You are on page 1of 80

Xiao-Liang Qi

Lecture notes for Phys370:


Theory of Many-Particle Systems
March 24, 2011
Contents

1 From particles to elds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1


1.1 Review of single-particle quantum mechanics . . . . . . . . . . . . . . . . 1
1.1.1 Single harmonic oscillator . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Two coupled Harmonic oscillators . . . . . . . . . . . . . . . . . . . 2
1.1.3 Thermodynamic properties: a comparison between
quantum and classical systems . . . . . . . . . . . . . . . . . . . . . . 2
1.2 From coupled harmonic oscillators to quantum eld theory . . . 3
1.2.1 One-d chain and continuum limit . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2.3 Density distribution and X-ray diraction . . . . . . . . . . . . 5
1.2.4 Specic heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.5 Symmetry breaking and Goldstone theorem . . . . . . . . . . . 8
1.3 Lagrangian Formulism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.1 Lagrangian Formulism of Field Theory . . . . . . . . . . . . . . . 8
1.3.2 Noethers Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.4 Maxwell theory and the quantization of photons . . . . . . . . . . . . . 11
1.4.1 Classical electromagnetism from symmetry principle . . . 11
1.4.2 Canonical quantization of the Maxwell theory . . . . . . . . . 14
1.4.3 Casimir eect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Second quantization and its applications . . . . . . . . . . . . . . . . . . . 17


2.1 Second quantization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.1.1 Second quantization for Bosons . . . . . . . . . . . . . . . . . . . . . 17
2.1.2 Second quantization for Fermions . . . . . . . . . . . . . . . . . . . . 20
2.2 Applications of second quantization . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.1 Free fermion gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Tight-binding models and graphene as an example . . . . . 24
2.2.3 Interaction eect in tight-binding systems . . . . . . . . . . . . 27
2.2.4 Quantum ferromagnetism and anti-ferromagnetism . . . . 29
2.2.5 Transverse-eld Ising model . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.6 Bosonization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4 Contents

3 Path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1 Single-particle path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1.1 Introduction to single-particle path integral . . . . . . . . . . . 41
3.1.2 Imaginary time path integral . . . . . . . . . . . . . . . . . . . . . . . 43
3.1.3 Harmonic oscillator and stationary phase approximation 44
3.1.4 Double-well problem and instantons . . . . . . . . . . . . . . . . . 46
3.1.5 Linear response of a harmonic oscillator . . . . . . . . . . . . . . 48
3.2 Functional path integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.1 Boson coherent state path integral . . . . . . . . . . . . . . . . . . . 53
3.2.2 Fermion coherent state path integral . . . . . . . . . . . . . . . . . 56

4 Perturbation Theory and Beyond . . . . . . . . . . . . . . . . . . . . . . . . . . 59


4.1 Discussion on the convergence of perturbation theory . . . . . . . . 59
4.2 More general cases and Feyman diagram . . . . . . . . . . . . . . . . . . . 61
4.3 Perturbative approach to interacting electron gas . . . . . . . . . . . . 65
4.3.1 Ground state energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.2 Perturbative theory of the dielectric response . . . . . . . . . 68
4.3.3 Beyond perturbation theory: RPA approximation . . . . . . 72
4.3.4 Eective eld theory approach and the plasmon mode . . 73
1
From particles to fields

1.1 Review of single-particle quantum mechanics


1.1.1 Single harmonic oscillator

We would like to review the harmonic oscillator because it will become a


foundation of a lot of discussions in many-body systems. Harmonic oscillator
has the Hamiltonian

p2 1
H= + m 2 x2 (1.1)
2m 2
The momentum operator p = i x
and the Schordinger equation can be
solved as an ordinary dierential equation problem. However it is more helpful
to solve this problem by dening
( )
1 p
a= mx + i (1.2)
2 m

It can be checked that


[ ]
[p, x] = i a, a = 1
( )
1
H = a a+
2
[
]
a, a a = a
[ ]
a , a a = a (1.3)

(Heisenburg algebra) These commutation relations tell us some important



information about the eigenvalues of the operator
( a a. ) If a state | is eigen-
state of a a so that a a | = |, then a a a | = (a a + 1) (a |)

a a (a |) = ( 1) a |. Thus for any eigenstate with eigenvalue , there


2 1 From particles to fields

exists an eigenvalue 1, unless a | = 0. Because a a is nonnegative, we


obtain the conclusion that Z, 0. (Otherwise, there will be innite
number of states of eigenvalue of a a negative.) Thus all states form towers
connected by a and a . The bottom of the tower are ground states annihi-
lated by a.
Going back to the simple harmonic oscillator case, a |0 = 0 is written in
wave-function form as
( )
1 x
mx + (x) = 0 (1.4)
2 m
which has only one solution:

0 (x) = e 2 mx (unormalized)
1 2
(1.5)

Thus there is only one tower in the Hilbert space, with the basis |n =
n
1 a |0.
n!

1.1.2 Two coupled Harmonic oscillators

Why harmonic oscillator is special compared to other quantum mechanical


problems? To see that one can study two coupled harmonic oscillators (with
equal mass for simplicity).

p21 1 p2 1
H= + m12 x21 + 2 + m22 x22 + kx1 x2
2m 2 ( )2m
( )2 ( )( )
1 10 p1 1 m12 k x1
= (p1 , p2 ) + (x1 , x2 ) (1.6)
2m 01 p2 2 k m22 x2

The 2 2 matrix can be diagonalized by orthorgonal rotation. As an example


we consider the special case 1 = 2 = , in which case the diagonalization
is explicit
1 ( 2 ) 1( ) 1( )
H= p+ + p2 + m 2 + k x2+ + m 2 k x2 (1.7)
2m 2 2
1 1
x = (x1 x2 ) , p = (p1 p2 )
2 2
From this example we see that the special property of harmonic oscillator is
that coupled harmonic oscillators are equivalent to decoupled har-
monic oscillators, by coordinate transformation.

1.1.3 Thermodynamic properties: a comparison between


quantum and classical systems

Even for a single harmonic oscillator, we can calculate the thermodynamic


properties such as heat capacity. The calculation is more physically relevant
1.2 From coupled harmonic oscillators to quantum field theory 3

for a many-body system but the calculation is the same. The specic heat is
also one probe to see the dierence between quantum and classical systems.
If we consider a classical harmonic oscillator, the Hamiltonian is a function
of p and x which are real-number valued variables. The Boltzman partition
function is
( )
p2
2m + 12 m 2 x2 2
H(p,x)
Z = dpdxe = dpdxe = (1.8)

The energy
log Z 1 U
U = H = = =T C= =1 (1.9)
T
This is the equipartition theorem in classical statistical mechanics. For N
harmonic oscillators C = N and U = N T , independent from any detail of the
coupling.
On comparison, the quantum system has states labeled integer n = 0, 1, 2...
which have energy (n + 1/2). Thus the partition function is

e/2 1
Z= e(n+1/2) =
= (1.10)
n=0
1e 2 sinh
2
2 log Z 2 2
C = 2 = ( )2 (1.11)
2
4 sinh 2

At high temperature the result is equivalent asymptotically to classical


result. At low temperature the eect of discrete energy spectrum appears.

1.2 From coupled harmonic oscillators to quantum field


theory
1.2.1 One-d chain and continuum limit

Consider a one-dimensional chain of atoms.


p2
H= i
+ V (xi xi+1 ) (1.12)
i
2m i

As an approximation, we expand the interaction energy and get the harmonic


oscillator form
p2 1 2
H= i
+ ks (xi xi+1 a) (1.13)
i
2m i
2

Dene xi ia = i the displacement around the equilibrium position, we have


4 1 From particles to fields
p2 1 2
H= i
+ ks (i i+1 ) (1.14)
i
2m i
2

Similar to the two-oscillator case the Hamiltonian can be diagonalized by


unitary rotation

2 1 0 ... 0 1
1 2 1 0 ... 0

1 T 1
T

H= p Ip + ks (1.15)
2m 2 0 ... ... 0

i
0 ... 0 1 2 1
1 0 .. 0 1 2

The diagonalization can be done by Fourier transformation directly, but


it is even simpler to take a continuum limit

1
i a(r), pi a(r), dr , i+1 i ar (r)(1.16)
i
a

The continuum Hamiltonian


[ ]
(r)2 ks a2 2
H = dr + (r (r)) (1.17)
2m 2

Notice that ks a2 needs to remain nite in the limit a 0.


(Btw, this is actually the simplest string theory!)

1.2.2 Quantization

The commutation relation in continuum limit is given by


1
[pi , j ] = iij [(x), (y)] = i ij = i(x y) (1.18)
a
The Hamiltonian can be diagonalized by Fourier transformation
1 ikr
(r) = e k ,
L k
1 ikr
(r) = e k , (1.19)
L k

(This is for a nite chain with length L. x [0, L] with periodic boundary
condition.) It can be checked that

[k , k ] = ikk (1.20)

The Hamiltonian is transformed to


1.2 From coupled harmonic oscillators to quantum field theory 5
[ k k ks a2 2
]
H= + k k k (1.21)
2m 2
k

The Hamiltonian is diagonalized to independent harmonic oscillators with


frequency m 2 = ks a2 k 2 = ks a2 /mk. Similar to single harmonic
oscillator case, we can dene
( )
1 k
ak = mk k + i (1.22)
2 mk
( 1
)
H= k ak ak + (1.23)
2
k
( )
Since k = 1
2mk
ak + ak , we have

1 1 ( )
(r) = ak + ak eikr (1.24)
L k 2mk

Notice that k = 0 is NOT a harmonic oscillator because k = 0. Instead


it is a free particle with only 02 /2m term.
The energy-momentum relation k = ks a2 /mk = vk is called disper-
sion relation. k is the energy of a single-particle excitation with momen-
tum k, i.e. the energy for the creation of one mode with momentum k. The
name dispersion relation comes from the fact that light will have disper-
sion if light with dierent wavevector have dierent velocity. Thus a linear
dispersion means no dispersion.

1.2.3 Density distribution and X-ray diraction

With the quantum eld theory description of the one-dimensional chain ob-
tained above, one can study the physical properties of the chain. As an ex-
ample, we study the density distribution of atoms which is related to X-ray
scattering experiments. (This is also related to the question we discussed on
the class last week, about measuring m and ks separately.) In X-ray diraction
experiment, the signal observed is determined by the fourier transformation
of the electron density. For a single crystal, the electron density distribution
is periodic in the lattice constant a. Here we study the eect of the atom
oscillation (phonon) and assume the electrons follow the nuclei in the motion.
To determine the density distribution we temporarily return to the lattice
formulation (1.14). The atom density n(r) is determined by the average value
of the following operator:


N
n(r) = (na + n r)
n=1
n(r) = n(r) (1.25)
6 1 From particles to fields

The X-ray diraction signal can be obtained from the Fourier transformation
N
1 iqr 1
F (q) = dre n(r) = dreiqr (na + n r)
L i=1
L

1 iq(na+n )
N
= e (1.26)
L n=1

Due to translation symmetry eiqn = f (q) is independent from n, so that

N 2m
F (q) = f (q) , for q = , mZ (1.27)
L a

For q belongs to the reciprocal lattice, F (q) is determined by

q2 2
f (q) = eiqn 1 iq n n + ...
2
q2
e 2 n
2
(1.28)

(Actually the equality holds rigorously, which is a special property of free eld
theory. Proof of this wont be shown here.)
Before going to calculation of 2n we would like to say
that this expression
is generic. The information about the uctuation 2n can be obtained by
studying the ratio of two peaks
[ ]
f (q = 2/a) 2 2
= exp 2 2n (1.29)
f (q = 0) a

To calculate 2n we take n = 0, go to the continuum limit and use Eq.
(1.24):
a 1 ( )( )
2n=0 = a (r = 0)2 = ak + ak ak + ak
L 2mk 2mk
k,k

a 1 a 1 a L
= = dk = log (1.30)
L 2mk 2 2mvk 4mv a
k

Thus
[ ]
q2 a L
f (q) = exp log (1.31)
8mv a

Notice that there is a log divergence in (r)2 , which is specic for one-
dimensional systems. The same calculation as above can be studied in higher
dimensions. For example in 2d there will be two displacement elds x,y (r)
and we have
1.2 From coupled harmonic oscillators to quantum field theory 7
2
f (q) e 4 x +y
q 2 2

2 a2 1
x + 2y = a2 2x + 2y (r = 0) = 2
L 2mvk
k
2 2
( )
a 1 a 2 2 a
= 2
2
d k = (1.32)
(2) 2mvk 4mv a L 2mv
We see that the uctuation of atom position is determined by the geometrical
average of lattice constant a and
h h
= = (1.33)
mv mks a2
Physically is the wavelength of the atom corresponding to the energy ks a2 .
Further remarks. Compare the results in 1d and 2d we see that the
uctuation range = 2n is diverging for a large system L in 1d,
but remains nite at 2d. It can be checked that in higher than 2d its always
nite. And interesting question is what happens at nite temperature. The

average value in Eq. (1.30) involves terms such as ak ak which is temperature

dependent. It can be shown that at nite temperature, even for 2d = 2n
is divergent. This is indicating that (1) in 1d there is no crystal; (2) in 2d there
is crystal at T = 0 but it melts at any nite temperature; (3) in 3d and higher
dimensions there is real well-dened crystal.

1.2.4 Specic heat

The dierence between the quantum and classical theories can be seen from
specic heat. Simply applying the result of single harmonic oscillator we get

Cclassical = 1 (1.34)
k

which gives innity in the continuum theory. If we go back to the discrete


theory, the result is C = N (number of sites) so that the specic heat cV =
C/(N a) = 1/a. Historically this is the ultraviolet catastrophe in blackbody
radiation, the solution of which by Planck is the rst achievement of quantum
mechanics.
The quantum specic heat
( k /2 )2
Cquantum =
sinh (k /2)
k
( )2
Cquantum dk vk
cV = =
L 2 2 sinh (vk/2)
( )2
1 t
= dt T (1.35)
v sinh t
8 1 From particles to fields
L

(Notice k = 2 dk) High momentum contribution is suppressed.
How to cross over to the classical behavior? Cut-o k = 2/a needs to be
considered. We need to go back to the discrete system. Around temperature
T 2v/a, cV has a cross-over to classical behavior.

1.2.5 Symmetry breaking and Goldstone theorem

Although the one-d chain model is very simple, it provides a clear example
of spontaneous symmetry breaking and Goldstone theorem. The dispersion
relation k = vk is linear, and gapless, which means a harmonic oscillator
mode can be created with arbitrarily low energy (in large systems). Why is
that? k is a continuous function of k. k is gapless because k has to go
to 0 when k 0. In other words, the excitation spectrum is gapless because
the mode at k = 0 does not cost any energy. The k = 0 mode k=0 is the
translation of the whole chain. If we take k=0 = 0 for all k = 0, then

1 1
(r) = k eikr = k=0 (1.36)
L k L

is a constant independent from x. Thus the fact that such a translation does
not cost energy is actually nothing but translation symmetry. By choosing
an equilibrium position of each atom, the translation symmetry is broken
spontaneously. This is an example that spontaneous symmetry breaking leads
to gapless excitations of the system. In general, gapless excitations always
exist if a continuous symmetry is broken spontaneously. This is also the rst
of several mechanisms which can protect gapless excitations in a quantum
many-body system. (In general, gaplessness always needs a reason.)
I would like to mention that at this moment we cannot see from the the-
ory presented above the reason why such a spontaneous symmetry breaking
occurs. Why will the atoms pick a specic location which breaks the trans-
lation symmetry? If some parameters in the Hamiltonian are changed, they
may form a liquid instead, which preserves the translation symmetry. Such
questions need to be answered in the interacting theory which we will discuss
in later part of this course.

