Lecture 3: PATH INTEGRALS IN QUANTUM MECHANICS

The path integral formalism, which we explain now in the context of quan-
tum mechanics, leads to a representation of physical quantities as averages
with an appropriate weight (real or complex) over a set of paths or trajec-
tories. Thus, it gives a rather concrete meaning to the concept of quantum
fluctuations.
In fact, to a large extent, one could define quantum mechanics directly
in terms of path integrals, quite interesting a viewpoint, but which would
require the introduction of a mathematical formalism more complicated
than differential equations for the study even of simple quantum systems.
Note, however, that when the number of degrees of freedom is large, as in
statistical physics or in quantum field theory, the advantage of the functional
integral formalism becomes overwhelming.
Here instead, for pedagogical reasons, we assume some minimal background
in quantum mechanics, as the notion of Hilbert space (the notation of bras
and kets will be used for vectors), on which act hermitian or unitary oper-
ators corresponding to the physical observables, like position, momentum,
the time evolution operator or the density matrix.
Moreover, the evolution of wave functions is governed by the Schrödinger
equation.
We first describe a general strategy that leads to a representation of the
matrix elements (in the position basis) of some evolution operators by path
integrals. The existence of such a representation rely on two basic properties:
(i) A Markovian evolution, that is, without memory, a characteristic prop-
erty of isolated systems or without influence on their environment.
(ii) The locality of the evolution for short time intervals, a property that
we explain later. The locality of quantum evolution is in correspondence
with the locality of classical evolution.
These two properties are satisfied both by the unitary evolution operator
e−itH/~ (this form assumes that H is a time-independent Hamiltonian) that
describes quantum evolution and the statistical operator e−βH , which is
proportional to the density matrix at thermal equilibrium.
They are also shared by some stochastic processes of diffusion type that
are not related to quantum physics, like Brownian motion, which historically
has led to the introduction of path integrals (Wiener’s integral).
We then construct a path integral representation of the matrix elements
of the statistical operator in the simple example of Hamiltonians of the form
H = p̂2 /2m + V (q̂), where q̂, p̂ are the position and momentum operators,
respectively.
3.1 Position, momentum operators and matrix elements

The notion of locality makes reference to positions. In quantum mechanics,

to each physical observable, like position, is associated a hermitian operator.
We introduce the basis in Hilbert space in which the position operator q̂ is
diagonal. The eigenvalues q of q̂ are real and, in the standard bra and ket
notation, the corresponding eigenvectors (in some extended sense) |qi are
orthogonal:
q̂ |qi = q |qi , hq ′ |qi = δ(q − q ′ ). (3.1)
(δ is Dirac’s function). Moreover, the basis is complete, which implies
Z
dq |qi hq| = 1 . (3.2)

In terms of matrix elements in this basis, using the decomposition (3.2) of

the identity, one can write the product of two operators O1 and O2 as
Z
hq ′′ | O2 O1 |q ′ i = dq hq ′′ | O2 |qi hq| O1 |q ′ i .

85
3.1.1 Momentum operator and Fourier transformation
We now assume that the position operator q̂ is a d-component vector
(q ∈ Rd ) and denote by p̂ the corresponding momentum operator. Their
components have the canonical commutation relations
[q̂α , p̂β ] = i~δαβ 1 .
(1 is the identity operator).
Fourier transformation: convention. Denoting by |epi the vectors of the
basis in which the momentum operator p̂ is diagonal, we define the Fourier
transformation by Z
dd q eiq·p/~ |qi = |e
pi. (3.3)

The matrix elements of the identity operator in this basis are then
p′′ | 1 |e
he p′ i = he
p′′ |e
p′ i = (2π~)d δ (d) (p′′ − p′ ).

86
Therefore, in a product of operators in momentum representation, the in-
tegration measure is dd p/(2π~)d :
Z
′′ ′ dd p ′′ ′
he
p | U2 U1 |e
pi= he
p | U2 |e
p i he
p | U1 |e
p i.
(2π~)d

Moreover, this normalization implies for the wave functions ψ(q) ≡ hq|ψi:
Z Z
1
he
p| ψi ≡ ψ̃(p) = dd q e−iq·p/~ ψ(q) ⇒ ψ(q) = dd p eiq·p/~ ψ̃(p).
(2π~)d

87
3.2 Markov process

We consider a bounded operator in a Hilbert space, U (t, t′ ), t ≥ t′ , which

describes the evolution from time t′ to time t and which satisfies a Markov
property in time:
U (t, t′′ )U (t′′ , t′ ) = U (t, t′ ), for t ≥ t′′ ≥ t′ and U (t′ , t′ ) = 1 . (3.4)
This property, which is also characteristic of some stochastic processes,
means that the evolution associated with the operator U has no memory.
The evolution from time t′′ to time t depends only on the state of the system
at time t′′ , and not on the preceding evolution.
The Markov property (3.4) allows writing U (t′′ , t′ ) as a product, time-
ordered as in equation (3.4), of operators corresponding to time intervals ε
that we choose equal and which can be made arbitrarily small:
n
Y
U (t′′ , t′ ) = U [t′ + mε, t′ + (m − 1)ε], ε = (t′′ − t′ )/n . (3.5)
m=1

88
When the evolution operator then takes the special form U (t, t′ ) ≡ U (t−t′ ),
the evolution is time-translation invariant and the Markov property (3.4)
becomes a semi-group property: U (t)U (t′ ) = U (t + t′ ) with t ≥ 0 , t′ ≥ 0 .

3.2.1 Differentiable evolution operators

We assume now that U (t, t′ ) is differentiable with a continuous derivative.
We set
∂U (t, t′ )
= −K(t) .
∂t t=t′

Differentiating equation (3.4) with respect to t and taking the t′′ = t limit,
one finds
∂U
(t, t′ ) = −K(t)U (t, t′ ) . (3.6)
∂t
When the operator K is time-independent, it is the generator of time trans-
lations, U (t + ε, t) = e−εK and one recognizes in equation (3.5) Trotter’s
formula.

89
The operator K. In these lectures, we meet two types of operators:
If the operator U describes quantum evolution in time, it is unitary and
the operator K = iH/~ where H is the hermitian quantum Hamiltonian.
If the operator K itself is hermitian and time-independent, K = H and
the operator U (β, 0) is the statistical quantum operator, proportional to the
density matrix ρ(β) at thermal equilibrium at a temperature T = 1/β (the
equilibrium can be induced by a weak coupling to a thermal bath):
ρ(β) = U (β, 0)/Z(β), Z(β) = tr U (β, 0)
and Z is the partition function.
Nevertheless, we will still call the evolution variable time (or Euclidean
time), even though, from the viewpoint of quantum evolution, it is an imag-
inary time. Indeed, if one sets β = it/~, one recovers the usual evolution op-
erator of quantum mechanics. The same analytic continuation allows trans-
posing the algebraic part of the calculations that follow to quantum evolu-
tion.

90
3.2.2 Matrix elements
Introducing matrix elements in the position basis, we can rewrite identity
(3.4) for times t3 ≥ t2 ≥ t1 as
Z
hq3 | U (t3 , t1 ) |q1 i = dq2 hq3 | U (t3 , t2 ) |q2 i hq2 | U (t2 , t1 ) |q1 i ,

where we have used the decomposition (3.2) of the identity.

Generalizing this identity, we rewrite equation (3.5) as
Z n−1
Y n
Y
hq ′′ |U (t′′ , t′ )| q ′ i = dqk hqk |U (tk , tk−1 )| qk−1 i (3.7)
k=1 k=1
with the conventions
tk = t′ + kε , q0 = q ′ , qn = q ′′ .
In this expression, we can take the limit n → ∞, or ε → 0, reducing the
evaluation of expression (3.7) to the asymptotic evaluation (but with enough
precision) of matrix elements hq|U (t+ε, t)|q ′ i for infinitesimal time intervals.

91
3.3 Locality of short-time evolution

Local short time evolution means that the support of the matrix elements
hq ′′ |U (t′′ , t′ )|q ′ i vanish for |q ′′ − q ′ | > 0 when t′′ − t′ goes to zero. An explicit
and very relevant example is provided by the random walk where the role
of hq ′′ |U (t′′ , t′ )|q ′ i is played by the transition probability ρ(q ′′ − q ′ ) and, in
macroscopic variables, for t′′ − t′ = ε ≪ 1,
2
ρ(q) ∝ e−q /2ε
.
Assuming that the operator U is differentiable, we have introduced its
derivative −K(t). If U is local, the matrix elements of the operator hq ′′ |K|q ′ i
have a support limited to q ′ = q ′′ . This property is satisfied by differential
operators, which can be represented by a polynomial in the momentum op-
erator p̂, K(t) ≡ K(p̂, q̂; t). In what follows we mainly consider operators
proportional to quantum Hamiltonians of the form p̂2 /2m + V (q̂), which
clearly satisfy the locality condition.

