Professional Documents
Culture Documents
(A Work in Progress)
Daniel Arovas
Department of Physics
University of California, San Diego
March 14, 2010
Contents
0.1
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
0 Introductory Information
0.1
vi
1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 Boltzmann Transport
2
5
1.1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
1.3.1
1.3.2
Local Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
11
1.4.1
11
1.4.2
14
1.4.3
16
1.4.4
18
19
1.5.1
19
1.5.2
23
25
1.6.1
Properties of Holes . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
28
1.4
1.5
1.6
1.7
ii
CONTENTS
1.8
1.9
1.7.1
28
1.7.2
31
1.7.3
32
1.7.4
34
Thermal Transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
1.8.1
Boltzmann Theory . . . . . . . . . . . . . . . . . . . . . . . . . . .
36
1.8.2
40
1.8.3
41
1.8.4
Onsager Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
43
Electron-Phonon Scattering . . . . . . . . . . . . . . . . . . . . . . . . . . .
45
1.9.1
Introductory Remarks
. . . . . . . . . . . . . . . . . . . . . . . . .
45
1.9.2
Electron-Phonon Interaction . . . . . . . . . . . . . . . . . . . . . .
45
1.9.3
48
2 Mesoscopia
51
2.1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
2.2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
2.3
51
2.3.1
54
Multichannel Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
2.4.1
60
2.4.2
62
2.4.3
64
2.4.4
67
75
2.5.1
Weak Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . .
78
Anderson Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
80
2.6.1
82
2.6.2
83
2.4
2.5
2.6
CONTENTS
iii
2.6.3
85
2.6.4
Finite Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . .
89
91
91
3.1.1
Energy Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
3.2
Kramers-Kronig Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . .
93
3.3
95
3.3.1
Spectral Representation . . . . . . . . . . . . . . . . . . . . . . . . .
97
3.3.2
Energy Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
3.3.3
Correlation Functions . . . . . . . . . . . . . . . . . . . . . . . . . .
100
3.3.4
Continuous Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .
101
3.4
Example: S =
101
Bloch Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
102
Electromagnetic Response . . . . . . . . . . . . . . . . . . . . . . . . . . . .
104
3.5.1
106
3.5.2
A Sum Rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
106
3.5.3
107
3.5.4
Neutral Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
107
3.5.5
108
Density-Density Correlations . . . . . . . . . . . . . . . . . . . . . . . . . .
110
3.6.1
Sum Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
112
113
3.7.1
Explicit T = 0 Calculation . . . . . . . . . . . . . . . . . . . . . . .
114
119
3.8.1
119
3.8.2
120
3.8.3
121
3.8.4
122
3.4.1
3.5
3.6
3.7
3.8
1
2
iv
CONTENTS
3.8.5
Plasmons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Magnetism
125
127
4.1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
127
4.2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
127
4.2.1
129
129
4.3.1
129
4.3.2
130
4.3.3
130
4.3.4
132
4.3.5
133
4.3.6
Spin-Orbit Interaction . . . . . . . . . . . . . . . . . . . . . . . . .
135
4.3.7
136
137
4.4.1
138
4.4.2
139
141
4.5.1
Pauli Paramagnetism . . . . . . . . . . . . . . . . . . . . . . . . . .
141
4.5.2
Landau Diamagnetism . . . . . . . . . . . . . . . . . . . . . . . . .
143
145
4.6.1
145
4.6.2
146
4.6.3
Antiferromagnetic Solution . . . . . . . . . . . . . . . . . . . . . . .
150
4.6.4
151
4.3
4.4
4.5
4.6
4.7
. . . . . . . . . . . .
153
4.7.1
153
4.7.2
155
4.7.3
157
CONTENTS
4.7.4
Herrings approach . . . . . . . . . . . . . . . . . . . . . . . . . . .
157
158
4.8.1
Ferromagnets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
161
4.8.2
Antiferromagnets . . . . . . . . . . . . . . . . . . . . . . . . . . . .
161
4.8.3
Susceptibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
162
4.8.4
163
Magnetic Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
166
4.9.1
168
4.9.2
Quasi-1D Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . .
169
170
4.8
4.9
4.10.1
171
4.10.2
173
4.10.3
173
4.10.4
178
vi
0.1
CONTENTS
Preface
This is a proto-preface. A more complete preface will be written after these notes are
completed.
These lecture notes are intended to supplement a graduate level course in condensed matter
physics.
Chapter 0
Introductory Information
Instructor:
Contact :
Lectures:
Office Hours:
Daniel Arovas
Mayer Hall 5671 / 534-6323 / darovas@ucsd.edu
Tu Th / 9:30 am - 10:50 am / Mayer Hall 5301
W 2:00 pm - 3:30 pm / Mayer Hall 5671
A strong emphasis of this class will be on learning how to calculate. I plan to cover the
following topics this quarter:
Transport: Boltzmann equation, transport coefficients, cyclotron resonance, magnetoresistance, thermal transport, electron-phonon scattering
Mesoscopic Physics: Landauer formula, conductance fluctuations, Aharonov-Bohm effect, disorder, weak localization, Anderson localization
Magnetism: Weak vs. strong, local vs. itinerant, Hubbard and Heisenberg models, spin
wave theory, magnetic ordering, Kondo effect
Other: Linear response theory, Fermi liquid theory (time permitting)
There will be about four assignments and a take-home final examination. I will be following
my own notes, which are available from the course web site.