1.3 Lagrangian Formulism


1.3.1 Lagrangian Formulism of Field Theory

The quantization of one-dimensional chain discussed above is based on Hamil-


tonian formulism. The advantage of Hamiltonian formulism is that the quan-
tum mechanical properties of the system are well-dened once we identify
momentum and coordinate operators. In this subsection we will introduce an
1.3 Lagrangian Formulism 9

alternative formulism based on Lagrangian. The advantage of Lagrangian for-


mulism is that the symmetry properties are more transparent, and the it is
suitable for the path integral quantization we will discuss later.
p2
For a single particle with Hamiltonian H = 2m + V (x) we know that the
Lagrangian is given by
1
L(x, x) = mx2 V (x) (1.37)
2
and the action is dened by
[ ]
1
S [x(t)] = L(x, x)dt = dt mx V (x)
2
(1.38)
2
S[x(t)] is a functionala map from functions x(t) to real number R. The
equation of motion is given by minimizing S[x(t)]. Consider an arbitrary small
variation x(t) x(t) + x(t), we have
( )
L L dt
S[x(t) + x(t)] = L(x + x, x + x)dt L(x, x) + x + x
x x
( ( ))
L d L
= S[x(t)] + dtx(t) (1.39)
x dt x
S = S[x(t) + x(t)] S[x(t)] = 0 leads to the condition
( )
L d L
=0 (1.40)
x dt x
which is the Lagrange equation of motion.
The Lagrangian and Hamiltonian are related by the equation
L
p=
x
H(p, r) = px L (1.41)

It should be noticed that there is a change of variable between p and x by using


the rst equation. In the second equation, x on the right side is considered as
a function of p and x.
For more general multiple dimensional system with Lagrangian L(qi , qi )
the Lagrange equation is
( )
L d L
=0 (1.42)
qi dt qi
Now we take the one-d chain and study the Lagrangian in the continuum
limit. The Lagrangian of the one-d chain with Hamiltonian (1.14) is
1 2 1 2
L(n , n ) = mn ks (n+1 n ) (1.43)
n
2 2 n
10 1 From particles to fields

In the continuum limit n a(rn ) we have
[ ]
1 2 1 2
L = dr m (t ) ks a (x )
2
(1.44)
2 2

The integrand is named as Lagrangian density


1 2 1 2 m
L (, ) (r) = m (t ) ks a2 (x ) (r) = ( g ) (1.45)
2 2 2
with g = diag[1, v 2 ] the Lorenz metric. Notice that L (r) is a function of
(r) and its gradient. The Lagrange equation can be expressed by Lagrangian
density by minimizing the action

S[(r, t)] = dtL = dtdxL (1.46)

The result is
( ) ( )
L L L
=0 (1.47)
t t r r

or equivalently
( )
L L
=0 (1.48)

1.3.2 Noethers Theorem

As an important application of the Lagrangian and variational principle, we in-


troduce Noethers theorem on the relation between continuous symmetry and
conservation law. As a simplest example, consider the Lagrangian density of
the one-d chain (1.45). As discussed above, the translation symmetry requires
L(, ) = L( + , ) with any R. Consequently L(, ) = L( )
is independent from . Thus the Lagrange equation (1.48) is simplied to
( )
L
= m g = 0 (1.49)

If we dene
L
j = (1.50)

we have j = 0 so that j is a conserved current. This is the simplest case of


Noethers theorem, where a symmetry combined with Lagrange equation gives
a conservation law. More generally, consider a Lagrangian density L(a , a )
with multiple-component eld a . If there is a symmetry transformation a
1.4 Maxwell theory and the quantization of photons 11

a + Fa (), (with innitesimal parameter ) the Lagrangian density should


be preserved in this transformation

L(a , a ) = L (a + Fa (), a + Fa ())


L L
Fa () + Fa () = 0 (1.51)
a a

On the other hand we have the Lagrange equation (1.48). Combine (1.48)
with the equation above we get
( ) ( )
L L L
Fa () + Fa () = Fa () = 0 (1.52)
a a a

More generally if the coordinates are also changed in the symmetry transfor-
mation (such as (x) (x + C)) the expression is more complicated, but
can be derived in the same way.

1.4 Maxwell theory and the quantization of photons

1.4.1 Classical electromagnetism from symmetry principle

In this section we generalize the quantization scheme above to photons. Clas-


sical electromagnetism (in vacuum) is described by Maxwells equations

E = 0, B t E = 0
B = 0, E + t B = 0 (1.53)

By introducing gauge potential A = (A0 , A) and B = A, E = A0


t A, the Maxwell equations can be written in the covariant form

F = 0 (1.54)

with F = A A and = /x , = g . g = diag[1, 1, 1, 1]


is the Lorenz metric. The Lorenz transformation A = L A preserves the
metric g in the sense of

A g A = A A g L L g = g (1.55)

In the expression F = A A there is a redundancy because


A = A + and A corresponds to the same F , for any arbitrary
function . This is called gauge symmetry.
Although from classical electromagnetism we know the Hamiltonian (i.e.,
energy functional) of electromagnetic eld, the quantization of this theory
is not as straightforward as the one-d chain, because its not obvious which
12 1 From particles to fields

elds should be taken as canonical momentum and coordinate. Now we wan-


t to take a dierent point of view and ask the following question: What is
the possible action of a vector eld A which preserves Lorentz symmetry
and gauge symmetry? The Lagrange formulism is specially suitable for such
an approach because the action is always invariant in a symmetry operation,
while the Hamiltonian is not always invariant. The Lagrangian density satis-
fying Lorenz invariance can be expanded in powers of A :

L = a A + bA A + c A A + d( A )2 + ... (1.56)

The terms with more than two derivatives or more than two A are neglected.
After considering gauge invariance L(A ) = L(A + ) we see a = 0, b =
0, c + d = 0 so that
c
L= F F (1.57)
2
Its straightforward to check that the Lagrange equation of motion leads
to Maxwells equations (1.54) correctly. Thus we see that the two symmetries
completely determine the form of Maxwells equations to the leading order. If
there is a material eld with a conserved current j , one more term f j A is
possible, which leads to
c
L= F F + f A j (1.58)
2
The Lagrange equation is j = (2c/f ) F , which recovers the Maxwell
equations with source correctly, if we rescale the elds A 2c
1
A , J

2c
f J . The action becomes

1
L = F F + A j (1.59)
4
It seems that the action is completely unique without any parameter. How-
ever, there is a hidden parameter due to charge quantization. For exam-
ple, for
electrons the charge density = j 0 integrated over any area gives
Q = d rj = en, n Z. During the rescaling the unit of charge needs
3 0

to be rescaled. To see that explicitly we can introduce j = eJ with J the


number current of electrons. After a rescaling eA A , the Maxwell theory
can be written as
1
L = F F + A J (1.60)
4e2
where we see explicitly that e2 plays the role of a coupling constant. When
recovering the full units, the dimensionless coupling constant e2 is actually
e2
the ne structure constant = hc (in CGS unit). In the following we will use
the standard Lagrangian (1.59) for simplicity.
1.4 Maxwell theory and the quantization of photons 13

Its interesting to notice that the coupling j A preserves gauge invariance


only if j = 0, which is exactly the charge conservation condition. The gauge
elds are deeply related to conserved currents. The coupling A j between
gauge eld and conserved current is called minimal coupling.
Further remarks:
Is there a Noether theorem for gauge symmetry? Gauge symmetry re-
quires the Lagrangian density to satisfy L(A , A ) = L(A + , A +
). Naively, Noether theorem can be obtained. For example, if we
take = x1 so that = 1 , we obtain a translation symmetry
A A + 1 . The dierence from the case of global symmetry is that
there are innite number of symmetries dened in this way, which cor-
respond to innite number of conservation laws. In global symmetry the
innitesimal transformation is a number, while in gauge symmetry is
an arbitrary function. Expand the Lagrangian density we obtain
L L
+ = 0 (1.61)
A A

for arbitrary function (x, t). By taking = x and = x x we obtain


L
=0
A
L L
and + =0 (1.62)
A A

respectively. These are local constraints satised at each point, rather


than conservation laws.
There is another term

LME = F F (1.63)

which almost preserves all the symmetries. The only symmetry it breaks
is the discrete subgroup of Lorenz groupthe spatial inversion. The spatial
inversion is determined by A0 A0 , A A, 0 0 , . This
term corresponds to an isotropic magneto-electric eect and it has the
interesting property of topological invariance. Recently this term plays an
important role in the topological eld theory of topological insulators. For
an introduction see e.g. Qi and Zhang, Phys Today Jan 2010.
If we study 2+1 rather than 3+1 dimensions, there is one additional term
which preserves the special Lorenz group but not the spatial inversion,
which is the Chern-Simons term
n
LCS = A A (1.64)
4
This term plays an essential role in the quantum Hall eect and fractional
statistics in 2+1 dimensions. It can be easily seen that this Lagrangian
14 1 From particles to fields

density is not gauge invariant. However, the action is gauge invariant be-
cause the Lagrangian density changes by a total derivative under gauge
transformation. This is an example that the invariance of action and the
invariance of Lagrangian density are not equivalent. The Chern-Simons
term (and its counter-part in non-Abelian gauge theories) plays an essen-
tial role in the research of quantum Hall eect, the rst known topological
state of matter in Nature.

1.4.2 Canonical quantization of the Maxwell theory

Now we study the quantization of the Maxwell theory in vacuum without


source term. The Lagrangian density can be written as
1 1 1 1 2 1 2
L = F F = F0i F 0i Fij F ij = (t A A0 ) ( A)
(1.65)
4 2 4 2 2
Thus we see A0 does not have a canonical momentum, and A has the canonical
momentum
L
i = = t Ai i A0 (1.66)
t Ai
Thus the Hamiltonian
[ ]
( ) 1 1 2
H = d3 x i t Ai L = d3 x 2 + ( A) A0 (1.67)
2 2
A0 only appears in the last term, and it does not have a conjugate momentum.
Thus the equation of motion of A0 is just the classical one
H
= = 0 (1.68)
A0
If we take the fourier transform into momentum space, we have
[1 1
]
H= k k + (k Ak ) (k Ak ) (1.69)
2 2
k

with the constraint

k k = 0 (1.70)

Dene the orthogonal basis e1 , e2 , e3 so that e3 = k = k/ |k| and e1,2 are


the other two orthogonal
vectors, we can decompose the vectors A and on
3
these directions as Ak = s=1 es Ask . Thus
[1 1 2
]
H= s,k s,k + k As,k As,k (1.71)
s=1,2
2 2
k
3,k 0
1.4 Maxwell theory and the quantization of photons 15

Thus the Hamiltonian becomes two branches of decoupled harmonic oscilla-


tors, both with the dispersion k = |k|. The third branch (which is called lon-
gitudinal mode disappears because of the constraint on 3,k = k k = 0.
(The quantization of the Maxwell theory in a wave-guide can be found in
the text book, which I wont repeat here.)

1.4.3 Casimir eect

In Sec. 1.2.4 we have discussed an important dierence between quantum


and classical theory, which is the specic heat at low temperature. Another
important signature of the quantum theory is the Casimir eect: Two parallel
metal plates have a interaction. The energy of the system with two innite
large parallel plates in a distance d is proportional to the area A of the plates.
The energy per area E/A is only determined by d. Since there is no other
length scales in this system, dimensional analysis shows
E hc
3 (1.72)
A d
The result obtained by Casimir is

E 2 hc
= (1.73)
A 720d3
so that the interaction is attractive.
Without going to detail of the derivation, it should be noticed that the
coupling constant = e2 /hc does not enter the result, which is due to the
scaling invariance of the problem. Rescaling A A with any keeps
invariant both the photon dispersion and also the boundary condition set by
the metallic plates, so that the Casimir force remains the same. When the two
plates are very closed to each other, the eect of the metallic plate cannot be
treated as a simple boundary condition, and the higher order corrections such
as the Van der Waals force between plates has to be considered. The higher
order terms are in general dependent on the ne structure constant . Casimir
force can also be repulsive if the media is not vacuum. For an example, see
Munday et al, Nature 457 170 (2009).
2
Second quantization and its applications

2.1 Second quantization


In Chapter 1 we have actually already used a lot of concepts in second quan-
tization, but we havent introduced the systematic framework of it. In short,
second quantization is a formulism to describe a generic quantum many-body
system using annihilation and creation operators a, a , and their fermionic
analogue.

2.1.1 Second quantization for Bosons

In Chapter 1 we see that in quantum eld theory approach the fundamental


(bosonic) particles such as photons can be described in the same way as the
energy quanta in harmonic oscillators. From the discussion in Chapter 1 it
is clear that such a formulism applies to a generic quantum eld theory as
long as we know the Hamiltonian H = H [(x), (x)]. Harmonic oscillators
can be dened from linear combinations of momentum eld (x) and coordi-
nation eld (x). On the other hand, in the particle picture we say that the
particles described by such quantum eld theory, such as photons, are Bosons,
i.e., identical particles with wavefunction (x1 , x2 , ...) symmetric under per-
mutation of any two coordinates xi , xj . To see clearly the connection between
these two pictures, we can consider the simple one-dimensional chain again.
For the Hamiltonian (1.14), instead of dening harmonic oscillator operators
in the momentum space (Eq. (1.22)), we can also dene harmonic oscillator
operators in real space. The Hamiltonian can be written as
p2
H= i
+ ks 2i ks i i+1 (2.1)
i
2m i i

whichcan be considered as harmonic oscillators at sites i = 1, ..., N with


2ks
= m with the last term as their coupling. Annihilation and creation
18 2 Second quantization and its applications

operators can be dened for each site in the same way as in single harmonic
oscillator case
( )
1 pi 2ks
ai = mi + i , = (2.2)
2 m m
The harmonic oscillator at i site has the Hilbert space expanded by the eigen-
states of ai ai which are labeled as |ni i , ni = 0, 1, 2, ..., so that a basis in the
many-body Hilbert space is


N
|{ni } = |ni i (2.3)
i=1

If we interprete ni = ai ai as the particle number at i-th site, such a basis


corresponds to the occupation number representation. When viewing the sys-
tem as a gas of many such particles, another representation is the particle
coordinate representation in which the wave function ({xi }) is a function of
particle coordinates xi . To see the relation between these two representations
we can take the simple case of two particles at i and j sites. This corresponds
to the state |00..010..010.. in occupation number representation, where the
two 1s appear at i and j sites. In a state | the possibility amplitude of such
as state is given by (i, j) = 00..010..010..| . On comparison, in the parti-
cle coordinate representation this state can be described as the rst particle
at i site and the second particle at j site, or the other way around. Thus the
wavefunction is given by

(x1 = i, x2 = j) = (x1 = j, x2 = i) = 00..010..010..| (2.4)

By this comparison we see that there is a redundancy in the particle coordi-


nate representation. To remove the redundancy we require the wavefunction
({xi }) to be symmetric under permutation of any two particle coordinates
xi and xj . This is exactly the requirement of these particles having bosonic
statistics. In summary, from the discussion above we have seen why identical
bosonic particles have the same Hilbert space as harmonic oscillators, and
thus can be described by the harmonic oscillator operators ai , ai .
Besides removing the redundancy, another important advantage of occu-
pation number representation is that states with dierent total number of
particles can be described on equal footing. The Hilbert space expanded by
|{ni } contains the subspaces with dierent total particle number N = i ni .

Such as Hilbert space is named as Fock space H = N HN .
The annihilation and creation operators ai , ai are very helpful in this for-
mulism because all operators in Fock space can be written as functions of ai , ai .
To prove this conclusion, consider the matrix representation of an operator O
in the occupation number basis

Onm = {ni }| O |{mi } (2.5)


2.1 Second quantization 19

with n and m stands for the occupation number sets {ni } , {mi } respectively.
The operator is written as

O = |{ni } Onm {mi }| (2.6)
{ni },{mi }


Since ai |ni = ni |ni 1 , ai |ni = ni + 1 |ni + 1, we have
1
|{ni } = aini |0 (2.7)
i
ni !

Thus
1 1
O = Onm an i
i
|0 0| am
i
i
(2.8)
{ni },{mi } i
ni ! i
mi !

