Joseph Maciejko

An Introduction to
Nonequilibrium Many-Body
Theory
October 25, 2007

Springer
Contents

1 Nonequilibrium Perturbation Theory . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Failure of Conventional Time-Ordered Perturbation Theory . . . 1
1.1.1 Equilibrium Many-Body Theory . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Failure of Conventional Techniques . . . . . . . . . . . . . . . . . . 2
1.2 The Schwinger-Keldysh Contour . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Closed Time Path Contour . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Interactions: the Kadanoff-Baym Contour . . . . . . . . . . . . . . . . . . 8
1.4.1 Neglect of Initial Correlations and Schwinger-Keldysh
Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Contour Dyson Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.6 Initial Correlations with Arbitrary Initial Density Matrix . . . . . 12
1.7 Relation to Real-Time Green’s Functions . . . . . . . . . . . . . . . . . . . 12
1.7.1 Larkin-Ovchinnikov Representation . . . . . . . . . . . . . . . . . . 13
1.7.2 Langreth Theorem of Analytic Continuation . . . . . . . . . . 14

2 Quantum Kinetic Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.1 Keldysh Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Kadanoff-Baym Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Wigner Representation and Gradient Expansion . . . . . . . 23
2.3 Quantum Boltzmann Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.1 QBE with Electric Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.2 QBE with Electric and Magnetic Field . . . . . . . . . . . . . . . 28
2.3.3 One-Band Spinless Electrons . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.1 Nonequilibrium Transport through a Quantum Dot . . . . . . . . . . 33
3.1.1 Expression for the Current . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.2 Perturbation Expansion for the Mixed Green’s Function 35
3.1.3 Path Integral Derivation of the Mixed Green’s Function 37
3.1.4 General Expression for the Current . . . . . . . . . . . . . . . . . . 39
3.1.5 Noninteracting Quantum Dot . . . . . . . . . . . . . . . . . . . . . . . 41
VI Contents

3.1.6 Interacting Quantum Dot: Anderson Model and

Coulomb Blockade . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Linear Response for Steady-State and Homogeneous Systems . 44
3.2.1 Example: Conductivity from Impurity Scattering . . . . . . 48
3.3 One-Band Electrons with Spin-Orbit Coupling . . . . . . . . . . . . . . 50
3.4 Classical Boltzmann Limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
1
Nonequilibrium Perturbation Theory

The goal of this chapter is to construct the perturbation expansion for the
1-particle contour-ordered Green’s function. First, we explain why the usual
perturbation expansion on the Feynman contour (the real-time axis from −∞
to ∞) or on the Matsubara contour (a segment on the imaginary-time axis
from −iβ to iβ) fails in general nonequilibrium situations.

1.1 Failure of Conventional Time-Ordered Perturbation

Theory
The central goal of nonequilibrium many-body theory is to calculate real-time
correlation functions. For example, we might want to calculate the 1-particle
time-ordered Green’s function,

iG(x, t; x0 , t0 ) = hT [ψ(x, t)ψ † (x0 , t0 )]i = Tr ρT [ψ(x, t)ψ † (x0 , t0 )] (1.1)

in the Heisenberg picture, where ρ is an arbitrary nonequilibrium density

matrix and the Hamiltonian H(t) is in general time-dependent.

1.1.1 Equilibrium Many-Body Theory

In conventional equilibrium many-body theory, we know how to setup a per-

turbation theory to calculate this quantity. At zero temperature, the density
matrix is
ρ = |Ψ0 ihΨ0 | (1.2)
where |Ψ0 i is the exact ground state of a full interacting but time-independent
Hamiltonian H. Then the time-ordered Green’s function

iG(x, t; x0 , t0 ) = hΨ0 |T [ψ(x, t)ψ † (x0 , t0 )]|Ψ0 i (1.3)

is given by conventional Feynman-Dyson perturbation theory,

2 1 Nonequilibrium Perturbation Theory

hΦ0 |T [S(∞, −∞)ψ̂(x, t)ψ̂ † (x0 , t0 )]|Φ0 i

iG(x, t; x0 , t0 ) = (1.4)
hΦ0 |S(∞, −∞)|Φ0 i

where the caret denotes operators in the interaction picture with respect to
a quadratic Hamiltonian H0 where H = H0 + V , |Φ0 i is the ground state of
H0 , and the S-matrix is
 Z ∞ 
S(∞, −∞) = T exp −i dt1 V̂ (t1 ) (1.5)
−∞

The power series expansion of the S-matrix and the subsequent use of Wick’s
theorem generates the usual Feynman diagrams, of which the disconnected
ones cancel against the phase factor in the denominator.
At finite temperatures, we use the coincidence in functional form of the
thermal density matrix
e−βH
ρ= (1.6)
Z
where Z = Tr e−βH is the partition function, and the evolution operator
U (t) = e−iHt to setup a perturbation theory in imaginary time. Here again,
Wick’s theorem can be used since each term in the perturbation expansion
is an average over a noninteracting density matrix ρ0 = e−βH0 /Z0 where
Z0 = Tr e−βH0 .

1.1.2 Failure of Conventional Techniques

These techniques seem quite powerful, so why can’t we apply them to the
more general problem of Eq. (1.1)?
Let us consider the Matsubara technique first. Out of equilibrium, there
is no such thing as a temperature. As a result, in general the density matrix
is not of exponential form, so there is no way we can use the Matsubara
trick which consists in a simultaneous expansion of the density matrix and
the time-evolution operator allowed by the coincidence in functional form of
these two operators.
In the Feynman case, once again the density matrix is not a simple pro-
jector on the ground state as in Eq. (1.2).
These arguments are correct, but it is instructive to see in more detail
where exactly does the mathematical construction of the above equilibrium
perturbation expansions fail. We will try to construct an expansion of the
nonequilibrium Green’s function using the ordinary Feynman approach and
see that it fails.
Consider a generic density matrix
X
ρ= pΦ |ΦihΦ| (1.7)
Φ
 1.1 Failure of Conventional Time-Ordered Perturbation Theory 3

where the |Φi can be arbitrary quantum states. First recall that in the Heisen-
berg picture, the density matrix does not evolve in time since its time evolu-
tion, given by the quantum Liouville equation, goes in a way opposite to that
given by the Heisenberg equation of motion, so that the time evolution of ρ
cancels out altogether. For convenience, we study one state at a time, namely
we want to calculate the expectation value

iGΦ (x, t; x0 , t0 ) = hΦ|T [ψ(x, t)ψ † (x0 , t0 )]|Φi (1.8)

P
and we can obtain the full Green’s function by G = Φ pΦ GΦ .
Now consider the following partition of the Hamiltonian H(t) = H + H 0 (t)
where H is the unperturbed equilibrium Hamiltonian (but may still contain
interactions) and all the time dependence is included in the nonequilibrium
perturbation H 0 (t). The field operator in the interaction picture is

ψ̂(x, t) = eiHt ψ(x)e−iHt (1.9)

and in the Heisenberg picture,

ψ(x, t) = U † (t)ψ(x)U(t) = S(0, t)ψ̂(x, t)S(t, 0) (1.10)

where S(t, 0) = eiHt U(t) and the evolution operator is

 Z t 
U(t) = T exp −i dt1 H(t1 ) (1.11)
0

more generally, we have the S-matrix

 Z t 
0 iHt 0 −iHt0 0
S(t, t ) = e U(t, t )e = T exp −i dt1 Ĥ (t1 ) (1.12)
t0
 R 
t
where U(t, t0 ) = T exp −i t0 dt1 H(t1 ) and Ĥ 0 (t) = eiHt H 0 (t)e−iHt is the
nonequilibrium perturbation in the interaction picture1 .
We substitute these relations into Eq. (1.8),

iGΦ (x, t; x0 , t0 ) = hΦ|T [S(0, t)ψ̂(x, t)S(t, 0)S(0, t0 )ψ̂ † (x0 , t0 )S(t0 , 0)]|Φi (1.13)

Now the interaction picture state is given by

|Φ(t)iI = S(t, 0)|Φ(0)iI = S(t, 0)|Φi (1.14)

since all pictures coincide at t = 0. We can then say that

1
We use curly letters for U and S to indicate that these operators take care of the
full nonequilibrium Hamiltonian H with H 0 as the perturbation. Later upright
letters U and S will be used for the corresponding operators taking care of the
equilibrium Hamiltonian H with whatever interactions it may contain as the
perturbation.
4 1 Nonequilibrium Perturbation Theory

|Φi = S(0, ±∞)|Φ(±∞)iI (1.15)

Putting everything together, we have

iGΦ (x, t; x0 , t0 ) = hΦ(∞)|I S(∞, 0)

×T [S(0, t)ψ̂(x, t)S(t, t0 )ψ̂ † (x0 , t0 )S(t0 , 0)]S(0, −∞)|Φ(−∞)iI (1.16)

Now, the S-matrix S(∞, −∞) is itself a time-ordered product, so that we can
write

iGΦ (x, t; x0 , t0 ) = hΦ(∞)|I T [S(∞, −∞)ψ̂(x, t)ψ̂ † (x0 , t0 )]|Φ(−∞)iI (1.17)

making use of the group property of S(t, t0 ).

So far everything is correct in a general nonequilibrium setting. At this
point, the conventional Feynman expression Eq. (1.4) is obtained only provided
that the following crucial step holds: |Φ(∞)iI and |Φ(−∞)iI differ only by a
phase factor. This is true in equilibrium but breaks down out of equilibrium.
The reason is the following. In the conventional zero-temperature theory,
we adiabatically switch on and off the interaction:

H = H + e−|t| H 0 (1.18)

where  is a positive infinitesimal. We also choose |Φi to be |Ψ0 i, the ground

state of H = H + H 0 . We assume this ground state to be nondegenerate. Since
the time evolution is adiabatic, we can use the adiabatic theorem to say that
|Φ(±∞)iI = lim→0 S (±∞, 0)|Ψ0 i are both eigenstates of H. Since |Ψ0 i is
nondegenerate, these two states can differ by at most a phase factor,

|Φ(∞)iI = eiL |Φ(−∞)iI (1.19)

For definiteness, using the notation |Φ0 i ≡ |Φ(−∞)iI it is easy to see that

eiL = hΦ0 |S(∞, −∞)|Φ0 i (1.20)

so that Eq. (1.17) gives Eq. (1.4) directly.

Out of equilibrium, the use of the adiabatic theorem is unjustified since
under the assumption of a general time-dependent Hamiltonian H(t) the
time evolution is not adiabatic. In the case that the Hamiltonian is time-
independent but the density matrix is still arbitrary – the important case of
a nonequilibrium steady state, we still cannot use the adiabatic theorem be-
cause the generic quantum states |Φi are not necessarily eigenstates of H and
the adiabatic theorem applies only to eigenstates of the Hamiltonian. If we
insist and expand the density matrix in the basis of eigenstates of H, it is in
general not diagonal in this basis,
X X
Tr ρO = pΦ hΦ|O|Φi = ρnn0 hn|O|n0 i (1.21)
Φ nn0
 1.2 The Schwinger-Keldysh Contour 5

so that even if |n(∞)iI = eiLn |n(−∞)iI , the appearance of crossed terms

spoils the expansion. Finally, even if the ground state is assumed to be nonde-
generate, excited states |ni appearing in the nonequilibrium density matrix ρ
can be degenerate so they can be mixed by a non-Abelian Berry phase under
adiabatic evolution, which once again invalidates the conventional procedure.
In brief, out of equilibrium |Φ(∞)iI and |Φ(−∞)iI are not simply related.
This is obvious physically since when a system is driven out of equilibrium,
its asymptotic future is quite different from its initial preparation in the re-
mote past. The purpose of the various nonequilibrium contours such as the
two-branch Schwinger-Keldysh contour is precisely to avoid any reference to
the state in the asymptotic future |Φ(∞)iI and base the expansion solely on
|Φ(−∞)iI , the state in the asymptotic past.

1.2 The Schwinger-Keldysh Contour

This idea was pioneering by Schwinger [1], Craig [2], Mills [3], and popularized
by Keldysh [4]. To get rid of the unwanted asymptotic future state |Φ(∞)iI
we simply use
|Φ(∞)iI = S(∞, −∞)|Φ(−∞)iI (1.22)
namely, we rewind the time evolution back to the asymptotic past. Equation
(1.17) then becomes

iGΦ (x, t; x0 , t0 ) = hΦ(−∞)|I S(−∞, ∞)T [S(∞, −∞)ψ̂(x, t)ψ̂ † (x0 , t0 )]|Φ(−∞)iI
(1.23)

Fig. 1.1. Schwinger-Keldysh contour C.

6 1 Nonequilibrium Perturbation Theory

Now we cannot simply push S(−∞, ∞) past the time-ordering operator T

and merge it with the forward evolution S(∞, −∞), since the whole backward
evolution S(−∞, ∞) lies to the left of the T -product. Left means later as far
as time ordering is concerned. A clever way to kill these two birds with one
stone is to introduce a two-branch contour C and ordering along this contour
enforced by a contour-ordering operator Tc . This Schwinger-Keldysh contour
(Fig. 1.1) consists of two oriented branches C = C+ ∪ C− , the forward branch
C+ extending from −∞ to ∞ and the backward branch C− extending from
∞ to −∞. We now extend the time variables t, t0 to variables defined on this
contour τ, τ 0 , and define the fundamental object of nonequilibrium many-body
theory, the contour-ordered Green’s function,

iG(x, τ ; x0 , τ 0 ) = hTc [ψ(x, τ )ψ † (x0 , τ 0 )]i = Tr ρTc [ψ(x, τ )ψ † (x0 , τ 0 )] (1.24)

With variables and ordering along the contour C, the interaction picture ex-
pression corresponding to Eq. (1.23) is

iGΦ (x, τ ; x0 , τ 0 ) = hΦ(−∞)|I Tc [Sc (−∞, −∞)ψ̂(x, τ )ψ̂ † (x0 , τ 0 )]|Φ(−∞)iI

(1.25)
where the contour S-matrix is
 I 
0
Sc (−∞, −∞) ≡ Tc exp −i dτ1 Ĥ (τ1 ) (1.26)
C

where the integral is a line integral along the contour C. Note that there is
no phase factor in the denominator as in the equilibrium theories since the
contour S-matrix by itself (i.e. not in a Tc -ordered product) is actually unity.
According to our previous discussion, the full Green’s function is now given
by

iG(x, τ ; x0 , τ 0 ) = Tr ρ(−∞)Tc [Sc (−∞, −∞)ψ̂(x, τ )ψ̂ † (x0 , τ 0 )] (1.27)

where we have a density matrix defined by

X
ρ(−∞) = pΦ |Φ(−∞)iI hΦ(−∞)|I (1.28)
Φ

where now the states are not the Heisenberg states |Φi but the interaction
picture states in the remote past |Φ(−∞)iI . From Eqs. (1.7) and (1.15), it is
not hard to show that

ρ = S(0, −∞)ρ(−∞)S(−∞, 0) (1.29)

So far this is rather general. Now, in the Keldysh theory we do adiabatically

switch on the nonequilibrium perturbation V according to et from the remote
past t = −∞ to the present t = 0, but we do not switch it off. This enables us
to study stationary nonequilibrium states, a long time after the system has
 1.3 The Closed Time Path Contour 7

been driven out of equilibrium2 . Because of the adiabatic switch-on procedure,

ρ(−∞) can be identified as the density matrix of the system in the remote past
when the nonequilibrium perturbation is turned off. It is then evolved in the
usual way by the S-matrix S(0, −∞) to a nonequilibrium density matrix ρ.
One thus usually chooses ρ(−∞) to be an equilibrium distribution, ρ(−∞) =
e−βH /Z.