92
3.3.1 Free evolution or Brownian motion
As an illustration, we first consider the example of the statistical operator
corresponding to a free particle of mass m in d dimensions, which is mathe-
matically equivalent to the example of Brownian motion already discussed.
The quantum Hamiltonian of the free motion can be written as

H0 = p̂2 /2m .

The matrix elements of the operator U (β, 0) = e−βH0 are simple in the basis
in which p̂ is diagonal since

′′ ′ (d) ′′ ′ −βp′ 2
he
p | U (β, 0) |e
pi=δ (p − p ) e .

They are related with the position basis by Fourier transformation.

93
Inverting the Fourier transformation, one concludes
Z d
 ′′ ′ 2

d p (q − q ) · p p
hq′′ |U (β, 0)| q′ i = exp i − β
(2π~)d ~ 2m
 d/2 !
2
m m (q′′ − q′ )
= exp − . (3.8)
2πβ 2β ~2
The expression clearly exhibits a locality property: when β → 0, the region
′′ ′ ′′ ′
√
in which hq |U (β, 0)|q i remains finite decreases as |q − q | = O( β).

3.3.2 The statistical operator for short time intervals

From now on, the operator K, which appears in equation (3.6), is identified
with a quantum Hamiltonian of the form
K ≡ H = p̂2 /2m + V (q̂) , (3.9)
(where p, q are d-component vectors) where the potential V is differentiable
and, temporarily, assumed to be time-independent.

94
Then, time-translation invariance implies U (t′′ , t′ ) = U (t′′ − t′ , 0).
When the potential is differentiable, for short time intervals, one can use
the Baker-Campbell-Hausdorff formula at leading order,
−εH −εp̂2 /2m −εV (q̂)+O(ε2 )
U (ε, 0) = e =e e ,

where the term of order ε2 is proportional to the commutator [p̂2 , V (q̂)].

Taking the matrix elements of both sides, one finds
" #
 m d/2 ′′
m (q − q )′ 2
hq′′ | U (ε, 0) |q′ i = exp − 2
− εV (q ′
) .
2πε 2ε~

Basic observation. In the exponent, the most singular term for ε → 0 is

(q − q′ )2 /2ε (which is independent of the potential). This implies that the
support of the matrix element hq|U (β, 0)|q′ i corresponds to values |q′ −q| =
√
O( ε), as in the Brownian motion.

95
 ′
√
Then, for |q − q| = O( ε),
′ 3/2


εV (q ) = εV (q) + O(ε ) = 1
2εV (q) + V (q ) + O(ε3/2 )
′
Z ε


= dt V (q + (q − q )t/ε + O(ε3/2 ).
′ ′
0

Introducing the trajectory

q(t) = (q′ + (q − q′ )t/ε
and the quantity (q̇(t) ≡ dq/dt)
Z ε h
m 2
i
S(ε, 0) = dt 2
q̇ (t) + V q(t) ,
0 2~
one can rewrite the matrix elements of the statistical operator
 m d/2 3/2
′
hq| U (ε, 0) |q i = e−S(ε,0)+O(ε ) .
2πε

96
3.3.3 Path integral representation of the statistical operator
We now combine equation (3.7) in the form,
Z n−1
Y n
Y
hq′′ |U (β, 0)| q′ i = dd qk hqk |U (tk , tk−1 )| qk−1 i ,
k=1 k=1
in the limit n → ∞, ε → 0 at nε = β fixed, where tk = kε, with the short
time evaluation of the matrix elements of the statistical operator.
In terms of the piecewise linear, continuous trajectory,
q(t) = qk + (qk+1 − qk )(t − tk )/ε for tk ≤ t ≤ tk+1 ,
we obtain
 m dn/2 Z n−1
Y
′′ ′
hq |U (β, 0)| q i = lim dd qk e−S(q,ε) , (3.10)
n→∞ 2πε
k=1
n
X √
with S(q, ε) = S(tk , tk−1 ) + O(nε3/2 = β ε).
k=1

97
 q(t) q ′′
qn−1

qk
q1 q2
qk+1 qn−2
′
q qk−1

t
0 t1 t2 tk−1 tk tk+1 tn−2 tn−1 β

Fig. 3.1 – A path contributing to the integral (3.10): qk ≡ q(tk ).

The integration over the variables qk is thus equivalent to an integration

over the points of a piecewise linear path of the kind displayed in Fig. 3.1.

98
Summing over k, one obtains

X Z tk+1
n−1 h m i √
2
S(q, ε) = dt q̇ (t) + V q(t)) + O(β ε)
tk 2~2
k=0
Z β h m i
2
√
= dt 2
q̇ (t) + V q(t)) + O(β ε).
0 2~

In the limit n → ∞, ε → 0 with nε = β fixed, S(q, ε) converges toward the

action Z β h
m 2
i
S(q) ≡ dt 2
q̇ (t) + V q(t) . (3.11)
0 2~
For ~ = 1, we recognize the integral of the Euclidean Lagrangian associated
with the Hamiltonian. The Euclidean Lagrangian differs from the usual
action of classical mechanics by the relative sign between the kinetic and
potential terms.

99
From a formal viewpoint, the Euclidean action corresponds to a motion in
imaginary time and this explains the sign difference.
We then take the formal limit of expression (3.10):

Z q(β)=q′′
hq′′ |U (β, 0)| q′ i = [dq(t)] e−S(q) , (3.12)
q(0)=q′

and call it path integral because the right hand side involves a sum over all
paths satisfying the prescribed boundary conditions, with the weight e−S(q) .
Remarks.
(i) We use the symbol [dq(t)], with brackets, throughout these lectures,
to distinguish path integrals from ordinary integrals.
(ii) We have assumed the potential time-independent but a little more
labour one can show that the path integral representation is also valid for
time-dependent potentials, a situation mainly relevant for real-time.

100
 (iii) In the symbol [dq(t)] is hidden a normalization
 m dn/2
N =
2π~ε
that diverges for n → ∞ but does not depend on the potential. Therefore,
in explicit calculations, path integrals are generally normalized by dividing
them by reference path integrals whose value is already known (the free
motion V ≡ 0, for example).
(iv) In contrast to a more abstract construction of quantum mechanics,
the path integral representation relies on a specific basis, the basis in which
the position operator is diagonal.

101
 Generalization. The generalization of the preceding construction to a sys-
tem with several particles corresponding to a Hamiltonian

X p̂2
a
H= + V (q̂),
a
2m a

is immediate and leads to a path integral

Z q(t′′ )=q′′
hq′′ |U (t′′ , t′ )| q′ i = [dq(t)] e−S(q) ,
q(t′ )=q′

which involves the corresponding action

Z " #
X ma

2
S(q) = dt 2
q̇ a (t) + V q(t) .
a
2~

102
 Euclidean or imaginary-time action. If the action we change variables,
setting t 7→ τ = ~t, the path integral becomes
Z q(τ ′′ )=q′′
hq′′ |U (τ ′′ , τ ′ )| q′ i = [dq(τ )] e−SE (q)/~ ,
q(τ ′ )=q′

where SE (q) is now the Euclidean action,

Z τ ′′ 1 2


SE (q) = dτ 2 mq̇ + V q(τ ) ,
τ′

integral of the Euclidean Lagrangian.

103
3.3.4 Real-time or classical action
Finally, if instead we change variables, setting t 7→ τ = −i~t, we obtain the
matrix elements of the evolution operator,
Z q(t′′ )=q′′
hq′′ |U (t′′ , t′ )| q′ i = [dq(t)] eiA(q)/~ , (3.13)
q(t′ )=q′

where A is the usual classical action, integral of the classical Lagrangian:

Z t′′
1 2


A(q) = dt 2 mq̇ − V q(τ ) .
t′

The expression (3.13) establishes a beautiful relation between classical and

quantum mechanics. In the quantum evolution all paths contribute but they
are weighted with the complex weight eiA/~ . Therefore, paths close to solu-
tions of the classical equations of motion, which thus leave the action sta-
tionary, give the largest contributions to the path integral, and completely
dominate the integral when the action is large compared to ~.

104
3.3.5 Discussion
In the action (3.11), the two terms play quite different roles. The potential
determines the contribution of paths as a function of the value of q(t) at
each time, and determines the physical properties of the quantum system.
R
The kinetic term dt q̇2 , in contrast, determines the space of paths that
contribute to the integral and, thus, is essential for the very existence of
the path integral. It selects paths regular enough, that is, those for which
[q(t + ε) − q(t)]2 /ε remains finite when ε goes to zero. These paths are
typical of Brownian motion or random walk.
The kinetic term really belongs to the integration measure. Moreover, the
explicit expression of the kinetic term, which involves q̇, is somewhat formal
since the typical paths that contribute to the path integral are continuous
(they satisfy a Hölder condition of order 1/2) but not differentiable.