The state |0 is the vacuum state with ni = 0 for all i. The projector to this
state |0 0| can be considered as a delta function of the operator ni = ai ai :

|0 0| = a ai (2.9)
i
i

Consequently
[ 1
]
O = Onm ani ami (2.10)
{ni },{mi } i
ni !mi ! i ai ai i

is a function of ai , ai . Thus no matter what kind of interaction is present, the


Hamiltonian of this system can be written in terms of ai , ai .
Specically, the ordinary terms in a Hamiltonian such as kinetic energy
and potential energy have simple form in this representation. For example
consider a single particle potential energy term, which in the single particle
Hilbert space is an operator

v |i = v(ri ) |i (2.11)

The corresponding many-body operator V in the Fock space acts on each


boson independently, so that for a state with occupation numbers {ni },

V |{ni } = ni v(ri ) |{ni } = ai ai v(ri ) |{ni }
i i

V = ai ai v(ri ) (2.12)
i

Such a relation between single-body operator and many-body operator can


be generalized to the single-body operators which are not diagonal in basis
20 2 Second quantization and its applications

|i. Such a generalization is possible because the unitary rotation of harmon-


ic oscillator operators are still harmonic oscillator operators. For a generic
unitary rotation |n = Uni |i, an = Uni ai and an satisfy the harmonic oscil-
lator commutation relation. Occupation number basis can be dened in an an
eigenstates. Consequently for any given single-body operator

o |i = oij |j (2.13)

one can always rotate to the eigenstate basis of o and repeat the procedure
above to obtain the many-body operator O. After rotating back to the original
i basis we get

O = ai oij aj = ai i| o |j aj (2.14)
i,j i,j

Similarly, two-body operators such as interaction between two particles


can be written as
1
O = ai aj i| j| o |k |l ak al (2.15)
2
i,j,k,l

with o the two-body operator dened in two-particle Hilbert space. Such an


explicit relation between few-body terms in the Hamiltonian and correspond-
ing terms in many-body Hamiltonian is an important advantage of second
quantized formulism, which greatly simplied the description of many-body
problem. For example, the Hamiltonian of many atoms with kinetic energy

p2 2
hK = =
2M 2M
and two-body interaction energy v(x1 x2 ) has the many-body Hamiltonian
( )
2
H = dxa (x) a(x)
2M

1
+ dxdya (x)a (y)v(x y)a(y)a(x) (2.16)
2

2.1.2 Second quantization for Fermions

We have seen that coupled harmonic oscillators are equivalent to identical


bosons with their wavefunction symmetric under permutation. If we take the
particle coordinate picture from the begining and consider the wavefunction
(x1 , x2 , ..., xN ) for a many-particle system, there is another possibility for
identical particles. The wavefunction can be anti-symmetric in permutation,
which corresponds to Fermions. Fermions also have a second quantized de-
scription. To see that we rst notice that the occupation number basis |{ni }
can also be dened for fermions. The only dierence from boson case is that
2.1 Second quantization 21

ni 1 because no two particles can stay at the same site. Consequently the
fermion Fock space is nite dimensional. When the single particle Hilbert s-
pace is N dimensional, the many-body Hilbert space is 2N dimensional. For
each site with ni = 0, 1 one can also dene the creation and annihilation
operators ci , ci :

c = |0 1| , c = |1 0| (2.17)

It is straightforward to check that c c = |1 1| = n is the particle number,


and the commutation relation between boson operators ai , ai is replaced by
{ }
ci , ci ci ci + ci ci = 1 (2.18)

Using operators ci , we can also obtain all the states in the Fock space from
the vacuum state |0 with ni = 0, i. Consequently all many body operators
are also functions of ci , ci . Especially, a single-body operator

v |i = v(ri ) |i (2.19)

also corresponds to

V = ci v(ri )ci (2.20)
i

However, when we consider the o-diagonal operators we need to be more


careful. The
[ important
] property of bosonic system is that the commutation

relation ai , aj = ij , [ai , aj ] = 0 is invariant under unitary rotation be-
tween ai , which enables the expression of a generic single-body operator in
Eq. (2.14). To obtain the same simple expression here, we need to specify the
commutation relation between operators of dierent sites to be
{ }
{ci , cj } = 0, ci , cj = ij (2.21)

which means the operators at dierent sites are not independent to each other.
It should be noticed that particle number ni = ci ci is still commute with all
other cj , j = i. In this denition we have in general

O = ci i| o |j cj (2.22)
i,j

for single-body operators, and similar for two-body operators.


It should be noticed that the fermion anti-commutation relation (2.21)
is consistent with the anti-symmetric property of wavefunction in particle
coordinate representation. If we dene the wavefunction (x1 , x2 , ..., xN ) as

(x1 , x2 , ..., xN ) = cxN ...cx2 cx1 |0 (2.23)


22 2 Second quantization and its applications

the anti-symmetry is satised.


Further remarks about dimensionality and fractional statistics
What does the particle statistics really mean? The symmetry property
of wavefunction under permutation operation is not really physical. Indeed
the boson and fermion Fock space are dierent, with dierent dimensions.
However we can also consider a boson system in which ni > 1 states have
very high energy so that eectively the Hilbert space is the same as that of
Fermion. Even in that case, the system behaves very dierent from Fermions.
To see the dierence, consider a toy model with time-dependent Hamiltonian
[ ( ) ]
2
H(t) = d r a
d
a(r) + V (r, t)a (r)a(r)
2m
( )2
+U dd ra a(r) 2 (2.24)

in which V (r, t) has the shape of two wells far from each other, and the well
position is time dependent. The last term is an interaction energy which only
depends on the total number of particles. If U is very large, we only need to
consider the states with two particles. No matter whether the particles are
fermions or bosons, the lowest energy state will contain two particles trapped
in the two wells. The only dierence between boson and fermion systems
occur when we move the wells slowly until their position are exchanged at
time t = T . Thus H(t = T ) = H(t = 0). For boson system, the ground
state also comes back to itself at time T . On comparison, the ground state of
fermion system obtains a minus sign. In other words,
T {
i dtH(t) |G(0) , boson
|G(T ) = T e 0 |G(0) = (2.25)
|G(0) , fermion
T
i dtH(t)
in which T e 0 is the time-evolution operator. Such a sign is a Berrys
phase which is well-dened and physical. In this way we see that fermion and
boson correspond to dierent representation of permutation group, since each
permutation of N particles will be assigned a sign 1 depending on its parity.
Mathematically, such a point of view leads to an interesting generalization in
two spatial dimensions. In two dimensions there are two permutation process
(clockwise and counterclockwise) which cannot be smoothly deformed to each
other. Consequently the permutation group is enlarged to braid group which
has more complicated representations. This concept was introduced to physics
by Wilczek (Phys Rev Lett 49 957 (1982)). Such a dierence between two and
higher dimensions enable particles which are neither boson nor fermion in two
dimensions, which are called anyons.
2.2 Applications of second quantization 23

2.2 Applications of second quantization


2.2.1 Free fermion gas

Since we have studied two boson systems in the last Chapter, as application
of second quantization we rst study some fermion systems. The electrons in
a crystal can be described by the following general Hamiltonian if we ignore
the phonons for a moment
(( ) )
2
H = d3 r (r) + U (r) (r)
2m

1
+ d3 rd3 r (r) (r ) V (r r ) (r ) (r) (2.26)
2

U (r) is the ion potential and V (r r ) is the Coulomb interaction between


electrons. is called chemical potential which will be explained later.
We rst study the free electron gas with U (r) = 0, V (r r ) = 0. In this
case the Hamiltonian has translation symmetry and is quadratic. It can be
diagonalized in the same way as bosons:
( k2 )
H= k k (2.27)
2m
k

These are fermionic analog of decoupled harmonic oscillators. For each k there
are two states labeled by occupation number nk = k k = 0, 1. Thus we
immediately know the ground state of the system is given by
{
0, k > 0
nk = (2.28)
1, k < 0

with k = k2 /2m . This result already shows an important dierence


between boson and fermion systems: the energy of harmonic oscillator k for
boson case must be non-negative. Otherwise the energy wont be bound from
below. On comparison Fermion single particle energy can have either sign
while the Hamiltonian is always bound from below (because the Fock space
is nite dimensional).
For > 0, k = 0 gives |k| = kF = 2m, and all the states with |k| < kF
is occupied in the ground state. Such a ground state is called a fermi sea
while the boundary of fermi sea at |k| = kF is called the fermi surface.
For a free fermion system with having intersection to the energy dispersion,
there is always a fermi sea and a fermi surface. The low energy excitations of
the fermion system are excitations near fermi surface, rather than near k = 0
like in the boson case. There are also much more excitations compared to the
boson case, which can be seen by studying the specic heat. The partition
function of a free fermion gas is given by
24 2 Second quantization and its applications
{ }
( )
Z () = exp nk k = 1 + ek (2.29)
nk =0,1 k k

from which the specic heat can be computed.

2.2.2 Tight-binding models and graphene as an example

Now we consider the eect of ion potential U (r), but still ignore the electron-
electron interaction V (r r ) term. When the lattice vibration is ignored,
U (r) is periodic in a crystal, i.e. U (r + ai ) = U (r) for any Brava lattice
vector ai . Dierent treatments and approximations can be applied to such a
Hamiltonian. For example we can assume U (r) is a small perturbation so that
the eigenstates are similar to the plane-wave obtained in the free electron gas.
Alternatively, one can take the opposite point of view that the traps in U (r)
are very deep so that the electrons are almost trapped around each ion. Here
we will study the latter case since it leads to a simplied discrete description
of the system.
We start from the atomic limit where the ions are innitely far away,
so that dierent atoms can be treated independently. Each atom has its own
atomic levels, such as s, p, d orbitals. The second quantized Hamiltonian is
simply a sum of all atoms

H= n cin cin (2.30)
i n

where n labels the atomic energy levels. Now consider the atoms with nite
distance, in which case the atomic orbitals are not eigenstates of the Hamil-
tonian any more. However, we can still use the atomic orbitals as a basis to
express the Hamiltonian. The matrix elements of the single particle Hamilto-
nian in this basis is given by
( 2 )
p
hin,jm = in| + U (r) |jm (2.31)
2m
which corresponds to the second quantized Hamiltonian

H= cin hin,jm cjm (2.32)
i,n;j,m

Such a Hamiltonian in discrete basis is called a tight-binding model. Because


the atomic orbitals are localized, hin,jm decays exponentially versus the dis-
tance between i and j sites, so that only several nearest neighbors are impor-
tant. This is why the model is called a tight-binding model.
It should be noticed that the description above is not rigorous because
the atomic levels are not orthogonal, and also we have ignored the deep core
levels and higher unoccupied levels of each atom. Rigorously speaking, all lev-
els have to be included to provide a complete basis of the continuous system
2.2 Applications of second quantization 25

with single-particle Hamiltonian (2.31). Alternatively, one can stay with nite
number of states per site by using Wannier function basis rather than atomic
levels. Wannier functions are dened to be local single-particle states which
form a complete basis of the relevant energy bands. They are modied ver-
sion of the atomic levels which are orthogonalized while remain local. More
concretely, they are fourier transform of the Bloch states
1 ikR
|Rn = e |kn (2.33)
N kBZ

Intuitively, Wannier function is a set of basis in the Hilbert space of low energy
states which is most closed to the coordinate basisi.e., the basis formed by
Dirac functions. For the purpose of qualitative discussions below, we can
safely neglect the dierence between Wannier functions and atomic orbitals.
Now we take the interesting two-dimensional material graphene as an ex-
ample for tight-binding models. Graphene (for which Andre Geim and Kon-
stantin Novoselov were awarded the 2010 Nobel Prize for Physics) is a two-d
honeycomb lattice of carbon atoms. The relevant carbon atomic orbitals are
2s and 2p orbitals. In the honeycomb lattice the s and px , py orbitals are hy-
bridized and form 3 sp2 hybridized orbitals named as orbitals. pz does not
involve in the hybridization and is called orbital. The orbitals are oriented
in-plane and strongly coupled to each other, and the orbital wavefunction is
more localized in the in-plane direction so that they are coupled with weaker
strength. The low energy behavior of graphene is determined by orbitals.
The simplest tight-binding model of orbitals is given by
( )
H = t ci cj + h.c. (2.34)
ij

with ij standing for neighbor sites in the honeycomb lattice. Similar to


the treatment to coupled harmonic oscillators in Chapter 1, we use fourier
transformation to diagonalize the Hamiltonian. It should be noticed that there
are two atoms in each unit cell of the honeycomb lattice. In other words there
are two atoms which are not equivalent by translation. Dene the two unit
vectors of the honeycomb lattice as a1 and a2 with
1 ( )
a1,2 = 3x y (2.35)
2
(For simplicity we have taken the lattice constant a = 1.) Correspondingly the
momentum k is dened in a Brillouin zone with the periodic boundary con-
dition ck+b1 = ck , ck+b2 = ck . Here b1 and b2 are unit vectors of reciprocal
lattice satisfying

bi aj = 2ij (2.36)

Explicitly
26 2 Second quantization and its applications

2 ( )
b1,2 = x 3y (2.37)
3
The Hamiltonian is written as
( )
H = t ciA ciB + ciA cia1 ,B + ciA ci+a2 a1 ,B + h.c. (2.38)
i

After fourier transformation


( )( 0 f (k)
)(
ckA
)

H= ckA , ckB (2.39)
f (k) 0 ckB
kBZ
( )
f (k) = t 1 + eika1 + eik(a2 a1 )

The 2 2 matrix can be easily diagonalized to obtain



H= Eks cks cks
kinBZ s=1


Eks = st 1 + eika1 + eik(a2 a1 ) (2.40)

Thus there are two energy bands with opposite dispersion relation. It should
also be noticed that the -orbital is half-lled, so that the fermi level is exactly
in the middle of the two bands. Because Ek+ 0 Ek , the two bands are
either well-separated or touching at zero energy. It can be checked that there
are two points in the Brillouin zone where Ek+ = 0, with the coordinates
K = (b1 b2 ) /3 and K = K. Near K and K points Eks is linear in
|k K| or |k K |, which is why electrons behave like massless relativistic
particle in graphene.
Further remarks:
What makes graphene so special? The simple tight-binding model dis-
cussed above works very well for graphene because the coupling beyond n-
earest neighbor is negligible. The bands which are strongly bonded make
graphene very robust and very clean. (Before graphene was discovered it was
widely believed that a 2d crystal was not stable.) The bonding between
orbitals studied above leads to a linear dispersion of electrons with the high
velocity of 106 m/s, 1/300 of the speed of light. One of the interesting conse-
quences of the relativistic dispersion is observed in the quantum Hall eect
of graphene, which was among the earliest experiments on graphene. (See
Novoselov, K. S. et al. Two-dimensional gas of massless Dirac fermions in
graphene. Nature 438, 197C200 (2005); Zhang, Y. et al, Experimental obser-
vation of the quantum Hall eect and Berrys phase in graphene. Nature 438,
201C204 (2005).) There is a simple understanding on the dierence between
relativistic and non-relativistic fermions in a strong magnetic eld. There is
a universal unit of the magnetic ux, the ux quanta 0 = hc/e. Thus a
magnetic eld B strength corresponds to a density B/0 which has the u-
nit of length2 . The discrete energy levels in a magnetic eld corresponds
2.2 Applications of second quantization 27

to the electron density values = (n + )B/0 with n integer and some


constant that cannot be obtained in this semi-classical picture. Each density
corresponds to a fermi surface size according to the formula
AF S
= (2.41)
(2)2

For circular Fermi surface AF S = kF2 so that

B
kF2 = 4 = 4(n + ) (2.42)
0
Consequently, for a non-relativistic fermion with parabolic dispersion

h2 kF2 2h2 B eB
E= = (n + ) = (n + ) h (2.43)
2m m0 mc
For the fermion with relativistic dispersion

2Be
E = hvkF = hv (n + ) (2.44)
hc
Interestingly,
= 1/2 for parabolic case but = 0 for relativistic dispersion.
Thus E nB for graphene. An important consequence is that the Landau
level gap between two neighbor energy levels split much faster for graphene
than for an ordinary electron gas, so that graphene quantum Hall eect can
be observed even at room temperature.