1.3 The Closed Time Path Contour

In the previous section, we constructed the perturbation expansion such that
the Green’s function would be expressed as an average over the density matrix
of the system in the remote past because this is what we know: because of the
adiabatic switch-on procedure, we have ρ(−∞) = e−βH /Z which is explicitly
known. Indeed, the idea of the Keldysh technique is that we want to avoid
reference to the general nonequilibrium density matrix ρ which is not usually
explicitly known. However, this is not mandatory. If we actually know the
density matrix ρ of the system at t = 0, then the early Eq. (1.13) is already
enough:
iG(x, τ ; x0 , τ 0 ) = Tr ρTc [Sc (0, 0)ψ̂(x, τ )ψ̂ † (x0 , τ 0 )] (1.30)
where ρ is now the full density matrix Eq. (1.7) at t = 0, and the contour over
which the Tc operator and the S-matrix Sc are defined is the so-called closed

Fig. 1.2. Closed time path contour C0 .

2
Actually, we may still study transient behavior in the Keldysh formalism, but
provided we use an explicitly time-dependent Hamiltonian and there are no initial
correlations, namely, the initial density matrix has to be noninteracting.
8 1 Nonequilibrium Perturbation Theory

time path contour C0 going from 0 to the latest of t or t0 , and then back to
0. Furthermore, there is nothing special about the time t = 0: we might have
defined the Heisenberg and interaction pictures with respect to some other
initial time t0 . Then the average would be over the density matrix at time t0 :
iG(x, τ ; x0 , τ 0 ) = Tr ρ(t0 )Tc [Sc (t0 , t0 )ψ̂(x, τ )ψ̂ † (x0 , τ 0 )] (1.31)
and the closed time path contour [5, 6] now goes from t0 to the latest of
t or t0 , and then back to t0 (Fig. 1.2). To recover the Schwinger-Keldysh
contour, we take the limit t0 → −∞ and insert a factor of S(t, ∞)S(∞, t)
or S(t0 , ∞)S(∞, t0 ) in the perturbation expansion Eq. (1.13) to extend the
contour to ∞, depending on whether t or t0 is the latest. The density matrix
is now ρ(−∞) and we invoke the adiabatic switch-on procedure to choose
ρ(−∞) = e−βH /Z.

1.4 Interactions: the Kadanoff-Baym Contour

Equation (1.31) is in a form suitable for a perturbation expansion provided
that ρ(t0 ) is a noninteracting (i.e., 1-particle) density matrix and that the
field operators ψ̂, ψ̂ † in the interaction picture evolve with a noninteracting
Hamiltonian H. Indeed, these are the conditions of applicability of Wick’s the-
orem [7]. They are satisfied in the Keldysh theory if H is noninteracting since
then ρ(−∞) = e−βH /Z is a 1-particle density matrix. What if H contains
interactions?
Let us first keep a general t0 and assume that ρ(t0 ) = e−βH /Z with H a
general interacting Hamiltonian. We can always take t0 → −∞ at the end to
recover the Keldysh theory. We first break further the equilibrium Hamiltonian
H = H0 + V into a noninteracting part H0 and the interactions V . We want
to express the interacting density matrix e−βH in terms of a noninteracting
one e−βH0 . We make use of Eqs. (1.11) and (1.12), but in imaginary time
and with a general initial time t0 . Now the full (equilibrium) Hamiltonian
is H. As mentioned earlier, we will therefore use upright letters U and S
for the evolution and S-matrix respectively. The interaction and Heisenberg
pictures are now defined with respect to the initial time t0 , meaning that
A(t) = eiH(t−t0 ) Ae−iH(t−t0 ) and Â(t) = eiH0 (t−t0 ) Ae−iH0 (t−t0 ) . Since H is
time-independent, the evolution operator is simply U (t, t0 ) = e−iH(t−t0 ) and
we have
S(t, t0 ) = eiH0 (t−t0 ) U (t, t0 ) (1.32)
 R 
t
where the S-matrix is S(t, t0 ) = T exp −i t0 dt1 V̂ (t1 ) . We then have

e−βH = e−βH0 S(t0 − iβ, t0 ) (1.33)

where here the S-matrix S(t0 − iβ, t0 ) evolves the density matrix along a
contour [t0 , t0 − iβ] on the imaginary axis. The Green’s function Eq. (1.31)
now assumes the form
 1.4 Interactions: the Kadanoff-Baym Contour 9

1 †
iG(1, 10 ) = Tr e−βH0 S(t0 − iβ, t0 )Tc [Sc (t0 , t0 )ψ̂H (1)ψ̂H (10 )] (1.34)
Z
where for simplicity we denote space and time arguments by a collective in-
dex 1 ≡ (x, τ ), 10 ≡ (x0 , τ 0 ). We have also explicitly indicated that the field
operators still evolve according to the full H, see Eq. (1.9). They have to be
brought to the current interaction picture defined with respect to H0 . This
can be achieved as previously through the S-matrix,
ψ̂H (x, t) = S(t0 , t)ψ̂(x, t)S(t, t0 ) (1.35)
where the caret without the subscript denotes the interaction picture with
respect to H0 . Equation (1.34) thus becomes
1
iG(1, 10 ) = Tr e−βH0 S(t0 − iβ, t0 )
Z
×Tc [Sc (t0 , t0 )S(t0 , t)ψ̂(1)S(t, t0 )S(t0 , t0 )ψ̂ † (10 )S(t0 , t0 )] (1.36)
where we now see that the S-matrix for interactions S evolves along a three-
branch contour C ∗ = C0 ∪ [t0 , t0 − iβ] which is the Kadanoff-Baym contour [8]
(Fig. 1.3). This contour goes from t0 to the latest of t and t0 , back to t0 , and

Fig. 1.3. Kadanoff-Baym contour C ∗ .

then down to t0 − iβ. We define an ordering operator Tc∗ along this contour so
that we can move the thermal S-matrix S(t0 −iβ, t0 ) into the contour-ordered
product. We then obtain
Tr e−βH0 Tc∗ [Sc∗ (t0 − iβ, t0 )Sc (t0 , t0 )ψ̂(1)ψ̂ † (10 )]
iG(1, 10 ) = (1.37)
Tr e−βH0 Tc∗ [Sc∗ (t0 − iβ, t0 )Sc (t0 , t0 )]
10 1 Nonequilibrium Perturbation Theory

where in analogy with Eq. (1.26) we define a new contour-ordered S-matrix

defined along the Kadanoff-Baym contour,
 Z 
Sc∗ (t0 − iβ, t0 ) ≡ Tc∗ exp −i dτ1 V̂ (τ1 ) (1.38)
C∗

We have written the partition function as

Z = Tr e−βH = Tr e−βH0 S(t0 − iβ, t0 ) = Tr e−βH0 Tc∗ [Sc∗ (t0 − iβ, t0 )Sc (t0 , t0 )]
(1.39)
since the S-matrices are already time-ordered on their respective contours,
and since times on the [t0 , t0 − iβ] strip are always later than times on C0 , we
simply have Sc (t0 , t0 ) = 1 and Sc∗ (t0 − iβ, t0 ) = S(t0 − iβ, t0 ) even inside the
Tc∗ -ordered product. Note that the time arguments in Eq. (1.37) are defined
on the three-branch Kadanoff-Baym contour C ∗ . Equation (1.37) is now fit
for perturbation theory since by assumption the averages are with respect to
a 1-particle density matrix e−βH0 and the field operators evolve according to
the noninteracting H0 , so Wick’s theorem can be applied.
The Kadanoff-Baym formalism is adequate for the study of initial corre-
lations, namely the effect for times t > t0 of having an interacting density
matrix at time t0 , without the assumption t  t0 . The price to pay for this
powerful formalism is that the Green’s function is defined on a three-branch
contour and has a complicated expression in terms of the simultaneous per-
turbation expansion of two S-matrices. However for many practical purposes
this is overkill, and for steady-state problems we do not care about the effect
of initial correlations. In many cases we can assume that correlations decay in
time so that if we take the limit t0 → −∞, at any finite time t  t0 there is
no signature left of the correlations in the initial density matrix ρ(t0 ). This is
the Bogoliubov principle of weakening correlations, a general principle in non-
equilibrium statistical mechanics. It is however advised to keep in mind that
in some cases initial correlations can persist at long times due for example to
the presence of metastable states.

1.4.1 Neglect of Initial Correlations and Schwinger-Keldysh Limit

For most practical purposes we can safely ignore the initial correlations when
taking the limit t0 → −∞. It has been shown [9, 10, 11] that neglecting
initial correlations amounts to neglecting the imaginary strip [t0 , t0 −iβ] in the
Kadanoff-Baym contour, so we are back with the Schwinger-Keldysh contour
(after having extended the closed time path contour as explained in section
1.3). In this limit, the S-matrices in the denominator of Eq. (1.37) are trivial
and the denominator is just Tr e−βH0 . We thus have

iG(1, 10 ) = Tr ρ0 Tc [Sc (−∞, −∞)Sc (−∞, −∞)ψ̂(1)ψ̂ † (10 )] (1.40)

where the noninteracting density matrix is just

 1.5 Contour Dyson Equation 11

e−βH0
ρ0 ≡ (1.41)
Tr e−βH0
and Sc (−∞, −∞) is just the Kadanoff-Baym contour-ordered S-matrix of
Eq. (1.38) but neglecting the third branch of the contour and taking the limit
t0 → −∞,  I 
Sc (−∞, −∞) ≡ Tc exp −i dτ1 V̂ (τ1 ) (1.42)
C
so that now all the ordering takes place along the Schwinger-Keldysh contour
C. If the system is initially at zero temperature, we simply have
iG(1, 10 ) = hΦ0 |Tc [Sc (−∞, −∞)Sc (−∞, −∞)ψ̂(1)ψ̂ † (10 )]|Φ0 i (1.43)
with |Φ0 i the noninteracting ground state of H0 . We see that we have now
almost recovered Eq. (1.27), albeit with an additional S-matrix Sc contain-
ing the effect of the interactions V . We summarize both zero and finite-
temperature results Eqs. (1.43) and (1.40) in the formula
iG(1, 10 ) = hTc [Sc (−∞, −∞)Sc (−∞, −∞)ψ̂(1)ψ̂ † (10 )]i (1.44)
with the suitable expectation value. This is the essential starting point for cal-
culations in Keldysh theory. The perturbation expansion can now be carried
out, with both nonequilibrium terms Ĥ 0 from Sc (Eq. (1.26)) and interaction
terms V̂ from Sc (Eq. (1.38)) appearing in the expansion and in the ensuing
Feynman diagrams. Let us remark here that the power of the Kadanoff-Baym
and Keldysh approaches to nonequilibrium many-body theory – or more gen-
erally, to nonequilibrium field theory – lies in its structure being formally
identical to that of usual equilibrium many-body theory, albeit with a time
evolution and the corresponding perturbative expansion defined on a more
general contour. Then most of the tools of quantum field theory can be ap-
plied: Feynman diagrams, integral equations for vertex functions such as the
Dyson equation, etc.

1.5 Contour Dyson Equation

The whole perturbation expansion for the contour-ordered 1-particle Green’s
function on the nonequilibrium contour C can be resummed in the form of an
integral equation, the Dyson equation:
Z
G(1, 10 ) = G0 (1, 10 ) + d2 G0 (1, 2)U (2)G(2, 10 )
Z Z
+ d2 d3 G0 (1, 2)Σ(2, 3)G(3, 10 ) (1.45)
Z
G(1, 10 ) = G0 (1, 10 ) + d2 G(1, 2)U (2)G0 (2, 10 )
Z Z
+ d2 d3 G(1, 2)Σ(2, 3)G0 (3, 10 ) (1.46)
12 1 Nonequilibrium Perturbation Theory

where G(1, 10 ) ≡ −ihTc [ψ(1)ψ † (10 )]i is the exact Green’s function Eq. (1.24)
and G0 (1, 10 ) ≡ −ihTc [ψ̂(1)ψ̂ † (10 )]i is the unperturbed Green’s function with
field operators in the interaction picture, U (2) is a 1-particle potential and
Σ(2, 3) is the 1-particle irreducible
R Pself-energy.
R R The integral sign means a sum
over all internal variables, d2 ≡ σ2 dx2 C dτ2 . For simplicity, we will use
the following notation,

G = G0 + G0 U G + G0 ΣG (1.47)
G = G0 + GU G0 + GΣG0 (1.48)

1.6 Initial Correlations with Arbitrary Initial Density

Matrix
For the sake of completeness, let us just mention a few words about the
most general problem (as far as I know) although we won’t actually try to
solve it in these lectures. Consider Eq. (1.31). In the Keldysh theory we let
t0 → −∞ and neglected initial correlations. In the Kadanoff-Baym theory we
considered initial correlations by keeping a finite t0 , but we assumed that the
initial density matrix ρ(t0 ) had the equilibrium form ρ(t0 ) = e−βH /Z with H
an interacting Hamiltonian. Now, what if ρ(t0 ) is a general, nonequilibrium,
interacting density matrix? This corresponds to preparing the system in a
correlated nonequilibrium state at a given time t0 and observing the time
evolution of the system for finite times t > t0 without assuming t  t0 . People
need that in the study of correlated plasmas for example. This problem has
been studied by Fujita [9], Hall [10], Kukharenko and Tikhodeev [11], and
Wagner [12] who gives a rather clear and comprehensive discussion. The idea
is to use a modified Kadanoff-Baym contour. First, the general nonequilibrium
density matrix can still be written in the form

e−λB
ρ(t0 ) = (1.49)
Tr e−λB
since it is positive definite and Hermitian, where λ is not the temperature
and B is not the Hamiltonian but some general quantities. But because of the
formal analogy to β and H, we can setup a perturbation expansion on a mod-
ified Kadanoff-Baym contour with the imaginary strip [t0 − iλ, t0 ]. This very
general approach encompasses the Feynman, Matsubara, Schwinger-Keldysh,
and Kadanoff-Baym approaches as special cases [12].

