PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . .

The Bogoliubov–de Gennes and Andreev Equations:

Andreev Reflection.
References: de Gennes ch. 5, Tinkham ch. 10 §1.

Consider the problem of a system described by a Hamiltonian of the generic form

Ĥ = Ĥ0 + V̂ (1)

where Ĥ0 is a general single-particle Hamiltonian of the form

X 1 
2
Ĥ0 = (pi − eA(ri )) + U (ri , σi ) − µ (2)
2m
i

and V̂ is an interaction which may (inter alia) lead to pairing: we shall often use the
specific ‘contact’ form
1X
V = V (ri − rj ), V (ri − rj ) = −gδ(ri − rj ) (3)
2
ij

with some phenomenological energy cutoff, but the argument is actually more general.
Up to now we assumed that when Cooper pairs form, they form in eigenstates of Ĥ0
related by time reversal, i.e. that the many-body wave function is of the form

(un + vn a†n a†n̄ )|vaci

Y
Ψ= (4)
n

where n and n̄ are eigenstates of Ĥ0 corresponding to the same eigenvalue n = n̄ .
In general, however, this scheme is either not possible or not optimal. If either A 6= 0
or U is a (time reversal violating) function of the spin σ, then in general eigenstates of
Ĥ0 related by time reversal do not exist. Even for the case of no vector potential and
a spin-independent U (r), when time-reversed pairs do exist, the pairing scheme used in
the last lecture may not be optimal, even for the ground state; generally speaking, this
tends to happen when the physical conditions vary over a length scale . ξ0 . Finally,
even when the conditions are slowly varying over ∼ ξ0 , we may wish to discuss states
where the condensate is moving, which clearly breaks T-invariance. So a more general
scheme is called for.
This more general scheme goes under the name of the Bogoliubov–de Gennes equa-
tions (or technique). For simplicity, I first just quote the principal results of the discus-
sion ∗ : The Hamiltonian can be cast in the form
X
†
Ĥ = E0 + En αnσ αnσ (5)
nσ
∗
For simplicity I assume at this point (with de Gennes, section 5.1) that the potential term is spin-
independent, U (r, σ) ≡ U (r). The opposite case will be discussed in lecture 12.
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 2

where the α’s are fermion quasiparticles operators satisfying the anti-commutation rela-
tions
[αnσ , αn† 0 σ0 ]+ = δnn0 δσσ0 (6)
and the En are the positive solutions of the pair of equations (the BdG equations) for
two functions un (r), vn (r):

Ĥ0 un (r) + ∆(r)vn (r) = En un (r) (7)

−Ĥ0∗ vn (r) + ∆∗ (r)un (r) = En vn (r) (8)

where the local energy gap must be self-consistently determined according to the pre-
scription (given explicitly for the contact potential −gδ(r))
X
∆(r) = g vn∗ (r)un (r) tanh βEn /2 (9)
n

The ground state energy is given by the expression (relative to N µ)

Z Z
X
2 1
E0 = − En |vn (r)| dr + |∆(r)|2 dr (10)
n
2g

Note that in the limit ∆(r) → 0,the vn (r) are simply the single-particle eigenfunctions
of the one-particle Schrödinger equation with negative energy (relativePto µ), and since
these are normalized (cf. below) the expression for E0 reduces to − n <0 |n |, as it
should.
Although the eigenvalues En of the BdG equations are independent of normalization,
the value of ∆(r) in eqn. (9) depends on it: we have implicitly chosen the normalization
Z
|un (r)|2 + |vn (r)|2 dr = 1
 
(11)

This is a special case of a more general orthonormality relation, see below.

In the literature the BdG equations are almost invariably derived by writing down the
Hamiltonian in second-quantized form, making a generalized mean-field approximation
so that

gψ↑† (r)ψ↓† (r)ψ↓ (r)ψ↑ (r) = ∆(r)ψ↑† (r)ψ↓† (r) + h.c., where ∆(r) ≡ ghψ↓ (r)ψ↑ (r)i (12)

and diagonalizing the resulting ‘mean-field’ Hamiltonian (which contains of course both
the above terms and the usual ones ∝ ψ † (r)ψ(r)) by the generalized Bogoliubov trans-
formations Xh i
†
ψσ (r) = αnσ un (r) − σαnσ vn∗ (r) (13)
n
etc.
This procedure is very standard and is well explained e.g. in de Gennes §5.1; however,
it does not by itself give much insight into the nature of the groundstate, so I shall take
an alternative route.
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 3

