Gaussian Integrals

H. Sonoda 12 May 2019

last update 21 June 2019

1 Real symmetric matrices and hermitian matrices

1.1 Real symmetric matrices
Let S be an N -by-N real symmetric matrix:

Sij = Sji ∈ R (i, j = 1, · · · , N )

S has N real eigenvectors v (i) (i = 1, · · · , N ) with a real eigenvalue λi :

N
(i) (i)
X
Sv (i) = λi v (i) ⇐⇒ Sjk vk = λi vj
k=1

We call S positive when all the eigenvalues are positive:

(∀i) λi > 0

We normalize the eigenvectors so that

N
(i) (j)
X
v (i)T v (j) = δij ⇐⇒ vk vk = δij
k=1

Completeness of the basis v (i) (i = 1, · · · , N ) implies

N N
(i) (i)
X X
v (i) v (i)T = 1N ⇐⇒ vj vk = δjk
i=1 i=1

This means that the N -by-N matrix defined by

 
(j)
O = v (1) , v (2) , · · · , v (N ) ⇐⇒ Oij = vi

is a real orthogonal matrix:

v (1)T
 
 v (2)T  
OT O = 
 (1) (2)
 v , v , · · · , v (N ) = 1N

..
 . 
v (N )T

1
 We wish to show that O diagonalizes S:
 
λ1
 λ2 
OT SO = diag(λ1 , λ2 , · · · , λN ) = 
 
.. 
 . 
λN

(Proof)    
SO = S v (1) , · · · , v (N ) = λ1 v (1) , · · · , λN v (N )

Hence,
 
λ1
v (1)T
 
..
   λ2 
OT SO =  (1)
 λ1 v , · · · , λN v
(N )
=
   
. .. 
 . 
v (N )T
λN

(QED)

1.2 Hermitian matrices

Let H be a hermitian N -by-N matrix satisfying

H † = H ⇐⇒ Hji
∗
= Hij

It has N eigenvectors:
Hv (i) = λi v (i) (i = 1, · · · , N )
with real eigenvalues λi . We call H positive if (∀i) λi > 0. We normalize the eigenvectors
as
v (i)† v (j) = δij
so that
N
X
v (i) v (i)† = 1N
i=1

Let  
U = v (1) , · · · , v (N )

so that
v (1)†
 

U † =  ... 
 

v (N )†
U is unitary since  
U †U = v (i)† v (j) = δij
ij

2
Let us show that U diagonalizes H:

U † HU = diag(λ1 , · · · , λN )

(Proof)    
HU = H v (1) , · · · , v (N ) = λ1 v (1) , · · · , λN v (N )
Hence,
v (1)†
  
λ1
 .. 
 
† (1)
U HU =  .  λ1 v , · · · , λN v (N )
=
 .. 
. 
v (N )† λN

2 Gaussian integrals
2.1 Real Gaussian integrals
We now consider the Gaussian integral
 
Z N Z  
N 1 X
N 1 T
I(A) = d x exp −  xi Aij xj = d x exp − x Ax

2 2
i,j=1

where A is a positive symmetric matrix.

Let O be a real orthogonal matrix that diagonalizes A:

OT AO = diag (a1 , · · · , aN ) (ai > 0)

We change variables from x to y so that

Oy = x ⇐⇒ y = OT x

Under the orthogonal change of variables, the volume element is invariant:

dN x = dN y

(The determinant of O is ±1.)

Since
N
X
T T T
x Ax = y O AOy = ai yi2
i=1
we obtain
N
! N Z N  1
Z ∞   
1X Y 1 Y 2π 2
I(A) = dN y exp − ai yi2 = dyi exp − ai yi2 =
2 −∞ 2 ai
i=1 i=1 i=1
N N
(2π) 2 (2π) 2
= QN √ =√
i=1 ai det A

3
2.2 Complex Gaussian integrals
We next compute the complex Gaussian integral
 
Z N Z  
2N 1 X
∗ 2N 1 †
J(B) = d z exp −  zi Bij zj = d z exp − z Bz

2 2
i,j=1

where B is a positive hermitian matrix, and d2 z = d<zd=z.

This can be calculated analogously. Let U be a unitary matrix that diagonalizes B:
U † BU = diag (b1 , · · · , bN ) (bi > 0)
We then introduce new complex variables by
w = U † z ⇐⇒ z = U w
so that
N
X
z † Bz = w† U † BU w = bi |wi |2
i=1
Since the volume element dN w
is invariant under the unitary transformation, we obtain
N
Z   Z !
2N 1 † 2N 1X 2
J(B) = d z exp − z Bz = d w exp − bi |wi |
2 2
i=1
N
Y 2π (2π)N (2π)N
= = QN =
bi i=1 bi
det B
i=1

3 Exercises
1. Compute Z  
N 1 T T
d x exp − x Ax + j x
2
(Solution) We complete the square:
1 1
T
1
− xT Ax + j T x = − x − A−1 j A x − A−1 j + j T A−1 j
2 2 2
Shifting the integration variables, we obtain
Z  
1
dN x exp − xT Ax + j T x
2
Z  
N 1 −1 T

