221A Lecture Notes

Path Integral

1 Feynmans Path Integral Formulation

Feynmans formulation of quantum mechanics using the so-called path inte-
gral is arguably the most elegant. It can be stated in a single line:
Z
hxf , tf |xi , ti i = Dx(t)eiS[x(t)]/h . (1)

The meaning of this equation is the following. If you want to know the
quantum mechanical amplitude for a point particle at a position xi at time
ti to reach a position xf at time tf , you integrate over all possible paths
connecting the points with a weight factor given by the classical action for
each path. Hence the name path integral. This is it. Note that the position
kets form a complete set of basis, and knowing this amplitude for all x is
enough information to tell you everything about the system. The expression
is generalized for more dimensions and more particles in a straightforward
manner.
As we will see, this formulation is completely equivalent to the usual
formulation of quantum mechanics. On the other hand, there are many
reasons why this expression is just beautiful.
First, the classical equation of motion comes out in a very simple way.
If you take the limit h 0, the weight factor eiS/h oscillates very rapidly.
Therefore, we expect that the main contribution to the path integral comes
from paths that make the action stationary. This is nothing but the deriva-
tion of EulerLagrange equation from the classical action. Therefore, the
classical trajectory dominates the path integral in the small h limit.
Second, we dont know what path the particle has chosen, even when we
know what the initial and final positions are. This is a natural generalization
of the two-slit experiment. Even if we know where the particle originates from
and where it hit on the screen, we dont know which slit the particle came
through. The path integral is an infinite-slit experiment. Because you cant
specificy where the particle goes through, you sum them up.
Third, we gain intuition on what quantum fluctuation does. Around the
classical trajectory, a quantum particle explores the vicinity. The trajec-
tory can deviate from the classical trajectory if the difference in the action is

1
roughly within h. When a classical particle is confined in a potential well, a
quantum particle can go on an excursion and see that there is a world outside
the potential barrier. Then it can decide to tunnel through. If a classical
particle is sitting at the top of a hill, it doesnt fall; but a quantum particle
realizes that the potential energy can go down with a little excursion, and
decides to fall.
Fourth, whenever we have an integral expression for a quantity, it is often
easier to come up with an approximation method to work it out, compared
to staring at a differential equation. Good examples are the perturbative
expansion and the steepest descent method. One can also think of useful
change of variables to simplify the problem. In fact, some techniques in
quantum physics couldnt be thought of without the intuition from the path
integral.
Fifth, the path integral can also be used to calculate partition functions
in statistical mechanics.
Sixth, the connection between the conservation laws and unitarity trans-
formations become much clearer with the path integral. We will talk about
it when we discuss symmetries.
There are many more but I stop here.
Unfortunately, there is also a downside with the path integral. The ac-
tual calculation of a path integral is somewhat technical and awkward. One
might even wonder if an integral over paths is mathematically well-defined.
Mathematicians figured that it can actually be, but nonetheless it makes some
people nervous. Reading off energy eigenvalues is also less transparent than
with the Schrodinger equation.
Below, we first derive the path integral from the conventional quantum
mechanics. Then we show that the path integral can derive the conventional
Schrodinger equation back. After that we look at some examples and actual
calculations.
By the way, the original paper by Feynman on the path integral Rev.
Mod. Phys. 20, 367-387 (1948) is quite readable for you, and I recommend
it. Another highly recommended read is Feynmans Nobel Lecture. You
will see that Feynman invented the path integral with the hope of replacing
quantum field theory with particle quantum mechanics; he failed. But the
path integral survived and did mighty good in the way he didnt imagine.

2
2 Physical Intuition
Take the two-slit experiment. Each time an electron hits the screen, there is
no way to tell which slit the electron has gone through. After repeating the
same experiment many many times, a fringe pattern gradually appears on the
screen, proving that there is an interference between two waves, one from one
slit, the other from the other. We conclude that we need to sum amplitudes
of these two waves that correspond to different paths of the electron. Now
imagine that you make more slits. There are now more paths, each of which
contributing an amplitude. As you increase the number of slits, eventually
the entire obstruction disappears. Yet it is clear that there are many paths
that contribute to the final amplitude of the electron propagating to the
screen.
As we generalize this thought experiment further, we are led to conclude
that the amplitude of a particle moving from a point xi to another point
xf consists of many components each of which corresponds to a particular
path that connects these two points. One such path is a classical trajectory.
However, there are infinitely many other paths that are not possible classi-
cally, yet contribute to the quantum mechanical amplitude. This argument
leads to the notion of a path integral , where you sum over all possible paths
connecting the initial and final points to obtain the amplitude.
The question then is how you weight individual paths. One point is clear:
the weight factor must be chosen such that the classical path is singled out
in the limitR h 0. The correct choice turns out to be eiS[x(t)]/h , where
t
S[x(t)] = tif dtL(x(t), x(t)) is the classical action for the path x(t) that
satisfies the boundary condition x(ti ) = xi , x(tf ) = xf . In the limit h 0,
the phase factor oscillates so rapidly that nearly all the paths would cancel
each other out in the final amplitude. However, there is a path that makes the
action stationary, whose contribution is not canceled. This particular path
is nothing but the classical trajectory. This way, we see that the classical
trajectory dominates the path integral in the h 0 limit.
As we increase h, the path becomes fuzzy. The classical trajectory
still dominates, but there are other paths close to it whose action is within
S ' h and contribute significantly to the amplitude. The particle does an
excursion around the classical trajectory.