1.7 Relation to Real-Time Green’s Functions

In all that follows, we will confine ourselves exclusively to the Keldysh ap-
proach [4, 14, 16, 13, 15] (but the equilibrium Hamiltonian H can still contain
 1.7 Relation to Real-Time Green’s Functions 13

interactions). Let us recap. Our initial goal is to calculate the real-time Green’s
function of Eq. (1.1),

iG(x, t; x0 , t0 ) = hT [ψ(x, t)ψ † (x0 , t0 )]i = Tr ρT [ψ(x, t)ψ † (x0 , t0 )] (1.50)

with operators in the Heisenberg picture with respect to the full nonequilib-
rium Hamiltonian H and a nonequilibrium density matrix ρ. We have defined
a contour-ordered Green’s function Eq. (1.24) with time arguments on a con-
tour C,
iG(x, τ ; x0 , τ 0 ) = hTc [ψ(x, τ )ψ † (x0 , τ 0 )]i (1.51)
and obtained a perturbation expansion for that Green’s function in Eq. (1.44),

iG(x, τ ; x0 , τ 0 ) = hTc [Sc (−∞, −∞)Sc (−∞, −∞)ψ̂(x, τ )ψ̂ † (x0 , τ 0 )]i (1.52)

with operators in the interaction picture with respect to the noninteract-

ing part H0 of the equilibrium Hamiltonian H. Now, how is the contour-
ordered Green’s function Eq. (1.51) related to the real-time Green’s function
Eq. (1.50)?
The answer is that depending on which branch the contour arguments τ, τ 0
belong to, different real-time Green’s functions are obtained:

G (x, t; x0 , t0 ) ≡ −ihT [ψ(x, t)ψ † (x0 , t0 )]i if τ, τ 0 ∈ C+ ,

 T

 G< (x, t; x0 , t0 ) ≡ ihψ † (x0 , t0 )ψ(x, t)i

if τ ∈ C+ , τ 0 ∈ C− ,
G(x, τ ; x0 , τ 0 ) = 0 0 † 0 0


>
G (x, t; x , t ) ≡ −ihψ(x, t)ψ (x , t )i if τ ∈ C− , τ 0 ∈ C+ ,
G (x, t; x , t ) ≡ −ihT̃ [ψ(x, t)ψ (x , t )]i if τ, τ 0 ∈ C− ,
0 0 † 0 0
 T̃
(1.53)
where GT is the time-ordered Green’s function of Eq. (1.50) and T̃ is an anti-
time ordering operator which orders operators in the opposite way as T , so
that GT̃ is the anti-time-ordered Green’s function. G<,> are the lesser and
greater Green’s functions, respectively. One then defines a 2×2 matrix Green’s
function in real time,
 T <  
G G G11 G12
G= = (1.54)
G> GT̃ G21 G22

It is not hard to show from the definitions that the following identities hold,

GR = GT − G< = G> − GT̃ (1.55)

A T > < T̃
G = G −G =G −G (1.56)
GK = G> + G< = GT + GT̃ (1.57)

1.7.1 Larkin-Ovchinnikov Representation

In the so-called Larkin-Ovchinnikov representation [16], we perform a linear

transformation on G to obtain another matrix Ĝ,
14 1 Nonequilibrium Perturbation Theory

GR GK
 
Ĝ = Lτ 3 GL† = (1.58)
0 GA

where L = √12 (τ 0 − iτ 2 ) where τ i are the Pauli matrices (τ 0 is unity). GR,A

are the retarded and advanced Green’s functions,

GR (x, t; x0 , t0 ) = −iθ(t − t0 )h{ψ(x, t), ψ † (x0 , t0 )}i (1.59)

GA (x, t; x0 , t0 ) = iθ(t0 − t)h{ψ(x, t), ψ † (x0 , t0 )}i (1.60)

and GK is the Keldysh Green’s function,

GK (x, t; x0 , t0 ) = −ih[ψ(x, t), ψ † (x0 , t0 )]i (1.61)

1.7.2 Langreth Theorem of Analytic Continuation

Consider the ‘matrix products’ occurring in Eqs. (1.45) and (1.46). These are
convolution integrals on the two-branch contour. Consider first the following
product of two contour-ordered quantities (such as Green’s functions or self-
energies), Z
C(1, 10 ) = d2A(1, 2)B(2, 10 ) (1.62)

or in simple matrix notation,

C = AB (1.63)
How do we convert this product to an integral over the real axis involving
the real-time components of A and B? This is accomplished by the Langreth
theorem [18, 19] which consists in a series of rules:
< < <
(AB)> = AR B > + A> B A (1.64)
(AB)R,A = AR,A B R,A (1.65)
< < < <
(ABC)> = AR B R C > + AR B > C A + A> B A C A (1.66)
(ABC)R,A = AR,A B R,A C R,A (1.67)

where the ‘matrix products’ on the right-hand side consist in summations

over internal degrees of freedom (space and spin) and convolution integrals
on the real axis from −∞ to ∞. We are now completely rid of the contour;
the 2 × 2 matrix structure takes care of the nonequilibrium physics and we
can use ordinary real-time integrals. In particular, we can Fourier transform
to frequency space in the case of a nonequilibrium steady state in which the
Green’s functions are time-translationally invariant, Gij (t, t0 ) = Gij (t − t0 ).
We will not derive all these rules but only show how the first one can be
obtained. Consider C = AB on the Schwinger-Keldysh contour C. Then
 1.7 Relation to Real-Time Green’s Functions 15
Z
C < (t, t0 ) = dτ1 A(t, τ1 )B(τ1 , t0 )
C
Z ∞ Z −∞
= dt1 AT (t, t1 )B < (t1 , t0 ) + dt1 A< (t, t1 )B T̃ (t1 , t0 )
−∞ ∞
Z ∞ h i
= dt1 AT (t, t1 )B < (t1 , t0 ) − A< (t, t1 )B T̃ (t1 , t0 )
−∞

From Eqs. (1.55) and (1.56), we know that AT = AR +A< and B T̃ = B < −B A ,
hence
Z ∞
C < (t, t0 ) = dt1 AR (t, t1 ) + A< (t, t1 ) B < (t1 , t0 ) − A< (t, t1 ) B < (t1 , t0 ) − B A (t1 , t0 )



−∞
Z ∞
dt1 AR (t, t1 )B < (t1 , t0 ) + A< (t, t1 )B A (t1 , t0 )
 
=
−∞

The other rules are obtained in a similar fashion. Usually, one writes down
the diagrams as in conventional many-body theory and the corresponding
analytic expressions involving contour-ordered Green’s functions. Typically,
these involve integrals over the contour. Then one translates these expressions
in the real-time language using the Langreth rules. This is often referred to
as ‘analytic continuation’ from the contour to the real-time axis in analogy
with the Matsubara formalism. We feel this is somewhat pedantic since the
Schwinger-Keldysh contour is not really a contour in the complex plane; it is
only two counter-propagating copies of the real axis.
2
Quantum Kinetic Equations

We have now set up the necessary machinery to construct quantum kinetic

equations which govern the evolution of correlation functions in a quantum
many-body system driven out of equilibrium. We first say ‘correlation func-
tions’ which are two-time objects instead of ‘distribution functions’ which
are one-time objects because equations typically do not close for distribution
functions. However equations for correlation functions are very complicated
integral equations which can usually only be solved numerically. There how-
ever exists a variety of useful approximations which make the equations for
distribution functions close.
We will thus first discuss the Keldysh and Kadanoff-Baym equations which
are exact quantum kinetic equations for the nonequilibrium correlation func-
tions of the system. Then we will introduce the Wigner transform, the gradient
expansion and the gradient approximation which yield the quantum Boltz-
mann equation (QBE). As this equation is still rather complicated, further
approximations can be made such as the quasiparticle and quasiclassical ap-
proximations. The QBE is still an equation for a four-parameter (two-point)
correlation function G< (p, ω, R, T ). Expressed in terms of the Wigner dis-
tribution function f W (p, R, T ) ≡ −iG< (p, t = 0, R, T ) which is a one-time
function, the QBE does not close. One can however generate a closed equation
for the Wigner distribution function by introducing the so-called generalized
Kadanoff-Baym ansatz [17]. We will also recover the classical Boltzmann equa-
tion as a limiting case of the QBE.
We can actually understand from simple arguments why a kinetic equation
for a quantum distribution function should have more terms than that for a
classical distribution function. Recall how the classical Boltzmann equation
is derived. Because of the incompressibility of phase space, the total rate of
change of the classical distribution function
 f (p, R, T ) is equal to the rate of
∂f
change of f due to collisions, say ∂T ≡ I[f ]:
coll

df
= I[f ]
dT
18 2 Quantum Kinetic Equations

The Boltzmann equation is then obtained simply by using the chain rule for
derivatives,  
∂
+ v · ∇R + F · ∇p f = I[f ] (2.1)
∂T
since dR/dT = v is the velocity and dp/dT = F is the force from Newton’s
second law. In the quantum mechanical case, the distribution function also
has an energy argument ω: f = f (p, ω, R, T ). We thus generate an additional
term in the corresponding Boltzmann equation:
 
∂ dω ∂
+ v · ∇ R + F · ∇p + f = I[f ]
∂T dT ∂ω

But the rate of change of energy is just the power, dω/dT = P = F · v.

Considering electric and magnetic fields, the force is just the Lorentz force
F = e(E + v × B), so dω/dT = eE · v since the magnetic force does no work.
Hence we obtain [13]
   
∂ ∂
+ v · ∇R + eE + e(E + v × B) · ∇p f = I[f ] (2.2)
∂T ∂ω

We will see later in this chapter that this is essentially the structure of the
QBE.

2.1 Keldysh Equation

We first apply the fourth Langreth rule Eq. (1.67) to the contour Dyson
equations (1.47) and (1.48). We obtain

GR,A = GR,A
0 + GR,A
0 U GR,A + GR,A
0 Σ R,A GR,A (2.3)
G R,A
= GR,A
0 +G R,A
U GR,A
0 +G R,A
Σ R,A
GR,A
0 (2.4)

because a one-body potential U is proportional to a delta function in time (in-

stantaneous) and is thus neither retarded nor advanced. Hence we see that the
retarded and advanced nonequilibrium Green’s functions obey Dyson equa-
tions which are formally identical to the equilibrium Dyson equations. If we
now apply the third Langreth rule Eq. (1.66), we obtain
< < < < < < <
G> = G> R > > A R R > R > A > A A
0 + G0 U G + G0 U G + G0 Σ G + G0 Σ G + G0 Σ G
< < < < < < <
G> = G> R > > A R R > R > A > A A
0 + G U G0 + G U G0 + G Σ G 0 + G Σ G 0 + G Σ G 0

because the instantaneous one-body potential is diagonal in Keldysh space

<
(U > = 0). These two equations are the Keldysh equations. Integration over
the real axis is understood, for example
 2.1 Keldysh Equation 19
Z ∞ Z
0 0 0 0
(GR <
0 U G )(x, t; x , t ) = dt1 dx1 GR <
0 (x, t; x1 , t1 )U (x1 , t1 )G (x1 , t1 ; x , t )
−∞

and
Z ∞ Z Z ∞ Z
0 0
(GR R <
0 Σ G )(x, t; x , t ) = dt1 dx1 dt2 dx2 GR
0 (x, t; x1 , t1 )
−∞ −∞
×Σ (x1 , t1 ; x2 , t2 )G< (x2 , t2 ; x0 , t0 )
R

The Keldysh equations can be iterated. The infinite order iterate of either
equation gives the same following explicit equation for G< ,
< < <
G> = [1 + GR (U + Σ R )]G> A A R > A
0 [1 + (U + Σ )G ] + G Σ G (2.5)

which is the usual general form of the Keldysh equation.

We now digress a moment to study the free Green’s functions G<,>,R,A
0
more closely. They are propagators for a noninteracting Hamiltonian H0 . To be
fully general, we consider a multiband one-particle Hamiltonian H0µν (x, −i∇)
where µ, ν are band (and/or spin) indices. We however require that each H0µν
be a properly symmetrized function of the conjugate variables,
X µν
H0µν (x, −i∇) = {fi (x), giµν (−i∇)} + uµν (x) + v µν (−i∇) (2.6)
i

Then Hermiticity of the second quantized Hamiltonian,

XZ
H0 = dx ψµ† (x)H0µν (x, −i∇)ψν (x), (2.7)
µν

requires that1
H0µν (x, −i∇) = H0νµ (x, i∇)∗ (2.8)
We use a matrix notation where Gµν
is represented by a matrix G0 in band
0
space. It is easy to show by application of the equation of motion tech-
nique that the contour-ordered propagator G0 (x, τ ; x0 , τ 0 ) satisfies the fol-
lowing equations of motion2 where matrix multiplication in band space is
understood,

1
Two remarks. First, in the case that the only imaginary quantities in H0µν come
from −i∇ (for example for a spin system without Zeeman term and in the absence
of spin-orbit coupling), Eq. (2.8) requires the matrix H0µν to be symmetric. Sec-
ond, if H0µν depends only on x and not on −i∇ (for example for a tight-binding
model where x becomes a discrete variable and ∇ is replaced by finite differences),
Eq. (2.8) requires the matrix H0µν to be Hermitian.
2
The delta function δ(τ, τ 0 ) is defined on the contour, such that δ R,A (t, t0 ) = δ(t−t0 )
<
and δ > (t, t0 ) = 0.
20 2 Quantum Kinetic Equations
 
i∂τ − H0 (x, −i∇) G0 (x, τ ; x0 , τ 0 ) = δ(x − x0 )δ(τ, τ 0 ) (2.9)
 
G0 (x, τ ; x0 , τ 0 ) −i∂τ 0 − H0 (x0 , i∇0 ) = δ(x − x0 )δ(τ, τ 0 ) (2.10)

where in Eq. (2.10) the derivatives ∂τ 0 and ∇0 act to the left. We do not write
explicitly unit matrices in band space. If we take the real-time components of
these contour equations, we obtain
 
i∂t − H0 (x, −i∇) GR,A 0 (x, t; x0 , t0 ) = δ(x − x0 )δ(t − t0 ) (2.11)
 
i∂t − H0 (x, −i∇) G<,> 0 (x, t; x0 , t0 ) = 0 (2.12)
 
GR,A
0 (x, t; x0 , t0 ) −i∂t0 − H0 (x0 , i∇0 ) = δ(x − x0 )δ(t − t0 ) (2.13)
 
G<,>
0 (x, t; x0 , t0 ) −i∂t0 − H0 (x0 , i∇0 ) = 0 (2.14)

We can write in simplified notation

R,A
Ĝ−1
0 G0 =1 (2.15)
<
Ĝ−1
0 G0
>
=0 (2.16)
R,A −1
G0 Ĝ0 =1 (2.17)
<
−1
G>0 Ĝ0 =0 (2.18)

where the action of the operator Ĝ−1 0 on the right and on the left is defined
in Eqs. (2.11)-(2.14). These relations will be helpful in deriving the Kadanoff-
Baym and quantum Boltzmann equations in the next section.
The Keldysh equations are used directly in mesoscopic transport for ex-
ample, where one studies finite-sized nonequilibrium systems with important
spatial and/or temporal inhomogeneities. Then the continuous spatial argu-
ments x, x0 in the Green’s functions are typically replaced by discrete site
indices i, j (i.e. in a tight-binding representation) or by principal quantum
numbers n, n0 for discrete energy levels (i.e. in quantum dot physics). The
Green’s functions then become finite-sized matrices and the Keldysh equa-
tions become matrix equations. For a finite-sized system, it is possible to
show that the Keldysh equation (2.5) becomes
< < <
G> = GR Ĝ−1 > A A R > A
0 G0 [1 + (U + Σ )G ] + G Σ G (2.19)

but the first term vanishes by Eq. (2.16), so that we have

< <
G> = G R Σ > GA (2.20)

which is the form of the Keldysh equation that is used in mesoscopic transport.
 2.2 Kadanoff-Baym Equation 21

At the other side of the spectrum, if one wishes to study transport in

macroscopic samples, i.e. transport in metals or bulk semiconductors, non-
equilibrium (but not mesoscopic) superconductivity, etc., then it is usually
more practical to use the Kadanoff-Baym equation and its ensuing Boltzmann-
type approximations to be studied in the next sections [16]. Then one has to
solve an (integro-)differential equation. It is easier to make approximations
(such as near-homogeneous or slowly-varying perturbations, not too far from
equilibrium) on the Kadanoff-Baym equation than on the Keldysh equation.
Furthermore, a connection to conventional transport formalisms such as linear
response theory (Kubo formula) can be made: transport coefficients calculated
from the linearized quantum Boltzmann equation (which can be derived from
the Kadanoff-Baym equation) in the steady-state limit are identical to those
obtained in linear response theory.