105
However, a more mathematically correct notation would be less intuitive. In
particular, it is clear in expression (3.12) that the maxima of the integrand
correspond to paths that minimize the action (3.11), that is, that satisfy

δS δ2S
= 0 with ≥ 0,
δqi (t) δqi (t1 )δqj (t2 )

in the sense of operators. (δ denotes functional derivatives a generalization

of normal derivatives, a notion defined later). They are thus solutions of
the equation of the classical motion, here Euclidean (or in imaginary time),
which are differentiable functions. This observation is specially relevant for
~ → 0 where classical solutions are saddle points or stationary points of
the integrand in the path integral. Though only non-differentiable paths
contribute to a path integral, in the semi-classical limit ~ → 0, the leading
contributions come from paths close to differentiable classical paths.

106
3.4 Quantum partition function

We discuss now one of the simplest quantities: the quantum partition func-
tion. The potential is again assumed to be time-independent.
Trace of an operator. The trace of an operator O defined in terms of
matrix elements in the position basis is given by:
Z Z Z
tr O = tr dq |qi hq| O = dq tr (|qi hq| O) = dq hq| O |qi , (3.14)

where the completeness relation (3.2) has been used.

Partition function. The quantum partition function Z(β) = tr e−βH (β
is the inverse temperature) has a path integral representation that follows
directly from the representation of the matrix elements of the statistical
operator e−βH . Indeed, using relation (3.14),
Z
Z(β) = tr e−βH ≡ tr U (β, 0) = dq hq| U (β, 0) |qi .

107
Introducing the path integral representation (3.12) of hq′′ | U (β, 0) |q′ i for
q′ = q(0) and q′′ = q(β), one finds that the paths that contribute to the
partition function satisfy the periodic boundary conditions q(0) = q(β),
and one integrates over all values of q(0). We express these conditions by
the notation Z
Z(β) = [dq(t)] exp [−S(q)] . (3.15)
q(0)=q(β)

These conditions define closed paths, which also correspond to the set of
periodic functions q(t) with period β.
Low temperature limit. The ground state energy E0 of the Hamiltonian
can be derived from the partition function by studying the low temperature,
β → ∞ limit since
1
lim ln Z(β) = −E0 .
β→∞ β

This remark is specially relevant when the structure of the vacuum is not
trivial.

108
Note that in the path integral one can impose boundary conditions at t = 0
and t = β as well as at t = ±β/2 because the action is time-translation
invariant. However, in the limit β → ∞, in the first case one is led to
integrate over all paths with t ≥ 0 and an initial condition q(0) = 0, while in
the latter case the resulting formalism is explicitly time-translation invariant
since one integrates from −∞ to +∞. This formalism, clearly, is simpler.

109
3.5 Explicit calculation: Gaussian path integrals

We have defined path integrals as formal limits of integrals with discrete

times. One could thus fear that any calculation of a path integral would
require returning to its definition as a limit of discrete time integrals. This
is fortunately not the case; most calculations make no reference to the lim-
iting process, and this gives a practical justification to the introduction of
this new mathematical entity. This will be illustrated in this section, first
with the example of free motion, then with the harmonic oscillator, both
corresponding to Gaussian path integrals.

3.5.1 Free motion

In the example of free motion V ≡ 0 (in one dimension for simplicity and
with ~ = 1)), the Euclidean action is the simple quadratic form in q(t),
Z β
1 2
S(q) = 2 m q̇ (t)dt.
0

110
The corresponding path integral (3.12),
Z q(β)=q ′′
hq ′′ |U (β, 0)| q ′ i = [dq(t)] exp [−S(q)] ,
q(0)=q ′

is Gaussian and thus can be calculated explicitly. We first change variables,

q(t) 7→ r(t), (a translation at each time t):

q(t) = qc (t) + r(t),

where the function qc is determined below but satisfies the boundary con-
ditions
qc (0) = q ′ , qc (β) = q ′′ ⇒ r(0) = r(β) = 0 . (3.16)

The action becomes

Z β
S(qc + r) = S(qc ) + S(r) + m dt q̇c (t)ṙ(t) .
0

111
Since r(0) = r(β) = 0, the term linear in r can be written as
Z t′′ Z t′′
q̇c (t)ṙ(t)dt = − q̈c (t)r(t)dt . (3.17)
t′ t′

The linear term vanishes if the function qc (t) is a solution of the equation
of the free classical motion:

mq̈c (t) = 0 .

The solution that satisfies the conditions (3.16) is

t ′′
qc (t) = q ′ + (q − q ′ ).
β

One infers
m ′′
S(qc ) = (q − q ′ )2 .
2β

112
Since the change of variables is a translation, the Jacobian of the transfor-
mation is equal to 1 and [dq(t)] = [dr(t)]. We thus obtain
 
2 m ′′
hq ′′ | e−β p̂ /2m
|q ′ i = N exp − (q − q ′ )2 ,
2β

where Z " Z #
β
m
N = [dr(t)] exp − dt ṙ 2 (t)
2 0

with r(0) = r(β) = 0.

Since the normalization N does not depend on q ′ , q ′′ , the expression
is consistent with the exact result. The path integral N is the value of
2
hq ′′ | e−β p̂ /2m |q ′ i for q ′ = q ′′ = 0. To determine it, one must either return
to discrete time intervals, or use some independent information.

113
3.5.2 The quantum harmonic oscillator
The Hamiltonian of the one-dimensional harmonic oscillator can be written
as
1 2 1
H0 = p̂ + mω 2 q̂ 2 , (3.18)
2m 2
where the constant 2π/ω is the period of the classical oscillations. The
corresponding Euclidean action is
Z β 1 2 1 2 2

S0 (q) = 2 mq̇ (t) + 2 mω q (t) dt , (3.19)
0

which, again, is a quadratic form in q(t). Therefore, the path integral (3.12),
Z q(β)=q ′′
hq ′′ |U0 (β, 0)| q ′ i = [dq(t)] e−S0 (q) , (3.20)
q(0)=q ′

remains Gaussian.

114
It is convenient to divide the calculation into several steps. The easiest part,
and to some extent the most useful one, is the determination of the explicit
dependence on the boundary conditions, that is, on q ′ , q ′′ .
One changes variables q(t) 7→ r(t) (a translation at each time t), setting

q(t) = qc (t) + r(t),

where the functions qc (t) and r(t) satisfy the boundary conditions (3.16)
and qc (t) is determined below. The action becomes
Z β
2


S0 (qc + r) = S0 (qc ) + S0 (r) + m dt q̇c (t)ṙ(t) + ω qc (t)r(t) .
0

One integrates by parts the term proportional to ṙ q̇c as in equation (3.17).

One then chooses the function qc (t), solution of the equation

−mq̈c + mω 2 qc = 0 . (3.21)

The term linear in r(t) then cancels.

115
One notices that qc (t) satisfies the classical motion equation derived from
the action S0 corresponding to a motion in an inverted harmonic potential
− 21 mω 2 q 2 (the − sign reflecting the Euclidean time).
The action (3.19) then reduces to the sum of two terms,

S0 (q) = S0 (qc ) + S0 (r),

where S0 (qc ) is the classical action evaluated on the trajectory qc . The path
integral becomes (the Jacobian is 1)

hq ′′ | U0 (β, 0) |q ′ i = N (ω, β) e−S0 (qc ) (3.22)

with
Z " Z #
β

1 2
N (ω, β) = [dr(t)] exp − dt 1
2 m ṙ (t) + 12 mω 2 r 2 (t) , (3.23)
~ 0

where the paths now satisfy the boundary conditions r(0) = r(β) = 0.

116
The classical action. One can write the solution of equation (3.21) as
1  ′
′′


qc (t) = q sinh ω(β − t) + q sinh ωt .
sinh(ωβ)
In expression (3.19), an integration by parts,
Z Z
q̇ 2 dt = q q̇ − q q̈ dt ,

combined with equation (3.21), then simplifies the calculation of the classical
action S0 (qc ). One finds
mω  ′2 ′′2

′ ′′

S0 (qc ) = q +q cosh ωβ − 2q q .
2 sinh ωβ

Normalization. The remaining Gaussian integral (3.23) no longer depends

on q ′ , q ′′ and gives a simple normalization. Since, as expected, a direct cal-
culation involves an infinite factor, for essentially pedagogical reasons, we
postpone its calculation.