3
3 Propagator
The quantity
K(xf , tf ; xi , ti ) = hxf , tf |xi , ti i (2)
is called a propagator . It knows everything about how a wave function prop-
agates in time, because
Z
(xf , tf ) = hxf , tf |i = hxf , tf |xi , ti idxi hxi , ti |i
Z
= K(xf , tf ; xi , ti )(xi , ti )dxi . (3)

In other words, it is the Greens function for the Schrodinger equation.

The propagator can also be written using energy eigenvalues and eigen-
states (if the Hamiltonian does not depend on time),

K(xf , tf ; xi , ti ) = hxf |eiH(tf ti )/h |xi i = hxf |nieiEn (tf ti )/h hn|xi i
X

n
iEn (tf ti )/h
X
= e n (xf )n (xi ). (4)
n

In particular, Fourier analyzing the propagator tells you all energy eigenval-
ues, and each Fourier coefficients the wave functions of each energy eigen-
states.
The propagator is a nice package that contains all dynamical information
about a quantum system.

4 Derivation of the Path Integral

The basic point is that the propagator for a short interval is given by the
classical Lagrangian
2 )/h
hx1 , t + t|x0 , ti = cei(L(t)t+O(t) , (5)

where c is a normalization constant. This can be shown easily for a simple

Hamiltonian
p2
H= + V (x). (6)
2m

Note that the expression here is in the Heisenberg picture. The base kets |xi and the
operator x depend on time, while the state ket |i doesnt.

4
The quantity we want is
Z
hx1 , t + t|x0 , ti = hx1 |eiHt/h |x0 i = dphx1 |pihp|eiHt/h |x0 i. (7)

Because we are interested in the phase factor only at O(t), the last factor
can be estimated as

hp|eiHt/h |x0 i
= hp|1 iHt/h + (t)2 |x0 i
ipx0 /h
i p2
!
i 2 e
= 1 t V (x0 )t + O(t)
h 2m h 2h
2
!
1 i p
= exp px0 + t + V (x0 )t + O(t)2 . (8)
2h h 2m

Then the p-integral is a Fresnel integral

hx1 , t + t|x0 , ti
Z
dp ipx1 /h i(px0 + p2 t+V (x0 )t+O(t)2 )/h
= e e 2m
2h
i m (x1 x0 )2
!
m
r
= exp V (x0 )t + O(t)2 . (9)
2iht h 2 t

The quantity in the parentheses is nothing but the classical Lagrangian

times t by identifying x2 = (x1 x0 )2 /(t)2 .
Once Eq. (5) is shown, we use many time slices to obtain the propagator
for a finite time interval. Using the completeness relation many many times,
Z
hxf , tf |xi , ti i = hxf , tf |xN 1 , tN 1 idxN 1 hxN 1 , tN 1 |xN 2 , tN 2 idxN 2
dx2 hx2 , t2 |x1 , t1 idx1 hx1 , t1 |xi , ti i. (10)

The time interval for each factor is t = (tf ti )/N . By taking the limit
N , t is small enough that we can use the formula Eq. (5), and we
find
1
Z NY PN 1
hxf , tf |xi , ti i = dxi ei i=0
L(ti )t/h
, (11)
i=1

The expression is asymmetrical between x1 and x0 , but the difference between V (x0 )t
and V (x1 )t is approximately V 0 t(x0 x1 ) ' V 0 x(t)2 and hence of higher oder.

5
up to a normalization. (Here, t0 = ti .) In the limit N , the integral
over positions at each time slice can be said to be an integral over all possible
paths. The exponent becomes a time-integral of the Lagrangian, namely the
action for each path.
This completes the derivation of the path integral in quantum mechanics.
As clear from the derivation, the overall normalization of the path integral
is a tricky business.
A useful point to notice is that even matrix elements of operators can be
written in terms of path integrals. For example,
Z
hxf , tf |x(t0 )|xi , ti i = dx(t0 )hxf , tf |x(t0 ), t0 ix(t0 )hx(t0 ), t0 |xi , ti i
Z Z Z
iS[x(t)]/h
= dx(t0 ) Dx(t)e x(t0 ) Dx(t)eiS[x(t)]/h
tf >t>t0 t0 >t>ti
Z
= Dx(t)eiS[x(t)]/h x(t0 ), (12)
tf >t>ti

