πh̄ 2πh̄

with the periodic boundary condition p(ta ) = p(tb ), and the same right-hand side.
Hence, the quantum-mechanical partition function is given by the path integral

Dp iA[p,x]/h̄ Dp
I Z I Z
ZQM (tb , ta ) = Dx e = DxeiĀ[p,x]/h̄. (2.48)
2πh̄ 2πh̄

In the right-hand exponential, the action Ā[p, x] can be replaced by A[p, x], since
the extra terms in (2.39) are removed by the periodic boundary conditions. In the
time-sliced expression, the equality is easily derived from the rearrangement of the
sum (2.35), which shows that

N
X +1 N
X
pn (xn − xn−1 ) = − (pn+1 − pn )xn . (2.49)
n=1 xN+1 =x0 n=0 pN+1 =p0

In the path integral expression (2.48) for the partition function, the rules of quan-
tum mechanics appear as a natural generalization of the rules of classical statistical
mechanics, as formulated by Planck. According to these rules, each volume element
in phase space dxdp/h is occupied with the exponential probability e−E/kB T . In the
path integral formulation of quantum mechanics, each volume element in the path
phase space n dx(tn )dp(tn )/h is associated with a pure phase factor eiA[p,x]/h̄. We
Q

see here a manifestation of the correspondence principle which specifies the tran-
sition from classical to quantum mechanics. In path integrals, it looks somewhat
more natural than in the historic formulation, where it requires the replacement of
all classical phase space variables p, x by operators, a rule which was initially hard
to comprehend.

2.1.7 Feynman’s Configuration Space Path Integral

Actually, in his original paper, Feynman did not give the path integral formula in
the above phase space formulation. Since the kinetic energy in (2.7) has usually the
form T (p, t) = p2 /2M, he focused his attention upon the Hamiltonian

p2
H= + V (x, t), (2.50)
2M
98 2 Path Integrals — Elementary Properties and Simple Solutions

for which the time-sliced action (2.15) becomes

N +1
p2
" #
AN = pn (xn − xn−1 ) − ǫ n − ǫV (xn , tn ) .
X
(2.51)
n=1 2M

It can be quadratically completed to

N +1
" 2 2 #
ǫ xn − xn−1 M xn − xn−1
 
N
X
A = − pn − M + ǫ − ǫV (xn , tn ) . (2.52)
n=1 2M ǫ 2 ǫ

The momentum integrals in (2.14) may then be performed using the Fresnel integral
formula (1.337), yielding
" 2 #
∞ dpn i ǫ xn − xn−1 1
Z 
exp − pn − M =q , (2.53)
−∞ 2πh̄ h̄ 2M ǫ 2πh̄iǫ/M

and we arrive at the alternative representation

 
N
1 ∞ dxn i N
Y Z  
(xb tb |xa ta ) ≈ q  q  exp A , (2.54)
2πh̄iǫ/M n=1 −∞ 2πh̄iǫ/M h̄

where AN is now the sum

N +1
" 2 #
M xn − xn−1

N
X
A =ǫ − V (xn , tn ) , (2.55)
n=1 2 ǫ

with xN +1 = xb and x0 = xa . Here the integrals run over all paths in configuration
space rather than phase space. They account for the fact that a quantum-mechanical
particle starting from a given initial point xa will explore all possible ways of reaching
a given final point xb . The amplitude of each path is exp(iAN /h̄). See Fig. 2.1 for
a geometric illustration of the path integration. In the continuum limit, the sum
(2.55) converges towards the action in the Lagrangian form:
tb tb M 2
Z Z  
A[x] = dtL(x, ẋ) = dt ẋ − V (x, t) . (2.56)
ta ta 2
Note that this action is a local functional of x(t) in the temporal sense as defined in
Eq. (1.26).2
For the time-sliced Feynman path integral, one verifies the Schrödinger equation
as follows: As in (2.20), one splits off the last slice as follows:
Z ∞
(xb tb |xa ta ) ≈ dxN (xb tb |xN tN ) (xN tN |xa ta )
−∞
Z ∞
= d∆x (xb tb |xb −∆x tb −ǫ) (xb −∆x tb −ǫ|xa ta ), (2.57)
−∞

2
R
A functional F [x] is called
R local if it can be written as an integral dtf (x(t), ẋ(t)); it is called
ultra-local if it has the form dtf (x(t)).

H. Kleinert, PATH INTEGRALS

2.1 Path Integral Representation of Time Evolution Amplitudes 99

Figure 2.1 Zigzag paths, along which a point particle explores all possible ways of
reaching the point xb at a time tb , starting from xa at ta . The time axis is drawn from
right to left to have the same direction as the operator order in Eq. (2.2).

where
( " 2 #)
1 i M ∆x

(xb tb |xb −∆x tb −ǫ) ≈ q exp ǫ − V (xb , tb ) . (2.58)
2πh̄iǫ/M h̄ 2 ǫ

We now expand the amplitude in the integral of (2.57) in a Taylor series

1
 
(xb −∆x tb −ǫ|xa ta ) = 1 − ∆x ∂xb + (∆x)2 ∂x2b + . . . (xb , tb −ǫ|xa ta ). (2.59)
2
Inserting this into (2.57), the odd powers of ∆x do not contribute. For the even
powers, we perform the integrals using the Fresnel version of formula (1.339), and
obtain zero for odd powers of ∆x, and
( 2 ) !n
∞ d∆x iM ∆x h̄ǫ
Z 
2n
q (∆x) exp ǫ = i (2.60)
−∞ 2πh̄iǫ/M h̄ 2 ǫ M

for even powers, so that the integral in (2.57) becomes

" #
ih̄ 2 i

(xb tb |xa ta ) = 1 + ǫ ∂xb + O(ǫ2 ) 1 − ǫ V (xb , tb ) + O(ǫ2 ) (xb , tb −ǫ|xa ta ). (2.61)
2M h̄
In the limit ǫ → 0, this yields again the Schrödinger equation. (2.23).
In the continuum limit, we write the amplitude (2.54) as a path integral
Z x(tb )=xb
(xb tb |xa ta ) ≡ Dx eiA[x]/h̄ . (2.62)
x(ta )=xa
100 2 Path Integrals — Elementary Properties and Simple Solutions

This is Feynman’s original formula for the quantum-mechanical amplitude (2.1). It

consists of a sum over all paths in configuration space with a phase factor containing
the form of the action A[x].
We have used the same measure symbol Dx for the paths in configuration space
as for the completely different paths in phase space in the expressions (2.29), (2.38),
(2.46), (2.47). There should be no danger of confusion. Note that the q extra dpn -
integration in the phase space formula (2.14) results now in one extra 1/ 2πh̄iǫ/M
factor in (2.54) which is not accompanied by a dxn -integration.
The Feynman amplitude can be used to calculate the quantum-mechanical par-
tition function (2.43) as a configuration space path integral
I
ZQM = Dx eiA[x]/h̄ . (2.63)

As in (2.45), (2.46), the symbol Dx indicates that the paths have equal endpoints
H

x(ta ) = x(tb ), the path integral being the continuum limit of the product of integrals
NY
+1 Z ∞
dxn
I
Dx ≈ q . (2.64)
n=1 −∞ 2πih̄ǫ/M
q
There is no extra 1/ 2πih̄ǫ/M factor as in (2.54) and (2.62), due to the integration
over the initial (= final) position xb = xa representing the quantum-mechanical
trace. The use of the same symbol Dx as in (2.48)
H
should not cause any confusion
since (2.48) is always accompanied by an integral Dp.
R

For the sake of generality we might point out that it is not necessary to slice the
time axis in an equidistant way. In the continuum limit N → ∞, the canonical path
integral (2.14) is indifferent to the choice of the infinitesimal spacings

ǫn = tn − tn−1 . (2.65)

The configuration space formula contains the different spacings ǫn in the following
way: When performing the pn integrations, we obtain a formula of the type (2.54),
with each ǫ replaced by ǫn , i.e.,
 
N
1 ∞ dxn
Y Z
(xb tb |xa ta ) ≈ q  q 
2πh̄iǫb /M n=1 −∞ 2πih̄ǫn /M
N +1
(xn − xn−1 )2
( " #)
i X M
× exp − ǫn V (xn , tn ) . (2.66)
h̄ n=1 2 ǫn

To end this section, an important remark is necessary: It would certainly be

possible to define the path integral for the time evolution amplitude (2.29), without
going through Feynman’s time-slicing procedure, as the solution of the Schrödinger
differential equation [see Eq. (1.308))]:

[Ĥ(−ih̄∂x , x) − ih̄∂t ](x t|xa ta ) = −ih̄δ(t − ta )δ(x − xa ). (2.67)

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 101

If one possesses an orthonormal and complete set of wave functions ψn (x) solving
the time-independent Schrödinger equation Ĥψn (x)=En ψn (x), this solution is given
by the spectral representation (1.323)
ψn (xb )ψn∗ (xa )e−iEn (tb −ta )/h̄ ,
X
(xb tb |xa ta ) = Θ(tb − ta ) (2.68)
n

where Θ(t) is the Heaviside function (1.304). This definition would, however, run
contrary to the very purpose of Feynman’s path integral approach, which is to un-
derstand a quantum system from the global all-time fluctuation point of view. The
goal is to find all properties from the globally defined time evolution amplitude, in
particular the Schrödinger wave functions.3 The global approach is usually more
complicated than Schrödinger’s and, as we shall see in Chapters 8 and 12–14, con-
tains novel subtleties caused by the finite time slicing. Nevertheless, it has at least
four important advantages. First, it is conceptually very attractive by formulating
a quantum theory without operators which describe quantum fluctuations by close
analogy with thermal fluctuations (as will be seen later in this chapter). Second,
it links quantum mechanics smoothly with classical mechanics (as will be shown in
Chapter 4). Third, it offers new variational procedures for the approximate study of
complicated quantum-mechanical and -statistical systems (see Chapter 5). Fourth,
it gives a natural geometric access to the dynamics of particles in spaces with cur-
vature and torsion (see Chapters 10–11). This has recently led to results where
the operator approach has failed due to operator-ordering problems, giving rise to
a unique and correct description of the quantum dynamics of a particle in spaces
with curvature and torsion. From this it is possible to derive a unique extension
of Schrödinger’s theory to such general spaces whose predictions can be tested in
future experiments.4

2.2 Exact Solution for the Free Particle

In order to develop some experience with Feynman’s path integral formula we con-
sider in detail the simplest case of a free particle, which in the canonical form reads
p2
" !#
Z x(tb )=xb
′
Z
Dp i Z tb
(xb tb |xa ta ) = Dx exp dt pẋ − , (2.69)
x(ta )=xa 2πh̄ h̄ ta 2M
and in the pure configuration form:
x(tb )=xb i tb M
Z  Z 
(xb tb |xa ta ) = Dx exp dt ẋ2 . (2.70)
x(ta )=xa h̄ ta 2
Since the integration limits are obvious by looking at the left-hand sides of the equa-
tions, they will be omitted from now on, unless clarity requires their specification.
3
Many publications claiming to have solved the path integral of a system have violated this rule
by implicitly using the Schrödinger equation, although camouflaged by complicated-looking path
integral notation.
4
H. Kleinert, Mod. Phys. Lett. A 4 , 2329 (1989) (http://www.physik.fu-berlin.de/~klei-
nert/199); Phys. Lett. B 236 , 315 (1990) (ibid.http/202).
102 2 Path Integrals — Elementary Properties and Simple Solutions

2.2.1 Trivial Solution

Since the Hamiltonian of a free particle H = p2 /2M does not depend on x, the path
integral in momentum space yields the simple result (2.40). In configuration space
it reduces to simple Fourier integral Eq. (2.41). Inserting the above Hamiltonian,
the integral reads
dp ip(xb −xa )/h̄−i(tb −ta )p2 /2M h̄
Z
(xb tb |xa ta ) = e , (2.71)
2πh̄
and can be done with the help of the Fresnel formula (2.53). The result is
i M (xb − xa )2
" #
1
(xb tb |xa ta ) = q exp . (2.72)
2πih̄(tb − ta )/M h̄ 2 tb − ta

This result is easily generalized to D dimensions, where the free-particle ampli-

tude (2.40) reads
2 /2M h̄
(pb tb |pa ta ) = (2πh̄)D δ (D) (pb − pa )e−i(tb −ta )p , (2.73)

and the Fourier transform yields

i M (xb − xa )2
" #
1
(xb tb |xa ta ) = q D exp , (2.74)
2πih̄(tb − ta )/M h̄ 2 tb − ta

in agreement with the quantum-mechanical result in D dimensions (1.341).

2.2.2 Solution in Configuration Space

The problem can also be solved with little effort starting from Eq. (2.70). The time-
sliced expression to be integrated is given by Eqs. (2.54), (2.55) with V (x) set equal
to zero. The resulting product of Gaussian integrals can easily be done successively
using formula (1.337), from which we derive the simple rule
i M (x′′ − x′ )2 i M (x′ − x)2
" # " #
1 1
Z
′
dx q exp q exp
2πih̄Aǫ/M h̄ 2 Aǫ 2πih̄Bǫ/M h̄ 2 Bǫ
i M (x′′ − x)2
" #
1
=q exp , (2.75)
2πih̄(A + B)ǫ/M h̄ 2 (A + B)ǫ

which leads directly to the free-particle amplitude

i M (xb − xa )2
" #
1
(xb tb |xa ta ) = q exp . (2.76)
2πih̄(N + 1)ǫ/M h̄ 2 (N + 1)ǫ

After inserting (N + 1)ǫ = tb − ta , this agrees with the previous result (2.72). Note
that the free-particle amplitude happens to be independent of the number N + 1 of
time slices.

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 103

The calculation shows that the path integrals (2.69) and (2.70) possess a simple
solvable generalization to a time-dependent mass M(t) = Mg(t):

p2
" !#
x(tb )=xb Dp i tb
Z Z Z
′
(xb tb |xa ta ) = Dx exp dt pẋ − , (2.77)
x(ta )=xa 2πh̄ h̄ ta 2Mg(t)

and in the pure configuration form:

√ i tb M
x(tb )=xb
Z  Z 
(xb tb |xa ta ) = Dx g exp dt g(t)ẋ2 (t) . (2.78)
x(ta )=xa h̄ ta 2
√
Here the measure of integration Dx g symbolizes the continuum limit of the
R

product [compare (2.54)]:

 
N
x(tb )=xb √ 1 ∞ dxn
Z Y Z
Dx g ≡ q  q . (2.79)
x(ta )=xa 2πh̄iǫ/Mg(tb ) n=1 −∞ 2πh̄iǫ/Mg(tn )

The factor g(tn ) enters in each of the integrations (2.75), where the previous time
slicing parameters ǫ becomes now ǫn = ǫ/g(tn ), and we find instead of (2.72) the
amplitude

i M (xb − xa )2
" #
1
(xb tb |xa ta ) = q exp . (2.80)
h̄ 2 ttab g −1(t)
R
2πih̄M −1 ttab g −1 (t)
R

This has the Fourier representation

p2
( " #)
dp i tb
Z Z
(xb tb |xa ta ) = exp ip(xb − xa ) − g −1 (t) . (2.81)
2πh̄ h̄ 2M ta

2.2.3 Fluctuations around the Classical Path

There exists yet another method of calculating this amplitude which is somewhat
involved than the simple case at hand deserves, but which turns out to be useful for
the treatment of a certain class of more complicated path integrals, after a suitable
generalization. This method is based on summing aver all path deviations from the
classical path, i.e., all paths are split into the classical path
xb − xa
xcl (t) = xa + (t − ta ), (2.82)
tb − ta
along which the free particle would run following the equation of motion

ẍcl (t) = 0, (2.83)

plus deviations δx(t):

x(t) = xcl (t) + δx(t). (2.84)

104 2 Path Integrals — Elementary Properties and Simple Solutions

Since the initial and final points are fixed at xa , xb , respectively, the deviations vanish
at the endpoints:

δx(ta ) = δx(tb ) = 0. (2.85)

The deviations δx(t) are referred to as the quantum fluctuations of the particle
orbit. In mathematics, the boundary conditions (2.85) are referred to as Dirichlet
boundary conditions. When inserting the decomposition (2.84) into the action we
observe that due to the equation of motion (2.83) for the classical path, the action
separates into the sum of a classical and a purely quadratic fluctuation term
M tb
Z n o
dt ẋ2cl (t) + 2ẋcl (t)δ ẋ(t) + [δ ẋ(t)]2
2 ta
M Z tb tb
Z tb
M Z tb
= dtẋ2cl + M ẋδx − M dtẍcl δx + dt(δ ẋ)2
2 ta ta ta 2 ta
M
Z t Z tb 
b
= dtẋ2cl + dt(δ ẋ)2 .
2 ta ta

The absence of a mixed term is a general consequence of the extremality property

of the classical path,
δA = 0. (2.86)
x(t)=xcl (t)

It implies that a quadratic fluctuation expansion around the classical action

Acl ≡ A[xcl ] (2.87)

can have no linear term in δx(t), i.e., it must start as

1 Z tb Z tb ′ δ2A
A = Acl + dt dt ′
δx(t)δx(t′ ) + ... . (2.88)
2 ta ta δx(t)δx(t ) x(t)=xcl (t)

With the action being a sum of two terms, the amplitude factorizes into the product
of a classical amplitude eiAcl /h̄ and a fluctuation factor F0 (tb − ta ),
Z
(xb tb |xa ta ) = Dx eiA[x]/h̄ = eiAcl /h̄ F0 (tb , ta ). (2.89)

For the free particle with the classical action

tb M 2
Z
Acl = dt ẋ , (2.90)
ta 2 cl
the function factor F0 (tb − ta ) is given by the path integral
i tb M
Z  Z 
F0 (tb − ta ) = Dδx(t) exp dt (δ ẋ)2 . (2.91)
h̄ ta 2
Due to the vanishing of δx(t) at the endpoints, this does not depend on xb , xa but
only on the initial and final times tb , ta . The time translational invariance reduces

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 105

this dependence further to the time difference tb −ta . The subscript zero of F0 (tb −ta )
indicates the free-particle nature of the fluctuation factor. After inserting (2.82) into
(2.90), we find immediately
M (xb − xa )2
Acl = . (2.92)
2 tb − ta
The fluctuation factor, on the other hand, requires the evaluation of the multiple
integral
 
N
1 ∞ dδxn i N
Z  
F0N (tb − ta ) = q
Y
  exp A , (2.93)
h̄ fl
q
2πh̄iǫ/M n=1 −∞ 2πh̄iǫ/M

where AN
fl is the time-sliced fluctuation action
!2
M NX
+1
δxn − δxn−1
AN
fl = ǫ . (2.94)
2 n=1 ǫ
At the end, we have to take the continuum limit
N → ∞, ǫ = (tb − ta )/(N + 1) → 0.

2.2.4 Fluctuation Factor

The remainder of this section will be devoted to calculating the fluctuation factor
(2.93). Before doing this, we shall develop a general technique for dealing with
such time-sliced expressions. Due to the frequent appearance of the fluctuating
δx-variables, we shorten the notation by omitting all δ’s and working only with
x-variables.
A useful device for manipulating sums on a sliced time axis such as (2.94) is the
difference operator ∇ and its conjugate ∇, defined by
1 1
∇x(t) ≡ [x(t + ǫ) − x(t)], ∇x(t) ≡ [x(t) − x(t − ǫ)]. (2.95)
ǫ ǫ
They are two different discrete versions of the time derivative ∂t , to which both
reduce in the continuum limit ǫ → 0:
ǫ→0
∇, ∇ −
−−→ ∂t , (2.96)
if they act upon differentiable functions. Since the discretized time axis with N + 1
steps constitutes a one-dimensional lattice, the difference operators ∇, ∇ are also
called lattice derivatives.
For the coordinates xn = x(tn ) at the discrete times tn we write
1
∇xn = (xn+1 − xn ), N ≥ n ≥ 0,
ǫ
1
∇xn = (xn − xn−1 ), N + 1 ≥ n ≥ 1. (2.97)
ǫ
106 2 Path Integrals — Elementary Properties and Simple Solutions

The time-sliced action (2.94) can then be expressed in terms of ∇xn or ∇xn as
(writing xn instead of δxn )
N
M X M NX
+1
AN
fl = ǫ (∇xn )2 = ǫ (∇xn )2 . (2.98)
2 n=0 2 n=1

In this notation, the limit ǫ → 0 is most obvious: The sum ǫ n goes into the
P

integral ttab dt, whereas both (∇xn )2 and (∇xn )2 tend to ẋ2 , so that
R

Z tb M 2
AN
fl → dt ẋ . (2.99)
ta 2
Thus, the time-sliced action becomes the Lagrangian action.
Lattice derivatives have properties quite similar to ordinary derivatives. One
only has to be careful in distinguishing ∇ and ∇. For example, they allow for the
useful operation summation by parts which is analogous to integration by parts.
Recall the rule for the integration by parts
Z tb tb
Z tb
dtg(t)f˙(t) = g(t)f (t) − dtġ(t)f (t). (2.100)
ta ta ta

On the lattice, this relation yields for functions f (t) → xn and g(t) → pn :
N +1 N
pn ∇xn = pn xn |N +1
X X
ǫ 0 −ǫ (∇pn )xn . (2.101)
n=1 n=0

This follows directly by rewriting (2.35).

For functions vanishing at the endpoints, i.e., for xN +1 = x0 = 0, we can omit
the surface terms and shift the range of the sum on the right-hand side to obtain
the simple formula [see also Eq. (2.49)]
N
X +1 N
X N
X +1
pn ∇xn = − (∇pn )xn = − (∇pn )xn . (2.102)
n=1 n=0 n=1

The same thing holds if both p(t) and x(t) are periodic in the interval tb − ta , so that
p0 = pN +1 , x0 = xN +1 . In this case, it is possible to shift the sum on the right-hand
side by one unit arriving at the more symmetric-looking formula
N
X +1 N
X +1
pn ∇xn = − (∇pn )xn . (2.103)
n=1 n=1

In the time-sliced action (2.94) the quantum fluctuations xn (=δx

ˆ n ) vanish at the
ends, so that (2.102) can be used to rewrite
N +1 N
(∇xn )2 = −
X X
xn ∇∇xn . (2.104)
n=1 n=1

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 107

In the ∇xn -form of the action (2.98), the same expression is obtained by applying
formula (2.102) from the right- to the left-hand side and using the vanishing of x0
and xN +1 :
N N +1 N
(∇xn )2 = −
X X X
xn ∇∇xn = − xn ∇∇xn . (2.105)
n=0 n=1 n=1

The right-hand sides in (2.104) and (2.105) can be written in matrix form as
N
X N
X
− xn ∇∇xn ≡ − xn (∇∇)nn′ xn′ ,
n=1 n,n′ =1
N
X N
X
− xn ∇∇xn ≡ − xn (∇∇)nn′ xn′ , (2.106)
n=1 n,n′ =1

with the same N × N -matrix

 
−2 1 0 ... 0 0 0
1 −2 1 . . . 0 0 0
 
 
1 
.. ..

∇∇ ≡ ∇∇ ≡ 
. . .

(2.107)
ǫ2

 

 0 0 0 . . . 1 −2 1 

0 0 0 ... 0 1 −2

This is obviously the lattice version of the double time derivative ∂t2 , to which it
reduces in the continuum limit ǫ → 0. It will therefore be called the lattice Laplacian.
A further common property of lattice and ordinary derivatives is that they can
both be diagonalized by going to Fourier components. When decomposing
Z ∞
x(t) = dωe−iωt x(ω), (2.108)
−∞

and applying the lattice derivative ∇, we find

∞1  −iω(tn +ǫ)
Z 
∇x(tn ) = dω
e − e−iωtn x(ω) (2.109)
−∞ ǫ
1
Z ∞
= dωe−iωtn (e−iωǫ − 1) x(ω).
−∞ ǫ
Hence, on the Fourier components, ∇ has the eigenvalues
1 −iωǫ
(e − 1). (2.110)
ǫ
In the continuum limit ǫ → 0, this becomes the eigenvalue of the ordinary time
derivative ∂t , i.e., −i times the frequency of the Fourier component ω. As a reminder
of this we shall denote the eigenvalue of i∇ by Ω and have
i
(i∇x)(ω) = Ω x(ω) ≡ (e−iωǫ − 1) x(ω). (2.111)
ǫ
108 2 Path Integrals — Elementary Properties and Simple Solutions

For the conjugate lattice derivative we find similarly

i
(i∇x)(ω) = Ω x(ω) ≡ − (eiωǫ − 1) x(ω), (2.112)
ǫ
where Ω is the complex-conjugate number of Ω, i.e., Ω ≡ Ω∗ . As a consequence, the
eigenvalues of the negative lattice Laplacian −∇∇≡−∇∇ are real and nonnegative:
i −iωǫ i 1
(e − 1) (1 − eiωǫ ) = 2 [2 − 2 cos(ωǫ)] ≥ 0. (2.113)
ǫ ǫ ǫ
Of course, Ω and Ω̄ have the same continuum limit ω.
When decomposing the quantum fluctuations x(t) [=δx(t)] ˆ into their Fourier
components, not all eigenfunctions occur. Since x(t) vanishes at the initial time
t = ta , the decomposition can be restricted to the sine functions and we may expand
Z ∞
x(t) = dω sin ω(t − ta ) x(ω). (2.114)
0

The vanishing at the final time t = tb is enforced by a restriction of the frequencies

ω to the discrete values
πm πm
νm = = . (2.115)
tb − ta (N + 1)ǫ

Thus we are dealing with the Fourier series

∞
s
X 2
x(t) = sin νm (t − ta ) x(νm ) (2.116)
m=1 (tb − ta )

with real Fourier components x(νm ). A further restriction arises from the fact that
for finite ǫ, the series has to represent x(t) only at the discrete points x(tn ), n =
0, . . . , N + 1. It is therefore sufficient to carry the sum only up to m = N and to
expand x(tn ) as
s
N
X 2
x(tn ) = sin νm (tn − ta ) x(νm ), (2.117)
m=1 N +1
√
where a factor ǫ has been removed from the Fourier components, for convenience.
The expansion functions are orthogonal,
N
2 X
sin νm (tn − ta ) sin νm′ (tn − ta ) = δmm′ , (2.118)
N + 1 n=1

and complete:
N
2 X
sin νm (tn − ta ) sin νm (tn′ − ta ) = δnn′ (2.119)
N + 1 m=1

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 109

(where 0 < m, m′ < N + 1). The orthogonality relation follows by rewriting the
left-hand side of (2.118) in the form
N +1
iπ(m − m′ ) iπ(m + m′ )
( " # " #)
2 1 X
Re exp n − exp n , (2.120)
N +12 n=0 N +1 N +1

with the sum extended without harm by a trivial term at each end. Being of the
geometric type, this can be calculated right away. For m = m′ the sum adds up to
1, while for m 6= m′ it becomes
′ ′
1 − eiπ(m−m ) eiπ(m−m )/(N +1)
" #
2 1
Re − (m′ → −m′ ) . (2.121)
N +12 1 − eiπ(m−m′ )/(N +1)

The first expression in the curly brackets is equal to 1 for even m − m′ 6= 0; while
being imaginary for odd m − m′ [since (1 + eiα )/(1 − eiα ) is equal to (1 + eiα )(1 −
e−iα )/|1 − eiα |2 with the imaginary numerator eiα − e−iα ]. For the second term the
same thing holds true for even and odd m + m′ 6= 0, respectively. Since m − m′ and
m + m′ are either both even or both odd, the right-hand side of (2.118) vanishes for
m 6= m′ [remembering that m, m′ ∈ [0, N + 1] in the expansion (2.117), and thus in
(2.121)]. The proof of the completeness relation (2.119) can be carried out similarly.
Inserting now the expansion (2.117) into the time-sliced fluctuation action (2.94),
the orthogonality relation (2.118) yields
N
M X M NX
+1
AN
fl = ǫ (∇xn )2 = ǫ x(νm )Ωm Ωm x(νm ). (2.122)
2 n=0 2 m=1

Thus the action decomposes into a sum of independent quadratic terms involving
the discrete set of eigenvalues
1 1 πm
  
Ωm Ωm = 2 [2 − 2 cos(νm ǫ)] = 2 2 − 2 cos , (2.123)
ǫ ǫ N +1
and the fluctuation factor (2.93) becomes
 
N
1 ∞ dxn
Z
F0N (tb − ta ) = q
Y
 q 
2πh̄iǫ/M n=1 −∞ 2πh̄iǫ/M
N
iM
 
ǫΩm Ωm [x(νm )]2 .
Y
× exp (2.124)
m=1 h̄ 2

Before performing the integrals, we must transform the measure of integration from
the local variables xn to the Fourier components x(νm ). Due to the orthogonality
relation (2.118), the transformation has a unit determinant implying that
N
Y N
Y
dxn = dx(νm ). (2.125)
n=1 m=1
110 2 Path Integrals — Elementary Properties and Simple Solutions

With this, Eq. (2.124) can be integrated with the help of Fresnel’s formula (1.337).
The result is
N
1 1
F0N (tb − ta ) = q
Y
q . (2.126)
2πh̄iǫ/M m=1 ǫ2 Ωm Ωm
To calculate the product we use the formula5
N 
mπ x2(N +1) − 1

1 + x2 − 2x cos
Y
= . (2.127)
m=1 N +1 x2 − 1
Taking the limit x → 1 gives
N N
mπ
 
2
Y Y
ǫ Ωm Ωm = 2 1 − cos = N + 1. (2.128)
m=1 m=1 N +1
The time-sliced fluctuation factor of a free particle is therefore simply
1
F0N (tb − ta ) = q , (2.129)
2πih̄(N + 1)ǫ/M
or, expressed in terms of tb − ta ,
1
F0 (tb − ta ) = q . (2.130)
2πih̄(tb − ta )/M
As in the amplitude (2.72) we have dropped the superscript N since this final result
is independent of the number of time slices.
Note that the dimension of the fluctuation factor is 1/length. In fact, one may
introduce a length scale associated with the time interval tb − ta ,
q
l(tb − ta ) ≡ 2πh̄(tb − ta )/M, (2.131)
and write
1
F0 (tb − ta ) = √ . (2.132)
il(tb − ta )
With (2.130) and (2.92), the full time evolution amplitude of a free particle (2.89)
is again given by (2.72)
i M (xb − xa )2
" #
1
(xb tb |xa ta ) = q exp . (2.133)
2πih̄(tb − ta )/M h̄ 2 tb − ta

It is instructive to present an alternative calculation of the product of eigenvalues

in (2.126) which does not make use of the Fourier decomposition and works entirely
in configuration space. We observe that the product
N
ǫ2 Ωm Ωm
Y
(2.134)
m=1
5
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 1.396.2.

H. Kleinert, PATH INTEGRALS

2.2 Exact Solution for the Free Particle 111

is the determinant of the diagonalized N × N -matrix −ǫ2 ∇∇. This follows from
the fact that for any matrix, the determinant is the product of its eigenvalues.
The product (2.134) is therefore also called the fluctuation determinant of the free
particle and written
N
ǫ2 Ωm Ωm ≡ detN (−ǫ2 ∇∇).
Y
(2.135)
m=1

With this notation, the fluctuation factor (2.126) reads

1 h i−1/2
F0N (tb − tb ) = q detN (−ǫ2 ∇∇) . (2.136)
2πh̄iǫ/M

Now one realizes that the determinant of ǫ2 ∇∇ can be found very simply from the
explicit N × N matrix (2.107) by induction: For N = 1 we see directly that

detN =1 (−ǫ2 ∇∇) = |2| = 2. (2.137)

For N = 2, the determinant is

2 −1
detN =2 (−ǫ2 ∇∇) = = 3. (2.138)
−1 2

A recursion relation is obtained by developing the determinant twice with respect

to the first row:

detN (−ǫ2 ∇∇) = 2 detN −1 (−ǫ2 ∇∇) − detN −2 (−ǫ2 ∇∇). (2.139)

With the initial condition (2.137), the solution is

detN (−ǫ2 ∇∇) = N + 1, (2.140)

in agreement with the previous result (2.128).

2.2.5 Finite Slicing Properties of Free-Particle Amplitude

The time-sliced free-particle time evolution amplitude (2.76) happens to be inde-
pendent of the number N of time slices used for their calculation. We have pointed
this out earlier for the fluctuation factor (2.129). Let us study the origin of this
independence for the classical action in the exponent. The difference equation of
motion

−∇∇x(t) = 0 (2.141)

is solved by the same linear function

x(t) = At + B, (2.142)
112 2 Path Integrals — Elementary Properties and Simple Solutions

as in the continuum. Imposing the initial conditions gives

n
xcl (tn ) = xa + (xb − xa ) . (2.143)
N +1
The time-sliced action of the fluctuations is calculated, via a summation by parts
on the lattice [see (2.101)]. Using the difference equation ∇∇xcl = 0, we find
N +1
M
(∇xcl )2
X
Acl = ǫ (2.144)
n=1 2
N
!
M N +1 X
= xcl ∇xcl −ǫ xcl ∇∇xcl
2 n=0
n=0
M N +1 M (xb − xa )2
= xcl ∇xcl = .
2 n=0 2 tb − ta
This coincides with the continuum action for any number of time slices.
In the operator formulation of quantum mechanics, the ǫ-independence of the
amplitude of the free particle follows from the fact that in the absence of a potential
V (x), the two sides of the Trotter formula (2.26) coincide for any N.