2.2.3 Interaction eect in tight-binding systems

Now we consider the interaction term in Eq. (2.26):



1
Hint = d3 rd3 r (r) (r ) V (r r ) (r ) (r) (2.45)
2
In a tight-binding system we can expand

(r) = i (r)ci (2.46)
i

with i (r) the Wannier function at i site. Here we have consider a simple
tight-binding model with only one orbital per site, such as graphene. In this
basis
1
Hint = ci cj Vijkl cl ck (2.47)
2
ijkl

with
28 2 Second quantization and its applications

Vijkl = d3 rd3 r i (r)j (r )l (r )k (r)V (r r ) (2.48)

Notice that we have recovered the spin indices ignored in earlier discussion.
We will see below that with interaction, spin plays an important role.
Since i (r) is localized near site i, the leading term in Vijkl is i = j = k = l.
Dene Viiii = U we have
U
HU = ci ci ci ci = U ni ni (2.49)
2 i i

This is called Hubbard interaction. If we only keep the kinetic energy term
and this term, the model is called Hubbard model.
The important subleading terms are Vijij = V and Vijji = J F for i, j
nearest neighbors. The corresponding terms are

HV = V ni nj (2.50)
ij

and
( 1
)
Hex = J F ci cj ci cj = 2J F Si Sj + ni nj (2.51)
4
ij ij

Interestingly, we have obtained a spin-spin interacting from a purely charge


(Coulomb) interaction. This term is called exchange interaction and is the
origin of ferromagnetism. Physically such an interaction is due to Pauli ex-
clusion principle between fermionic electrons. Two electrons with the same
spin are repulsive to each other so that the Coulomb energy is lower than two
electrons with opposite spin.
Its easy to check that U, V, J F > 0. Thus it seems that electron spin
always prefers to align. However, there are also antiferromagnetic materials
in which the electron spins prefer to be anti-parallel to each other, which
corresponds to an JSi Sj interaction with positive coupling constant. Such a
term arises from a dierent mechanism as the exchange interaction, named as
superexchange. This process is important when V , J F terms can be ignored,
and the system is described by the Hubbard model:
( )
H = t ci cj + h.c. + U ni ni (2.52)
ij i

We rst study the U term. If the occupation of up and down electrons at a


given site is ni = ni + ni < 2, U ni ni = 0. If ni = ni = 1, U ni ni = U .
Consequently the Hubbard interaction U gives the double-occupied site a high
energy. If there are one electron per-site on average, such an interaction will
require the ground state to have exactly ni = 1 for each site. Thus each site
has an electron with spin up or down. Before considering the eect of t term,
2.2 Applications of second quantization 29

the spin of each electron is independent, leading to a huge 2N degeneracy in


the ground state of HU . If we consider the large U limit t U , the eect of t
term can be studied by perturbation theory. One can see that in the rst order
of perturbation, t term has no matrix element between any two of the 2N low
energy states. The leading non-trivial correction comes from the second order
perturbation. The eective Hamiltonian in the low energy subspace is given
by
1
H (2) = Ht |N U EU
N | Ht (2.53)
U >0
EG N
N,EN

U U
in which EG and EN are the energy of HU in the 2N low energy states and
the higher energy state |N , respectively. Since Ht acting on a single-occupied
state can only create one double-occupied site, the state |N must have energy
U
EN = U , so that

2t2
H (2) = ci cj cj ci = J Si Sj + constant (2.54)
U
ij ij

Here some constant proportional to particle number ni is neglected since ni =


2
1 for all low energy states. J = 4tU . Such a spin-spin interaction is called
superexchange interaction.
More generally, if we dont require the system to have one electron per
site, the low energy subspace with HU = 0 contains 3 states per site, with
the total dimension 3N . The same kind of perturbation theory can be applied
which leads to the so-called t J model:
( ) ( 1
)
HtJ = t ci cj + h.c. + J Si Sj ni nj (2.55)
4
ij ij

with ci the annihilation operator restricted in the low energy Hilbert space.
In other words ci and ci act in the subspace with no double-occupancy, and
ci |i = 0 instead of creating a double-occupied state. The Hubbard model
and t-J model are considered as relevant to the high-temperature supercon-
ductivity in cuprates such as YBa2 Cu3 O6+x , La2x Srx CuO4 , etc. Although
they look simple, a controllable way to solve them has not been found despite
of tremendous eorts.

2.2.4 Quantum ferromagnetism and anti-ferromagnetism

Having derived the exchange interaction (2.51) and super-exchange interac-


tion (2.54) for dierent physical systems, we still need to do more work to
obtain the low energy properties of the system. For simplicity, we consider
the Hamiltonian with only the exchange or superexchange term, which corre-
sponds to a system with all electrons localized and the lling is commensurate
30 2 Second quantization and its applications

(one electron per lattice site). Such a model with only spin interaction is called
a Heisenburg model. I would like to mention that such an approximation is
suitable for the superexchange case, as can be seen in the derivation from
Hubbard model to Heisenburg model. However for the case of exchange inter-
action, the system is usually still metallic, which means there are additional
electrons moving around beside the localized one forming the spin. Thus the
spin model only describes the physics approximately, when the metallic elec-
trons (usually called itinerant electrons) and localize electrons interact weakly.
We start from the exchange interaction

H = 2J F Si Sj (2.56)
ij

Since spins prefer to be parallel to each other, it is natural to expect that


the ground state is ferromagnetic, with all spins aligned to one direction.
Electrons have spin 1/2 but for generality we study the model above with
generic spin S. Physically, a spin large than 1/2 can arise from the spin-spin
coupling of multiple electrons on the same site, which is known as the Hunds
rule coupling. The spin SU (2) algebra is
[Si , Sj ] = iijk Sk (2.57)

with S = S (S + 1). Dening S = Si iSy we have
2
[ + ] [ ] [ ]
S , Sz = iS + , S , Sz = iS , S + , S = 2Sz (2.58)
Thus S are similar to the annihilation and creation operators and Sz is simi-
lar to the number operator of harmonic oscillator. However Sz has eigenvalues
S, S + 1, ..., S 1, S which is dierent from the harmonic oscillator case.
In the large S limit we can consider a ferromagnetic state with Sz = S, and
small uctuation near that state. In that case one can dene

Sz = S + a a, S 2Sa, S + 2Sa (2.59)
More rigorously, we can also include the dierence between the two algebras
to obtain

Sz = S + a a, S = 2S a aa, S + = a 2S a a (2.60)
with the constraint a a 2S. This representation is called Holstein-Primako
transformation. In the way the spin Hamiltonian can be reduced to a harmonic
oscillator Hamiltonian
[ 1( + )
]
H = 2J F Siz Sjz + Si Sj + h.c.
2
ij
[ ]
2J F Sai ai Saj aj + Sai aj + aj ai
ij
( )
F
= 2J S ai aj (ai aj ) (2.61)
ij
2.2 Applications of second quantization 31

We see that in the continuum limit we get



( )
H = 2J F Sa2d dd x a (a) (2.62)

so that the dispersion is Ek = 2J F Sa2d k2 . If we dont take the continuous


limit, on square lattice we have
2
H = 2J F S ak ak eiki 1 (2.63)
k i=x,y,..

Because of the spin SU (2) rotation symmetry is spontaneously broken in a


ferromagnetic phase (by choosing a direction of magnetization), the spin-wave
modes we obtain here are required to be gapless. Interestingly, the gapless spin
wave has quadratic dispersion rather than linear.
Now we study the anti-ferromagnetic system. With anti-ferromagnetic in-
teraction the spin on nearest neighbor sites prefer to be anti-parallel. This
is harder than being parallel because its not always possible to make each
two neighbor spins anti-parallel. It is only possible to have such a stateNeel
state if the lattice is bipartite, i.e. if the sites in the lattice can be colored to
two colors and only dierent color are adjacent to each other. For example
square lattice is bipartite, and triangular lattice is not. Classically, on a bipar-
tite lattice the ferromagnetic and anti-ferromagnetic states are simply related
by a rotation of spin on one of the sublattices. However quantum properties
of them are quite dierent.
Similar Holstein-Primako transformation can be dened. However, the
transformation needs to be dened dierently on the two sublattices since the
spin in the ground state is opposite for the two sublattices. Now we study
the antiferromagnetic spin system on two-dimensional square lattice. Take
Sz = S for sublattice A and Sz = S for sublattice B we dene

Siz = S + ai ai , Si = 2S ai ai ai , Si+ = ai 2S ai ai , i A

Siz = S ai ai , Si = ai 2S ai ai , Si+ = 2S ai ai ai , i B (2.64)
The Hamiltonian is expanded to
( ( ))
H=J Si Sj J Sai ai + Saj aj + S ai aj + aj ai (2.65)
ij ij

Compared to the ferromagnetic case, ai and ai are exchanged on B sublattices,


so that ai aj term appears. With Fourier transformation
[( ) ]
H = JS ak ak + ak ak (cos kx + cos ky ) + 4ak ak
k
( )( 2 cos kx + cos ky
)(
ak
)
= JS ak , ak (2.66)
cos kx + cos ky 2 ak
k
32 2 Second quantization and its applications

In this way the Hamiltonian looks like the one of graphene (2.39) which has a
2 2 matrix and needs to be diagonalized. However here the vector contains
ak and ak , so that we cannot use unitary transformation to diagonalize
the matrix. To gure out what transformations are allowed, we consider the
generic linear transformation to the vector:
( ) ( )
bk ak
=V (2.67)
bk ak

with V a generic 2 2 matrix. We want V hV to be diagonalized with h the


kernel matrix
( )
2 cos kx + cos ky
h=
cos kx + cos ky 2

On the other hand, we need to make sure bk is also a boson operator. Thus
we require
[( ) ] ( )
bk 1 0
, (b ,
k kb ) = z (2.68)
bk 0 1

This leads to the requirement

V z V = z , V hV = D (2.69)

with D a diagonal matrix. Thus V are the transformations which preserves


the matrix z . This is a generalized eigenvalue problem for matrix h which
can be also formulated to an eigenvalue problem:
1
hV = V D = z V z D z hV = V z D (2.70)

Thus z D contains the eigenvalues of


( )
2 cos kx + cos ky
z h = (2.71)
cos kx cos ky 2

which are

2
1,2 = Ek = 4 (cos kx + cos ky ) (2.72)

Correspondingly
( )
u k vk
V =
vk uk

1 1
uk = +
Ek 2

1 1
vk = sgn (cos kx + cos ky ) (2.73)
Ek 2
2.2 Applications of second quantization 33

The Hamiltonian is diagonalized to



2
H= bk 2JS 4 (cos kx + cos ky ) bk (2.74)
k

It should be noticed that during the derivation we always omit the constant
terms coming from commutators, which are irrelevant. It should be noticed
that this dispersion has a period in momentum space E(kx , ky ) = E(kx +
, ky + ), which is smaller than the period given by Brillouin zone of the
square lattice. Such a smaller momentum space unit cell is a consequence of
the larger real space unit cell since the two sublattices are not equivalent. The
low energy excitations are near k = (0, 0) and k = (, ). Near each point the
dispersion is linear:

Ek = 2 2JS |k| (2.75)

Thus the spin wave of antiferromagnetic spin state has linear dispersion, just
like phonons and photons. The transformation above is called bosonic Bo-
golubov transforation, which is the quantum version of canonical transfor-
mation of momentum and coordinate in classical mechanics.
Compared to ferromagnetic system where the Hamiltonian can be diag-
onalized without this Bogolubov transformation, there are some important
physical consequences beside the dierent dispersion. In the ferromagnetic s-
the spin is fully polarized given by Sz = S, since
tate, at zero temperature

Sz = S + a a and a a = 0. However in the antiferromagnetic case
1
Siz = S + ai ai = S + ak ak
N
k
1 ( )( )
= S + uk bk vk bk uk bk vk bk
N
k
1 2
= S + vk
N
k

d2 k 1 1
= S + 2
(2.76)
(2) 2 2
4 (cos kx + cos ky )

Thus even at zero temperature, the spin is not fully polarized.


34 2 Second quantization and its applications

2.2.5 Transverse-eld Ising model

In the subsection above we have shown how the ferromagnetic and anti-
ferromagnetic spin models are treated in the large-S approximation. Usually,
spin models cannot be solved exactly. However in one dimension it is possible
to obtain exact solution. A simple and interesting example is the transverse
eld Ising model:

H = J iz i+1
z
h ix (2.77)
i i

Here iz and ix are the Pauli matrices which are related to spin operators of
spin-1/2 particle by Sia = 21 ia , a = x, y, z.
If h = 0, the Hamiltonian is classical, since iz all commute with each
other. If J > 0, the ground state is ferromagnetic with all spin up or all
spin down. Notice that we only have a coupling between z component of
spin so that there is no continuous spin rotation symmetry. Consequently
the ferromagnetic ground state only breaks a discrete symmetry, which is
the rotation around x direction. This operation is dened by the following
operator

R= ix , R2 = 1, R1 HR = H (2.78)
i

For h = 0 this discrete symmetry is still preserved. If we consider a small


h term as perturbation to the h = 0 Hamiltonian, it can be seen that the
matrix element between the two ground states | ... and | ... is zero
until N -th order, with N the total number of sites.
( )m

... | h ix
| ... = 0, m < N (2.79)
i

Consequently it is natural to expect the symmetry breaking to be stable a-


gainst a small h. In other words, the symmetry wont be immediately recovered
for an innitesimal h because the two symmetry-breaking ground states are
only hybridized with an exponentially small matrix element.
On the other hand, if we consider the opposite limit J = 0, h = 0 we see
that all spins are aligned to +x direction if h > 0, so that the -rotation
symmetry along x is preserved. Thus there must be some critical value of h
where a phase transition occurs between the symmetry breaking and symme-
try preserving phases.
As has been discussed earlier, spin operators do not satisfy the boson or
fermion algebras, which make it dicult to solve it by unitary transformation.
The Hilbert space of the spin-1/2 model is 2N dimensional, with two states
on each site. Compared to a tight-binding model of fermions we see that the
states of spin-1/2 model have one-to-one correspondence to those of a fermion
2.2 Applications of second quantization 35

model, with only one orbital on each site (and no additioal spin degree of
freedom). The states | and | correspond to fermion states |0 and |1 in
occupation number basis. However, the Hamiltonian of a fermion and a spin
model is usually very dierent since the former is written as a function of
fermion operators ci , ci which anti-commute between dierent sites. In the
spin-1/2 language, this means the corresponding spin operators are non-local.
In generally, a fermion model can be mapped to a non-local spin model. The
non-locality makes this mapping useless. However, one-dimension is special
because two particles cannot exchange if they are not allowed to penetrate
into each other. The non-locality in the mapping above disappears in this case
which makes it possible to map a fermion model to a local spin model, and
vise versa. The transverse eld Ising model is solved because such a mapping
to a fermion model.
The mapping between fermion and spin is dened explicitly by
1 x
ni = ci ci = ( + 1)
2 i
1 y
i1
ci z
= (i + ii ) jx
2 j=1

1 y
i1
ci = (i ii )
z
jx (2.80)
2 j=1

One can check explicitly


{ }
{ci , cj } = 0, ci , cj = ij (2.81)
i1
The string operator j=1 jx is essential for the anti-commutation relation
ofci , ci
between dierent sites. This mapping from spin to fermions is called
Jordan-Wigner transformation. To map the Hamiltonian (2.77) to fermions,
we write down the inverse map

ix = 2ni 1
( ) i1

iy = ci + ci (2nj 1)
j=1
( ) i1

iz = i ci ci (2nj 1) (2.82)
j=1

In the fermion operators the Hamiltonian (2.77) can be written as


[ 1
]

N
N
H = J i i+1 h
z z
i JN
x z z
1
i=1 i=1
36 2 Second quantization and its applications


N 1 ( )( )
N
=J ci ci ci+1 ci+1 (2ni 1) h (2ni 1)
i=1 i=1
( ) N
1 ( )
+J cN cN (2nj 1) c1 c1
j=1


N 1 ( )( )
=J ci + ci ci+1 ci+1 2h ni
i=1 i
( )
N ( )
+J cN + cN (2nj 1) c1 c1 (2.83)
j=1

where a constant N h has been omitted. Interestingly, the model is mapped


to a non-interacting fermion model except the last term which is required by
the periodic boundary condition. If we study a system with open boundary
condition, the last term is absent and the spin model is completely equivalent
to a free fermion model, so that the problem can be exactly solved. For periodic
boundary condition we need to be more carefully about the last term. Notice
that the string factor


N
N
(2nj 1) = jx = R
j=1 j=1

is exactly the rotation operator we discussed at the beginning, which is a


symmetry of the Hamiltonian. Since [R, H] = 0, the Hamiltonian is diagonal-
ized in the eigenvalues of R. Since R2 = 1, R has only two eigenvalues 1.
We rst study the sub-Hilbert space with R = 1. In this case the Hamiltoni-
an is a free-fermion Hamiltonian which can be block-diagonalized by Fourier
transformation:
( )
H=J ci ci+1 + ci+1 ci + ci+1 ci + ci ci+1 2h ni
i i
[ ( )]
= (2J cos k 2h) ck ck + iJ sin k ck ck ck ck
k
( ) ( J cos k h iJ sin k
)(
ck
)
= ck , ck (2.84)
iJ sin k J cos k + h ck
k

This is the fermionic counter-part of Eq. (2.66) which is 2 and involves pair-
ing terms ck ck and ck ck . Similar to the boson case this Hamiltonian can be
diagonalized by Bogoliubov transformation. The fermionic Bogoliubov trans-
formation is simpler than boson because
{( ) } ( )
ck 10
, (ck , ck ) = (2.85)
ck 01
2.2 Applications of second quantization 37

so that the allowed transformations are simply unitary transformations. The


Hamiltonian is diagonalized to

2
H= ak ak 2 (J cos k h) + J 2 sin2 k (2.86)
k

with
( ) ( )( )
ak uk vk ck
= (2.87)
ak vk uk ck

with

1 k
uk = +
2 2Ek

1 k
vk = isgn (J sin k) (2.88)
2 2Ek

2
with Ek = (J cos k h) + J 2 sin2 k, k = J cos k h

The energy Ek is usually gapped, for both large and small h/J, but it is
gapless for h = J. Going back to the spin model, the critical value h = J
corresponds to the phase transition we expected from the two limiting cases
h = 0 and J = 0.
It should be noticed that the Fermion model has a unique ground state de-
termined by occupation number ak ak = 0, while the spin model with |h| < J
has two degenerate ground states. Such a dierence comes from the require-
ment R = 1 we take when reducing the spin model to the fermion model.
If we take R = 1 we will get a slightly dierent Fermion model with also
a unique ground state. These two ground states are eigenstates of R. Since
the R operator ips between the two fully polarized states, its eigenstates are
superpositions of the fully polarized states | ... and | ... :
1
|G(R = 1) = (| ... | ... ) (2.89)
2
The open boundary system is also interesting. With open boundary, the
last term in Eq. (2.83) is absent, and the mapping to fermion is straight-
forward. As we know from the h = 0 limit, the phase with |h| < J has two
degenerate ground states. Correspondingly, the fermion system must also have
two degenerate ground
states. In a free fermion system the energy eigenval-
ues are given by m nm Em with Em the single particle energy eigenvalues
and nm the occupation number. Thus the ground state is degenerate only if
some Em = 0. Thus we obtain the conclusion that the Fermion Hamiltonian
(2.83) with open boundary should have a zero energy quasi-particle state for
|h| < J. This statement can be directly veried by solving the open boundary
Hamiltonian numerically or analytically. The zero mode has a wave-function
38 2 Second quantization and its applications

localized on the two edges of the chain. Such zero modes are called Majo-
rana zero modes because their relation to Majorana fermion, which is its own
anti-particle.