2.2 Kadanoff-Baym Equation

With the help of these relations, we are now able to derive the Kadanoff-Baym
equation from the Keldysh equations. If we act with Ĝ−1 0 on the left of the
first Keldysh equation and on the right of the second Keldysh equation and
use relations (2.15)-(2.18), we obtain

(Ĝ−1 < R < < A

0 − U )G = Σ G + Σ G (2.21)
G <
(Ĝ0−1 R <
− U) = G Σ + G Σ < A
(2.22)

We now subtract these equations from one another:

[Ĝ−1 < R < < A R < < A

0 − U, G ] = Σ G + Σ G − G Σ − G Σ (2.23)

We now introduce the nonequilibrium spectral function A ≡ i(GR − GA )

and the scattering rate, or linewidth, or imaginary part of the self-energy
Γ ≡ i(Σ R − Σ A ). We also define symmetric combinations or real parts as
Re G ≡ 21 (GR + GA ) and Re Σ ≡ 21 (Σ R + Σ A ). In terms of these quantities,
Eq. (2.23) can be written as

[Ĝ−1 < < < i < <

0 − U, G ] − [Re Σ, G ] − [Σ , Re G] = − 2 {Γ, G } − {A, Σ }

However, as seen before, the Keldysh components GR,A,<,> and Σ R,A,<,>

are not independent. With the relations −iA = GR − GA = G> − G< and
−iΓ = Σ R − Σ A = Σ > − Σ < we can rewrite the equation as

[Ĝ−1 < < 1 > < > <

0 − U − Re Σ, G ] − [Σ , Re G] = 2 {Σ , G } − {G , Σ } (2.24)

which is the Kadanoff-Baym equation [8, 13]. As we will see, the term on
the left-hand side involving [Ĝ−1 <
0 − U, G ] is a driving term, while the terms
involving Re Σ and Re G describe renormalization effects. From a transport
22 2 Quantum Kinetic Equations

point of view, they lead to renormalized transport coefficients and are ne-
glected in the classical Boltzmann limit. The right-hand side is a collision
term as will be seen. A similar Kadanoff-Baym equation can be derived for
G> . By subtracting the two equations, one obtains an equation for the spectral
function,
[Ĝ−1
0 − U − Re Σ, A] − [Γ, Re G] = 0 (2.25)
which is used as a consistency check when one looks for approximate solutions
to the Kadanoff-Baym equation.
The Kadanoff-Baym equation has to be supplemented by an equation for
the retarded and advanced Green’s functions. It is obtained in the same way,
applying relations (2.15)-(2.18) to the nonequilibrium Dyson equations (2.3)-
(2.4):

(Ĝ−1 R R R
0 − U )G = 1 + Σ G (2.26)
G R
(Ĝ−1
0
R
− U) = 1 + G Σ R
(2.27)

Add these two equations and dividing by 2 gives

−1
1
2 {Ĝ0 − U, GR } = 1 + 21 {Σ R , GR } (2.28)

In practice, because self-energies are typically functionals of the Green’s func-

tion Σ = Σ[G], the Kadanoff-Baym equation and the equations of motion
(2.26)-(2.27) for GR have to be solved self-consistently. Needless to say, this is
a rather difficult task and one often has to resort to numerical techniques. Ap-
proximations have to be made for the self-energy functions. However, these
approximations cannot be made blindly by choosing an arbitrary subset of
diagrams since it is then possible to violate conservation laws as shown by
Baym and Kadanoff. Self-energies should be derived from the Luttinger-Ward
[20] functional Φ,
δΦ[G]
Σ[G] = (2.29)
δG
The Luttinger-Ward functional Φ[G] is defined diagrammatically as the sum
of all skeleton connected vacuum diagrams with free propagators G0 replaced
by exact propagators G. In practice, one keeps only a few diagrams in the
diagrammatic expansion of Φ and then obtains the corresponding diagram-
matic representation of the self-energy Σ[G] by removing a fermion line G
(which corresponds to functional differentiation in Eq. (2.29)). These so-called
Φ-derivable approximations preserve conservation laws and are termed con-
serving approximations [21, 22, 8]. For interacting systems involving coupled
degrees of freedom such as electrons and phonons, the electron self-energy Σ
and phonon polarization Π should be derived from the same electron-phonon
vertex Γ . This ensures that the same level of approximation is maintained
for both the electron and phonon subsystems [23]. In general, gauge invari-
ance implies the Ward identities which relate the self-energies and the vertex
so that a given Φ-derivable approximation for the self-energy Σ[G] defines a
 2.2 Kadanoff-Baym Equation 23

fully consistent level of approximation for Green’s functions, self-energies, and

the vertex.

2.2.1 Wigner Representation and Gradient Expansion

We now move on to the gradient expansion of the Kadanoff-Baym equation

[8, 16, 24]. The idea is to separate slow macroscopic variations from fast
microscopic variations, and then perform a gradient expansion on the slow
variables. In order to do that, we introduce mixed coordinates, the so-called
Wigner representation:

r ≡ x1 − x2 , R ≡ 21 (x1 + x2 ) (2.30)
t ≡ t1 − t2 , T ≡ 21 (t1 + t2 ) (2.31)

so that a function C(x1 , t1 ; x2 , t2 ) becomes a function C(r, t, R, T ). The fast

‘relative’ variables (r, t) will be Fourier transformed to (k, Ω) while the slow
‘center-of-mass’ variables3 (R, T ) will serve for the gradient expansion. The
Fourier transform is defined as
Z Z
C(k, Ω, R, T ) ≡ dt dr ei(Ωt−k·r) C(r, t, R, T ) (2.32)
Z Z
= dt dr ei(Ωt−k·r) C R + 12 r, T + 12 t; R − 12 r, T − 21 t

We first work out the Wigner transform of the driving commutator term
[Ĝ−1 <
0 − U, G ]. We have
 
−1 < ∂
[Ĝ0 − U, G ]x1 ,x2 = i − H0 (x1 , −i∇1 ) − U (x1 ) G< (x1 , x2 )
∂t1
 
< ∂
−G (x1 , x2 ) −i − H0 (x2 , i∇2 ) − U (x2 )
∂t2

and for convenience we use a four-vector notation xµ1,2 = (t1,2 , x1,2 ).

We want to study the dynamics of a system driven out of equilibrium
by constant and uniform external electric E and magnetic B fields4 . We first
concentrate on the electric field and the magnetic field will be added later. The

3
For example, consider a free propagator g(t1 , t2 ) = −ie−i$(t1 −t2 ) where $ is
some characteristic frequency. Considered as a function of t = 12 (t1 − t2 ) and
T = 12 (t1 + t2 ), this function has a fast variation with frequency 2$ in the t
variable and is constant (infinitely slow variation) in the T variable. If we add a
slow disturbance $ → $ + ∆(t), there will start to be a weak dependence in T ,
but slower than the dependence in t.
4
All the following derivations can be done for a general one-particle potential U (x)
but would lead to extremely cumbersome expressions. This doesn’t have much
interest anyway since the physically relevant perturbations are E and B fields.
24 2 Quantum Kinetic Equations

one-particle potential U (x) contains the electric field in the scalar potential
gauge:
U (x) = −eE · x (2.33)
We have
 
∂ ∂
[Ĝ−1 <
0 − U, G ]x1 ,x2 = i +i G< (x1 , x2 ) − H0 (x1 , −i∇1 )G< (x1 , x2 )
∂t1 ∂t2
+G< (x1 , x2 )H0 (x2 , i∇2 ) + eE · (x1 − x2 )G< (x1 , x2 )

The change of variables Eqs. (2.30) and (2.31) imply the following relations
for the derivatives,
 
∂ ∂ ∂ ∂ 1 ∂ ∂
= + , = − ,
∂T ∂t1 ∂t2 ∂t 2 ∂t1 ∂t2
and similar relations for the spatial derivatives. In Wigner coordinates we thus
have
∂ <
[Ĝ−1 <
G (r, t, R, T ) − H0 R + 21 r, −i( 12 ∇R + ∇r ) G<


0 − U, G ]r,t,R,T = i
∂T
+G< H0 R − 21 r, i( 12 ∇R − ∇r ) + eE · rG<

We now perform the Fourier transformation of the center-of-mass variables

Eq. (2.32), which corresponds to the following substitutions,
∂
i →Ω (2.34)
∂t
−i∇r → k (2.35)
∂
t → −i (2.36)
∂Ω
r → i∇k (2.37)

We obtain
∂ <
[Ĝ−1 <
G (k, Ω, R, T ) − H0 R + 2i ∇k , k − 2i ∇R G<


0 − U, G ]k,Ω,R,T = i
∂T
+G< H0 R − 2i ∇k , k + 2i ∇R + ieE · ∇k G<

We now perform a change of variables introduced by Mahan and Hänsch,

ω = Ω + eE · R (2.38)

which eliminates an unphysical ∝ E · R term in the equation for GR to be

derived later. This change of variables implies the following substitutions,

Ω → ω − eE · R (2.39)
∂
∇R → ∇R + eE (2.40)
∂ω
 2.2 Kadanoff-Baym Equation 25

The derivative ∂/∂ω is important and differentiates the QBE from the classical
Boltzmann equation. We thus obtain the following final exact form for the
driving commutator in the Wigner representation,

∂G<
  
−1 < ∂
[Ĝ0 − U, G ]k,ω,R,T = i i i
− H0 R + 2 ∇k , k − 2 ∇R + eE G<
∂T ∂ω
  
< ∂
i i
+G H0 R − 2 ∇k , k + 2 ∇R + eE + ieE · ∇k G<
∂ω
(2.41)

where G< = G< (k, ω, R, T ). We now see the importance of a proper ordering
of the Hamiltonian since [R, ∇R ] 6= 0 and [k, ∇k ] 6= 0. The driving anti-
commutator {Ĝ−1 R
0 − U, G } appearing in Eq. (2.28) for the retarded Green’s
function can be derived in a similar way. We simply quote the result:
  
−1 R R ∂
1
2 {Ĝ 0 − U, G }k,ω,R,T = ωG − 1
H
2 0 R + i
∇
2 k , k − i
2 ∇ R + eE GR
∂ω
  
∂
− 12 GR H0 R − 2i ∇k , k + 2i ∇R + eE (2.42)
∂ω

where as advertised the unphysical term E · R is removed by the Mahan-

Hänsch transformation Eq. (2.38).
Having derived the driving terms in the Wigner representation, we must
now see how to transform products of the form AB,
Z Z
0
(AB)x1 ,t1 ;x2 ,t2 = dt dx0 A(x1 , t1 ; x0 , t0 )B(x0 , t0 ; x2 , t2 ) (2.43)

From the definition of the Wigner coordinates Eqs. (2.30-2.31), we have

A(x1 , t1 ; x0 , t0 ) = A(x1 − x0 , t1 − t0 , 21 (x1 + x0 ), 12 (t1 + t0 ))

x0 − x2 t0 − t2
 
= A x1 − x0 , t1 − t0 , R + ,T + (2.44)
2 2

and similarly

B(x0 , t0 ; x2 , t2 ) = B(x0 − x2 , t0 − t2 , 21 (x0 + x2 ), 12 (t0 + t2 ))

x0 − x1 t0 − t1
 
= B x0 − x2 , t0 − t2 , R + ,T + (2.45)
2 2

Since the derivative is the generator of translations, we have the following

compact form for the usual Taylor expansion,

f (R + a, T + s) = ea·∇R es∂T f (R, T ) (2.46)

Using Eqs. (2.44,2.45,2.46), we can write Eq. (2.43) as

26 2 Quantum Kinetic Equations
Z Z
0 B 0 B
0
(AB)x1 ,t1 ;x2 ,t2 = dt dx0 A(x1 − x0 , t1 − t0 , R, T )e−(x1 −x )·∇R /2 e−(t1 −t )∂T /2
0 A 0 A
×e(x −x2 )·∇R /2 e(t −t2 )∂T /2 B(x0 − x2 , t0 − t2 , R, T ) (2.47)

where the superscripts A and B on the derivative operators ∇R , ∂T indicate

that they act on the function A and B, respectively. The integral now has the
form of a convolution product,
Z Z
dt0 dx0 Ã(x1 − x0 , t1 − t0 , R, T )B̃(x0 − x2 , t0 − t2 , R, T ) = Ã ∗ B̃,

where
B B
Ã(r, t, R, T ) = A(r, t, R, T )e−(r·∇R +t∂T )/2 (2.48)
(r·∇A A
R +t∂T )/2
B̃(r, t, R, T ) = e B(r, t, R, T ) (2.49)

We now take the Fourier transform, as defined in Eq. (2.32), of the convolution
product, so that we have

(AB)k,Ω,R,T = F{Ã ∗ B̃} = Ã(k, Ω, R, T )B̃(k, Ω, R, T ) (2.50)

since the Fourier transform of a convolution product is the ordinary product

of the Fourier transforms of the convoluted functions. We thus have
Z Z
B B
Ã(k, Ω, R, T ) = dt dr ei(Ωt−k·r) A(r, t, R, T )e−(r·∇R +t∂T )/2 (2.51)
Z Z
A A
B̃(k, Ω, R, T ) = dt dr ei(Ωt−k·r) e(r·∇R +t∂T )/2 B(r, t, R, T ) (2.52)

Since a phase factor in a Fourier transform generates a translation,

F{ea·r est f (r, t)} = f (k + ia, Ω − is) = eia·∇k e−is∂Ω f (k, Ω),

we obtain
B A B A
Ã(k, Ω, R, T ) = A(k, Ω, R, T )e−i∇R ·∇k /2 ei∂T ∂Ω /2 (2.53)
i∇A B A B
R ·∇k /2 −i∂T ∂Ω /2
B̃(k, Ω, R, T ) = B(k, Ω, R, T )e e (2.54)

Inserting these results in Eq. (2.50), we have

(AB)k,Ω,R,T = A(k, Ω, R, T )Ĝ AB (k, Ω, R, T )B(k, Ω, R, T ) (2.55)

where we have now dropped the curly letter notation A and B for simplicity,
and we define the gradient operator Ĝ AB as
A B A B A B A B
Ĝ AB (k, Ω, R, T ) ≡ e−i(∂T ∂Ω −∂Ω ∂T −∇R ·∇k +∇k ·∇R )/2

As a mathematical aside, we note in passing that Eq. (2.55) corresponds to

the Moyal or star product of two functions which is used in noncommutative
 2.3 Quantum Boltzmann Equation 27

geometry and deformation quantization, and more generally in mathematics

in the construction of deformed algebras.
So far, everything has been exact. In the so-called gradient approximation,
we expand the exponential to first order in the gradients. We have

i ∂A ∂B ∂A ∂B
 
AB A B A B
Ĝ (k, Ω, R, T ) ' 1 − − − ∇R · ∇k + ∇k · ∇R
2 ∂T ∂Ω ∂Ω ∂T

so the basic expression of the gradient approximation is

(AB)k,Ω,R,T ' A(k, Ω, R, T )B(k, Ω, R, T )

 
i ∂A ∂B ∂A ∂B
− − − ∇R A · ∇ k B + ∇ k A · ∇R B
2 ∂T ∂Ω ∂Ω ∂T

from Eq. (2.55). We can now write down the Kadanoff-Baym commutators
and anticommutators in the Wigner representation under the gradient ap-
proximation as follows,
    
i ∂A ∂B ∂A ∂B
[A, B]k,ω,R,T = [A, B] + , − , − {∇k A, ∇R B} + {∇R A, ∇k B}
2 ∂ω ∂T ∂T ∂ω
   
i ∂A ∂B
+ eE · , ∇k B − ∇k A, (2.56)
2 ∂ω ∂ω
    
i ∂A ∂B ∂A ∂B
{A, B}k,ω,R,T = {A, B} + , − , − [∇k A, ∇R B] + [∇R A, ∇k B]
2 ∂ω ∂T ∂T ∂ω
   
i ∂A ∂B
+ eE · , ∇k B − ∇k A, (2.57)
2 ∂ω ∂ω

where on the right-hand side, these are ordinary matrix commutators and
anticommutators, and we have performed the Mahan-Hänsch transformation
Eq. (2.38).