2.3 Exact Solution for Harmonic Oscillator

A further problem to be solved along similar lines is the time evolution amplitude
of the linear oscillator
Dp i
Z Z  
(xb tb |xa ta ) = D′xexp A[p, x]
2πh̄ h̄
i
Z  
= Dx exp A[x] , (2.145)
h̄
with the canonical action
1 2 Mω 2 2
!
Z tb
A[p, x] = dt pẋ − p − x , (2.146)
ta 2M 2

and the Lagrangian action

tb M 2
Z
A[x] = dt (ẋ − ω 2 x2 ). (2.147)
ta 2

2.3.1 Fluctuations around the Classical Path

As before, we proceed with the Lagrangian path integral, starting from the time-
sliced form of the action

M NX
+1 h i
AN = ǫ (∇xn )2 − ω 2 x2n . (2.148)
2 n=1

H. Kleinert, PATH INTEGRALS

2.3 Exact Solution for Harmonic Oscillator 113

The path integral is again a product of Gaussian integrals which can be evaluated
successively. In contrast to the free-particle case, however, the direct evaluation
is now quite complicated; it will be presented in Appendix 2B. It is far easier to
employ the fluctuation expansion, splitting the paths into a classical path xcl (t) plus
fluctuations δx(t). The fluctuation expansion makes use of the fact that the action
is quadratic in x = xcl + δx and decomposes into the sum of a classical part
tb M 2
Z
Acl = dt (ẋ − ω 2x2cl ), (2.149)
ta 2 cl
and a fluctuation part
tb M
Z
Afl = dt [(δ ẋ)2 − ω 2(δx)2 ], (2.150)
ta 2
with the boundary condition

δx(ta ) = δx(tb ) = 0. (2.151)

There is no mixed term, due to the extremality of the classical action. The equation
of motion is

ẍcl = −ω 2 xcl . (2.152)

Thus, as for a free-particle, the total time evolution amplitude splits into a classical
and a fluctuation factor:
Z
(xb tb |xa ta ) = Dx eiA[x]/h̄ = eiAcl /h̄ Fω (tb − ta ). (2.153)

The subscript of Fω records the frequency of the oscillator.

The classical orbit connecting initial and final points is obviously
xb sin ω(t − ta ) + xa sin ω(tb − t)
xcl (t) = . (2.154)
sin ω(tb − ta )
Note that this equation only makes sense if tb − ta is not equal to an integer multiple
of π/ω which we shall always assume from now on.6
After an integration by parts we can rewrite the classical action Acl as
tb Mh M tb
Z i
Acl = dt xcl (−ẍcl − ω 2 xcl ) + xcl ẋcl . (2.155)
ta 2 2 ta

The first term vanishes due to the equation of motion (2.152), and we obtain the
simple expression
M
Acl = [xcl (tb )ẋcl (tb ) − xcl (ta )ẋcl (ta )]. (2.156)
2
6
For subtleties in the immediate neighborhood of the singularities which are known as caustic
phenomena, see Notes and References at the end of the chapter, as well as Section 4.8.
114 2 Path Integrals — Elementary Properties and Simple Solutions

Since
ω
ẋcl (ta ) = [xb − xa cos ω(tb − ta )], (2.157)
sin ω(tb − ta )
ω
ẋcl (tb ) = [xb cos ω(tb − ta ) − xa ], (2.158)
sin ω(tb − ta )
we can rewrite the classical action as
Mω h i
Acl = (x2b + x2a ) cos ω(tb − ta ) − 2xb xa . (2.159)
2 sin ω(tb − ta )

2.3.2 Fluctuation Factor

We now turn to the fluctuation factor. With the matrix notation for the lattice
operator −∇∇ − ω 2 , we have to solve the multiple integral
 
N
1 ∞ dδxn
Z
FωN (tb , ta ) = q
Y
 q 
2πh̄iǫ/M n=1 −∞ 2πh̄iǫ/M
 
i N
M X 
× exp  ǫ δxn [−∇∇ − ω 2]nn′ δxn′  . (2.160)
h̄ 2 n,n′ =1

When going to the Fourier components of the paths, the integral factorizes in the
same way as for the free-particle expression (2.124). The only difference lies in the
eigenvalues of the fluctuation operator which are now
1
Ωm Ωm − ω 2 = [2 − 2 cos(νm ǫ)] − ω 2 (2.161)
ǫ2
instead of Ωm Ωm . For times tb , ta where all eigenvalues are positive (which will
be specified below) we obtain from the upper part of the Fresnel formula (1.337)
directly
N
1 1
FωN (tb , ta ) = q
Y
q . (2.162)
2πh̄iǫ/M m=1 ǫ2 Ωm Ωm − ǫ2 ω 2

The product of these eigenvalues is found by introducing an auxiliary frequency ω̃

satisfying
ǫω̃ ǫω
sin ≡ . (2.163)
2 2
Then we decompose the product as
N N h N
ǫ2 Ωm Ωm − ǫ2 ω 2
" #
i Y
2 2 2 2
Y Y
[ǫ Ωm Ωm − ǫ ω ] = ǫ Ωm Ωm
m=1 m=1 m=1 ǫ2 Ωm Ωm
  
N h i N
sin2 ǫω̃
ǫ2 Ωm Ωm 
Y Y
1 − 2
= 2 mπ
 . (2.164)
m=1 m=1 sin 2(N +1)

H. Kleinert, PATH INTEGRALS

2.3 Exact Solution for Harmonic Oscillator 115

The first factor is equal to (N + 1) by (2.128). The second factor, the product of
the ratios of the eigenvalues, is found from the standard formula7
 
N 2
Y
1 −
sin x  1 sin[2(N + 1)x]
2 mπ = . (2.165)
m=1 sin 2(N +1) sin 2x (N + 1)

With x = ω̃ǫ/2, we arrive at the fluctuation determinant

N
2 2 2 sin ω̃(tb − ta )
[ǫ2 Ωm Ωm − ǫ2 ω 2 ] =
Y
detN (−ǫ ∇∇ − ǫ ω ) = , (2.166)
m=1 sin ǫω̃
and the fluctuation factor is given by
s
1 sin ω̃ǫ
FωN (tb , ta ) =q , tb − ta < π/ω̃, (2.167)
2πih̄/M ǫ sin ω̃(tb − ta )
√
where, as we have agreed earlier in Eq. (1.337), i means eiπ/4 , and tb − ta is always
larger than zero.

2.3.3 The iη -Prescription and Maslov-Morse Index

The result (2.167) is initially valid only for

tb − ta < π/ω̃. (2.168)

In this time interval, all eigenvalues in the fluctuation determinant (2.166) are pos-
itive, and the upper version of the Fresnel formula (1.337) applies to each of the
integrals in (2.160) [this was assumed in deriving (2.162)]. If tb − ta grows larger
than π/ω̃, the smallest eigenvalue Ω1 Ω1 − ω 2 becomes negative and the integration
over the associated Fourier component has to be done according to the lower case of
the Fresnel formula (1.337). The resulting amplitude carries an extra phase factor
e−iπ/2 and remains valid until tb − ta becomes larger than 2π/ω̃, where the second
eigenvalue becomes negative introducing a further phase factor e−iπ/2 .
All phase factors emerge naturally if we associate with the oscillator frequency ω
an infinitesimal negative imaginary part, replacing everywhere ω by ω − iη with an
infinitesimal η > 0. This is referred to as the iη-prescription. Physically, it amounts
to attaching an infinitesimal damping term to the oscillator, so that the amplitude
behaves like e−iωt−ηt and dies down to zero after a very long time (as opposed to
an unphysical antidamping term which would make it diverge after a long time).
Now, each time that tb − ta passes an integer multiple of π/ω̃, the square root of
sin ω̃(tb −ta ) in (2.167) passes a singularity in a specific way which ensures the proper
phase.8 With such an iη-prescription it will be superfluous to restrict tb − ta to the
7
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 1.391.1.
8
In the square root, we may equivalently assume tb − ta to carry a small negative imaginary
part. For a detailed discussion of the phases of the fluctuation factor in the literature, see Notes
and References at the end of the chapter.
116 2 Path Integrals — Elementary Properties and Simple Solutions

range (2.168). Nevertheless it will sometimes be useful to exhibit the phase factor
arising in this way in the fluctuation factor (2.167) for tb − ta > π/ω̃ by writing
s
1 sin ω̃ǫ
FωN (tb , ta ) =q e−iνπ/2 , (2.169)
2πih̄/M ǫ| sin ω̃(tb − ta )|

where ν is the number of zeros encountered in the denominator along the trajectory.
This number is called the Maslov-Morse index of the trajectory9 .

2.3.4 Continuum Limit

Let us now go to the continuum limit, ǫ → 0. Then the auxiliary frequency ω̃ tends
to ω and the fluctuation determinant becomes
ǫ→0 sin ω(tb − ta )
detN (−ǫ2 ∇∇ − ǫ2 ω 2) −
−−→ . (2.170)
ωǫ
The fluctuation factor FωN (tb − ta ) goes over into
s
1 ω
Fω (tb − ta ) = q , (2.171)
2πih̄/M sin ω(tb − ta )

with the phase for tb − ta > π/ω determined as above.

In the limit ω → 0, both fluctuation factors agree, of course, with the free-particle
result (2.130).
In the continuum limit, the ratios of eigenvalues in (2.164) can also be calculated
in the following simple way. We perform the limit ǫ → 0 directly in each factor.
This gives
ǫ2 Ωm Ωm − ǫ2 ω 2 ǫ2 ω 2
= 1−
ǫ2 Ωm Ωm 2 − 2 cos(νm ǫ)
ǫ→0 ω 2 (tb − ta )2
−
−−→ 1 − . (2.172)
π 2 m2
As the number N goes to infinity we wind up with an infinite product of these
factors. Using the well-known infinite-product formula for the sine function10
∞
x2
!
Y
sin x = x 1− 2 2 , (2.173)
m=1 mπ

we find, with x = ω(tb − ta ),

∞ 2
Y Ωm Ωm ǫ→0 Y νm ω(tb − ta )
−
− −→ 2 2
= , (2.174)
m Ωm Ωm − ω 2 m=1 νm − ω sin ω(tb − ta )
9
V.P. Maslov and M.V. Fedoriuk, Semi-Classical Approximations in Quantum Mechanics, Rei-
del, Boston, 1981.
10
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 1.431.1.

H. Kleinert, PATH INTEGRALS

2.3 Exact Solution for Harmonic Oscillator 117

and obtain once more the fluctuation factor in the continuum (2.171).
Multiplying the fluctuation factor with the classical amplitude, the time evolu-
tion amplitude of the linear oscillator in the continuum reads
i tb M 2
Z  Z 
(xb tb |xa ta ) = Dx(t) exp dt (ẋ − ω 2x2 )
h̄ ta 2
s
1 ω
= q (2.175)
2πih̄/M sin ω(tb − ta )
( )
i Mω
× exp [(x2 + x2a ) cos ω(tb − ta ) − 2xb xa ] .
2h̄ sin ω(tb − ta ) b
The result can easily be extended to any number D of dimensions, where the
action is
Z tb
M 2 
A= dt ẋ − ω 2 x2 . (2.176)
ta 2
Being quadratic in x, the action is the sum of the actions of each component leading
to the factorized amplitude:
D s D
1 ω
(xib tb |xia ta )
Y
(xb tb |xa ta ) = =q D
i=1 2πih̄/M sin ω(tb − ta )
( )
i Mω
× exp [(x2 + x2a ) cos ω(tb − ta ) − 2xb xa ] , (2.177)
2h̄ sin ω(tb − ta ) b
where the phase of the second square root for tb − ta > π/ω is determined as in the
one-dimensional case [see Eq. (1.546)].

2.3.5 Useful Fluctuation Formulas

It is worth realizing that when performing the continuum limit in the ratio of eigenvalues (2.174),
we have actually calculated the ratio of the functional determinants of the differential operators
det(−∂t2 − ω 2 )
. (2.178)
det(−∂t2 )
Indeed, the eigenvalues of −∂t2 in the space of real fluctuations vanishing at the endpoints are
simply
 2
2 πm
νm = , (2.179)
tb − ta
so that the ratio (2.178) is equal to the product
∞
det(−∂t2 − ω 2 ) Y 2
νm − ω2
= , (2.180)
det(−∂t2 ) m=1
2
νm

which is the same as (2.174). This observation should, however, not lead us to believe that the
entire fluctuation factor
 Z tb 
i M
Z
2 2 2
Fω (tb − ta ) = Dδx exp dt [(δ ẋ) − ω (δx) ] (2.181)
h̄ ta 2
118 2 Path Integrals — Elementary Properties and Simple Solutions

could be calculated via the continuum determinant

ǫ→0 1 1
Fω (tb , ta ) −
−−→ p p (false). (2.182)
2πh̄iǫ/M det(−∂t2 − ω 2 )

The product of eigenvalues in det(−∂t2 − ω 2 ) would be a strongly divergent expression

∞
Y
det(−∂t2 − ω 2 ) = 2
(νm − ω2) (2.183)
m=1
∞ ∞  2 ∞ 
− ω2 π 2 m2
 
Y
2

Y νm Y sin ω(tb − ta )
= νm 2
= 2
× .
m=1 m=1
νm m=1
(t b − t a ) ω(tb − ta )

Only ratios of determinants −∇∇ − ω 2 with different ω’s can be replaced by their differential
limits. Then the common divergent factor in (2.183) cancels.
Let us look at the origin of this strong divergence. The eigenvalues on the lattice and their
continuum approximation start both out for small m as

2 π 2 m2
Ωm Ωm ≈ νm ≈ . (2.184)
(tb − ta )2

For large m ≤ N , the eigenvalues on the lattice saturate at Ωm Ωm → 2/ǫ2 , while the νm 2
’s keep
growing quadratically in m. This causes the divergence.
The correct time-sliced formulas for the fluctuation factor of a harmonic oscillator is summa-
rized by the following sequence of equations:
N
"Z #  
N 1 Y dδxn iM T
Fω (tb − ta ) = p p exp δx (−ǫ2 ∇∇ − ǫ2 ω 2 )δx
2πh̄iǫ/M n=1 2πh̄iǫ/M h̄ 2ǫ
1 1
= p q , (2.185)
2πh̄iǫ/M det (−ǫ2 ∇∇ − ǫ2 ω 2 )
N

where in the first expression, the exponent is written in matrix notation with xT denoting the trans-
posed vector x whose components are xn . Taking out a free-particle determinant detN (−ǫ2 ∇∇),
formula (2.140), leads to the ratio formula
−1/2
detN (−ǫ2 ∇∇ − ǫ2 ω 2 )

1
FωN (tb − ta ) = p , (2.186)
2πh̄i(tb − ta )/M detN (−ǫ2 ∇∇)

which yields
s
1 sin ω̃ǫ
FωN (tb − ta ) = p , (2.187)
2πih̄/M ǫ sin ω̃(tb − ta )

If with ω̃ of Eq. (2.163). If we are only interested in the continuum limit, we may let ǫ go to zero
on the right-hand side of (2.186) and evaluate
−1/2
det(−∂t2 − ω 2 )

1
Fω (tb − ta ) =
det(−∂t2 )
p
2πh̄i(tb − ta )/M
∞  2 −1/2
1 Y νm − ω 2
= p 2
2πh̄i(tb − ta )/M m=1 νm
s
1 ω(tb − ta )
= . (2.188)
2πh̄i(tb − ta )/M sin ω(tb − ta )
p

H. Kleinert, PATH INTEGRALS

2.3 Exact Solution for Harmonic Oscillator 119

Let us calculate also here the time evolution amplitude in momentum space. The Fourier
transform of initial and final positions of (2.177) [as in (2.73)] yields
Z Z
(pb tb |pa ta ) = dD xb e−ipb xb /h̄ dD xa eipa xa /h̄ (xb tb |xa ta )

(2πh̄)D 1
= √ D p D
2πih̄ M ω sin ω(tb − ta )
 
i 1  2
(pb + p2a ) cos ω(tb − ta ) − 2pb pa .

× exp (2.189)
h̄ 2M ω sin ω(tb − ta )

The limit ω → 0 reduces to the free-particle expression (2.73), not quite as directly as in the
x-space amplitude (2.177). Expanding the exponent

1  2
(pb + pa ) cos ω(tb − ta ) − 2pb p2a

2M ω sin ω(tb − ta )
 
1 2 1 2 2 2
= (p b − p a ) − (p + p )[ω(t b − t a )] + . . . , (2.190)
2M ω 2 (tb − ta ) 2 b a

and going to the limit ω → 0, the leading term in (2.189)

(2π)D
 
i 1
D
exp 2
(pb − pa )2 (2.191)
p
2πiω 2 (tb − ta )h̄M h̄ 2M ω (tb − ta )

tends to (2πh̄)D δ (D) (pb − pa ) [recall (1.534)], while the second term in (2.190) yields a factor
2
e−ip (tb −ta )/2Mh̄ , so that we recover indeed (2.73).

2.3.6 Oscillator Amplitude on Finite Time Lattice

Let us calculate the exact time evolution amplitude for a finite number of time slices. In contrast to
the free-particle case in Section 2.2.5, the oscillator amplitude is no longer equal to its continuum
limit but ǫ-dependent. This will allow us to study some typical convergence properties of path
integrals in the continuum limit. Since the fluctuation factor was initially calculated at a finite ǫ in
(2.169), we only need to find the classical action for finite ǫ. To maintain time reversal invariance
at any finite ǫ, we work with a slightly different sliced potential term in the action than before in
(2.148), using

N +1
M X
AN = ǫ (∇xn )2 − ω 2 (x2n + x2n−1 )/2 ,

(2.192)
2 n=1

or, written in another way,

N
M X
AN = ǫ (∇xn )2 − ω 2 (x2n+1 + x2n )/2 .

(2.193)
2 n=0

This differs from the original time-sliced action (2.148) by having the potential ω 2 x2n replaced by
the more symmetric one ω 2 (x2n + x2n−1 )/2. The gradient term is the same in both cases and can
be rewritten, after a summation by parts, as
N
X +1 N
X N
X
(∇xn )2 = ǫ (∇xn )2 = xb ∇xb − xa ∇xa − ǫ
 
ǫ xn ∇∇xn . (2.194)
n=1 n=0 n=1
120 2 Path Integrals — Elementary Properties and Simple Solutions

This leads to a time-sliced action

N
N M M 2 2 2 M X
xn (∇∇ + ω 2 )xn .


A = xb ∇xb − xa ∇xa − ǫ ω (xb + xa ) − ǫ (2.195)
2 4 2 n=1

Since the variation of AN is performed at fixed endpoints xa and xb , the fluctuation factor is the
same as in (2.160). The equation of motion on the sliced time axis is
(∇∇ + ω 2 )xcl (t) = 0. (2.196)
Here it is understood that the time variable takes only the discrete lattice values tn . The solution
of this difference equation with the initial and final values xa and xb , respectively, is given by
1
xcl (t) = [xb sin ω̃(t − ta ) + xa sin ω̃(tb − t)] , (2.197)
sin ω̃(tb − ta )
where ω̃ is the auxiliary frequency introduced in (2.163). To calculate the classical action on the
lattice, we insert (2.197) into (2.195). After some trigonometry, and replacing ǫ2 ω 2 by 4 sin2 (ω̃ǫ/2),
the action resembles closely the continuum expression (2.159):
M sin ω̃ǫ
AN
 2
(xb + x2a ) cos ω̃(tb − ta ) − 2xb xa .

cl = (2.198)
2ǫ sin ω̃(tb − ta )
The total time evolution amplitude on the sliced time axis is
N
(xb tb |xa ta ) = eiAcl /h̄ FωN (tb − ta ), (2.199)
with sliced action (2.198) and the sliced fluctuation factor (2.169).

2.4 Gelfand-Yaglom Formula

In many applications one encounters a slight generalization of the oscillator fluc-
tuation problem: The action is harmonic but contains a time-dependent frequency
Ω2 (t) instead of the constant oscillator frequency ω 2. The associated fluctuation
factor is
i
Z  
F (tb , ta ) = Dδx(t) exp A , (2.200)
h̄
with the action
M
Z tb
A= dt [(δ ẋ)2 − Ω2 (t)(δx)2 ]. (2.201)
ta 2
Since Ω(t) may not be translationally invariant in time, the fluctuation factor de-
pends now in general on both the initial and final times. The ratio formula (2.186)
holds also in this more general case, i.e.,
#−1/2
detN (−ǫ2 ∇∇ − ǫ2 Ω2 )
"
N 1
F (tb , ta ) = q . (2.202)
2πh̄i(tb − ta )/M detN (−ǫ2 ∇∇)

Here Ω2 (t) denotes the diagonal matrix

Ω2N
 

Ω2 (t) = 
 .. 
, (2.203)
 . 
Ω21
with the matrix elements Ω2n = Ω2 (tn ).

H. Kleinert, PATH INTEGRALS

2.4 Gelfand-Yaglom Formula 121

2.4.1 Recursive Calculation of Fluctuation Determinant

In general, the full set of eigenvalues of the matrix −∇∇ − Ω2 (t) is quite difficult
to find, even in the continuum limit. It is, however, possible to derive a powerful
difference equation for the fluctuation determinant which can often be used to find its
value without knowing all eigenvalues. The method is due to Gelfand and Yaglom.11
Let us denote the determinant of the N × N fluctuation matrix by DN , i.e.,
 
DN ≡ detN −ǫ2 ∇∇ − ǫ2 Ω2 (2.204)
2 − ǫ2 Ω2N −1 0 ... 0 0 0
2 2
−1 2 − ǫ ΩN −1 −1 . . . 0 0 0
≡ .
.. .. .
.
2 2
0 0 0 . . . −1 2 − ǫ Ω2 −1
0 0 0 ... 0 −1 2 − ǫ2 Ω21
By expanding this along the first column, we obtain the recursion relation
DN = (2 − ǫ2 Ω2N )DN −1 − DN −2 , (2.205)
which may be rewritten as
1 DN − DN −1 DN −1 − DN −2
   
2
ǫ − + Ω2N DN −1 = 0. (2.206)
ǫ ǫ ǫ
Since the equation is valid for all N, it implies that the determinant DN satisfies
the difference equation
(∇∇ + Ω2N +1 )DN = 0. (2.207)
In this notation, the operator −∇∇ is understood to act on the dimensional label
N of the determinant. The determinant DN may be viewed as the discrete values of
a function of D(t) evaluated on the sliced time axis. Equation (2.207) is called the
Gelfand-Yaglom formula. Thus the determinant as a function of N is the solution
of the classical difference equation of motion and the desired result for a given N is
obtained from the final value DN = D(tN +1 ). The initial conditions are
D1 = (2 − ǫ2 Ω21 ),
(2.208)
D2 = (2 − ǫ2 Ω21 )(2 − ǫ2 Ω22 ) − 1.

2.4.2 Examples
As an illustration of the power of the Gelfand-Yaglom formula, consider the known case of a
constant Ω2 (t) ≡ ω 2 where the Gelfand-Yaglom formula reads
(∇∇ + ω 2 )DN = 0. (2.209)
This is solved by a linear combination of sin(N ω̃ǫ) and cos(N ω̃ǫ), where ω̃ is given by (2.163). The
solution satisfying the correct boundary condition is obviously
sin(N + 1)ǫω̃
DN = . (2.210)
sin ǫω̃
11
I.M. Gelfand and A.M. Yaglom, J. Math. Phys. 1 , 48 (1960).
122 2 Path Integrals — Elementary Properties and Simple Solutions

Figure 2.2 Solution of equation of motion with zero initial value and unit initial slope. Its
value at the final time is equal to ǫ times the discrete fluctuation determinant DN = D(tb ).

Indeed, the two lowest elements are

D1 = 2 cos ǫω̃,
D2 = 4 cos2 ǫω̃ − 1, (2.211)

which are the same as (2.208), since ǫ2 Ω2 ≡ ǫ2 ω 2 =2(1 − cos ω̃ǫ).

The Gelfand-Yaglom formula becomes especially easy to handle in the continuum limit ǫ → 0.
Then, by considering the renormalized function

Dren (tN ) = ǫDN , (2.212)

the initial conditions D1 = 2 and D2 = 3 can be re-expressed as

(ǫD)1 = Dren (ta ) = 0, (2.213)

ǫD2 − ǫD1 ǫ→0
= (∇ǫD)1 −
−−→ Ḋren (ta ) = 1. (2.214)
ǫ
The difference equation for DN turns into the differential equation for Dren (t):

[∂t2 + Ω2 (t)]Dren (t) = 0. (2.215)

The situation is pictured in Fig. 2.2. The determinant DN is 1/ǫ times the value of the function
Dren (t) at tb . This value is found by solving the differential equation starting from ta with zero
value and unit slope.
As an example, consider once more the harmonic oscillator with a fixed frequency ω. The
equation of motion in the continuum limit is solved by
1
Dren (t) = sin ω(t − ta ), (2.216)
ω
which satisfies the initial conditions (2.214). Thus we find the fluctuation determinant to become,
for small ǫ,
ǫ→0 1 sin ω(tb − ta )
det(−ǫ2 ∇∇ − ǫ2 ω 2 ) −
−−→ , (2.217)
ǫ ω
in agreement with the earlier result (2.210). For the free particle, the solution is Dren (t) = t − ta
and we obtain directly the determinant detN (−ǫ2 ∇∇) = (tb − ta )/ǫ.
For time-dependent frequencies Ω(t), an analytic solution of the Gelfand-Yaglom initial-value
problem (2.213), (2.214), and (2.215) can be found only for special classes of functions Ω(t). In
fact, (2.215) has the form of a Schrödinger equation of a point particle in a potential Ω2 (t), and
the classes of potentials for which the Schrödinger equation can be solved are well-known.

H. Kleinert, PATH INTEGRALS

2.4 Gelfand-Yaglom Formula 123

2.4.3 Calculation on Unsliced Time Axis

In general, the most explicit way of expressing the solution is by linearly combining
Dren =ǫDN from any two independent solutions ξ(t) and η(t) of the homogeneous
differential equation
[∂t2 + Ω2 (t)]x(t) = 0. (2.218)
The solution of (2.215) is found from a linear combination

Dren (t) = αξ(t) + βη(t). (2.219)

The coefficients are determined from the initial condition (2.214), which imply

αξ(ta ) + βη(ta ) = 0,
˙ a ) + β η̇(ta ) = 1,
αξ(t (2.220)

and thus
ξ(t)η(ta ) − ξ(ta )η(t)
Dren (t) = . (2.221)
˙ a )η(ta ) − ξ(ta )η̇(ta )
ξ(t
The denominator is recognized as the time-independent Wronski determinant of the
two solutions
↔
˙
W ≡ ξ(t) ∂ t η(t) ≡ ξ(t)η̇(t) − ξ(t)η(t) (2.222)

at the initial point ta . The right-hand side is independent of t.

The Wronskian is an important quantity in the theory of second-order differential
equations. It is defined for all equations of the Sturm-Liouville type
" #
d dy(t)
a(t) + b(t)y(t) = 0, (2.223)
dt dt
for which it is proportional to 1/a(t). The Wronskian is used to construct the Green
function for all such equations.12
In terms of the Wronskian, Eq. (2.221) has the general form
1
Dren (t) = − [ξ(t)η(ta ) − ξ(ta )η(t)] . (2.224)
W
Inserting t = tb gives the desired determinant
1
Dren = − [ξ(tb)η(ta ) − ξ(ta )η(tb )] . (2.225)
W
Note that the same functional determinant can be found from by evaluating the
function
1
D̃ren (t) = − [ξ(tb )η(t) − ξ(t)η(tb )] (2.226)
W
12
For its typical use in classical electrodynamics, see J.D. Jackson, Classical Electrodynamics,
John Wiley & Sons, New York, 1975, Section 3.11.
124 2 Path Integrals — Elementary Properties and Simple Solutions

at ta . This also satisfies the homogenous differential equation (2.215), but with the
initial conditions

D̃ren (tb ) = 0, ˙ (t ) = −1.

D̃ (2.227)
ren b

It will be useful to emphasize at which ends the Gelfand-Yaglom boundary condi-

tions are satisfied by denoting Dren (t) and D̃ren (t) by Da (t) and Db (t), respectively,
summarizing their symmetric properties as

[∂t2 + Ω2 (t)]Da (t) = 0 ; Da (ta ) = 0, Ḋa (ta ) = 1, (2.228)

[∂t2 + Ω2 (t)]Db (t) = 0 ; Db (tb ) = 0, Ḋb (tb ) = −1, (2.229)

with the determinant being obtained from either function as

Dren = Da (tb ) = Db (ta ). (2.230)

In contrast to this we see from the explicit equations (2.224) and (2.226) that the
time derivatives of two functions at opposite endpoints are in general not related.
Only for frequencies Ω(t) with time reversal invariance, one has

Ḋa (tb ) = −Ḋb (ta ), for Ω(t) = Ω(−t). (2.231)

For arbitrary Ω(t), one can derive a relation

Z tb
Ḋa (tb ) + Ḋb (ta ) = −2 dt Ω(t)Ω̇(t)Da (t)Db (t). (2.232)
ta

As an application of these formulas, consider once more the linear oscillator, for
which two independent solutions are

ξ(t) = cos ωt, η(t) = sin ωt. (2.233)

Hence

W = ω, (2.234)

and the fluctuation determinant becomes

1 1
Dren = − (cos ωtb sin ωta − cos ωta sin ωtb) = sin ω(tb − ta ). (2.235)
ω ω

2.4.4 D’Alembert’s Construction

It is important to realize that the construction of the solutions of Eqs. (2.228) and
(2.229) requires only the knowledge of one solution of the homogenous differential
equation (2.218), say ξ(t). A second linearly independent solution η(t) can always
be found with the help of a formula due to d’Alembert,
t dt′
Z
η(t) = w ξ(t) , (2.236)
ξ 2 (t′ )

H. Kleinert, PATH INTEGRALS

2.4 Gelfand-Yaglom Formula 125

where w is some constant. Differentiation yields

˙
ξη w ¨
ξη
η̇ = + , η̈ = . (2.237)
ξ ξ ξ

The second equation shows that with ξ(t), also η(t) is a solution of the homogenous
differential equation (2.218). From the first equation we find that the Wronski
determinant of the two functions is equal to w:
˙
W = ξ(t)η̇(t) − ξ(t)η(t) = w. (2.238)

Inserting the solution (2.236) into the formulas (2.224) and (2.226), we obtain
explicit expressions for the Gelfand-Yaglom functions in terms of one arbitrary so-
lution of the homogenous differential equation (2.218):
Z t dt′ Z tb dt′
Dren (t) = Da (t) = ξ(t)ξ(ta ) , D̃ren (t) = Db (t) = ξ(tb )ξ(t) . (2.239)
ta ξ 2 (t′ ) t ξ 2(t′ )

The desired functional determinant is

Z tb dt′
Dren = ξ(tb )ξ(ta ) . (2.240)
ta ξ 2 (t′ )

2.4.5 Another Simple Formula

There exists yet another useful formula for the functional determinant. For this we solve the
homogenous differential equation (2.218) for an arbitrary initial position xa and initial velocity
ẋa at the time ta . The result may be expressed as the following linear combination of Da (t) and
Db (t):

1 h i
x(xa , ẋa ; t) = Db (t) − Da (t)Ḋb (ta ) xa + Da (t)ẋa . (2.241)
Db (ta )

We then see that the Gelfand-Yaglom function Dren (t) = Da (t) can be obtained from the partial
derivative
∂x(xa , ẋa ; t)
Dren (t) = . (2.242)
∂ ẋa
This function obviously satisfies the Gelfand-Yaglom initial conditions Dren (ta ) = 0 and Ḋren (ta ) =
1 of (2.213) and (2.214), which are a direct consequence of the fact that xa and ẋa are independent
variables in the function x(xa , ẋa ; t), for which ∂xa /∂ ẋa = 0 and ∂ ẋa /∂ ẋa = 1.
The fluctuation determinant Dren = Da (tb ) is then given by

∂xb
Dren = , (2.243)
∂ ẋa

where xb abbreviates the function x(xa , ẋa ; tb ). It is now obvious that the analogous equations
(2.229) are satisfied by the partial derivative Db (t) = −∂x(t)/∂ ẋb , where x(t) is expressed in terms
of the final position xb and velocity ẋb as x(t) = x(xb , ẋb ; t)

1 h i
x(xb , ẋb ; t) = Da (t) + Db (t)Ḋa (tb ) xb − Db (t)ẋb , (2.244)
Da (tb )
126 2 Path Integrals — Elementary Properties and Simple Solutions

so that we obtain the alternative formula

∂xa
Dren = − . (2.245)
∂ ẋb
These results can immediately be generalized to functional determinants of differential opera-
tors of the form −∂t2 δij −Ω2ij (t) where the time-dependent frequency is a D ×D-dimensional matrix
Ω2ij (t), (i, j = 1, . . . , D). Then the associated Gelfand-Yaglom function Da (t) becomes a matrix
Dij (t) satisfying the initial conditions Dij (ta ) = 0, Ḋij (tb ) = δij , and the desired functional
determinant Dren is equal to the ordinary determinant of Dij (tb ):

Dren = Det −∂t2 δij − Ω2ij (t) = det Dij (tb ).

 
(2.246)

The homogeneous differential equation and the initial conditions are obviously satisfied by the
partial derivative matrix Dij (t) = ∂xi (t)/∂ ẋja , so that the explicit representations of
 Dij (t) in
terms of the general solution of the classical equations of motion −∂t2 δij − Ω2ij (t) xj (t) = 0
become !
∂xib ∂xia
Dren = det j = det − j . (2.247)
∂ ẋa ∂ ẋb
A further couple of formulas for functional determinants can be found by constructing a solution
of the homogeneous differential equation (2.218) which passes through specific initial and final
points xa and xb at ta and tb , respectively:

Db (t) Da (t)
x(xb , xa ; t) = xa + xb . (2.248)
Db (ta ) Da (tb )

The Gelfand-Yaglom functions Da (t) and Db (t) can therefore be obtained from the partial deriva-
tives
Da (t) ∂x(xb , xa ; t) Db (t) ∂x(xb , xa ; t)
= , = . (2.249)
Da (tb ) ∂xb Db (ta ) ∂xa
At the endpoints, Eqs. (2.248) yield

Ḋb (ta ) 1
ẋa = xa + xb , (2.250)
Db (ta ) Da (tb )
1 Ḋa (tb )
ẋb = − xa + xb , (2.251)
Db (ta ) Da (tb )

so that the fluctuation determinant Dren = Da (tb ) = Db (ta ) is given by the formulas
 −1  −1
∂ ẋa ∂ ẋb
Dren = =− , (2.252)
∂xb ∂xa

where ẋa and ẋb are functions of the independent variables xa and xb . The equality of these ex-
pressions with the previous ones in (2.243) and (2.245) is a direct consequence of the mathematical
identity for partial derivatives
!−1
∂xb ∂ ẋa
= . (2.253)
∂ ẋa xa ∂xb xa
Let us emphasize that all functional determinants calculated in this Chapter apply to the
fluctuation factor of paths with fixed endpoints. In mathematics, this property is referred to
as Dirichlet boundary conditions. In the context of quantum statistics, we shall also need such
determinants for fluctuations with periodic boundary conditions, for which the Gelfand-Yaglom
method must be modified. We shall see in Section 2.11 that this causes considerable complications

H. Kleinert, PATH INTEGRALS

2.5 Harmonic Oscillator with Time-Dependent Frequency 127

in the lattice derivation, which will make it desirable to find a simpler derivation of both functional
determinants. This will be found in Section 3.27 in a continuum formulation.
In general, the homogenous differential equation (2.218) with time-dependent frequency Ω(t)
cannot be solved analytically. The equation has the same form as a Schrödinger equation for
a point particle in one dimension moving in a one dimensional potential Ω2 (t), and there are
only a few classes of potentials for which the solutions are known in closed form. Fortunately,
however, the functional determinant will usually arise in the context of quadratic fluctuations
around classical solutions in time-independent potentials (see in Section 4.3). If such a classical
solution is known analytically, it will provide us automatically with a solution of the homogeneous
differential equation (2.218). Some important examples will be discussed in Sections 17.4 and
17.11.

