Time-Dependent Statistical Mechanics

12. Quantum correlation functions

c Hans C. Andersen

November 4, 2009

1 Introduction
If A and B are classical variables and we are dealing with classical mechanics Continuation
for a system, it is fairly clear what of Lec-
ture 11
A(t)B(t
) 10/27/09

means for the physical system. It is a number that is a product of two

numbers, each of which means the value of a property at a specific time,
which can be calculated from the location of the system in phase space at
that time. Then it is fairly clear what

A(t)B(t
)

means.
The quantum analog of this is far from obvious from this way of discussing
it. Suppose t > t
. If we measure B at time t
, we perturb the state in an
unpredictable way. This makes it less clear that a subsequent measurement
of A at time t is a quantity that should appear in a physical theory.
We need the quantum analog of A(t)B(t
) and then of its ensemble average.
To define these quantities, we will use the Heisenberg picture of quantum
mechanics. Note that in the present discussion, we are merely laying the
foundations by defining a general set of correlation functions. We will later

1
do a theoretical analysis of various experiments and show that these cor-
relation functions are those that arise in a quantitative discussion of those
experiments.
slide

Quantum correlation functions

In classical mechanics we have:
• dynamical variables, e.g. A(Γ), B(Γ)
• time dependent dynamical variables, e.g. A(t, Γ)
• equilibrium ensemble averages, e.g. A, A(t)B(0)
• nonequilibrium ensemble averages
What are the quantum mechanical concepts that are most closely related to
these classical concepts?
We will define those concepts using the Heisenberg representation of quantum
mechanics.
Then we shall show how these quantities are related to experiments.
end of slide

2 Various formulations of quantum mechan-

ics
General structure of quantum mechanics
In all versions of quantum mechanics,
• the state of the system is represented by a vector in a Hilbert space,
e.g. |ψ,
• a dynamical variable is represented by an operator Ô

2
Schrodinger picture
• states depend on time according to the Schrodinger equation
∂ i
|ψS (t) = − ĤS (t)|ψS (t)
∂t h̄
where
– |ψS (t) is the Schrodinger state vector.
– ĤS (t) = Hamiltonian operator in the Schrodinger picture. (The
only time dependence is explicit, as in the time dependence of an
external field.)
• operators are independent of time (except for any explicit time depen-
dence because of aspects of the system that are under external control).
The typical operator might be denoted ÔS .
• the expectation value of any physical quantity at time t is calculated
from its operator ÔS and the state |ψS (t) as

ψS (t)|ÔS |ψS (t)

The solution of the Schrodinger equation can be expressed in the following

form.
|ψS (t) = Û (t, t
)|ψS (t
) (1)
U (t, t
) is sometimes called the time evolution operator or propagator. It
operates on the wave function appropriate for time t
and generates the wave
function for time t.
We can get the equation of motion for Û by substituting this into the Start of
Schrodinger equation, which gives material
not cov-
∂

i


Û (t, t )|ψS (t ) = − ĤS (t)Û (t, t )|ψS (t ) ered in
∂t h̄ lecture
We want this to be true for all initial states ψS (t ), so the following operator in 2009

equation must hold.

∂ i
Û (t, t
) = − ĤS (t)Û (t, t
)
∂t h̄

3
Moreover, we must have that

Û (t
, t
) = Iˆ

where Iˆ is the identity operator. End of

material
In the special case that the HS is independent of time, the time evolution
not cov-
operator takes on an especially simple form.
ered in
Û (t, t
) = exp (−iHS (t − t
)/h̄) lecture
in 2009
1
as can easily be verified.

Heisenberg picture
In the Heisenberg picture, the state vector is independent of time but the
operators depend on time.
• The Heisenberg state vector |ψH  (independent of time) is the t = 0
value of the Schrodinger state vector.

|ψH  = |ψS (0)

• The Heisenberg operator ÔH is time dependent and is to be defined so

that the expectation value of Ô at time t is

ψH |ÔH (t)|ψH 

We demand that the expectation value calculated in this way should

be consistent with what is obtained in the Schrodinger picture.
So we demand that

ψH |ÔH (t)|ψH  = ψS (t)|ÔS |ψS (t) (2)

1
Exponential operators can be defined in terms of their power series or in terms of their
effects on energy eigenstates. In either case, if ĤS |φ = |φ, then

exp(−iĤS t/h̄)|φ = e−it/h̄ |φ

4
It follows from Eq. (1) that

|ψS (t) = Û (t, 0)|ψS (0)