Let’s first consider an even number of fermions at T = 0 and thus consider a single
state of the system (which need however not necessarily be the ground state). As in
lecture 5 we assume the general form of the wave function corresponds to the formation
of Cooper pairs; explicitly

ΨN = N A [φ(r1 σ1 ; r2 σ2 )φ(r3 σ3 ; r4 σ4 ) . . . φ(rN −1 σN −1 ; rN σN )] (14)

where all the φ’s are exactly the same. In second-quantized notation this reads:
( )N/2
X ZZ
ΨN = N drdr0 Kσσ0 (r, r0 )ψσ† (r)ψσ† 0 (r0 ) |vaci (15)
σσ 0

where Kσσ0 (r, r0 ) may be taken antisymmetric under the exchange rσ

r0 σ 0 . If we
introduce an arbitrary orthonormal one-particle basis χi (r, σ) then
nX oN/2
ΨN = N Kij a†i a†j |vaci (16)
ij

where Kij = −Kji . Now, there is a general theorem∗ that any antisymmetric even-rank
square matrix can be ‘skew-diagonalized’, that is written in the form
 
0 K1 0 0
 −K1 0 0 0 
 
 0 0 0 K2 (17)

 
 0
 0 −K 2 0 

..
.

Hence we can rewrite ΨN in the form

nX oN/2
ΨN = N cn a†n a†n̄ |vaci (18)
n

where the sets n and n̄ have zero intersection and together form an orthonormal basis.
The simple BCS form of ΨN is clearly a special form of (18) with n = (k ↑), n̄ = (−k ↓).
Just as in that case, we can relax particle number conservation and replace (18) by the
form
vn
(un + vn a†n a†n̄ )|vaci, |un |2 + |vn |2 = 1,
Y
Ψ= = cn (19)
n
u n

We use capital letters for the un and vn to emphasize that these will not necessarily
have all the properties associated with un and vn which we met e.g. in lecture 9, and
moreover in general bear no simple relation to the quantities un (r), vn (r) introduced in
eqns. (7-8), c.f. below.
∗
A proof can be found in the well-known paper of Yang. Rev. Mod. Phys 34, 694 (1962)
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 4

We can follow the original argument a little further. We define the ‘pair wave func-
tion’ F (rr0 , σσ 0 ) ≡ hψσ0 (r0 )ψσ (r)i, and find that it is given by the expression
X
F (rr0 , σσ 0 ) = un vn χn (rσ)χn̄ (r0 σ 0 ) (20)
n

We now write down the expectation value of the Hamiltonian as a function of the trial
functions χn (r), χn̄ (r) and the constants (which, we must remember, are in general
complex, though the un can always be chosen real by convention). The pairing term is
a simple generalization of the form in lecture 10:
Z Z X 2
2
hV ipair = −g |F (r, r)| dr = −g dr  un vn χn (r)χn̄ (r) (21)
 
n

The single-particle energy has the form

X
|vn |2 (χn , Ĥ0 , χn ) + (χn̄ , Ĥ0 , χn̄ )
 
hĤ0 i = (22)
nσ

This can be cast in a more useful form by adding and subtracting a term
1 Xh i
K0 = |un |2 (χn , Ĥ0 , χn ) + |vn |2 (χn̄ , Ĥ0 , χn̄ ) (23)
2 nσ

The quantity in brackets turns out (see below) to be just the trace of Ĥ0 which is simply
a constantP(infinite, if no cutoff is imposed): compare the addition and subtraction of
the term k k in the translation-invariant case. Thus we can ignore this term in the
expectation value of hĤ0 i which then takes the form
Xh i
hĤ0 i = − |un |2 (χn , Ĥ0 , χn ) + |vn |2 (χn̄ , Ĥ0 , χn̄ ) (24)
n

where the factor of 1/2 has been cancelled against the sum over spins.
Thus, the expectation value of the total Hamiltonian in the ground state∗ has the
form
Xh Z Z i Z X 2
∗ ∗
hĤi = 2 2
− |un | χn Ĥ0 χn dr + |vn | χn̄ Ĥ0 χn̄ dr − g dr  un v∗n χn (r)χn̄ (r)
 