At the last step, we used the fact that an integral over all paths from xi
to x(t0 ), all paths from x(t0 ) to xf , further integrated over the intermediate
position x(t0 ) is the same as the integral over all paths from xi to xf . The
last expression is literally an expectation value of the position in the form of
an integral. If we have multiple insertions, by following the same steps,
Z
hxf , tf |x(t2 )x(t1 )|xi , ti i = Dx(t)eiS[x(t)]/h x(t2 )x(t1 ). (13)
tf >t>ti

Here we assumed that t2 > t1 to be consistent with successive insertion of

positions in the correct order. Therefore, expectation values in the path
integral corresponds to matrix elements of operators with correct ordering
in time. Such a product of operators is called timed-ordered T x(t2 )x(t1 )
defined by x(t2 )x(t1 ) as long as t2 > t1 , while by x(t1 )x(t2 ) if t1 > t2 .
Another useful point is that EulerLagrange equation is obtained by the
change of variable x(t) x(t) + x(t) with xi = x(ti ) and xf = x(tf ) held
fixed. A change of variable of course does not change the result of the integral,
and we find Z Z
iS[x(t)+x(t)]/h
Dx(t)e = Dx(t)eiS[x(t)]/h , (14)
and hence
Z
iS[x(t)+x(t)]/h
Z
iS[x(t)]/h
Z
i
Dx(t)e Dx(t)e = Dx(t)eiS[x(t)]/h S = 0. (15)
h
6
Recall (see Note on Classical Mechanics I)
!
Z
L d L
S = S[x(t) + (t)] S[x(t)] = x(t)dt. (16)
x dt x

Therefore,
!
Z
iS[x(t)]/h iZ L d L
Dx(t)e x(t)dt = 0. (17)
h x dt x

Because x(t) is an arbitrary change of variable, the expression must be zero

at all t independently,
!
Z
iS[x(t)]/h i L d L
Dx(t)e = 0. (18)
h x dt x

Therefore, the EulerLagrange equation must hold as an expectation value,

nothing but the Ehrenfests theorem.

5 Schrodinger Equation from Path Integral

Here we would like to see that the path integral contains all information we
need. In particular, we rederive Schrodinger equation from the path integral.
Let us first see that the momentum is given by a derivative. Starting
from the path integral
Z
hxf , tf |xi , ti i = Dx(t)eiS[x(t)]/h , (19)

we shift the trajectory x(t) by a small amount x(t)+x(t) with the boundary
condition that xi is held fixed (x(ti ) = 0) while xf is varied (x(tf ) 6= 0).
Under this variation, the propagator changes by

hxf + x(tf ), tf |xi , ti i hxf , tf |xi , ti i = hxf , tf |xi , ti ix(tf ). (20)
xf

On the other hand, the path integral changes by

Z
iS[x(t)+x(t)]/h
Z
iS[x(t)]/h
Z
iS
Dx(t)e Dx(t)e = Dx(t)eiS[x(t)]/h . (21)
h

7
Recall that the action changes by (see Note on Classical Mechanics II)
S = S[x(t) + x(t)] S[x(t)]
Z tf !
L L
= dt x + x
ti x x
x !
L  f Z tf L d L
= x + dt x. (22)
x xi ti x dt x
The last terms vanishes because of the equation of motion (which holds as
an expectation value, as we saw in the previous section), and we are left with
L
S = x(tf ) = p(tf )x(tf ). (23)
x
By putting them together, and dropping x(tf ), we find
Z
i
hxf , tf |xi , ti i = Dx(t)eiS[x(t)]/h p(tf ). (24)
xf h
This is precisely how the momentum operator is represented in the position
space.
Now the Schrodinger equation can be derived by taking a variation with
respect to tf . Again recall (see Note on Classical Mechanics II)
S
= H(tf ) (25)
tf
after using the equation of motion. Therefore,
Z
i
hxf , tf |xi , ti i = Dx(t)eiS[x(t)]/h H(tf ). (26)
tf h
If
p2
H= + V (x), (27)
2m
the momentum can be rewritten using Eq. (24), and we recover the Schrodinger
equation,
!2
1 h
ih hxf , tf |xi , ti i = + V (xf ) hxf , tf |xi , ti i. (28)
t 2m i xf

In other words, the path integral contains the same information as the
conventional formulation of the quantum mechanics.

8
6 Examples
6.1 Free Particle and Normalization
Here we calculate the path integral for a free particle in one-dimension. We
need to calculate
Z R m 2
hxf , tf |xi , ti i = Dx(t)ei 2
x dt/h
, (29)

over all paths with the boundary condition x(ti ) = xi , x(tf ) = xf .

