S.H.

Lim

Rigorous Treatment of the Grand Hamiltonian

Students of condensed matter physics may first encounter the Grand Hamiltonian when
studying the microscopic theory of superfluidity and superconductivity. The Grand Hamil-
tonian is defined as:

ĤG = Ĥ–uN̂

As an operator, we are told that it should be diagonalized in order to calculate the

ground state and excitation energies of the system. However, treatments of the subject
give little or no justification on why this is so. Some authors may allude to the fact that
when Ĥ is approximated as a mean field operator, it no longer preserves particle number.
But students may still struggle to see how including the term uN̂ fixes the problem.
Furthermore, it is not clear what is being achieved mathematically, as the operator ĤG
still contains bad terms that do not conserve particle number. Therefore it is the aim of
this paper to understand the origin of the term uN̂, and provide step-by-step reasoning to
justify using ĤG in place of just Ĥ. It will also be shown that up until now, u has not been
treated rigorously, which leads to conceptual difficulties about particle conservation. The
key question to answer is how do we determine what is the appropriate Hamiltonian to
use for a particular problem, and the answer should come from Ritz Variational method.

The Hamiltonian of Schrodinger’s Equation

Students of Quantum Mechanics may be used to thinking of the Schrodinger equation as
a postulate. More advanced students may also be aware of the Ritz Variational principle,
and understand that the formulation of Quantum Mechanics as a minimization problem
gives the same solutions as solving the Schrodinger equation. Indeed both formulation’s
are equivalent, but when we are unsure about what is the Schrodinger equation we should
be looking at, it is helpful to start from the minimization problem.
 S.H. Lim

Take a N particle wavefunction ψ(x1 , x2 , ..., xn ) and Hamiltonian operator H. The

energy is given by hψ|Ĥ|ψi. The ground state energy (E0 ) is then the minimum energy
for any possible wavefunction ψ, subject to the normalization constraint hψ|ψi = 1. The
ground state wavefunction is by definition the wavefunction that minimizes the energy.
The term hψ|Ĥ|ψi is a functional with degrees of freedom ψ and ψ ∗ . By calculas of
variation, the ground state is found by solving:
∂ hψ|Ĥ|ψi ∗ ∂ hψ|Ĥ|ψi
δψ + δψ = 0 (1)
∂ψ ∗ ∂ψ
The second term is just the complex conjugate of the first and it can be ignored1 . At
this point, we can already see a problem. We get the trivial solution ψ = 0 because we
have not incorporated the normalization constraint hψ|ψi = 1. To formally impose the
constraint, we can apply the method of Lagrange multipliers. The steps are as follows,
first formulate the constraint equation such that the right side equals zero:

hψ|ψi − 1 = 0 (2)

Multiply the left hand side with a term (ǫ) and subtract it from hψ|Ĥ|ψi. ǫ is called the
Lagrange multiplier and is to be treated as an additional degree of freedom.

hψ|Ĥ|ψi − ǫ(hψ|ψi − 1) (3)

Minimization of the term (3) with respect to the degrees of freedom (ψ, ǫ) ensures that
the solution truly minimizes hψ|Ĥ|ψi, subject to the normalization constraint. That is
the result according to the method of Lagrange multipliers. The student should review
text concerning the calculus of variations if they are unfamiliar with this method. The
point is that this step is a purely mathematical one. If we now try to minimize (3) with
respect to ψ ∗ , we get:
∂ hψ|Ĥ|ψi ∂ hψ|ψi
− ǫ =0 (4)
∂ψ ∗ ∂ψ ∗
which can be rewritten as:
Ĥ |ψi = ǫ |ψi (5)

Finally, we have derived the Schrodinger equation.