2.4.6 Generalization to D Dimensions

The above formulas have an obvious generalization to a D-dimensional version of
the fluctuation action (2.201)
tb M
Z
2
A= dt [(δ ẋ)2 − δxT (t)δx], (2.254)
ta 2

where 2 (t) is a D × D matrix with elements Ω2ij (t). The fluctuation factor (2.202)
generalizes to
2
#−1/2
detN (−ǫ2 ∇∇ − ǫ2
"
N 1 )
F (tb , ta ) = q D . (2.255)
2πh̄i(tb − ta )/M detN (−ǫ2 ∇∇)

The fluctuation determinant is found by Gelfand-Yaglom’s construction from a for-

mula
Dren = det Da (tb ) = det Db (ta ), (2.256)
with the matrices Da (t) and Db (t) satisfying the classical equations of motion and
initial conditions corresponding to (2.228) and (2.229):
2
[∂t2 + (t)]Da (t) = 0 ; Da (ta ) = 0, Ḋa (ta ) = 1, (2.257)
2
[∂t2 + (t)]Db (t) = 0 ; Db (tb ) = 0, Ḋb (tb ) = −1, (2.258)

where 1 is the unit matrix in D dimensions. We can then repeat all steps in the last
section and find the D-dimensional generalization of formulas (2.252):
!−1 !#−1
∂ ẋi ∂ ẋi
"
Dren = det aj = det − jb . (2.259)
∂xb ∂xa

2.5 Harmonic Oscillator with Time-Dependent Frequency

The results of the last section put us in a position to solve exactly the path integral
of a harmonic oscillator with arbitrary time-dependent frequency Ω(t). We shall
first do this in coordinate space, later in momentum space.
128 2 Path Integrals — Elementary Properties and Simple Solutions

2.5.1 Coordinate Space

Consider the path integral
i
Z  
(xb tb |xa ta ) = Dx exp A[x] , (2.260)
h̄
with the Lagrangian action
M Z tb h 2 i
A[x] = dt ẋ (t) − Ω2 (t)x2 (t) , (2.261)
2 ta
which is harmonic with a time-dependent frequency. As in Eq. (2.14), the result can
be written as a product of a fluctuation factor and an exponential containing the
classical action:
Z
(xb tb |xa ta ) = Dx eiA[x]/h̄ = FΩ (tb , ta )eiAcl /h̄ . (2.262)

From the discussion in the last section we know that the fluctuation factor is, by
analogy with (2.171), and recalling (2.243),
1 1
FΩ (tb , ta ) = q q . (2.263)
2πih̄/M Da (tb )

The determinant Da (tb ) = Dren may be expressed in terms of partial derivatives

according to formulas (2.243) and (2.252):
!−1/2 !1/2
1 ∂xb 1 ∂ ẋa
FΩ (tb , ta ) = q =q , (2.264)
2πih̄/M ∂ ẋa 2πih̄/M ∂xb

where the first partial derivative is calculated from the function x(xa , ẋa ; t), the
second from ẋ(xb , xa ; t). Equivalently we may use (2.245) and the right-hand part
of Eq. (2.252) to write
!−1/2 !1/2
1 ∂xa 1 ∂ ẋb
FΩ (tb , ta ) = q − =q − . (2.265)
2πih̄/M ∂ ẋb 2πih̄/M ∂xa

It remains to calculate the classical action Acl . This can be done in the same
way as in Eqs. (2.155) to (2.159). After a partial integration, we have as before
M
Acl = (xb ẋb − xa ẋa ). (2.266)
2
Exploiting the linear dependence of ẋb and ẋa on the endpoints xb and xa , we may
rewrite this as
!
M ∂ ẋb ∂ ẋa ∂ ẋb ∂ ẋa
Acl = xb xb − xa xa + xb xa − xa xb . (2.267)
2 ∂xb ∂xa ∂xa ∂xb

H. Kleinert, PATH INTEGRALS

2.5 Harmonic Oscillator with Time-Dependent Frequency 129

Inserting the partial derivatives from (2.250) and (2.251) and using the equality of
Da (tb ) and Db (ta ), we obtain the classical action
M h 2 i
Acl = xb Ḋa (tb ) − x2a Ḋb (ta ) − 2xb xa . (2.268)
2Da (tb )
Note that there exists another simple formula for the fluctuation determinant Dren :
!−1
∂2
Dren = Da (tb ) = Db (ta ) = −M Acl . (2.269)
∂xb ∂xa

For the harmonic oscillator with time-independent frequency ω, the Gelfand-

Yaglom function Da (t) of Eq. (2.235) has the property (2.231) due to time reversal
invariance, and (2.268) reproduces the known result (2.159).
The expressions containing partial derivatives are easily extended to D dimen-
sions: We simply have to replace the partial derivatives ∂xb /∂ ẋa , ∂ ẋb /∂ ẋa , . . . by
the corresponding D × D matrices, and write the action as the associated quadratic
form.
The D-dimensional versions of the fluctuation factors (2.264) are
#−1/2 #1/2
∂xi ∂ ẋi
" "
1 1
FΩ (tb , ta ) = q D det jb =q D det aj . (2.270)
2πih̄/M ∂ ẋa 2πih̄/M ∂xb

All formulas for fluctuation factors hold initially only for sufficiently short times
tb − ta . For larger times, they carry phase factors determined as before in (2.169).
The fully defined expression may be written as
−1/2 1/2
1 ∂xib −iνπ/2 1 ∂ ẋia
FΩ (tb , ta ) = q D det j e =q D det j e−iνπ/2 ,
2πih̄/M ∂ ẋa 2πih̄/M ∂xb
(2.271)
where ν is the Maslov-Morse index. In the one-dimensional case it counts the turn-
ing points of the trajectory, in the multidimensional case the number of zeros in
determinant det ∂xib /∂ ẋja along the trajectory, if the zero is caused by a reduction
of the rank of the matrix ∂xib /∂ ẋja by one unit. If it is reduced by more than one
unit, ν increases accordingly. In this context, the number ν is also called the Morse
index of the trajectory.
The zeros of the functional determinant are also called conjugate points. They
are generalizations of the turning points in one-dimensional systems. The surfaces
in x-space, on which the determinant vanishes, are called caustics. The conjugate
points are the places where the orbits touch the caustics.13
Note that for infinitesimally short times, all fluctuation factors and classical ac-
tions coincide with those of a free particle. This is obvious for the time-independent
harmonic oscillator, where the amplitude (2.177) reduces to that of a free particle
13
See M.C. Gutzwiller, Chaos in Classical and Quantum Mechanics, Springer, Berlin, 1990.
130 2 Path Integrals — Elementary Properties and Simple Solutions

in Eq. (2.74) in the limit tb → ta . Since a time-dependent frequency is constant over

an infinitesimal time, this same result holds also here. Expanding the solution of
the equations of motion for infinitesimally short times as

xb ≈ (tb − ta )ẋa + xa , xa ≈ −(tb − ta )ẋb + xb , (2.272)

we have immediately

∂xib ∂xa
= δij (tb − ta ), = −δij (tb − ta ). (2.273)
∂ ẋja ∂ ẋjb

Similarly, the expansions

xb − xa
ẋb ≈ ẋa ≈ (2.274)
tb − ta
lead to
∂ ẋib 1 ∂ ẋia 1
j = −δij , j = δij . (2.275)
∂xa tb − ta ∂xb tb − ta

Inserting the expansions (2.273) or (2.274) into (2.266) (in D dimensions), the action
reduces approximately to the free-particle action

M (xb − xa )2
Acl ≈ . (2.276)
2 tb − ta

2.5.2 Momentum Space

Let us also find the time evolution amplitude in momentum space. For this we write the classical
action (2.267) as a quadratic form
 
M xb
Acl = (xb , xa ) A (2.277)
2 xa

with a matrix
∂ ẋb ∂ ẋb
 
 ∂xb ∂xa 
A= . (2.278)
 ∂ ẋa ∂ ẋ 
a
− −
∂xb ∂xa
The inverse of this matrix is
∂xb ∂xb
 
−
 ∂ ẋb ∂ ẋa 
A−1 =
 ∂x
. (2.279)
a ∂x 
a
−
∂ ẋb ∂ ẋa
The partial derivatives of xb and xa are calculated from the solution of the homogeneous differential
equation (2.218) specified in terms of the final and initial velocities ẋb and ẋa :
1
x(ẋb , ẋa ; t) =
Ḋa (tb )Ḋb (ta ) + 1
nh i h i o
× Da (t) + Db (t)Ḋa (tb ) ẋa + −Db (t) + Da (t)Ḋb (ta ) ẋb , (2.280)

H. Kleinert, PATH INTEGRALS

2.5 Harmonic Oscillator with Time-Dependent Frequency 131

which yields
1 h i
xa = Db (ta )Ḋa (ta )ẋb − Db (ta )ẋb , (2.281)
Ḋa (tb )Ḋb (ta ) + 1
1 h i
xb = Da (tb )ẋa + Da (tb )Ḋb (ta )ẋb , (2.282)
Ḋa (tb )Ḋb (ta ) + 1
so that  
Ḋb (ta ) −1
Da (tb )
A−1 =  . (2.283)
Ḋa (tb )Ḋb (ta ) + 1 −1 −Ḋa (tb )
The determinant of A is the Jacobian
∂(ẋb , ẋa ) Ḋa (tb )Ḋb (ta ) + 1
det A = − =− . (2.284)
∂(xb , xa ) Da (tb )Db (ta )
We can now perform the Fourier transform of the time evolution amplitude and find, via a quadratic
completion,
Z Z
(pb tb |pa ta ) = dxb e−ipb xb /h̄ dxa eipa xa /h̄ (xb tb |xa ta ) (2.285)
r s
2πh̄ Da (tb )
=
iM Ḋa (tb )Ḋb (ta ) + 1
 i
i 1 Da (tb ) h
2 2
× exp −Ḋb (ta )pb + Ḋa (tb )pa − 2pb pa .
h̄ 2M Ḋa (tb )Ḋb (ta ) + 1

Inserting here Da (tb ) = sin ω(tb − ta )/ω and Ḋa (tb ) = cos ω(tb − ta ), we recover the oscillator result
(2.189).
In D dimensions, the classical action has the same quadratic form as in (2.277)
 
M T T
xb
Acl = xb , xa A (2.286)
2 xa

with a matrix A generalizing (2.278) by having the partial derivatives replaced by the corresponding
D × D-matrices. The inverse is the 2D × 2D-version of (2.279), i.e.
∂ ẋb ∂ ẋb ∂xb ∂xb
   
−
 ∂xb ∂xa   ∂ ẋb ∂ ẋa 
A=  ∂ ẋ
, A−1 =  . (2.287)
a ∂ ẋa   ∂xa ∂xa 
− − −
∂xb ∂xa ∂ ẋb ∂ ẋa
The determinant of such a block matrix
 
a b
A= (2.288)
c d

is calculated after a triangular decomposition

a−1 b a − bd−1 c
       
a b a 0 1 1 b 0
A= = = (2.289)
c d c 1 0 d − ca−1 b 0 d d−1 c 1

in two possible ways as

 
a b
det = det a · det (d − ca−1 b) = det (a − bd−1 c) · det d, (2.290)
c d
132 2 Path Integrals — Elementary Properties and Simple Solutions

depending whether det a or det b is nonzero. The inverse is in the first case

1 −a−1 bx
 −1   −1 −1
a +a bxca−1 −a−1 bx
 
a 0
A= = , x ≡(d−ca−1b)−1. (2.291)
0 x −ca−1 1 −xca−1 x

The resulting amplitude in momentum space is

Z Z
−ipb xb /h̄
(pb tb |pa ta ) = dxb e dxa eipa xa /h̄ (xb tb |xa ta )
   
2π 1 i 1 T T

−1 pb
=√ √ exp pb , pa A . (2.292)
2πih̄M Dren det A h̄ 2M pa

Also in momentum space, the amplitude (2.292) reduces to the free-particle one in Eq. (2.73)
in the limit of infinitesimally short time tb − ta : For the time-independent harmonic oscillator, this
was shown in Eq. (2.191), and the time-dependence of Ω(t) becomes irrelevant in the limit of small
tb − ta → 0.

2.6 Free-Particle and Oscillator Wave Functions

In Eq. (1.335) we have expressed the time evolution amplitude of the free particle
(2.72) as a Fourier integral

dp ip(xb −xa )/h̄ −ip2 (tb −ta )/2M h̄

Z
(xb tb |xa ta ) = e e . (2.293)
(2πh̄)

This expression contains the information on all stationary states of the system. To
find these states we have to perform a spectral analysis of the amplitude. Recall
that according to Section 1.7, the amplitude of an arbitrary time-independent system
possesses a spectral representation of the form
∞
ψn (xb )ψn∗ (xa )e−iEn (tb −ta )/h̄ ,
X
(xb tb |xa ta ) = (2.294)
n=0

where En are the eigenvalues and ψn (x) the wave functions of the stationary states.
In the free-particle case the spectrum is continuous and the spectral sum is an
integral. Comparing (2.294) with (2.293) we see that the Fourier decomposition
itself happens to be the spectral representation. If the sum over n is written as an
integral over the momenta, we can identify the wave functions as
1
ψp (x) = √ eipx . (2.295)
2πh̄
For the time evolution amplitude of the harmonic oscillator
1
(xb tb |xa ta ) = q (2.296)
2πih̄ sin [ω(tb − ta )] /Mω
( )
iMω h i
× exp (x2b + x2a ) cos ω(tb − ta ) − 2xb xa ,
2h̄ sin [ω(tb − ta )]

H. Kleinert, PATH INTEGRALS

2.6 Free-Particle and Oscillator Wave Functions 133

the procedure is not as straight-forward. Here we must make use of a summation

formula for Hermite polynomials (see Appendix 2C) Hn (x) due to Mehler:14
( )
1 1
√ exp − 2)
[(x2 + x′2 )(1 + a2 ) − 4xx′ a]
1−a2 2(1 − a
∞
2 ′2 an
Hn (x)Hn (x′ ),
X
= exp(−x /2 − x /2) n
(2.297)
n=0 2 n!

with
2 dn −x2
H0 (x) = 1, H1 (x) = 2x, H2 (x) = 4x2 − 2, . . . , Hn (x) = (−1)n ex e . (2.298)
dxn
Identifying
q q
x≡ Mω/h̄ xb , x′ ≡ Mω/h̄ xa , a ≡ e−iω(tb −ta ) , (2.299)
so that
a 1 1 + a2 1 + e−2iω(tb −ta ) cos [ω(tb − ta )]
2
= , 2
= −2iω(t −t )
=
1−a 2i sin [ω(tb − ta )] 1−a 1−e b a i sin [ω(tb − ta )]
we arrive at the spectral representation
∞
ψn (xb )ψn (xa )e−i(n+1/2)ω(tb −ta ) .
X
(xb tb |xa ta ) = (2.300)
n=0

From this we deduce that the harmonic oscillator has the energy eigenvalues
En = h̄ω(n + 1/2) (2.301)
and the wave functions
2 /2λ2
ψn (x) = Nn λ−1/2
ω e−x ω
Hn (x/λω ). (2.302)
Here, λω is the natural length scale of the oscillator
s
h̄
λω ≡ , (2.303)
Mω
and Nn the normalization constant
√
Nn = (1/2n n! π)1/2 . (2.304)
It is easy to check that the wave functions satisfy the orthonormality relation
Z ∞
dx ψn (x)ψn′ (x)∗ = δnn′ , (2.305)
−∞

using the well-known orthogonality relation of Hermite polynomials15

1
Z ∞
2

n
√ dx e−x Hn (x)Hn′ (x) = δn,n′ . (2.306)
2 n! π −∞
14
See P.M. Morse and H. Feshbach, Methods of Theoretical Physics, McGraw-Hill, New York,
Vol. I, p. 781 (1953).
15
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 7.374.1.
134 2 Path Integrals — Elementary Properties and Simple Solutions

2.7 General Time-Dependent Harmonic Action

A simple generalization of the harmonic oscillator with time-dependent frequency allows also for
a time-dependent mass, so that the action (2.307) becomes
Z tb
M 
g(t)ẋ2 (t) − Ω2 (t)x2 (t) ,

A[x] = dt (2.307)
ta 2
with some dimensionless time-dependent factor g(t). This factor changes the measure of path
integration so that the time evolution amplitude can no longer be calculated from (2.260). To find
the correct measure we must return to the canonical path integral (2.29) which now reads
Z x(tb )=xb
Dp iA[p,x]/h̄
Z
(xb tb |xa ta ) = D′ x e , (2.308)
x(ta )=xa 2πh̄
with the canonical action
tb
p2
 
M 2
Z
2
A[p, x] = dt pẋ − − Ω (t)x (t) . (2.309)
ta 2M g(t) 2
Integrating the momentum variables out in the sliced form of this path integral as in Eqs. (2.51)–
(2.53) yields
N
"Z #
∞  
1 Y dxn i N
(xb tb |xa ta ) ≈ p p exp A . (2.310)
2πh̄iǫ/M g(tN +1) n=1 −∞ 2πh̄iǫ/M g(tn ) h̄

The continuum limit of this path integral is written as

 
√ i
Z
(xb tb |xa ta ) = Dx g exp A[x] , (2.311)
h̄
with the action (2.307).
The classical orbits solve the equation of motion
−∂t g(t)∂t − Ω2 (t) x(t) = 0,
 
(2.312)
which, by the transformation
ġ 2 (t) g̈(t)
 
p 1
x̃(t) = g(t)x(t), Ω̃2 (t) = Ω2 (t) + − , (2.313)
g(t) 4g(t) 2
can be reduced to the previous form
p h i
g(t) −∂t2 − Ω̃2 (t) x̃(t) = 0. (2.314)

The result of the path integration is therefore

√
Z
(xb tb |xa ta ) = Dx g eiA[x]/h̄ = F (xb , tb ; xa , ta )eiAcl /h̄ , (2.315)

with a fluctuation factor [compare (2.263)]

1 1
F (xb , tb ; xa , ta ) = p p , (2.316)
2πih̄/M Da (tb )
where Da (tb ) is found from a generalization of the formulas (2.264)–(2.269). The classical action
is
M
Acl = (gb xb ẋb − ga xa ẋa ), (2.317)
2

H. Kleinert, PATH INTEGRALS

2.8 Path Integrals and Quantum Statistics 135

where gb ≡ g(tb ), ga ≡ g(ta ). The solutions of the equation of motion can be expressed in terms
of modified Gelfand-Yaglom functions (2.228) and (2.229) with the properties
[∂t g(t)∂t + Ω2 (t)]Da (t) = 0 ; Da (ta ) = 0, Ḋa (ta ) = 1/ga , (2.318)
2
[∂t g(t)∂t + Ω (t)]Db (t) = 0 ; Db (tb ) = 0, Ḋb (tb ) = −1/gb, (2.319)
as in (2.248):
Db (t) Da (t)
x(xb , xa ; t) = xa + xb . (2.320)
Db (ta ) Da (tb )
This allows us to write the classical action (2.317) in the form
M h i
Acl = gb x2b Ḋa (tb ) − ga x2a Ḋb (ta ) − 2xb xa . (2.321)
2Da (tb )
From this we find, as in (2.269),
−1
∂ 2 Acl

Dren = Da (tb ) = Db (ta ) = −M , (2.322)
∂xb ∂xa
so that the fluctuation factor becomes
s
1 ∂ 2 Acl
F (xb , tb ; xa , ta ) = √ − . (2.323)
2πih̄ ∂xb ∂xa
As an example take a free particle with a time-dependent mass term, where
Z t Z tb Z tb
Da (t) = dt′ g −1 (t′ ), Db (t) = dt′ g −1 (t′ ), Dren = Da (tb ) = Db (ta ) = dt′ g −1 (t′ ), (2.324)
ta t ta

and the classical action reads

M (xb − xa )2
Acl = . (2.325)
2 Da (tb )
The result can easily be generalized to an arbitrary harmonic action
Z tb
M 
g(t)ẋ2 + 2b(t)xẋ − Ω2 (t)x2 ,

A= dt (2.326)
ta 2
which is extremized by the Euler-Lagrange equation [recall (1.8)]
h i
∂t g(t)∂t + ḃ(t) + Ω2 (t) x = 0. (2.327)

The solution of the path integral (2.315) is again given by (2.315), with the fluctuation factor
(2.323), where Acl is the action (2.326) along the classical path connecting the endpoints.
A further generalization to D dimensions is obvious by adapting the procedure in Subsec-
tion 2.4.6, which makes Eqs. (2.318)–(2.320). matrix equations.

2.8 Path Integrals and Quantum Statistics

The path integral approach is useful to also understand the thermal equilibrium
properties of a system. We assume the system to have a time-independent Hamilto-
nian and to be in contact with a reservoir of temperature T . As explained in Sec-
tion 1.7, the bulk thermodynamic quantities can be determined from the quantum-
statistical partition function
 
Z = Tr e−Ĥ/kB T = e−En /kB T .
X
(2.328)
n
136 2 Path Integrals — Elementary Properties and Simple Solutions

This, in turn, may be viewed as an analytic continuation of the quantum-mechanical

partition function  
ZQM = Tr e−i(tb −ta )Ĥ/h̄ (2.329)
to the imaginary time
ih̄
tb − ta = − ≡ −ih̄β. (2.330)
kB T
In the local particle basis |xi, the quantum-mechanical trace corresponds to an
integral over all positions so that the quantum-statistical partition function can be
obtained by integrating the time evolution amplitude over xb = xa and evaluating
it at the analytically continued time:
Z ∞ Z ∞ Z ∞
−β Ĥ
Z≡ dx z(x) = dx hx|e |xi = dx (x tb |x ta )|tb −ta =−ih̄β . (2.331)
−∞ −∞ −∞

The diagonal elements

z(x) ≡ hx|e−β Ĥ |xi = (x tb |x ta )|tb −ta =−ih̄β (2.332)

play the role of a partition function density. For a harmonic oscillator, this quantity
has the explicit form [recall (2.175)]
s !
1 ω Mω h̄βω 2
zω (x) = q exp − tanh x . (2.333)
2πh̄/M sinh h̄βω h̄ 2

By splitting the Boltzmann factor e−β Ĥ into a product of N + 1 factors e−ǫĤ/h̄

with ǫ = h̄/kB T (N + 1), we can derive for Z a similar path integral representa-
tion just as for the corresponding quantum-mechanical partition function in (2.42),
(2.48):
NY
+1 Z ∞ 
Z≡ dxn (2.334)
n=1 −∞

× hxN +1 |e−ǫĤ/h̄ |xN ihxN |e−ǫĤ/h̄ |xN −1 i × . . . × hx2 |e−ǫĤ/h̄ |x1 ihx1 |e−ǫĤ/h̄ |xN +1 i.

As in the quantum-mechanical case, the matrix elements hxn |e−ǫĤ/h̄ |xn−1 i are re-
expressed in the form
∞ dpn ipn (xn −xn−1 )/h̄−ǫH(pn ,xn )/h̄
Z
−ǫĤ/h̄
hxn |e |xn−1 i ≈ e , (2.335)
−∞ 2πh̄
with the only difference that there is now no imaginary factor i in front of the
Hamiltonian. The product (2.334) can thus be written as
NY
+1
"Z #
∞ ∞ dpn 1
Z  
Z≈ dxn exp − AN , (2.336)
n=1 −∞ −∞ 2πh̄ h̄ e

H. Kleinert, PATH INTEGRALS

2.8 Path Integrals and Quantum Statistics 137

where AN
e denotes the sum

N +1
AN
X
e = [−ipn (xn − xn−1 ) + ǫH(pn , xn )] . (2.337)
n=1

In the continuum limit ǫ → 0, the sum goes over into the integral
Z h̄β
Ae [p, x] = dτ [−ip(τ )ẋ(τ ) + H(p(τ ), x(τ ))], (2.338)
0

and the partition function is given by the path integral

Dp −Ae [p,x]/h̄
Z Z
Z= Dx e . (2.339)
2πh̄
In this expression, p(τ ), x(τ ) may be considered as paths running along an “imag-
inary time axis” τ = it. The expression Ae [p, x] is very similar to the mechanical
canonical action (2.27). Since it governs the quantum-statistical path integrals it
is called quantum-statistical action or Euclidean action, indicated by the subscript
e. The name alludes to the fact that a D-dimensional Euclidean space extended
by an imaginary-time axis τ = it has the same geometric properties as a D + 1-
dimensional Euclidean space. For instance, a four-vector in a Minkowski spacetime
has a square length dx2 = −(cdt)2 + (dx)2 . Continued to an imaginary time, this
becomes dx2 = (cdτ )2 + (dx)2 which is the square distance in a Euclidean four-
dimensional space with four-vectors (cτ, x).
The integrand of the Euclidean action (2.339) is the Euclidean Lagrangian Le . It
is related to the Hamiltonian by the Euclidean Legendre transform [compare (1.9)]
∂Le
H = Le + i ẋ = Le + ipẋ (2.340)
∂ ẋ
in which ẋ is eliminated in favor of p = ∂Le /∂ ẋ [compare (1.10)].
Just as in the path integralH for the quantum-mechanical partition function (2.48),
the measure of integration Dx Dp/2πh̄ in the quantum-statistical expression
R

(2.339) is automatically symmetric in all p’s and x’s:

NY
+1 Z ∞
I Z
Dp I
Dp Z Z ∞
dpn
Dx = Dx = dxn . (2.341)
2πh̄ 2πh̄ n=1 −∞ −∞ 2πh̄

The symmetry is of course due to the trace integration over all initial ≡ final posi-
tions.
Most remarks made in connection with Eq. (2.48) carry over to the present
case. The above path integral (2.339) is a natural extension of the rules of clas-
sical statistical mechanics. According to these, each cell in phase space dxdp/h is
occupied with equal statistical weight, with the probability factor e−E/kB T . In quan-
tum statistics, the paths of all particles fluctuate evenly over the cells in path phase
space n dx(τn )dp(τn )/h (τn ≡ nǫ), each path carrying a probability factor e−Ae /h̄
Q

involving the Euclidean action of the system.

138 2 Path Integrals — Elementary Properties and Simple Solutions

2.9 Density Matrix

The partition function does not determine any local thermodynamic quantities. Im-
portant local information resides in the thermal analog of the time evolution ampli-
tude hxb |e−Ĥ/kB T |xa i. Consider, for instance, the diagonal elements of this amplitude
renormalized by a factor Z −1 :

ρ(xa ) ≡ Z −1 hxa |e−Ĥ/kB T |xa i. (2.342)

They determine the thermal average of the particle density of a quantum-statistical

system. Due to (2.334), the factor Z −1 makes the spatial integral over ρ equal to
unity: Z ∞
dx ρ(x) = 1. (2.343)
−∞

By inserting into (2.342) a complete set of eigenfunctions ψn (x) of the Hamiltonian

operator Ĥ, we find the spectral decomposition
.X
|ψn (xa )|2 e−βEn e−βEn .
X
ρ(xa ) = (2.344)
n n

Since |ψn (xa )|2 is the probability distribution of the system in the eigenstate |ni,
while the ratio e−βEn / n e−βEn is the normalized probability to encounter the sys-
P

tem in the state |ni, the quantity ρ(xa ) represents the normalized average particle
density in space as a function of temperature.
Note the limiting properties of ρ(xa ). In the limit T → 0, only the lowest energy
state survives and ρ(xa ) tends towards the particle distribution in the ground state
T →0
−−→ |ψ0 (xa )|2 .
ρ(xa ) − (2.345)

In the opposite limit of high temperatures, quantum effects are expected to become
irrelevant and the partition function should converge to the classical expression
(1.538) which is the integral over the phase space of the Boltzmann distribution
T →∞ ∞ ∞ dp −H(p,x)/kB T
Z Z
Z−
−−→ Zcl = dx e . (2.346)
−∞ −∞ 2πh̄
We therefore expect the large-T limit of ρ(x) to be equal to the classical particle
distribution
T →∞ ∞ dp −H(p,x)/kB T
Z
ρ(x) −
−−→ ρcl (x) = Zcl−1 e . (2.347)
−∞ 2πh̄
Within the path integral approach, this limit will be discussed in more detail in
Section 2.13. At this place we roughly argue as follows: When going in the original
time-sliced path integral (2.334) to large T , i.e., small τb − τa = h̄/kB T , we may
keep only a single time slice and write
Z ∞ 
Z≈ dx hx|e−ǫĤ/h̄ |xi, (2.348)
−∞

H. Kleinert, PATH INTEGRALS

2.9 Density Matrix 139

with
∞ dpn −ǫH(pn ,x)/h̄
Z
−ǫĤ
hx|e |xi ≈ e . (2.349)
−∞ 2πh̄
After substituting ǫ = τb − τa this gives directly (2.347). Physically speaking, the
path has at high temperatures “no (imaginary) time” to fluctuate, and only one
term in the product of integrals needs to be considered.
If H(p, x) has the standard form
p2
H(p, x) =
+ V (x), (2.350)
2M
the momentum integral is Gaussian in p and can be done using the formula
∞ dp −ap2 /2h̄ 1
Z
e =√ . (2.351)
−∞ 2πh̄ 2πh̄a
This leads to the pure x-integral for the classical partition function
∞ dx ∞ dx −βV (x)
Z Z
−V (x)/kB T
Zcl = q e = e . (2.352)
−∞ 2
2πh̄ /MkB T −∞ le (h̄β)

In the second expression we have introduced the length

q
le (h̄β) ≡ 2πh̄2 β/M. (2.353)
It is the thermal (or Euclidean) analog of the characteristic length l(tb − ta ) intro-
duced before in (2.131). It is called the de Broglie wavelength associated with the
temperature T = 1/kB β or, in short, the thermal de Broglie wavelength.
Omitting the x-integration in (2.352) renders the large-T limit ρ(x), the classical
particle distribution
T →∞ 1
−−→ ρcl (x) = Zcl−1
ρ(x) − e−V̄ (x) . (2.354)
le (h̄β)
For a free particle, the integral over x in (2.352) diverges. If we imagine the length
of the x-axis to be very large but finite, say equal to L, the partition function is
equal to
L
Zcl = . (2.355)
le (h̄β)
In D dimensions, this becomes
VD
Zcl = D
, (2.356)
le (h̄β)
where VD is the volume of the D-dimensional system. For a harmonic oscillator
with potential Mω 2 x2 /2, the integral over x in (2.352) is finite and yields, in the
D-dimensional generalization
lD
Zcl = D ω , (2.357)
l (h̄β)
140 2 Path Integrals — Elementary Properties and Simple Solutions

where s
2π
lω ≡ (2.358)
βMω 2
denotes the classical length scale defined by the frequency of the harmonic oscillator.
It is related to the quantum-mechanical one λω of Eq. (2.303) by

lω le (h̄β) = 2π λ2ω . (2.359)

Thus we obtain the mnemonic rule for going over from the partition function of a
harmonic oscillator to that of a free particle: we must simply replace

lω −
−−→ L, (2.360)
ω→0

or s
1 βM
−
−−→ L. (2.361)
ω ω→0 2π
The real-time version of this is, of course,
s
1 (tb − ta )M
−
−−→ L. (2.362)
ω ω→0 2πh̄
Let us write down a path integral representation for ρ(x). Omitting in (2.339)
the final trace integration over xb ≡ xa and normalizing the expression by a factor
Z −1 , we obtain
x(h̄β)=xb Dp −Ae [p,x]/h̄
Z Z
ρ(xa ) = Z −1 D′x e
x(0)=xa 2πh̄
Z x(h̄β)=xb
= Z −1 Dxe−Ae [x]/h̄ . (2.363)
x(0)=xa

The thermal equilibrium expectation of an arbitrary Hermitian operator Ô is

given by
hÔiT ≡ Z −1 e−βEn hn|Ô|ni.
X
(2.364)
n

In the local basis |xi, this becomes

ZZ ∞
−1
hÔiT = Z dxb dxa hxb |e−β Ĥ |xa ihxa |Ô|xb i. (2.365)
−∞

An arbitrary function of the position operator x̂ has the expectation

ZZ ∞ Z
hf (x̂)iT = Z −1 dxb dxa hxb |e−β Ĥ |xa iδ(xb − xa )f (xa ) = dxρ(x)f (x). (2.366)
−∞

The particle density ρ(xa ) determines the thermal averages of local observables.
If f depends also on the momentum operator p̂, then the off-diagonal matrix
elements hxb |e−β Ĥ |xa i are also needed. They are contained in the density matrix

H. Kleinert, PATH INTEGRALS

2.9 Density Matrix 141

introduced for pure quantum systems in Eq. (1.221), and reads now in a thermal
ensemble of temperature T :

ρ(xb , xa ) ≡ Z −1 hxb |e−β Ĥ |xa i, (2.367)

whose diagonal values coincide with the above particle density ρ(xa ).
It is useful to keep the analogy between quantum mechanics and quantum statis-
tics as close as possible and to introduce the time translation operator along the
imaginary time axis

Ûe (τb , τa ) ≡ e−(τb −τa )Ĥ/h̄ , τb > τa , (2.368)

defining its local matrix elements as imaginary or Euclidean time evolution
amplitudes16
(xb τb |xa τa ) ≡ hxb |Ûe (τb , τa )|xa i, τb > τa . (2.369)
As in the real-time case, we shall only consider the causal time-ordering τb > τa .
Otherwise the partition function and the density matrix do not exist in systems with
energies up to infinity. Given the imaginary-time amplitudes, the partition function
is found by integrating over the diagonal elements
Z ∞
Z= dx(x h̄β|x 0), (2.370)
−∞

and the density matrix

ρ(xb , xa ) = Z −1 (xbh̄β|xa 0). (2.371)
For the sake of generality we may sometimes also consider the imaginary-time
evolution operators for time-dependent Hamiltonians and the associated amplitudes.
They are obtained by time-slicing the local matrix elements of the operator
1 τb
 Z 
Û (τb , τa ) = Tτ exp − dτ Ĥ(−iτ ) . (2.372)
h̄ τa

Here Tτ is an ordering operator along the imaginary-time axis.