First, we do it somewhat sloppily, but we can obtain dependences on
important parameters nonetheless. We will come back to much more careful
calculation with close attention to the overall normalization later on.
The classical path is
xf xi
xc (t) = xi + (t ti ). (30)
tf ti

We can write x(t) = xc (t) + x(t), where the left-over piece (a.k.a. quantum
fluctuation) must vanish at the initial and the final time. Therefore, we can
expand x(t) in Fourier series

X n
x(t) = an sin (t ti ). (31)
n=1 tf ti

The integral over all paths can then be viewed as integrals over all an ,
Z
Z Y
Dx(t) = c dan . (32)
n=1

The overall normalization factor c that depends on m and tf ti will be fixed

later. The action, on the other hand, can be calculated. The starting point
is
xf xi X n n
x = xc + x = + an cos (t ti ). (33)
tf ti n=1 tf ti tf ti
Because different modes are orthogonal upon t-integral, the action is

m (xf xi )2 m X 1 (n)2 2
S= + a . (34)
2 tf ti 2 n=1 2 tf ti n

9
The first term is nothing but the classical action. Then the path integral
reduces to an infinite collection of Fresnel integrals,

(xf xi )2 X 1 (n)2 2
" !#
Z Y
im
c dan exp + an . (35)
n=1 h 2 tf ti n=1 2 tf ti

We now obtain

!1/2
i m 1 (n)2 i m (xf xi )2
" #
Y
hxf , tf |xi , ti i = c exp . (36)
n=1 h 2 2 tf ti h 2 tf ti

Therefore, the result is simply

i m (xf xi )2
" #
0
hxf , tf |xi , ti i = c (tf ti ) exp . (37)
h 2 tf ti

The normalization constant c can depend only on the time interval tf ti ,

and is determined by the requirement that
Z
dxhxf , tf |x, tihx, t|xi , ti i = hxf , tf |xi , ti i. (38)

And hence
v
u 2ih(tf t)(t ti )
u
0 0
c (tf t)c (t ti )t = c0 (tf ti ). (39)
m(tf ti )

We therefore find
m
r
c0 (t) =
, (40)
2iht
recovering Eq. (2.5.16) of Sakurai precisely.
In order to obtain this result directly from the path integral, we should
have chosen the normalization of the measure to be
s s
m Y m n
Dx(t) = dan . (41)
2ih(tf ti ) n=1 2ih(tf ti ) 2

This argument does not eliminate the possible factor ei0 (tf ti ) , which would cor-
respond to a zero point in the energy h0 . Mahiko Suzuki pointed out this problem to
me. Here we take the point of view that the normalization is fixed by comparison to the
conventional method.

10
 Now we do it much more carefully with close attention to the overall
normalization. The path integral for the time interval tf ti is divided up
into N time slices each with t = (tf ti )/N ,
1
N/2 Z NY
m

K= dxn eiS/h , (42)
2iht n=1

where
1
1 NX (xn+1 xn )2
S= m . (43)
2 n=1 t
The point is that this is nothing but a big Gaussian (to be more precise,
Fresnel) integral, because the action is quadratic in the integration variables
x1 , , xN 1 . To make this point clear, we rewrite the action into the follow-
ing form:

xN

0


1m 2 m 0 1m
S= (x + x20 ) (xN 1 , xN 2 , xN 3 , , x2 , x1 ) .. +


2 t N

t
.
2 t
0


x0

2 1 0 0 0 xN 1

1 2 1 0 0
xN 2

0 1 2 0

0 xN 3
(xN 1 , xN 2 , xN 3 , , x2 , x1 ) .. .. .. . . .. .. ..



. . . . . .
.

0 0 0 2 1 x2


0 0 0 1 2 x1
(44)

Comapring to the identity Eq. (106),

N
Z Y
1 T AxxT y 1 T A1 y
dxn e 2 x = (2)N/2 (detA)1/2 e+ 2 y (45)
n=1

11
we identify

1 1/2 0 0 0

1/2 1 1/2 0 0

1/2

i 2m 0 1 0 0
A =
.. .. .. .. .. (46)


h t ...
. . . . .

0 0 0 1 1/2


0 0 0 1/2 1
This is nothing but the matrix KN 1 in Eq. (106) with a = 1/2 up to a
factor of hi 2m
t
. One problem is that, for a = 1/2, + = = 1/2 and hence
Eq. (111) is singular. Fortunately, it can be rewritten as
 N 1
N N
1 2 2 1 1
detKN 1 = + = N
+ +N
+ ++ ++ N
+N
=N .
+ 2
(47)
Therefore,
2m N 1
N 1
1 m
  
detA = N N 1 = N . (48)
iht 2 iht
Therefore, the prefactor of the path integral is
N/2
m

(2)(N 1)/2 (detA)1/2
2iht
m N/2 m (N 1)/2
   
= (2)1/2 N 1/2
iht iht
1/2 !1/2
m m

= = . (49)
2ihN t 2ih(tf ti )
The exponent is given by

xN

0


i 1 im 2 1 m
 2
1
0
(xN + x20 ) + (xN , 0, 0, , 0, x0 )A .. (50)


h 2 ht 2 t
.