2.3 Quantum Boltzmann Equation

We first discuss the QBE [24] in zero magnetic field, then add the magnetic
field.

2.3.1 QBE with Electric Field

The QBE is now obtained merely as the Kadanoff-Baym equation (2.24) in

the Wigner representation:

[Ĝ−1 < < <

0 − U, G ]k,ω,R,T = [Re Σ, G ]k,ω,R,T + [Σ , Re G]k,ω,R,T
+ 2 {Σ , G }k,ω,R,T − {G> , Σ < }k,ω,R,T (2.58)
1 > <


28 2 Quantum Kinetic Equations

where the left-hand side is given by Eq. (2.41) and the terms on the right-hand
side are given by Eqs. (2.56) and (2.57). So far we have only rephrased the
Kadanoff-Baym equation in Wigner coordinates and made the gradient ap-
proximation. Equation (2.58) is still valid to arbitrary order in the electric field
and for both time and space-dependent (but slowly varying) perturbations.
Note that this very general equation has in principle to be solved together
with the equation for the retarded Green’s function Eq. (2.28),
−1
1
2 {Ĝ0 − U, GR }k,ω,R,T = 1 + 21 {Σ R , GR }k,ω,R,T (2.59)
where the left-hand side is given in Eq. (2.42) and the anticommutator on the
right-hand side is given by Eq. (2.57). Indeed, the nonequilibrium GR enters
the renormalization terms Re G and Re Σ (since Σ R = Σ R [GR ]) in Eq. (2.58).
It is obvious that this primary form of the QBE is extremely complicated
and that not much progress can be made unless further approximations are
introduced, otherwise one has to resort to numerical techniques. In addition,
we have kept a fully general multiband Hamiltonian H0 so far, but one can
also restrict the analysis to a given class of Hamiltonians.

2.3.2 QBE with Electric and Magnetic Field

We now add the effect of a constant uniform magnetic field B. Whereas the
electric field E was introduced in Eq. (2.33) through the scalar potential, we
now introduce the magnetic field through a vector potential,
A(x) = − 21 x × B
We also perform the Peierls substitution in the Hamiltonian,
H0 (−i∇) → H0 (−i∇ − eA) = H0 −i∇ + 12 ex × B


(2.60)
We now proceed along similar steps as before. In Wigner coordinates, we have
H0 (x1 , −i∇1 − eA1 ) → H0 R + 12 r, −i( 12 ∇R + ∇r ) + 12 e(R + 12 r) × B

H0 (x2 , i∇2 − eA2 ) → H0 R − 12 r, i( 12 ∇R − ∇r ) + 21 e(R − 12 r) × B

After Fourier transformation, we obtain terms like

H0 R ± 2i ∇k , k + 21 eR × B ∓ 2i (∇R + 12 eB × ∇k )

As before, we perform the Mahan-Hänsch transformation Eq. (2.38) which

changes the derivative with respect to R according to Eq. (2.40), and obtain
  
i 1 i ∂ 1
H0 R ± 2 ∇k , k + 2 eR × B ∓ 2 ∇R + eE + eB × ∇k
∂ω 2
The Mahan-Hänsch transformation got rid of the ∝ E · R term. We now
perform a similar extra transformation to get rid of the unphysical R × B
term by introducing the kinematical momentum,
 2.3 Quantum Boltzmann Equation 29

p = k + 21 eR × B = k − eA(R) (2.61)

which implies ∇k → ∇p but modifies the R derivative once more,

∇R → ∇R + 21 eB × ∇p

so that the final expression reads

  
i i ∂
H0 R ± 2 ∇p , p ∓ 2 ∇R + eE + eB × ∇p
∂ω

Hence the generalization of Eq. (2.41) to include the magnetic field B is

∂G<
  
−1 < ∂
[Ĝ0 − U, G ]p,ω,R,T = i i i
− H0 R + 2 ∇p , p − 2 ∇R + eE + eB × ∇p G<
∂T ∂ω
  
∂
+G< H0 R − 2i ∇p , p + 2i ∇R + eE + eB × ∇p
∂ω
<
+ieE · ∇p G (2.62)

and similarly the generalization of Eq. (2.42) is

  
−1 R R ∂
1
2 {Ĝ 0 − U, G }p,ω,R,T = ωG − 1
2 H 0 R + i
2 ∇ p , p − i
2 ∇ R + eE + eB × ∇ p GR
∂ω
  
∂
− 12 GR H0 R − 2i ∇p , p + 2i ∇R + eE + eB × ∇p (2.63)
∂ω

All Green’s functions and self-energies are now functions of (p, ω, R, T ). The
last thing to figure out is how Eqs. (2.56) and (2.57) are modified in the
presence of the magnetic field. It is straightforward to show that the only
changes are the substitution k → p and a new term linear in the magnetic
field,

([A, B]± )p,ω,R,T = ([A, B]± )k,ω,R,T + 2i eB · [∇p A × ∇p B]± (2.64)

k→p ,

where we use the notation

[∇p A × ∇p B]± ≡ ∇p A × ∇p B ± ∇p B × ∇p A (2.65)

,

Then the QBE in the presence of both electric E and magnetic B fields is
formally identical to Eq. (2.58),

[Ĝ−1 < < <

0 − U, G ]p,ω,R,T = [Re Σ, G ]p,ω,R,T + [Σ , Re G]p,ω,R,T
+ 21 {Σ > , G< }p,ω,R,T − {G> , Σ < }p,ω,R,T (2.66)

but with the definitions in Eqs. (2.62) and (2.64). Similarly, the equation for
the retarded Green’s function Eq. (2.59) becomes
30 2 Quantum Kinetic Equations
−1
1
2 {Ĝ0 − U, GR }p,ω,R,T = 1 + 21 {Σ R , GR }p,ω,R,T (2.67)

together with Eqs. (2.63) and (2.64).

Equation (2.66) is the most general form of the quantum Boltzmann equa-
tion and makes no assumptions other than that of slowly varying disturbances.
It is obvious that it is a rather complicated equation. Approximations should
be introduced to bring it to a more tractable form. The first obvious approx-
imation is to restrict the analysis to linear transport, that is, keep only terms
to linear order in the electric field, while keeping terms to all orders in the
magnetic field. Afterwards, approximations can branch off in different direc-
tions and depend on the problem at hand. In the next chapter, we discuss
several useful approximate forms of the QBE.

2.3.3 One-Band Spinless Electrons

Most analytical work using the QBE is done for one-band spinless electrons.
The single-particle Hamiltonian H0 is of course

(−i∇)2
H0 (−i∇) =
2m
The driving terms Eqs. (2.62) and (2.63) become
   
∂ ∂
[Ĝ−1
0 − U, G <
]p,ω,R,T = i + vp · ∇ R + eE + e(E + vp × B) · ∇ p G
<
∂T ∂ω
"  2 #
−1 R 1 ∂
1
2 {Ĝ0 − U, G }p,ω,R,T = ω − p + 8m ∇R + eE ∂ω + eB × ∇p GR

where we have the usual definitions of the single-particle energies and velocities

p2 p
p = , vp = ∇p p = .
2m m

We see that the driving term [Ĝ−1 <

0 − U, G ] has indeed the form of Eq. (2.2)
derived from simple considerations. We now consider the Kadanoff-Baym com-
mutators Eq. (2.64). Consider first the anticommutator {A, B}p,ω,R,T . All the
commutators in Eq. (2.57) vanish since the quantities are all scalars. Further-
more, the term proportional to B in Eq. (2.64) vanishes because of the anti-
symmetry of the cross product. Hence we have simply {A, B}p,ω,R,T = 2AB,
and Eq. (2.67) becomes
"  2 #
1 ∂
ω − p + ∇R + eE + eB × ∇p − Σ GR = 1
R
(2.68)
8m ∂ω

The Kadanoff-Baym commutator in Eq. (2.64) becomes

 2.3 Quantum Boltzmann Equation 31
 
∂A ∂B ∂A ∂B
[A, B]p,ω,R,T = i − − ∇p A · ∇ R B + ∇ R A · ∇ p B
∂ω ∂T ∂T ∂ω
 
∂A ∂B
+ieE · ∇p B − ∇p A + ieB · (∇p A × ∇p B)
∂ω ∂ω

so that the QBE Eq. (2.66) becomes

   
∂ Re Σ ∂ ∂ Re Σ ∂ ∂
i 1− + + (vp + ∇p Re Σ) · ∇R + eE
∂ω ∂T ∂T ∂ω ∂ω
   
∂ Re Σ
+e 1− E + (vp + ∇p Re Σ) × B · ∇p − ∇R Re Σ · ∇p G<
∂ω
∂Σ <
 
∂ Re G <
−ieE · ∇p Re G − ∇p Σ − ieB · (∇p Σ < × ∇p Re G)
∂ω ∂ω
∂Σ < ∂ Re G ∂Σ < ∂ Re G
 
> < > <
=Σ G −G Σ +i − (2.69)
∂pµ ∂Xµ ∂X µ ∂pµ

where we use the four-vector notation pµ = (ω, p) and X µ = (T, R) with the
Minkowski metric ηµν = (+ − −−).
Equation (2.69) is valid to arbitrary order in E and B fields. We see that
the self-energy term [Re Σ, G< ] in the right-hand side of Eq. (2.66) has renor-
malized the velocity in the driving term,

vp → vp + ∇p Re Σ = ∇p (p + Re Σ),

and introduced the factor 1 − ∂ Re Σ

∂ω reminiscent of the wavefunction renor-
malization factor Z −1 in Fermi liquid theory. The other term [Σ < , Re G]
gives some additional terms proportional to the E and B fields. The term
Σ > G< − G> Σ < on the left-hand side is a collision term.
3
Applications

We now consider a few applications of the nonequilibrium formalism.

3.1 Nonequilibrium Transport through a Quantum Dot

In this section we study nonequilibrium transport through a quantum dot
connected to two external metallic leads (two-probe system). Let us first derive
a general expression for the current through the dot in terms of Keldysh
nonequilibrium Green’s functions.
The Hamiltonian of the whole two-probe system is
X  
kα c†kα ckα +HC [{d†n }, {dn }]+ tkα,n c†kα dn + t∗kα,n d†n ckα ,
X
Ĥ =
kα∈L,R kα∈L,R,n
(3.1)
where c†kα (ckα ) is a fermionic creation (annihilation) operator for a single-
particle momentum state k in channel α in the left or right metallic lead, and
d†n , dn are creation/annihilation operators for states in the quantum dot. HC
is the Hamiltonian of the quantum dot. For a noninteracting dot, HC would
be X
HC = hmn d†m dn (3.2)
mn

For a dot with on-site repulsive interactions and a single site with spin-split
levels σ , we use the Anderson model,
X
HC = σ d†σ dσ + U n↑ n↓ (3.3)
σ

with nσ = d†σ dσ is the number operator on the dot.

In any case, let us first derive a general expression for the current regardless
of the presence of interactions in the dot. We however do require that the
interactions, if any, be limited to the electrons inside the dot: the leads should
34 3 Applications

be noninteracting and there should be no interactions between electrons in

the leads and electrons in the dot.
The full Hamiltonian is partitioned as explained in Chapter 1, i.e. H =
H + H 0 where H is the equilibrium Hamiltonian and H 0 is the nonequilibrium
perturbation. We also have to specify the initial noninteracting density matrix
ρ(−∞). In the present problem, we have two possible ways of doing this
partitioning:
• Choose H to be the Hamiltonian of the isolated leads and dot, and H 0 to
be the coupling between the dots. The initial density matrix in the remote
past ρ(−∞) is the product of the density matrices of the isolated systems
in equilibrium:

kα c†kα ckα + HC [{d†n }, {dn }]

X
H=
kα∈L,R
 
tkα,n c†kα dn + t∗kα,n d†n ckα
X
H0 =
kα∈L,R,n
1  −β(HL −µL NL ) (0)

ρ(−∞) = e ⊗ ρC ⊗ e−β(HR −µR NR ) (3.4)
Z
where HL,R are the Hamiltonians of the isolated leads in equilibrium each
with chemical potential µL,R , and the initial density matrix of the dot1
(0)
ρC involves only the noninteracting part of the Hamiltonian of the dot
HC .
• Alternatively, we can choose H to be the Hamiltonian of the connected
system in equilibrium at a single chemical potential µ. The nonequilibrium
perturbation H 0 is then a one-body potential term which raises the single-
particle energies by ∆L,R in the leads. This represents the shift2 in chemical
potential µ → µL,R ≡ µ + ∆L,R . We then have

(0)
(0) (0)
1
The obvious choice is ρC = e−β(HC −µC NC )
where HC is the noninteracting
<
part of HC but µC is unspecified. However this initial condition is included in G> 0
which drops out of the Keldysh equation, see Section 2.1, so that the steady-state
at any finite time t > −∞ after the Keldysh adiabatic evolution is independent of
µC . In particular, the steady-state current is independent of µC . This is a example
of washing out of the initial conditions in the Keldysh formalism (even if here
we are not even considering initial correlations). The steady-state nonequilibrium
population of the dot will be given by the nonequilibrium Green’s function of the
dot G< .
2
To get the correct nonequilibrium self-consistent potential profile throughout the
device including the dot, one would have to include Coulomb interactions in the
Hamiltonian, at least at the Hartree (mean-field) level. Otherwise one can neglect
interactions and add a term in the nonequilibrium perturbation to mimick the
self-consistent potential profile, δH 0 = ∆d† d where ∆ = ∆L +∆ 2
R
, assuming a
symmetric device.
 3.1 Nonequilibrium Transport through a Quantum Dot 35
 
kα c†kα ckα + HC [{d†n }, {dn }] + tkα,n c†kα dn + t∗kα,n d†n ckα
X X
H=
kα∈L,R kα∈L,R,n

∆L,R c†kα ckα

X
H0 =
kα∈L,R
1 −β(H−µN )
ρ(−∞) = e
Z
We will choose the first method which is well-suited to steady-state problems,
while the second method is easier for time-dependent transport.