It must be emphasized that the usefulness of the operator (2.372) in describing
thermodynamic phenomena is restricted to the Hamiltonian operator Ĥ(t) depend-
ing very weakly on the physical time t. The system has to remain close to equilibrium
at all times. This is the range of validity of the so-called linear response theory (see
Chapter 18 for more details).
The imaginary-time evolution amplitude (2.369) has a path integral representa-
tion which is obtained by dropping the final integration in (2.336) and relaxing the
condition xb = xa :
N Z  NY
+1 Z
" #
∞ ∞ dpn  
exp −AN
Y
(xb τb |xa τa ) ≈ dxn e /h̄ . (2.373)
n=1 −∞ n=1 −∞ 2πh̄
16
The imaginary-time amplitude (xb β|xa 0) = hxb |e−β Ĥ |xa i is often referred to as heat kernel .
142 2 Path Integrals — Elementary Properties and Simple Solutions

The time-sliced Euclidean action is

N +1
AN
X
e = [−ipn (xn − xn−1 ) + ǫH(pn , xn , τn )] (2.374)
n=1

(we have omitted the factor −i in the τ -argument of H). In the continuum limit
this is written as a path integral
Dp 1
Z Z  
′
(xb τb |xa τa ) = Dx exp − Ae [p, x] (2.375)
2πh̄ h̄
[by analogy with (2.339)]. For a Hamiltonian of the standard form (2.7),

p2
H(p, x, τ ) = + V (x, τ ),
2M
with a smooth potential V (x, τ ), the momenta can be integrated out, just as in
(2.53), and the Euclidean version of the pure x-space path integral (2.54) leads to
(2.55):
( )
1 h̄β M
Z Z 
(xb τb |xa τa ) = Dx exp − dτ (∂τ x)2 + V (x, τ )
h̄ 0 2
 
N
1 Y Z ∞ dxn
≈ q  q  (2.376)
2πh̄ǫ/M n=1 −∞ 2πβ/M
1 NX+1
( " 2 #)
M xn − xn−1

× exp − ǫ + V (xn , τn ) .
h̄ n=1 2 ǫ

From this we calculate the quantum-statistical partition function

Z ∞
Z = dx (x h̄β|x 0)
−∞
Z Z x(h̄β)=x I
= dx Dx e−Ae [x]/h̄ = Dx e−Ae [x]/h̄ , (2.377)
x(0)=x

where Ae [x] is the Euclidean version of the Lagrangian action

τb M ′2
Z  
Ae [x] = dτ x + V (x, τ ) . (2.378)
τa 2
The prime denotes differentiation with respect to the imaginary time. As in the
quantum-mechanical partition function in (2.63), the path integral Dx now stands
H

for
NY
+1 Z ∞
I
dxn
Dx ≈ q . (2.379)
n=1 −∞ 2πh̄ǫ/M
q
It contains no extra 1/ 2πh̄ǫ/M factor, as in (2.376), due to the trace integration
over the exterior x.

H. Kleinert, PATH INTEGRALS

2.10 Quantum Statistics of the Harmonic Oscillator 143

The condition x(h̄β) = x(0) is most easily enforced by expanding x(τ ) into a
Fourier series
∞
1
e−iωm τ xm ,
X
x(τ ) = √ (2.380)
m=−∞ N +1
with the Matsubara frequencies
2πm
ωm ≡ 2πmkB T /h̄ = , m = 0, ±1, ±2, . . . . (2.381)
h̄β
When considered as functions on the entire τ -axis, the paths are periodic in h̄β at
any τ , i.e.,
x(τ ) = x(τ + h̄β). (2.382)
Thus the path integral for the quantum-statistical partition function comprises all
periodic paths with a period h̄β. In the time-sliced path integral (2.376), the coor-
dinates x(τ ) are needed only at the discrete times τn = nǫ. Correspondingly, the
sum over m in (2.380) can be restricted to run from m = −N/2 to N/2 for even N
and from −(N − 1)/2 to (N + 1)/2 for odd N (see Fig. 2.3). In order to have a real
x(τn ), we must require that

xm = x∗−m (modulo N + 1). (2.383)

Note that the Matsubara frequencies in the expansion of the paths x(τ ) are now
twice as big as the frequencies νm in the quantum fluctuations (2.115) (after analytic
continuation of tb − ta to −ih̄/kB T ). Still, they have about the same total number,
since they run over positive and negative integers. An exception is the zero frequency
ωm = 0, which is included here, in contrast to the frequencies νm in (2.115) which
run only over positive m = 1, 2, 3, . . . . This is necessary to describe paths with
arbitrary nonzero endpoints xb = xa = x (included in the trace).

2.10 Quantum Statistics of the Harmonic Oscillator

The harmonic oscillator is a good example for solving the quantum-statistical path integral. The
τ -axis is sliced at τn = nǫ, with ǫ ≡ h̄β/(N + 1) (n = 0, . . . , N + 1), and the partition function is
given by the N → ∞ -limit of the product of integrals
N
"Z #
∞
N
Y dxn
exp −AN


Zω = p e /h̄ , (2.384)
n=0 −∞ 2πh̄ǫ/M

where AN
e is the time-sliced Euclidean oscillator action

N +1
M X
AN
e = xn (−ǫ2 ∇∇ + ǫ2 ω 2 )xn . (2.385)
2ǫ n=1

Integrating out the xn ’s, we find immediately

1
ZωN = q . (2.386)
detN +1 (−ǫ2 ∇∇ + ǫ2 ω 2 )
144 2 Path Integrals — Elementary Properties and Simple Solutions

Figure 2.3 Illustration of the eigenvalues (2.387) of the fluctuation matrix in the action
(2.385) for even and odd N .

Let us evaluate the fluctuation determinant via the product of eigenvalues which diagonalize
the matrix −ǫ2 ∇∇ + ǫ2 ω 2 in the sliced action (2.385). They are
ǫ2 Ωm Ωm + ǫ2 ω 2 = 2 − 2 cos ωm ǫ + ǫ2 ω 2 , (2.387)
with the Matsubara frequencies ωm . For ω = 0, the eigenvalues are pictured in Fig. 2.3. The
action (2.385) becomes diagonal after going to the Fourier components xm . To do this we arrange
the real and imaginary parts Re xm and Im xm in a row vector
(Re x1 , Im x1 ; Re x2 , Im x2 ; . . . ; Re xn , Im xn ; . . .),
and see that it is related to the time-sliced positions xn = x(τn ) by a transformation matrix with
the rows
Tmn xn = (Tm )n xn
r
2  1 m m
= √ , cos 2π · 1, sin 2π · 1,
N +1 2 N +1 N +1
m m
cos 2π · 2, sin 2π · 2, . . .
N +1 N +1
m m 
. . . , cos 2π · n, sin 2π · n, . . . xn . (2.388)
N +1 N +1 n

For each row index m = 0, . . . , N, the column index n runs from zero to N/2 for even N , and to
(N + 1)/2 for odd N . In the odd case, the last column sin Nm +1 2π · n with n = (N + 1)/2 vanishes
identically and must be dropped, so that the number of columns in Tmn is in both cases N + 1,
as it should be. For odd N , the second-last column of Tmn is an alternating sequence √ ±1. Thus,
for a proper normalization, it has to be multiplied by an extra normalization factor 1/ 2, just as
the elements in the first column. An argument similar to (2.120), (2.121) shows that the resulting
matrix is orthogonal. Thus, we can diagonalize the sliced action in (2.385) as follows
 h PN/2 i

 ω 2 x20 + 2 m=1 (Ωm Ωm + ω 2 )|xm |2 for N = even,
N M   2 2
Ae = ǫ 2 2
ω x0 + (Ω(N +1)/2 Ω(N +1)/2 + ω )xN +1 (2.389)
2  P(N −1)/2 i
2 2
(Ω Ω + ω )|x |

 +2 m=1 m m for
m N = odd.
Q R∞
Thanks to the orthogonality of Tmn , the measure n −∞ dx(τn ) transforms simply into
Z ∞ N/2 ∞ ∞
YZ Z
dx0 d Re xm d Im xm for N = even,
−∞ m=1 −∞ −∞

H. Kleinert, PATH INTEGRALS

2.10 Quantum Statistics of the Harmonic Oscillator 145

(2.390)
Z ∞ Z ∞ (N −1)/2 Z ∞ Z ∞
Y
dx0 dx(N +1)/2 d Re xm d Im xm for N = odd.
−∞ −∞ m=1 −∞ −∞

By performing the Gaussian integrals we obtain the partition function

" N #−1/2
2 2 −1/2
Y
N 2 2 2 2
 
Zω = detN +1 (−ǫ ∇∇ + ǫ ω ) = (ǫ Ωm Ωm + ǫ ω )
m=0
)−1/2 " N
N  −1/2
( #
2 ωm ǫ
Y Y
2 2 2 2
 
= 2(1 − cos ωm ǫ) + ǫ ω = 4 sin +ǫ ω . (2.391)
m=0 m=0
2

Thanks to the periodicity of the eigenvalues under the replacement n → n + N + 1, the result has
become a unique product expression for both even and odd N .
It is important to realize that contrary to the fluctuation factor (2.162) in the real-time am-
plitude, the partition function (2.391) contains the square root of only positive eigenmodes as a
unique result of Gaussian integrations. There are no phase subtleties as in the Fresnel integral
(1.337).
To calculate the product, we observe that upon decomposing
ωm ǫ  ωm ǫ   ωm ǫ 
sin2 = 1 + cos 1 − cos , (2.392)
2 2 2
the sequence of first factors
ωm ǫ πm
1 + cos ≡ 1 + cos (2.393)
2 N +1
runs for m = 1, . . . N through the same values as the sequence of second factors
ωm ǫ πm N +1−m
1 − cos = 1 − cos ≡ 1 + cos π , (2.394)
2 N +1 N +1
except in an opposite order. Thus, separating out the m = 0 -term, we rewrite (2.391) in the form
" N
#−1 " N
!#−1/2
1 Y ωm ǫ  Y ǫ2 ω 2
ZωN = 2 1 − cos 1+ . (2.395)
ǫω m=1
2 m=1
4 sin2 ωm 2
ǫ

The first factor on the right-hand side is the quantum-mechanical fluctuation determinant of the
free-particle determinant detN (−ǫ2 ∇∇) = N + 1 [see (2.128)], so that we obtain for both even and
odd N
" N !#−1/2
2 2
k B T Y ǫ ω
ZωN = 1+ . (2.396)
h̄ω m=1 4 sin2 ωm2
ǫ

To evaluate the remaining product, we must distinguish again between even and odd cases of N .
For even N , where every eigenvalue occurs twice (see Fig. 2.3), we obtain

N/2
!−1
2 2
k B T Y ǫ ω
ZωN =  1+  . (2.397)
h̄ω m=1 4 sin2 Nmπ+1

For odd N , the term with m = (N + 1)/2 occurs only once and must be treated separately so that
 !−1
2 2 1/2 (NY −1)/2 2 2
 
kB T  ǫ ω ǫ ω
ZωN = 1+ 1+ 2 πm
 . (2.398)
h̄ω 4 m=1
4 sin N +1
146 2 Path Integrals — Elementary Properties and Simple Solutions

We now introduce the parameter ω̃e , the Euclidean analog of (2.163), via the equations
ω̃e ǫ ωǫ ω̃e ǫ ωǫ
sin i ≡i , sinh ≡ . (2.399)
2 2 2 2
In the odd case, the product formula17
(N −1)/2
" #
Y sin2 x 2 sin[(N + 1)x]
1− 2 mπ = (2.400)
m=1
sin (N +1) sin 2x (N + 1)

[similar to (2.165)] yields, with x = ω̃e ǫ/2,

 −1
N kB T 1 sinh[(N + 1)ω̃e ǫ/2]
Zω = . (2.401)
h̄ω sinh(ω̃e ǫ/2) N +1

In the even case, the formula18

N/2
" #
Y sin2 x 1 sin[(N + 1)x]
1− = , (2.402)
m=1
sin2 (Nmπ
+1)
sin x (N + 1)

produces once more the same result as in Eq. (2.401). Inserting Eq. (2.399) leads to the partition
function on the sliced imaginary time axis:
1
ZωN = . (2.403)
2 sinh(h̄ω̃e β/2)
The partition function can be expanded into the following series

ZωN = e−h̄ω̃e /2kB T + e−3h̄ω̃e /2kB T + e−5h̄ω̃e /2kB T + . . . . (2.404)

By comparison with the general spectral expansion (2.328), we display the energy eigenvalues of
the system:
 
1
En = n + h̄ω̃e . (2.405)
2
They show the typical linearly rising oscillator sequence with
2 ωǫ
ω̃e = arsinh (2.406)
ǫ 2
playing the role of the frequency on the sliced time axis, and h̄ω̃e /2 being the zero-point energy.
In the continuum limit ǫ → 0, the time-sliced partition function ZωN goes over into the usual
oscillator partition function
1
Zω = . (2.407)
2 sinh(βh̄ω/2)
In D dimensions this becomes, of course, [2 sinh(βh̄ω/2)]−D , due to the additivity of the action in
each component of x.
Note that the continuum limit of the product in (2.396) can also be taken factor by factor.
Then Zω becomes
" ∞  #−1
kB T Y ω2
Zω = 1+ 2 . (2.408)
h̄ω m=1 ωm

17
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 1.391.1.
18
ibid., formula 1.391.3.

H. Kleinert, PATH INTEGRALS

2.10 Quantum Statistics of the Harmonic Oscillator 147

 
Q∞ x2
According to formula (2.173), the product m=1 1+ m2 π 2 converges rapidly against sinh x/x
and we find with x = h̄ωβ/2
kB T h̄ω/2kB T 1
Zω = = . (2.409)
h̄ω sinh(h̄ω/2kB T ) 2 sinh(βh̄ω/2)
As discussed after Eq. (2.183), the continuum limit can be taken in each factor since the product
in (2.396) contains only ratios of frequencies.
Just as in the quantum-mechanical case, this procedure of obtaining the continuum limit can
be summarized in the sequence of equations arriving at a ratio of differential operators
−1/2
ZωN detN +1 (−ǫ2 ∇∇ + ǫ2 ω 2 )

=
" #−1/2
 ′ 2
−1/2 detN +1 (−ǫ2 ∇∇ + ǫ2 ω 2 )
= detN +1 (−ǫ ∇∇)
det′N +1 (−ǫ2 ∇∇)
−1/2 ∞  2 −1
kB T det(−∂τ2 + ω 2 ) + ω2

ǫ→0 kB T Y ωm
−
−−→ = . (2.410)
h̄ det′ (−∂τ2 ) h̄ω m=1 ωm2

In the ω = 0 -determinants, the zero Matsubara frequency is excluded to obtain a finite expression.
This is indicated by a prime. The differential operator −∂τ2 acts on real functions which are periodic
2
under the replacement τ → τ +h̄β. Remember that each eigenvalue ωm of −∂τ2 occurs twice, except
for the zero frequency ω0 = 0, which appears only once.
Let us finally mention that the results of this section could also have been obtained directly
from the quantum-mechanical amplitude (2.175) [or with the discrete times from (2.199)] by an
analytic continuation of the time difference tb − ta to imaginary values −i(τb − τa ):
1 ω
r
(xb τb |xa τa ) = p
2πh̄/M sinh ω(τb − τa )
 
1 Mω
× exp − [(x2b + x2a ) cosh ω(τb − τa ) − 2xb xa ] . (2.411)
2h̄ sinh ω(τb − τa )
By setting x = xb = xa and integrating over x, we obtain [compare (2.333)]
Z ∞ s
1 ω(τb − τa )
Zω = dx (x τb |x τa ) = p
−∞ 2πh̄(τb − τa )/M sinh[ω(τ b − τa )]
p
2πh̄ sinh[ω(τb − τa )]/ωM 1
× = . (2.412)
2 sinh[ω(τb − τa )/2] 2 sinh[ω(τb − τa )/2]
Upon equating τb − τa = h̄β, we retrieve the partition function (2.407). A similar treatment of
the discrete-time version (2.199) would have led to (2.403). The main reason for presenting an
independent direct evaluation in the space of real periodic functions was to display the frequency
structure of periodic paths and to see the difference with respect to the quantum-mechanical paths
with fixed ends. We also wanted to show how to handle the ensuing product expressions.
For applications in polymer physics (see Chapter 15) one also needs the partition function of
all path fluctuations with open ends
Z ∞ Z ∞ s
1 ω(τb − τa ) 2πh̄
Zωopen = dxb dxa (xb τb |xa τa ) = p
−∞ −∞ 2πh̄(τb − τa )/M sinh[ω(τb − τa )] M ω
r
2πh̄ 1
= p . (2.413)
M ω sinh[ω(τb − τa )]
√
The prefactor is 2π times the length scale λω of Eq. (2.303).
148 2 Path Integrals — Elementary Properties and Simple Solutions

2.11 Time-Dependent Harmonic Potential

It is often necessary to calculate thermal fluctuation determinants for the case of a time-dependent
frequency Ω(τ ) which is periodic under τ → τ + h̄β. As in Section 2.3.6, we consider the amplitude
Dp − ττb dτ [−ipẋ+p2 /2M +MΩ2 (τ )x2 /2]/h̄
Z Z R
′
(xb τb |xa τa ) = Dx e a
2πh̄
Z R τb
− dτ [M ẋ2 +Ω2 (τ )x2 ]/2h̄
= Dxe τa . (2.414)

The time-sliced fluctuation factor is [compare (2.202)]

F N (τa − τb ) = detN +1 [−ǫ2 ∇∇ + ǫΩ2 (τ )]−1/2 , (2.415)
with the continuum limit
−1/2
det(−∂τ2 + Ω2 (τ ))

kB T
F (τa − τb ) = . (2.416)
h̄ det′ (−∂τ2 )
Actually, in the thermal case it is preferable to use the oscillator result for normalizing the
fluctuation factor, rather than the free-particle result, and to work with the formula
−1/2
det(−∂τ2 + Ω2 (τ ))

1
F (τb , τa ) = . (2.417)
2 sinh(βh̄ω/2) det(−∂τ2 + ω 2 )
This has the advantage that the determinant in the denominator contains no zero eigenvalue which
would require a special treatment as in (2.410); the operator −∂τ2 + ω 2 is positive.
As in the quantum-mechanical case, the spectrum of eigenvalues is not known for general Ω(τ ).
It is, however, possible to find a differential equation for the entire determinant, analogous to the
Gelfand-Yaglom formula (2.209), with the initial condition (2.214), although the derivation is now
much more tedious. The origin of the additional difficulties lies in the periodic boundary condition
which introduces additional nonvanishing elements −1 in the upper right and lower left corners of
the matrix −ǫ2 ∇∇ [compare (2.107)]:
 
2 −1 0 ... 0 0 −1
 −1 2 −1 ... 0 0 0 
 .. ..  .
 
2
−ǫ ∇∇ =  . .  (2.418)
 
 0 0 0 . . . −1 2 −1 
−1 0 0 . . . 0 −1 2
To better understand the relation with the previous result we shall replace the corner elements
−1 by −α which can be set equal to zero at the end, for a comparison. Adding to −ǫ2 ∇∇ a
time-dependent frequency matrix we then consider the fluctuation matrix
2 + ǫ2 Ω2N +1
 
−1 0 ... 0 −α
 −1 2 + ǫ2 Ω2N −1 . . . 0 0 
2 2 2
−ǫ ∇∇ + ǫ Ω =  .. .. .
 
 . . 
−α 0 0 . . . −1 2 + ǫ2 Ω21
(2.419)
Let us denote the determinant of this (N + 1) × (N + 1) matrix by D̃N +1 . Expanding it along the
first column, it is found to satisfy the equation
D̃N +1 = (2 + ǫ2 Ω2N +1 ) (2.420)
2 + ǫ2 Ω2N
 
−1 0 . . . 0 0
×detN 
 .. .. 
. . 
0 0 0 . . . −1 2 + ǫ2 Ω21

H. Kleinert, PATH INTEGRALS

2.11 Time-Dependent Harmonic Potential 149

 
−1 0 0 0 ... 0 −α
 −1 2 + ǫ2 Ω2N −1 −1 0 ... 0 0 
2 + ǫ2 Ω2N −2
 
+detN 
 0 −1 −1 . . . 0 0 
.. ..

 
 . . 
0 0 0 0 . . . −1 2 + ǫ2 Ω21

 
−1 0 0 ... 0 −α
 2 + ǫ2 Ω2N −1 0 ... 0 0 
2 + ǫ2 Ω2N −1
 
+(−1) N +1
αdetN 
 −1 −1 . . . 0 0 .

 .. .. 
 . . 
0 0 0 . . . 2 + ǫ2 Ω22 −1

The first determinant was encountered before in Eq. (2.204) (except that there it appeared with
−ǫ2 Ω2 instead of ǫ2 Ω2 ). There it was denoted by DN , satisfying the difference equation

−ǫ2 ∇∇ + ǫ2 Ω2N +1 DN = 0,


(2.421)

with the initial conditions

D1 = 2 + ǫ2 Ω21 ,
D2 = (2 + ǫ2 Ω21 )(2 + ǫ2 Ω22 ) − 1. (2.422)

The second determinant in (2.420) can be expanded with respect to its first column yielding

−DN −1 − α. (2.423)

The third determinant is more involved. When expanded along the first column it gives

(−1)N 1 + (2 + ǫ2 Ω2N )HN −1 − HN −2 ,

 
(2.424)

with the (N − 1) × (N − 1) determinant

HN −1 ≡ (−1)N −1 (2.425)
 
0 0 0 ... 0 0 −α
 2 + ǫ2 Ω2N −1 −1 0 ... 0 0 0 
2 + ǫ2 Ω2N −2
 
×detN −1 
 −1 −1 . . . 0 0 0 .

 .. .. 
 . . 
0 0 0 . . . −1 2 + ǫ2 Ω22 −1

By expanding this along the first column, we find that HN satisfies the same difference equation
as DN :

(−ǫ2 ∇∇ + ǫ2 Ω2N +1 )HN = 0. (2.426)

However, the initial conditions for HN are different:

0 −α
H2 = = α(2 + ǫ2 Ω22 ), (2.427)
2 + ǫ2 Ω22 −1

0 −α0
H3 = −
2 + ǫ2 Ω23 −1
0
−1 2 + ǫ2 Ω22
−1
= α (2 + ǫ2 Ω22 )(2 + ǫ2 Ω23 ) − 1 .
 
(2.428)
150 2 Path Integrals — Elementary Properties and Simple Solutions

They show that HN is in fact equal to αDN −1 , provided we shift Ω2N by one lattice unit upwards
to Ω2N +1 . Let us indicate this by a superscript +, i.e., we write
+
HN = αDN −1 . (2.429)

Thus we arrive at the equation

D̃N +1 = (2 + ǫ2 Ω2N )DN − DN −1 − α

+ +
−α[1 + (2 + ǫ2 Ω2N )αDN −2 − αDN −3 ]. (2.430)
+
Using the difference equations for DN and DN , this can be brought to the convenient form
+
D̃N +1 = DN +1 − α2 DN −1 − 2α. (2.431)

For quantum-mechanical fluctuations with α = 0, this reduces to the earlier result in Section 2.3.6.
For periodic fluctuations with α = 1, the result is
+
D̃N +1 = DN +1 − DN −1 − 2. (2.432)
+
In the continuum limit, DN +1 − DN −1 tends towards 2Ḋren , where Dren (τ ) = Da (t) is the
imaginary-time version of the Gelfand-Yaglom function in Section 2.4 solving the homogenous
differential equation (2.215), with the initial conditions (2.213) and (2.214), or Eqs. (2.228). The
corresponding properties are now:
 2
−∂τ + Ω2 (τ ) Dren (τ ) = 0, Dren (0) = 0,

Ḋren (0) = 1. (2.433)

In terms of Dren (τ ), the determinant is given by the Gelfand-Yaglom-like formula

ǫ→0
det(−ǫ2 ∇∇ + ǫΩ2 )T −
−−→ 2[Ḋren(h̄β) − 1], (2.434)

and the partition function reads

1
ZΩ = r h i. (2.435)
2 Ḋren (h̄β) − 1

The result may be checked by going back to the amplitude (xb tb |xa ta ) of Eq. (2.262), continuing
it to imaginary times t = iτ , setting xb = xa = x, and integrating over all x. The result is
1
ZΩ = q , tb = ih̄β, (2.436)
2 Ḋa (tb ) − 1

in agreement with (2.435).

As an example, take the harmonic oscillator for which the solution of (2.433) is
1
Dren (τ ) = sinh ωτ (2.437)
ω
[the analytically continued (2.216)]. Then

2[Ḋren (τ ) − 1] = 2(cosh βh̄ω − 1) = 4 sinh2 (βh̄ω/2), (2.438)

and we find the correct partition function:

n o−1/2
Zω = 2[Ḋren(τ ) − 1]
τ =h̄β
1
= . (2.439)
2 sinh(βh̄ω/2)

H. Kleinert, PATH INTEGRALS

2.12 Functional Measure in Fourier Space 151

On a sliced imaginary-time axis, the case of a constant frequency Ω2 ≡ ω 2 is solved as follows.

From Eq. (2.210) we take the ordinary Gelfand-Yaglom function DN , and continue it to Euclidean
ω̃e , yielding the imaginary-time version
sinh(N + 1)ω˜e ǫ
DN = . (2.440)
sinh ω˜e ǫ
+
Then we use formula (2.432), which simplifies for a constant Ω2 ≡ ω 2 for which DN −1 = DN −1 ,
and calculate
1
D̃N +1 = [sinh(N + 2)ω˜e ǫ − sinh N ω˜e ǫ] − 2
sinh ω˜e ǫ
= 2 [cosh(N + 1)ω˜e ǫ − 1] = 4 sinh2 [(N + 1)ω˜e ǫ/2]. (2.441)

Inserting this into Eq. (2.386) yields the partition function

1 1
Zω = q = , (2.442)
2 sinh(h̄ω˜e β/2)
D̃N +1

in agreement with (2.403).

2.12 Functional Measure in Fourier Space

There exists an alternative definition for the quantum-statistical path integral which
is useful for some applications (for example in Section 2.13 and in Chapter 5). The
limiting product formula (2.410) suggests that instead of summing over all zigzag
configurations of paths on a sliced time axis, a path integral may be defined with
the help of the Fourier components of the paths on a continuous time axis. As in
(2.380), but with a slightly different normalization of the coefficients, we expand
these paths here as
∞  
xm eiωm τ + c.c. , x−m ≡ x∗m . (2.443)
X
x(τ ) = x0 + η(τ ) ≡ x0 + x0 = real,
m=1

Note that the temporal integral over the time-dependent fluctuations η(τ ) is zero,
R h̄/kB T
0 dτ η(τ ) = 0, so that the zero-frequency component x0 is the temporal average
of the fluctuating paths:
kB T h̄/kB T
Z
x0 = x̄ ≡ dτ x(τ ). (2.444)
h̄ 0

In contrast to (2.380) which was valid on a sliced time axis and was therefore
subject to a restriction on the range of the m-sum, the present sum is unrestricted
and runs over all Matsubara frequencies ωm = 2πmkB T /h̄ = 2πm/h̄β. In terms of
xm , the Euclidean action of the linear oscillator is
M h̄/kB T
Z
Ae = dτ (ẋ2 + ω 2 x2 )
2 0"
∞
Mh̄ ω 2 2 X
#
2 2 2
= x + (ω + ω )|xm | . (2.445)
kB T 2 0 m=1 m
152 2 Path Integrals — Elementary Properties and Simple Solutions

The integration variables of the time-sliced path integral were transformed to the
Q R∞
Fourier components xm in Eq. (2.388). The product of integrals n −∞ dx(τn )
turned into the product (2.390) of integrals over real and imaginary parts of xm . In
the continuum limit, the result is
Z ∞ ∞ Z
Y ∞ Z ∞
dx0 d Re xm d Im xm . (2.446)
−∞ m=1 −∞ −∞

Placing the exponential e−Ae /h̄ with the frequency sum (2.445) into the integrand,
2
the product of Gaussian integrals renders a product of inverse eigenvalues (ωm +ω 2 )−1
for m = 1, . . . , ∞, with some infinite factor. This may be determined by comparison
with the known continuous result (2.410) for the harmonic partition function. The
infinity is of the type encountered in Eq. (2.183), and must be divided out of the
measure (2.446). The correct result (2.408) is obtained from the following measure
of integration in Fourier space
∞
"Z #
∞ dx0 Y ∞ ∞ d Re xm d Im xm
I Z Z
Dx ≡ 2
. (2.447)
−∞ le (h̄β) m=1 −∞ −∞ πkB T /Mωm
2
The divergences in the product over the factors (ωm + ω 2 )−1 discussed after
2
Eq. (2.183) are canceled by the factors ωm in the measure. It will be convenient to
introduce a short-hand notation for the measure on the right-hand side, writing it
as
∞ dx0
I Z I
Dx ≡ D ′ x. (2.448)
−∞ le (h̄β)

The denominator of the x0 -integral is the length scale le (h̄β) associated with β
defined in Eq. (2.353).
Then we calculate
∞ Z
" #
∞ d Re xm d Im xm −M h̄[ω2 x20 /2+P∞ (ωm
∞
I Z
m ] B
2 +ω 2 )|x |2 /k T
Zωx0 ≡ ′ −Ae /h̄
Y
D xe = 2
e m=1

m=1 −∞ −∞ πkB T /Mωm

∞
#−1
2
+ ω2
"
−M ω 2 x20 /2kB T
Y ωm
=e 2
. (2.449)
m=1 ωm

The final integral over the zero-frequency component x0 yields the partition function
∞
#−1
2
+ ω2
"
∞ dx0 x0 kB T Y ωm
I Z
−Ae /h̄
Zω = Dx e = Zω = 2
, (2.450)
−∞ le (h̄β) h̄ω m=1 ωm

as in (2.410).
The same measure can be used for the more general amplitude (2.414), as is
obvious from (2.416). With the predominance of the kinetic term in the measure
of path integrals [the divergencies discussed after (2.183) stem only from it], it can

H. Kleinert, PATH INTEGRALS

2.12 Functional Measure in Fourier Space 153

easily be shown that the same measure is applicable to any system with the standard
kinetic term.
It is also possible to find a Fourier decomposition of the paths and an associated
integration measure for the open-end partition function in Eq. (2.413). We begin by
considering the slightly reduced set of all paths satisfying the Neumann boundary
conditions
ẋ(τa ) = va = 0, ẋ(τb ) = vb = 0. (2.451)
They have the Fourier expansion
∞
X
x(τ ) = x0 + η(τ ) = x0 + xn cos νn (τ − τa ), νn = nπ/β. (2.452)
n=1

The frequencies νn are the Euclidean version of the frequencies (3.64) for Dirichlet
boundary conditions. Let us calculate the partition function for such paths by
analogy with the above periodic case by a Fourier decomposition of the action
∞
Mh̄ ω 2 2 1 X
" #
M h̄/kB T
Z
2 2 2
Ae = dτ (ẋ + ω x ) = x0 + (νn2 + ω 2)x2n , (2.453)
2 0 kB T 2 2 n=1

and of the measure

∞
"Z #
∞ dx0 Y ∞ ∞ d xn
I Z Z
Dx ≡
−∞ le (h̄β) n=1 −∞ −∞ πkB T /2Mνn2
∞ dx0
Z I
≡ D ′ x. (2.454)
−∞ le (h̄β)
We now perform the path integral over all fluctuations at fixed x0 as in (2.449):
∞ Z
" #
∞ ∞ d xn
I Z
−M h̄[ω 2 x20 /2+ n=1 (νn n ] B
P∞ 2 +ω 2 )|x |2 /k T
ZωN,x0 ≡ ′ −Ae /h̄
Y
D xe = 2
e
n=1 −∞ −∞ πkB T /2Mνn
∞
#−1
νn2 + ω 2
"
−M ω 2 x20 /2kB T
Y
=e . (2.455)
n=1 νn2

Using the product formula (2.183), this becomes

s
ωh̄β M
 
ZωN,x0 = exp −β ω 2 x20 . (2.456)
sinh ωh̄β 2
The final integral over the zero-frequency component x0 yields the partition function
s
1 2πh̄ 1
ZωN = √ . (2.457)
le (h̄β) Mω sinh ωh̄β
We have replaced the denominator in the prefactor 1/le (h̄β) by the length scale
1/le (h̄β) of Eq. (2.353). Apart from this prefactor, the Neumann partition function
coincides precisely with the open-end partition function Zωopen in Eq. (2.413).
154 2 Path Integrals — Elementary Properties and Simple Solutions

What is the reason for this coincidence up to a trivial factor, even though the
paths satisfying Neumann boundary conditions do not comprise all paths with open
ends? Moreover, the integrals over the endpoints in the defining equation (2.413) do
not force the endpoint velocities, but rather endpoint momenta to vanish. Indeed,
recalling Eq. (2.189) for the time evolution amplitude in momentum space we can
see immediately that the partition function with open ends Zωopen in Eq. (2.413) is
identical to the imaginary-time amplitude with vanishing endpoint momenta:

Zωopen = (pb h̄β|pa 0)|pb =pa =0 . (2.458)

Thus, the sum over all paths with arbitrary open ends is equal to the sum of all
paths satisfying Dirichlet boundary conditions in momentum space. Only classically,
the vanishing of the endpoint momenta implies the vanishing of the endpoint veloc-
ities. From the general discussion of the time-sliced path integral in phase space in
Section 2.1 we know that fluctuating paths have M ẋ 6= p. The fluctuations of the
difference are controlled by a Gaussian exponential of the type (2.53). This leads to
the explanation of the trivial factor between Zωopen and ZωN . The difference between
M ẋ and p appears only in the last short-time intervals at the ends. But at short
time, the potential does not influence the fluctuations in (2.53). This is the reason
why the fluctuations at the endpoints contribute only a trivial overall factor le (h̄β)
to the partition function ZωN .