2.2.6 Bosonization

From the example of transverse-eld Ising model we have seen the special
property of one-dimensional system which makes boson (spin) and fermion
systems equivalent to each other. Another consequence of such equivalence
is the bosonization method which describes the low energy excitations of an
interacting fermion system by bosons. The intuitive reason why bosonization is
possible is that fermions do not penetrate into each other, so that the fermionic
statistics is not important. Thus the fermion with interaction behaves similarly
as coupled harmonic oscillators.
To be more precise, consider a one-d tight-binding Hamiltonian
( )
H = t ci ci+1 + h.c. ci ci
i i
1
+ V (ri rj ) (ni n) (nj n) (2.90)
2 i,j

in which n = N1 i ni is the average fermion density. Since total electron
number is conserved, we can consider n as a classical number. We use ni
n instead of ni in the interaction term so that the interaction term wont
change the fermion density for xed chemical potential . For simplicity we
consider spinless fermions so that the interaction only contains density-density
interaction. If we instead introduce an interaction term V ni nj , the fermion
density will depend on V , so that Fermi momentum kF and Fermi velocity
vF will be interaction dependent.
Consider the Fourier transformation of density operator ni which corre-
sponds to density uctuation:
1 1
(q) = ni eiqri = ck ck+q (2.91)
N i N k

Thus the operator (q) creates a particle-hole excitation by removing an elec-


tron at k + q and create one at k. The free fermion energy spectrum is

k = 2t cos k (2.92)

For 2t < < 2t there is a lled Fermi sea with Fermi surface consisting of
two points kF and kF . The states kF k kF are lled. Thus for the
ground state ck ck+q |G = 0 only if

|k + q| kF , |k| > kF (2.93)


2.2 Applications of second quantization 39

This leads to
kF q < k < kF , if q > 0
kF < k < kF + q, if q < 0 (2.94)
Consequently, for small q satisfying |q| kF , the operator (q) acting on
the ground state |G only creates some excitation around kF (kF ) if q > 0
(q < 0). Moreover, the energy of the excited state ck ck+q |G is given by
Ekq = k k+q . For q > 0 and kF q < k < kF we have approximately
k vF k, so that Ekq = vF q. Similarly for q < 0 we have Ekq = vF q.
Consequently the energy of the excited state (q) |G is simply Eq = vF |q|.
Such an observation strongly indicate that the density uctuations created
by (q) are linear dispersing bosons, just like phonons in a one-dimensional
chain. From the argument above one can see clearly that such a property is
special for 1d. For higher dimensions Ekq = k k+q depends on both k and
q even if |q| kF . To describe the density uctuations as bosons, we rst
separate the left movers and right movers.

1 1
(q) = ck ck+q =
ck ck+q +
ck ck+q + ck ck+q
N k N |k+k |< |kk |< other k
F F

= L + R + H (2.95)
with L and R containing the excitations near left and right Fermi surfaces,
and H the higher energy parts which are not important in low energy behav-
ior of the system. kF is a cut-o in momentum. The following treatment
is only correct for long wavelength limit q < . According to the discus-
sion above, L (q) creates particle-hole excitation for q > 0 and R (q) creates
excitation for q < 0. Thus we can naturally dene
L (q) = f (q)bq , q > 0
R (q) = f (q)bq , q < 0 (2.96)
The function f (q) can be determined by commutation relations.
[ ] 1 [ ] 1 ( )
L (q), L (q) = ck ck+q , ck +q ck = ck ck ck+q ck+q (2.97)
L L
k,k k

Taking the expectation value of the right-side in ground state we have


[ ] q
L (q), L (q) (2.98)
2

The same is true for R (q). Thus f (q) = |q| /2. As analyzed above, the
boson has the energy vF |q| so that the kinetic energy is eectively written as

Ht vF |q| bq bq (2.99)
|q|<
40 2 Second quantization and its applications

Interestingly, the interaction term is also quadratic in these bosons.


1 1
HV = V (ri rj ) (ni n) (nj n) = (q)(q)V (q)(2.100)
2 i,j 2
q=0

V (q) = V (ri )eiri q
i

Thus
1
HV (L (q) + R (q)) (L (q) + R (q)) V (q)
2
q=0
1 q ( )( ) 1 q ( )( )
= V (q) bq + bq bq + bq V (q) bq + bq bq + bq
2 q>0 2 2 q<0 2
( ) ( )
1 ( ) g4 g2 bq
= q bq , bq (2.101)
2 g2 g4 bq
q>0

in the last line we have dened the parameters g2 and g4 . For this interaction
g2 = g4 = V (q) but its helpful to keep it general since other type of interaction
may make them dierent. Physically, g4 corresponds to forward scattering and
g2 corresponds to back-scattering.
Once we get this form we can take H = Ht +HV and apply the Bogoliubov
transformation
( ( )( )

) vF + g4 /2 g2 /2 bq
H= q bq , bq
g2 /2 vF + g4 /2 bq
q>0

= Eq aq aq (2.102)
q=0

with
(
g4 )2 ( g2 )2
Eq = |q| vF + (2.103)
2 2
The relation between aq and bq can be obtained in the same way as in Eq.
(2.66). From this result we see that the interacting system is similar to non-
interacting system, with a renormalized Fermi velocity. However, it should be
remember that the ground state is not the non-interacting ground state any
more, because of the Bogoliubov transformation. (Recall that the same thing
happens for anti-ferromagnet, where the ground state is dierent from the
classical Neel ground state because of the Bogoliubov transformation.)
3
Path integral

3.1 Single-particle path integral


3.1.1 Introduction to single-particle path integral

Path integral is an alternative formulation of quantum mechanics developed


by R. P. Feynman. The advantage of path integral approach is its intuitive
relation to classical mechanics and its ability to provide some important non-
perturbative methods.
The idea of path integral is to focus on the time-evolution operator:

U (t) = eiHt (3.1)

The time evolution of any given state | is given by |(t) = U (t) |. Cer-
tainly, U (t) for a given t contains all information about the Hamiltonian H,
so that U (t) contains full information of a quantum mechanical system.
Now consider the decomposition
( )N
U (t) = eiHt/N = U ()N , = t/N (3.2)

If we write down the matrix of operator U () and U (t) in some given basis |j
we see

UjN j0 (t) = UjN jN 1 ()UjN 1 jN 2 ()UjN 2 jN 3 ()...Uj1 j0 () (3.3)
j1 ,j2 ,...jN 1

It can be seen that in term of U () the matrix element of U (t) is given by


a summation over all paths j0 j1 ...jN which ends at j0 and jN . This is the
simplest example of a path integral. What we really use in path integral is
a bit dierent, using overcomplete basis rather than complete ones, but the
idea is the same as above: the time-evolution operator U (t) can be obtained
by splitting into small time slices and summing over paths. This is possible
because the linearity of quantum mechanics.
42 3 Path integral

Now we consider the path integral for a single particle

p2
H= + V (x) (3.4)
2m
The time-evolution operator U () for a small time is

U () = eiH I iH (3.5)

If we write U () in an orthogonal basis such as x basis, the matrix is almost di-


agonal, and hard to write down. Instead, we use both momentum and position
eigenstates. The completeness of each basis requires

dx |x x| = I, dp |p p| = I (3.6)

In term of wave-functions in coordinate basis x| y = (x y) and x| p =


1 eipx . We insert the basis
2

dxdp ipx
I= dxdp |x x| p p| = e |x p| (3.7)
2

between each two U () in U (t) = U ()N . This leads to


N
1
N
N
xN | U (t) |x0 = dxi dpi eipi xi pi | U () |xi1 (3.8)
i=1 i=1 i=1

The advantage of using both |p and |x is that both kinetic energy and
potential energy terms in the Hamiltonian has a simple form in the matrix
element pi | U () |xi1 .

pi | U () |xi1 pi | (I iH) |xi1 = eipi xi1 (1 iH(pi , xi1 ))


eipi xi1 iH(pi ,xi1 ) (3.9)

Thus
N
1
{ }

N
N
xN | U (t) |x0 = dxi dpi exp i [pi (xi xi1 ) H(pi , xi1 )]
i=1 i=1 i=1

D[x(t)]|x(0)=x0 ,x(t)=xN D[p(t)]ei dt(pxH(p,x))
(3.10)

One can see that the term appears on the exponential is nothing but La-
grangian in classical mechanics. However Lagrangian is usually a function of
x, x but here we have L as a function of p and x, x. In the last step we have take
the approximation 0, N . The integral over xi , pi has been replaced
by functional integral which is integrated over the spaces of all paths. This
3.1 Single-particle path integral 43

is the path integral representation of the single particle quantum mechanics.


The functional on the exponential

S[p(t), x(t)] = dt (px H(p, x)) (3.11)

is the action of this system.


An alternative representation can be obtained by integrating over pi ex-
p2i
plicitly. Since H(pi , xi1 ) = 2m + V (xi1 ), the action only contains quadratic
and linear terms in p so that the integration over p can be done explicitly to
obtain
N1
{ N [ ( 2 )]}
N pi
xN | U (t) |x0 = dxi dpi exp i pi (xi xi1 ) + V (xi1 )
i=1 i=1 i=1
2m
N1 ( )N/2 { N ( )}
2m m (xi xi1 )2
= dxi exp i V (xi1 )
i=1
i=1
2

= Const D[x(t)]ei dtL(x,x) (3.12)

with L(x, x) the Lagrangian. In this form, S = dtL(x, x) is exactly the same
quantity as the classical action. This expression tells us that the possibility
amplitude of a particle at x0 (at time t = 0) to appear at position xN at
time t is given by a superposition of innite number of phase factors, each
of which is determined by the action of a classical path. One nice property
of path integral approach is that the classical limit is explicitly recovered.
In the derivation above we have omitted the Planck constant h. Recovering
h the phase factor in the integrand should be eiS/h . In the classical limit
of h 0 (which physically means action S is very large compared to h),
the main contribution comes from stationary point of S[x(t)], since all other
contributions are oscillating in very fast frequency. The stationary point of
S[x(t)] recovers the classical path correctly.

3.1.2 Imaginary time path integral

Another application of the path integral formulism is to the thermodynamics.


Generically, a system with Hamiltonian H at temperature T is described by
the following partition function
( )
Z = Tr eH (3.13)

with = 1/T the inverse temperature. The partition function describes the
thermodynamic properties of the system, such as the specic heat we com-
puted in the rst Chapter. One can easily see the similarity between eH
and eitH , so that eH can also be expressed in path integral, with it .
For the same single particle system
44 3 Path integral

( H ) d L(p,x,x)
Z = Tr e = D[p( )]D[x( )]e 0 (3.14)
( 2 )
p
L(p, x, x) = ipx + + V (x) (3.15)
2m
Due to the trace in the denition of Z, the path integral in this case is carried
over p( ), x( ) with periodic boundary condition. Dierent from the unitary
evolution case, L is not real. If we integrate over momentum p, the sign of
kinetic energy term is changed:

d 1 mx2 +V (x))
Z = Const D[x( )]e 0 ( 2 (3.16)

The path integral expressions (3.15) and (3.16) can be obtained from substi-
tution dt id in the real time expressions (3.10) and (3.12) respectively.

3.1.3 Harmonic oscillator and stationary phase approximation

As an example, we study the single harmonic oscillator with V (x) = 12 m 2 x2 .


We consider the imaginary time path integral

d 1 mx2 + 12 m 2 x2 )
Z = D[x( )]e 0 ( 2 (3.17)

The action is quadratic in x and x so that the integral can be exactly com-
puted. Take a fourier transform
1 in 2n
x( ) = e xn , n (3.18)
nZ

Then

Z= dxn e n
( 21 mn2 |xn |2 + 12 m2 |xn |2 ) (3.19)
n

The Gauss integral to all xn leads to


( 2
)1/2 1
Z= = Const. (3.20)
m ( 2 + n2 ) + in
nZ nZ

To simplify the expression of Z we notice that



log Z = log ( + in ) (3.21)
n

up to a constant, so that
log Z 1
= (3.22)
n
+ in
3.1 Single-particle path integral 45

The function
1
f () = (3.23)
n
+ in

has poles at = in = i2n/, from which we can determine the expres-


sion
( )
1 1
f () = + = coth (3.24)
e 1 2
2 2
(The constant 1/2 is determined by the requirement of no pole at = .)
Thus
( )
1
log Z = log sinh Z= (3.25)
2 sinh
2
This is consistent with the direct calculation of trace
[ ] [ ] e/2 1
Z = Tr eH = Tr e(n+1/2) = = (3.26)
1 e 2 sinh
2
It should be noticed that a constant in front of Z does not change any physical
quantity, which can always be omitted.
For more general potential V (x), the path integral is not Gaussian and
cannot be calculated rigorously. In the semi-classical limit, the most important
path is the classical path xcl (t), so that we can expand around xcl (t) as x(t) =
xcl (t) + x(t). The action can be expanded to the second order

S[x(t)] = S[xcl (t) + x(t)] = dtL(x, x)
[ ( ( ))
L L
dt L(xcl , xcl ) + t x
x x
( 2 ( 2 ) ( 2 ))]
L 2 2 L L
+ x + x + 2x x (3.27)
x2 x2 x x
where the second line vanishes according to Lagrange equation. In this expan-
sion, one obtains a quadratic approximate action for which the path integral
can be calculated. For real time, the result is
( [ 2 ( 2 ) ( 2 ) ])1/2
L L L
xf | U (t) |xi e iS(xcl )
Det + 2
t + 2 t (3.28)
x2 x2 x x
where the determinant comes from the Gaussian integral. For the action L =
2 mx V (x), the determinant is dened as
1 2

[ ]
Det mt2 V (x) = n
n

with n dened by the eigenvalues of the operator t2 V (x):


( 2 )
mt V (x) fn (x) = n fn (x)
46 3 Path integral

3.1.4 Double-well problem and instantons

As an interesting application of the path integral formulism, we consider a


particle in the double well potential V (x) with two minima at x = a. We
want to calculate the possibility of a particle tunneling from one well to the
other, given by

a| U (t) |a (3.29)

The stationary phase approximation does not apply since there is no classical
path connecting the two minima. This is an example where the property
we are interested (quantum tunneling) cannot be obtained from semi-classical
picture. The trick to solve this problem is to introduce a Wick rotation t i
which rotates the time-evolution operator U (t) = eitH to the imaginary time
evolution operator U ( ) = e H . In the imaginary time the Lagrangian
1
LE = mx2 + V (x) (3.30)
2
so that the barrier between the two wells becomes a potential trap, and there
is a classical path from a to a, satisfying the equation of motion

mx = x V (x) (3.31)

which is dierent from the real time equation of motion by a 1 on the right
side. This equation has a classical solution starting from a and ending at a.
According to the equation of motion

mxx = x V (x)x
(m )
d x2 = dxx V (x)
2
m
x2cl = V (x) (3.32)
2
in the last step we have used V (a) = 0 and xcl = 0 when x = a. Thus the
classical action
( ) a a
m 2
S= d x + V (xcl ) = dxmx = dx 2mV (x) (3.33)
0 2 cl a a