n n
(25)
It is now tempting to try to minimize expression (25) with respect to a) the form of the
functions χn (rσ), χn̄ (rσ) and b) the coefficients vn and un . Indeed we shall see below
that the result of doing so without additional constraints is precisely the BdG equations.
Unfortunately, however, we now come to a major difference with the simple BCS case
(or generalizations of it such as that of lecture 9): In general, the BdG equations do not
guarantee that the χn and χn̄ individually form a complete orthonormal set. Thus, if
∗
or more generally in the ‘completely paired’ state we are considering.
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 5

we want to preserve this state of affairs (as we must), it is necessary to introduce the
constraint of orthogonality explicitly, e.g. by means of appropriate Lagrange multipliers,
and the resulting equations are so messy as to make explicit solution usually impossible.
Thus, although we can always write the ground state in the form (19), there is in general
no simple relation between the quantities occurring in that equation and the solution of
the BdG equations.
It may however be instructive to consider the special case defined by the requirement
that the solutions un (r), vn (r) of the BdG equations have the properties (which, we
repeat, is not guaranteed by the equations themselves)




un (rσ), un0 (rσ) = vn (rσ), vn0 (rσ) = δnn0 , un (rσ), vn0 (rσ) = 0 (26)

(The second equation is automatically satisfied if, for example, all the un correspond
to spin up and all the vm to spin down; but the first need not be satisfied even under
that condition). If these conditions are satisfied, then it turns out (and can be checked
a posteriori) that the orthogonality of the χn ’s and χn̄ ’s individually is automatically
satisfied and does not have to be enforced by Lagrange multipliers (the normalization
condition still needs one, however: cf. below). It is then convenient to introduce the
notation
un (r) ≡ un χn (r), vn (r) ≡ v∗n χ∗n̄ (r) (27)
Notice that association of un (r) and vn (r) with χn (r) and χn̄ (r) respectively (rather
than the other way around) is purely conventional, and that even though we can always
choose the un real by convention, un (r) may be nontrivially complex because we are not
guaranteed that χn (r) is real. In terms of the functions un (r), vn (r) the expectation
value of the energy takes the form
X Z Z Z X 2
∗ ∗ ∗

hĤi = − dr un (r)Ĥ0 un (r) + dr vn (r)Ĥ0 vn (r) − g dr  un (r)vn (r) (28)
 
n n

and the normalization conditions on the χ’s and un , vn above are equivalent to the single
condition Z
Kn ≡ dr |un (r)|2 + |vn (r)|2 = 1

(29)

As usual, we handle the constraint by P introducing a Lagrange multiplier λn for each n

and minimizing the quantity hĤi − n λn Kn . Anticipating the result, it is convenient
P
to relabel the λn as En . Then explicit minimization of hĤi − n λn Kn with respect to
u∗n (r) and vn (r) separately gives two equations for each n:

Ĥ0 un (r) + ∆(r)vn (r) = En un (r)

(30)
∆(r)un (r) − Ĥ0 vn∗ (r) = En vn∗ (r)

where ∆(r) is a shorthand for the quantity

X
un (r)vn∗ (r) ≡ gF ∗ (r, r)


∆(r) ≡ g (31)
n
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 6

The first of eqns. (30) is the first BdG equation, and the second is turned into the second
BdG equation by complex conjugation (note that e.g. in the presence of a vector poten-
tial, the Hamiltonian is Hermitian but not real, so in the general case it is necessary to
complex-conjugate it.) Thus for the special case considered we recover in this way com-
plete argument with the usual textbook treatment, and there is a simple correspondence
between the GSWF and the solution of the BdG equations. Again we emphasize this is
not the generic case.
At this point it is useful to note some general properties of the solutions to the
BdG equations. We can regard the un (r) and vn (r) as respectively the upper and lower
components of a ‘spinor’ ψn (r) and the effective Hamiltonian acting on this spinor is
then of the form of a matrix:  
Ĥ0 ∆(r)
(32)
∆∗ (r) −Ĥ0∗
This matrix operator is Hermitian, and therefore by the usual theorems the solutions
can be chosen to form a complete orthonormal set, in the sense that