3.1.1 Expression for the Current

The current flowing through lead α is defined as

Jα (t) ≡ −ehṄα i, (3.5)

where Nα = k c†kα ckα is the number operator for lead α and e > 0 is the
P

electron charge. From the Heisenberg equation of motion ih̄Ṅα = [Nα , H] and
Eq. (3.1), we obtain
2e X
Jα (t) = Re tkα,n G<
n,kα (t, t), (3.6)
h̄
k,n

0 † 0
where we have the mixed lesser Green’s function G< n,kα (t, t ) = ihckα (t )dn (t)i.
Using the Keldysh technique, we will obtain an expression for the mixed
contour-ordered Green’s function Gn,kα (τ, τ 0 ) = −ihTc {dn (τ )c†kα (τ 0 )}i and
perform analytic continuation to real time to obtain the lesser function
0
G<n,kα (t, t ).
For pedagogical reasons, let us first derive the expression from the usual
perturbation expansion of the S-matrix and the subsequent application of
Wick’s theorem, and then obtain it from the path integral method which is
more transparent.

3.1.2 Perturbation Expansion for the Mixed Green’s Function

As explained, we define the unperturbed Hamiltonian as the first two terms

of Eq. (3.1),
kα c†kα ckα + HC [{d†n }, {dn }],
X
H=
kα

and the perturbation as the tunneling term,

X 
H0 = tkα,n c†kα dn + t∗kα,n d†n ckα .
kα,n

The perturbation expansion becomes

36 3 Applications
∞
(−i)l+1
I I
dτl hTc {dˆn (τ )Ĥ 0 (τ1 ) · · · Ĥ 0 (τl )ĉ†kα (τ 0 )}i,
X
0
Gn,kα (τ, τ ) = dτ1 · · ·
l! C C
l=0
(3.7)
where we have
l X D n 
hTc {dˆn (τ )Ĥ 0 (τ1 ) · · · Ĥ 0 (τl )ĉ†kα (τ 0 )}i = Tc dˆn (τ ) ĉ†ki αi (τi )tki αi ,ni dˆni (τi )
Y

i=1 ki αi ,ni
 oE
+dˆ†ni (τi )t∗ki αi ,ni ĉki αi (τi ) ĉ†kα (τ 0 ) .

Wick’s theorem can now be applied to correlators of ĉ, ĉ† fields since H is
quadratic in these fields, whence the importance of the requirement that the
leads be noninteracting. All the contractions involving the first term in Ĥ 0 (τi )

vanish, since dˆn ĉ†ki αi = 0 and dˆni ĉ†kα = 0 for the unconnected system, and the

anomalous contractions dˆn dˆni and ĉ†ki αi ĉ†kα vanish because particle number is
conserved. By Wick’s theorem we thus have
hTc {dˆn (τ )Ĥ 0 (τ1 ) · · · Ĥ 0 (τl )ĉ†kα (τ 0 )}i
l D n oE
(τi )ĉ†kα (τ 0 )
Y X
= t∗ ki αi ,ni Tc dˆn (τ )dˆ† (τi )ĉk
ni i αi
i=1 ki αi ,ni

t∗kl αl ,nl hTc {dˆn (τ )Ĥ 0 (τ1 ) · · · Ĥ 0 (τl−1 )ĉ†kl αl ,nl (τl )}ihTc {ĉkl αl (τl )ĉ†kα (τ 0 )}i
X
=
kl αl ,nl

t∗kl−1 αl−1 ,nl−1 hTc {dˆn (τ )Ĥ 0 (τ1 ) · · · Ĥ 0 (τl−2 )Ĥ 0 (τl )ĉ†kl−1 (τl−1 )}i
X
+
kl−1 αl−1 ,nl−1

×hTc {ĉ†kl−1 αl−1 (τl−1 )ĉ†kα (τ 0 )}i + . . .

t∗k1 α1 ,n1 hTc {dˆn (τ )Ĥ 0 (τ2 ) · · · Ĥ 0 (τl )ĉ†k1 α1 (τ1 )}ihTc {ĉk1 α1 (τ1 )ĉ†kα (τ 0 )}i,
X
+
k1 α1 ,n1

where all the l terms are seen to yield the same contribution to the pertur-
bation expansion Eq. (3.7) if the dummy integration variables τ1 , . . . , τl are
relabeled. Choosing the labeling of the last term, we obtain
∞
X I (−i)l−1
X I I
0
Gn,kα (τ, τ ) = dτ1 (−i) dτ2 · · · dτl
C (l − 1)! C C
k1 α1 ,n1 l=1

×hTc {dˆn (τ )Ĥ 0 (τ2 ) · · · Ĥ 0 (τl )ĉ†k1 α1 (τ1 )}it∗k1 α1 ,n1 (−i)hTc {ĉk1 α1 (τ1 )ĉ†kα (τ 0 )}i.
After relabeling the dummy integration variables once more, we see that the
perturbation terms Ĥ 0 · · · Ĥ 0 give rise to the contour S-matrix Eq. (1.26), so
that we finally have
XI
Gn,kα (τ, τ 0 ) = dτ1 Gnm (τ, τ1 )t∗kα,m gkα (τ1 , τ 0 ), (3.8)
m C
 3.1 Nonequilibrium Transport through a Quantum Dot 37

where

Gnm (τ, τ 0 ) ≡ −ihTc {dn (τ )d†m (τ 0 )}i = −ihTc {Sc (−∞, −∞)dˆn (τ )dˆ†m (τ 0 )}i,
(3.9)
is the contour-ordered Green’s function of the scattering region, and we have

−ihTc {ĉkα (τ )ĉ†k0 α0 (τ 0 )}i = δαα0 δkk0 gkα (τ, τ 0 ),

because of translational invariance in the leads, where gkα (τ, τ 0 ) ≡ −ihTc {ĉkα (τ )ĉ†kα (τ 0 )}i
is the contour-ordered Green’s function of the (isolated) leads.

3.1.3 Path Integral Derivation of the Mixed Green’s Function

The path integral can also be used in the Keldysh technique. The only dif-
ference is that the action is obtained by integrating the Lagrangian over the
Schwinger-Keldysh contour C instead of over the real axis. The Lagrangian is
X X X
¯ d) = d¯n i∂τ dn −HC [{d¯n }, {dn }]− c̄kα tkα,n dn +d¯n t∗kα,n ckα


L(c̄, c, d, c̄kα (i∂τ −kα )ckα +
kα n kα,n
(3.10)
To calculate the propagator Gn,kα (τ, τ 0 ), it is appropriate to define the
Keldysh generating functional Z[η̄, J],
 H P P †

−i dτ :H:− η̄n dn − ckα Jkα
Z[η̄, J] = Tr ρTc e C n kα

where η̄n and Jkα are Grassmann sources, and the normalization is such that
Z[0, 0] = Tr ρ ≡ Z. The generating functional is useful because it can be used
to generate the mixed correlation function,

1 δ 2 Z[η̄, J]
iGn,kα (τ, τ 0 ) = (3.11)
Z δ η̄n (τ )δJkα (τ 0 ) η̄=0,J=0

The generating functional has a path integral representation,

Z H

¯
i S[c̄,c,d,d]+ dτ (η̄d+c̄J)
¯
Z[η̄, J] = D[c̄, c, d, d]e C

P P
where we use the shorthand notation η̄d ≡ n η̄n dn and c̄J ≡ kα c̄kα Jkα .
The d and the c fields are entangled in the Lagrangian Eq. (3.10). To
disentangle them, we perform a shift of variables,
XI
c̄0kα (τ ) ≡ c̄kα (τ ) − dτ1 d¯m (τ1 )t∗kα,m gkα (τ1 , τ )
m C
XI
c0kα (τ ) ≡ ckα (τ ) − dτ1 gkα (τ, τ1 )tkα,m dn (τ1 )
m C
38 3 Applications

where gkα (τ, τ1 ) is the contour-ordered Green’s function of the isolated leads
defined earlier. It satisfies (see Eqs. (2.9) and (2.10) of the notes)

(i∂τ − kα )gkα (τ, τ1 ) = δ(τ, τ1 )

gkα (τ, τ1 )(−i∂τ1 − kα ) = δ(τ, τ1 )

where in the second equation the derivative acts to the left. With this shift of
variables, the Lagrangian becomes
X X I
¯ d) =
L(c̄0 , c0 , d, c̄0kα (i∂τ −kα )c0kα +L0 ({d¯n }, {dn })− dτ1 d¯m (τ )t∗kα,m gkα (τ, τ1 )tkα,n dn (τ1 )
kα kα,mn C

where L0 ({d¯n }, {dn }) =

P ¯ ¯
n dn i∂τ dn − HC [{dn }, {dn }] is the Lagrangian of
the isolated dot.

EXERCISE. Show this (hint: use integration by parts).

As a result, the action becomes

¯ d] = Sleads [c̄0 , c0 ] + SQD [d,
S[c̄0 , c0 , d, ¯ d]

where
XI
Sleads [c̄0 , c0 ] ≡ dτ c̄0kα (i∂τ − kα )c0kα
kα C

is the action of the isolated leads, and

I X I I
¯
SQD [d, d] ≡ ¯
dτ L0 ({dn }, {dn })− dτ dτ1 d¯m (τ )t∗kα,m gkα (τ, τ1 )tkα,n dn (τ1 )
C kα,mn C C

is an effective action for the quantum dot, which contains the effect of the
external leads (bath). The source term η̄d is unaffected by the change of
variables, but the source term c̄J becomes
!
X X XI
c̄J = c̄kα (τ )Jkα (τ ) = 0
c̄ (τ ) + ¯
dτ1 dm (τ1 )t ∗
gkα (τ1 , τ ) Jkα (τ )
kα kα,m
kα kα m C

The generating functional thus becomes

Z H
dτ η̄d+(c̄0 +dt
¯ ∗ g)J )
¯ d]eiSleads [c̄0 ,c0 ] eiSQD [d,d] e C (
¯ i
Z[η̄, J] = D[c̄0 , c0 , d,

where the integration measure D[c̄, c] = D[c̄0 , c0 ] is invariant under the change
of variables since it is only a shift.
We now perform the functional differentiation Eq. (3.11). This generates
two terms. The first term is
 3.1 Nonequilibrium Transport through a Quantum Dot 39
Z
1 ¯ d]eiSleads eiSQD dn (τ )c̄0 (τ 0 ) = hTc {dn (τ )ĉ† (τ 0 )}i
D[c̄0 , c0 , d, kα kα
Z
which is zero because it corresponds to the unconnected system (the action
in the path integral is that of isolated leads). The second term is
Z XI
1 0 0 ¯
D[c̄ , c , d, d]eiSleads iSQD
e dτ1 dn (τ )d¯m (τ1 )t∗kα,m gkα (τ1 , τ 0 )
Z m
XI
= dτ1 iGnm (τ, τ1 )tkα,m gkα (τ1 , τ 0 )
∗

m C

where
Z
1 ¯ d]eiSleads eiSQD dn (τ )d¯m (τ1 )
iGnm (τ, τ1 ) ≡ hTc {dn (τ )d†m (τ1 )}i = D[c̄0 , c0 , d,
Z
Z
1 ¯ d]eiS dn (τ )d¯m (τ1 )
= D[c̄, c, d,
Z
is the exact Green’s function of the connected dot. Hence we obtain
XI
Gn,kα (τ, τ 0 ) = dτ1 Gnm (τ, τ1 )t∗kα,m gkα (τ1 , τ 0 )
m C

which is just Eq. (3.8).

3.1.4 General Expression for the Current

In Eq. (3.6) for the current, we need the lesser mixed Green’s function. We
therefore apply the Langreth analytic continuation theorem Eq. (1.64) to Eq.
(3.8) to get
XZ
Gn,kα (t, t0 ) = dt1 GR ∗ < 0 < ∗ A 0


nm (t, t1 )tkα,m gkα (t1 , t )+Gnm (t, t1 )tkα,m gkα (t1 , t )
m

We now assume steady state so that all the Green’s functions depend only on
the time difference, i.e. G(t, t0 ) = G(t − t0 ). Then we get
X Z dω
∗ ∗
GR < < A


Gn,kα (t, t) = nm (ω)tkα,m gkα (ω) + Gnm (ω)tkα,m gkα (ω) (3.12)
m
2π

The Green’s functions of the isolated leads are

<
gkα (ω) = 2πifα (kα )δ(ω − kα ) (3.13)
A 1
gkα (ω) = (3.14)
ω − iδ − kα

where fα () = (eβ(−µα ) + 1)−1 and the chemical potentials are those of the
initial density matrix Eq. (3.4). We now substitute Eq. (3.12) in Eq. (3.6),
40 3 Applications

2e X Z dω
t∗kα,m tkα,n gkα
<
(ω)GR A <


Jα = Re nm (ω) + gkα (ω)Gnm (ω) (3.15)
h̄ 2π
k,mn

It is not hard to show from the definition of the Green’s functions that the
following relations hold,

G< (ω)† = −G< (ω)

GR (ω)† = GA (ω)

considering the Green’s functions as matrices G = Gnm . Using these relations

and Eqs. (3.13) and (3.14), we can show that Eq. (3.15) becomes [25]
Z
ie dω
Tr Γα (ω) fα (ω)[GR (ω) − GA (ω)] + G< (ω)

Jα = (3.16)
h̄ 2π
where we define a linewidth function

Γα,mn (ω) ≡ 2πρα (ω)t∗α,m (ω)tα,n (ω)

P
and we have introduced the density of states ρα (ω) ≡ k δ(ω−kα ) to convert
the sum over k in Eq. (3.15) to an integral,
X Z
F (kα ) = d ρα ()F ()
k

Equation (3.16) is an exact expression for the steady-state current through

lead α of an arbitrary multiprobe system with interactions inside the dot (but
not in the leads). However, the nonequilibrium Green’s functions have to be
calculated in the presence of interactions, and in the presence of the leads held
at different chemical potentials.
For a two-probe system with proportionate couplings to the leads, ΓL (ω) =
λΓR (ω), one can arrive at a simpler expression. We first use the fact that in
steady state, J = JL = −JR , so that we can write J = xJL − (1 − x)JR for
an arbitrary x. The current then reads
Z
ie dω
Tr ΓR (λx − (1 − x))G< + (λxfL − (1 − x)fR )(GR − GA )
 
J=
h̄ 2π
1
We then fix the arbitrary parameter x = 1+λ so that the first term vanishes
<
and the current does not depend on G anymore. We then obtain
Z  
ie dω ΓL (ω)ΓR (ω)
J= [fL (ω) − fR (ω)] Tr [GR (ω) − GA (ω)] (3.17)
h̄ 2π ΓL (ω) + ΓR (ω)

where the ratio is well-defined since the matrices ΓL and ΓR were assumed to
be proportional.
 3.1 Nonequilibrium Transport through a Quantum Dot 41

3.1.5 Noninteracting Quantum Dot

We derive an alternate expression for the current in the case that there are
no interactions in the dot. In this case, the retarded Green’s function satisfies
the Dyson equation
GR = G R R R R
0 + G0 Σ G , (3.18)
where X
R
Σmn (ω) = t∗kα,m gkα
R
(ω)tkα,n (3.19)
kα

is the noninteracting tunneling self-energy, with Γ = β Γβ = i(Σ R − Σ A ).