2.13 Classical Limit

The alternative measure of the last section serves to show, somewhat more convinc-
ingly than before, that in the high-temperature limit the path integral representation
of any quantum-statistical partition function reduces to the classical partition func-
tion as stated in Eq. (2.346). We start out with the Lagrangian formulation (2.376).
Inserting the Fourier decomposition (2.443), the kinetic term becomes
∞
h̄β M 2 Mh̄ X
Z
dτ ẋ = ω 2 |xm |2 , (2.459)
0 2 kB T m=1 m

and the partition function reads

∞ ∞
" #
M X 1 h̄/kB T
I Z
2 2 ′
xm e−iωm τ ) . (2.460)
X
Z= Dx exp − ωm |xm | − dτ V (x0 +
kB T m=1 h̄ 0 m=−∞

The summation symbol with a prime implies the absence of the m = 0 -term. The
measure is the product (2.447) of integrals of all Fourier components.
We now observe that for large temperatures, the Matsubara frequencies for m 6= 0
diverge like 2πmkB T /h̄ . This has the consequence that the Boltzmann factor for
the xm6q=0 fluctuations becomes sharply peaked around xm= 0. The average size of
√
xm is kB T /M/ωm = h̄/2πm MkB T . If the potential V x0 + ′ ∞ ′
m=−∞ xm e
−iωm τ
P

is a smooth function of its arguments, we can approximate it by V (x0 ) plus terms

H. Kleinert, PATH INTEGRALS

2.14 Calculation Techniques on Sliced Time Axis via the Poisson Formula 155

containing higher powers of xm . For large temperatures, these are small on the
average and can be ignored. The leading term V (x0 ) is time-independent. Hence
we obtain in the high-temperature limit
∞
" #
T →∞ M X 1
I
2
Z−
−−→ Dx exp − ωm |xm |2 − V (x0 ) . (2.461)
kB T m=1 kB T

The right-hand side is quadratic in the Fourier components xm . With the measure
of integration (2.447), we perform the integrals over xm and obtain
T →∞ ∞ dx0 −V (x0 )/kB T
Z
Z−
−−→ Zcl = e . (2.462)
−∞ le (h̄β)

This agrees with the classical statistical partition function (2.352).

The derivation reveals an important prerequisite for the validity of the classical
limit: It holds only for sufficiently smooth potentials. We shall see in Chapter 8
that for singular potentials such as −1/|x| (Coulomb), 1/|x|2 (centrifugal barrier),
1/ sin2 θ (angular barrier), this condition is not fulfilled and the classical limit is no
longer given by (2.462). The particle distribution ρ(x) at a fixed x does not have
this problem. It always tends towards the naively expected classical limit (2.354):
T →∞
−−→ Zcl−1 e−V (x)/kB T .
ρ(x) − (2.463)

The convergence is nonuniform in x, which is the reason why the limit does not
always carry over to the integral (2.462). This will be an important point in deriving
in Chapter 12 a new path integral formula valid for singular potentials. At first, we
shall ignore such subtleties and continue with the conventional discussion valid for
smooth potentials.

2.14 Calculation Techniques on Sliced Time Axis

via the Poisson Formula
In the previous sections we have used tabulated product formulas such as (2.127), (2.165), (2.173),
(2.400), (2.402) to find fluctuation determinants on a finite sliced time axis. With the recent
interest in lattice models of quantum field theories, it is useful to possess an efficient calculational
technique to derive such product formulas (and related sums). Consider, as a typical example, the
quantum-statistical partition function for a harmonic oscillator of frequency ω on a time axis with
N + 1 slices of thickness ǫ,
N
Y
Z= [2(1 − cos ωm ǫ) + ǫ2 ω 2 ]−1/2 , (2.464)
m=0

with the product running over all Matsubara frequencies ωm = 2πmkB T /h̄. Instead of dealing
with this product it is advantageous to consider the free energy
N
1 X
F = −kB T log Z = kB T log[2(1 − cos ωm ǫ) + ǫ2 ω 2 ]. (2.465)
2 m=0
156 2 Path Integrals — Elementary Properties and Simple Solutions

We now observe that by virtue of Poisson’s summation formula (1.205), the sum can be rewritten
as the following combination of a sum and an integral:
∞ Z 2π
1 X dλ iλn(N +1)
F = kB T (N + 1) e log[2(1 − cos λ) + ǫ2 ω 2 ]. (2.466)
2 n=−∞ 0
2π

The sum over n squeezes λ to integer multiples of 2π/(N + 1) = ωm ǫ which is precisely what we
want.
We now calculate the integrals in (2.466):
Z 2π
dλ iλn(N +1)
e log[2(1 − cos λ) + ǫ2 ω 2 ]. (2.467)
0 2π
For this we rewrite the logarithm of an arbitrary positive argument as the limit
 Z ∞ 
dτ −τ a/2
log a = lim − e + log(2δ) + γ, (2.468)
δ→0 δ τ
where
N
!
′
X 1
γ ≡ −Γ (1)/Γ(1) = lim − log N ≈ 0.5773156649 . . . (2.469)
N →∞
n=1
n
is the Euler-Mascheroni constant. Indeed, the function
Z ∞
dt −t
E1 (x) = e (2.470)
x t
is known as the exponential integral with the small-x expansion19
∞
X (−x)k
E1 (x) = −γ − log x − . (2.471)
kk!
k=1

With the representation (2.468) for the logarithm, the free energy can be rewritten as
∞  Z ∞
dτ 2π dλ iλn(N +1)−τ [2(1−cos λ)+ǫ2 ω2 ]/2

1 X
Z
F = lim − e − δn0 [log(2δ) + γ] .
2ǫ n=−∞ δ→0 δ τ 0 2π
(2.472)

The integral over λ is now performed20 giving rise to a modified Bessel function In(N +1) (τ ):
∞  Z ∞ 
1 X dτ 2 2
F = lim − In(N +1) (τ )e−τ (2+ǫ ω )/2 − δn0 [log(2δ) + γ] .
2ǫ n=−∞ δ→0 δ τ
(2.473)

If we differentiate this with respect to ǫ2 ω 2 ≡ m2 , we obtain

∞ Z ∞
∂F 1 X 2
= dτ In(N +1) (τ )e−τ (2+m )/2 (2.474)
∂m2 4ǫ n=−∞ 0

and perform the τ -integral, using the formula valid for Re ν > −1, Re α > Re µ
Z ∞ p p
−τ α ν (α − α2 − µ2 )−ν , −ν (α − α2 − µ2 )ν
dτ Iν (µτ )e =µ p =µ p , (2.475)
0 α2 − µ2 α2 − µ2
19
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 8.214.2.
20
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formulas 8.411.1 and 8.406.1.

H. Kleinert, PATH INTEGRALS

2.14 Calculation Techniques on Sliced Time Axis via the Poisson Formula 157

to find
∞
" p #|n|(N +1)
∂F 1 X 1 m2 + 2 − (m2 + 2)2 − 4
= . (2.476)
∂m2
p
2ǫ n=−∞ (m2 + 2)2 − 4 2

From this we obtain F by integration over m2 + 1. The n = 0 -term under the sum gives
p
log[(m2 + 2 + (m2 + 2)2 − 4 )/2] + const (2.477)
and the n 6= 0 -terms:
1 p
− [(m2 + 2 + (m2 + 2)2 − 4 )/2]−|n|(N +1) + const , (2.478)
|n|(N + 1)
where the constants of integration can depend on n(N + 1). They are adjusted by going to the
limit m2 → ∞ in (2.473). There the integral is dominated by the small-τ regime of the Bessel
functions
1  z α
Iα (z) ∼ [1 + O(z 2 )], (2.479)
|α|! 2
and the first term in (2.473) becomes
Z ∞
1 dτ  τ |n|(N +1) −τ m2 /2
− e
(|n|(N + 1))! δ τ 2
log m2 + γ + log(2δ)
 
n=0
≈ . (2.480)
−(m2 )−|n|(N +1) /|n|(N + 1) n =
6 0
The limit m2 → ∞ in (2.477), (2.478) gives, on the other hand, log m2 + const and
−(m2 )−|n|(N +1) /|n|(N + 1) + const , respectively. Hence the constants of integration must be
zero. We can therefore write down the free energy for N + 1 time steps as
N
1 X
F = log[2(1 − cos(ωm ǫ)) + ǫ2 ω 2 ]
2β m=0
(
1 h p  i
= log ǫ2 ω 2 + 2 + (ǫ2 ω 2 + 2)2 − 4 2 (2.481)
2ǫ
∞
)
2 X 1 h 2 2 p   i−|n|(N +1)
− ǫ ω + 2 + (ǫ2 ω 2 + 2)2 − 4 2 .
N + 1 n=1 n
Here it is convenient to introduce the parameter
nh p i o
ǫω̃e ≡ log ǫ2 ω 2 + 2 + (ǫ2 ω 2 + 2)2 − 4 2 , (2.482)
which satisfies
p
cosh(ǫω̃e ) = (ǫ2 ω 2 + 2)/2, sinh(ǫω̃e ) = (ǫ2 ω 2 + 2)2 − 4/2, (2.483)
or
sinh(ǫω̃e /2) = ǫω/2.
Thus it coincides with the parameter introduced in (2.399), which brings the free energy (2.481)
to the simple form
∞
" #
h̄ 2 X 1 −ǫω̃e n(N +1)
F = ω̃e − e
2 ǫ(N + 1) n=1 n
1
h̄ω̃e + 2kB T log(1 − e−βh̄ω̃e )

=
2
1
= log [2 sinh(βh̄ω̃e /2)] , (2.484)
β
158 2 Path Integrals — Elementary Properties and Simple Solutions

whose continuum limit is

ǫ→0 1 h̄ω 1
F = log [2 sinh (βh̄ω/2)] = + log(1 − e−βh̄ω ). (2.485)
β 2 β

2.15 Field-Theoretic Definition of Harmonic Path Integrals

by Analytic Regularization
A slight modification of the calculational techniques developed in the last section
for the quantum partition function of a harmonic oscillator can be used to define
the harmonic path integral in a way which neither requires time slicing, as in the
original Feynman expression (2.66), nor a precise specification of the integration
measure in terms of Fourier components, as in Section 2.12. The path integral for
the partition function
I R h̄β I R h̄β
M [ẋ2 (τ )+ω 2 x2 (τ )]/2 M x(τ )[−∂τ2 +ω 2 ]x(τ )/2
Zω = Dxe− 0 = Dxe− 0 (2.486)

is formally evaluated as
1 1 2 2
Zω = q = e− 2 Tr log(−∂τ +ω ) . (2.487)
Det(−∂τ2 + ω 2 )

Since the determinant of an operator is the product of all its eigenvalues, we may
write, again formally,
Y 1
Zω = √ 2 . (2.488)
ω′ ω + ω2
′

The product runs over an infinite set of quantities which grow with ω ′2, thus being
certainly divergent. It may be turned into a divergent sum by rewriting Zω as
1
log(ω ′ 2 +ω 2 )
P
Zω ≡ e−Fω /kB T = e− 2 ω′ . (2.489)

This expression has two unsatisfactory features. First, it requires a proper definition
of the formal sum over a continuous set of frequencies. Second, the logarithm of
2
the dimensionful arguments ωm + ω 2 must be turned into a meaningful expression.
The latter problem would be removed if we were able to exchange the logarithm
by log[(ω ′ 2 + ω 2 )/ω 2]. This would require the formal sum ω′ log ω 2 to vanish. We
P

shall see below in Eq. (2.514) that this is indeed one of the pleasant properties of
analytic regularization.
At finite temperatures, the periodic boundary conditions along the imaginary-
time axis make the frequencies ω ′ in the spectrum of the differential operator −∂τ2 +ω 2
discrete, and the sum in the exponent of (2.489) becomes a sum over all Matsubara
frequencies ωm = 2πkB T /h̄ (m = 0, ±1, ±2, . . .):
∞
" #
1 X 2
Zω = exp − log(ωm + ω2) . (2.490)
2 m=−∞

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 159

For the free energy Fω ≡ (1/β) log Zω , this implies

∞
1 2 2 1 X 2
Fω = Tr log(−∂τ + ω ) = log(ωm + ω 2 ). (2.491)
2β per 2β m=−∞

where the subscript per emphasizes the periodic boundary conditions in the τ -
interval (0, h̄β).

2.15.1 Zero-Temperature Evaluation of the Frequency Sum

In the limit T → 0, the sum in (2.491) goes over into an integral, and the free energy
becomes
1 h̄ ∞ dω ′
Z
Fω ≡ Tr log(−∂τ2 + ω 2 ) = log(ω ′2 + ω 2 ), (2.492)
2β ±∞ 2 −∞ 2π
where the subscript ±∞ indicates the vanishing boundary conditions of the eigen-
functions at τ = ±∞. Thus, at low temperature, we can replace the frequency sum
in the exponent of (2.489) by
∞ dω ′
X Z
−
−−→ h̄β . (2.493)
ω′
T →0 −∞ 2π

This could have been expected on the basis of Planck’s rules for the phase space
invoked earlier on p. 97 to explain the measure of path integration. According to
theseRrules, the volume
R
element in the phase space of energy and time has the mea-
sure dt dE/h = dt dω/2π. If the integrand is independent of time, the temporal
integral produces an overall factor , which for the imaginary-time interval (0, h̄β) of
statistical mechanics is equal to h̄β = h̄/kB T , thus explaining the integral version
of the sum (2.493).
The integral on the right-hand side of (2.492) diverges at large ω ′ . This is called
an ultraviolet divergence (UV-divergence), alluding to the fact that the ultraviolet
regime of light waves contains the high frequencies of the spectrum.
The important observation is now that the divergent integral (2.492) can be
made finite by a mathematical technique called analytic regularization.21 This is
based on rewriting the logarithm log(ω ′2 + ω 2) in the derivative form:

d ′2
log(ω ′2 + ω 2 ) = − (ω + ω 2 )−ǫ . (2.494)
dǫ ǫ=0

Equivalently, we may obtain the logarithm from an ǫ → 0 -limit of the function

1 1
lMS (ǫ) = − (ω ′2 + ω 2 )−ǫ + . (2.495)
ǫ ǫ
21
G. ’t Hooft and M. Veltman, Nucl. Phys. B 44 , 189 (1972). Analytic regularization is at
present the only method that allows to renormalize nonabelian gauge theories without destroying
gauge invariance. See also the review by G. Leibbrandt, Rev. Mod. Phys. 74 , 843 (1975).
160 2 Path Integrals — Elementary Properties and Simple Solutions

The subtraction of the pole term 1/ǫ is commonly referred to a minimal subtraction.
Indicating this process by a subscript MS, we may write
1
lMS (ǫ) = − (ω ′2 + ω 2 )−ǫ . (2.496)
ǫ MS, ǫ→0

Using the derivative formula (2.494), the trace of the logarithm in the free energy
(2.492) takes the form

1 d ∞ dω ′ ′2
Z
Tr log(−∂τ2 + ω 2 ) = − (ω + ω 2 )−ǫ . (2.497)
h̄β dǫ −∞ 2π ǫ=0

We now set up a useful integral representation, due to Schwinger, for a power a−ǫ
generalizing (2.468). Using the defining integral representation for the Gamma func-
tion Z ∞
dτ µ −τ ω2
τ e = ω −µ/2 Γ(µ), (2.498)
0 τ
the desired generalization is
1 ∞ dτ ǫ −τ a
Z
−ǫ
a = τ e . (2.499)
Γ(ǫ) 0 τ
This allows us to re-express (2.497) as

1 2 2 d 1 Z ∞ dω ′ Z ∞ dτ ǫ −τ (ω′ 2 +ω2 )
Tr log(−∂τ + ω ) = − τ e . (2.500)
h̄β dǫ Γ(ǫ) −∞ 2π 0 τ ǫ=0

As long as ǫ is larger than zero, the τ -integral converges absolutely, so that we can
interchange the τ - and ω ′ -integrations, and obtain
1 d 1 Z ∞ dτ ǫ Z ∞ dω ′ −τ (ω′ 2 +ω2 )
Tr log(−∂τ2 + ω 2 ) = − τ e . (2.501)
h̄β dǫ Γ(ǫ) 0 τ −∞ 2π ǫ=0

At this point we can perform the Gaussian integral over ω ′ using formula (1.338),
and find
1 2 2 d 1 Z ∞ dτ ǫ 1 2
Tr log(−∂τ + ω ) = − τ √ e−τ ω . (2.502)
h̄β dǫ Γ(ǫ) 0 τ 2 τ π ǫ=0

For small ǫ, the τ -integral is divergent at the origin. It can, however, be defined by
an analytic continuation of the integral starting from the regime ǫ > 1/2, where it
converges absolutely, to ǫ = 0. The continuation must avoid the pole at ǫ = 1/2.
Fortunately, this continuation is trivial since the integral can be expressed in terms of
the Gamma function, whose analytic properties are well-known. Using the integral
formula (2.498), we obtain

1 1 d 1
Tr log(−∂τ2 + ω 2 ) = − √ ω 1−2ǫ Γ(ǫ − 1/2) . (2.503)
h̄β 2 π dǫ Γ(ǫ) ǫ=0

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 161

The right-hand side has to be continued analytically from ǫ > 1/2 to ǫ = 0. This is
easily done using the defining property of the Gamma
√ function Γ(x) = Γ(1 + x)/x,
from which we find Γ(−1/2) = −2Γ(1/2) = −2 π, and 1/Γ(ǫ) ≈ ǫ/Γ(1 + ǫ) ≈ ǫ.
The derivative with respect to ǫ leads to the free energy of the harmonic oscillator
at low temperature via analytic regularization:
1 ∞ dω ′
Z
Tr log(−∂τ2 + ω 2) = log(ω ′2 + ω 2) = ω, (2.504)
h̄β −∞ 2π
so that the free energy of the oscillator at zero-temperature becomes
h̄ω
Fω = . (2.505)
2
This agrees precisely with the result obtained from the lattice definition of the path
integral in Eq. (2.407), or from the path integral (3.808) with the Fourier measure
(2.447).
With the above procedure in mind, we shall often use the sloppy formula ex-
pressing the derivative of Eq. (2.499) at ǫ = 0:
∞ dτ −τ a
Z
log a = − e . (2.506)
0 τ
This formula differs from the correct one by a minimal subtraction and can be
used in all calculations with analytic regularization. Its applicability is based on
the possibility of dropping the frequency integral over 1/ǫ in the alternative correct
expression
1 1 Z ∞ dω ′ 1 ′2 1
 
2 2
Tr log(−∂τ + ω ) = − (ω + ω 2 )−ǫ − . (2.507)
h̄β ǫ −∞ 2π ǫ ǫ ǫ→0

In fact, within analytic regularization one may set all integrals over arbitrary pure
powers of the frequency equal to zero:
Z ∞
dω ′ (ω ′)α = 0 for all α. (2.508)
0

This is known as Veltman’s rule.22 It is a special limit of a frequency integral which

is a generalization of the integral in (2.497):
∞ dω ′ (ω ′2)γ Γ(γ + 1/2) 2 γ+1/2−ǫ
Z
′2 2 ǫ
= (ω ) . (2.509)
−∞ 2π (ω + ω ) 2πΓ(ǫ)
This equation may be derived by rewriting the left-hand side as
1 ∞ dω ′ ′2 γ ∞ dτ ǫ −τ (ω′ 2 +ω2 )
Z Z
(ω ) τ e . (2.510)
Γ(ǫ) −∞ 2π 0 τ
22
See the textbook H. Kleinert and V. Schulte-Frohlinde, Critical Properties of φ4 -Theories,
World Scientific, Singapore, 2001 (http://www.physik.fu-berlin.de/~kleinert/b8).
162 2 Path Integrals — Elementary Properties and Simple Solutions

The integral over ω ′ is performed as follows:

∞ dω ′ ′2 γ −τ (ω′ 2 +ω2 ) 1 ∞ dω ′ 2 ′2 γ+1/2 −τ (ω′ 2 +ω2 ) τ −γ−1/2

Z Z
(ω ) e = (ω ) e = Γ(γ + 1/2),
−∞ 2π 2π 0 ω′ 2 2π
(2.511)
leading to a τ -integral in (2.510)
Z ∞ dτ ǫ−γ−1/2 −τ ω2
τ e = (ω 2 )γ+1/2+ǫ , (2.512)
0 τ

and thus to the formula (2.509). The Veltman rule (2.508) follows from this directly
in the limit ǫ → 0, since 1/Γ(ǫ) → 0 on the right-hand side. This implies that the
subtracted 1/ǫ term in (2.507) gives no contribution.
The vanishing of all integrals over pure powers by Veltman’s rule (2.508) was
initially postulated in the process of developing a finite quantum field theory of
weak and electromagnetic interactions. It has turned out to be extremely useful
for the calculation of critical exponents of second-order phase transitions from field
theories.22
An important consequence of Veltman’s rule is to make the logarithms of di-
mensionful arguments in the partition functions (2.489) and the free energy (2.491)
meaningful quantities. First, since d(ω ′/2π) log ω 2 = 0, we can divide the argu-
R

ment of the logarithm in (2.492) by ω 2 without harm, and make them dimensionless.
At finite temperatures, we use the equality of the sum and the integral over an ωm -
independent quantity c
∞ ∞ dωm
X Z
kB T c= c (2.513)
m=−∞ −∞ 2π

to show that also

∞
log ω 2 = 0,
X
kB T (2.514)
m=−∞

so that we have, as a consequence of Veltman’s rule, that the Matsubara frequency

sum over the constant log ω 2 vanishes for all temperatures. For this reason, also the
argument of the logarithm in the free energy (2.491) can be divided by ω 2 without
change, thus becoming dimensionless.

2.15.2 Finite-Temperature Evaluation of the Frequency Sum

At finite temperature, the free energy contains an additional term consisting of the
difference between the Matsubara sum and the frequency integral
∞ 2 2
! !
kB T ωm h̄ ∞ dωm ωm
X Z
∆Fω = log 2
+1 − log +1 , (2.515)
2 m=−∞ ω 2 −∞ 2π ω2

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 163

where we have used dimensionless logarithms as discussed at the end of the last
subsection. The sum is conveniently split into a subtracted, manifestly convergent
expression
∞ 2 2 ∞
ω2
" ! # !
X ωm ωm X
∆1 Fω = kB T log + 1 − log = k B T log 1 + , (2.516)
m=1 ω2 ω2 m=1
2
ωm
and a divergent sum
∞ 2
X ωm
∆2 Fω = kB T log . (2.517)
m=1 ω2
The convergent part is most easily evaluated. Taking the logarithm of the product
in Eq. (2.408) and recalling (2.409), we find
∞
ω2
!
Y sinh(βh̄ω/2)
1+ 2 = , (2.518)
m=1 ωm βh̄ω/2
and therefore
1 sinh(βh̄ω/2)
∆F1 = log . (2.519)
β βh̄ω/2
The divergent sum (2.517) is calculated by analytic regularization as follows: We
rewrite
 !−ǫ 
∞ 2 ∞  ∞
"  #
ωm d ωm −ǫ d 2π
m−ǫ 
X X X
log =− 2 = − 2 , (2.520)
m=1 ω2 dǫ m=1 ω ǫ→0
dǫ βh̄ω m=1 ǫ→0

and express the sum over m−ǫ in terms of Riemann’s zeta function
∞
m−z .
X
ζ(z) = (2.521)
m=1

This sum is well defined for z > 1, and can be continued analytically into the entire
complex z-plane. The only singularity of this function lies at z = 1, where in the
neighborhood ζ(z) ≈ 1/z. At the origin, ζ(z) is regular, and satisfies23
1
ζ(0) = −1/2, ζ ′ (0) = − log 2π, (2.522)
2
such that we may approximate
1
ζ(z) ≈ − (2π)z , z ≈ 0. (2.523)
2
Hence we find
 !−ǫ 
∞
ω2 d 2π d
log m2 = − 2 (βh̄ω)ǫ
X
ζ(ǫ) = = log h̄ωβ. (2.524)
m=1 ω dǫ βh̄ω dǫ ǫ→0
ǫ→0
23
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 9.541.4.
164 2 Path Integrals — Elementary Properties and Simple Solutions

thus determining ∆2 Fω in Eq. (2.517).

By combining this with (2.519) and the contribution −h̄ω/2 from the integral
(2.515), the finite-temperature part (2.491) of the free energy becomes
1
∆Fω = log(1 − e−h̄βω ). (2.525)
β

Together with the zero-temperature free energy (2.505), this yields the dimensionally
regularized sum formula
∞
1 1 X h̄ω 1
Fω = Tr log(−∂τ2 + ω 2 ) = 2
log(ωm + ω 2) = + log(1 − e−h̄ω/kB T )
2β 2β m=−∞ 2 β
!
1 h̄ωβ
= log 2 sinh , (2.526)
β 2

in agreement with the properly normalized free energy (2.485) at all temperatures.
Note that the property of the zeta function ζ(0) = −1/2 in Eq. (2.522) leads once
more to our earlier result (2.514) that the Matsubara sum of a constant c vanishes:
∞
X −1
X ∞
X
c= c+c+ c = 0, (2.527)
m=−∞ m=−∞ m=1

since
∞
X −∞
X
1= 1 = ζ(0) = −1/2. (2.528)
m=1 m=−1

As mentioned before, this allows us to divide ω 2 out of the logarithms in the sum
in Eq. (2.526) and rewrite this sum as
∞ 2 ∞ 2
! !
1 X ωm 1 X ωm
log + 1 = log +1 (2.529)
2β m=−∞ ω2 β m=1 ω2

2.15.3 Quantum-Mechanical Harmonic Oscillator

This observation leads us directly to the analogous quantum-mechanical discussion.
Starting from the fluctuation factor (2.91) of the free particle which can formally be
written as
i tb M 1
Z  Z 
F0 (∆t) = Dδx(t) exp dt δx(−∂t2 )δx = q , (2.530)
h̄ ta 2 2πh̄i∆t/M

where ∆t ≡ tb − ta [recall (2.130)]. The path integral of the harmonic oscillator has
the fluctuation factor [compare (2.188)]
#−1/2
Det(−∂t2 − ω 2 )
"
Fω (∆t) = F0 (∆t) . (2.531)
Det(−∂t2 )

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 165

The ratio of the determinants has the Fourier decomposition

∞ h
Det(−∂t2 − ω 2 )
( )
i
2 2 2
X
= exp log(νn − ω ) − log νn , (2.532)
Det(−∂t2 ) n=1

where νn = nπ/∆t [recall (2.115)], and was calculated in Eq. (2.188) to be

Det(−∂t2 − ω 2) sin ω∆t

2
= . (2.533)
Det(−∂t ) ω∆t

This result can be reproduced with the help of formulas (2.529) and (2.526). We
raplace β by 2∆t, and use again Σn 1 = ζ(0) = −1/2 to obtain
∞ ∞
νn2
" ! #
Det(−∂t2 2
log(νn2 2
+ 1 + log ω 2
X X
−ω ) = +ω ) = log 2
n=1
ω→iω
n=1 ω ω→iω
" ∞
ν2
! #
1 sin ω∆t
 
log n2 + 1 − log ω 2
X
= = log 2 . (2.534)
n=1 ω 2 ω→iω
ω

For ω = 0 this reproduces Formula (2.524). Inserting this and (2.534) into (2.532),
we recover the result (2.533). Thus we find the amplitude

1 1 ω
r
−1/2
(xb tb |xa ta ) = q Det (−∂t2 2
− ω )e iAcl /h̄
=q eiAcl /h̄ , (2.535)
πi/M πi/M 2 sin ω∆t

in agreement with (2.175).

2.15.4 Tracelog of the First-Order Differential Operator

The trace of the logarithm in the free energy (2.492) can obviously be split into two
terms
Tr log(−∂τ2 + ω 2 ) = Tr log(∂τ + ω) + Tr log(−∂τ + ω). (2.536)
Since the left-hand side is equal to βh̄ω by (2.504), and the two integrals must be
the same, we obtain the low-temperature result

dω∞ ∞ dω
Z Z
Tr log(∂τ + ω) = Tr log(−∂τ + ω) = h̄β log(−iω ′ + ω) = h̄β log(iω ′ + ω)
−∞ 2π −∞ 2π
h̄βω
= . (2.537)
2
The same result could be obtained from analytic continuation of the integrals over
∂ǫ (±iω ′ + ω)ǫ to ǫ = 0.
For a finite temperature, we may use Eq. (2.526) to find
!
1 βh̄ω
Tr log(∂τ + ω) = Tr log(−∂τ + ω) = Tr log(−∂τ2 + ω 2 ) = log 2 sinh , (2.538)
2 2
166 2 Path Integrals — Elementary Properties and Simple Solutions

which reduces to (2.537) for T → 0.

The result is also the same if there is an extra factor i in the argument of
the tracelog. To see this we consider the case of time-independent frequency where
Veltman’s rule (2.508) tells us that it does not matter whether one evaluates integrals
over log(iω ′ ∓ ω) or over log(ω ′ ± iω).
Let us also replace ω ′ by iω ′ in the zero-temperature tracelog (2.504) of the
second-order differential operator (−∂τ2 + ω 2 ). Then we rotate the contour of inte-
gration clockwise in the complex plane to find
∞ dω ′
Z
log(−ω ′2 + ω 2 − iη) = ω, ω ≥ 0, (2.539)
−∞ 2π

where an infinitesimal positive η prescribes how to bypass the singularities at ω ′ =

±ω ∓ iη along the rotated contour of integration. Recall the discussion of the iη-
prescription in Section 3.3. The integral (2.539) can be split into the integrals
∞ dω ′ ω
Z
log[ω ′ ± (ω − iη)] = i , ω ≥ 0. (2.540)
−∞ 2π 2

Hence formula (2.537) can be generalized to arbitrary complex frequencies ω =

ωR + iωI as follows:
dω ′ ω
Z ∞
log(ω ′ ± iω) = ∓ǫ(ωR ) , (2.541)
−∞ 2π 2
and
∞ dω ′ ω
Z
log(ω ′ ± ω) = −iǫ(ωI ) , (2.542)
−∞ 2π 2
where ǫ(x) = Θ(x)−Θ(−x) = x/|x| is the antisymmetric Heaviside function (1.315),
which yields the sign of its argument. The formulas (2.541) and (2.542) are the
large-time limit of the more complicated sums
∞
" #
kB T X kB T h̄ω
log(ωm ± iω) = log 2ǫ(ωR ) sinh , (2.543)
h̄ m=−∞ h̄ 2kB T

and
∞
" #
kB T X kB T h̄ω
log(ωm ± ω) = log −2iǫ(ωI ) sin . (2.544)
h̄ m=−∞ h̄ 2kB T

The first expression is periodic in the imaginary part of ω, with period 2πkB T , the
second in the real part. The determinants possess a meaningful large-time limit only
if the periodic parts of ω vanish. In many applications, however, the fluctuations will
involve sums of logarithms (2.544) and (2.543) with different complex frequencies
ω, and only the sum of the imaginary or real parts will have to vanish to obtain a
meaningful large-time limit. On these occasions we may use the simplified formulas
(2.541) and (2.542). Important examples will be encountered in Section 18.9.2.

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 167

In Subsection 3.3.2, Formula (2.538) will be generalized to arbitrary positive

time-dependent frequencies Ω(τ ), where it reads [see (3.133)]
( " #)
1 h̄β ′′
Z
Tr log [±∂τ + Ω(τ )] = log 2 sinh dτ Ω(τ ′′ )
2 0
1 Z h̄β ′′
 R h̄β ′′ 
′′
= dτ Ω(τ ′′ ) + log 1 − e− 0 dτ Ω(τ ) . (2.545)
2 0

2.15.5 Gradient Expansion of the One-Dimensional Tracelog

Formula (2.545) may be used to calculate the trace of the logarithm of a second-
order differential equation with arbitrary frequency as a semiclassical expansion. We
introduce the Planck constant h̄ and the potential w(τ ) ≡ h̄Ω(τ ), and factorize as
in (2.536):
h i
Det −h̄2 ∂τ2 + w 2 (τ ) = Det [−h̄∂τ − w̄(τ )] × Det [h̄∂τ − w̄(τ )] , (2.546)

where the function w̄(τ ) satisfies the Riccati differential equation:24

h̄∂τ w̄(τ ) + w̄ 2 (τ ) = w 2 (τ ). (2.547)

By solving this we obtain the trace of the logarithm from (2.545):

( " #)
1 h̄β
h i Z
Tr log −h̄2 ∂τ2 2
+ w (τ ) = log 4 sinh 2 ′
dτ w̄(τ ) ′
. (2.548)
2h̄ 0

The exponential of this yields the functional determinant. For constant w̄(τ ) =
ω this agrees with the result (2.433) of the Gelfand-Yaglom formula for periodic
boundary conditions.
This agreement is no coincidence. We can find the solution of any Riccati differ-
ential equation if we know how to solve the second-order differential equation (2.433).
Imposing the Gelfand-Yaglom boundary conditions in (2.433), we find Dren (τ ) and
from this the functional determinant 2[Ḋren (h̄β)−1]. Comparison with (2.548) shows
that the solution of the Riccati differential equation (2.547) is given by
q
w̄(τ ) = 2h̄∂τ arsinh [Ḋren (τ ) − 1]/2. (2.549)

For the harmonic oscillator where Ḋren (τ ) is equal to (2.437), this leads to the
constant w̄(τ ) = h̄ω, as it should.
If we cannot solve the second-order differential equation (2.433), a solution to
the Riccati equation (2.547) can still be found as a power series in h̄:
∞
w̄n (τ )h̄n ,
X
w̄(τ ) = (2.550)
n=0

24
Recall the general form of the Riccati differential equation y ′ = f (τ )y + g(τ )y 2 + h(y), which
is an inhomogeneous version of the Bernoulli differential equation y ′ = f (τ )y + g(τ )y n for n = 2.
168 2 Path Integrals — Elementary Properties and Simple Solutions

which provides us with a so-called gradient expansion of the trace of the logarithm.
The lowest-order coefficient function w̄0 (τ ) is obviously equal to w(τ ). The higher
ones obey the recursion relation
n−1
!
1
˙
X
w̄n (τ ) = − w̄n−1 (τ ) + w̄n−k (τ )w̄k (τ ) , n ≥ 1. (2.551)
2w(τ ) k=1

These are solved for n = 0, 1, 2, 3 by

(
2 3
p v ′ (τ ) 5 v ′ (τ ) v ′′ (τ ) 15 v ′ (τ ) 9 v ′ (τ ) v ′′ (τ ) v (3) (τ )
v(τ ), − , − 5/2
+ 3/2
, − 4 + 3 − 2,
4 v(τ ) 32 v(τ ) 8 v(τ ) 64 v(τ ) 32 v(τ ) 16 v(τ )
)
4 2 2
1105 v ′(τ ) 221 v ′ (τ ) v ′′ (τ ) 19 v ′′ (τ ) 7 v ′ (τ ) v (3) (τ ) v (4) (τ )
− 11/2
+ 9/2
− 7/2
− 7/2
+ 5/2
,(2.552)
2048 v(τ ) 256 v(τ ) 128 v(τ ) 32 v(τ ) 32 v(τ )

where v(τ ) ≡ w 2 (τ ). The series can, of course, be trivially extended to any desired
orders.