ψS (t)| = ψS (0)|Û † (t, 0)

Thus the right side of Eq. (2) can be expressed as

ψS (t)|ÔS |ψS (t) = ψS (0)|Û † (t, 0)ÔS U (t, 0)|ψS (0)
= ψS (0)|Û † (t, 0)ÔS U (t, 0)|ψS (0)
= ψH |Û † (t, 0)ÔS U (t, 0)|ψH 

Eq. (2) can be satisfied if we define the Heisenberg operators in the following
way.
ÔH (t) ≡ Û † (t, 0)ÔS U (t, 0)
We adopt this definition.
In the special case that HS is independent of time, this reduces to

ÔH (t) ≡ exp(iĤS t/h̄)ÔS exp(−iĤS t/h̄)

Incidentally, in this special case

ĤH (t) = exp(iĤS t/h̄)ĤS exp(−iĤS t/h̄) = HS

(The exponential of ĤS commutes with ĤS itself, and this result follows.)
Note that the result is independent of time. Thus it would be OK to
write
ÔH (t) ≡ exp(iĤt/h̄)ÔS exp(−iĤt/h̄)
without specifying whether the Ĥ in the exponent is HS or HH (t), since the
two are the same operator. Note that the simple exponential representation
of U holds only in the case that HS is independent of time.

3 Quantum ensembles and quantum ensem-

ble averages
Earlier in the course, we defined precisely what we meant by classical en-
semble averages. We now want to do the same thing for quantum ensemble

5
averages. Let’s restrict our discussion to the canonical ensemble in which N
and V are fixed. Classically, this has the effect of saying that all systems in
the ensemble have the same Hamiltonian and can be represented as points in
the same phase space. Quantum mechanically, this has the consequence that
all systems in the ensemble have the same Hamiltonian and can be expanded
in terms of the same eigenfunctions of the Hamiltonian.

In our discussion of classical ensemble averages, the basic elements of the

discussion were:
1. The state of a system prepared according to a specific prescription is a
random variable.
2. The state is specified by a point in phase space, Γ.
3. The quantity being averaged is a dynamical variable represented by a
function of Γ, e.g. A(Γ).
4. To calculate averages we need the probability distribution function of
Γ. Call it P (Γ).
5. The ensemble average of A(Γ) is

A = dΓ P (Γ)A(Γ)

6. For a preparation method that involves fixing the temperature of the

system, the appropriate P is

P (Γ) = (const.)e−H(Γ)/kB T

These are the essential postulates of the classical canonical ensemble. End of
Lecture
11
10/27/09

6
We now want to discuss the quantum mechanical analog of this. Lecture
12
1. The state of a system prepared according to a specific prescription is a
10/29/09
random variable. (This is the same as the classical statement.)
2. The state of the system is specified by a wave function. In general the
wave function can be expanded in eigenstates of the Hamiltonian. Let
φi satisfy
Hφi = Ei φi
The wave function can then be expressed as

ψ= ci φi
i

where the sum is over all of the eigenstates of the Hamiltonian. Thus
the set of coefficients, ci for all i, is what it takes to specify the wave
function.
3. The quantity being averaged is a dynamical variable represented by a
quantum mechanical operator Â. The quantum expectation value of A
for a well defined quantum state ψ is

ψ|Â|ψ

Note that this is a quantum mechanical expectation value, so it already

is an average of some sort even for one system, but this is the quantity
that must averaged over all systems in the ensemble.
Using the basis function expansion for the wave function we get

ψ|Â|ψ = c∗i cj φi |Â|φj 
i,j

and this is to be averaged over all systems in the ensemble.

4. Note that all systems in the ensemble have the same basis functions
and the same matrix elements. The only thing that varies from one
system to another is the values of the coefficients ci . To calculate the
desired average, we need the ensemble average of c∗i cj .
  
ψ|Â|ψ = c∗i cj  φi |Â|φj 
i,j

7
 (By analogy with the classical case, you might have thought that we
needed the distribution function of the set of coefficients ci , i = 1 . . . n,
but it is now clear that all we really need is averages of products of the
form c∗i cj .)
Let us define
ρji = cj c∗i  = c∗i cj 
(Note the order of the factors and note which is complex conjugated.)
This might or might not be the quantity you might initially want to
define, but with this definition, we have the following compact result.
  