117
The complete expression is
 1/2
mω
hq ′′ | U0 (β, 0) |q ′ i = e−S0 (qc ) .
2π sinh ωβ
By contrast, the matrix density at thermal equilibrium is independent of
the normalization since
1
hq ′′ | ρ(β) |q ′ i = hq ′′ | U0 (β, 0) |q ′ i with Z0 (β) = tr U0 (β, 0).
Z0 (β)
Then,
Z r
 2
 π
Z0 (β) = N dq exp −mω tanh(ωβ/2)q = N
mω tanh(ωβ/2)

and thus
′′ ′[mω tanh(ωβ/2)]1/2 −S0 (qc )
hq | ρ(β) |q i = 1/2
e .
π

118
3.6 Correlation functions: generating functional

In the path integral (3.15), the integrand e−S(q) defines a positive measure
for paths. To this measure correspond expectation values, defined by
Z Z
1
hF(q)i ≡ [dq(t)]F(q) e−S(q) with Z = [dq(t)] e−S(q) .
Z
The normalization Z is chosen in such a way that h1i = 1.

3.6.1 Correlation functions

Moments of the measure of the form
Z
hq(t1 )q(t2 ) . . . q(tn )i = Z −1 [dq(t)]q(t1 )q(t2 ) . . . q(tn ) exp[−S(q)], (3.24)

where h•i means expectation value, are called correlation functions: a cor-
relation function that depends on n different times is called an n-point
function.

119
On a finite time interval β, the path integral definition requires some bound-
ary conditions. For simplicity, we choose in most of these lectures periodic
boundary conditions: q(β/2) = q(−β/2). Then, the normalization Z is the
quantum partition function.
Of course, it is possible to define correlation functions corresponding to
other boundary conditions in the path integral but we will not discuss them
here, all explicit expressions in the case of Gaussian integrals being more
complicated.
A physical interpretation of these correlation functions will be given later.
Here, they appear only as useful mathematical quantities.

120
3.6.2 Generating functionals
In the discussion of algebraic properties of correlation functions, generating
functionals and functional derivatives provide very useful tools. We define
them now.
We consider an infinite set {F (n) (t1 , . . . , tn )}, n = 0, 1, . . ., of symmetric
functions of their arguments. We introduce an additional function f (t) of
one variable and the following formal series in f :
X∞ Z
1
F(f ) = dt1 . . . dtn F (n) (t1 , . . . , tn )f (t1 ) . . . f (tn ).
n=0
n!

The sum F(f ) is a generating functional of the set of functions F (n) .

In fact, more generally, we will admit for F (n) ’s also mathematical distri-
butions (like the Dirac δ-function and its derivatives). Then, the function
f (t) must belong to the corresponding set of test functions and, thus, must
be sufficiently differentiable.

121
The correlation functions that we have defined,

Z (n) (t1 , . . . , tn ) ≡ hq(t1 ) . . . q(tn )i ,

are symmetric functions of their arguments. A generating functional of cor-

relation functions is then
X 1 Z
Z(f ) = dt1 . . . dtn Z (n) (t1 , . . . , tn )f (t1 ) . . . f (tn )
n=0
n!
X 1 Z
= dt1 . . . dtn hq(t1 ) . . . q(tn )i f (t1 ) . . . f (tn ).
n=0
n!

The latter series can be formally summed, by exchanging integrals and ex-
pectation value (because it is a linear operation). One finds the simple for-
mula  Z 
Z(f ) = exp dt q(t)f (t) . (3.25)

122
3.6.3 Functional derivative
To recover the functions F (n) from the generating functional F(f ), one can
use functional differentiation, an operation that we now define.
Functional differentiation with respect to a function f (t), which we denote
by δ/δf (t) to distinguish it from normal derivatives, is defined by the usual
algebraic properties, linearity and Leibnitz rule:

δ δ δ
[F1 (f ) + F2 (f )] = F1 (f ) + F2 (f ),
δf (t) δf (t) δf (t)
δ δ δ
[F1 (f )F2 (f )] = F1 (f ) F2 (f ) + F2 (f ) F1 (f )
δf (t) δf (t) δf (t)

together with
δ
f (t) = δ(t − u),
δf (u)
where δ(t) is Dirac’s function (more accurately distribution).

123
 For example, the first derivative of F(f ) is
X∞ Z
δ 1
F(f ) = dt1 . . . dtn F (n+1) (u, t1 , . . . , tn )f (t1 ) . . . f (tn ).
δf (u) n=0
n!

Differentiating now p times and setting f ≡ 0, one obtains

( p ! )
Y δ
(p)
F (t1 , . . . , tp ) = F .
i=1
δf (t i )
f ≡0

In the example (3.25), these identities become

 Z 
δZ(f )
= q(t1 ) exp dt q(t)f (t) ,
δf (t1 )
and, more generally,
p
! * p Z +
Y δ Y
Z(f ) = q(ti ) exp dt q(t)f (t) .
i=1
δf (ti ) i=1

124
In the limit f ≡ 0, one thus obtains
* p
+ p
!
Y Y δ
q(ti ) = Z(f ) .
i=1 i=1
δf (ti )
f ≡0

Remark. This formalism also applies when the F (n) are not functions in
the strict sense but also distributions. For example,
Z
δ df (t) δ d
= dv δ(t − v)f (v)
δf (u) dt δf (u) dt
d
= δ(t − u),
dt

where δ(t) is Dirac’s δ-function. With this extension, the classical equa-
tions of motion in the Lagrangian formalism can be obtained by functional
differentiation of the action.

125
 For example, the functional derivative of the classical action
Z h i

2

A(q) = dt 12 q̇(t) − V q(t)

is
Z  
δA d ′ ′


= dt q̇(t) δ(t − τ ) − V (q)δ(t − τ ) = −q̈(τ ) − V q(τ ) .
δq(τ ) dt

The equation of the classical motion can thus be written as

δA(q)
= 0.
δq(τ )

126
3.7 Correlation functions and time-ordered products

We consider the n-point function

Z
Z (n) (t1 , t2 , . . . , tn ) = [dq] q(t1 ) . . . q(tn ) e−S(q) , (3.26)

with periodic boundary conditions q(β/2) = q(−β/2 and

Z +∞
1 2

S(q) = dt 2 mq̇ (t) + V q(t) .
−∞

The corresponding correlation function is then

hq(t1 ) . . . q(tn )i = Z (n) (t1 , . . . , tn )/Z(β) ,

where Z(β) = tr e−βH is the partition function. We then order times:

−β/2 ≤ t1 ≤ t2 ≤ · · · ≤ tn ≤ β/2 . (3.27)

127
We decompose the time interval (−β/2, β/2) into (n + 1) subintervals,
(−β/2, t1 ), (t1 , t2 ),..., (tn , +β/2).
The total action is the sum of the corresponding contributions:
n+1
X Z τ′
′
1 2 
S(q) = S(ti−1 , ti ) with S(τ, τ ) = 2 mq̇ + V (q) dt
i=1 τ

and t0 = −β/2 , tn+1 = +β/2.

We rewrite the path integral (3.26) with the help of the identity
Yn Z Yn
q(ti ) = dqi δ [q(ti ) − qi ] qi .
i=1 i=1
The path integral then factorizes into a product of path integrals corre-
sponding to the different subintervals.
Returning to the very definition of the path integral (equation (3.12)),
using the ordering (3.4), we see that we can rewrite the n-function as
h i
Z (n) = tr e−βH etn H q̂ e−(tn −tn−1 )H q̂ . . . e−(t2 −t1 )H q̂ e−t1 H .

128
For example, for n = 1, one obtains (t0 = −β/2, t2 = β/2, q0 = q(±β/2)
Z Z Z q0


Z (1) = dq0 dq1 q1 [dq] e−S(t0 ,t1 ) δ q(t1 ) − q1 [dq] e−S(t1 ,t2 )
q0
Z
= dq0 dq1 q1 hq1 | e−(t1 +β/2)H |q0 i hq0 | e−(β/2−t1 )H |q1 i .

Then, Z Z
dq0 |q0 i hq0 | = 1 , dq1 |q1 i q1 hq1 | = q̂

Therefore, h i
(1) −(β/2−t1 )H −(t1 +β/2)H
Z = tr e q̂ e
 −βH t1 H −t1 H 
= tr e e q̂ e .
Quite generally, introducing the the Heisenberg representation in imaginary
time of the operator q̂,
Q(t) = etH q̂ e−tH ,

129
we can rewrite Z (n) as
(n)
 −βH 
Z (t1 , t2 , . . . , tn ) = tr e Q(tn ) . . . Q(t1 ) . (3.28)

The order of the operators in the right hand side reflects the time-ordering
(3.27).
We now introduce a time-ordering operator T, which to a set of time-
dependent operators A1 (t1 ),...,Al (tl ) associates the time-ordered product
(T-product) of these operators. For example, for l = 2,

T [A1 (t1 )A2 (t2 )] = A1 (t1 )A2 (t2 )θ(t1 − t2 ) + A2 (t2 )A1 (t1 )θ(t2 − t1 ).