1
https : //[Link].f amu.f [Link]/ dommelen/quantum/stylea/[Link]
 S.H. Lim

Superconducting ground state

After understanding how to derive the Schrodinger equation, we can now turn to the
superconducting ground state, as first analyzed by Schriefer1 . We can quickly show how
to derived ĤG from Schriefer’s energy minimization problem.
According to BCS theory, the electron-phonon coupling gives rise to an effective
electron-electron interaction of the form:
X
Vkq P̂k+ P̂q (6)
k,q

P̂k+ = â+ +
k↑ â−k↓ (7)

The Schrodinger equation is then:

!
X X
ǫk (ĉ+ +
k↑ ĉk↑ + ĉk↓ ĉk↓ ) + Vkq P̂k+ P̂q |ψi = Ĥ |ψi = E |ψi (8)
k k,q

Consider the solutions that we will get if we could solve this exactly. Since we formulated
the problem using second quantization, we will get all possible particle number wave-
functions, meaning solutions with 0, 1 particle, 2 particles, 3 particles,...,N, N+1,... and
associated with each N particle state will also be the full multiplicity of all possible energy
eigenstates. The ground state has zero particles, which is the price we pay for using the
second quantization formulation. This inconvenience can be remedied if we could somehow
constrain the problem so that the lowest energy wavefunction has N particles. In order
to do that we start from the energy minimization problem, that is we seek to minimize
hψ|Ĥ|ψi with some constraints on ψ.2 As before, we can write a constraint equation:

hψ|N̂|ψi − N = 0 (9)

â+
P
where N̂ is the particle number operator k k âk . This does not guarantee that the

ground state of the system will be in a state of N particles, but it will have N particles
on average. This is the limitation given our formulation using second quantization, but is
this really the best that we can do? We will see that we a forced to accept this limitation
when we turn on the mean field approximation.
2
Scriefer himself explained that this is a Lagrange multiplier problem, so this is nothing new
 S.H. Lim

Now, on top of the first constraint, lets not forget that we also have the normalization
constraint:

hψ|ψi − 1 = 0 (10)

Using the method of Lagrange multipliers, we reformulate the problem as finding the
minimum of:

hψ|Ĥ|ψi − u( hψ|N̂|ψi − N) − E(hψ|ψi − 1) (11)

where u and E are Lagrange multipliers. Taking the variation with respect to ψ ∗ and
equating to zero, we get:

Ĥ |ψi − uN̂ |ψi = ĤG |ψi = E |ψi (12)

In other words, instead of solving (8), we are told that we should solve (12). Note that
in the original BCS theory, the term that Scriefer minimizes is just hψ|ĤG |ψi, which
ignores the E hψ|ψi term. The reason they are able to do so is because the normalization
constraint is explicitly included in the BCS wavefunction, as we will see shortly.
At this point, it would be instructive to examine what is the difference between the
solutions of (8) vs (12). Equation (12) includes just one additional term, uN̂ |ψi. However,
we observe that the N̂ operator commutes with Ĥ. Therefore, the solutions of both
equations are exactly the same! The only difference will be that the eigenenergies of (12)
are shifted by u. Therefore the effect of uN̂ seems to be trivial. This should be reassuring;
remember that u was introduced for computational convenience. We do not actually want
the physics to change! But at this point, the introduction of u has no computational benefit
either, since we are calculating the exact solution. However, once we take the mean field
approximation, the effect of u will be non-trivial. It actually affects the accuracy of the
approximation, and if we adjust u correctly, we will get the closest approximation to the
true N-particle ground state.
Next we introduce the BCS wavefunction, which is a wavefunction of uncorrelated
 S.H. Lim

pairs.
Y
|ψBCS i = |ψk i (13)
k

|ψk i = (uk + vk P̂k+ ) |0i (14)

P̂k+ = â+ +
k↑ â−k↓ (15)

The function to minimize is:

X
hψBCS |Ĥ − uN̂|ψBCS i − Ek hψk |ψk i
k
X X X (16)
= 2(ǫk − u)|vk |2 + Vkq u∗q v∗k uk vq − Ek (u2k + v2k )
k k,q k

To get the Schrodinger equation, we take the variation with respect to the degrees of
freedom, u, v:
X
2(ǫk − u)vk + Vkq u∗q vq uk = Ek vk (17)
q
X
Vkq v∗k uk vq = Eq uq (18)
k

A very strange looking Schrodinger equation(s) indeed! As in the toy problem [Appendix
A], we can massage the equations to make them into an operator × ket equation. First,
notice we can substitute v∗ u with hPˆ+ i and u∗ v with hPˆ i. Also we multiply (17) with
k k k q q q

|Pk i and (18) with |0i. The equations become:

" #
Vkq hPˆq i uk |Pk i = Ek vk |Pk i
X
2(ǫk − u)vk + (19)
q
X
Vkq hP̂q+ i vk |0i = Ek uk |0i (20)
q

where we have swapped the indices of k and q for eqn (18). Also we utilized the fact that
Vkq is symmetric. A few more tricks we need:

2vk |Pk i = (â+ + +

k↑ âk↑ + â−k↓ â−k↓ )(uk + vk P̂k ) |0i (21)
uk |Pk i = P̂k+ (uk + vk P̂k+ ) |0i (22)
vk |0i = Pˆk (uk + vk P̂k+ ) |0i (23)
 S.H. Lim

Putting it all together, we have:

" #
hP̂q+ i Pˆk )
X
(ǫk − u)(â+
k↑ âk↑ + â+
−k↓ â−k↓ ) + Vkq ( hP̂q i P̂k+ + (uk + vk P̂k+ ) |0i
q (24)
= Ek (uk + vk Pˆk+ ) |0i

Notice that the operators in the bracket [...] only operate on the k state. Therefore we
can just as well replace the single particle wavefunction (u + v Pˆ+ ) |0i with |ψ i on k k k BCS

the left and right side of equation (24). Finally, taking a summation over k, we get:
" #
(△∗k P̂k+ + △k Pˆk − △k hPˆk i) |ψBCS i = Etot |ψBCS i
X X
(ǫk − u)(â+ +
k↑ âk↑ + â−k↓ â−k↓ ) −
k k

(25)

Vkq hPˆq+ i
X
△k = − (26)
q

where we subtract a term like △k hPˆk i to avoid double counting the interaction energy.
We recover the mean field Hamiltonian of the Bogoliubov theory.3

Particle conservation
The BCS wavefunction is a superposition of particle number states, meaning it is not an
eigenstate of N̂ . But as advertised we can make the wavefunction have a expectation
value of N particles by introducing the u term and tuning it appropriately. The reason
why we can’t have a wavefunction with fixed particle number is because of the mean field
approximation. Notice that the interaction terms in (25) couple states of different particle
number. The utility of the mean field Hamiltonian is that it can be easily diagonalized
using the Bogoliubov transformation:
3
xx
 S.H. Lim

γ̂k1 = uk âk↑ − vk â+

−k↓ (27)
γ̂k2 = uk â−k↓ + vk â+
k↑ (28)
1 1
u2k = (1 + (ǫk − u)/Ek ); v2k = (1 − (ǫk − u)/Ek ) (29)
2 q 2
Ek = (ǫk − u)2 + △2k (30)

With these new operators, we can easily calculate the ground state energy and excita-
tion levels. The question is what happens to the particle number when we introduce an
excitation? In order to excite the field, we apply the creation operator one time to the
ground state:
+
γ̂k1 (uk + vk P̂k+ ) |0i (31)
= (uk â+ +
k↑ − vk â−k↓ )(uk + vk P̂k ) |0i (32)
= (u2k â+ 2 + + 2 + 2 + +
k↑ − vk â−k↓ âk↑ â−k↓ ) |0i = (uk âk↑ + vk âk↑ â−k↓ â−k↓ ) |0i (33)
= (u2k + v2k )â+
k↑ |0i = |k ↑i (34)

Close to the fermi sea, there will be a level where |uk |2 = |vk |2 = 1/2. Therefore the
ground state of the k-state has probability 1/2 of being unoccupied and probability 1/2 of
being occupied by a pair of electrons with opposite k vector and spin. So on average there
+
is 1 electron. After applying the γ̂k1 operator, the state will change to |k ↑i. So there will
still be one electron.
Now what happens if we create an excitation far above the fermi sea? The ground
state wavefunction will have |uk | ≈ 1 and |vk | ≈ 0. So on average, almost zero electrons
+
in this state. But after applying γ̂k1 , there will be exactly one electron! So it seems that
particle number is not conserved. Observations such as these have given rise to conceptual
difficulties in interpreting the meaning of the condensate wavefunction of superfluidity, and
some have even gone so far as to propose replacing the free field creation and annihilation
operators with particle preserving operators that restore or remove a particle from the
condensate. While this works for the case of superfluidity, the mathematics are somewhat
awkward to adapt to the case superconductivity, due to the fact that the macroscopic
condensate state is replaced with the fermi sea.
 S.H. Lim