2.15.6 Duality Transformation and Low-Temperature Expansion

There exists another method of calculating the finite-temperature part of the free
energy (2.491) which is worth presenting at this place, due to its broad applicability
in statistical mechanics. For this we rewrite (2.515) in the form
∞
!
kB T h̄ Z ∞ dωm 2
+ ω 2 ).
X
∆Fω = − log(ωm (2.553)
2 m=−∞ kB T −∞ 2π

Changing the integration variable to m, this becomes

 !2 
∞
!
kB T 2πkB T
∞
Z
m2 + ω 2  .
X
∆Fω = − dm log  (2.554)
2 m=−∞ −∞ h̄

Within analytic regularization, this expression is rewritten with the help of formula
(2.506) as
∞
!
kB T ∞ dτ ∞
Z Z
dm e−τ [(2πkB T /h̄) ].
X 2 m2 +ω 2
∆Fω = − − (2.555)
2 0 τ m=−∞ −∞

The duality transformation proceeds by performing the sum over the Matsubara
R
frequencies with the help of Poisson’s formula (1.205) as an integral dµ plus by an
extra sum over integer numbers n. This brings (2.555) to the form (expressing the
temperature in terms of β),
∞
!
1 ∞ dτ ∞
Z Z
e2πµni − 1 e−τ [(2π/h̄β) ].
X 2 µ2 +ω 2
∆Fω = − dµ (2.556)
2β 0 τ −∞ n=−∞

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 169

P∞
The parentheses contain the sum 2 n=1 e2πµni . After a quadratic completion of the
exponent
!2 !2 " #2
2π 2π nh̄2 β 2 1
2πµni − τ 2
µ = −τ µ−i − (h̄βn)2 , (2.557)
h̄β h̄β 4πτ 4τ
the integral over µ can be performed, with the result
∞
h̄ ∞ dτ −1/2 X
Z
2 2
∆Fω = − √ τ e−(nh̄β) /4τ −τ ω . (2.558)
2 π 0 τ n=1

Now we may use the integral formula [compare (1.347)]25

 ν
∞ dτ ν −a2 /τ −b2 τ a
Z
τ e =2 Kν (2ab), Kν (2ab) = K−ν (2ab), (2.559)
0 τ b
to obtain the sum over modified Bessel functions
∞
h̄ω X √
∆Fω = − √ 2 (nβh̄ω)−1/2 2K1/2 (nβh̄ω). (2.560)
2 π n=1

The modified Bessel functions with index 1/2 are particularly simple:
π −z
r
K1/2 (z) = e . (2.561)
2z
Inserting this into (2.560), the sum is a simple geometric one, and may be performed
as follows:
1 X 1 −βh̄ωn 1  
∆Fω = − e = log 1 − e−βh̄ω , (2.562)
β n=1 n β
in agreement with the previous result (2.525).
The effect of the duality tranformation may be rephrased in another way. It
converts the sum over Matsubara frequencies ωm in (2.516):
∞
ω2
!
X
S(βh̄ω) = kB T log 1 + 2 (2.563)
m=1 ωm

into a sum over the quantum numbers n of the harmonic oscillator:

∞
βh̄ω X 1 −nβh̄ω
S(βh̄ω) = − log βh̄ω − e . (2.564)
2 n=1 n

The sum (2.563) converges fast at high temperatures, where it can be expanded in
powers of ω 2 :
∞ ∞
! !2 k
(−1)k
X X1 βh̄ω
S(βh̄ω) = − 2k
  . (2.565)
k=1
k m=1 m 2π
25
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formulas 3.471.9 and 8.486.16.
170 2 Path Integrals — Elementary Properties and Simple Solutions

The expansion coefficients are equal to Riemann’s zeta function ζ(z) of Eq. (2.521)
at even arguments z = 2k, so that we may write
∞
!2k
X (−1)k βh̄ω
S(βh̄ω) = − ζ(2k) . (2.566)
m=1 k 2π

At even positive arguments, the values of the zeta function are related to the
Bernoulli numbers by26

(2π)2n
ζ(2n) = |B2n |. (2.567)
2(2n)!

The Bernoulli numbers are defined by the expansion

∞
t X tn
= Bn . (2.568)
et − 1 n=0 n!

The lowest nonzero Bernoulli numbers are B0 = 1, B1 = −1/2, B2 = 1/2, B4 =

−1/30, . . . . The Bernoulli numbers determine also the values of the zeta functions
at negative odd arguments:
B2n
ζ(1 − 2n) = − , (2.569)
2n
this being a consequence of the general identity27
zπ
ζ(z) = 2z π z−1 sin(πz/2)Γ(1 − z)ζ(1 − z) = 2z−1 π z ζ(1 − z)/Γ(z) cos . (2.570)
2
Typical values of ζ(z) which will be needed here are28

π2 π4 π6
ζ(2) = , ζ(4) = , ζ(6) = , . . . , ζ(∞) = 1. (2.571)
6 90 945
In contrast to the Matsubara frequency sum (2.563) and its expansion (2.566),
the dually transformed sum over the quantum numbers n in (2.564) converges rapidly
for low temperatures. It converges everywhere except at very large temperatures,
where it diverges logarithmically. The precise behavior can be calculated as follows:
For large T there exists a large number N which is still much smaller than 1/βh̄ω,
such that e−βh̄ωN is close to unity. Then we split the sum as
∞
X 1 −nβh̄ω NX−1
1 X∞
1 −nβh̄ω
e ≈ + e . (2.572)
n=1 n n=1 n n=N n

26
ibid., Formulas 9.542 and 9.535.
27
ibid., Formula 9.535.2.
28
Other often-needed values are ζ(0) = −1/2, ζ ′ (0) = − log(2π)/2, ζ(−2n) = 0, ζ(3) ≈
1.202057, ζ(5) ≈ 1.036928, . . . .

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 171

Since N is large, the second sum can be approximated by an integral

∞ dn −nβh̄ω ∞ dx −x
Z Z
e = e ,
N n N βh̄ω x
which is an exponential integral E1 (Nβh̄ω) of Eq. (2.470) with the large-argument
expansion −γ − log(Nβh̄ω) of Eq. (2.471).
The first sum in (2.572) is calculated with the help of the Digamma function
Γ′ (z)
ψ(z) ≡ . (2.573)
Γ(z)
This has an expansion29
∞ 
1 1
X 
ψ(z) = −γ − − , (2.574)
n=0 n+z n+1

which reduces for integer arguments to

N −1
X 1
ψ(N) = −γ + , (2.575)
n=1 n

and has the large-z expansion

∞
1 X B2n
ψ(z) ≈ log z − − . (2.576)
2z n=1 2nz 2n

Combining this with (2.471), the logarithm of N cancels, and we find for the sum
in (2.572) the large-T behavior
∞
X 1 −nβh̄ω
e ≈ − log βh̄ω + O(β). (2.577)
n=1 n
T →∞

This cancels the logarithm in (2.564).

The low-temperature series (2.564) can be used to illustrate the power of analytic
regularization. Suppose we want to extract from it the large-T behavior, where the
sum ∞
X 1 −nβh̄ω
g(βh̄ω) ≡ e (2.578)
n=1 n

converges slowly. We would like to expand the exponentials in the sum into powers
of ω, but this gives rise to sums over positive powers of n. It is possible to make
sense of these sums by analytic continuation. For this we introduce a generalization
of (2.578):
∞
1 −nβh̄ω
ζν (eβh̄ω ) ≡
X
ν
e , (2.579)
n=1 n

29
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 1.362.1.
172 2 Path Integrals — Elementary Properties and Simple Solutions

which reduces to g(βh̄ω) for ν = 1. This sum is evaluated by splitting it into an

integral over n and a difference between sum and integral:
∞
!
∞ 1 1 −nβh̄ω
Z Z ∞
βh̄ω
dn ν e−nβh̄ω +
X
ζν (e )= − e . (2.580)
0 n n=1 0 nν

The integral is convergent for ν < 1 and yields Γ(1 − ν) (βh̄ω)ν via the integral
formula (2.498). For other ν’s it is defined by analytic continuation. The remainder
may be expanded sloppily in powers of ω and yields
∞ ∞ ∞
(−1)k
! " ! #
∞ 1 1
Z Z ∞ Z ∞
βh̄ω
dn ν e−nβh̄ω + (βh̄ω)k . k−ν
X X X
ζν (e )= − + − n
0 n n=1 0 nν k=1 n=1 0 k!
(2.581)
The second term is simply the Riemann zeta function ζ(ν) [recall (2.521)]. Since
the additional integral vanishes due to Veltman’ rule (2.508), the zeta function may
also be defined by
∞
!
X Z ∞ 1
− = ζ(ν), (2.582)
n=1 0 νk

If this formula is applied to the last term in (2.581), we obtain the so-called Robinson
expansion 30
∞
1
ζν (eβh̄ω ) = Γ(1 − ν)(βh̄ω)ν−1 + ζ(ν) + (−βh̄ω)k ζ(ν − k).
X
(2.583)
k=1 k!

This expansion will later play an important role in the discussion of Bose-Einstein
condensation [see Eq. (7.38)].
For various applications it is useful to record also the auxiliary formula
∞ ∞
enβh̄ω
!
∞ 1
Z
(−βh̄ω)k ζ(ν − k) ≡ ζ̄ν (eβh̄ω ),
X X
− ν
= (2.584)
n=1 0 n k=1 k!

since in the sum minus the integral, the first Robinson terms are absent and the
result can be obtained from a naive Taylor expansion of the exponents enβh̄ω and
the summation formula (2.582).
From (2.583) we can extract the desired sum (2.578) by going to the limit ν → 1.
Close to the limit, the Gamma function has a pole Γ(1−ν) = 1/(1−ν)−γ +O(ν −1).
From the identity
πz
2z Γ(1 − z)ζ(1 − z) sin = π 1−z ζ(z) (2.585)
2
and (2.522) we see that ζ(ν) behaves near ν = 1 like
1
ζ(ν) = + γ + O(ν − 1) = −Γ(1 − ν) + O(ν − 1). (2.586)
ν−1
30
J.E. Robinson, Phys. Rev. 83 , 678 (1951).

H. Kleinert, PATH INTEGRALS

2.15 Field-Theoretic Definition of Harmonic Path Integrals by Analytic Regularization 173

Hence the first two terms in (2.583) can be combined to yield for ν → 1 the finite
result limν→1 Γ(1 − ν) [(βh̄ω)ν−1 − 1]=− log βh̄ω. The remaining terms contain in
the limit the values ζ(0) = −1/2, ζ(−1), ζ(−2), etc. Here we use the property of
the zeta function that it vanishes at even negative arguments, and that the function
at arbitrary negative argument is related to one at positive argument by the identity
(2.585). This implies for the expansion coefficients in (2.583) with k = 1, 2, 3, . . . in
the limit ν → 1:
1 (2p)!
ζ(−2p) = 0, ζ(1 − 2p) = (−1)p ζ(2p), p = 1, 2, 3, . . . . (2.587)
p (2π)2p
Hence we obtain for the expansion (2.583) in the limit ν → 1:
∞
βh̄ω X (−1)k
g(βh̄ω) = ζ1 (eβh̄ω ) = − log βh̄ω + + ζ(2k) (βh̄ω)2k . (2.588)
2 k=1 k!
This can now be inserted into Eq. (2.564) and we recover the previous expansion
(2.566) for S(βh̄ω) which was derived there by a proper duality transformation.
It is interesting to observe what goes wrong if we forget the separation (2.580)
of the sum into integral plus sum-minus-integral and its regularization. For this we
re-expand (2.578) directly, and illegally, in powers of ω. Then we obtain for ν = 1
the formal expansion
∞ ∞ ∞
(−1)p (−1)p
!
ζ1 (eβh̄ω ) = np−1 (βh̄ω)p = −ζ(1) + (βh̄ω)p ,
X X X
ζ(1 − p)
p=0 n=1 p! p=1 p!
(2.589)
which contains the infinite quantity ζ(1). The correct result (2.588) is obtained from
this by replacing the infinite quantity ζ(1) by − log βh̄ω, which may be viewed as a
regularized ζreg (1):
ζ(1) → ζreg (1) = − log βh̄ω. (2.590)
The above derivation of the Robinson expansion can be supplemented by a dual
version as follows. With the help of Poisson’s formula (1.197) we rewrite the sum
(2.579) as an integral over n and an auxiliary sum over integer numbers m, after
which the integral over n can be performed yielding
∞
βh̄ω
Z ∞ 1
dn e(2πim+βh̄ω)n = Γ(1 − ν)(−βh̄ω)ν−1
X
ζν (e ) ≡ ν
m=−∞ 0 n
∞
(−βh̄ω − 2πim)ν−1 .
X
+ Γ(1 − ν) 2 Re (2.591)
m=1

The sum can again be expanded in powers of ω

∞
!ν−1
X ν−1 βh̄ω
2 Re (−2πim) 1+
m=1 2πim
∞ 
ν −1

cos[(1 − ν − k)π/2] (2π)ν−1−k ζ(1 − ν + k) (βh̄ω)k .(2.592)
X
=2
k=0 k
174 2 Path Integrals — Elementary Properties and Simple Solutions

Using the relation (2.585) for zeta-functions, the expansion (2.591) is seen to coincide
with (2.579).
Note that the representation (2.591) of ζν (eβh̄ω ) is a sum over Matsubara fre-
quencies ωm = 2πm/β [recall Eq. (2.381)]:
∞ Z ∞ 1
ζν (eβh̄ω ) ≡ dn e(iωm +h̄ω)βn
X

m=−∞ 0 nν
∞
" #
ν−1 ν−1
X
= Γ(1 − ν)(−βh̄ω) 1 + 2 Re (1 + iωm /h̄ω) . (2.593)
m=1

The first term coming from the integral over n in (2.580) is associated with the
zero Matsubara frequency. This term represents the high-temperature or classical
limit of the expansion. The remainder contains the sum over all nonzero Matsubara
frequencies, and thus the effect of quantum fluctuations.
It should be mentioned that the first two terms in the low-temperature expansion
(2.564) can also be found from the sum (2.563) with the help of the Euler-Maclaurin
formula31 for a sum over discrete points t = a + (k + κ)∆ of a function F (t) from
k = 0 to K ≡ (b − a)/∆:
K
1 b 1
X Z
F (a + k∆) = dt F (t) + [F (a) + F (b)]
k=0 ∆ a 2
∞
∆2p−1 h i
B2p F (2p−1) (b) − F (2p−1) (a) ,
X
+ (2.594)
p=1 (2p)!

or, more generally for t = a + (k + κ)∆,

K−1 ∞
1 Zb ∆p−1 h i
Bp (κ) F (p−1) (b) − F (p−1) (a) ,
X X
F (a + (k + κ)∆) = dt F (t) +
k=0 ∆ a p=1 p!
(2.595)
where Bn (κ) are the Bernoulli functions defined by a generalization of the expansion
(2.568):
∞
teκt X tn
= Bn (κ) . (2.596)
et − 1 n=0 n!
At κ = 0, the Bernoulli functions start out with the Bernoulli numbers: Bn (0) = Bn .
The function B0 (κ) is equal to 1 everywhere.
Another way of writing formula (2.597) is
 
K−1 ∞
1 Zb  ∆p
Bp (κ)∂tp  F (t).
X X
F (a + (k + κ)∆) = dt 1 + (2.597)
k=0 ∆ a p=0 p!

This implies that a sum over discrete values of a function can be replaced by an
integral over a gradient expansion of the function.
31
M. Abramowitz and I. Stegun, op. cit., Formulas 23.1.30 and 23.1.32.

H. Kleinert, PATH INTEGRALS

2.16 Finite-N Behavior of Thermodynamic Quantities 175

Using the first Euler-Maclaurin formula (2.594) with a = ω12 , b = ωM

2
, and
∆ = ω1 , we find
M h i m=M
( ) m=M
ω ωm h
i  
2 2 2 2
+ ω 2 )−2
X
log(ωm + ω )− log(ωm ) = π + log(ωm − ω=0
m=0 ω1 ω1 m=1 m=1
1n o n o
+ log(ω12 + ω 2 ) + log(ωM2
+ ω 2 ) − ω = 0 . (2.598)
2
For small T , the leading two terms on the right-hand side are

ω 1 ω2
π − log 2 , (2.599)
ω1 2 ω1

in agreement with the first two terms in the low-temperature series (2.564). Note
that the Euler-Maclaurin formula is unable to recover the exponentially small terms
in (2.564), since they are not expandable in powers of T .
The transformation of high- into low-temperature expansions is an important
tool for analyzing phase transitions in models of statistical mechanics.32

2.16 Finite-N Behavior of Thermodynamic Quantities

Thermodynamic fluctuations in Euclidean path integrals are often imitated in computer simula-
tions. These are forced to employ a sliced time axis. It is then important to know in which way
the time-sliced thermodynamic quantities converge to their continuum limit. Let us calculate the
internal energy E and the specific heat at constant volume C for finite N from (2.484). Using
(2.484) we have

∂(β ω̃e ) ω
= ,
∂β cosh(ǫω̃e /2)
∂(ǫω̃e ) 2
= tanh(ǫω̃e /2), (2.600)
∂β β

and find the internal energy

∂ h̄ ∂(β ω̃e )
E = (βF ) = coth(βh̄ω̃e /2)
∂β 2 ∂β
h̄ω coth(βh̄ω̃e /2)
= . (2.601)
2 cosh(ǫω̃e /2)

The specific heat at constant volume is given by

1 ∂2 ∂
C = −β 2 2 (βF ) = −β 2 E (2.602)
kB ∂β ∂β
 
1 1 ǫ 1
= β 2 h̄2 ω 2 + coth(βh̄ω̃ e /2) tanh(ǫω̃ e /2) .
4 sinh2 (βh̄ω̃e /2) h̄β cosh2 (ǫω̃e /2)

Plots are shown in Fig. 2.5 for various N using natural units with h̄ = 1, kB = 1. At high
32
See H. Kleinert, Gauge Fields in Condensed Matter , Vol. I Superflow and Vortex Lines, World
Scientific, Singapore, 1989, pp. 1–742 (http://www.physik.fu-berlin.de/~kleinert/b1).
176 2 Path Integrals — Elementary Properties and Simple Solutions

Figure 2.4 Finite-lattice effects in internal energy E and specific heat C at constant volume,
as a function of the temperature for various numbers N + 1 of time slices. Note the nonuniform
way in which the exponential small-T behavior of C ∝ e−ω/T is approached in the limit N → ∞.

temperatures, F, E, and C are independent of N :

1
F → log β, (2.603)
β
1
E → = T, (2.604)
β
C → 1. (2.605)
These limits are a manifestation of the Dulong-Petit law : An oscillator has one kinetic and one
potential degree of freedom, each carrying an internal energy T /2 and a specific heat 1/2. At low
temperatures, on the other hand, E and C are strongly N -dependent (note that since F and E
are different at T = 0, the entropy of the lattice approximation does not vanish as it must in the
continuum limit). Thus, the convergence N → ∞ is highly nonuniform. After reaching the limit,
the specific heat goes to zero for T → 0 exponentially fast, like e−ω/T . The quantity ω is called
activation energy.33 It is the energy difference between the ground state and the first excited state
of the harmonic oscillator. For large but finite N , on the other hand, the specific heat has the
large value N + 1 at T = 0. This is due to ω̃e and cosh2 (ǫω̃e /2) behaving, for a finite N and T → 0
(where ǫ becomes large) like
1
ω̃e →log(ǫ2 ω 2 ),
ǫ
cosh(ǫω̃e /2) → ǫω/2. (2.606)
Hence
T →0 1 T →0
E −
−−→ coth[(N + 1) log(ǫω)] −
−−→ 0, (2.607)
β
33
Note that in a D-dimensional solid the lattice vibrations can be considered as an ensemble of
harmonic oscillators with energies ω ranging from zero to the Debye frequency. Integrating over
the corresponding specific heats with the appropriate density of states, dωω D−1 e−ω/kB T , gives
R

the well-known power law at low temperatures C ∝ T D .

H. Kleinert, PATH INTEGRALS

2.17 Time Evolution Amplitude of Freely Falling Particle 177

T →0
C −
−−→ N + 1. (2.608)
The reason for the nonuniform approach of the N → ∞ limit is obvious: If we expand (2.484) in
powers of ǫ, we find
 
1 2 2
ω̃e = ω 1 − ǫ ω + . . . . (2.609)
24
When going to low T at finite N the corrections are quite large, as can be seen by writing (2.609),
with ǫ = h̄β/(N + 1), as
h̄2 ω 2
 
1
ω̃e = ω 1 − + ... . (2.610)
24 kb2 T 2 (N + 1)2
Note that (2.609) contains no corrections of the order ǫ. This implies that the convergence of
all thermodynamic quantities in the limit N → ∞, ǫ → 0 at fixed T is quite fast — one order in
1/N faster than we might at first expect [the Trotter formula (2.26) also shows the 1/N 2 -behavior].

2.17 Time Evolution Amplitude of Freely Falling Particle

The gravitational potential of a particle on the surface of the earth is
V (x) = V0 + M g · x, (2.611)
where −g is the earth’s acceleration vector pointing towards the ground, and V0
some constant. The equation of motion reads
ẍ = −g, (2.612)
which is solved by
g
x = xa + va (t − ta ) + (t − ta )2 , (2.613)
2
with the initial velocity
xb − xa g
va = − (tb − ta ). (2.614)
tb − ta 2
Inserting this into the action
tb M 2
Z  
A= dt ẋ − V0 − g · x , (2.615)
ta 2
we obtain the classical action
M (xb − xa )2 1 1
Acl = −V0 (tb − ta ) + − (tb − ta )g · (xb + xa ) − (tb − ta )3 g2 . (2.616)
2 tb − ta 2 24
Since the quadratic part of (2.615) is the same as for a free particle, also the fluctu-
ation factor is the same [see (2.130)], and we find the time evolution amplitude
1 − h̄i V0 (tb −ta )
(xb tb |xa ta ) = q 3e
2πih̄(tb − ta )/M
iM (xb − xa )2
( " #)
1
× exp − (tb − ta )g · (xb + xa ) − (tb − ta )3 g2 . (2.617)
2h̄ tb − ta 12
178 2 Path Integrals — Elementary Properties and Simple Solutions

The potential (2.611) can be considered as a limit of a harmonic potential

M 2
V (x) = V0 + ω (x − x0 )2 (2.618)
2
for

ω → 0, x0 = −g/ω 2 → −∞ ≈ ĝ, V0 = −Mx20 /2 = −Mg 2 /2ω 4 → −∞, (2.619)

keeping
g = −Mω 2 x0 , (2.620)
and
M 2 2
v0 = V0 +
ω x0 (2.621)
2
fixed. If we perform this limit in the amplitude (2.177), we find of course (2.617).
The wave functions can be obtained most easily by performing this limiting
procedure on the wave functions of the harmonic oscillator. In one dimension, we
set n = E/ω and find that the spectral representation (2.294) goes over into
Z
(xb tb |xa ta ) = dEAE (xb )A∗E (xa )e−i(E−v0 )(tb −ta )/h̄ , (2.622)

with the wave functions

1 x E
 
AE (x) = √ Ai − . (2.623)
lε l ε
Here ε ≡ (h̄2 g 2M/2)1/3 and l ≡ (h̄2 /2M 2 g)1/3 = ε/Mg are the natural units of
energy and length, respectively, and Ai(z) is the Airy function solving the differential
equation
Ai′′ (z) = zAi(z), (2.624)
For positive z, the Airy function can be expressed in terms of modified Bessel func-
tions Iν (ξ) and Kν (ξ):34
√
z 1 z
r  
Ai(z) = [I−1/3 (2z 3/2 /3) − I1/3 (2z 3/2 /3)] = K1/3 2z 3/2 /3 . (2.625)
2 π 3
For large z, this falls off exponentially:
1 3/2
Ai(z) → √ 1/4 e−2z /3 , z → ∞. (2.626)
2 πz
For negative z, an analytic continuation35

Iν (ξ) = e−πνi/2 J(eπi/2 ξ), − π < argξ ≤ π/2,

Iν (ξ) = e−πνi/2 J(eπi/2 ξ), π/2 < argξ ≤ π, (2.627)
34
A compact description of the properties of Bessel functions is found in M. Abramowitz and I.
Stegun, op. cit., Chapter 10. The Airy function is expressed in Formulas 10.4.14.
35
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formulas 8.406.

H. Kleinert, PATH INTEGRALS

2.18 Charged Particle in Magnetic Field 179

leads to
1√ h i
Ai(z) = z J−1/3 (2(−z)3/2 /3) + J1/3 (2(−z)3/2 /3) , (2.628)
3
where J1/3 (ξ) are ordinary Bessel functions. For large arguments, these oscillate like
s
2
Jν (ξ) → cos(ξ − πν/2 − π/4) + O(ξ −1 ), (2.629)
πξ
from which we obtain the oscillating part of the Airy function
1 h
3/2
i
Ai(z) → √ sin 2(−z) /3 + π/4 , z → −∞. (2.630)
πz 1/4
The Airy function has the simple Fourier representation
∞ dk i(xk+k3 /3)
Z
Ai(x) = e . (2.631)
−∞ 2π
In fact, the momentum space wave functions of energy E are
s
l −i(pE−p3 /6M )l/εh̄
hp|Ei = e (2.632)
ε
fulfilling the orthogonality and completeness relations
dp
Z Z
hE ′ |pihp|Ei = δ(E ′ − E), dE hp′ |EihE|pi = 2πh̄δ(p′ − p). (2.633)
2πh̄
The Fourier transform of (2.632) is equal to (2.623), due to (2.631).

2.18 Charged Particle in Magnetic Field

Having learned how to solve the path integral of the harmonic oscillator we are ready
to study also a more involved harmonic system of physical importance: a charged
particle in a magnetic field. This problem was first solved by L.D. Landau in 1930
in Schrödinger theory.36

2.18.1 Action
The magnetic interaction of a particle of charge e is given by
e Z tb
Amag = dt ẋ(t) · A(x(t)), (2.634)
c ta
where A(x) is the vector potential of the magnetic field. The total action is
tb M 2 e
Z  
A[x] = dt ẋ (t) + ẋ(t) · A(x(t)) . (2.635)
ta 2 c
36
L.D. Landau and E.M. Lifshitz, Quantum Mechanics, Pergamon, London, 1965.
180 2 Path Integrals — Elementary Properties and Simple Solutions

Suppose now that the particle moves in a homogeneous magnetic field B pointing
along the z-direction. Such a field can be described by a vector potential

A(x) = (0, Bx, 0). (2.636)

But there are other possibilities. The magnetic field

B(x) = ∇ × A(x) (2.637)

as well as the magnetic interaction (2.634) are invariant under gauge transformations

A(x) → A(x) + ∇Λ(x), (2.638)

where Λ(x) are arbitrary single-valued functions of x. As such they satisfy the
Schwarz integrability condition [compare (1.40)–(1.41)]

(∂i ∂j − ∂j ∂i )Λ(x) = 0. (2.639)

For instance, the axially symmetric vector potential

1
Ã(x) = B×x (2.640)
2
gives the same magnetic field; it differs from (2.636) by a gauge transformation

Ã(x) = A(x) + ∇Λ(x), (2.641)

with the gauge function

1
Λ(x) = − B xy. (2.642)
2
In the canonical form, the action reads
( 2 )
tb 1 e
Z 
A[p, x] = dt p · ẋ − p − A(x) . (2.643)
ta 2M c

The magnetic interaction of a point particle is thus included in the path integral by
the so-called minimal substitution of the momentum variable:
e
p → P ≡ p − A(x). (2.644)
c
For the vector potential (2.636), the action (2.643) becomes
Z tb
A[p, x] = dt [p · ẋ − H(p, x)] , (2.645)
ta

with the Hamiltonian

p2 1 1
H(p, x) = + MωL2 x2 − ωL lz (p, x), (2.646)
2M 8 2

H. Kleinert, PATH INTEGRALS

2.18 Charged Particle in Magnetic Field 181

where x = (x, y) and p = (px , py ) and

lz (p, x) = (x × p)z = xpy − ypx (2.647)

is the z-component of the orbital angular momentum. In a Schrödinger equation, the

last term in H(p, x) is diagonal on states with a given angular momentum around
the z-axis. We have introduced the field-dependent frequency
e
ωL = B, (2.648)
Mc
called Landau frequency or cyclotron frequency. This can also be written in terms
of the Bohr magneton
h̄e
µB ≡ , (2.649)
Mc
as
ωL = µB B/h̄. (2.650)
The first two terms in (2.646) describe a harmonic oscillator in the xy-plane with
a field-dependent magnetic frequency
ωL
ωB ≡ . (2.651)
2
Note that in the gauge gauge (2.636), the Hamiltonian would have the rotation-
ally noninvariant form

p2 1
H(p, x) = + MωL2 x2 − ωL xpy (2.652)
2M 2
rather than (2.646), implying oscillations of frequency ωL in the x-direction and a
free motion in the y-direction.
The time-sliced form of the canonical action (2.643) reads
N +1 
1 h 2 i
AN 2 2
X
e = pn (xn − xn−1 ) − p + (py n − Bxn ) + pzn , (2.653)
n=1 2M x n

and the associated time-evolution amplitude for the particle to run from xa to xb is
given by
N Z  NY
+1 Z
d 3 pn
" #
i N
 
3
Y
(xb tb |xa ta ) = d xn exp A , (2.654)
n=1 n=1 (2πh̄) 3 h̄ e

with the time-sliced action

N +1 
N
X 1 h 2 2 2
i
A = pn (xn − xn−1 ) − p + (py n − Bxn ) + pzn . (2.655)
n=1 2M x n
182 2 Path Integrals — Elementary Properties and Simple Solutions

2.18.2 Gauge Properties

Note that the time evolution amplitude is not gauge-invariant. If we use the vector
potential in some other gauge

A′ (x) = A(x) + ∇Λ(x), (2.656)

the action changes by a surface term

e tb e
Z
∆A = dt ẋ · ∇Λ(x) = [Λ(xb ) − Λ(xa )]. (2.657)
c ta c
The amplitude is therefore multiplied by a phase factor on both ends

(xb tb |xa ta )A → (xb tb |xa ta )A′ = eieΛ(xb )/ch̄ (xb tb |xa ta )A e−ieΛ(xa )/ch̄ . (2.658)

For the observable particle distribution (x tb |x ta ), the phase factors are obviously
irrelevant. But all other observables of the system must also be independent of the
phases Λ(x) which is assured if they correspond to gauge-invariant operators.