At equal times the expression is not defined if the operators then do not
commute.
We can then rewrite expression (3.28), irrespective now of the order be-
tween the times t1 , . . . , tn , as
(n)
 −βH
Z (t1 , t2 , . . . , tn ) = tr e T [Q(t1 )Q(t2 ) . . . Q(tn )] . (3.29)

130
Therefore,

−1

hq(t1 )q(t2 ) . . . q(tn )i = Z (β) tr e−βH T [Q(t1 )Q(t2 ) . . . Q(tn )] .

After continuation to real time ti 7→ iti /~, but β remaining real, Q(t) be-
comes the usual Heisenberg representation of the operator q̂,

Q(t) = eitH q̂ e−itH

and one obtains the thermal average of the n-point function at temperature
1/β.

131
The zero temperature limit. We now assume that the Hamiltonian H has a
unique normalizable ground state |0i. In the limit β → ∞, the statistical
operator becomes a projector onto the ground state,
e−βH ∼ e−βE0 |0ih0|
β→∞

and one obtains

hq(t1 )q(t2 ) . . . q(tn )i = h0| T [Q(t1 )Q(t2 ) . . . Q(tn )] |0i . (3.30)
For example, for the two-point function, the relation becomes
hq(t1 )q(t2 )i = h0| q̂ e−|t2 −t1 |(H−E0 ) q̂ |0i . (3.31)
After continuation to real time, we find a relation between the path integral
and the ground state (or vacuum in the quantum field theory language) ex-
pectation value of the time ordered product of Heisenberg operators. These
expectation values generate Green’s functions from which one can, for in-
stance, calculate scattering amplitudes, as we shall see later.

132
3.8 General Gaussian path integral and correlation functions

We now show how to derive all expectation values with a Gaussian weight,
here correlation functions, from a Gaussian integral where a linear term is
added to the quadratic form.

3.8.1 General Gaussian path integral

We consider the Gaussian path integral
Z
tr UG (β/2, −β/2; b) = [dq(t)] exp[−SG (q, b)] (3.32)

with periodic boundary conditions: q(β/2) = q(−β/2) and

Z β/2
SG (q, b) = S0 (q) − dt b(t)q(t)
−β/2
Z β/2 1 2 1 2 2

= dt 2 mq̇ (t) + 2 mω q (t) − b(t)q(t) . (3.33)
−β/2

133
Due to periodic boundary conditions, we can extend q(t) and b(t) as periodic
functions of t of period β and t as a cyclic variable. This also implies time
translation invariance on the circle.
Calculation. The integral, which is a functional of b(t), can be calculated
by first eliminating the linear term in SG (q, b).
Again, the calculation is based on a change variables, a translation of the
path q(t) 7→ r(t) with

q(t) = qc (t) + r(t), qc (β/2) = qc (−β/2) ⇒ r(β/2) = r(−β/2), (3.34)

where the function qc (t) is determined below. Then,

SG (q, b) = S0 (r) + SG (qc , b) + Slin. (r, b),

where Slin. , the linear term in r, is

Z β/2
 2

Slin. (r, b) = dt mq̇c (t)ṙ(t) + mω qc (t)r(t) − b(t)r(t) .
−β/2

134
One integrates by parts q̇c ṙ, taking into account r(β/2) = r(−β/2),
Z β/2 Z β/2
dt q̇c (t)ṙ(t) = [q̇c (β/2) − q̇c (−β/2)] r(β/2) − dt q̈c (t)r(t).
−β/2 −β/2

One obtains
Z β/2  2

Slin. (r, b) = dt −mq̈c (t) + mω qc (t) − b(t) r(t)
−β/2

+ mr(β/2) [q̇c (β/2) − q̇c (−β/2)] .

The coefficient of the term linear in r vanishes if the function qc (t) is a

solution of the classical equation of motion

−q̈c (t) + ω 2 qc (t) = b(t)/m

and satisfies q̇c (β/2) = q̇c (−β/2) (in addition to qc (β/2) = qc (−β/2)).

135
The solution can be written as (time translation invariance is used)
Z β/2
1
qc (t) = ∆(t − u)b(u)du . (3.35)
m −β/2

The function ∆(t) is the solution of the equation (δ(t) is Dirac’s function)

¨
−∆(t) + ω 2 ∆(t) = δ(t) (3.36)

with the periodic boundary conditions, ∆(β/2) = ∆(−β/2), ˙

∆(β/2) =
˙
∆(−β/2). One finds

1

∆(t) = cosh ω(β/2 − |t|) . (3.37)
2ω sinh(ωβ/2)

136
To verify the result directly one needs the relations between distributions,
d d
|t| = sgn(t), sgn(t) = 2δ(t),
dt dt
where sgn(t) is the sign function, sgn(t) = 1 for t > 0, sgn(t) = −1 for t < 0.
The function ∆(t), which plays an essential role in perturbative expan-
sions around the harmonic oscillator, is also called propagator. Its limit
when β → ∞ (assuming ω > 0) is the solution that decreases at infinity,
Z +∞
1 −ω|t| 1 eiκt
∆(t) = e = dκ 2 2
. (3.38)
2ω 2π −∞ κ +ω
Note that the kernel ∆(t) has no ω = 0 limit.
The kernel ∆(t1 − t2 ) is the inverse of the differential operator −d2t + ω 2
with periodic boundary conditions. It depends only on the difference t2 − t1
because periodic boundary conditions are time-translation invariant. We
have already pointed out that the interval [−β/2, β/2] can be identified
with a circle and time with an angular variable.

137
3.8.2 Remark
A general remark simplifies the calculation of the classical action. One cal-
culates the action corresponding to λqc (t):
Z β/2  2
1 2 1 2 2

SG (λqc , b) = dt λ 2 m q̇c (t) + 2 mω qc (t) − λb(t)qc (t) .
−β/2

Since qc is a solution of the classical equation, the action SG (λqc , b) is sta-

tionary at the solution, that is for λ = 1:

d
SG (λqc , b) = 0.
dλ λ=1

One infers
Z β/2 Z β/2
1 1 
dt 2 mq̇c2 (t) + 21 mω 2 qc2 (t) = dt b(t)qc (t).
−β/2 2 −β/2

138
Thus,
Z β/2 Z
1 1
SG (qc , b) = − dt qc (t)b(t) = − dt du b(t)∆(t − u)b(u)
2 −β/2 2m B

with B ≡ t, u ∈ [−β/2, β/2].

The remaining integration over the path r(t) yields a normalization that
is identical to the partition function of the harmonic oscillator, Z0 (β) =
tr e−βH0 (H0 is the Hamiltonian of the harmonic oscillator). The final ex-
pression is

ZG (β, b) = tr UG (β/2, −β/2; b) = Z0 (β) e−SG (qc ,b)

 Z 
1
= Z0 (β) exp du dv b(v)∆(v − u)b(u) . (3.39)
2m B

139
3.8.3 Gaussian correlation functions. Wick’s theorem
We now introduce Gaussian expectation values, which we denote by h•i0 ,
corresponding to the normalized Gaussian distribution e−S0 /Z0 (β) with
periodic boundary conditions.
The functional (3.32), divided by the partition function Z0 (β) which can
then be written as
* "Z #+
β/2
ZG (β, b)Z0−1 (β) = exp dt b(t)q(t) ,
−β/2
0

generates the corresponding correlation functions (equation (3.25)). Acting

by functional differentiation (functional derivatives have been defined in
section 3.6.3) on both sides, in the limit b ≡ 0 one obtains
 
Yℓ
δ 
Z0−1 (β)  ZG (β, b) = hq(t1 )q(t2 ) . . . q(tℓ )i0 . (3.40)
j=1
δb(t j )
b≡0

140
Replacing ZG (b, β) by the explicit result (3.39) in the left hand side of
equation (3.40), one then finds
Yℓ  Z 
δ 1
hq(t1 ) . . . q(tℓ )i0 = exp du dv b(v)∆(v − u)b(u) .
j=1
δb(t j ) 2m b≡0
(3.41)
In particular, the second derivative yields the two-point correlation function
δ2 1
hq(t)q(u)i0 = Z0−1 (β) ZG (β, b) = ∆(t − u). (3.42)
δb(t)δb(u) b≡0 m

Wick’s theorem. Quite generally, each functional differentiation of the

expression (3.41) generates a factor b:
 Z 
δ 1
exp du dv b(v)∆(v − u)b(u)
δb(t1 ) 2m
Z  Z 
1 1
= du1 ∆(t1 − u1 )b(u1 ) exp du dv b(v)∆(v − u)b(u) .
m 2m

141
In the limit b ≡ 0, the only surviving terms correspond to pairings of
functional differentiations. We find a characteristic property of the centred
Gaussian measure: all correlation functions can be expressed in terms of the
two-point function in a form expressed by Wick’s theorem:
X 1
hq(t1 )q(t2 ) . . . q(tℓ )i0 = ∆(tP1 − tP2 ) . . . ∆(tPℓ−1 − tPℓ )
mℓ
P {1,2,...,ℓ}
X
= hq(tP1 )q(tP2 )i0 · · · q(tPℓ−1 )q(tPℓ ) 0
,(3.43)
P {1,2,...,ℓ}

where P {1, 2, . . . , ℓ} are all possible (unordered) pairings of {1, 2, . . . , ℓ}.

