15

Functional Derivatives
15.1 Functionals
A functional G[f] is a map from a space of functions to a set of numbers.
For instance, the action functional S[q] for a particle in one dimension maps
the coordinate q(t), which is a function of the time t, into a numberthe
action of the process. If the particle of mass m is moving slowly and freely,
then for the interval (t
1
, t
2
) its action is
S
0
[q] =
_
t
2
t
1
dt
m
2

dq(t)
dt

2
. (15.1)
If the particle is moving in a potential V (q(t)), then its action is
S[q] =
_
t
2
t
1
dt
_
m
2

dq(t)
dt

2
V (q(t))
_
. (15.2)
15.2 Functional Derivatives
A functional derivative is a functional
G[f][h] =
d
d
G[f + h]

=0
(15.3)
of a functional. For instance, if G
n
[f] is the functional
G
n
[f] =
_
dxf
n
(x) (15.4)
15.2 Functional Derivatives 619
then its functional derivative is the functional that maps the pair of functions
f, h to the number
G
n
[f][h] =
d
d
G
n
[f + h]

=0
=
d
d
_
dx(f(x) + h(x))
n

=0
=
_
dxnf
n1
(x)h(x). (15.5)
Physicists often use the less elaborate notation
G[f]
f(y)
= G[f][(x y)] (15.6)
in which the second function h(x) is (xy). Thus, in the preceding example
G[f]
f(y)
=
_
dxnf
n1
(x)(x y) = nf
n1
(y). (15.7)
Functional derivatives of functionals that involve powers of derivatives
also are easily dealt with. Suppose that the functional involves the square
of the derivative f
0
(x)
G[f] =
_
dx
_
f
0
(x)
_
2
. (15.8)
Then its functional derivative is
G[f][h] =
d
d
G[f + h]

=0
=
d
d
_
dx
_
f
0
(x) + h
0
(x)
_
2

=0
=
_
dx2f
0
(x)h
0
(x) = 2
_
dxf
00
(x)h(x) (15.9)
in which we have integrated by parts and used suitable boundary conditions
to drop the surface terms. In physics notation, we have
G[f]
f(y)
= 2
_
dxf
00
(x)(x y) = 2f
00
(y). (15.10)
620 Functional Derivatives
Lets now compute the functional derivative of the action (15.2), which in-
volves the square of the time-derivative q(t) and the potential energy V (q(t))
S[q][h] =
d
d
S[q + h]

=0
=
d
d
_
dt
_
m
2

q(t) +

h(t)

2
V (q(t) + h(t))
_

=0
=
_
dt
_
m q(t)

h(t) V
0
(q(t))h(t)
_
=
_
dt

m q(t) V
0
(q(t))

h(t) (15.11)
where we once again have integrated by parts and used suitable boundary
conditions to drop the surface terms. In physics notation, this is
S[q]
q(t
0
)
=
_
dt

m q(t) V
0
(q(t))

(t t
0
) = m q(t
0
) V
0
(q(t
0
)). (15.12)
In these terms, the stationarity of the action S[q] is the vanishing of its
functional derivative either in the form
S[q][h] = 0 (15.13)
for arbitrary functions h(t) (that satisfy the boundary conditions) or equiv-
alently in the form
S[q]
q(t)
= 0 (15.14)
which is Lagranges equation of motion
m q = V
0
(q). (15.15)
Heres a shortcut to the functional derivative in the notation of physics:
G[f]
f(y)
=
d
d
G[f +
y
]

=0
(15.16)
in which the function h(x) is replaced by
y
(x) = (x y).
Example 15.1 (Shortest Path is a Straight Line) On a plane, the length
of the path (x, y(x)) from (x
0
, y
0
) to (x
1
, y
1
) is
L[y] =
_
x
1
x
0
_
dx
2
+ dy
2
=
_
x
1
x
0
_
1 + y
02
dx. (15.17)
15.3 Higher-Order Functional Derivatives 621
The shortest path y(x) minimizes this length L[y]
L[y]
y
=
d
d
L[y + h]

=0
=
d
d
_
x
1
x
0
_
1 + (y
0
+ h
0
)
2
dx

=0
=
_
x
1
x
0
y
0
h
0
_
1 + y
02
dx =
_
x
1
x
0
h
d
dx
y
0
_
1 + y
02
dx (15.18)
since h(x) satises h(x
0
) = h(x
1
) = 0. Dierentiating, we nd
L[y]
y
=
_
x
1
x
0
h
y
00
1 + y
02
dx (15.19)
which can vanish for arbitrary functions h(x) only if y
00
= 0, which is to say
only if y(x) is a straight line, y = mx + b.
15.3 Higher-Order Functional Derivatives
The second functional derivative is