0


x0
We need only (1, 1), (N 1, N 1), (1, N 1), and (N 1, 1) components
of A1 .
iht detKN 2 iht N 1
(A1 )1,1 = (A1 )N 1,N 1 = = . (51)
2m detKN 1 m N

12
On the other hand,

(A1 )1,N 1 = (A1 )N 1,1

1/2 1/2

1 0
iht 1

0 1/2 1 0
(1)N 2 det

= .. .. .. .. ..
2m detKN 1
. . . . .


0 0 0 1/2
iht 2N 1 1
=
2m N 2N 2
iht
= . (52)
Nm
We hence find the exponent

xN

0


i1m 2 1 m
 2
1
0
(xN + x20 ) + (xN , 0, 0, , 0, x0 )A ..


h 2 t 2 t
.

0


x0
2 !
i1m 2 1 im iht N 1 2 iht

= (xN + x20 ) + (xN + x20 ) + 2 xN x0
h 2 t 2 ht m N Nm
i1m 1 2 1 im
= (x + x20 ) 2xN x0
h 2 t N N 2 hN t
i m (xN x0 )2
=
h 2 N t
i m (xf xi )2
= . (53)
h 2 tf ti

Putting the prefactor Eq. (49) and the exponent Eq. (53) together, we find
the propagator
!1/2 2
m i m (xf xi )
h 2 tf ti
K= e . (54)
2ih(tf ti )
We recovered Eq. (2.5.16) of Sakurai without any handwaving this time!

13
6.2 Harmonic Oscillator
Now that we have fixed the normalization of the path integral, we calculate
the path integral for a harmonic oscillator in one-dimension. We need to
calculate Z R m 2 1 2 2
hxf , tf |xi , ti i = Dx(t)ei ( 2 x 2 m x )dt/h , (55)
over all paths with the boundary condition x(ti ) = xi , x(tf ) = xf . The
classical path is
sin (tf t) sin (t ti )
xc (t) = xi + xf . (56)
sin (tf ti ) sin (tf ti )
The action along the classical path is
Z tf
m 2 1

2 2

1 (x2i + x2f ) cos (tf ti ) 2xi xf
Sc = x m x dt = m .
ti 2 2 2 sin (tf ti )
(57)
The quantum fluctuation around the classical path contributes as the
O(h) correction to the amplitude, relative to the leading piece eiSc /h . We
expand the quantum fluctuation in Fourier series as in the case of free particle.
Noting that the action is stationary with respect to the variation of x(t)
around xc (t), there is no linear piece in an . Because different modes are
orthogonal upon t-integral, the action is

(n)2
!
m 1 2
2 (tf ti )
X
S = Sc + a . (58)
n=1 2 tf ti 2 n
Therefore the path integral is an infinite number of Fresnel integrals over an
using the measure in Eq. (41)
s
iSc /h m
hxf , tf |xi , ti i = e
2ih(tf ti )

(n)2
s " ! #
m n Z im 1 2
2 (tf ti )
Y
dan exp a
n=1 2ih(tf ti ) 2 h 2 tf ti 2 n

!2 1/2
(tf ti )
s
m
= eiSc /h
Y
1 . (59)
2ih(tf ti ) n=1 n
Now we resort to the following infinite product representation of the sine
function

x2
!
Y sin x
1 2 = . (60)
n=1 n x

14
We find
v
u (tf ti )
s
m
u
hxf , tf |xi , ti i = eiSc /h t
2ih(tf ti ) sin (tf ti )
s
iSc /h m
= e (61)
2ih sin (tf ti )
This agrees with the result from the more conventional method as given in
Sakurai Eq. (2.5.18).
Again, we can pay closer attention to the normalization by going through
the same steps as for a free particle using discretized time slices. The action
is
1 1
1 NX (xn+1 xn )2 1 N
1 1 2 1 2
 
m 2 x2n m 2
X
S= m x + x . (62)
2 n=1 t 2 n=1 2 2 0 2 N
The last term is there to correctly acccount for all N 1 time slices for
the potential energy term V dt, but to retain the symmetry between the
R

initial and final positions, we took their average. Once again, this is nothing
but a big Fresnel (complex Gaussian) integral:

xN

0


1m 2 2 m 0 1m
S= (x + x0 ) (xN 1 , xN 2 , xN 3 , , x2 , x1 ) .. +


2 t N

t
.
2 t
0


x0

2 1 0 0 0 xN 1

1 2 1 0 0 xN 2


0 1 2 0

0 xN 3

(xN 1 , xN 2 , xN 3 , , x2 , x1 ) .. .. .. . . . .. ..

. ..


. . . . .