For example,

hq(t1 )q(t2 )q(t3 )q(t4 )i = hq(t1 )q(t2 )ihq(t3 )q(t4 )i + hq(t1 )q(t3 )ihq(t2 )q(t4 )i
+ hq(t1 )q(t4 )ihq(t3 )q(t2 )i.

142
Regularity of generic paths revisited. The preceding results allow calculating
the expectation value of the quantity
D
2 E 2
q(t + ε) − q(t) = [∆(0) − ∆(ε)] .
0 m

For ε → 0, one finds

D
2 E 1
q(t + ε) − q(t) ∼ |ε| . (3.44)
0 m

This result confirms that generic trajectories contributing to the path inte-
gral are not differentiable (though continuous), and that the behaviour of
q(t + ε) − q(t) for ε → 0 does not depend on the potential but only on the
kinetic term in the action. It also implies that the average action hS(q)i0 is
infinite, because the kinetic term diverges.

143
3.9 Quantum harmonic oscillator: the partition function

As an exercise, we can now complete the calculation of the partition function

Z0 (β) of the harmonic oscillator, which appears in the normalization of
correlation functions in equations (3.24).
For this purpose, we use the important property that while the partition
function, which is just a normalization, is not an observable, its logarith-
mic derivatives with respect to any parameter, are. Indeed, if σ is such a
parameter, Z
∂ ∂S0
Z0 (β) = − [dq] e−S0 (q) .
∂σ ∂σ
Dividing both sides by Z0 (β), we obtain
 
∂ ∂S0
ln Z0 (β) = − .
∂σ ∂σ 0
Quite generally, the quantity W = ln Z is called the free energy in the
context of statistical physics and will play an important role later.

144
We now determine the dependence of the partition function on the param-
eter ω. Differentiating the path integral (3.20) with boundary conditions
(3.15), one obtains
  Z β/2
∂ ∂S0
ln Z0 (β) = − = −mω dt q 2 (t) 0 ,
∂ω ∂ω 0 −β/2

where h•i0 stands for Gaussian expectation value with the weight e−S0 (q) /Z0 .
Thus,
∂ β cosh(ωβ/2)
ln Z0 (β) = −βω∆(0) = − , (3.45)
∂ω 2 sinh(ωβ/2)
where the explicit form (3.37) of ∆(t) has been used. We infer
1
Z0 (β) = N ′ .
sinh(βω/2)
The constant N ′ is a pure number. Since in the limit β → ∞, Z0 ∼ e−βE) ,
one concludes N ′ = 21 . This then yields the exact partition function.

145
3.10 Perturbed harmonic oscillator

We now consider a potential that is the sum of a harmonic potential and a

perturbation. The Hamiltonian can be written as
1 2 m 2 2
H= p + ω q + VI (q). (3.46)
2m 2
We assume that the perturbation is a polynomial in the variable q:
X
VI (q) = vn q n .
n=1

The corresponding partition function is given by (we still set ~ = 1)

Z ( Z )
β/2 

1 2 1 2 2
Z(β) = [dq] exp − 2 mq̇ (t) + 2 mω q (t) + VI q(t) dt
−β/2
(3.47)
with q(−β/2) = q(β/2).

146
3.10.1 Perturbative expansion
The integrand (3.47) can be expanded in powers of VI (q) and this leads to
*Z k +
Z(β) X (−1) k

= dt VI q(t)
Z0 (β) k!
k=0 0
X (−1)k Z



= dt1 dt2 . . . dtk VI q(t1 ) . . . VI q(tk ) 0
,
k!
k=0

where h•i0 means expectation value with respect to the Gaussian measure
e−S0 /Z0 (equation (3.40)) with periodic boundary conditions.
Since VI (q) is a polynomial, successive terms in the expansion can be
calculated using Wick’s theorem in the form (3.43). This is the basis of
perturbation theory.

147
3.10.2 Perturbative expansion and the minimum of the potential
To each decomposition of the potential into a sum of a quadratic term
and a remainder VI (q) is associated a perturbative expansion. However, the
integrand is the largest in the vicinity of the paths that minimize the action.
Clearly, periodic functions that minimize the action are constant functions
q(t) ≡ q0 , to minimize the kinetic term, where the value q0 must minimize
the potential V (q) and thus

V ′ (q0 ) = 0 , V ′′ (q0 ) > 0 .

Therefore, the optimal decomposition is

2
V (q) = V (q0 ) + 21 V ′′ (q0 ) (q − q0 ) + VI (q).

Specific problems associated with a degeneracy of the minimum of the po-

tential will be discussed somewhere else.

148
3.10.3 Perturbative expansion in powers of ~ and steepest descent method
In this section, we reinstate Planck’s constant ~ and consider the Euclidean
or imaginary time action. For example, the partition function takes the form
Z
Z= [dq(t)] exp [−S(q)/~] ,
q(−τ /2)=q(τ /2)

where τ = ~β will be considered fixed when ~ → 0. We consider a general

Hamiltonian of the form
1 2
H= p + V (q),
2m
where the potential V (q) is expandable in powers of q (and not only a
polynomial) with a unique minimum at q = 0 such that

V (q) = 21 mω 2 q 2 + VI (q), VI (q) = O(q 3 ).

149
We have already discussed perturbative calculations of the partition func-
tion in the form of expansions in powers of some polynomial perturbation
VI (q). We have pointed out that an optimal harmonic approximation is pro-
vided by a quadratic approximation of the potential around its minimum.
A perturbative expansion around this harmonic approximation can be
reorganized in a simple way as an expansion in powers of ~. The latter
expansion differs from the simple perturbative expansion as soon as the
perturbation VI is no longer a monomial.
An expansion in powers of ~ can be called semi-classical but, in fact, it
is the expansion in ~ based on the representation of the quantum partition
function that corresponds better to the idea of a semi-classical approach.
We consider an Euclidean action S(q) of the form
Z τ /2
1 2

S(q) = 2 mq̇ (t) + V q(t) dt .
−τ /2

We evaluate the path integral in the formal limit ~ → 0, with τ fixed.

150
From a semi-classical viewpoint, this limit corresponds indeed to ~ → 0, but
simultaneously β = τ /~ → ∞, that is, in a correlated way the temperature
goes to zero, a limit that enhances small oscillations around the classical
minimum and excitations close to the quantum ground state.
In this limit, the path integral has exactly a form that justifies using
the steepest descent method. Since q = 0 corresponds to an absolute and
non-degenerate minimum of the potential V , the path integral is dominated
for ~ → 0 by the trivial saddle point q(t) ≡ 0, which minimizes both the
√
potential and the kinetic term. The change of variables q(t) 7→ ~q(t) in
the path integral, allows a direct identification of the leading terms in the
action. After the change,
Z τ /2 Z τ /2 √

m  2 2 2
 1
S(q)/~ = q̇ (t) + ω q (t) + dt VI q(t) ~ .
2 −τ /2 ~ −τ /2
√ √
Since VI (q ~)/~ is at least of order ~, at leading order one recovers the
partition function of a harmonic oscillator.

151
At higher orders, the integral can be calculated by first expanding the in-
tegrand in powers of VI and then collecting terms with the same powers of
~.
For any potential expandable in powers of q and at any finite order in
~, the integrand remains the product of a Gaussian factor by a polynomial
in q since the terms of degree q n in the potential start contributing only
at order ~n/2−1 . Therefore, for any expandable potential, the expansion of
Z(τ /~) in powers of ~ relies on Gaussian expectation values of polynomials,
which can be calculated with the help of Wick’s theorem.