2
G[f][h] =
d
2
d
2
G[f + h]|
=0
. (15.20)
So if G
N
[f] is the functional
G
N
[f] =
_
f
N
(x)dx (15.21)
then

2
G
N
[f][h] =
d
2
d
2
G
N
[f + h]|
=0
=
d
2
d
2
_
(f(x) + h(x))
N
dx

=0
=
d
2
d
2
_
N
2

2
h
2
(x)f
N2
(x) dx

=0
= N(N 1)
_
f
N2
(x)h
2
(x)dx. (15.22)
Example 15.2 (
2
S
0
) The second functional derivative of the action S
0
[q]
(15.1) is

2
S
0
[q][h] =
d
2
d
2
_
t
2
t
1
dt
m
2

dq(t)
dt
+
dh(t)
dt

=0
=
_
t
2
t
1
dt m

dh(t)
dt

2
0 (15.23)
622 Functional Derivatives
and is positive for all functions h(t). The stationary classical trajectory
q(t) =
t t
1
t
2
t
1
q(t
2
) +
t
2
t
t
2
t
1
q(t
1
) (15.24)
is a minimum of the action S
0
[q].
The second functional derivative of the action S[q] (15.2) is

2
S[q][h] =
d
2
d
2
_
t
2
t
1
dt
_
m
2

dq(t)
dt
+
dh(t)
dt

2
V (q(t) + h(t))
_

=0
=
_
t
2
t
1
dt
_
m

dh(t)
dt

2
V (q(t))
q
2
(t)
h
2
(t)
_
(15.25)
and it can be positive, zero, or negative. Chaos sometimes arises in systems
of several particles when the second variation of S[q] about a stationary path
is negative,
2
S[q][h] < 0 while S[q][h] = 0.
The nth functional derivative is dened as

n
G[f][h] =
d
n
d
n
G[f + h]|
=0
. (15.26)
The nth functional derivative of the functional (15.21) is

n
G
N
[f][h] =
N!
(N n)!
_
f
Nn
(x)h
n
(x)dx. (15.27)
15.4 Functional Taylor Series
It follows from the Taylor-series theorem (section 4.6) that
e

G[f][h] =

n=0

n
n!
G[f][h] =

n=0
1
n!
d
n
d
n
G[f + h]

=0
= G[f + h] (15.28)
which illustrates an advantage of the present mathematical notation.
The functional S
0
[q] of Eq.(15.1) provides a simple example of the func-
15.5 Functional Dierential Equations 623
tional Taylor series (15.28):
e

S
0
[q][h] =

1 +
d
d
+
1
2
d
2
d
2

S
0
[q + h]

=0
=
m
2
_
t
2
t
1

1 +
d
d
+
1
2
d
2
d
2

q(t) +

h(t)

2
dt

=0
=
m
2
_
t
2
t
1

q
2
(t) + 2 q(t)

h(t) +

h
2
(t)

dt
=
m
2
_
t
2
t
1

q(t) +

h(t)

2
dt = S
0
[q + h]. (15.29)
If the function q(t) makes the action S
0
[q] stationary, and if h(t) is smooth
and vanishes at the endpoints of the time interval, then
S
0
[q + h] = S
0
[q] + S
0
[h] (15.30)
in which the functions q(t) and h(t) respectively satisfy the boundary con-
ditions q(t
i
) = q
i
and h(t
1
) = h(t
2
) = 0.
More generally, if q(t) makes the action S[q] stationary, and h(t) is any
loop from and to the origin, then
S[q + h] = e

S[q][h] = S[q] +

n=2
1
n!
d
n
d
n
S[q + h]|
=0
. (15.31)
If further S
2
[q] is purely quadratic in q and q, like the harmonic oscillator,
then
S
2
[q + h] = S
2
[q] + S
2
[h]. (15.32)
15.5 Functional Dierential Equations
In inner products like hq
0
|fi, we represent the momentum operator as
p =
~
i
d
dq
0
(15.33)
because then
hq
0
|p q|fi =
~
i
d
dq
0
hq
0
|q|fi =
~
i
d
dq
0
_
q
0
hq
0
|fi
_
=

~
i
+ q
0
~
i
d
dq
0

hq
0
|fi (15.34)
which respects the commutation relation [q, p] = i~.
624 Functional Derivatives
So too in inner products h
0
|fi of eigenstates |
0
i of (x, t)
(x, t)|
0
i =
0
(x)|
0
i (15.35)
we can represent the momentum (x, t) canonically conjugate to the eld
(x, t) as the functional derivative
(x, t) =
~
i

0
(x)
(15.36)
because then
h
0
|(x
0
, t)(x, t)|fi =
~
i

0
(x
0
)
h
0
|(x, t)|fi
=
~
i

0
(x
0
)
_

0
(x)h
0
|fi
_
(15.37)
=
~
i

0
(x
0
)
_
(x x
0
)
0
(x
0
) d
3
x
0
h
0
|fi

=
~
i

(x x
0
) +
0
(x)