However, we point out here that the conceptual difficulty arises simply because we
were sloppy about defining the chemical potential u. In all treatments of the subject,
whether in superconductivity of superfluidity, u is taken to be a constant. Usually it is
calculated by first assuming there is no interaction term, and filling the energy levels up to
N electrons. While most authors would correctly recognize this as an approximation, it is
mostly assumed that u is a fixed value which we input into the Schrodinger equation. Then
we diagonalize the Schrodinger equation to get eigenvalues. Because the diagonlization of
the Schrodinger equation is merely a proxy for the actual computation that we want to
do, which is the minimization of the constrained energy functional, we lose track of the
true meaning of u. When we do the minimization, we are looking for stationary solutions
ψi along with two Lagrange multipliers, u and E. However, we assume u is a fixed value
which we have a right to set, whereas E is something we solve for. Calculus theory makes
no such assumption, u and E are both degrees of freedom that we should solve for. In other
words, for every Eigenenergy state that we find, Ei , there is a distinct chemical potential
ui , even for states of the same particle number. Therefore u(N, ψ) is really function of
N and ψ (and also T , but here we only consider T = 0). To be extra clear, we mean
that u of the ground state is not the same as u of the excited state. When we excite an
electron from the ground state, u is lowered ever so slightly, which modifies the uk and vk
coefficients (through eqn 29) of all the other states. This change is extremely small, but
just enough to keep the average particle count unchanged.
 S.H. Lim

Appendix A: From E Minimization to the Mean Field Hamilto-

nian
Lets study a toy problem of two different particles, A and B. To be precise, we mean that
A(x) is the wavefunction of the A particle and B(y) is the wavefunction of the B particle.
And to be absolutely clear, in this problem we will assume there is exactly one A particle
and exactly one B particle. So no multiple particles in the same state and also the total
state has a definite particle number (one of each).
The Hamiltonian with interaction is:4

H = H0 + Vint (35)
X X
H0 = ǫak â+
k âk + ǫbq b̂+
q b̂q (36)
k q
X
Vint = Vkq,q′k′ â+ +
k b̂q b̂q ′ âk ′ (37)
kk ′ ,qq ′

To make a mean field approximation of the Hamiltonian, we first imagine we already

know the solution of the Schrodinger equation, call it ψ(x, y). This gives us the ability to
define the expectation values of certain operators.

â+ +
k ′ âk = hψ|âk ′ âk |ψi (38)
hb̂+ +
q ′ b̂q i = hψ|b̂q ′ b̂q |ψi (39)

Nevermind for the moment that we don’t actually know what these expectation values are
given that our knowledge of the solution ψ remains fictitious. Now we can mathematically
define what it means for particle A and B to be uncorrelated. The correlation of an
operator product such as XY is hXY i − hXi hY i. We say X and Y are uncorrelated if
the correlation is zero. For our two particle Hamiltonian, this translates to:

hâ+ + + +
k âk ′ b̂q b̂q ′ i = âk âk ′ hb̂q b̂q ′ i (40)

But what does this imply about the wavefunction? If we truncate our Hilbert space to
the set of wavefunctions that can be written as ψ(x, y) = A(x)B(y), we see that the
4
Bruus, Flansberg
 S.H. Lim

above equation is automatically satisfied. This is why physicist say that the mean field
approximation is related to the assumption of uncorrelated particles.
Now we can replace the original interaction term (37) with a mean field interaction:
X
MF
Vint = Vkq,q′ k′ ( â+ + + + + +
k âk ′ b̂q b̂q ′ + âk âk ′ hb̂q b̂q ′ i − âk âk ′ hb̂q b̂q ′ i) (41)
kk ′ ,qq ′

MF
Solving the Schrodinger equation with Vint is a lot easier than with Vint and that is the
MF
motivation for making the mean field approximation. Both Vint and Vint are equivalent
in the sense that if we restrict ourselves to the truncated Hilbert space of uncorrelated
particles then:

MF
hψ1 |Vint |ψ2 i = hψ1 |Vint |ψ2 i (42)

for any pair of wavefunctions ψ1 (x, y) and ψ2 (x, y) that can be written in the form
A(x)B(y). A good way to see this is first expand A(x) and B(y) in some orthogonal
basis set, and then use that to prove that the above equality holds.
Next, we want to see how the mean field Hamiltonian pops out naturally from the
MF
variational problem. In other words, we want to derive Vint from first principles. As usual
we have the expectation energy hψ|H|ψi. But now we are making the approximation that
the solution ψ is of the form A(x)B(y). Next expand A and B in the basis set |ak i, |bq i
corresponding to the operators âk and b̂q respectively.
X
A= ck |ak i (43)
k
X
B= dq |bq i (44)
q