152
3.11 Quantum evolution and Scattering matrix

In the preceding lectures, we have mainly discussed quantum statistical

operators, and thus, formally, evolution in imaginary or euclidean time.
Here, we give a few elements of real-time evolution.
Scattering S-matrix elements, quantities relevant to microscopic physics,
are related to the quantum evolution operator in real time.
Therefore, we first derive the path integral representation of the evolu-
tion operator and the S-matrix in non-relativistic quantum mechanics. To
illustrate the power of the formalism, we show how the perturbative expan-
sion of the scattering amplitude, as generated by the Lippman–Schwinger’s
equation, can be recovered.

153
3.11.1 Quantum evolution and Scattering matrix
Evolution in quantum mechanics is associated with an operator acting lin-
early on the Hilbert space of states. Conservation of probabilities for an
isolated system implies that the evolution operator must be unitary. Fi-
nally, one assumes that the evolution of an isolated system is Markovian
(without memory effects). Then the evolution operator U (t′′ , t′ ) from time
t′ to time t′′ satisfies the relation
U (t3 , t2 )U (t2 , t1 ) = U (t3 , t1 ).
Assuming the evolution differentiable, we expand U (t + ε, t) for an infinites-
imal time interval ε:
U (t + ε, t) = 1 − iεH(t)/~ + O(ε2 ) ,
where H(t) is the quantum Hamiltonian. Equation (3.4) then implies
∂U
i~ (t, t′ ) = H(t)U t, t′ ), U (t′ , t′ ) = 1 .
∂t
154
When the operator H(t) is time-independent, U (t′′ , t′ ) = U (t′′ − t′ ) =
′′ ′
e−i(t −t )H/~ . The evolution operators U (t) then form a representation of
the symmetry group of time translations. The generator of the Lie algebra
∂/∂t is represented by the operator −iH/~.
The S-matrix. The scattering S-matrix is obtained by comparing the
quantum evolution with the free evolution in the absence of interactions.
More precisely, the S-matrix can be defined as the limit of the evolution
operator in the so-called interaction representation:

iH0 t′′ /~ ′′ ′ −iH0 t′ /~

S= ′
lim e U (t , t ) e , (3.48)
t →−∞
t′′ →+∞

where the Hamiltonian H0 is the free Hamiltonian corresponding to H:

its eigenstates are the free or asymptotic states at large time of quantum
evolution.

155
3.11.2 Scattering matrix in quantum particle mechanics
In the simple quantum mechanics of a particle in a potential, examples of
free and interacting Hamiltonians are simply

H0 = p2 /2m, H = p2 /2m + V (q, t). (3.49)

The large time limit and, therefore, the S-matrix exist only if the potential
decreases fast enough at large distance, or large time in such a way that
for |t| → ∞ the evolution converges fast enough toward the free evolution
corresponding to H0 .
′′ ′
The matrix elements of the free evolution operator U0 = e−i(t −t )H0 /~ ,
corresponding to the Hamiltonian H0 in (3.49), in d space dimensions, is
given by
 d/2 " #
′′ ′ 2
′′ ′′ ′ ′ m i m (q − q )
hq | U0 (t , t ) |q i = exp .
2iπ~(t′′ − t′ ) ~ 2(t′′ − t′ )

156
In the momentum basis, the relation (3.48) between the S-matrix and the
evolution operator takes the form
′′ ′′ ′ ′
′′ ′
hp̃ | S |p̃ i = ′
lim eiE t /~
hp̃ | U (t , t ) |p̃ i e−iE
′′ ′′ ′ ′ t /~
(3.50)
t →−∞
t′′ →+∞

with E ′ = E(p′ ), E ′′ = E(p′′ ), E(p) = p2 /2m.

The limits t′ → −∞, t′′ → +∞ have to be understood mathematically in
the sense of distributions (one should use test functions in the form of wave
packets).
The S-matrix in the momentum basis is in general parametrized in terms
of the scattering matrix T :
S = 1 − iT , ⇒ hp̃′′ |S| p̃′ i = (2π~)d δ (d) (p′′ − p′ ) − i hp̃′′ |T | p̃′ i . (3.51)
When the potential is time-independent, energy is conserved and one can
write
hp̃′′ |T | p̃′ i = −2πδ(E ′′ − E ′ )T (p′′ , p′ ). (3.52)

157
3.12 Path integral representation: Evolution and S-matrix

The matrix elements of the evolution operator, corresponding to a Hamil-

tonian of the form
H = p2 /2m + V (q, t),

are given by the path integral (expression (3.13))

Z q(t′′ )=q ′′
hq ′′ | U (t′′ , t′ ) |q ′ i = [dq(t)] exp [iA(q)/~]
q(t′ )=q ′

with Z t′′ 1

2
A(q) = dt 2 mq̇ (t) − V q(t), t .
t′

158
Phase space formulation. In the real time formulation, the path integral over
phase space is replaced by
Z
hq ′′ | U (t′′ , t′ ) |q ′ i = [dp(t)dq(t)] exp [iA(p, q)/~] . (3.53)

The quantity A(p, q), which replaces the Euclidean action in the path inte-
gral, is again the classical action in the Hamiltonian formalism:
Z t′′ 

A(p, q) = p(t)q̇(t) − H p(t), q(t), t dt .
t′

Even in this more general situation the quantum evolution is obtained by

summing over all paths weighted with the complex weight eiA/~ , where A
is the classical action in the phase space formalism.

159
3.12.1 Path integral and S-Matrix
We now show that the representation (3.13) of the evolution operator leads
to a path integral representation for elements of the scattering S-matrix,
which is particularly well suited to the study of the semi-classical limit.
The elements of the S-matrix between two wave packets are given by
Z
′′
hψ2 | S |ψ1 i = ′ lim dq ′ dq ′′ hψ2 | eiH0 t /~ |q′′ i hq′′ | U (t′′ , t′ ) |q′ i
t →−∞
t′′ →+∞
′ −iH0 t′ /~
× hq | e |ψ1 i . (3.54)

Introducing the two wave functions ψ̃1 (p) and ψ̃2 (p) in the momentum
basis, we define (i = 1, 2)
Z d
  2
 
d p p
ψi (q, t) = hq| e−iH0 t/~ |ψi i = ψ̃i (p) exp i p · q − t /~ .
(2π)d 2m
(3.55)

160
When t becomes large, the phase in expression (3.55) varies rapidly, and
the integral is then dominated by the stationary points of the phase:
 2

∂ p p
p·q−t = 0 =⇒ q = t .
∂p 2m m
The evaluation of the integral (3.55) thus yields the estimate
 d/2  
1 m~ iπ p2
ψ1 (q, t) ∼ ψ̃1 (p) exp sgn t + it
|t|→∞ (2π)d/2 |t| 4 2m~
with p = mq/t. We then change variables in integral (3.54), setting
t′ ′
′ ′′t′′ ′′
q = p, q = p ,
m m
and obtain
Z   ′′2 ′2

i ′′ p p
hψ2 | S |ψ1 i ∝ lim dp′ dp′′ ψ̃2∗ (p′′ )ψ̃1 (p′ ) exp t − t′
′
t →−∞ ~ 2m 2m
t′′ →+∞
× ht p /m| U (t′′ , t′ ) |t′ p′ /mi .
′′ ′′
(3.56)

161
In the equation, we introduce the path integral representation (3.13) of the
evolution operator:
Z q(t′′ )=t′′ p′′ /m
′′ ′′ ′′ ′ ′ ′


ht p /m| U (t , t ) |t p /mi = [dq(t)] exp iA(q)/~ .
q(t′ )=t′ p′ /m

We conclude that the S-matrix can be derived from the calculation of the
path integral with classical scattering boundary conditions, that is, by sum-
ming over paths solutions at large time of the free classical equations of
motion.
In particular, if we know how to solve the classical equations of motion
with such boundary conditions we can calculate the evolution operator and
thus the S-matrix for ~ small. This leads to semi-classical approximations
of the S-matrix.