0 0 0 2 1 x2


0 0 0 1 2 x1

xN 1

xN 2


1 2
xN 3 1
m (xN 1 , xN 2 , xN 3 , , x2 , x1 ) .. t m 2 (x2N + x20 )t

2
.
4
x2


x1

15
 (63)

Comapring to the identity Eq. (106),

N
Z Y
1 T AxxT y 1 T A1 y
dxn e 2 x = (2)N/2 (detA)1/2 e+ 2 y (64)
n=1

we identify

1 1/2 0 0 0

1/2 1 1/2 0 0

1/2

i i 2m 0 1 0 0
A= m 2 tIN 1 .. .. .. .. ..

h

h t ..
. . . . . .

0 0 0 1 1/2


0 0 0 1/2 1
(65)
This is nothing but the matrix KN 1 in Eq. (106) up to an overall factor of
i 2m i m
+ m 2 t = (2 2 (t)2 ) (66)
h t h iht
with
1 2m/t 1
a= 2m = . (67)
2
2 t m t 2 2 (t)2
Using Eq. (111),
s !
1 4
= 1 1 (68)
2 (2 2 (t)2 )2

Therefore,

N N
N 1
m +

detA = (2 2 (t)2 )N 1 . (69)
iht +
Therefore, the prefactor of the path integral is
N/2
m

(2)(N 1)/2 (detA)1/2
2iht
!1/2
r
m  (N 1)/2 N N
+
= 2 2 (t)2 . (70)
2iht +

16
Now we use the identity
x N
 
lim 1 + = ex . (71)
N N
We find in the limit of t = (tf ti )/N and N ,
s !
1 4 1
= 1 1 = (1 it) + O(N 3 ), (72)
2 (2 (t)2 )2
2 2
and hence
N N
+ 1 ei(tf ti ) ei(tf ti )
+ O(N 1 )
+ 2N it
N sin (tf ti )
= + O(N 1 ). (73)
2N 1 (tf ti )
In the same limit,
(N 1)/2
1

2 (N 1)/2
 
2 (N 1)/2
2 (t) = 2 1 2 (t)2
2
(N 1)/2
= 2 + O(N 1 ). (74)
Therefore, the prefactor Eq. (70) is
!1/2
N sin (tf ti )
s
m m
r
2(N 1)/2 N 1 = . (75)
2iht 2 (tf ti ) 2ih sin (tf ti )
The exponent is given by

xN

0

2
i 1 im 2 2 i1 2 2 2 1 im

1
0
(xN +x0 ) m (xN +x0 )+ (xN , 0, 0, , 0, x0 )A ..


h 2 ht h 4 2 ht
.

0


x0
(76)
We need only (1, 1), (N 1, N 1), (1, N 1), and (N 1, 1) components
of A1 .
(A1 )1,1 = (A1 )N 1,N 1
iht detKN 2 iht N 1 N

1
= (2 2 (t)2 )1 = (2 2 (t)2 )1 + N
m detKN 1 m + N
(77)

17
On the other hand,

(A1 )1,N 1 = (A1 )N 1,1

a 1 a 0

iht 1

0 a 1 0
(2 2 (t)2 )1 (1)N 2 det

= .. .. .. . . .
m detKN 1
. . . . ..

0 0 0 a
iht + N 2
= (2 2 (t)2 )1 N a
m + N

iht 1 2N 1 (tf ti ) 1 iht (tf ti )
2 N 2
= . (78)
m N sin (tf ti ) 2 N m sin (tf ti )
We hence find the exponent

xN

0


i 1 im

1

1 m
 2
1
0
m 2 (x2N 2
+ x0 ) + (xN , 0, 0, , 0, x0 )A ..


h 2 ht 4 2 t
.

0


x0
2
i 1 im 1 1 im iht
  
= m 2 (x2N + x20 ) +
h 2 ht 4 2 ht m
N 1 1
N
!
2 2 1 + 2 2 2 (tf ti )
(2 (t) ) (xN + x0 ) + xN x0
N
+
N
N sin (tf ti )
(79)

Taking the limit N , the second term is obviously regular, and the last
term in the parentheses is also:
2
1 im iht 2 (tf ti ) i m

xN x0 = xN x0 . (80)
2 ht m N sin (tf ti ) h sin (tf ti )
The other terms are singular and we need to treat them carefully.
i1m 2 1 im N 1 N
1
(xN + x20 ) (2 2 (t)2 )1 + N (x2N + x20 )
h 2 t 2 ht + N
N 1 1
N
!
im 2 2 1 +
= 1 (2 (t) ) (x2N + x20 ). (81)
2ht N+ N

18
We need O(N 1 ) term in the parentheses. The factor (2 2 (t)2 )1 =
21 + O(N 2 ) is easy. Using Eq. (72), the next factor is
1 N 1
N
+ 2N (1 + it)N 1 (1 it)N 1 + O(N 2 )
= (82)
N
+
N
2N 1 (1 + it)N (1 it)N + O(N 2 )
Therefore, dropping O(N 2 ) corrections consistently,
N 1 1
+ N