ψn (r), ψn0 (r) ≡ un (r), un0 (r) + vn (r), vn0 (r) = δnn0 (33)

Because of the completeness, the trace of any single-particle operator Q̂ can be written
in the form X
Tr Q̂ = (un , Q̂ un ) + (vn , Q̂ vn ) (34)
n

which is what justified us earlier in asserting that the expression (23) above was just the
trace of Ĥ0 . A further conclusion follows from the observation that if (un , vn ) is a spinor
satisfying the BdG equations with eigenvalue En , then the spinor (vn∗ , −u∗n ) is equally
a solution with eigenvalue −En . Since spinors with eigenvalues En and −En0 must be
orthogonal, this gives a second orthogonality condition, namely

(un , vn∗ 0 ) − (vn , u∗n0 ) = 0 (all n, n0 ) (35)

Note that eqns. (33) and (35), while generally true, do not in general imply eqns. (26).
It is often asserted in the literature that the negative-energy solutions of the BdG
equations correspond to the states occupying the ‘filled Fermi sea’. Personally I find this
point of view more confusing than helpful, and find it more informative simply to regard
the ‘creation operators’ associated with them as annihilating the groundstate.

The Andreev equations

A.F. Andreev, Soviet Physics JETP 19, 1228 (1964)

The Andreev equations are the ‘semiclassical limit’ of the more general BdG equations.
Consider a situation where the ‘diagonal’ potential U (r) can be taken as a constant
(and thus absorbed in the zero of energy) and the gap ∆(r) is slowly varying over
scales of the order of kF −1 (though it may vary appreciably over scales ∼ ξ0 ). Such a
situation may obtain, for example, in the intermediate state of a type-I superconductor
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 7

near an N-S interface (Andreev’s original problem), in the case of a vortex in a type-II
superconductor or (a case of great current interest) in an ‘exotic’ superconductor near a
so-called pair-breaking surface.
Under these conditions the single-particle Hamiltonian reduces simply to the kinetic
energy operator −(~2 /2m)∇2 − µ. It is convenient to take out a ‘fast oscillating’ piece
of the BdG functions un (r), vn (r). Let us in fact concentrate on a particular pair n and
make the ansatz
u(r) ≡ f (r) exp ikF n̂ · r
(36)
v(r) ≡ g(r) exp ikF n̂ · r
where n̂ is a unit vector on the Fermi surface. On substituting these forms into the BdG
equations and noting that ~2 kF 2 /2m = µ we obtain a term proportional to kF |∇f | (etc.)
and one proportional to ∇2 f . Assuming that f and g themselves vary on the scale of
the variation of ∆(r), i.e. slowly on the scale of kF−1 , we can drop the second term in
comparison to the first. Then, introducing the notation vF ≡ ~kF /m as usual, we obtain
the Andreev equations §
−i~vF n̂ · ∇f (r) + ∆(r)g(r) = Ef (r)
(37)
−i~vF n̂ · ∇g(r) + ∆∗ (r)f (r) = Eg(r)
Being linear rather than quadratic in the gradients, these equations are usually easier to
solve than the full BdG equations. A crucial point is that since the combined equations
are second order, we expect in general 2 solutions for each energy eigenvalue E.

Andreev reflection
Consider a situation in which the gap ∆(r) is either constant (in magnitude and phase) or
very slowly varying, let us say for definiteness in the z-direction. Under these conditions
the Andreev equations can be combined to give a single Schrödinger-like equation for
f (r), namely

2
−(~vF )2 n̂ · ∇ f + |∆(r)|2 f = E 2 f (38)
and an identical equation for g. The general solution involves a dependence on x and y
of the form exp iq⊥ · r⊥ (r⊥ ≡ (x, y)) with q⊥ a constant vector, but it is clear that this
can be removed by an appropriate choice of n̂(⊥ q⊥ ), so we can take f to be a function
only of z and to satisfy the equation
∂2f
−(~vF nz )2
+ |∆(z)|2 f = E 2 f (z) (39)
∂z 2
This second-order equation has two independentRsolutions, which in the limit of slowly
varying ∆(z) have the semiclassical form exp ±i k(z) dz, where