P

GR0 is the equilibrium Green’s function of the unconnected dot,

−1
GR
0 (ω) = (ω + iδ − h) .
In other words, in the noninteracting case, the self-energy contains only the
effect of the leads. From these equations it is easy to derive the identity,
GR − GA = −iGR Γ GA
The lesser Green’s function follows from the Keldysh equation,
G< = GR Σ < GA
where the lesser self-energy is
X
Σ< = i fβ Γβ (3.20)
β

Consider a system with a single level, such that all quantities G, Γ, Σ are
scalars. If we write G< in pseudoequilibrium form
G< = if¯A
with A = i(GR −GA ) = GR Γ GA the spectral function, the pseudodistribution
f¯ is
fL ΓL + fR ΓR
f¯ =
ΓL + ΓR
which is clearly not of equilibrium form if fL 6= fR .
Using the previous results, one can show that
Z
eX
Jα = dω[fα (ω) − fβ (ω)]Tαβ (ω)
h
β

where the transmission coefficient Tαβ is

Tαβ (ω) = Tr Γα (ω)GR (ω)Γβ (ω)GA (ω)
which is the usual Landauer formula. Note that Landauer obtained this result
heuristically while here it has been derived rigorously from the nonequilibrium
formalism. Also, the Green’s functions GR and GA have to be calculated out
of equilibrium.
42 3 Applications

3.1.6 Interacting Quantum Dot: Anderson Model and Coulomb

Blockade

In this section we consider on-site repulsive interactions in the dot [26, 27]. We
assume that the couplings are proportionate ΓL ∝ ΓR so we use Eq. (3.17).
The Hamiltonian of the quantum dot is the Anderson model Eq. (3.3),
X
HC = σ d†σ dσ + U n↑ n↓
σ

Assuming that the couplings are diagonal in spin space, the current is
Z  
eX 1 R
J= dω[fL (ω) − fR (ω)]Γσσ (ω) − Im Gσσ (ω)
h̄ σ π

where
†
GR
σσ (t) = −iθ(t)h{dσ (t), dσ (0)}i

is the nonequilibrium retarded Green’s function of the dot (in the presence of
the leads). This Green’s function also satisfies a Dyson equation Eq. (3.18),
but the self-energy now contains both contributions from the leads and from
interactions in the dot. Both contributions are entangled in the exact self-
energy which is not known. Obviously because of interactions the problem
cannot be solved exactly. We will solve the problem approximately by using
R
the equation of motion technique to obtain the exact Green’s function gσσ of
R R
the isolated interacting system, and calculate Gσσ from gσσ by using the non-
interacting tunneling self-energy Σ of the noninteracting Dyson equation Eq.
(3.18). This is actually equivalent to performing a Hartree-Fock factorization
of the higher correlation functions.
For simplicity, we denote the Green’s function of the isolated dot (governed
only by the Hamiltonian HC ) by
R
gσσ (t) = −iθ(t)h{dσ (t), d†σ (0)}i0

where the subscript h· · ·i0 indicates that the average is taken for the isolated
dot. By taking a time derivative, we get
R
∂gσσ (t)
i = δ(t) + θ(t)h{d˙σ (t), d†σ (0)}i0 (3.21)
∂t

We use the Heisenberg equation of motion to obtain d˙σ ,

id˙σ = [dσ , HC ] = σ dσ + U [dσ , n↑ d↓ ] = σ dσ + U dσ nσ̄ (3.22)

where σ̄ ≡ −σ. Substituting this result in Eq. (3.21), we get

R
∂gσσ (t) R (2)
i = δ(t) + σ gσσ (t) + U gσσ (t) (3.23)
∂t
 3.1 Nonequilibrium Transport through a Quantum Dot 43

where we define a 2-particle correlation function

(2)
gσσ (t) ≡ −iθ(t)h{dσ (t)nσ̄ (t), d†σ (0)}i0

We can also write down its own equation of motion,

(2)
∂gσσ
i = δ(t)h{dσ nσ̄ , d†σ }i0 +θ(t)h{d˙σ (t)nσ̄ (t), d†σ (0)}i0 +θ(t)h{dσ (t)ṅσ̄ (t), d†σ (0)}i0
∂t
Since [nσ̄ , dσ ] = 0, the first term is simply δ(t)hnσ̄ i. The last term is zero
since nσ̄ commutes with HC . In the second term we substitute the result of
Eq. (3.22), so that we finally get
(2)
∂gσσ (2) (2)
i = δ(t)hnσ̄ i + σ gσσ (t) + U gσσ (t) (3.24)
∂t
In Fourier space, Eqs. (3.23) and (3.24) become
R (2)
(ω + iδ − σ )gσσ (ω) = 1 + U gσσ (ω)
(2)
(ω + iδ − σ − U )gσσ (ω) = hnσ̄ i

Substituting the second equation in the first, we get

R hnσ̄ i 1 − hnσ̄ i
gσσ (ω) = +
ω + iδ − σ − U ω + iδ − σ
which is the exact Green’s function of the isolated dot. The corresponding
spectral function has two peaks: one peak at ω = σ with spectral weight
1 − hnσ̄ i corresponding to the probability of the site being occupied by a
single electron of spin σ, and a second peak at ω = σ + U with spectral
weight hnσ̄ i corresponding to double occupancy of the site with electrons of
opposite spins.
To obtain the nonequilibrium Green’s function GR σσ of the connected dot,
we make the following ansatz: we take the interacting Green’s function of
R
the isolated dot gσσ as the unperturbed Green’s function GR 0 , and add the
noninteracting tunneling self-energy Eq. (3.19):
1
GR
σσ (ω) ' R (ω)]−1 − Σ R (ω)
[gσσ
which gives the following result,
ω + iδ − σ − [1 − hnσ̄ i]U
GR
σσ (ω) =
(ω + iδ − σ )(ω + iδ − σ − U ) − (ω + iδ − σ − [1 − hnσ̄ i]U )Σ R (ω)
(3.25)
In the simplest case, we neglect the energy dependence of the self-energy, and
have Σ R = −i(ΓL + ΓR )/2. Note that Eq. (3.25) is actually a self-consistent
equation, if we take into account the change in spin population hnσ i out of
equilibrium. Indeed, hnσ i is given out of equilibrium by
44 3 Applications
Z
dω <
hnσ i = hd†σ dσ i = G (ω)
2πi σσ
where the lesser Green’s function of the dot is given by the Keldysh equation
which again involves GRσσ (ω),

G< R < A
σσ (ω) = Gσσ (ω)Σσσ (ω)Gσσ (ω)

where Σ < is the lesser noninteracting tunneling self-energy Eq. (3.20).

Once GR σσ (ω) is obtained by a numerical solution of the self-consistent
equations, we can calculate the current J and the differential conductance
dJ/dV as a function of bias voltage V = (µL − µR )/e. We can also add a gate
potential Vg such that the single-particle energies are shifted, σ → σ + eVg .
In the following figures, we illustrate results for ΓL = ΓR = 0.02 eV,
↑ = −0.65 eV, ↓ = −0.45 eV and U = 1 eV at zero temperature. The
chemical potentials are defined symmetrically µL = −µR = eVb /2. The figures
illustrate clearly the Coulomb blockade effect.
In Fig. 3.1, we display the Coulomb blockade diamond plot, the differential
conductance dJ/dV as function of gate voltage and bias voltage. Note that
linear response theory would give only the Vb = 0 line whereas with the
nonequilibrium theory, we get the full bias-gate voltage diagram.
In Fig. 3.2, we plot two cuts of Fig. 3.1 along Vg , for two values of bias
voltage: Vb = 0 (equilibrium case for which the differential conductance dJ/dV
is just the equilibrium two-terminal conductance G), and Vb = 1 V (out of
equilibrium: the equilibrium conductance peaks split due to the applied bias).
Finally, in Fig. 3.3 we plot the current J itself as a function of Vg and Vb .

3.2 Linear Response for Steady-State and Homogeneous

Systems
We now consider the limit of linear transport, that is, we keep only terms to
first order in the electric field. We will study the QBE Eq. (2.69) in the
absence3 of magnetic field B = 0, and for steady-state (∂/∂T = 0) and
homogeneous (∇R = 0) systems. Consider first the equation for the retarded
Green’s function Eq. (2.68). Neglecting the possible dependence of the self-
energy on the electric field which is a small effect, we have
1
GR = + O(E 2 ) (3.26)
ω − p − Σ R
which to linear order in the electric field is just the equilibrium Green’s func-
tion GR (p, ω). This simplifies drastically the solution of the QBE since all the
retarded and advanced quantities can be taken in equilibrium. The QBE is
3
The B 6= 0 case can also be considered, but then the equilibrium Green’s function
has a dependence on the magnetic field which in general breaks translational
invariance and makes things more complicated.
 3.2 Linear Response for Steady-State and Homogeneous Systems 45

Fig. 3.1. Coulomb blockade diamond: differential conductance dJ/dV in units of

e2 /h, as function of gate voltage Vg and bias voltage Vb . Note the doubling of the
lines due to the level spacing ∆ = ↓ − ↑ = 0.2 eV, and the large U = 1 eV spacing
of the conductance peaks. The equilibrium conductance peaks at Vb = 0 split for
nonzero applied bias Vb 6= 0 (see also Fig. 3.2).
46 3 Applications

ε↑=−0.65 eV, ε↓=−0.45 eV, U=1 eV,ΓL=ΓR=0.02 eV, T=0

0.8
Vb = 0
0.7 Vb = 1 V

0.6

0.5
dJ/dV (e2/h)

0.4

0.3

0.2

0.1

0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
Vg (V)

Fig. 3.2. Differential conductance dJ/dV as function of gate voltage Vg , both at

equilibrium (Vb = 0) and out of equilibrium (Vb 6= 0).
 3.2 Linear Response for Steady-State and Homogeneous Systems 47

Fig. 3.3. Current J in units of e/h, as function of gate voltage Vg and bias voltage
Vb .
48 3 Applications
  
∂ Re Σ ∂
eE · 1− · ∇p + (vp + ∇p Re Σ) G<
∂ω ∂ω
∂Σ <

∂ Re G
− ∇p Re G + ∇p Σ < = Σ < A − Γ G< (3.27)
∂ω ∂ω
This equation can be drastically simplified in the linear response regime since
the terms in the bracket on the left-hand side can be taken in equilibrium,
since they are already multiplied by the electric field. This term can then be
simplified by using the well-known equilibrium expressions,

G<
eq (p, ω) = inF (ω)A(p, ω) ,
<
Σeq (p, ω) = inF (ω)Γ (p, ω) ,
Γ σ
A(p, ω) = 2 2
, Re G(p, ω) = 2 , (3.28)
σ + (Γ/2) σ + (Γ/2)2
where σ(p, ω) ≡ ω − p − Re Σ. In Eq. (3.27), all terms proportional to nF (ω)
vanish and we are left with
i 2 ∂nF   < <
2 A (p, ω) ∂ω eE · (vp + ∇p Re Σ)Γ + σ∇p Γ = Σ A − Γ G (3.29)

Using equation (3.29) to derive transport coefficients is equivalent to using

the Kubo formalism.

3.2.1 Example: Conductivity from Impurity Scattering

To illustrate how transport coefficients are calculated from the QBE, we study
the simple example of the conductivity from impurity scattering in the di-
lute limit. The left-hand side of Eq. (3.29) contains equilibrium quantities as
discussed before. The right-hand side (collision term) vanishes identically in
equilibrium as is immediately seen from the relations (3.28). We keep only
the terms O(E) in the collision term. As seen in Eq. (3.26), retarded quan-
tities and the derived quantities A = −2 Im GR and Γ = −2 Im Σ R have no
O(E) term, hence they are taken in equilibrium. Consequently, only the O(E)
terms in the nonequilibrium functions Σ < and G< in the collision term are
needed. Consider first the collision term. We consider the self-consistent Born
approximation in which the usual equilibrium time-ordered self-energy is (Fig.
3.4)
d3 q
Z
Σ(p, ω) = ni V0 + ni |Vp−q |2 G(q, ω) (3.30)
(2π)3
The nonequilibrium contour-ordered self-energy has the same structure, hence
the nonequilibrium lesser self-energy is
d3 q
Z
Σ < (p, ω) = ni |Vp−q |2 G< (q, ω)
(2π)3
Consider expanding the nonequilibrium Green’s function G< (p, ω) as the equi-
<
librium piece G<
eq (p, ω) plus a nonequilibrium piece G1 (p, ω):
 3.2 Linear Response for Steady-State and Homogeneous Systems 49

Fig. 3.4. Self-energy for impurity scattering in the self-consistent Born approxima-
tion.