−1

1 T −1
= d x exp − x − A j A x − A j + j A j
2 2
 Z  
1 T −1 1
= exp j A j dN x exp − xT Ax
2 2
  N
1 T −1 (2π) 2
= exp j A j √
2 det A

4
 2. Compute Z  
2N 1 † † †
d z exp − z Bz + j z + z j
2
(Solution) We complete the square:
1 1
†
1
− z † Bz + j † z + z † j = − z − B −1 j B z − B −1 j + j † B −1 j
2 2 2
Shifting the integration variables, we obtain
Z  
2N 1 † † †
d z exp − z Bz + j z + z j
2
Z  
2N 1 −1 †

−1

1 † −1
= d z exp − z − B j B z − B j + j B j
2 2
 Z  
1 † −1 1
= exp j B j d2N z exp − z † Bz
2 2
(2π)N
 
1 † −1
= exp j B j
2 det B

4 Application to path integrals

We first compute the euclidean path integral of a free particle:

1 T
Z  Z 
2 1
[dx] x(0)=0 exp − dt ẋ(t) = √
x(T )=0 2 0 2πT
(Derivation by the operator formalism)
1 T
Z  Z 
[dx] x(0)=0 exp − dt ẋ(t) = hx = 0| e−T H |x = 0i
2
x(T )=0 2 0
where
1
H = p2
2
Hence,
Z
−T H dp
h0| e |0i = h0| e|−T H
{z } |pi |hp{z
| 0i
2π }
p2 =1
=e−T 2
Z
dp −T p2 1
= e 2 =√ QED
2π 2πT
Using the above result, let us calculate the path integral for a harmonic oscillator:
1 T
Z  Z 
2 2 2


Zω ≡ [dx] x(0)=0 exp − dt ẋ(t) + ω x(t)
x(T )=0 2 0
1 T d2
Z  Z   
= [dx] x(0)=0 exp − dt x(t) − 2 + ω 2 x(t)
x(T )=0 2 0 dt

5
The differential operator
d2
− + ω2
dt2
plays the role of a symmetric matrix A.
We expand r ∞
2X πnt
x(t) = xn sin
T T
n=1
in terms of the normalized sine functions:
πn0 t
Z T
2 πnt
dt sin sin = δn,n0
0 T T T
Using
∞
dx
√ n
Y
[dx] = A
n=1
2π
we obtain
1 T d2
Z  Z   
2
Zω = [dx] x(0)=0 exp − dt x(t) − 2 + ω x(t)
x(T )=0 2 0 dt
∞
Z Y "  #
dxn 1X  πn 2

2
=A √ · exp − +ω
2π 2 T
n=1 n=1
∞
Y 1
=A q
πn 2


n=1
T + ω2

Hence, we obtain the ratio

v !1
∞ u πn 2 ∞

2
r
Zω Y u
T
Y 1 ωT
= t = =
Z0 πn 2 ω2 T 2 sinh ωT


n=1 T + ω2 n=1 1+ π 2 n2

Using
1
Z0 = √
2πT
we obtain r
ω
Zω =
2π sinh ωT
Let us now consider a euclidean path integral:
Z  Z T  
1 2 1 2 2
Z[j] ≡ [dx] x(0)=0 exp − dt ẋ(t) + ω x(t) − j(t)x(t)
x(T )=0 0 2 2
where r
ω
Z[0] = Zω =
2π sinh ωT

6
Integration by parts gives
T
d2
Z  Z    
1 2
Z[j] = [dx] x(0)=0 exp − dt x(t) − 2 + ω x(t) − j(t)
x(T )=0 0 2 dt

As an analog of A−1 , we define G(t, t0 ) by

∂2
 
− 2 + ω G(t, t0 ) = δ(t − t0 )
2
∂t

and
G(0, t0 ) = G(T, t0 ) = 0
This gives
1
G(t, t0 ) = θ(t0 − t) sinh ωt sinh ω(T − t0 ) + θ(t − t0 ) sinh ω(T − t) sinh ωt0


ω sinh ωT
It is legitimate to change variables as
Z T
x(t) −→ x(t) − dt0 G(t, t0 )j(t0 )
0

which satisfies the boundary conditions at t = 0, T . Since

Z T Z T Z T
d2
    
0 0 0 1 2 0 0 0
dt x(t) + dt G(t, t )j(t ) − 2 +ω x(t) + dt G(t, t )j(t )
0 0 2 dt 0
Z T
1 T
Z T
d2
    Z
1
= dt x(t) 2
− 2 + ω x(t) + j(t) + dt dt0 j(t)G(t, t0 )j(t0 )
0 2 dt 2 0 0

we obtain

 Z T Z T 
1 0 0 0
Z[j] = Z[0] exp dt dt j(t)G(t, t )j(t )
2 0 0
r  Z T Z T 
ω 1 0 0 0
= exp dt dt j(t)G(t, t )j(t )
2π sinh ωT 2 0 0