1 p
k(z) ≡ E 2 − |∆(z)|2 (40)
~vF nz
§
In the literature one sometimes finds slightly different (but equivalent) forms, which result e.g. from
a different choice of phase for ∆(r).
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 8

The two solutions represent ‘particle-like’ and ‘hole-like’ states: if nz is positive, the
+ sign corresponds to the particle. This can be seen most easily by taking the limit
∆(z) → 0, when E → |k |.
Consider the momentum and velocity of these two
k(z)
states. The momentum is ~[kF n̂ ± k(z)ẑ], and thus
is little different for the two. However, the velocity is kF n̂ !k
given by the standard expression 1/~ ∂E(k)/∂k, where
k = kF n̂ ± k(z)ẑ. Here we have to be a little careful,
since a dependence on k arises through n̂. The eas-
iest technique is to define the normal-state KE k by
k ≡ ±~vF nz k(z); since ∂k /∂k = ~2 k/m ≈ ~vF n̂, we
have
∂Ek k ~vF k(z)
v≈ vF n̂ = vF n̂ = ± vF n̂ (41)
∂k Ek Ek
Thus the velocity is (nearly) opposite on the two branches.
Lets consider a situation where ∆(z) → 0 as z → −∞ and → some finite value
∆0 as z → +∞ (such a situation might arise, for example, near an N-S interface in a
type-I superconductor). It is clear that for E < ∆0 , there exist solutions of (39) for
z → −∞ but not for z → +∞, so that the packet properties towards positive z: in the
region ∆(z) = 0, which is simply a normal metal, these are simply ‘electron-like’ wave
packets. Now, it is clear that for E < ∆0 the packet cannot propagate to z = +∞.
On the other hand, it apparently cannot be reflected in any ordinary sense either: a
‘single-shot’ reflection in which k → −k corresponds (at best) to a matrix element of the
form ∆(z) exp 2ikF z dz, which since ∆ is assumed slowly varying over a range ∼ kF −1
R

is exponentially small, while a process of gradual reflection in which k changed step by

step to −k is blocked by the Fermi sea (Pauli principle). So the only possibility is that
the ‘electron-like’ wave packet is reflected as a ‘hole-like’ packet! (‘Andreev reflection’)
Formally, this is possible because the true eigenstates are of the form of definite linear
combinations of the electronlike and holelike solutions, i.e.
Z Z
f (z) = A exp +i k(z) dz + B exp −i k(z) dz (42)

where the magnitudes of A and B are equal but the relative phase is determined by the
condition that f (z) → 0 for z → +∞ (where of course the simple form (42) is not valid).
A rather easier way of understanding Andreev reflection is to treat the wave packet
semiclassically. Write

1/2
E(k, z) = 2k + |∆(z)|2 , (k ≡ ~nz k) (43)

(so that conservation of E leads to equation 40)). For simplicity let’s take ∆(z) real and
nz = +1 (particle propagating normal to interface). Then we can write for the position
PHYS598 A.J.Leggett Lecture 11 The Bogoliubov–de Gennes and Andreev Equations . . . 9

z(t) and k-vector (relative to kF ) k(t) the semiclassical equations

dz(t) 1 ∂E ~vF
= = vF k(t)
dt ~ ∂k E (44)
dk(t) 1 ∂E 1 ∆(z) ∂∆(z)
=− =−
dt ~ ∂z ~ E ∂z

It is clear that k(t) decreases uniformly in time,

and eventually goes through zero at the point where
∆(z) = E (‘classical turning point’). At this point
the velocity of the packet reverses, and it emerges as
a hole-like packet propagating back towards z = −∞.
It is important to note that in the more general case
not only the z-component but also the transverse com-
ponents of the velocity are reversed, in contrast to the
familiar case of specular reflection. Note also that al-
though the packet undergoes some change of momentum on Andreev reflection, this is
small compared to that (∼ 2~kF ) on ordinary reflection.
The semiclassical description is of course not a complete account: even for E < ∆ it
must fail very close to the classical turning point, and moreover it would predict that for
E > ∆ we get 100% transmission. Whereas a proper quantum-mechanical solution of the
Andreev equations show that even for E > ∆ we can get a finite reflection amplitude(the
exact value depends on the shape of the function ∆(z); generally speaking, the sharper
the spatial variation of ∆ the greater the reflection amplitude).