This proves that the action of this classical path is determined by the potential
V (x). Near x = a and x = a, V (x) is quadratic. Near x = a V (x)
1 2 2
2 m (x + a) , so that the dierential equation

m 2
x = V (x) xcl = a + ce (3.34)
2 cl
and similar for the neighborhood of a. Thus we see that the solution stays near
x = a and x = +a for most of the time, and only for a time window of the
3.1 Single-particle path integral 47

width 1/ the particle moves from a to a. Such an observation is very helpful


in the treatment of this problem. In the imaginary time, the simplest path
from a to a is thus such a fast transition which is localized in time. Such a
process is called an instanton. Since the instanton process happens in a time
window of 1/, if we are interested in much longer time scale there can be
processes of multiple instantons. If we make an analogy from the imaginary
time to a one-dimensional space, the instantons are like particles and the
multiple instanton process is like a gas of multiple particles. (The instanton
going from a to a are called anti-instantons.) The tunneling amplitude can
be obtained by summing over the instanton contributions. There are always
odd number of instantons so that the particle goes from a to a.


a| U ( ) |a A2n1 (3.35)
n=1

with A2n1 the 2n 1 instanton process given by


t1 t2n2
A2n1 = dt1 dt2 ... dt2n1 K 2n1 e(2n1)Sins
0 0 0
1 ( )2n1
= KeSins U0 ( ) (3.36)
(2n 1)!
a
where Sins = a dx 2mV (x) is the single instanton contribution to the ac-
tion. U0 ( ) is the evolution amplitude without instanton U0 ( ) a| U ( ) |a
which can be approximated by a harmonic oscillator. K is the determinant
from the stationary phase approximation. The integral over time is done over
2n 1 instantons and anti-instantons which are aligned alternatively at time
t1 , t2 , ...t2n1 . Thus

1 ( )2n1
a| U ( ) |a U0 ( ) KeSins
n
(2n 1)!
( )
= U0 ( ) sinh KeSins (3.37)

For long time ( 1/) we have



|a| n| e(n+1/2) e /2 |a| 0| e /2
2 2
U0 ( ) = a| U ( ) |a =
n

Thus
( )
a| U ( ) |a e /2 sinh KeSins (3.38)

Rotating back to the real time it we see


Sins
a| U (t) |a eit( 2 Ke

) + eit( 2 +KeSins ) (3.39)
48 3 Path integral

The two Fourier components with frequency 2 KeSins are the bounding
and anti-bounding states of the ground states of the two well.
Although in this simple problem the overlap can be obtained by simpler
approach, the instanton approach is generally more helpful in more compli-
cated theories such as eld theories with nontrivial instanton congurations.

3.1.5 Linear response of a harmonic oscillator

The path integral formulism is useful for both perturbative and nonpertur-
bative treatments to a quantum system. Although we will study more about
the perturbation theory in the next Chapter, its helpful to show how physi-
cal quantities are calculated in path integral formulism in a simple example.
We consider a single electron in a harmonic oscillator potential and apply
an external electric eld. We know that if we apply a static electric eld, a
charge dipole moment will be induced because the electron position is shifted.
However, if we apply a time-dependent electric eld and want to know the
response, it is not so straightforward to see the answer. In the following we
will show how to obtain the answer in path integral.
The system has the time-dependent Hamiltonian
p2 1
H= + m 2 x2 E(t)x (3.40)
2m 2
Here we have taken the electron charge e = 1. The electric eld has a generic
time dependence. The problem we want to solve is the charge polarization,
which is also time dependent and given by

P (t) = G(t)| x |G(t) (3.41)

The time-dependent state |G(t) is determined by Schordinger equation. If we


assume the electric eld is turned on gradually starting from time , the
time-evolution is
t
i H(t )dt
|G(t) = T e |G0 U (t, ) |G0 (3.42)

where |G0 is the E = 0 ground state, and the T labels time-ordering:


t
i H(t )dt
Te = lim eiH(t) eiH(t) eiH(t2) ... (3.43)
0

Time ordering is necessary since the time-dependent Hamiltonian at dierent


time does not commute. Then

P (t) = G0 | U (, t)xU (t, ) |G0 (3.44)

where U (, t) = U 1 (t, ). The linear response is an expansion of P (t)


to the linear order of the eld E(t). In general the linear response has the
form
3.1 Single-particle path integral 49

P (t) = dt D(t, t )E(t ) (3.45)

with
P (t)
D(t, t ) = (t t ) (3.46)
E(t )

with (t t ) the step function required by causality. To calculate this varia-


tion, we notice that

U (t, )
= U (t, t )ixU (t , ) (3.47)
E(t )

Thus

D(t, t ) = (t t ) [G0 | U (, t)xU (t, t )ixU (t , ) |G0


G0 | U (, t )ixU (t , t)xU (t, ) |G0 ] (3.48)

In Heisenburg picture we can absorb the time evolution into operators x:

x(t) = U (, t)xU (t, ) (3.49)

so that

D(t, t ) = (t t )i G0 | [x(t), x(t )] |G0 (3.50)

Eq. (3.50) is a very generic result. From the derivation we can see that this
kind of expression occurs for a generic linear response problem. In general, an
external eld F is included in the Hamiltonian H = H(F ). The response of
some physical quantity O to the external eld F is given by

O(t) = G| O(t) |G = O0 + dt D(t t )F (t )
[ ]
H(F )
D(t t ) = i(t t ) G| O(t), (t ) |G (3.51)
F

To relate this expression to path integral, we assume that the electric eld
is gradually turned on starting from t = , and also gradually turned o at
t +. Then the ground state |G0 will nally return to the ground state.
In other words,

U (+, ) |G0 = ei |G0 (3.52)

Thus we can write


G0 | U (+, t)xU (t, ) |G0
P (t) = (3.53)
G0 | U (+, ) |G0
50 3 Path integral

This is almost the transition amplitude calculated in path integral. However


we still need to input |G0 . To further simplify the calculation, we introduce
the following operator
t2
i dtH(t)(1i)
U (t2 , t1 ) = T e t1
, for t2 > t1 , = 0+ (3.54)

Here is a small positive number which is a regularization. To see the eect


of this modication, we take E = 0 case as a simplest example, in which case

U (t2 , t1 ) = ei(t2 t1 )H(1i) = ei(t2 t1 )H e(t2 t1 )H (3.55)

If we take t1 while keeping nite, we have

lim e(t2 t1 )H |G0 G0 | (3.56)


t1

since all excitation states are suppressed exponentially. When electric eld is
considered, but the electric eld goes to zero for t , the term does not
change anything during the time interval where the electric eld is important,
but it provides the projection to the ground state near , where electric
eld is absent and the discussion above applies.
Thus we can write
Tr [U (+, t)xU (t, )]
P (t) = (3.57)
Tr [U (+, )]

which can be written as path integral form. The denominator

Z[E(t)] = Tr [U (+, )]
{ + ( ( 2 ) )}
p 1
= D[x(t)]D[p(t)] exp i dt px + m 2 x2 E(t)x ei
2m 2
+
i dtL (t)
= Const. D[x(t)]e (3.58)
( )
1 1
L (t) = mx2 ei m 2 x2 E(t)x ei
2 2

Here we have used ei 1 + i for innitesimal . The numerator is only


dierent from the denominator by inserting a x at time t. It is helpful to write
it as a variation to the denominator:
+
i dtL (t)
Tr [U (+, t)xU (t, )] = D[x(t)]x(t)e

1 Z[E(t)]
= (3.59)
i E(t)

Thus
3.1 Single-particle path integral 51

1 Z[E(t)] log Z[E(t)]


P (t) = = i (3.60)
iZ[E(t)] E(t) E(t)

P (t) for any time t depends on the electric eld E(t ) for all time t . In other
words, it is a functional of E(t). For the harmonic oscillator, the path integral
is Gaussian, so that we can obtain the functional P (t) = P (t)[E(t )] rigorously.
However it is helpful to keep the discussion general so that it can be applied
to other systems. In general, we can expand P (t) to the linear order of E(t).
Since P = 0 if E = 0, the zeroth order vanishes. In the linear order,

P (t) dt G(t, t )E(t )

P (t) 2 log Z[E(t)]


with D(t, t ) = = i (3.61)
E(t ) E(t )E(t)

In other words, if we expand the functional Se [E(t)] = i log Z[E(t)], D(t, t )


is the second order coecient.

1
Se [E(t)] i log Z[E(t)] dtdt G(t, t )E(t)E(t ) (3.62)
2
The quantity Se is called eective action which governs the response of
the system to electric eld. Eq. (3.61) can be written as
( )
2 log Z[E(t)] 1 Z
G(t, t ) = i = i
E(t )E(t) E(t ) Z E(t)
[ ( )( )]
1 2Z 1 Z 1 Z
= i
Z E(t )E(t) Z E(t) Z E(t )
= i (T x(t)x(t ) x(t) x(t )) (3.63)

Here T is again the time-ordering, and

T x(t)x(t ) = G0 | (x(t)x(t )(t t ) + x(t )x(t)(t t)) |G0 (3.64)

with (x) the step function dened by (x) = 1 for x > 0 and (x) = 0 for
x < 0. It should be noticed that now all average values are calculated in the
ground state with E = 0 since the E dependence has been expanded to the
second order and so that the second order coecient G(t, t ) is calculated at
E = 0.
Now the question is how to relate the time-ordered Green function (3.64)
which can be obtained from the path integral and the response coecient
(3.50). D(t, t ) can be rewritten as

D(t, t ) = i(t t ) (G0 | x(t)x(t ) |G0 h.c.)


= 2ReGR (t, t ) (3.65)

with GR the retarded Green function dened by


52 3 Path integral

GR (t, t ) = G(t, t )(t t ) = i G0 | x(t)x(t )(t t ) |G0 (3.66)

Thus we can obtain the time-ordered Green function and take the retarded
sector to obtain the physical quantity we want.
For the harmonic oscillator, the eective action Se can be obtained ex-
plicitly by integrating over x(t).
{ [ ( ) ]}
1 1
Z[E(t)] = D[x(t)] exp i dt mx2 ei m 2 x2 E(t)x ei
2 2
{ [ ]}
1 ( )
= D[x ] exp i d m 2 ei 2 ei x x + ei E x
2
{ }
e2i
= Const. exp i d E E (3.67)
2m ( 2 ei 2 ei )

Here we have used 1 + i ei for innitesimal , just to save some space.


Thus

1
Se = i log Z[E(t)] = d E E (3.68)
2m ( e 2 e3i )
2 i

d 1
D(t, t ) = e3i 2i
ei (tt ) (3.69)
2
2 m ( e )
2

The overall e3i factor goes to 1 at 0+ limit and can be omitted, but
the e2i at the denominator cannot be omitted because the function has
singularity at = where the regularization is important.
( )
1 d 1 1

G(t, t ) = i
i
ei (tt ) (3.70)
2m 2 e + e
The rst term

1 d 1
GR (t, t ) = i
ei (tt )
2m 2 e
i i(tt )
= e (t t ) (3.71)
2m

The factor (t t ) appears because for t t > 0, ei(tt ) is analytic at
lower half plane of , so that the integration path can be closed at lower
half plane and obtain the result from the residual theorey. For t t < 0 the
integration path can be closed at upper half plane and the result is 0 because
the pole of the function is at ei below the real axis. Due to the factor
(t t ) this part is called retarded correlation function, or Green function.
Similarly the second term in Eq. (3.70)

1 d 1
GA (t, t ) = i
ei (tt )
2m 2 + e
i i(t t)
= e (t t) (3.72)
2m
3.2 Functional path integral 53

which is called advanced Green function.


Thus we have
1
D(t, t ) = 2ReGR (t t ) = sin (t t )(t t ) (3.73)
m
There is a nice expression of this relation between D(t, t ) and G(t, t )
in the Fourier transformed frequency space. This expression can be obtained
generally using the following general mathematical fact:

For a function f (t), t R, dene f () = d 2 f (t)e
it
. f () can be ana-
lytically continuated to the complex plane. Dene f () = fR () + fA (),
with fR () analytical on the top half plane, and fA () analytical on the
bottom half plane. Thus we have

fR (t) f (t)(t) = dtfR ()eit ,

fA (t) f (t)(t) = dtfA ()eit , (3.74)

For D(t, t ) = 2ReG(t, t )(t t ) we have

D(t, t ) = (G(t, t ) + G (t, t )) (t t )



= GR (t t ) + [GR (t t )] (3.75)

D() = GR () + [GR ()] (3.76)

On the other hand, the time-ordered Green function is symmetric G(t, t ) =


G(t , t) because of the time-ordering. (It should be noticed that this is only true
when the two operators in the Green function are the same.) Thus GR (t, t ) =
GA (t , t) and GR () = GA (). Consequently

D() = GR () + [GA ()] (3.77)

Compared to G() = GR ()+GA () we see that D() can always be obtained


from G() by taking conjugation to the advanced part GA (). In other words,
D() can be obtained by moving all the poles of G() above the real axis to
the mirror position below the real axis.

3.2 Functional path integral


3.2.1 Boson coherent state path integral

The path integral description of quantum mechanical system can be gener-


alized to many-body systems. For boson systems such as the coupled har-
monic oscillators, the generalization is straight forward. For example the one-
dimensional phonon problem with the Lagrangian density
54 3 Path integral
m
L (, ) = (3.78)
2
which corresponds to the path integral
f
f | U (t) |i = D[]e dtdxL
(3.79)
i

In more generic systems, it is convenient to do path integral in a dierent


basiscoherent state basis, which also helps the generalization to fermion sys-
tems. We start from a single harmonic oscillator. The path integral (3.10) is
integrated over both momentum p and position x. The harmonic oscillator
operator is dened by
( )
1 1
a= mx + i p (3.80)
2 m

Now in the path integral formulism x and p become classical numbers, so that
we can also group them into a and a . The transformation from x, p to a, a
is unitary, so that no additional factors appear from Jacobian. The harmonic
oscillator action can be transformed as

p2 m 1
(a a ) t (a + a )
2
L = px V (x) = i |a|
2m 2 2m
= a (it ) a (3.81)

In the last step, integration by part has been applied. One can check straight-
forwardly that this expression leads to the same partition function as Eq.
(3.26).
Such a unitary transformation from (x, p) to (a, a ) corresponds to a basis
change in the denition of path integral. Instead of inserting p and x eigen-
states, eigenstates of the operator a are used, which are called coherent states.
The coherent states are dened by

a | = | (3.82)

Since each state in the harmonic oscillator Hilbert space can be written as
| = A(a ) |0, the function A(a ) satises
[ ]
a, A(a ) = A(a ) (3.83)
[ ]
Due to a, a = 1 we have the general expression

[ ]
a, A(a ) = A(a ) (3.84)
a
Thus the function A(a ) satises the dierential equation
3.2 Functional path integral 55


A(a ) = A(a )
a

A(a ) = Cea (3.85)

The state

1
| = Cea |0 = C n |n (3.86)
n=0 n!

which determines the normalization factor

C = e 2 ||
1 2
(3.87)

The coherent states are not orthogonal to each other:


1 ||2 +||2 ||2 +||2
( ) e 2
n
| = = e 2 (3.88)
n
n!

However the coherent states form an overcomplete basis



2 1
d d | | = d de|| n m |n m|
n,m n!m!
1
d de|| || |n n|
2 2n
=
n
n!

= |n n| = I (3.89)
n

By inserting the basis



d d
I= | | (3.90)

we can obtain the path integral of a generic harmonic oscillator system with
the Hamiltonian H = H(a, a ). Dene H in normal order so that a is always
on the left of a, then
N
1
dn dn ( )
N
f | U (t) |i = n | I iH(a , a) |n1
n=1
n=1
t
i 0 dt ( t H( ,))
= DD e (3.91)

Such an expression can be easily generalized to multiple harmonic oscillators,


and thus the quantum bosonic eld theories, such as Maxwell theories.
56 3 Path integral

3.2.2 Fermion coherent state path integral

Similar coherent states can be dened for fermions, so as to obtain a path-


integral expression of fermion systems. We consider the eigenstates of fermion
annihilation operator:

c | = | (3.92)

Since for fermion there are only two states, | = a |0 + b |1, then

c | = b |0 = (a |0 + b |1) (3.93)

From this equation it seems that the only eigenstate of c is |0 with eigenvalue
0. To obtain a nontrivial coherent state for fermions, we dene the formal
Grassman numbers which satisfy the Grassman algebra: = , 2 =
0. Such an algebra allows a nontrivial solution of the eigenvalue equation:
b = , a = 1. The state
( )
| = |0 |1 = 1 c |0 = ec |0 (3.94)

is the eigenstate of c. (It should be noticed that c and also anti-commute.