162
3.12.2 Path integral and S-Matrix: perturbation theory.
We now show how the path integral representing the evolution operator can
be calculated in the form of an expansion in powers of the potential. The
path integral formalism actually organizes the perturbative expansion in a
way similar to the operator formalism. From the expansion of the evolution
operator we then derive the expansion of the elements of the S-matrix.
Warning. This part is quite technical, though straightforward, and re-
quires a careful reading.
We consider a time-independent Hamiltonian

H = p2 /2m + V (x)

(setting for convenience ~ = 1). The classical actions corresponding to the

free Hamiltonian H0 and to H, respectively, are
Z t′′ Z t′′
1 2
1 2


A0 (x) = mẋ (t)dt , A(x) = 2 mẋ (t) − V x(t) dt .
t′ 2 t′

163
We now expand the path integral (3.13) in powers of V :

Z x(t′′ )=x′′ X
′′ ′′ ′ ′
hx |U (t , t )| x i = [dx(t)] exp [iA(x)] = hx′′ | U (ℓ) (t′′ , t′ ) |x′ i ,
x(t′ )=x′ ℓ
Z "Z #ℓ
x(t′′ )=x′′ ℓ t′′

′′ (ℓ) ′′ ′ ′ iA0 (x) (−i)
hx | U (t , t ) |x i = [dx(t)] e V x(t) dt .
x(t′ )=x′ ℓ! t′

Since the potential must vanish at large distances, it makes sense to assume
that it has a Fourier representation:

Z
V (x) = (2π)−d ddk eikx Ṽ (k). (3.57)

164
Introducing the representation (3.57) in the path integral,
ℓ Z t Y Z Y ℓ
′′

′′ (ℓ) ′′ ′ ′ (−i) ddkj

hx | U (t , t ) |x i = dτj Ṽ (kj ) d
ℓ! t′
j j=1
(2π)
 
Z x(t′′ )=x′′ Z t′′ X
2
× [dx(t)] exp i  1
2 mẋ (t)dt + kj x(τj ) ,
x(t′ )=x′ t′ j

we see that each term in the perturbative expansion then involves only the
calculation of a Gaussian path integral. The integrand is symmetric in the
times τ1 , τ2 , . . . , τℓ . Therefore, we can order them as t′′ ≥ τℓ ≥ τℓ−1 · · · ≥
τ1 ≥ t′ and suppress the factor 1/ℓ!.
The term of order zero in V yields a contribution (2π)d δ(p′′ − p′ ) to the
S-matrix.
To calculate the general term explicitly, we introduce δ-functions:
Z
d


exp[ikj x(τj )] = d xj δ xj − x(τj ) exp (ikj xj ) .

165
Then, in each interval τj ≤ t ≤ τj+1 , we recognize the matrix elements
of the free evolution operator (see also section 3.7), which we write in the
Fourier representation.
We calculate the Fourier transform with respect to x′ and x′′ , calling p′
and p′′ , respectively, the corresponding momenta. We find
Z Yℓ d Yℓ
′′ (ℓ) ′′ ′ ′ ℓ d kj d ddpj
hp̃ | U (t , t ) |p̃ i = (−i) dτj Ṽ (kj ) d
d xj d
j=1
(2π) j=2
(2π)
X
ℓ+1 
× exp −ip2j (τj − τj−1 )/2m + ipj (xj − xj−1 ) + ikj xj
j=1

with the conventions

τ 0 = t′ , τℓ+1 = t′′ , pℓ+1 = p′′ , p1 = p′ , x0 = xℓ+1 = 0 .
The integration over the variables xj yields δ-functions which determine the
variables kj as kj = pj+1 − pj .

166
After factorization of the free evolution operator on both sides, the limits
t′′ → +∞, t′ → −∞ can be taken.
The corresponding S-matrix elements follow:
Z Yℓ Yℓ
′′ (ℓ) ′ ℓ ddpj ip′′ 2 τℓ /2m ′′
hp̃ | S |p̃ i = (−i) dτj d
e Ṽ (p − pℓ )
j=1 j=2
(2π)
2 2 ′2
× eipℓ (τℓ−1 −τℓ )/2m · · · × eip2 (τ1 −τ2 )/2m Ṽ (p2 − p′ ) e−ip τ1 /2m
. (3.58)

We still have to integrate over the times τj . We set

τj+1 = τj + uj , uj ≥ 0 .

The remaining integral over τ1 yields a δ-function of energy conservation:

2πδ(E ′′ − E ′ ), which follows from time translation invariance. The integrals
over the variables uj on the positive axis yield mathematical distributions:
Z +∞
i(E ′′ −E(pj ))uj i
duj e = ′′ , E(p) ≡ p2 /2m .
0 E − E(pj ) + iǫ

167
The iǫ, ǫ → 0+ , addition identifies the distribution as a boundary value of
an analytic function and indicates how to avoid the pole at a p2j = p′′2 .
The final result is then
Z Y ddpj Ṽ (pj − pj−1 )
′′ (ℓ) ′ ′′ ′ ′′
hp̃ | S |p̃ i = −2iπδ(E − E ) Ṽ (p − pℓ ) d ′′
.
j
(2π) E + iǫ − E(pj )

In this form the perturbation series is a geometric series whose sum satisfies
an integral equation, called Lippman–Schwinger’s equation. In terms of the
operator T (E), where E is generically complex, solution of

T (E) = V − V G0 (E)T (E) , with G0 (E) = (H0 − E)−1 ,

the quantity T (p′′ , p′ ) which appears in equation (3.52) is given by

T (p′′ , p′ ) = hp′′ | T (E + iǫ) |p′ i for E = p′2 /2m = p′′2 /2m .

168
Exercises
Exercise 3.1
Locality. One considers the operator U (t) defined by its matrix elements:

′ t/π
hq| U (t) |q i = 2 .
t + (q − q ′ )2

Verify the semi-group property U (t1 )U (t2 ) = U (t1 + t2 ).

Show that the corresponding Hamiltonian (as defined by equation (3.6)
with ~ = 1) is non-local, that is has a support for q 6= q ′ .
Solution. The matrix elements of the Hamiltonian are
1
hq| H |q ′ i = − ′ 2
, for q 6
= q ′
.
π(q − q )

169
Exercise 3.2
The square-well potential. Use the expression obtained for the statistical
operator e−βH of the harmonic oscillator to derive the spectrum of the
attractive square-well potential H = p2 /2 + V(x) with

V(x) = 0 , for |x| > a/2 , V(x) = V < 0 , for |x| < a/2 .

(Exercise requiring some ingenuity).

Solution. The basic idea is to calculate the Fredholm determinant D(V, E)
of the operator H − E, whose zeros determine the spectrum. The determi-
nant can be expressed in terms of a Gaussian path integral:
Z
D−1/2 (V, E) ∝ [dq(x)] exp[−S(q)],
Z h
2
2 i
S(q) = dx 1
2 q ′ (x) + V(x) − E q (x) .

170
This Gaussian integral can be calculated explicitly in various ways. One
method is as follows. First, it is convenient to consider the integral as a
limit for β → ∞ of the partition function in Euclidean time β (that is at
temperature 1/β), which is obtained by imposing periodic boundary condi-
tions q(−β/2) = q(β/2):

D−1/2 (V, E) ∝ lim tr U (β)

β→∞

with
Z q(β/2)=q ′′
hq ′′ | U (β) |q ′ i = [dq(x)] exp [−S(q)] .
q(−β/2)=q ′

We then split the interval [−β/2, β/2] into three sub-intervals in which the
potential is constant: [−β/2, −a/2], [−a/2, a/2], [a/2, β/2]. We write U (β)
as a product of statistical operators corresponding to these various intervals.
In each interval, the path integral corresponds to the statistical operator of

171
a harmonic oscillator. For what follows, we introduce the notation
−E = 21 ω12 , V − E = 12 ω22 .
We denote by U1 and U2 the operators corresponding to ω1 and ω2 , respec-
tively. Then,
tr U (β) = tr U1 (β/2 − a/2)U2 (a)U1 (β/2 − a/2) = tr U1 (β − a)U2 (a),
where the cyclic property of the trace has been used. In the limit β →
∞, the operator U1 becomes the projector on to the ground state of the
corresponding oscillator:
r Z
ω1 −ω1 (β−a)/2 ′2 ′′ 2
tr U (β) ∼ e dq ′ dq ′′ e−ω1 q /2 e−ω1 q /2 hq ′′ | U2 (a) |q ′ i .
β→∞ π
One then uses the explicit result (3.22) and performs the two Gaussian
integrations over q ′ and q ′′ . One finds
 1/2
2ω1 ω2 2 2

−1/2 −ω1 (β−a)/2
tr U (β) ∼ ω1 + ω2 + 2ω1 ω2 coth(aω2 ) e .
β→∞ sinh(aω2 )

172
To obtain a result that has a finite limit, it is necessary to normalize the
determinant. Dividing it by its value for vanishing potential, one finds

aω1 ω12 + ω22

e D(V, E)/D(0, E) = cosh(aω2 ) + sinh(aω2 ).
2ω1 ω2

We have calculated for V − E > 0. The determinant can only vanish for
E > V . We thus set ω2 = iκ2 . The equation that gives the energies of the
bound states can then be written as
2ω1 κ2
tan(aκ2 ) = 2 2 ,
ω1 − κ2

an equation that can easily be verified by solving the Schrödinger equation

for the square-well directly and by combining the two spectral equations
that correspond to even and odd eigenfunctions.

173