0
(x
0
)

h
0
|fi
= h
0
| i~(x x
0
) + (x, t) (x
0
, t)|fi
which respects the equal-time commutation relation
[(x, t), (x
0
, t)] = i ~ (x x
0
). (15.38)
We can use the representation (15.36) for (x) to nd the wave function
of the ground state |0i of the hamiltonian
H =
1
2
_

2
+ ()
2
+ m
2

d
3
x (15.39)
where we set ~ = c = 1. We will use the trick we used in section 2.11 to nd
the ground state |0i of the harmonic-oscillator hamiltonian
H
0
=
p
2
2m
+
m
2
q
2
2
. (15.40)
In that trick, one writes
H
0
=
1
2m
(mq ip)(mq + ip) +
i
2
[p, q]
=
1
2m
(mq ip)(mq + ip) +
1
2
~ (15.41)
and seeks a state |0i that is annihilated by mq + ip
hq
0
|mq + ip|0i =

mq
0
+ ~
d
dq
0

hq
0
|0i = 0. (15.42)
15.5 Functional Dierential Equations 625
The solution to this dierential equation
d
dq
0
hq
0
|0i =
mq
0
~
hq
0
|0i (15.43)
is
hq
0
|0i =

m
~

1/4
exp

mq
02
2~

(15.44)
in which the prefactor is a constant of normalization.
So extending that trick to the hamiltonian (15.39), we factor H
H =
1
2
_
_
_

2
+ m
2
i
_ _
_

2
+ m
2
+ i
_
d
3
x + C (15.45)
in which C = (i/2)
_
[,
_
4+ m
2
] d
3
x is an (innite) constant. The
ground state |0i of H therefore must satisfy the functional dierential equa-
tion h
0
|

2
+ m
2
+ i|0i = 0 or
h
0
|0i

0
(x)
=
_

2
+ m
2

0
(x) h
0
|0i. (15.46)
The solution to this functional dierential equation is
h
0
|0i = N exp

1
2
_

0
(x)
_

2
+ m
2

0
(x) d
3
x

(15.47)
in which N is a normalization constant. To see that this functional does
satisfy equation (15.46), we compute the derivative
dh
0
+ h|0i
d
= N
d
d
exp
_

1
2
_
_

0
+ h
_
_
4+ m
2
_

0
+ h
_
d
3
x
_
(15.48)
which at = 0 is
dh
0
+ h|0i
d

=0
=
1
2
__
h(x)
_
4+ m
2

0
(x)
3
x
+
_

0
(x)
_
4+ m
2
h(x) d
3
x
_
h
0
|0i.
(15.49)
We integrate the second term by parts and drop the surface terms because
the smooth function h goes to zero quickly as its arguments go to innity.
We then have
dh
0
+ h|0i
d

=0
=
_
h(x
0
)
_
4+ m
2

0
(x
0
) d
3
x
0
h
0
|0i. (15.50)
Letting h(x
0
) =
(3)
(x
0
x), we arrive at (15.46).
626 Functional Derivatives
The spatial Fourier transform

0
(p)

0
(x) =
_
e
ipx

0
(p)
d
3
p
(2)
3
(15.51)
satises

0
(p) =

0
(p) since
0
is real. In terms of it, the ground-state
wave-function is
h
0
|0i = N exp

1
2
_
|

0
(p)|
2
_
p
2
+ m
2
d
3
p

. (15.52)
Example 15.3 (Other Theories, Other Vacua) We can nd exact ground
states for interacting theories with hamiltonians like
H =
1
2
_
_
_

2
+ m
2
+ c
n

n
i
_ _
_

2
+ m
2
+ c
n

n
+ i
_
d
3
x.
(15.53)
The state |i will be an eigenstate of H with eigenvalue zero if
h
0
|i

0
(x)
=
_
_

2
+ m
2

0
(x) + c
n

0n
_
h
0
|i. (15.54)
By extending the argument of equations (15.4515.50), one may show (ex-
ercise 15.3) that the wave functional of the vacuum is
h
0
|i = N exp
_

_
1
2

0
_

2
+ m
2

0
+
c
n
n + 1

0n+1

d
3
x
_
(15.55)
which is normalizable only when n is odd.
Exercises
15.1 Compute the action S
0
[q] (15.1) for the classical path (15.24).
15.2 Use (15.25) to nd a formula for the second functional derivative of the
action (15.2) of the harmonic oscillator for which V (q) = m
2
q
2
/2.
15.3 Show that (15.55) satises (15.54).
15.4 Derive (15.52) from equations (15.47 & 15.51).