G< (p, ω) = inF (ω)A(p, ω) + G<

1 (p, ω) (3.31)

then the equilibrium part of the collision term vanishes as mentioned earlier,
and the collision term is

Σ < A − Γ G< = Σ1< A − Γ G< <

1 ≡ I[G1 ]

where
d3 q
Z
Σ1< (p, ω) = ni |Vp−q |2 G<
1 (q, ω)
(2π)3
We now make the following ansatz for the nonequilibrium part G<
1,

∂nF
G<
1 (p, ω) = −iA(p, ω) eE · vp Λ(p, ω) (3.32)
∂ω
where Λ(p, ω) is a function to be determined. The collision term is then
∂nF
I[G<
1 ] = iA(p, ω)Γ (p, ω)
∂ω
d3 q
 Z 
ni 2
×eE · vp Λ(p, ω) − |Vp−q | A(q, ω)vq Λ(q, ω)
Γ (p, ω) (2π)3
(3.33)

We now consider the left-hand side of Eq. (3.29). In the dilute limit, we
neglect terms involving products of Γ and Re Σ since they are O(n2i ). Hence
we have
50 3 Applications

(vp + ∇p Re Σ)Γ + σ∇p Γ = vp Γ + (ω − p )∇p Γ + O(n2i ) → vp Γ

where in the last step we have neglected the term proportional to ω − p since
it is multiplied by A2 (p, ω) in Eq. (3.29) which is a strongly peaked function
around ω = p in the dilute limit (Re Σ ∝ ni is small). The QBE thus becomes

i 2 ∂nF ∂nF
eE · vp Γ (p, ω) = iA(p, ω)Γ (p, ω)
2 A (p, ω) ∂ω ∂ω
d3 q
 Z 
ni 2
×eE · vp Λ(p, ω) − |Vp−q | A(q, ω)vq Λ(q, ω)
Γ (p, ω) (2π)3

from which we extract an integral equation for the unknown function Λ(p, ω),

d3 q
Z
ni
1
vp Λ(p, ω) = 2 A(p, ω)vp + |Vp−q |2 A(q, ω)vq Λ(q, ω)
Γ (p, ω) (2π)3

which is reminiscent of the integral equation satisfied by the vector vertex

function in the ladder approximation of the diagrammatic Kubo analysis. In
fact, the approximations we have made on the QBE are equivalent to the
ladder approximation in diagrammatic language.
The electric charge current j is given in general by

d3 p
Z Z
dω <
j(R, T ) = −ie vp G (p, ω, R, T )
(2π)3 2π

From the decomposition Eq. (3.31), the equilibrium piece will give a vanishing
contribution and only G< 1 in Eq. (3.32) will give a nonvanishing contribution
to the current. We therefore obtain
d3 p
Z Z  
2 dω ∂nF
j=e vp (vp · E) − A(p, ω)Λ(p, ω)
(2π)3 2π ∂ω

from which the conductivity tensor is easily extracted,

e2 d3 p
Z Z  
dω ∂nF
σµν = 2 pµ pν − A(p, ω)Λ(p, ω)
m (2π)3 2π ∂ω

which is equivalent to the Kubo formula result.

3.3 One-Band Electrons with Spin-Orbit Coupling

Consider a generic translationally invariant spin- 21 Hamiltonian with spin-

orbit coupling,

(−i∇)2
H0 (−i∇) = 1 + λ(−i∇) · σ + µB B · σ
2m
 3.3 One-Band Electrons with Spin-Orbit Coupling 51

where λ(p) is odd in p in general due to time-reversal symmetry, but for

simplicity we neglect higher-order terms and assume it is only linear in p.
Consider for example a two-dimensional electron gas with Rashba and Dres-
selhaus spin-orbit coupling,

HSO = α(py σx − px σy ) + β(px σx − py σy )

where α is the Rashba coupling and β is the linear Dresselhaus coupling. Then
we simply have

λx (p) = αpy + βpx

λy (p) = −(αpx + βpy )

Orbital magnetic effects are taken care of by the Peierls substitution Eq.
(2.60). The Green’s functions and self-energies in the Wigner representation
are now 2×2 matrices. The Kadanoff-Baym driving term [Ĝ−1 <
0 −U, G ]p,ω,R,T
in the presence of electric and magnetic fields can be derived from the general
expression Eq. (2.62). We obtain
   
∂ ∂
[Ĝ−1
0 − U, G <
]p,ω,R,T = i + vp · ∇ R + eE + e(E + vp × B) · ∇ p G
<
∂T ∂ω
−[(λ + µB B) · σ, G< ] + 2i {∇p (λ · σ), ∇R G< }
∂G<
 
+ 2i eE · ∇p (λ · σ), − 2i eB · [∇p (λ · σ) × ∇p G< ]
∂ω ,

(3.34)

where vp = p/m as before, we write simply λ ≡ λ(p) and we use the notation
of Eq. (2.65). The anticommutator is derived in a similar manner from Eq.
(2.63),
"  2 #
−1 R 1 ∂
1
2 {Ĝ0 − U, G }p,ω,R,T = ω − p + 8m ∇R + eE ∂ω + eB × ∇p GR

− 21 {(λ + µB B) · σ, GR } + 4i [∇p (λ · σ), ∇R GR ]

∂GR
 
+ 4i eE · ∇p (λ · σ), − 4i eB · {∇p (λ · σ) × ∇p GR }
∂ω ,

(3.35)

The terms on the right-hand side of the QBE Eq. (2.66) and the equation
for the retarded Green’s function Eq. (2.67) are still given by the general
Kadanoff-Baym commutators Eq. (2.64). Equations (3.34) and (3.35) are valid
to arbitrary order in E and B.
In the case of homogeneous (∇R = 0) transport in zero magnetic field
B = 0, the retarded Green’s function picks up a contribution linear in the
electric field, as opposed to the case of spinless electrons in section 3.2 where
52 3 Applications

the leading order was O(E 2 ), Eq. (3.26). This is because the commutator in
Eq. (3.35)
∂GR
 
∇p (λ · σ), 6= 0
∂ω
does not vanish in general, hence the solution of Eq. (3.35) to linear order in
E is not just the equilibrium Green’s function.

3.4 Classical Boltzmann Limit

In this section, we derive the classical Boltzmann equation (2.1) from the QBE
for 1-band spinless electrons, Eq. (2.69). First, we assume that the quantities
Re Σ and Re G are essentially constant so that their derivatives vanish. In the
QBE, these functions renormalize the transport coefficients that are found
from the classical Boltzmann equation. We also perform the inverse of the
change of variables Eq. (2.40), so that we are back with the untransformed
frequency Ω. Hence we have
 
∂
i + vp · ∇R + e(E + vp × B) · ∇p G< (p, Ω, R, T ) = Σ > G< − G> Σ <
∂T
(3.36)
Consider Eq. (2.25) for the spectral function. It has been shown by Kadanoff
and Baym that Eq. (2.25) for the nonequilibrium spectral function can be
solved in the gradient approximation by the following ansatz,
Γ
A(p, Ω, R, T ) =
(Ω − p − Re Σ − U (R))2 + (Γ/2)2

where here U (R) = −eE · R. In the quasiparticle approximation, we assume

that Γ is very small so that the spectral function approaches a delta function,

A(p, Ω, R, T ) = 2πδ(Ω − p − Re Σ − U (R)) (3.37)

It is well-known that for equilibrium Green’s functions, the fluctuation-

dissipation theorem holds,

G< (p, Ω) = iA(p, Ω)f (Ω) (3.38)

G> (p, Ω) = −iA(p, Ω)[1 − f (Ω)] (3.39)

where f (Ω) is the distribution function. Out of equilibrium we define a non-

equilibrium quantity F (p, Ω, R, T ) such that the nonequilibrium Green’s func-
tions G<,> (p, Ω, R, T ) satisfy analogous relations,

G< (p, Ω, R, T ) = iA(p, Ω, R, T )F (p, Ω, R, T ) (3.40)

G> (p, Ω, R, T ) = −iA(p, Ω, R, T )[1 − F (p, Ω, R, T )] (3.41)
 3.4 Classical Boltzmann Limit 53

Note that these relations are consistent with the exact relation G< − G> = iA
and merely constitute a definition of F . However, within the quasiparticle
approximation Eq. (3.37), the Ω dependence of F is seen to be redundant
because of the delta function and it is sufficient to consider a three-variable
function f (p, R, T ) ≡ F (p, Ω, R, T ) which plays the role of a nonequilibrium
distribution function. Equations (3.40,3.41) thus become

G< (p, Ω, R, T ) = 2πiδ(Ω − p − Re Σ − U (R))f (p, R, T ) (3.42)

>
G (p, Ω, R, T ) = −2πiδ(Ω − p − Re Σ − U (R))[1 − f (p, R, T )](3.43)

where it is seen that the correlation functions G<,> contain information about
the energy spectrum in addition to information about the nonequilibrium dis-
tribution. For a general spectral function A(p, Ω, R, T ), the ansatz (3.40,3.41)
with F (p, Ω, R, T ) ≡ f (p, R, T ) independent of Ω is called the Kadanoff-
Baym ansatz [8]. We now substituting Eqs. (3.42) and (3.43) in Eq. (3.36)
and integrate over all Ω, which is easy because of the delta functions. We
obtain  
∂
+ vp · ∇R + e(E + vp × B) · ∇p f = I[f ]
∂T
which is the classical Boltzmann equation (2.1), where the collision term I[f ]
is

I[f ] = −iΣ < (p, Ω, R, T ) Ω=p +Re Σ−eE·R

[1 − f (p, R, T )]
>
−iΣ (p, Ω, R, T ) Ω=p +Re Σ−eE·R
f (p, R, T )

We can also perform again the Mahan-Hänsch transformation Eq. (2.38),

I[f ] = −iΣ < (p, ω, R, T ) ω=p +Re Σ

[1 − f (p, R, T )]
>
−iΣ (p, ω, R, T ) ω=p +Re Σ
f (p, R, T ) (3.44)

which can also be written simply as

I[f ] = S in [1 − f ] − S out f

where S in = −iΣ < is an in-scattering rate and S out = iΣ > is an out-scattering

rate.
For the sake of definiteness, we investigate a simple example, where colli-
sions are due to electrons scattering off static impurities in the dilute limit. We
treat the interaction at the level of the self-consistent Born approximation. At
this level the self-energy consists of one diagram with one resummed fermion
line and two connected impurity lines. The equilibrium result for time-ordered
quantities is well-known (Eq. (3.30)),

d3 q
Z
Σ(p, ω) = ni V0 + ni |Vp−q |2 G(q, ω)
(2π)3
54 3 Applications

where Vk is the Fourier transform of the impurity scattering potential and ni is

the density of impurities. Since diagrammatic perturbation theory is formally
the same in or out of equilibrium, the relation still holds for the Keldysh
contour-ordered functions and the real-time components follow by the usual
analytic continuation (which is trivial here since there are no products of
contour-ordered quantities):

d3 q
Z
<,>
Σ (p, ω, R, T ) = ni |Vp−q |2 G<,> (q, ω, R, T )
(2π)3
Substituting into the collision term Eq. (3.44) and using Eqs. (3.42,3.43), we
obtain
d3 q
Z
I[f ] = 2πni |Vp−q |2 δ(p − q )[f (q, R, T ) − f (p, R, T )] (3.45)
(2π)3
which is the standard result for impurity scattering. For weak electric fields,
this collision term can be brought in the following form. Assume a homoge-
neous system. To linear order in E, the distribution function is

f (p, T ) = f0 (p) + C(p, T )p · E (3.46)

where f0 (p) ≡ feq (p ) is the equilibrium distribution and C(p, T ) is a scalar
function. Then

f (q, T ) − f (p, T ) = C(p, T )(q − p) · E

since the delta function δ(p − q ) in Eq. (3.45) enforces p = q and p = q.

Hence the collision term becomes
d3 q
Z
I[f ] = −C(p, T )2πni |Vp−q |2 δ(p − q )(p − q) · E (3.47)
(2π)3
Choose p along ẑ, and parameterize E with (θ, φ) and q with (θ0 , φ0 ):

p = pẑ
E = E(sin θ cos φ x̂ + sin θ sin φ ŷ + cos θ ẑ)
q = p(sin θ0 cos φ0 x̂ + sin θ0 sin φ0 ŷ + cos θ0 ẑ)

where we see that cos θ0 = p · q ≡ cos θp,q . We get

(p − q) · E = pE cos θ(1 − cos θ0 ) − sin θ sin θ0 cos(φ − φ0 )

In the integral over q, the term cos(φ − φ0 ) will vanish upon integration over
φ0 , hence it can be omitted. Inserting the remaining first term in Eq. (3.47)
and using Eq. (3.46), we get
f (p, T ) − f0 (p)
I[f ] = −
τp
 3.4 Classical Boltzmann Limit 55

which is the usual form of the collision term in the relaxation time approxi-
mation, and the inverse relaxation time is

d3 q
Z
1
|Vp−q |2 δ(p − q ) 1 − cos θp,q


= 2πni 3
τp (2π)

Notice the important factor (1 − cos θp,q ) which describes the effectiveness of
large-angle scattering in destroying momentum.
References

1. J. Schwinger, J. Math. Phys. 2, 407 (1961).

2. R. A. Craig, J. Math. Phys. 9, 605 (1968).
3. R. Mills, unpublished first, then Propagators for Many-Particle Systems (Gor-
don and Breach, New York, 1969).
4. L. V. Keldysh, Zh. Eksp. Teor. Fiz. 47, 1515 (1964) [Sov. Phys. JETP 20, 1018
(1965)].
5. K.-C. Chou, Z.-B. Su, B.-L. Hao, and L. Yu, Phys. Rep. 118, 1 (1985).
6. E. Calzetta and B. L. Hu, Phys. Rev. D 37, 2878 (1988).
7. P. Danielewicz, Ann. Phys. 152, 239 (1984).
8. L. P. Kadanoff and G. Baym, Quantum Statistical Mechanics (Benjamin, New
York, 1962).
9. S. Fujita, Introduction to Nonequilibrium Quantum Statistical Mechanics
(Saunders, Philadelphia, 1966), and papers.
10. A. G. Hall, J. Phys. A 8, 214 (1975).
11. Yu. A. Kukharenko and S. G. Tikhodeev, Zh. Eksp. Teor. Fiz 83, 1444 (1982)
[Sov. Phys. JETP 56, 831 (1982)].
12. M. Wagner, Phys. Rev. B 44, 6104 (1991).
13. H. Haug and A.-P. Jauho, Quantum Kinetics in Transport and Optics of Semi-
conductors (Springer-Verlag, Berlin, 1998).
14. A. Kamenev, cond-mat/0109316 (2001); cond-mat/0412296 (2004).
15. A. M. Zagoskin, Quantum Theory of Many-Body Systems (Springer, New York,
1998).
16. J. Rammer and H. Smith, Rev. Mod. Phys. 58, 323 (1986).
17. P. Lipavský, V. Špička, and B. Velický, Phys. Rev. B 34, 6933 (1986).
18. D. C. Langreth and J. W. Wilkins, Phys. Rev. B 6, 3189 (1972).
19. D. C. Langreth, in Linear and Nonlinear Electron Transport in Solids (Plenum
Press, New York, 1976), vol. 17 of NATO Advanced Study Institute, Series B:
Physics, edited by J. T. Devreese and V. E. van Doren.
20. J. M. Luttinger and J. C. Ward, Phys. Rev. 118, 1417 (1960).
21. G. Baym and L. P. Kadanoff, Phys. Rev. 124, 287 (1961).
22. G. Baym, Phys. Rev. 127, 1391 (1962).
23. M. Bonitz, Quantum Kinetic Theory (Teubner Stuttgart, Leipzig, 1998).
24. G. D. Mahan, Phys. Rep. 145, 253 (1987).
25. Y. Meir and N. S. Wingreen, Phys. Rev. Lett. 68, 2512 (1992).
58 References

26. Y. Meir, N. S. Wingreen, and P. A. Lee, Phys. Rev. Lett. 66, 3048 (1991).
27. Y. Meir, N. S. Wingreen, and P. A. Lee, Phys. Rev. Lett. 70, 2601 (1993).