2.18.3 Time-Sliced Path Integration

Since the action AN contains the variables yn and zn only in the first term
PN +1
n=1 ipn xn , we can perform the yn , zn integrations and find a product of N ∆-
functions in the y- and z-components of the momenta pn . If the projections of p to
the yz-plane are denoted by p′ , the product is
 
(2πh̄)2 δ (2) p′N +1 − p′N · · · (2πh̄)2 δ (2) (p′2 − p′1 ) . (2.659)

These allow performing all py n , pz n -integrals, except for one overall py , pz . The path
integral reduces therefore to
N Z ∞  NY
+1 Z
" #
∞ dpy dpz Y dpx n
Z
(xb tb |xa ta ) = dxn (2.660)
−∞ (2πh̄)2 n=1 −∞ n=1 2πh̄
p2z
( " #)
i i N
 
× exp py (yb − ya ) + pz (zb − za ) − (tb − ta ) exp A ,
h̄ 2M h̄ x

where AN x is the time-sliced action involving only a one-dimensional path integral

over the x-component of the path, x(t), with the sliced action
N +1 h
p2x n
2 i
1 e

AN
X
x = px n (xn − xn−1 ) − − py − Bxn . (2.661)
n=1 2M 2M c

This is the action of a one-dimensional harmonic oscillator with field-dependent

frequency ωB whose center of oscillation depends on py and lies at

x0 = py /MωL . (2.662)

H. Kleinert, PATH INTEGRALS

2.18 Charged Particle in Magnetic Field 183

The path integral over x(t) is harmonic and known from (2.175):
s
MωL
(xb tb |xa ta )x0 =
2πih̄ sin ωL (tb − ta )
i MωL
 nh i
× exp (xb − x0 )2 + (xa − x0 )2 cos ωL (tb − ta )
h̄ 2 sin ωL (tb − ta )

− 2(xb − x0 )(xa − x0 )} . (2.663)

Doing the pz -integral in (2.660), we arrive at the formula

(z −z )2
1 iM b a
(xb tb |xa ta ) = q e 2h̄ tb −ta (x⊥ ⊥
b tb |xa ta ), (2.664)
2πih̄(tb − ta )/M
with the amplitude orthogonal to the magnetic field
⊥ ⊥ MωL Z ∞
(xb tb |xa ta ) ≡ dx0 eiM ωL x0 (yb −ya )/h̄ (xb tb |xa ta )x0 . (2.665)
2πh̄ −∞
After a quadratic completion in x0 , the total exponent in (2.665) reads
" #
iMωL 1
−(x2b + x2a ) tan[ωL (tb − ta )/2] + (xb − xa )2
2h̄ sin ωL (tb − ta )
!2
MωL xb + xa yb − ya
−i tan[ωL (tb − ta )/2] x0 − −
h̄ 2 2 tan[ωL (tb − ta )/2]
MωL (xb + xa )2 (yb − ya )2
" #
+i tan[ωL (tb − ta )/2] +
2h̄ 2 2 tan[ωL (tb − ta )/2]
MωL
+i (xb + xa )(yb − ya ). (2.666)
2h̄
R∞
The integration MωL −∞ dx0 /2πh̄ removes the second term and results in a factor
s
MωL πh̄
. (2.667)
2πh̄ iMωL tan[ωL (tb − ta )/2]
By rearranging the remaining terms, we arrive at the amplitude
s 3
M ωL (tb − ta )/2 i
 
(xb tb |xa ta ) = exp (Acl + Asf ) ,
2πih̄(tb − ta ) sin[ωL (tb − ta )/2] h̄
(2.668)
with an action
M (zb − za )2 ωL
 h i
Acl = + cot[ωL (tb − ta )/2] (xb − xa )2 + (yb − ya )2
2 tb − ta 2

+ ωL (xa yb − xb ya ) , (2.669)

and the surface term

MωL e b
Asf = (xb yb − xa ya ) = B xy . (2.670)
2 2c a
184 2 Path Integrals — Elementary Properties and Simple Solutions

2.18.4 Classical Action

Since the action is harmonic, the amplitude is again a product of a phase eiAcl and
a fluctuation factor. A comparison with (2.641) and (2.658) shows that the surface
term would be absent if the amplitude (xb tb |xa ta )Ã were calculated with the vector
potential in the axially symmetric gauge (2.640). Thus Acl must be equal to the
classical action in this gauge. Indeed, the orthogonal part can be rewritten as
( )
tb M d M
Z
A⊥
cl = dt (xẋ + y ẏ) + [x(−ẍ + ωL ẏ) + y(−ÿ − ωL ẋ)] . (2.671)
ta 2 dt 2
The equations of motion are
ẍ = ωL ẏ, ÿ = −ωL ẋ, (2.672)
reducing the action of a classical orbit to
M tb M
A⊥
cl = (xẋ + y ẏ) = ([xb ẋb − xa ẋa ] + [yb ẏb − ya ẏa ]) . (2.673)
2 ta 2
The orbits are easily determined. By inserting the two equations in (2.672) into
each other we see that ẋ and ẏ perform independent oscillations:
ẍ˙ + ωL2 ẋ = 0, ÿ˙ + ωL2 ẏ = 0. (2.674)
The general solution of these equations is
1
x= [(xb − x0 ) sin ωL (t − ta ) − (xa − x0 ) sin ωL (t − tb ) ] + x0 , (2.675)
sin ωL (tb − ta )
1
y= [(yb − y0 ) sin ωL (t − ta ) − (ya − y0 ) sin ωL (t − tb ) ] + y0 , (2.676)
sin ωL (tb − ta )
where we have incorporated the boundary condition x(ta,b ) = xa,b , y(ta,b )
= ya,b . The constants x0 , y0 are fixed by satisfying (2.672). This gives
1 ωL
 
x0 = (xb + xa ) + (yb − ya ) cot (tb − ta ) , (2.677)
2 2
1 ωL
 
y0 = (yb + ya ) − (xb − xa )cot (tb − ta ) . (2.678)
2 2
Now we calculate
ωL
xb ẋb = xb [(x0 − xa ) + (xb − x0 ) cos ωL (tb − ta )] , (2.679)
sin ωL (tb − ta )
ωL
xa ẋa = xa [(x0 − xa ) cos ωL (tb − ta ) + (xb − x0 )] , (2.680)
sin ωL (tb − ta )
and hence
ωL
xb ẋb − xa ẋa = ωL x0 (xb + xa ) tan (tb − ta )
2
ωL h i
+ (x2b + x2a ) cos ωL (tb − ta ) − 2xb xa
sin ωL (tb − ta )
ωL ωL
 
2
= (xb − xa ) cot (tb − ta ) + (xb + xa )(yb − ya ) . (2.681)
2 2

H. Kleinert, PATH INTEGRALS

2.18 Charged Particle in Magnetic Field 185

Similarly we find
ωL
ybẏb − ya ẏa = ωL y0 (yb + ya ) tan (tb − ta )
2
ωL
+ [(y 2 + ya2 ) cos ωL (tb − ta ) − 2yb ya ]
sin ωL (tb − ta ) b
ωL ωL
 
2
= (yb − ya ) cot (tb − ta ) − (xb − xa )(yb + ya ) . (2.682)
2 2

Inserted into (2.673), this yields the classical action for the orthogonal motion

M ωL
 h i 
A⊥
cl = cot[ωL (tb − ta )/2] (xb − xa )2 + (yb − ya )2 + ωL (xa yb − xb ya ) ,
2 2
(2.683)

which is indeed the orthogonal part of the action (2.669).

2.18.5 Translational Invariance

It is interesting to see how the amplitude ensures the translational invariance of all
physical observables. The first term in the classical action is trivially invariant. The
last term reads
MωL
∆A = (xa yb − xb ya ). (2.684)
2
Under a translation by a distance d,

x → x + d, (2.685)

this term changes by

MωL MωL
[dx (yb − ya ) + dy (xa − xb )] = [(d × x)b − (d × x)a ]z (2.686)
2 2
causing the amplitude to change by a pure gauge transformation as

(xb tb |xa ta ) → eieΛ(xb )/ch̄ (xb tb |xa ta )e−ieΛ(xa )/ch̄ , (2.687)

with the phase

MωL h̄c
Λ(x) = − [d × x]z . (2.688)
2e
Since observables involve only gauge-invariant quantities, such transformations are
irrelevant.
This will be done in 2.23.3.
186 2 Path Integrals — Elementary Properties and Simple Solutions

2.19 Charged Particle in Magnetic Field plus

Harmonic Potential
For application in Chapter 5 we generalize the above magnetic system by adding a
harmonic oscillator potential, thus leaving the path integral solvable. For simplicity,
we consider only the orthogonal part of the resulting system with respect to the
magnetic field. Omitting orthogonality symbols, the Hamiltonian is the same as in
(2.646). Without much more work we may solve the path integral of a more general
system in which the harmonic potential in (2.646) has a different frequency ω 6= ωL ,
and thus a Hamiltonian
p2 1
H(p, x) = + Mω 2 x2 − ωB lz (p, x). (2.689)
2M 2
The associated Euclidean action
Z τb
Ae [p, x] = dτ [−ip · ẋ + H(p, x)] (2.690)
τa

has the Lagrangian form

h̄β M 2 1
Z  
Ae [x] = dτ ẋ (τ ) + M(ω 2 − ωB2 )x2 (τ ) − iMωB [x(τ ) × ẋ(τ )]z .
0 2 2
(2.691)

At this point we observe that the system is stable only for ω ≥ ωB . The action
(2.691) can be written in matrix notation as
" #
h̄β M d M
Z
Acl = dτ (xẋ) + xT Dω2 ,B x , (2.692)
0 2 dτ 2
where Dω2 ,B is the 2 × 2 -matrix

−∂τ2 + ω 2 − ωB2
!
′ −2iωB ∂τ
Dω2 ,B (τ, τ ) ≡ δ(τ − τ ′ ). (2.693)
2iωB ∂τ −∂τ2 + ω 2 − ωB2

Since the path integral is Gaussian, we can immediately calculate the partition
function
1 −1/2
Z= det Dω2 ,B . (2.694)
(2πh̄/M)2
By expanding Dω2 ,B (τ, τ ′ ) in a Fourier series
∞
1 X ′
Dω2 ,B (τ, τ ′ ) = D̃ω2 ,B (ωm )e−iωm (τ −τ ) , (2.695)
h̄β m=−∞

we find the Fourier components

!
2
ωm + ω 2 − ωB2 −2ωB ωm
D̃ω2 ,B (ωm ) = , (2.696)
2ωB ωm ωm + ω 2 − ωB2
2

H. Kleinert, PATH INTEGRALS

2.19 Charged Particle in Magnetic Field plus Harmonic Potential 187

with the determinants

2
det D̃ω2 ,B (ωm ) = (ωm + ω 2 − ωB2 )2 + 4ωB2 ωm
2
. (2.697)
These can be factorized as
2 2 2 2
det D̃ω2 ,B (ωm ) = (ωm + ω+ )(ωm + ω− ), (2.698)
with
ω± ≡ ω ± ωB . (2.699)
The eigenvectors of D̃ω2 ,B (ωm ) are
! !
1 1 1 1
e+ = √ , e− = − √ , (2.700)
2 i 2 −i
with eigenvalues
2
d ± = ωm + ω 2 ± 2iωm ωB = (ωm + iω± )(ωm − iω∓ ). (2.701)
√
Thus the right- and left-circular combinations x± = ±(x ± iy)/ 2 diagonalize the
Lagrangian (2.692) to
(
h̄β M d  ∗
Z 
Acl = dτ x+ ẋ+ + x∗− ẋ−
0 2 dτ
Mh ∗ i 
+ x+ (−∂τ − ω+ )(−∂τ + ω− )x+ + x∗− (−∂τ − ω− )(−∂τ + ω+ )x− . (2.702)
2
Continued back to real times, the components x± (t) are seen to oscillate indepen-
dently with the frequencies ω± .
The factorization (2.698) makes (2.702) an action of two independent harmonic
oscillators of frequencies ω± . The associated partition function has therefore the
product form
1 1
Z= . (2.703)
2 sinh (h̄βω+ /2) 2 sinh (h̄βω− /2)
For the original system of a charged particle in a magnetic field discussed in
Section 2.18, the partition function is obtained by going to the limit ω → ωB in
the Hamiltonian (2.689). Then ω− → 0 and the partition function (2.703) diverges,
since the system becomes translationally invariant in space. From the mnemonic
replacement rule (2.361) we see that in this limit we must replace the relevant
vanishing inverse frequency by an expression proportional to the volume of the
system. The role of ω 2 in (2.361) is played here by the frequency in front of the
x2 -term of the Lagrangian (2.691). Since there are two dimensions, we must replace
1 1 β
2
−
−−→ −
−−→ V2 , (2.704)
ω 2 − ωB ω→ωB 2ωω− ω− →0 2π/M
and thus
1 V2
Z −
−−→ , (2.705)
ω− →0 2 sinh (h̄βω) λ2ω
where λω is the quantum-mechanical length scale in Eqs. (2.303) and (2.359) of the
harmonic oscillator.
188 2 Path Integrals — Elementary Properties and Simple Solutions

2.20 Gauge Invariance and Alternative Path Integral

Representation
The action (2.635) of a particle in an external ordinary potential V (x, t) and a vector potential
A(x, t) can be rewritten with the help of an arbitrary space- and time-dependent gauge function
Λ(x, t) in the following form:
tb  
M 2 e e
Z
A[x] = dt ẋ + ẋ(t)[A(x, t) + ∇Λ(x, t)]−V (x, t)+ ∂t Λ(x, t)
ta 2 c c
e
− [Λ(xb , tb ) − Λ(xa , ta )] . (2.706)
c
The Λ(x, t)-terms inside the integral are canceled by the last two surface terms making the action
independent of Λ(x, t).
We may now choose a particular function Λ(x, t) equal to c/e-times the classical action A(x, t)
which solves the Hamilton-Jacobi equation (1.65), i.e.,

1 h e i2
∇A(x, t) − A(x, t) + ∂t A(x, t) + V (x, t) = 0. (2.707)
2M c
Then we obtain the following alternative expression for the action:
tb i2
1 h e
Z
A[x] = dt M ẋ − ∇A(x, t) + A(x, t)
ta 2M c
+ A(xb , tb ) − A(xa , ta ). (2.708)

For two infinitesimally different solutions of the Hamilton-Jacobi equation, the difference between
the associated action functions δA satisfies the differential equation

v · ∇δA + ∂t δA = 0, (2.709)

where v is the classical velocity field

h e i
v(x, t) ≡ (1/M ) ∇A(x, t) − A(x, t) . (2.710)
c
The differential equation (2.709) expresses the fact that two solutions A(x, t) for which the particle
energy and momenta at x and t differ by δE and δp, respectively, satisfy the kinematic relation
δE = p · δp/M = ẋcl · ∇δA. This follows directly from E = p2 /2M . The so-constrained variations
δE and δp leave the action (2.708) invariant.
A sequence of changes δA of this type can be used to make the function A(x, t) coincide with
the action A(x, t; xa , ta ) of paths which start out from xa , ta and arrive at x, t. In terms of this
action function, the path integral representation of the time evolution amplitude takes the form
Z x(tb )=xb
(xb tb |xa ta ) = eiA(xb ,tb ;xa ,ta )/h̄ Dx (2.711)
x(ta )=xa
 Z tb i2 
i 1 h e
×exp dt M ẋ − ∇A(x, t; xa , ta ) + A(x, t)
h̄ ta 2M c

or, using v(x(t), t),

x(tb )=xb  Z tb 
i M
Z
iA(xb ,tb ;xa ,ta )/h̄ 2
(xb tb |xa ta ) = e Dx exp dt (ẋ − v) .
x(ta )=xa h̄ ta 2
(2.712)

H. Kleinert, PATH INTEGRALS

2.21 Velocity Path Integral 189

The fluctuations are now controlled by the deviations of the instantaneous velocity ẋ(t) from local
value of the classical velocity field v(x, t). Since the path integral attempts to keep the deviations
as small as possible, we call v(x, t) the desired velocity of the particle at x and t. Introducing
momentum variables p(t), the amplitude may be written as a phase space path integral
x(tb )=xb
Dp
Z Z
iA(xb ,tb ;xa ,ta )/h̄ ′
(xb tb |xa ta ) = e Dx
x(ta )=xa 2πh̄
 Z tb  
i 1 2
× exp dt p(t) [ẋ(t) − v(x(t), t)] − p (t) , (2.713)
h̄ ta 2M

which will be used in Section 18.15 to give a stochastic interpretation of quantum processes.

2.21 Velocity Path Integral

There exists yet another form of writing the path integral in which the fluctuating velocities play a
fundamental role and which will later be seen to be closely related to path integrals in the so-called
stochastic calculus to be introduced in Sections 18.13 and 18.640. We observe that by rewriting
the path integral as
 Z tb   Z tb  
i M 2
Z
3
(xb tb |xa ta ) = D x δ xb −xa − dt ẋ(t) exp dt ẋ −V (x) , (2.714)
ta h̄ ta 2

the δ-function allows us to include the last variable xn in the integration measure of the time-sliced
version of the path integral. Thus all time-sliced time derivatives (xn+1 − xn )/ǫ for n = 0 to N
are integrated over implying that they can be considered as independent fluctuating variables vn .
In the potential, the dependence on the velocities can be made explicit by inserting
Z tb
x(t) = xb − dt v(t), (2.715)
t
Z t
x(t) = xa + dt v(t), (2.716)
ta
tb
1
Z
x(t) = X+ dt′ v(t′ )ǫ(t′ − t), (2.717)
2 ta

where
xb + xa
X≡ (2.718)
2
is the average position of the endpoints and ǫ(t − t′ ) is the antisymmetric combination of Heaviside
functions introduced in Eq. (1.315).
In the first replacement, we obtain the velocity path integral
 Z tb   Z tb   Z tb 
i M 2
Z
(xb tb |xa ta ) = D3 v δ xb −xa− dt v(t) exp dt v −V xb − dt v(t) . (2.719)
ta h̄ ta 2 t

The measure of integration is normalized to make

 Z tb  
i M 2
Z
3
D v exp dt v = 1. (2.720)
h̄ ta 2

The correctness of this normalization can be verified by evaluating (2.719) for a free particle.
Inserting the Fourier representation for the δ-function
Z tb Z tb
d3 p
  Z   
i
δ xb − xa − dt v(t) = exp p xb − xa − dt v(t) , (2.721)
ta (2πi)3 h̄ ta
190 2 Path Integrals — Elementary Properties and Simple Solutions

we can complete the square in the exponent and integrate out the v-fluctuations using (2.720) to
obtain
d3 p p2
  
i
Z
(xb tb |xa ta ) = exp p (xb − xa ) − (t b − t a ) . (2.722)
(2πi)3 h̄ 2M
This is precisely the spectral representation (1.333) of the free-particle time evolution amplitude
(1.335) [see also Eq. (2.53)].
A more symmetric velocity path integral is obtained by choosing the third replacement (2.717).
This leads to the expression
 Z tb   Z tb 
i M 2
Z
3
(xb tb |xa ta ) = D v δ ∆x − dt v(t) exp dt v
ta h̄ ta 2
 Z tb  Z tb 
i 1
× exp − dt V X + dt′ v(t′ )ǫ(t′ − t) . (2.723)
h̄ ta 2 ta
The velocity representations are particularly useful if we want to know integrated amplitudes such
as
 Z tb 
i M 2
Z Z
3 3
d xa (xb tb |xa ta ) = D v exp dt v .
h̄ ta 2
Z tb
1 tb ′
  
i
Z
× exp − dtV xb − dt v(t′ ) , (2.724)
h̄ ta 2 t
which will be of use in the next section.

2.22 Path Integral Representation of the Scattering Matrix

In Section 1.16 we have seen that the description of scattering processes requires several nontrivial
limiting procedures on the time evolution amplitude. Let us see what these procedures yield when
applied to the path integral representation of this amplitude.

2.22.1 General Development

Formula (1.474) for the scattering matrix expressed in terms of the time evolution operator in
momentum space has the following path integral representation:
Z Z
hpb |Ŝ|pa i ≡ lim ei(Eb tb −Ea ta )/h̄ d3 xb d3 xa e−i(pb xb −pa xa )/h̄ (xb tb |xa ta ). (2.725)
tb −ta →∞

Introducing the momentum transfer q ≡ (pb − pa ), we rewrite e−i(pb xb −pa xa )/h̄ as

e−iqxb /h̄ e−ipa (xb −xa )/h̄ , and observe that the amplitude including the exponential prefactor
e−ipa (xb −xa )/h̄ has the path integral representation:
 Z tb  
i M 2
Z
−ipa (xb −xa )/h̄ 3
e (xb tb |xa ta ) = D x exp dt ẋ − pa ẋ − V (x) . (2.726)
h̄ ta 2

The linear term in ẋ is eliminated by shifting the path from x(t) to

pa
y(t) = x(t) − t (2.727)
M
leading to
 Z tb   
i M 2
Z
−ipa (xb −xb )/h̄ −ip2a (tb −ta )/2Mh̄ 3 P
e (xb tb |xa ta ) = e D y exp dt ẏ −V y+ t .
h̄ ta 2 M
(2.728)

H. Kleinert, PATH INTEGRALS

2.22 Path Integral Representation of the Scattering Matrix 191

Inserting everything into (2.725) we obtain

Z Z
2
hpb |Ŝ|pa i ≡ eiq tb /2Mh̄ d3 yb e−iqyb /h̄ d3 ya
lim
tb −ta →∞
 Z tb  
i M 2
Z  pa 
× D3 y exp dt ẏ −V y + t . (2.729)
h̄ ta 2 M

In the absence of an interaction, the path integral over y(t) gives simply
" #
2
1 i M (yb − ya )
Z
3
d ya p exp = 1, (2.730)
2πh̄i(tb − ta )/M h̄ 2 tb − ta )

and the integral over ya yields

2
hpb |Ŝ|pa i|V ≡0 = lim eiq (tb −ta )/8Mh̄
(2πh̄)3 δ (3) (q) = (2πh̄)3 δ (3) (pb − pa ), (2.731)
tb −ta →∞

which is the contribution from the unscattered beam to the scattering matrix in Eq. (1.477).
The first-order contribution from the interaction reads, after a Fourier decomposition of the
potential,

i d3 Q
Z Z Z
iq2 tb /2Mh̄ 3 −iqyb /h̄
hpb |Ŝ1 |pa i = − lim e d yb e V (Q) d3 ya
h̄ tb −ta →∞ (2πh̄)3
Z tb  Z  Z tb  
′ i pa Q ′ 3 i M 2 ′
× dt exp t D y exp dt ẏ + δ(t −t)Q y . (2.732)
ta h̄ M h̄ ta 2

The harmonic path integral was solved in one dimension for an arbitrary source j(t) in Eq. (3.168).
For ω = 0 and the particular source j(t) = δ(t′ − t)Q the result reads, in three dimensions,

i M (yb − ya )2
 
1
3 exp
h̄ 2 tb − ta
p
2πih̄(tb − ta )/M
  
i 1 1
× exp [yb (t′ − ta ) + ya (tb − t′ )] Q − (tb − t′ )(t′ − ta )Q2 . (2.733)
h̄ tb − ta 2M

Performing here the integral over ya yields

   
i i 1 ′ 2
exp Q yb exp − (tb − t )Q . (2.734)
h̄ h̄ 2M

The integral over yb in (2.732) leads now to a δ-function (2πh̄)3 δ (3) (Q − q), such that the expo-
nential prefactor in (2.732) is canceled by part of the second factor in (2.734).
In the limit tb − ta → ∞, the integral over t′ produces a δ-function 2πh̄δ(pb Q/M + Q2 /2M ) =
2πh̄δ(Eb − Ea ) which enforces the conservation of energy. Thus we find the well-known Born
approximation

hpb |Ŝ1 |pa i = −2πiδ(Eb − Ea )V (q). (2.735)

In general, we subtract the unscattered particle term (2.731) from (2.729), to obtain a path
integral representation for the T -matrix [for the definition recall (1.477)]:
Z Z
iq2 tb /2Mh̄ 3 −iqyb /h̄
2πh̄iδ(Eb − Ea )hpb |T̂ |pa i ≡ − lim e d yb e d3 ya
tb −ta →∞
 Z tb
i tb
   
i M 2
Z Z
3
 pa 
× D y exp dt ẏ exp − dt V y + t −1 . (2.736)
h̄ ta 2 h̄ ta M
192 2 Path Integrals — Elementary Properties and Simple Solutions

It is preferable to find a formula which does not contain the δ-function of energy conservation as
a factor on the left-hand side. In order to remove this we observe that its origin lies in the time-
translational invariance of the path integral in the limit tb − ta → ∞. If we go over to a shifted time
variable t → t + t0 , and change simultaneously y → y − pa t0 /M , then the path integral remains
the same Rexcept for shifted initial and final times tb + t0 and ta + t0 . In the limit tb − ta → ∞, the
t +t Rt
integrals tab+t00 dt can be replaced again by tab dt. The only place where a t0 -dependence remains
is in the prefactor e−iqyb /h̄ which changes to e−iqyb /h̄ eiqpa t0 /Mh̄ . Among all path fluctuations, there
exists one degree of freedom which is equivalent to a temporal shift of the path. This is equivalent to
an integral over t0 which yields a δ-function 2πh̄δ (qpa /M ) = 2πh̄δ (Eb − Ea ). We only must make
sure to find the relation between this temporal shift and the corresponding measure in the path
integral. This is obviously a shift of the path as a whole in the direction p̂a ≡ pa /|pa |. The formal
way of isolating this degree of freedom proceeds according to a method developed by Faddeev and
Popov37 by inserting into the path integral (2.729) the following integral representation of unity:
|pa | ∞
Z
1= dt0 δ (p̂a (yb + pa t0 /M )) . (2.737)
M −∞
In the following, we shall drop the subscript a of the incoming beam, writing

p ≡ pa , p ≡ |pa | = |pb |. (2.738)

After the above shift in the path integral, the δ-function in (2.737) becomes δ (p̂a yb ) inside the path
integral, with no t0 -dependence. The integral over t0 can now be performed yielding the δ-function
in the energy. Removing this from the equation we obtain the path integral representation of the
T -matrix
p
Z Z
2
hpb |T̂ |pa i ≡ i lim eiq (tb −ta )/8Mh̄ d3 yb δ (p̂a yb ) e−iqyb /h̄ d3 ya
M tb −ta →∞
 Z tb
i tb
    
i M
Z Z
P
× D3 y exp dt ẏ2 exp − dt V y+ t −1 . (2.739)
h̄ ta 2 h̄ ta M

At this point it is convenient to go over to the velocity representation of the path integral (2.723).
This enables us to perform trivially the integral over yb , and we obtain the y version of (2.724).
The δ-function enforces a vanishing longitudinal component of yb . The transverse component of
yb will be denoted by b:
b ≡ yb − (p̂a yb )p̂/a. (2.740)
Hence we find the path integral representation
p
Z
2
hpb |T̂ |pa i ≡ i lim eiq tb /2Mh̄ d2 b e−iqb/h̄
M tb −ta →∞
 Z tb h
i M
Z i
× D3 v exp dt v2 eiχb,p [v] − 1 , (2.741)
h̄ ta 2
where the effect of the interaction is contained in the scattering phase

1 tb
Z  Z tb 
p ′ ′
χb,p [v] ≡ − dt V b + t− dt v(t ) . (2.742)
h̄ ta M t

We can go back to a more conventional path integral by replacing the velocity paths v(t) by
Rt
ẏ(t) = − t b v(t). This vanishes at t = tb . Equivalently, we can use paths z(t) with periodic
boundary conditions and subtract from these z(tb ) = zb .
From hpb |T̂ |pa i we obtain the scattering amplitude fpb pa , whose square gives the differential
cross section, by multiplying it with a factor −M/2πh̄ [see Eq. (1.497)].
37
L.D. Faddeev and V.N. Popov, Phys. Lett. B 25 , 29 (1967).

H. Kleinert, PATH INTEGRALS

2.22 Path Integral Representation of the Scattering Matrix 193

Note that in the velocity representation, the evaluation of the harmonic path integral integrated
over ya in (2.732) is much simpler than before where we needed the steps (2.733), (2.734). After
the Fourier decomposition of V (x) in (2.742), the relevant integral is
 Z tb   R tb
i M 2
Z
3 ′
i
− 2M dtΘ2 (tb−t′)Q2 i ′ 2
D v exp dt v −Θ(tb− t )Q v = e h̄ t
a = e− 2Mh̄ (tb −t )Q . (2.743)
h̄ ta 2

The first factor in (2.734) comes directly from the argument Y in the Fourier representation of the
potential
 Z tb 
p ′ ′
V yb + t− dt v(t )
M t

in the velocity representation of the S-matrix (2.729).

2.22.2 Improved Formulation

2
The prefactor eiq tb /2Mh̄ in Formula (2.741) is an obstacle to taking a more explicit limit tb −ta → ∞
on the right-hand side. To overcome this, we represent this factor by an auxiliary path integral38
over some vector field w(t):
 tb  R tb
i M 2
Z Z
iq2 tb /2Mh̄ 3 i dt Θ(t)w(t)q/h̄
e = D w exp − dt w (t) e ta . (2.744)
h̄ ta 2
h R tb i
−iq b+ dt Θ(t)w(t) /h̄
The last factor changes the exponential e−iqb/h̄ in (2.741) into e ta
. Since b is
Rt
a dummy variable of integration, we can equivalently replace b → bw ≡ b − tab dt Θ(t)w(t) in the
scattering phase χb,p [v] and remain with

p
Z Z
fpb pa = lim d2 b e−iqb/h̄ D3 w
tb −ta →∞ 2πih̄
 Z ∞ 
i M 2
Z
D3 v exp v − w2 [exp (iχbw ,p ) − 1] .


× dt (2.745)
h̄ −∞ 2

The scattering phase in this expression can be calculated from formula (2.742) with the integral
taken over the entire t-axis:

1 ∞
Z  Z tb 
p
χbw ,p [v, w] = − dt V b + t− dt′ [Θ(t′ −t)v(t′ ) − Θ(t′ )w(t′ )] . (2.746)
h̄ −∞ M ta

The fluctuations of w(t) are necessary to correct for the fact that the outgoing particle does not
run, on the average, with the velocity p/M = pa /M but with velocity pb /M = (p + q)/M .
Rt
We may also go back to a more conventional path integral by inserting y(t) = − t b v(t) and
R tb
setting similarly z(t) = − t w(t). Then we obtain the alternative representation

p
Z Z Z
2 −iqb/h̄
fpb pa = lim d be 3
d ya d3 za
tb −ta →∞ 2πih̄
 Z tb 
i M 2
Z Z

h iχ i
× 3 3
D y D z exp dt ẏ − ż2
e bz ,p [y] − 1 , (2.747)
h̄ ta 2

38
See R. Rosenfelder, notes of a lecture held at the ETH Zürich in 1979: Pfadintegrale in der
Quantenphysik , 126 p., PSI Report 97-12, ISSN 1019-0643, and Lecture held at the 7th Int. Conf.
on Path Integrals in Antwerpen, Path Integrals from Quarks to Galaxies, 2002; Phys. Rev. A 79,
012701 (2009).
194 2 Path Integrals — Elementary Properties and Simple Solutions

with
1 tb
Z  p 
χbz ,p [y] ≡ − dt V b + t + y(t) − z(0) , (2.748)
h̄ ta M
where the path integrals run over all paths with yb = 0 and zb = 0. In Section 3.26 this path
integral will be evaluated perturbatively.

2.22.3 Eikonal Approximation to the Scattering Amplitude

To lowest approximation, we neglect the fluctuating variables y(t) and z(t) in (2.748). Since the
integral
 Z tb 
i M 2
Z Z Z Z
3 3 3 3 2


d ya d za D y D z exp dt ẏ − ż (2.749)
h̄ ta 2
in (2.747) has unit normalization [recall the calculation of (2.734)], we obtain directly the eikonal
approximation to the scattering amplitude
p
Z
fpeib pa ≡ d2 b e−iqb/h̄ exp iχei


b,p − 1 , (2.750)
2πih̄
with
1 ∞
Z  p 
χei
b,p ≡ − dt V b + t . (2.751)
h̄ −∞ M
The time integration can be converted into a line integration along the direction of the incoming
particles by introducing a variable z ≡ pt/M . Then we can write
M1 ∞
Z
χei
b,p ≡ − dz V (b + p̂z) . (2.752)
p h̄ −∞
If V (x) is rotationally symmetric, it depends only on r ≡ |x|. Then we shall write the potential
as V (r) and calculate (2.752) as the integral
M1 ∞
Z p 
ei
χb,p ≡ − dz V b2 + z 2 . (2.753)
p h̄ −∞
Inserting this into (2.750), we can perform the integral over all angles between q and b using the
formula Z π  
1 i
dθ exp qb cos θ = J0 (qb), (2.754)
2π −π h̄
where J0 (ξ) is the Bessel function, and find
p
Z
ei
db b J0 (qb) exp iχei


fpb pa = b,p − 1 . (2.755)
ih̄
The variable of integration b coincides with the impact parameter b introduced in Eq. (1.500). The
result (2.755) is precisely the eikonal approximation (1.500) with χeib,p /2 playing the role of the
scattering phases δl (p) of angular momentum l = pb/h̄:
χei
b,p = 2iδpb/h̄ (p). (2.756)

2.23 Heisenberg Operator Approach to Time Evolution

Amplitude
An interesting alternative to the path integral derivation of the time evolution amplitudes of
harmonic systems is based on quantum mechanics in the Heisenberg picture. It bears a close
similarity with the path integral derivation in that it requires solving the classical equations of
motion with given initial and final positions to obtain the exponential of the classical action eiA/h̄ .
The fluctuation factor, however, which accompanies this exponential is obtained quite differently
from commutation rules of the operatorial orbits at different times as we shall now demonstrate.

H. Kleinert, PATH INTEGRALS

2.23 Heisenberg Operator Approach to Time Evolution Amplitude 195

2.23.1 Free Particle

We want to calculate the matrix element of the time evolution operator

(x t|x′ 0) = hx|e−iĤt/h̄ |x′ i, (2.757)

where Ĥ is the Hamiltonian operator

p̂2
Ĥ = H(p̂) = . (2.758)
2M
We shall calculate the time evolution amplitude (2.757) by solving the differential equation
h i
ih̄∂t hx t|x′ 0i ≡ hx|Ĥ e−iĤt/h̄ |x′ i = hx|e−iĤt/h̄ eiĤt/h̄ Ĥ e−iĤt/h̄ |x′ i
= hx t|H(p̂(t))|x′ 0i. (2.759)

The argument contains now the time-dependent Heisenberg picture of the operator p̂. The evalu-
ation of the right-hand side will be based on re-expressing the operator H(p̂(t)) as a function of
initial and final position operators in such a way that all final position operators stand to the left
of all initial ones:

Ĥ = H(x̂(t), x̂(0); t). (2.760)

Then the matrix elements on the right-hand side can immediately be evaluated using the eigenvalue
equations
hx t|x̂(t) = xhx t|, x̂(0)|x′ 0i = x′ |x′ 0i, (2.761)
as being

hx t|H(x̂(t), x̂(0); t)|x̂ 0i = H(x, x′ ; t)hx t|x′ 0i, (2.762)

and the differential equation (2.759) becomes

ih̄∂t hx t|x′ 0i ≡ H(x, x′ ; t)hx t|x′ 0i, (2.763)

or R
−i dt′ H(x,x′ ;t′ )/h̄
hx t|x′ 0i = C(x, x′ )E(x, x′ ; t) ≡ C(x, x′ )e t . (2.764)
The prefactor C(x, x′ ) contains a possible constant of integration resulting from the time integral
in the exponent.
The Hamiltonian operator is brought to the time-ordered form (2.760) by solving the Heisen-
berg equations of motion

dx̂(t) i h i p̂(t)
= Ĥ, x̂(t) = , (2.765)
dt h̄ M
dp̂(t) i h i
= Ĥ, p̂(t) = 0. (2.766)
dt h̄
The second equation shows that the momentum is time-independent:

p̂(t) = p̂(0), (2.767)

so that the first equation is solved by

p̂(t)
x̂(t) − x̂(0) = t , (2.768)
M
196 2 Path Integrals — Elementary Properties and Simple Solutions

which brings (2.758) to

M 2
Ĥ = [x̂(t) − x̂(0)] . (2.769)
2t2
This is not yet the desired form (2.760) since there is one factor which is not time-ordered. The
proper order is achieved by rewriting Ĥ as
M 
Ĥ = 2 x̂2 (t) − 2x̂(t)x̂(0) + x̂2 (0) + [x̂(t), x̂(0)] , (2.770)
2t
and calculating the commutator from Eq. (2.768) and the canonical commutation rule [p̂i , x̂j ] =
−ih̄δij as
ih̄
[x̂(t), x̂(0)] = − Dt, (2.771)
M
so that we find the desired expression
M  D
Ĥ = H(x̂(t), x̂(0); t) = 2 x̂2 (t) − 2x̂(t)x̂(0) + x̂2 (0) − ih̄ .