ψ|Â|ψ = ρji Aij
i,j

The set of quantities ρij represent a matrix that is called the ‘density
matrix’. The set of matrix elements Aij are also a matrix, and the sum
on the right is the trace of the matrix product of the ρ and A matrices.2
Thus we can write  
ψ|Â|ψ = Tr ρA
But any matrix can also be regarded as an operator (i.e. the operator
whose matrix elements are the elements of the matrix). Thus, the
density matrix can be regarded as the matrix of a density operator ρ̂.
So we also can write
 
ψ|Â|ψ = Tr(ρ̂Â)

where the quantity in parentheses is an operator product of two oper-

ators and hence is an operator. Here we are using the convention that
2
The trace of a matrix is the sum of its diagonal elements. Thus, the trace of a matrix
B is 
TrB = Bjj
j

If the matrix B is the matrix product of two matrices C and D, then


Bjk = Cji Dik
i

and 
TrB = Cji Dij
ji

8
 the trace of an operator is the trace of the matrix of the operator, or
the sum of the diagonal elements of the operator.
slide
Let us define
ρji = cj c∗i  = c∗i cj 
Then   
ψ|Â|ψ = ρji Aij
i,j

The sum on the right is the trace of the matrix product of the ρ and
A matrices. Thus  
ψ|Â|ψ = Tr ρA
But any matrix can also be regarded as an operator (i.e. the operator
whose matrix elements are the elements of the matrix). Thus we also
can write  
ψ|Â|ψ = Tr(ρ̂Â)

end of slide
5. For a preparation method that involves fixing the temperature of the
system, the appropriate density matrix is (assumed to be)

ρ̂ = (const) × exp(−β Ĥ)

In other words,

ρii = (const) × exp(−Ei /kB T )

ρij = 0 for i = j

The ii diagonal element of ρ has meaning as the ensemble average of

the probability that the system is in state i, and this probability is
assumed to be proportional to the Boltzmann factor for the energy of
state i. The assumption that the off-diagonal elements are zero would
require some discussion to justify. We will not discuss this here. This
assumption for the form of the density matrix in equilibrium quantum
systems with specific N , V , and T is a basic postulate of equilibrium
quantum statistical mechanics.

9
Thus, if we have any quantum mechanical operator Ô, we define the ensemble
average of the operator Ô as

Ô = Tr ρ̂Ô = ρij Oji
ij

If the ensemble is an equilibrium canonical ensemble, then

ρ̂ = (const) × exp(−Ĥ/kB T )

4 Definition of a quantum mechanical corre-

lation function
Proposed definition of a quantum mechanical correlation function

For a system at equilibrium we define an average of the form

ÂH (t)B̂H (t
)

to be a quantum time correlation function of A and B.

You might think that this is a reasonable candidate for the definition of
a quantum equilibrium correlation function. In fact it is reasonable, and
functions of this type have some but not all of the properties of classical
equilibrium correlation functions.

Stationarity property. Such correlation functions depend only on the

time interval t − t
, not the individual time arguments. Let’s show this.
 
ÂH (t)B̂H (t
) = T r ρ̂ÂH (t)B̂H (t
)
 
= (const)T r exp(−β Ĥ) exp(iĤt)ÂS exp(−iĤt) exp(iĤt
)B̂S exp(−iĤt
)

10
Some elementary properties of operators
1. For any pair of operators Ĉ and D̂,

T r(Ĉ D̂) = T r(D̂Ĉ)

This is true even if the operators do not commute.

2. For any pair of operators Ĉ and D̂ that commute,

exp Ĉ exp D̂ = exp D̂ exp Ĉ

3. For any operator Ĉ and any two numbers a and b

exp(aĈ) exp(bĈ) = exp((a + b)Ĉ)

Back to the calculation above.

Move the exponential on the far right to the far left. This does not change
the value, because of the first property.
Commute it with the exp(−β Ĥ). This does not change the value because of
the second property.
We get

ÂH (t)B̂H (t
)

= (const)T r(exp(−β Ĥ) exp(−iĤt
) exp(iĤt)ÂS exp(−iĤt) exp(iĤt
)B̂S )
= (const)T r(exp(−β Ĥ) exp(iĤ(t − t
))ÂS exp(−iĤ(t − t
))B̂S )

Thus we have shown that the equilibrium correlation function depends on

the time interval not the absolute times. This is analogous to the behavior
of classical time correlation functions.

11
Thus, we can define
 
SAB (t − t
) ≡ ÂH (t)B̂H (t
)

But this function lacks one important feature that the classical correlation
function has.

Another elementary property of operators.