) The overlap

| = + 1 (3.95)

To normalize these states one can introduce a factor


( )
1
| = 1 |0 |1 = e /2 ec |0 (3.96)
2
To use these eigenstates in the path integral we dene the integral over Grass-
man number

d = 0, d = 1 (3.97)

and consider and as completely independent variables. Thus



d d | | = d d [(1 /2) |0 |1] [0| (1 /2) 1| ]

=I (3.98)

In this derivation it should be noticed that dd = d d. For Grassman


number, dierentiation and integral are equivalent.
In this formulation, a fermion system with Hamiltonian H = H(c , c)
(normal ordered in the same way as boson case) has the path integral repre-
sentation
t
dt ( t H( ,))
f | eitH |i = D[(t)]e 0
i
(3.99)
3.2 Functional path integral 57

Similar expression can be obtained for imaginary time path integral repre-
sentation of eH . However there is one important dierence from the boson
case
[ when] we consider the imaginary time path integral representation of
Tr eH . We have
[ ]
Tr eH = 0| eH |0 + 1| eH |1

= d d | eH | (3.100)

Compared to the overcompleteness condition (3.98), there is an additional


1 sign coming from exchanging and in | and |. Consequently,
the path integral representation is integrated over ( ) with anti-periodic
boundary condition () = (0). This is the key dierence between boson
and fermion.
Similar to boson case, the Gauss integral for fermions can be obtained
explicitly:


d dea = d (1 a ) = a (3.101)

For multiple components of fermions this generalizes to



Aij j
di di e i,j i = det A (3.102)
i

For example, the Fermionic analogy of harmonic oscillator with

H = c c (3.103)

has the partition function



( +)
Z= D De 0


= dn dn e n
n n (in +)

n

= (in + ) (3.104)
n
(2n 1)
with n = ,n Z (3.105)

Notice the frequency shift of / compared to the boson case, which leads to
dierent behavior of the partition function

Z = 1 + e (3.106)
4
Perturbation Theory and Beyond

4.1 Discussion on the convergence of perturbation theory


The free boson and free fermion theories can be solved exactly by diagonal-
ization. In the path integral formulism, they corresponds to the Gaussian
integral which can be carried out
( exactly. For
) example a free boson theory
with Lagrangian density L = 2m 1
2 has the partition function
[ ( )]1/2
d dxL 1 2
Z= De 0 = Det (4.1)
2m
If there is a non-harmonic term, for example
( )
1 2 4
L = + g || (4.2)
2m
If we interpret this theory as the massive phonon modes for two chain system
we studied in Homework 1, such an 4 term corresponds to a non-harmonic
interaction between the atoms in the two chains.
With the 4 term we cannot obtain the partition function and other phys-
ical quantities rigorously. Thus the simplest thing is to expand the partition
function Z(, g) in g, which is the perturbation theory that we are going to
study in this Chapter. However, this expansion is a bit tricky. To understand
whether this perturbation is well-dened, we study a simple example by re-
placing the path integral by a nite dimensional integral of the same form.
Consider the integral

Z(g) = dxe 2 x gx
a 2 4
(4.3)

Naively we can expand




1 ( )n
dxe 2 x
a 2
Z(g) = gn x4 fn g n (4.4)
n=0
n! n
60 4 Perturbation Theory and Beyond

The coecient
( )
n
(1) a x 2
4n (1)n 2n a 2
x 2n (1)n 22n 2n 2
fn = dxe 2 x = dxe 2 (2) =
n! n! a2n n! a2n a

2 2n (4n 1)!!
= (1)n a (4.5)
a n!
To see if this expansion converges, we notice that the error of n-th order
expansion is

n 2 ( )m
m m 1 a
Dn+1 (g) = Z(g) fm g = g dxe 2 x
x 4

m!
m=0 m=n+1

1 a 2 ( )n+1
g n+1 dxe 2 x x4 = g n+1 |fn+1 |
(4.6)
(n + 1)!
The function Dn (g) vs n is not monotonously decreasing:
g
log Dn (g) = n log 2 + log ((4n 1)!!) log n!
a
g ( )
= n log 2 + log ((4n 1)!) log 22n1 (2n 1)! log n!(4.7)
a
A constant has been omitted in the last step. Use the asymptotic formula
log(n!) n (log n 1) we have
( g)
log Dn (g) 6 log 2 + log 2 n + n (log n 1) (4.8)
a
Thus
( g)
log Dn (g) 6 log 2 + log 2 + log n (4.9)
n a
Consequently the minimal occurs at
26 a2
n= (4.10)
g
where
26 a2
Dn = en = e g (4.11)
This derivation shows that the error of the expansion decreases until the order
6 2
of n = 2 g a , and then starts to increase again. It is understandable that the
perturbation theory does not converge, because we see that the function Z(g)
6 2
diverges for g < 0. However if 2 g a 1, the minimal error we can get is
exponentially small. Since high order perturbation theory is usually dicult,
for small g all calculations we can do will be safely in the perturbative
region where the expansion look like converging. However for large value of
g, this derivation above shows that the expansion does not make sense even
if you can calculate to high order of perturbation.
4.2 More general cases and Feyman diagram 61

4.2 More general cases and Feyman diagram


To see the general structure of the perturbation theory, we consider a simple
generalization of the 1d integral given above:

n
dxi e 2 ai x2i gijkl xi xj xk xl
1
Z(g) = i (4.12)
i=1

Here x becomes a vector and g becomes a rank-4 tensor which determines


the deviation from Gaussian integral. The same perturbative expansion can
be made as above:
2 (g
n n
ijkl xi xj xk xl )
dxi e 2 i ai xi
1
Z(g) =
i=1 n
n!
n
1 n
( )
= Z(0) gi4a3 i4a2 i4a1 i4a xi4a3 xi4a2 xi4a1 xi4a
n
n! a=1 a=1
(4.13)

Here we have dened the average value


n
1
n 2
n
dxi e 2 i ai xi
1
xi4a3 xi4a2 xi4a1 xi4a = xi4a3 xi4a2 xi4a1 xi4a
a=1
Z(0) i=1 a=1
(4.14)

Instead of directly calculate the integral which is not as easy as the single
avor case, we introduce a general formulism:

1 ( 4
)n n
21

ai x2i +

Ji xi
Z(g) = gijkl dxi e i i
n
n! Ji Jj Jk Jl
i=1
n
J=0
1 n 4n
2 J /2ai
2
= gi4a3 i4a2 i4a1 i4a e i i
n! J i b i=1
a i
n a=1 b=1
J=0
1 ( ) Ji2 /2ai
n 4n
= Z(0) gi4a3 i4a2 i4a1 i4a e i (4.15)
n
n! Jib
a=1 b=1 J=0

In this way, each term in the expansion is a derivative to the Gaussian function
of the source eld Ji . According to the property of the Gaussian function,
the derivative is only nonzero if the derivative to each Ji is done even number
of times.


Ji2 /2ai
4n
i1 i2 i3 i4 i4n1 i4n
e i = ... + permutations (4.16)
Jib ai1 ai3 ai4n1
b=1 J=0
62 4 Perturbation Theory and Beyond

where permutations stands for all other ways of pairing up the 4n indices
ib into 2n pairs. This is the Wick Theorem:
2n
n
1
xib = ic ic (4.17)
a
c=1 ic
b=1 pairing

Here a pairing is a way to pair the 2n labels ib into n pairs (ic ic ).


Applying the Wick theorem, the perturbative expansion (4.15) can be
represented by Feynman diagrams. The 4n indices of n interaction terms need
to be paired up. Thus we can simply draw each interaction gijkl as a vertex
with four lines meeting. Pairing up the indices is done by connecting lines
pairwise, as shown in Fig. 4.2 (a). For each of such diagrams D we have 1/ai
for each line, and gijkl for each vertex, as shown in Fig. 4.2 (b). The factor
n! can be canceled by the permutation of dierent interaction vertices. In
other words if we dene the diagrams without a labeling of the vertices, the
perturbation expansion becomes

Z(g) = Z(0) V (D) (4.18)
Diagrams D

where V (D) = Ivertex of D giI jI kI lI IJ link of D a1i .
I
One interesting observation from the diagrammatic point of view is that
a diagram which is disconnected denoted by D = D1 D2 simply contributes
as V (D) = V (D1 )V (D2 ). Consequently each connected diagram is counted
many times in higher order. This can be written as an equation


V (D) = exp V (D) (4.19)

Diagrams D connected diagrams D

Consequently

log Z(g) = log Z(0) + V (D) (4.20)
connected diagrams D

The calculation above can be generalized to correlation functions. For example


the two-point function

2 log Z(g, J)
xi xj xi xj = (4.21)
Ji Jj J=0

Similar diagrammatic expansion can be made, but with two more derivatives
in each term. This corresponds to the diagrams with external lines which
are not contracted.
2
Z(g, J)|J=0 = V (Dij ) (4.22)
Ji Jj
Dij
4.2 More general cases and Feyman diagram 63

Fig. 4.1. Feynman diagrams.

with V (Dij ) the diagram with external lines i, j, as shown in Fig. 4.2 (c). In
general, we have the Taylor expansion
[ ]
1 1 ( ) Ji2 /2ai
m n 4n

Z(g, J) = Jj Z(0) gi4a3 i4a2 i4a1 i4a e i
m
m! a=1 a Jja n
n! a=1 Jib
b=1 J=0
1 m
= Jj V (Dj1 j2 ...jm ) (4.23)
m
m! a=1 a
D

Similar to the case without external lines, we can also relate the diagrammatic
sum to summation over connected diagrams:
64 4 Perturbation Theory and Beyond

1 m
log Z(g, J) = Jj V (Dj1 j2 ...jm ) (4.24)
m
m! a=1 a
connected diagrams

so that any correlation function can be calculated by connected diagrams



n log Z(g, J)
= V (Di1 ...in ) (4.25)
Ji1 Ji2 ...Jin J=0
connected Di1 ..in

The discussion above can be easily generalized to the physically interesting


case of functional path integral. For example for the theory in Eq. (4.2), we
can diagonalize the quadratic part by Fourier transformation, so that

(i+p2 /2m)|p |2
dp dp e
d dxL
Z(g = 0) = De 0 = ,p

,p
(4.26)

The interaction term is transformed to



1
1 p1 2 p2 3 p3 4 p4 ei(1 +2 3 4 ) (p1 +p2 p3 p4 )x
4
d dx || = 2
d dx
(L) i ,pi
1
= p p (p1 + p2 p3 p4 )(1 + 2 3 4 )
L ,p 1 p1 2 p2 3 3 4 4
i i

(4.27)

Thus the lines in the Feynman diagrams are labeled by Green functions
1
Gp = p2
(4.28)
i + 2m

and the interaction vertex is given by


g
gpi i = (p1 + p2 p3 p4 )(1 + 2 3 4 ) (4.29)
L
with the function determined by the energy and momentum conservation
law. It should be noticed that the lines in the Feynman diagram now has an
arrow, to distinguish and . The Feynman rule in this case is summarized
in Fig. 4.2 (d).
In the discussion above we have been using imaginary time path integral.
For the real time the formulas look similar, but we need to remember the
regulator ei we introduced before. The lines correspond to the time-ordered
Green function
1
Gp = p2 i
(4.30)
2m e
4.3 Perturbative approach to interacting electron gas 65

4.3 Perturbative approach to interacting electron gas


Instead of studying the boson theory above in more details, we study the
interacting electron gas as a more physical example to apply the perturbation
theory. The perturbation theory and diagrammatic representation above can
be generalized to Fermion system straightforwardly.
n
di di e 2 i ai i i gijkl i j k l
1
Z(g) =
i=1

1 ( 4
)n n


i i +i i )
= gijkl di di e i
ai i i +
i
(
n
n! i j k l
i=1
J=0
1 n
( )
2n
i i /ai
= Z(0) gi4a3 i4a2 i4a1 i4a e i (4.31)
n
n! a=1 ib jb
b=1 J=0

It should be noticed that we have to use and in the quadratic term rather
than 2 because i2 = 0. The source terms i is coupled to the Grassman eld
i so that i must be also Grassman-valued.
The derivative to Gaussian function can be calculated by Wick theorem in
the same way as boson case, but the Fermi sign needs to be treated carefully.
Since the derivative operators /i and / i are all anti-commuting to each
other, the dierent terms in the Wick theorem expansion is assigned dierent
sign:



n
n n
1
ia jb = (1)NP ic P (jc ) (4.32)
a
c=1 ic
a=1 b=1 P pairing

Here P labels the way to pair the indices ia of i with the indices jb of i . NP
is the number of permutations for a given pairing P respecting to a reference
conguration. P (jc ) stands for the label which is paired with ic . For example,

1 2 3 4 = 1 4 2 3 1 3 2 4
1
= (14 23 13 24 ) (4.33)
a1 a2
Now we apply this mechanism to the electron gas with Coulomb interac-
tion. In the coherent state path integral,

Z = DDeS[,]
{ [ ( ) ]
2
S[, ] = d d x +
3
(x)
0 2m
}
1
+ d xd yV (x y) (x) (y) (y) (x)
3 3
(4.34)
2
66 4 Perturbation Theory and Beyond

After Fourier transform


( )
p2 1
S[, ] = p in + p + 3
p+q, p q, V (q)p p,
2m 2L
p,n , p,p ,q,,

with V (q) = d3 xV (x)eiqx (4.35)

Here we have noted momentum p and frequency n = (2n1)/ together as


p. The momentum and energy conservation condition in the interaction term
has been explicitly assured. Similar to the discussion for the boson eld theory,
the perturbation theory to the interacting electron gas can be described by
Feynman diagrams. Each line stands for the free fermion Greens function
1
G(p, in ; ) = (4.36)
in + p2 /2m

The line is labeled by momentum, energy and spin =, . The interaction


term is a degree-4 vertex with the value of V (q). The Feynman rule is
summarized in Fig. 4.3.1 (a).

4.3.1 Ground state energy

Now we calculate the correction of the interaction to the free energy:


1 1
F = log Z = V (D) (4.37)

Dconnecteddiagrams

To the leading order of the perturbation theory, we consider the diagrams


with only one interaction vertex, which include the two diagrams as shown in
Fig. 4.3.1 (b). It should be noticed that these two diagrams are not equivalent
because the outer two lines have the same spin index while the inner two
lines have a dierent spin . The two diagrams correspond to the values
1
V (D1 ) = 3
V (q = 0) G(p, in , )G(p , in , )
2L p,p
1
V (D2 ) = V (q)G(p + q, in , )G(p, in , ) (4.38)
2L3 p,q,in ,

Going back to the real space, the rst term is proportional to d3 xd3 yV (x
y) (x) (y) which only depends on V (q = 0) because (x) is independent
from x in a uniform system. This term is called Hartree term which only plays
a role in an inhomogeneous system.
The second term is called Fock term which is non-vanishing and gives a
correction to the free energy
4.3 Perturbative approach to interacting electron gas 67

1
F (1) = V (q)G(p + q, in + in , )G(p, in , )
2 2 L3
p,q,in ,in ,
1 1 1
= V (q) (4.39)
2 L3 i (n + n ) + p+q in + p
p,q,in ,in

in which the sum over spin simply gives a factor of 2, and p p2 /2m .
Using
1
= nF (), (4.40)
in + e +1
n =(2n1)/

with nF () = e1+1 the Fermi distribution function, the sum over frequency
can be done explicitly:
1
F (1) = nF (p+q )nF (p )V (q) (4.41)
L3 p,q

In the low temperature limit nF (p ) = (p ) becomes the step function


which is 1 for p2 /2m < 0 or equivalently |p| < kF = 2m, and 0
otherwise. Thus

d3 p d3 p
F (1) = L3 3 3
V (p p ) (4.42)
|p|<kF (2) |p |<kF (2)

For Coulomb interaction



e2 4e2
V (q) = d3 xeiqx = (4.43)
|x| |q|
2

which leads to
e2 L3 kF4
F (1) = (4.44)
(2)4
To understand this result more physically, we compare it with the free electron
kinetic energy
d3 p p2
kF
p4 kF5
F (0) = L3 3
= L 3
dp 2
= L3 (4.45)
|p|<kF (2) 2m 0 2m 10m 2

(One may wonder whether there is an ambiguity in the denition of the kinetic
energy because the arbitrary choice of zero point. The zero point is dened
with reference to the band bottom, i.e the lowest energy of each fermion. )
Thus the ratio between interaction and kinetic energy is given by
(1)
F 5e2 m
rs = (4.46)
F (0) 8 2 kF
68 4 Perturbation Theory and Beyond

As has been discussed in the second chapter, the electron density is given by
the volume enclosed by the Fermi surface:

4 1 kF3
ne = 2 kF3 3 = (4.47)
3 (2) 3 2

Thus the average distance between two electrons


( )1/3 1
r0 = n1/3
e = 3 2 kF (4.48)

so that
r0
rs =
a0
8 2 ( 2 )1/3 1
with a0 = 2 3 2 (Bohr radius) (4.49)
5e m e m
The comparison of the length scale a0 and the average distance r0 determines
whether the interaction energy is large or small compared to the kinetic en-
ergy. Qualitatively, we can understand this result intuitively. The interaction
energy per particle F (1) /L3 ne e2 /r0 while the kinetic energy per particle
1/3
F (0) /L3 ne kF2 /2m r02 , which leads to the ratio rs r0 ne . Conse-
quently, the lower is the density of electrons, the stronger is the interaction.
Interestingly, from this estimation one can see that this behavior rs r0 is
independent from spatial dimension, but determined by the parabolic disper-
sion p = p2 /2m. For graphene with linear dispersion, F (0) /L3 ne kF r01
so that the ratio rs is independent from the electron density.