1 (2 2 (t)2 )1
N
+
N

(1 + it)N (1 it)N (1 + it)N 1 + (1 it)N 1

=
(1 + it)N (1 it)N
(1 + it)N 1 + (1 it)N 1
= it
(1 + it)N (1 it)N
(tf ti ) 2 cos (tf ti ) 1 (tf ti ) cos (tf ti )
= i = . (83)
N 2i sin (tf ti ) N sin (tf ti )
Collecting all the terms, the exponent is
i i m im 1 (tf ti ) cos (tf ti ) 2
Sc = xN x0 + (xN + x20 )
h h sin (tf ti ) 2ht N sin (tf ti )
i m (x2N + x20 ) cos (tf ti ) 2xN x0
= (84)
h 2 sin (tf ti )
which is the same as the classical action Eq. (57).

7 Partition Function
The path integral is useful also in statistical mechanics to calculate partition
functions. Starting from a conventional definition of a partition function
eEn ,
X
Z= (85)
n

where = 1/(kB T ), we rewrite it in the following way using completeness

relations in both energy eigenstates and position eigenstates.
Z
H
hn|xihx|eH |ni
X X
Z= hn|e |ni = dx
n n
Z Z
hx|eH |nihn|xi = dxhx|eH |xi.
X
= dx (86)
n

19
The operator eH is the same as eiHt/h except the analytic continuation
t i = ih. Therefore, the partition function can be written as a path
integral for all closed paths, i.e., paths with the same beginning and end
points, over a time interval ih. For a single particle in the potential
V (x), it is then
!2
Z
1 I h m x
Z= Dx( ) exp d + V (x) . (87)
h 0 2

The position satisfies the periodic boundary condition x(h) = x(0).

For example, in the case of a harmonic oscillator, we can use the result
in the previous section

hxf , tf |xi , ti i
(x2 + x2f ) cos (tf ti ) 2xi xf
s " #
m i1
= exp m i ,
2ih sin (tf ti ) h 2 sin (tf ti )
(88)

and take xf = xi = x, tf ti = ih
s " #
H m 11 2 cosh h 2
hx|e |xi = exp mx2 . (89)
2h sinh h h 2 sinh h

Then integrate over x,

Z
Z = dxhx|eH |xi
s s
m 2 sinh h
= h
2h sinh h m 2 cosh h 2
s
1
= 2
4 sinh h/2
1
=
2 sinh h/2
eh/2
=
1 eh

1
e(n+ 2 )h .
X
= (90)
n=0

20
Of course, we could have redone the path integral to obtain the same result.
The only difference is that the classical path you expand the action around
is now given by sinh rather than sin.
Let us apply the path-integral representation of the partition function to
a simple harmonic oscillator. The expression is
!2
Z
1 I h m x m
Z= Dx( ) exp d + 2 x2 . (91)
h 0 2 2

We write all possible paths x( ) in Fourier series as

X 2n X 2n
x( ) = a0 + an cos + bn sin , (92)
n=1 h n=1 h

satisfying the periodic boundary condition x(h) = x(0). Then the exponent
of the path integral becomes
!2
1 I h m x m
d + 2 x2
h 0 2 2
!2

m 2 1 X 2n
= a0 + + 2 (a2n + b2n ) . (93)
2 2 n=1 h

The integration over all possible paths can be done by integrating over the
Fourier coefficients an and bn . Therefore the partition function is
!2

Z
m 2 2 1 X 2n 2
(a2n b2n ) .
Y
Z=c da0 dan dbn exp a0 + + +
n=1 2 2 n=1 h
(94)
This is an infinite collection of Gaussian integrals and becomes
!2 1

s
2 Y 4 2n
Z=c + 2 . (95)
m 2 n=1 m h

Here, c is a normalization constant which can depend on because of nor-

malization of Fourier modes, but not on Lagrangian parameters such as m or
. Now we use the infinite product representation of the hyperbolic function

x2
!
Y sinh x
1+ 2 = . (96)
n=1 n x

21
The partition function Eq. (95) can be rewritten as

!2 1
h2
s
2 Y h
Z = c 2 2
1+
m n=1 mn 2n
2
= c0
sinh h/2
eh/2
= c0 , (97)
1 eh
where c0 is an overall constant, which does not depend on , is not important
when evaluating various thermally averages quantities. If you drop c0 , this
is exactly the partition function for the harmonic oscillator, including the
zero-point energy.