(2.772)
2t 2t
Its matrix elements (2.762) can now immediately be written down:
M 2 D
H(x, x′ ; t) = (x − x′ ) − ih̄ . (2.773)
2t2 2t
From this we find directly the exponential factor in (2.764)
 
i M D
R ′ 2
E(x, x′ ; t) = e−i dt H(x,x ;t)/h̄ = exp (x − x′ ) − log t . (2.774)
h̄ 2t 2
Inserting (2.774) into Eq. (2.764), we obtain
 
′ ′ 1 iM ′ 2
hx t|x 0i = C(x, x ) exp (x − x ) . (2.775)
tD/2 h̄ 2t
A possible constant of integration in (2.774) depending on x, x′ is absorbed in the prefactor C(x, x′ ).
This is fixed by differential equations involving x:
h i
−ih̄∇hx t|x′ 0i = hx|p̂e−iĤt/h̄ |x′ i = hx|e−iĤt eiĤt/h̄ p̂e−iĤt/h̄ |x′ i = hx t|p̂(t)|x′ 0i.

ih̄∇′ hx t|x′ 0i = hx|e−iĤt/h̄ p̂|x′ ihx t|p̂(0)|x′ 0i. (2.776)

Inserting (2.768) and using the momentum conservation (2.767), these become
M
−ih̄∇hx t|x′ 0i = (x − x′ ) hx t|x′ 0i,
t
M
ih̄∇′ hx t|x′ 0i = (x − x′ ) hx t|x′ 0i. (2.777)
t
Inserting here the previous result (2.775), we obtain the conditions
−i∇C(x, x′ ) = 0, i∇′ C(x, x′ ) = 0, (2.778)
which are solved only by a constant C. The constant, in turn, is fixed by the initial condition
lim hx t|x′ 0i = δ (D) (x − x′ ), (2.779)
t→0
to be r D
M
C= , (2.780)
2πih̄
so that we find the correct free-particle amplitude (2.74)
r D  
′ M iM
hx t|x 0i ≡ exp (x − x′ )2 . (2.781)
2πih̄t h̄ 2t
Note that the fluctuation factor 1/tD/2 emerges in this approach as a consequence of the commu-
tation relation (2.771).

H. Kleinert, PATH INTEGRALS

2.23 Heisenberg Operator Approach to Time Evolution Amplitude 197

2.23.2 Harmonic Oscillator

Here we are dealing with the Hamiltonian operator

p̂2 M ω2 2
Ĥ = H(p̂, x̂) = + x , (2.782)
2M 2
which has to be brought again to the time-ordered form (2.760). We must now solve the Heisenberg
equations of motion

dx̂(t) i h i p̂(t)
= Ĥ, x̂(t) = , (2.783)
dt h̄ M
dp̂(t) i h i
= Ĥ, p̂(t) = −M ω 2 x̂(t). (2.784)
dt h̄
By solving these equations we obtain [compare (2.158)]
ω
p̂(t) = M [x̂(t) cos ωt − x̂(0)] . (2.785)
sin ωt
Inserting this into (2.782), we obtain

M ω2 n 2 2 2
o
Ĥ = [x̂(t) cos ωt − x̂(0)] + sin ωt x̂ (t) , (2.786)
2 sin2 ωt
which is equal to

M ω2  2
Ĥ = x̂ (t) + x̂2 (0) − 2 cos ωt x̂(t)x̂(0) + cos ωt [x̂(t), x̂(0)] . (2.787)
2 sin2 ωt
By commuting Eq. (2.785) with x̂(t), we find the commutator [compare (2.771)]

ih̄ sin ωt
[x̂(t), x̂(0)] = − D , (2.788)
M ω
so that we find the matrix elements of the Hamiltonian operator in the form (2.762) [compare
(2.773)]

M ω2  2 ′2
 D
H(x, x′ ; t) = 2 x + x − 2 cos ωt xx′
− ih̄ ω cot ωt. (2.789)
2 sin ωt 2
This has the integral [compare (2.774)]

M ω h 2 D sin ωt
Z  i
2
dt H(x, x′ ; t) = − x + x′ cos ωt − 2 x x′ − ih̄ log . (2.790)
2 sin ωt 2 ω

Inserting this into Eq. (2.764), we find precisely the harmonic oscillator amplitude (2.177), apart
from the factor C(x, x′ ). This is again determined by the differential equations (2.776), leaving
only a simple normalization factor fixed by the initial condition (2.779) with the result (2.780).
Again, the fluctuation factor has its origin in the commutator (2.788).

2.23.3 Charged Particle in Magnetic Field

We now turn to a charged particle in three dimensions in a magnetic field treated in Section 2.18,
where the Hamiltonian operator is most conveniently expressed in terms of the operator of the
covariant momentum (2.644),
e
P̂ ≡ p̂ − A(x̂), (2.791)
c
198 2 Path Integrals — Elementary Properties and Simple Solutions

as [compare (2.643)]
P̂2
.
Ĥ = H(p̂, x̂) = (2.792)
2M
In the presence of a magnetic field, its components do not commute but satisfy the commutation
rules:
e e eh̄ eh̄
[P̂i , P̂j ] = − [p̂i , Âj ] − [Âi , p̂j ] = i (∇i Aj − ∇j Ai ) = i Bij , (2.793)
c c c c
where Bij = ǫijk BK is the usual antisymmetric tensor representation of the magnetic field.
We now have to solve the Heisenberg equations of motion

dx̂(t) i h i P̂(t)
= Ĥ, x̂ (t) = (2.794)
dt h̄ M
dP̂(t) i h i e eh̄
= Ĥ, P̂(t) = B(x̂(t))P̂(t) + i ∇j Bji (x̂(t)), (2.795)
dt h̄ Mc Mc

where B(x̂(t))P̂(t) is understood as the product of the matrix Bij (x̂(t)) with the vector P̂. In a
constant field, where Bij (x̂(t)) is a constant matrix Bij , the last term in the second equation is
absent and we find directly the solution

P̂(t) = eΩL t P̂(0), (2.796)

where ΩL is a matrix version of the Landau frequency (2.648)

e
ΩL ij ≡ Bij , (2.797)
Mc
which can also be rewritten with the help of the Landau frequency vector
e
!L ≡ B (2.798)
Mc
and the 3 × 3-generators of the rotation group

(Lk )ij ≡ −iǫkij (2.799)

as
ΩL = i L · !L . (2.800)
Inserting this into Eq. (2.794), we find

eΩL t − 1 P̂(0)
x̂(t) = x̂(0) + , (2.801)
ΩL M
where the matrix on the right-hand side is again defined by a power series expansion

eΩL t − 1 t2 t3
= t + ΩL + Ω2L + . . . . (2.802)
ΩL 2 3!

We can invert (2.801) to find

P̂(0) ΩL /2
= e−ΩL t/2 [x̂(t) − x̂(0)] . (2.803)
M sinh ΩL t/2

Using (2.796), this implies

P̂(t) = M N (ΩL t) [x̂(t) − x̂(0)] , (2.804)

H. Kleinert, PATH INTEGRALS

2.23 Heisenberg Operator Approach to Time Evolution Amplitude 199

with the matrix

ΩL /2
N (ΩL t) ≡ eΩL t/2 . (2.805)
sinh ΩL t/2
By squaring (2.804) we obtain

P̂2 (t) M T
= [x̂(t) − x̂(0)] K(ΩL t) [x̂(t) − x̂(0)] , (2.806)
2M 2
where
K(ΩL t) = N T (ΩL t)N (ΩL t). (2.807)
Using the antisymmetry of the matrix ΩL , we can rewrite this as
Ω2L /4
K(ΩL t) = N (−ΩL t)N (ΩL t) = . (2.808)
sinh2 ΩL t/2
The commutator between two operators x̂(t) at different times is, due to Eq. (2.801),
 ΩL t 
i e −1
[x̂i (t), x̂j (0)] = − , (2.809)
M ΩL ij

and
T
!
  i eΩL t − 1 eΩL t − 1
x̂i (t), x̂j (0) + [x̂j (t), x̂i (0)] = − +
M ΩL ΩTL
ij
 ΩL t
− e−ΩL t
  
i e i sinh ΩL t
= − = −2 . (2.810)
M ΩL ij M ΩL ij

Respecting this, we can expand (2.806) in powers of operators x̂(t) and x̂(0), thereby time-ordering
the later operators to the left of the earlier ones as follows:
M  T
x̂ (t)K(ΩL t)x̂(t) − 2x̂T K(ΩL t)x̂(0) + x̂T K(ΩL t)x̂(0)

H(x̂(t), x̂(0)) =
2  
ih̄ ΩL ΩL t
− tr coth . (2.811)
2 2 2
This has to be integrated in t, for which we use the formulas
Ω2L /2 ΩL ΩL t
Z Z
dt K(ΩL t) = dt 2 =− coth , (2.812)
sinh ΩL t/2 2 2

and
 
1 ΩL ΩL t sinh ΩL t/2 sinh ΩL t/2
Z
dt tr coth = tr log = tr log + 3 log t, (2.813)
2 2 2 ΩL /2 ΩL t/2
these results following again from a Taylor expansion of both sides. The factor 3 in the last term
is due to the three-dimensional trace. We can then immediately write down the exponential factor
E(x, x′ ; t) in (2.764):
   
′ 1 iM ′ T ΩL ΩL t ′ 1 sinh ΩL t/2
E(x, x ; t) = 3/2 exp (x−x ) coth (x−x ) − tr log . (2.814)
t h̄ 2 2 2 2 ΩL t/2
The last term gives rise to a prefactor
 −1/2
sinh ΩL t/2
det . (2.815)
ΩL t/2
200 2 Path Integrals — Elementary Properties and Simple Solutions

As before, the time-independent integration factor C(x, x′ ) in (2.764) is fixed by differential

equations in x and x′ , which involve here the covariant derivatives:
h e i h i
−ih̄∇− A(x) hx t|x′ 0i = hx|P̂e−iĤt/h̄ |x′ i = hx|e−iĤt/h̄ eiĤt/h̄ P̂e−iĤt/h̄ |x′ i
c
= hx t|P̂(t)|x′ 0i = L(ΩL t)(x − x′ )hx t|x′ 0i, (2.816)
h e i
ih̄∇′ − A(x) hx t|x′ 0i = hx|e−iĤt/h̄ P̂|x′ i
c
= hx t|P̂(0)|x′ 0i = L(ΩL t)(x − x′ )hx t|x′ 0i. (2.817)

Calculating the partial derivative we find

−ih̄∇hx t|x′ 0i = [−ih̄∇C(x, x′ )]E(x, x′ ; t)+C(x, x′ )[−ih̄∇E(x, x′ ; t)]

 
′ ′ ′ ΩL ΩL t
= [−ih̄∇C(x, x )]E(x, x ; t)+C(x, x )M coth (x−x′ )E(x, x′ ; t).
2 2

Subtracting the right-hand side of (2.816) leads to

 
ΩL ΩL t M
M coth (x − x′ ) − M L(ΩL t)(x − x′ ) = − ΩL (x − x′ ), (2.818)
2 2 2

so that C(x, x′ ) satisfies the time-independent differential equation

 
e M
−ih̄∇ − A(x) − ΩL (x − x′ ) C(x, x′ ) = 0. (2.819)
c 2

A similar equation is found from the second equation (2.817):

 
′ e M
ih̄∇ − A(x) − ΩL (x − x ) C(x, x′ ) = 0.
′
(2.820)
c 2
These equations are solved by
 Z x  
i e M
′
C(x, x ) = C exp d A() + ΩL ( − x ) .
′
(2.821)
h̄ x′ c 2
The contour of integration is arbitrary since the vector field in brackets,
 
e ′ e ΩL e 1
A () ≡ A() + ( − x′ ) = A() − B × ( − x′ ) (2.822)
c c 2 c 2

has a vanishing curl, ∇ × A′ (x) = 0. We can therefore choose the contour to be a straight line
connecting x′ and x, in which case d points in the same direction of x − x′ as  − x′ so that the
cross product vanishes. Hence we may write for a straight-line connection the ΩL -term
e x
 Z 
C(x, x′ ) = C exp i d A() . (2.823)
c x′

Finally, the normalization constant C is fixed by the initial condition (2.779) to have the value
(2.780).
Collecting all terms, the amplitude is
−1/2
e x
  
1 sinh ΩL t/2
Z
hx t|x′ 0i = 3 det exp i d A()
ΩL t/2 c x′
q
2πih̄2 t/M
   
iM ′ T ΩL ΩL t ′
× exp (x − x ) coth (x − x ) . (2.824)
h̄ 2 2 2

H. Kleinert, PATH INTEGRALS

Appendix 2A Baker-Campbell-Hausdorff Formula and Magnus Expansion 201

All expressions simplify if we assume the magnetic field to point in the z-direction, in which
case the frequency matrix becomes
 
0 ωL 0
ΩL =  −ωL 0 0 , (2.825)
0 0 0

so that  
cos ωL t/2 0 0
ΩL t 
cos = 0 cos ωL t/2 0 , (2.826)
2
0 0 1
and  
0 sin ωL t/2 0
sinh ΩL t/2 
= − sin ωL t/2 0 0 , (2.827)
ΩL t/2
0 0 1
whose determinant is
 2
sinh ΩL t/2 sinh ωL t/2
det = . (2.828)
ΩL t/2 ωL t/2
Let us calculate the exponential involving the vector potential in (2.824) explicitly. We choose
the gauge in which the vector potential points in the y-direction [recall (2.636)], and parametrize
the straight line between x′ and x as

 = x′ + s(x − x′ ), s ∈ [0, 1]. (2.829)

Then we find
Z x Z 1
d A() = B(y − y ′ ) ds [x′ + s(x − x′ )] = B(y − y ′ )(x + x′ )
x′ 0
= B(xy − x y ) + B(x′ y − xy ′ ).
′ ′
(2.830)

Inserting this and (2.828) into (2.764), we recover the earlier result (2.668).

Appendix 2A Baker-Campbell-Hausdorff Formula and

Magnus Expansion
The standard Baker-Campbell-Hausdorff formula, from which our formula (2.9) can be derived,
reads
e eB̂ = eĈ , (2A.1)
where
Z 1
Ĉ = B̂ + dtg(ead A t ead B )[Â], (2A.2)
0

and g(z) is the function

∞
log z X (1 − z)n
g(z) ≡ = , (2A.3)
z − 1 n=0 n + 1

and adB is the operator associated with B̂ in the so-called adjoint representation, which is defined
by
adB[Â] ≡ [B̂, Â]. (2A.4)
202 2 Path Integrals — Elementary Properties and Simple Solutions

One also defines the trivial adjoint operator (adB)0 [Â] = 1[Â] ≡ Â. By expanding the exponentials
in Eq. (2A.2) and using the power series (2A.3), one finds the explicit formula
∞
X (−1)n X 1
Ĉ = B̂ + Â + Pn
n=1
n+1 1+ i=1 pi
pi ,qi ;pi +qi ≥1
(adA)p1 (adB) q1
(adA)pn (adB)qn
× ··· [Â]. (2A.5)
p1 ! q1 ! pn ! qn !
The lowest expansion terms are
1 1 1 2 1 1 2

Ĉ = B̂ + Â− 2 adA + adB + 6 (adA) + 2 adA adB + 2 (adB) +. . . [Â]
2
1 1
+ 3 (adA)2 + 12 adA adB + 21 adB adA + (ad B)2 + . . . [Â]

(2A.6)
3
1 1 1
= Â + B̂ + [Â, B̂] + ([Â, [Â, B̂]] + [B̂, [B̂, Â]]) + [Â, [[Â, B̂], B̂]] . . . .
2 12 24
The result can be rearranged to the closely related Zassenhaus formula

eÂ+B̂ = e eB̂ eẐ2 eẐ3 eẐ4 · · · , (2A.7)

where
1
Ẑ2 = [B̂, Â] (2A.8)
2
1 1
Ẑ3 = − [B̂, [B̂, Â]] − [Â, [B̂, Â]]) (2A.9)
3 6
1  1
Ẑ4 = [[[B̂, Â], B̂], B̂] + [[[B̂, Â], Â], B̂] + [[[B̂, Â], Â], Â] (2A.10)
8 24
..
. .

To prove formula (2A.2) and thus the expansion (2A.6), we proceed similar to the derivation
of the interaction formula (1.295) by deriving and solving a differential equation for the operator
function
Ĉ(t) = log(eÂt eB̂ ). (2A.11)
Its value at t = 1 will supply us with the desired result Ĉ in (2A.5). The starting point is the
observation that for any operator M̂ ,

eĈ(t) M̂ e−Ĉ(t) = ead C(t) [M̂ ], (2A.12)

by definition of adC. Inserting (2A.11), the left-hand side can also be rewritten as
Ât B̂ −B̂ −Ât
e e M̂ e e , which in turn is equal to ead A t ead B [M̂ ], by definition (2A.4). Hence we have

ead C(t) = ead A t ead B . (2A.13)

Differentiation of (2A.11) yields

d −Ĉ(t)
eĈ(t)
e = −Â. (2A.14)
dt
The left-hand side, on the other hand, can be rewritten in general as
d −Ĉ(t) ˙
eĈ(t) e = −f (adC(t))[Ĉ(t)], (2A.15)
dt
where
ez − 1
f (z) ≡ . (2A.16)
z

H. Kleinert, PATH INTEGRALS

Appendix 2A Baker-Campbell-Hausdorff Formula and Magnus Expansion 203

This will be verified below. It implies that

˙
f (adC(t))[Ĉ(t)] = Â. (2A.17)

We now define the function g(z) as in (2A.3) and see that it satisfies

g(ez )f (z) ≡ 1. (2A.18)

We therefore have the trivial identity

˙ ˙
Ĉ(t) = g(ead C(t) )f (adC(t))[Ĉ(t)]. (2A.19)

Using (2A.17) and (2A.13), this turns into the differential equation
˙
Ĉ(t) = g(ead C(t) )[Â] = ead A t ead B [Â], (2A.20)

from which we find directly the result (2A.2).

To complete the proof we must verify (2A.15). The expression is not simply equal to
Ĉ(t) ˙ ˙ ˙
−e Ĉ(t)M e−Ĉ(t) since Ĉ(t) does not, in general, commute with Ĉ(t). To account for this
consider the operator
d
Ô(s, t) ≡ eĈ(t)s e−Ĉ(t)s . (2A.21)
dt
Differentiating this with respect to s gives
d  −Ĉ(t)s  d  
∂s Ô(s, t) = eĈ(t)s Ĉ(t) e − eĈ(t)s Ĉ(t)e−Ĉ(t)s
dt dt
Ĉ(t)s ˙ −Ĉ(t)s
= −e Ĉ(t)e
˙
= −ead C(t)s [Ĉ(t)]. (2A.22)

Hence
Z s
Ô(s, t) − Ô(0, t) = ds′ ∂s′ Ô(s′ , t)
0
∞
X sn+1 n ˙
= − (adC(t)) [Ĉ(t)], (2A.23)
n=0
(n + 1)!

from which we obtain

d −Ĉ(t) ˙
Ô(1, t) = eĈ(t) e = −f (adC(t))[Ĉ(t)], (2A.24)
dt
which is what we wanted to prove.
Note that the final form of the series for Ĉ in (2A.6) can be rearranged in many different ways,
using the Jacobi identity for the commutators. It is a nontrivial task to find a form involving the
smallest number of terms.39
The same mathematical technique can be used to derive a useful modification of the Neumann-
Liouville expansion or Dyson series (1.239) and (1.251). This is the so-called Magnus expansion 40 ,
in which one writes Û (tb , ta ) = eÊ , and expands the exponent Ê as
 2 Z tb
i tb
Z t2
1 −i
Z h i
Ê = − dt1 Ĥ(t1 ) + dt2 dt1 Ĥ(t2 ), Ĥ(t1 )
h̄ ta 2 h̄ ta ta

39
For a discussion see J.A. Oteo, J. Math. Phys. 32 , 419 (1991).
40
See A. Iserles, A. Marthinsen, and S.P. Norsett, On the implementation of the method of
Magnus series for linear differential equations, BIT 39, 281 (1999) (http://www.damtp.cam.ac.uk/
user/ai/Publications).
204 2 Path Integrals — Elementary Properties and Simple Solutions

 3 ( tb t3 t2
1 −i
Z Z Z h h ii
+ dt3 dt2 dt1 Ĥ(t3 ), Ĥ(t2 ), Ĥ(t1 )
4 h̄ ta ta ta
)
tb tb tb
1
Z Z Z hh i i
+ dt3 dt2 dt1 Ĥ(t3 ), Ĥ(t2 ) , Ĥ(t1 ) + ... , (2A.25)
3 ta ta ta

which converges faster than the Neumann-Liouville expansion.

Appendix 2B Direct Calculation of the Time-Sliced

Oscillator Amplitude
After time-slicing, the amplitude (2.145) becomes a multiple integral over short-time amplitudes
[using the action (2.192)]
i M (xn − xn−1 )2
  
1 21 2 2
(xn ǫ|xn−1 0) = p exp − ǫω (xn + xn−1 ) . (2B.26)
2πh̄iǫ/M h̄ 2 ǫ 2
We shall write this as
 
i 
a1 (x2n + x2n−1 ) − 2b1 xn xn−1 ,

(xn ǫ|xn−1 0) = N1 exp (2B.27)
h̄
with
  ωǫ 2 
M M
a1 = 1−2 , b1 = ,
2ǫ 2 2ǫ
1
N1 = p . (2B.28)
2πh̄iǫ/M
When performing the intermediate integrations in a product of N such amplitudes, the result must
have the same general form
 
i  2 2

(xN ǫ|xN −1 0) = NN exp aN (xN + x0 ) − 2bN xN x0 . (2B.29)
h̄
Multiplying this by a further short-time amplitude and integrating over the intermediate position
gives the recursion relations
r
iπh̄
NN +1 = N1 NN , (2B.30)
aN + a1
a2N − b2N + a1 aN a2 − b21 + a1 aN
aN +1 = = 1 , (2B.31)
a1 + aN a1 + aN
b1 bN
bN +1 = . (2B.32)
a1 + aN
From (2B.31) we find
a2N = b2N + a21 − b21 , (2B.33)
and the only nontrivial recursion relation to be solved is that for bN . With (2B.32) it becomes
b b
bN +1 = p 1 N , (2B.34)
a1 + b2N − (b21 − a21 )
or
s !
1 1 a1 b 2 − a2
= + 1− 1 2 1 . (2B.35)
bN +1 b1 bN bN

H. Kleinert, PATH INTEGRALS

Appendix 2C Derivation of Mehler Formula 205

We now introduce the auxiliary frequency ω̃ of Eq. (2.163). Then

M
a1 = cos ω̃, (2B.36)
2ǫ
and the recursion for bN +1 reads
s
1 cos ω̃ǫ 2ǫ M 2 sin2 ω̃ǫ
= + 1− . (2B.37)
bN +1 bN M 4ǫ2 b2N

By introducing the reduced quantities

2ǫ
βN ≡ bN , (2B.38)
M
with
β1 = 1, (2B.39)
the recursion becomes
s
1 cos ω̃ǫ sin2 ω̃ǫ
= + 1− 2 . (2B.40)
βN +1 βN βN

For N = 1, 2, this determines

1 p sin 2ω̃ǫ
= cos ω̃ǫ + 1 − sin2 ω̃ǫ = ,
β2 sin ω̃ǫ
s
1 sin 2ω̃ǫ sin2 2ω̃ǫ sin 3ω̃ǫ
= cos ω̃ǫ + 1 − sin2 ω̃ǫ 2 = . (2B.41)
β3 sin ω̃ǫ sin ω̃ǫ sin ω̃ǫ
We therefore expect the general result
1 sin ω̃(N + 1)ǫ
= . (2B.42)
βN +1 sin ω̃ǫ
It is easy to verify that this solves the recursion relation (2B.40). From (2B.38) we thus obtain
M sin ω̃ǫ
bN +1 = . (2B.43)
2ǫ sin ω̃(N + 1)ǫ
Inserting this into (2B.30) and (2B.33) yields
M cos ω̃(N + 1)ǫ
aN +1 = sin ω̃ǫ , (2B.44)
2ǫ sin ω̃(N + 1)ǫ
s
sin ω̃ǫ
NN +1 = N1 , (2B.45)
sin ω̃(N + 1)ǫ

such that (2B.29) becomes the time-sliced amplitude (2.199).

Appendix 2C Derivation of Mehler Formula

Here we briefly sketch the derivation of Mehler’s formula.41 It is based on the observation that the
left-hand side of Eq. (2.297), let us call it F (x, x′ ), is the Fourier transform of the function
2
+k′2 +akk′ )/2
F̃ (k, k ′ ) = π e−(k , (2C.46)
41
See P.M. Morse and H. Feshbach, Methods of Theoretical Physics, McGraw-Hill, New York,
Vol. I, p. 781 (1953).
206 2 Path Integrals — Elementary Properties and Simple Solutions

as can easily be verified by performing the two Gaussian integrals in the Fourier representation
Z ∞Z ∞
dk dk ′ ikx+ik′ x
F (x; x′ ) = e F̃ (k, k ′ ). (2C.47)
−∞ −∞ 2π 2π

We now consider the right-hand side of (2.297) and form the Fourier transform by recognizing the
2
exponential ek /2−ikx as the generating function of the Hermite polynomials42
∞
2 X (−ik/2)n
ek /2−ikx
= Hn (x). (2C.48)
n=0
n!

This leads to
Z ∞ Z ∞
′ 2
+k′2 )/2
F̃ (k, k ′ ) = dx dx′ F (x, x′ )e−ikx−ik x = e−(k
−∞ −∞
Z ∞Z ∞ ∞ X∞ ′
′
X (−ik/2)n (−ik ′ /2)n
′
× dx dx F (x, x ) Hn (x)Hn′ (x). (2C.49)
−∞ −∞ n=0 ′
n! n′ !
n =0

Inserting here the expansion on the right-hand side of (2.297) and using the orthogonality relation
of Hermite polynomials (2.306), we obtain once more (2C.47).

Notes and References

The basic observation underlying path integrals for time evolution amplitudes goes back to the
historic article
P.A.M. Dirac, Physikalische Zeitschrift der Sowjetunion 3, 64 (1933).
He observed that the short-time propagator is the exponential of i/h̄ times the classical action.
See also
P.A.M. Dirac, The Principles of Quantum Mechanics, Oxford University Press, Oxford, 1947;
E.T. Whittaker, Proc. Roy. Soc. Edinb. 61, 1 (1940).

Path integrals in configuration space were invented by R. P. Feynman in his 1942 Princeton thesis.
The theory was published in 1948 in
R.P. Feynman, Rev. Mod. Phys. 20, 367 (1948).

The mathematics of path integration had previously been developed by

N. Wiener, J. Math. Phys. 2, 131 (1923); Proc. London Math. Soc. 22, 454 (1924); Acta Math. 55,
117 (1930);
N. Wiener, Generalized Harmonic Analysis and Tauberian Theorems, MIT Press, Cambridge,
Mass., 1964,
after some earlier attempts by
P.J. Daniell, Ann. Math. 19, 279; 20, 1 (1918); 20, 281 (1919); 21, 203 (1920);
discussed in
M. Kac, Bull. Am. Math. Soc. 72, Part II, 52 (1966).
Note that even the name path integral appears in Wiener’s 1923 paper.

Further important papers are

I.M. Gelfand and A.M. Yaglom, J. Math. Phys. 1, 48 (1960);
S.G. Brush, Rev. Mod. Phys. 33, 79 (1961);
E. Nelson, J. Math. Phys. 5, 332 (1964);
A.M. Arthurs, ed., Functional Integration and Its Applications, Clarendon Press, Oxford, 1975,
42
I.S. Gradshteyn and I.M. Ryzhik, op. cit., Formula 8.957.1.

H. Kleinert, PATH INTEGRALS

Notes and References 207

C. DeWitt-Morette, A. Maheshwari, and B.L. Nelson, Phys. Rep. 50, 255 (1979);
D.C. Khandekar and S.V. Lawande, Phys. Rep. 137, 115 (1986).

The general harmonic path integral is derived in

M.J. Goovaerts, Physica 77, 379 (1974); C.C. Grosjean and M.J. Goovaerts, J. Comput. Appl.
Math. 21, 311 (1988); G. Junker and A. Inomata, Phys. Lett. A 110, 195 (1985).

The Feynman path integral was applied to thermodynamics by

M. Kac, Trans. Am. Math. Soc. 65, 1 (1949);
M. Kac, Probability and Related Topics in Physical Science, Interscience, New York, 1959, Chap-
ter IV.

A good selection of earlier textbooks on path integrals is

R.P. Feynman, A.R. Hibbs, Quantum Mechanics and Path Integrals, McGraw Hill, New York 1965,
L.S. Schulman, Techniques and Applications of Path Integration, Wiley-Interscience, New York,
1981,
F.W. Wiegel, Introduction to Path-Integral Methods in Physics and Polymer Science, World Sci-
entific, Singapore, 1986.
G. Roepstorff, Path Integral Approach to Quantum Physics, Springer, Berlin, 1994.

The path integral in phase space is reviewed by

C. Garrod, Rev. Mod. Phys. 38, 483 (1966).

The path integral for the most general quadratic action has been studied in various ways by
D.C. Khandekar and S.V. Lawande, J. Math. Phys. 16, 384 (1975); 20, 1870 (1979);
V.V. Dodonov and V.I. Manko, Nuovo Cimento 44B, 265 (1978);
A.D. Janussis, G.N. Brodimas, and A. Streclas, Phys. Lett. A 74, 6 (1979);
C.C. Gerry, J. Math. Phys. 25, 1820 (1984);
B.K. Cheng, J. Phys. A 17, 2475 (1984);
G. Junker and A. Inomata, Phys. Lett. A 110, 195 (1985);
H. Kleinert, J. Math. Phys. 27, 3003 (1986) (http://www.physik.fu-berlin.de/~klei-
nert/144).
The caustic phenomena near the singularities of the harmonic oscillator amplitude at tb − ta =
integer multiples of π/ω, in particular the phase of the fluctuation factor (2.169), have been dis-
cussed by
J.M. Souriau, in Group Theoretical Methods in Physics, IVth International Colloquium, Nijmegen,
1975, ed. by A. Janner, Springer Lecture Notes in Physics, 50;
P.A. Horvathy, Int. J. Theor. Phys. 18, 245 (1979).
See in particular the references therein.
The amplitude for the freely falling particle is discussed in
G.P. Arrighini, N.L. Durante, C. Guidotti, Am. J. Phys. 64, 1036 (1996);
B.R. Holstein, Am. J. Phys. 69, 414 (1997).
For the Baker-Campbell-Hausdorff formula see
J.E. Campbell, Proc. London Math. Soc. 28, 381 (1897); 29, 14 (1898);
H.F. Baker, ibid., 34, 347 (1902); 3, 24 (1905);
F. Hausdorff, Berichte Verhandl. Sächs. Akad. Wiss. Leipzig, Math. Naturw. Kl. 58, 19 (1906);
W. Magnus, Comm. Pure and Applied Math 7, 649 (1954), Chapter IV;
J.A. Oteo, J. Math. Phys. 32, 419 (1991);
See also the internet address
E.W. Weisstein, http://mathworld.wolfram.com/baker-hausdorffseries.html.
The Zassenhaus formula is derived in
208 2 Path Integrals — Elementary Properties and Simple Solutions

W. Magnus, Comm. Pure and Appl. Mathematics, 7, 649 (1954); C. Quesne, Disentangling q-
Exponentials, (math-ph/0310038).
For Trotter’s formula see the original paper:
E. Trotter, Proc. Am. Math. Soc. 10, 545 (1958).
The mathematical conditions for its validity are discussed by
E. Nelson, J. Math. Phys. 5, 332 (1964);
T. Kato, in Topics in Functional Analysis, ed. by I. Gohberg and M. Kac, Academic Press, New
York 1987.
Faster convergent formulas:
M. Suzuki, Comm. Math. Phys. 51, 183 (1976); Physica A 191, 501 (1992);
H. De Raedt and B. De Raedt, Phys. Rev. A 28, 3575 (1983);
W. Janke and T. Sauer, Phys. Lett. A 165, 199 (1992).
See also
M. Suzuki, Physica A 191, 501 (1992).
The path integral representation of the scattering amplitude is developed in
W.B. Campbell, P. Finkler, C.E. Jones, and M.N. Misheloff, Phys. Rev. D 12, 12, 2363 (1975).
See also:
H.D.I. Abarbanel and C. Itzykson, Phys. Rev. Lett. 23, 53 (1969);
R. Rosenfelder, see Footnote 38.
The alternative path integral representation in Section 2.18 is due to
M. Roncadelli, Europhys. Lett. 16, 609 (1991); J. Phys. A 25, L997 (1992);
A. Defendi and M. Roncadelli, Europhys. Lett. 21, 127 (1993).

H. Kleinert, PATH INTEGRALS