4.3.2 Perturbative theory of the dielectric response

Besides the ground state energy, one can also compute other physical proper-
ties of the interacting electron gas. In the last chapter we have computed the
linear response of the harmonic oscillator to an external eld. Similarly, we can
compute the charge polarization of an interacting electron gas induced by an
external electric eld. The interaction to electric eld is given by the coupling
to the scalar potential. Now we need to use the real time path integral:

Z[] = DDeiS[,]
{ [ ( ) ]
2
S[, ] = dt d3 x it (x, t) (x)
2m
}
1
d3 xd3 yV (x y) (x) (y) (y) (x) (4.50)
2

To calculate the generic dielectric response isequivalent to calculate the re-


sponse of charge density (x, t) = (x, t) to a generic potential (x, t):
4.3 Perturbative approach to interacting electron gas 69

Fig. 4.2. (a) Feynman rule for interacting electron gas. (b) First order correction
to the ground state energy. (c) Feyman diagrams for the density-density correlation
function defined in Eq. 4.53. (d) The diagrams included in RPA approximation.


(x, t) dt d3 x D(x x , t t )(x , t ) (4.51)

As derived in Eq. (3.51), the response coecient is given by


[ ]
D(x x , t t ) = i(t t ) (x, t), (x , t )
= 2(t t )Re(x x , t t ) (4.52)

with

(x x , t t ) = i T (x, t) (x , t ) (4.53)

the time-ordered density-density correlation function. (x x , t t ) can be


calculated from the propagator Z[] as

2 log Z[]

(x x , t t ) = i (4.54)
(x, t)(x , t ) =0

After Fourier transform


70 4 Perturbation Theory and Beyond

d3 pdd3 p d

(q, ) = i T
p+q,,+, p,, p q, , p , ,
(2)4
(4.55)

Similar to the boson case discussed in the last subsection, such correlation
function can be expressed as a summation over connected diagrams with ex-
ternal lines. Now in the action (4.50) there are two kinds of interaction vertices,
the degree-3 vertex between and and the degree-4 vertex from Coulomb
interaction. We rst consider the leading order term contributed by the free
electron gas, as shown in the rst diagram in Fig. 4.3.1 (c).
3
d pd
(q, ) = 2i
(0)
G(p + q, + )G(p, ) (4.56)
(2)4

In the derivation above we havent written explicitly the regulator ei as we


did in Sec. 3.1.5. The only place that the regulator needs to be considered is
in the time-ordered Greens function:
1
G(p, ) = (4.57)
ei p

Thus

d3 p 1 1
(0) (q, ) = 2i d
(2)4 + ei p+q ei p
[ ]
d3 p 1 1 1
= 2i d
(2)4 + ei p+q ei p ei (p+q p )

d3 p (p+q ) (p )
= 4 (4.58)
(2)4 ei (p+q p )

(0) (q, ) can be decomposed into retarded and advanced parts

(0) (q, ) = (0)R (q, ) + (0)A (q, ) (4.59)



( p+q ) (p )
(0)R (q, ) = 4 d3 p (p+q p ) (4.60)
ei (p+q p )

(p+q ) (p )
(0)A (q, ) = 4 d3 p (p+q + p )
ei (p+q p )
= (0)R (q, ) (4.61)

The response function

D(x x , t t ) = 2(t t )Re(x x , t t )


( )
D(q, ) = R (q, ) + R (q, )
( )
= R (q, ) + A (q, ) (4.62)
4.3 Perturbative approach to interacting electron gas 71

Compare the two equations above we see that the only dierence between
D(q, ) and (q, ) is that the position of poles in the retard part is above
the real axis in while below the real axis in D. Thus we can simply obtain
D(q, ) by replacing the regulator ei (p+q p ) by (p+q p i). To
the leading order

d3 p (p+q ) (p )
D(0) (q, ) = 4 (4.63)
(2)4 + i (p+q p )
Such procedure is also applicable to a generic correlation function: the re-
sponse function is related to the time-ordered correlation function by moving
all the poles of the time-ordered correlation function to the lower half plane.
p2
D(0) (q, ) for the free Fermi gas with p = 2m is the Linhard function.
At low |q| we have

d3 p p p+q
D (q, ) 4
(0)
(p )
(2) 4 + i (p+q p )

kF2 kF q cos
= d sind
3
4 mvF + i kF qmcos
[ ]
kF m 1 + i
= d (cos ) 1
2 2 1 + i kF qmcos
[ ( ) ]
mkF + i + vF q
= log 1 (4.64)
2 2vF q + i vF q
In the last step we have used vF = kF /m and also omitted the innitesimal
in the prefactor + i of the log function, since its contribution is innites-
imal. On comparison, the eect of the regulator in the log function cannot
be ignored. Dierent limits of the function D(0) (q, ) can be studied. For
example, if is nite and q 0, the asymptotic behavior is
( )
+ i + vF q 2vF q 2 ( vF q )3
log +
+ i vF q 3
k F m v F q )2
(
D(0) (q, ) (4.65)
3 2
D(q, ) determines the response of the electron gas to electric eld. More
specically, consider the dielectric constant of the system. In the Fourier space,
we have

(q, ) = D(q, )(q, )


eE(q, ) = iq(q, )
D(q, )
(q, ) = ie q E(q, )
q2
D(q, ) 1
P(q, ) = e2 2
E(q, ) V (q)D(q, )E(q, ) (4.66)
q 4
72 4 Perturbation Theory and Beyond

with P(q, ) the charge polarization satisfying P(r) = (r) so that


(q) = iq P(q). To relate this equation to dielectric constant, we need
to remember that E(q, ) here is the external electric eld acting on the
electron system (which is usually denoted by D but we wont do that here, to
avoid confusion with the response function D(q, ). The total electric eld is

Etot (q, ) = E(q, ) 4P(q, ) (4.67)

Thus the dielectric constant is


E(q, ) 1
(q, ) = = (4.68)
Etot (q, ) 1 + V (q)D(q, )

For the electron gas we have


[ ]
1 4kF me2 ( vF )2
lim 1 = V (q)D(q, ) = (4.69)
|q|0 (q, ) 3

We see that the dielectric constant 0 when frequency 0. This means


that a small external electric eld can generate an arbitrarily large charge
polarization. Consequently electric eld cannot actually penetrate into the
system, which is the expected behavior of a metal.
The next terms in the perturbation theory contain one vertex of Coulomb
interaction, which contain two connected diagrams. One of the diagrams is
shown in the second diagram in Fig. 4.3.1 (c). Interestingly, the summation
over frequency and momentum in the two bubbles are completely indepen-
dent, so that the contribution of this diagram can be simply given by

(1) (q, ) = (0) (q, )V (q) (0) (q, )


D(1) (q, ) = D(0) (q, )V (q)D(0) (q, ) (4.70)

The small q behavior remains the same since the 1/q 2 in V (q) cancels the q 2
in D(0) (q, ). The correction to the dielectric constant is
[ ] ( )
1
lim 1 = V (q)D(0) (q, ) 1 + V (q)D(0) (q, )
|q|0 (q, )
( )
4kF me2 ( vF )2 4kF me2 vF2
= 1+ (4.71)
3 3 2

4.3.3 Beyond perturbation theory: RPA approximation

Due to the simple form of the rst order perturbation correction we obtained
above, such perturbation can be generalized to higher order. A series of dia-
grams with N bubbles can be summed, as shown in Fig. 4.3.1 (d). The result
is simply
4.3 Perturbative approach to interacting electron gas 73
[ ]2
DRP A (q, ) = D(0) (q, ) + D(0) (q, )V (q)D(0) (q, ) + D(0) (q, ) V (q)D(0) (q, ) + ...
D(0) (q, )
= (4.72)
1 V (q)D(0) (q, )
The summation over this particular series of diagrams is called Random Phase
Approximation (RPA). The dielectric constant in this approximation is
1
RP A (q, ) = = 1 V (q)D(0) (q, ) (4.73)
1 + V (q)DRP A (q, )
Thus
4kF me2 ( vF )2 p2
RP A (q = 0, ) = 1 1 2 (4.74)
3
Compared RP A with (0) and (1) from the perturbation theory, the key
dierence is that RP A has a sign change at some nite frequency named as
plasmon frequency:
4kF mvF2 e2 4ne e2
p2 = = (4.75)
3 m
For < p the dielectric constant is negative. Since the photon in the material
with dielectric constant and permeability needs to satisfy the dispersion
c2
relation 2 = (q,)(q,) q2 , for < 0 the electro-magnetic eld has an
imaginary wavevector, which cannot penetrate into the system. For > p ,
electromagnetic eld can penetrate into the system. In other words, for light
with frequency < p the electron gas is non-transparent and all lights are
reected (as we expect for a perfect metal). For light with frequency > p
the electron gas is transparent. The measurement of plasmon frequency p is
a helpful experimental way to determine the carrier density ne in the system.

4.3.4 Eective eld theory approach and the plasmon mode


The RPA approximation and the physical meaning of the plasmon frequen-
cy can be understood more physically in a dierent approachthe eective
eld theory approach. In the treatment above, the density operator (r) is a
composite operator, i.e., a function of electron creation and annihilation op-
erators. For the understanding of the electron density uctuation, it is more
convenient to treat the density eld (r) as an independent eld, just like
what we did in the bosonization approach of 1d fermions. The path integral
formalism is most suitable for this purpose. To see how this works, we rst
notice a simple identity:
[ ]
dy 4g 2
(y2igx2 ) egx4
egx =
4 1
e
4g

dy 4g 1 2
y +iyx2
= e (4.76)
4g
74 4 Perturbation Theory and Beyond

This identity holds because the bracket [..] in the rst line is equal to 1. Such
a transformation is helpful because it transforms the non-harmonic term x4
into a coupling between x and a dierent eld y. Such a transformation is
called Hubbard-Stratonovic (HS) transformation. Now we apply this to the
interacting electron gas. The interaction term in the action in Eq. (4.50) can
be written as

1
Sint = dt d3 xd3 yV (x y) (x) (y) (y) (x)
2

1
= d3 qdV (q)q, q, (4.77)
2

d3 pd
with q, = p,, p+q,+,
(2)4
(4.78)
Using the HS transformation, we have
3
eiSin = e 2 d qdV (q)q, q,
i

{ ( )}
1
= const. D [q, ] exp i d3 qd q, q, + q, q,
2V (q)
(4.79)
We rst study the propagator without the external source term in Eq.
(4.50), which can be rewritten as

Z = DDDeiS[,,]
[ ( 2 )]
p
S[, , ] = d3 pd p, p,
2m
( )
1
+ d3 qd q, q, + q, q,
2V (q)
[ ( 2 )]
p 1
= d3 pd p, p, + d3 qd q, q,
2m 2V (q)

d3 pd
+ d3 qd q, p,, p+q,+, (4.80)
(2)4
Thus we see that by HS transformation we have obtained an eective action
which describes the fermion eld , and the boson eld with a cubic
interaction term. For Coulomb interaction, V (q) 1/q2 so that the eld
has a kinetic energy term 1/2V (q) q2 . Physically, is nothing but the
scalar potential, which can be also seen from its coupling with the electrons.
The coupling of to fermions is the same as the coupling of the external
potential with fermions in Eq. (4.50).
From this equation we can see the advantage of HS transformation: the
action is now quadratic in the electron elds , so that we can carry out the
4.3 Perturbative approach to interacting electron gas 75

Gauss integral over fermions to obtain an eective action of the boson


eld

eiSeff [] = DDeiS[,,] (4.81)

We can denote

1
Se [] = d3 qd q, q, + S0 [] (4.82)
2V (q)

since the rst term is independent from , . The second term S0 is obtained
by integrating out fermions. Since couples with the fermion in the same way
as an external scalar potential, eiS0 is actually the same as Z[] in Eq. (4.50)
with replaced by and interaction term switched o. Consequently, the
quadratic term in S0 is determined by the bare (non-interacting) correlation
function (q, ) calculated above:

1
S0 [] = d3 qdq, q, (0) (q, ) + o(2 ) (4.83)
2
Thus the eective action of the boson eld is
[ ]
1 1
Se [] = d3 qdq, q, (0) (q, ) (4.84)
2 V (q)

In the limit we studied in last subsection |q| 0 with nite, we have

kF m ( vF q )2
(0) (q, ) = D(0) (q, ) (4.85)
3 2
so that
1 q2 kF m ( vF q )2
(0) (q, ) 2

V (q) 4e 3 2
( )
2
q2 p
= 2
1 2 (4.86)
4e

To see the physical meaning of this action, we rescale q, eld and dene
q, = |q|
q, , the action becomes a standard free boson action:

1 ( )
Se = 2
d3 qdq, 2 p2 q, (4.87)
8e
Thus we see that there is a well-dened collective mode with frequency p
propagating in the system, which is the plasmon. Physically, this mode is
a charge density uctuation, because the eld we introduced through HS
transformation is related to the charge density q, by q, = V (q)(q, ).
76 4 Perturbation Theory and Beyond

We can also directly see the relation of the plasmon mode to the dielectric
constant. Since q, is the scalar potential, the eective action can be written
in term of electric eld eEq, = iqq as
[ ]
1 e2 1
Se = d3 qdEq, Eq, 2 (0) (q, )
2 q V (q)
[ ]
1
= d3 qdEq, Eq, 1 V (q) (0) (q, )
8

1
d3 qdEq, Eq, q, (4.88)
8
from which we can see that the dielectric constant agrees with the RPA result
in Eq. (4.73), when the time-ordered correlation function (0) is replaced by
the retarded correlation function D(0) .
Further remarks on the eective eld theory approach
The eective eld theory approach we take above is useful in many physical
systems, and is one of the main advantages of the path integral formulation.
By introducing new elds through HS transformation (or using other integra-
tion identities) one can treat the collective modes such as density uctuation
and the fundamental degrees of freedom such as electrons at equal footing, and
obtain an eective description by integrating out some of the elds. For ex-
ample, in the ferromagnetic and anti-ferromagnetic spin models we discussed
in Chapter 2, the spins are formed by localized electrons, which is a collective
mode just like the charge density discussed here. In Chapter 2 we treat spins
as if they are fundamental degrees of freedom. In a more complicated system,
for example a system with both localized electrons and itinerant electrons
coupling to each other, one need to derive the spin model from the funda-
mental interacting electron model, where the same idea of HS transformation
and eective eld theory can be used. Compared to the Hamiltonian approach
where we deal with Hilbert space, Schrodinger equation and operators, the
path integral approach has the advantage that Hilbert space is not explicit-
ly used, so that it is easy to change variables by inserting new elds, which
eectively has changed the Hilbert space we are dealing with. For example
in the interacting electron gas we just studied, we obtain a boson eective
theory from a purely fermionic system. To be more precise, if we return to
the Hilbert space and Hamiltonian language, what we did is actually nding a
boson system which has the same dynamics as the charge density uctuation
of an interacting fermion system.

You might also like