4. ÂH (t) and B̂H (t
) do not in general commute. (The only common situa-
tion in which they commute is if ÂS commutes with B̂S and t = t
.)
Since these operators do not in general commute, it is not in general true
that
SAB (t − t
) = SBA (t
− t)

Accordingly, we define a symmetrized version of S by

1
φAB (t) ≡ (SAB (t) + SBA (−t))
2
It follows that
φAB (t) = φBA (−t)
which is analogous to the property of classical CAB (t).
We choose these definitions for several reasons.
• In quantum mechanical theories of absorption spectroscopy, scattering
experiments, and relaxation processes, it is found that the measured
quantities are often related to correlation functions defined in this way.
They play the same role as the classical correlation functions do in
classical theories of the same processes.
• In the limit that quantum effects are negligible, the behavior of φAB (t)
and SAB (t) are similar to that of the classical correlation function
CAB (t).
• The function φAB (t) has one of the symmetry properties of the classical
CAB (t).
Note that there is another quantum correlation function that is often called
CAB (t). We may or may not introduce it later.

12
5 General expression for a quantum correla-
tion function
We can obtain a useful general expression for an equilibrium time correlation
function in the canonical ensemble in the following way.
We use the fact that the Hamiltonian has no explicit time dependence.
We use a complete set of eigenfunctions |i of Ĥ.
Ĥ|i = Ei |i
The set is complete. We assume the states are normalized.
The identity operator can be expressed as

Iˆ = |ii|
i

Let ÂS and B̂S be Hermitian operators. Their matrix elements in this basis
are
Aij = i|ÂS |i
and similarly for Bij .
Then the S correlation function is

SAB (t − t
) = ρii i|ÂH (t)B̂H (t
)|i
i



= ρii i|eiĤt/h̄ ÂS e−iĤt/h̄ eiĤt /h̄ B̂S e−iĤt /h̄ |i
i



= ρii eiEi (t−t )/h̄ i|ÂS e−iĤt/h̄ eiĤt /h̄ B̂S |i
i



= ρii eiEi (t−t )/h̄ i|ÂS e−iĤ(t−t )/h̄ B̂S |i
i



= ρii eiEi (t−t )/h̄ i|ÂS |jj|e−iĤ(t−t )/h̄ B̂S |i
i j

iEi (t−t
)/h̄

= ρii e i|ÂS |jj|e−iEj (t−t )/h̄ B̂S |i
ij


= ρii e−i(Ej −Ei )(t−t )/h̄ i|ÂS |jj|B̂S |i
ij


= ρii e−i(Ej −Ei )(t−t )/h̄ Aij Bji
ij

13
It is clear that the result is a function of the time interval rather than the
two separate time arguments.
If we specialize to A = B, we have


SAA (t) = ρii e−i(Ej −Ei )(t−t )/h̄ Aij Aji
ij


= ρii e−i(Ej −Ei )(t−t )/h̄ Aij A∗ij
ij


= ρii e−i(Ej −Ei )(t−t )/h̄ |Aij |2
ij

• Every correlation function is a sum of complex exponentials

– with positive definite weights that come from the Boltzmann dis-
tribution of state energies
– with amplitudes that depend on the matrix elements of the two
operators.
• The frequency (Ej − Ei )/h̄ appears if:
– state i is occupied at the temperature of interest, and
– the matrix element Aij is nonzero, and
– the matrix element Bji is nonzero.
• For t = 0
– 
SAB (0) = ρii Aij Bji = ÂS B̂S 
ij

– 
SAA (0) = ρii |Aij |2
ij

The latter is positive definite.

14
slide



SAB (t − t
) = ρii e−i(Ej −Ei )(t−t )/h̄ Aij Bji
ij


SAA (t) = ρii e−i(Ej −Ei )(t−t )/h̄ |Aij |2
ij

end of slide
We will ultimately need the Fourier transform of the correlation function, so
we may as well calculate it now.

ŜAB (ω) = 2π ρii δ (ω − (Ej − Ei )/h̄) Aij Bji
ij


ŜAA (ω) = 2π ρii δ (ω − (Ej − Ei )/h̄) |Aij |2
ij

Note that the latter is nonnegative for all frequencies, just like in the classical
case. Lecture
12 con-
We have not yet related this correlation function to absorption spectrum, but
tinued
it will turn out that the ŜAA (ω) is closely related to absorption spectra, and
in N13
it should look familiar. The absorption spection is a set of sharp lines at the
usual frequencies with intensities that depend on the equilibrium populations
of the various states and the matrix elements of various operators.

15