A Path Integrals in Phase Space

As we saw above, the normalization of the path integral is a somewhat tricky issue. The
source of p
the problem was when we did p integral in Eq. (9), that produced a singular
prefactor m/2iht. On the other hand, if we do not do the p integral, and stop at
Z
dp ipx1 /h i(px0 + p2 t+V (x0 )t+O(t)2 )/h
hx1 , t + t|x0 , ti = e e 2m
2h
Z
dp i(p x1 x0 p2 V (x0 ))t
= e t 2m , (98)
2h

we do not obtain any singular prefactor except 1/2h for every momentum integral. The
path integral can then be written as
Z
hxf , tf |xi , ti i = Dx(t)Dp(t)eiS[x(t),p(t)]/h , (99)

where the action is given in the phase space

Z tf
S[x(t), p(t)] = (px H(x, p)) dt. (100)
ti

The expression (99) is an integral over all paths in the phase space (x, p). It is understood
that there is one more p integral than the x integral (albeit both infinite) because of the
number of times the completeness relations are inserted.
Using the path integral in the phase space, we can work out the precise normalization
of path integrals up to numerical constants that depend on t, but not on m, or any other

22
parameters in the Lagrangian. For instance, for the harmonic oscillator, we expand

s
2 X n
x(t) = xc (t) + xn sin (t ti ), (101)
tf ti n=1 tf ti

s s
1 2 X n
p(t) = pc (t) + p0 + pn cos (t ti ). (102)
tf ti tf ti n=1 t f ti

The action is then

!
1 n
p2

1X m tf ti pn
S = Sc 0 (pn , xn ) n 2 , (103)
2m 2 n=1 tf ti m xn

where Sc is the action for the classical solution Eq. (57). Using the integration volume
Z Z
dp0 Y dxn dpn
Dx(t)Dp(t) = c(tf ti ) , (104)
2h n=1 2h

we find the path integral to be

 1/2
1 2mh
hxf , tf |xi , ti i = eiSc /h c(tf ti )
2h i

Y 1   1/2  1/2
2mh 2h 1
. (105)
n=1
2h i im 2 (n/(tf ti ))2

Up to an overall constant c(tf ti ) that depends only on tf ti , both m and dependences

are obtained correctly.

B A Few Useful Identities

To carry out the integral, a few identities would be useful. For N 1-dimensional column
vectors x, y, and a symmetric matrix A,
N
Z Y
1 T
AxxT y 1 T
A1 y
dxn e 2 x = (2)N/2 (detA)1/2 e+ 2 y (106)
n=1

It is easy to prove this equation first without the linear term y by diagonalizatin the matrix
A = OT DO, where O is a rotation Q matrix and D = diag(1 , , N ) has N eigenvalues.
Under a rotation, the measure dxn is invariant because the Jacobian is detO = 1.
Therefore, after rotation of the integration variables, the integral is
N Z N r N
Y
n x2n /2
Y 2 Y
dxn e = = (2)N/2 ( n )1/2 (107)
n=1 n=1
n n=1

23
The product of eigenvalues is nothing but the determinant of the matrix detA = detOT detDdetO =
Q N
n=1 n , and hence

N
Z Y
1 T
dxn e 2 x Ax
= (2)N/2 (detA)1/2 . (108)
n=1

In the presence of the linear term, all we need to do is to complete the square in the
exponent,
1 1 1
xT Ax xT y = (xT yT A1 )A(x A1 y) + yT A1 y, (109)
2 2 2
and shift x to eliminate A1 y, and diagonalize A. This proves the identity Eq. (106).
For a N N matrix of the form

1 a 0 0 0
a 1 a 0 0


0 a 1 0 0

KN = . , (110)

.. .. .. .. ..
.. . . . . .

0 0 0 1 a
0 0 0 a 1

one can show

N
+
+1
N

+1
1 p
detKN = , = (1 1 4a2 ). (111)
+ 2
This is done by focusing on the top two-by-two block. The determinant that picks
the (1, 1) element 1 comes with the determinant of remaining (N 1) (N 1) block,
namely detKN 1 . If you pick the (1, 2) element a, it is necessary to also pick the (2, 1)
element which is also a and the rest comes from the remaining (N 2) (N 2) block,
namely detKN 2 . Therefore, we find a recursion relation

detKN = detKN 1 a2 detKN 2 . (112)

We rewrite this recursion relation as

1 a2
    
detKN detKN 1
= . (113)
detKN 1 1 0 detKN 2

The initial conditions are detK1 = 1, detK2 = 1 a2 . Using the recursion relation
backwards, it is useful to define detK0 = 1. We diagonalize the matrix

1 a2
     
+ + 0 1 1
= . (114)
1 0 1 1 0 1 + +
Hence,
    N 1    
detKN + + 0 1 1 detK1
=
detKN 1 1 1 0 1 + + detK0

24
 N 1
   
1 + + 0 1
= 1
+ 1 1 0 N
1 + +
N N
  
1 + +
= N 1 1
+ + N

N +1
N +1
 
1 +
= N N . (115)
+ +

25