To expand out hψ|H|ψi, lets focus on the H0 and Vint terms separately:
X X
hψ|H0 |ψi = hB|Bi |ck |2 ǫak + hA|Ai |dq |2 ǫbq (45)
k q
X
hψ|Vint |ψi = Vkq,q′ k′ c∗k ck′ d∗q dq′ (46)
kk ′ ,qq ′

We want to find the minimum with respect to the coefficients ck and dq , but we also have
 S.H. Lim

to remember the normalization constraints:

X
hA|Ai = 1 =⇒ λA ( |c2k | − 1) = 0 (47)
k
X
hB|Bi = 1 =⇒ λB ( |d2q | − 1) = 0 (48)
q

Now we add eqns 43-46 and minimize the whole thing. We first do a variation with respect
to c∗k :
∂ X X
→ hB|Bi ck ǫak + ck |dq |2 ǫbq + hB|Bi Vkq,q′ k′ ck′ d∗q dq′ + λA ck = 0 (49)
∂ck ∗ q k ′ ,qq ′

Based on the Variational principle, this should be the Schrodinger equation, but right
now its in the form of a matrix equation. To get it into the form we are more familiar
with (Ĥ |ψi = E |ψi), we have to massage the equations a bit. First notice that the above
equation consist purely of c-numbers, ie. there are no operators or state kets. Therefore we
can multiply the above equation with any state ket. Lets say we choose |ak i. Also notice
that d∗q dq′ = hB|b̂+
q b̂q ′ |Bi (should be emphasized that this is the key step!). Therefore, we

can rewrite the above equation as:

!
X X
ck ǫak + Vkq,q′k′ hb̂+
q b̂q ′ i ck ′ |ak i = ( |dq |2 ǫbq − λA )ck |ak i (50)
k ′ ,qq ′ q

|dq |2 ǫbk − λA ) on the RHS

P
In the above we had also substituted hB|Bi = 1. The term ( q

is just some undetermined number, so lets call it EA . Also, notice that:

ck |ak i = â+
k âk |Ai (51)
ck′ |ak i = â+
k âk ′ |Ai (52)

Substituting into (50) we get:

X
ǫak â+
k âk |Ai + Vkq,q′k′ hb̂+ + +
q b̂q ′ i âk âk ′ |Ai = EA âk âk |Ai (53)
k ′ ,qq ′

And summing over k, we finally get:

X X
ǫak â+
k âk |Ai + Vkq,q′k′ hb̂+ +
q b̂q ′ i âk âk ′ |Ai = EA |Ai (54)
k kq,q ′ k ′
 S.H. Lim

â+
P
where the RHS is due to I = k k âk . We can see that the interaction part is of the
MF
form that we would expect for Vint , (see eqn 41). Redoing the analysis for the B particle
MF
will give us the other piece of Vint , which is â+ +
k âk ′ b̂q b̂q ′ . And finally if we want to

write down the full Schrodinger equation we have to subtract a term like â+ +
k âk ′ hb̂q b̂q ′ i

to avoid double counting the interaction energy.

 S.H. Lim

Appendix B: Physicists’ Definition of Uncorrelated Particles

It is often said the BCS theory is a mean field theory, where individual particles are
correlated, and particle pairs are uncorrelated (taken as a pair, they behave as uncorrelated
particles). What does this mean exactly? Here we will establish a mathematical definition
of what physicist mean by uncorrelated particles.
Mathematically, given a function f (x, y) of two variables x and y, we say they are
uncorrelated if the function can be written as:

f (x, y) = A(x)B(y) (55)

However, a wavefunction of two identical particles can never be written in the form
above (unless of course A and B are the same), due to the symmetrization requirement:

ψ(x, y) = P± (A(x)B(y)) = A(x)B(y) ± A(y)B(x) (56)

where P± () is the symmetrization operator for bosons/fermions. So in physics the defini-

tion of uncorrelated is somewhat more loose than the mathematical definition. We call the
particles uncorrelated if the term inside the P± () operator can be written as A(x)B(y).
In our lingo, we can say they are uncorrelated apart from symmetrization.
Given a wavefunction written in second quantized form, how do we determine if the
particles are uncorrelated? A general two particle wavefunction in second quantization is:
X
|ψi = ck,q â+ +
k âq |0i (57)
k,q

Note that there is no need to symmetrize the wavefunction once it is written in second
quantized form because the creation operators obey commutation/anticommutation rela-
tions. Therefore, if we can factor the above expression:
X X X
ck,q â+ +
k âq |0i = dk â+
k eq â+
q |0i (58)
k,q k q

then the particles are uncorrelated. Whether or not we can do this factorization will
depend on the coefficients ck,q . A simple example would be having one particle in the
 S.H. Lim

state |1i + |2i and a second particle also in the same state (assume these are bosons).
Then the two particle wavefunction would be:
  
1 1 1 1 1 1 1 1
|ψi = √ |1i + √ |2i √ |1i + √ |2i = |1i |1i + |1i |2i + |2i |1i + |2i |2i
2 2 2 2 2 2 2 2
(59)

The wavefunction above is uncorrelated. However, the wavefunction below is not:

1 1
|ψi = √ |1i |1i + √ |2i |2i (60)
2 2
We call such a state an entangled state.
Next we can see what it means for the BCS wavefunction to have uncorrelated pairs.
First of all the BCS wavefunction is a superposition of all possible particle number wave-
functions.

ψBCS = ψ(x1 ) + ψ(x1 , x2 ) + ψ(x1 , x2 , x3 ) + ... (61)

Each of these wavefunctions lives in its own sector of Fock space. Therefore when we
talk about correlation, in only makes sense to analyze these wavefunctions separately.
For starters, lets simplify the problem greatly by assuming the superconductor is in an ex-
tremely tiny box, such that there are only three pair states allowed: (k1, −k1), (k2, −k2), (k3, −k3).
Also we pretend there is no spin. The BCS wavefunction is then just:

+ + +
|ψBCS i = (u1 + v1 P̂k1 )(u2 + v2 P̂k2 )(u3 + v3 P̂k3 ) |0i (62)

Taken as pairs, the claim is they are uncorrelated. In other words, the motion of the
pair of electrons (k1, −k1), when viewed as a single particle (think of a molecule if that
helps), is uncorrelated to the motion of the (k2, −k2) pair and (k3, −k3) pair. If we
expand out the above expression, we see that the BCS state is a superposition of no pairs,
a 1 pair wavefunction, 2 pair wavefunction and 3 pair wavefunction. For the case of 0
and 1 pair, obviously there is no sense in talking about correlation. For the 3 pair state,
+ + +
the wavefunction is just v1 v2 v3 P̂k1 P̂k2P̂k3 |0i. This is just 3 single particle wavefunctions
multiplied together, so they are uncorrelated. For the two particle wavefunction, we have:

+ + + + + +
(u1 v2 v3 P̂k2 P̂k3 + u2 v1 v3 P̂k1 P̂k3 + u3 v1 v2 P̂k1 P̂k2 ) |0i (63)
 S.H. Lim

Are the two particles entangled or not? The answer is no, because it can be factored into
two single particle wavefunctions:
  
v1 + v2 + v3 + v1 + v2 + v3 +
= u1 u2 u3 P̂ + P̂ + P̂ P̂ + P̂ + P̂ |0i (64)
u1 k1 u2 k2 u3 k3 u1 k1 u2 k2 u3 k3

Therefore the pairs are uncorrelated.

What about the individual electrons themselves? The electrons of opposite k and spin
are all entangle. Just look at the single pair wavefunction in the example above:

(uk2 uk3 vk1 â+ + + + + +

k1 â−k1 + uk1 uk3 vk2 âk2 â−k1 + uk1 uk2 vk3 âk3 â−k3 ) |0i (65)

It is impossible to factor this into two single electron wavefunctions. Therefore in the BCS
theory, pairs of electrons are correlated, but when a pair is viewed as a single particle,
it is uncorrelated to all the other pairs. That is why we are able to write a mean field
Hamiltonian for the BCS wavefunction.

Superfluid Hamiltonian
lorem ipsum