0.1
References
0.1. REFERENCES
Chapter 1
Boltzmann Transport
1.1
References
1.2
Introduction
1.3
1.3.1
= vn (k) =
1 n (k)
~ k
(1.1)
dk
e
e
= E(r, t) vn (k) B(r, t) .
(1.2)
dt
~
~c
Here, n is the band index and n (k) is the dispersion relation for band n. The wavevector
is k (~k is the crystal momentum), and n (k) is periodic under k k + G, where G is
any reciprocal lattice vector. These formulae are valid only at sufficiently weak fields. They
neglect, for example, Zener tunneling processes in which an electron may change its band
index as it traverses the Brillouin zone. We also neglect the spin-orbit interaction in our
discussion.
We are of course interested in more than just a single electron, hence to that end let us
consider the distribution function fn (r, k, t), defined such that1
fn (r, k, t)
3
d3r of r and wavevectors within d3k of k at time t.
(2)
(1.3)
Note that the distribution function is dimensionless. By performing integrals over the
distribution function, we can obtain various physical quantities. For example, the current
density at r is given by
X Z d3k
j(r, t) = e
fn (r, k, t) vn (k) .
(1.4)
(2)3
n,
=
, , ,
,
,
.
x y z kx ky kz
1
(1.6)
(1.7)
We will assume three space dimensions. The discussion may be generalized to quasi-two dimensional
and quasi-one dimensional systems as well.
Now note that as a consequence of the dynamics (1.1,1.2) that u = 0, i.e. phase space
flow is incompressible, provided that (k) is a function of k alone, and not of r. Thus, in
the absence of collisions, we have
f
+ u f = 0 .
t
(1.8)
f
t
Ik {f }
(1.9)
coll
where the right side is known as the collision integral . The collision integral is in general
a function of r, k, and t and a functional of the distribution f . As the k-dependence is
the most important for our concerns, we will write Ik in order to make this dependence
explicit. Some examples should help clarify the situation.
First, lets consider a very simple model of the collision integral,
Ik {f } =
f (r, k, t) f 0 (r, k)
.
((k))
(1.10)
This model is known as the relaxation time approximation. Here, f 0 (r, k) is a static distribution function which describes a local equilibrium at r. The quantity ((k)) is the
relaxation time, which we allow to be energy-dependent. Note that the collision integral indeed depends on the variables (r, k, t), and has a particularly simple functional dependence
on the distribution f .
A more sophisticated model might invoke Fermis golden rule, Consider elastic scattering
from a static potential U(r) which induces transitions between different momentum states.
We can then write
Ik {f } =
=
2 X
0 2
| k U k | (fk0 fk ) (k k0 )
~
k0
Z 3 0
2
dk
k0 )|2 (fk0 fk ) (k k0 )
| U(k
~V (2)3
(1.11)
(1.12)
where we abbreviate fk f (r, k, t). In deriving the last line weve used plane wave wave-
(1.13)
for smooth functions A(k). Note the factor of V 1 in front of the integral in eqn. 1.12.
What this tells us is that for a bounded localized potential U(r), the contribution to the
collision integral is inversely proportional to the size of the system. This makes sense
because the number of electrons scales as V but the potential is only appreciable over a
region of volume V 0 . Later on, we shall consider a finite density of scatterers, writing
PNimp
U(r) = i=1
U (r Ri ), where the impurity density nimp = Nimp /V is finite, scaling as
0
k0 ) apparently scales as V , which would mean I {f } scales as V ,
V . In this case U(k
k
which is unphysical. As we shall see, the random positioning of the impurities means that
k0 )|2 is incoherent and averages out to zero. The coherent
the O(V 2 ) contribution to |U(k
piece scales as V , canceling the V in the denominator of eqn. 1.12, resulting in a finite value
for the collision integral in the thermodynamic limit (i.e. neither infinite nor infinitesimal).
Later on we will discuss electron-phonon scattering, which is inelastic. An electron with
wavevector k0 can scatter into a state with wavevector k = k0 q mod G by absorption of
a phonon of wavevector q or emission of a phonon of wavevector q. Similarly, an electron
of wavevector k can scatter into the state k0 by emission of a phonon of wavevector q or
absorption of a phonon of wavevector q. The matrix element for these processes depends
on k, k0 , and the polarization index of the phonon. Overall, energy is conserved. These
considerations lead us to the following collision integral:
n
2 X
Ik {f, n} =
|g (k, k0 )|2 (1 fk ) fk0 (1 + nq, ) (k + ~q k0 )
~V 0
k ,
+(1 fk ) fk0 nq (k ~q k0 )
fk (1 fk0 ) (1 + nq ) (k ~q k0 )
o
fk (1 fk0 ) nq (k + ~q k0 ) q,k0 k
mod G
, (1.14)
which is a functional of both the electron distribution fk as well as the phonon distribution
nq . The four terms inside the curly brackets correspond, respectively, to cases (a) through
(d) in fig. 1.1.
While collisions will violate crystal momentum conservation, they do not violate conservation of particle number. Hence we should have3
Z Z 3
dk
Ik {f } = 0 .
(1.15)
d3r
(2)3
2
Rather than plane waves, we should use Bloch waves nk (r ) = exp(ik r ) unk (r ), where cell function
unk (r ) satisfies unk (r + R) = unk (r ), where R is any direct lattice vector. Plane waves do not contain
the cell functions, although they do exhibit Bloch periodicity nk (r + R) = exp(ik R) nk (r ).
R d3k
3
If collisions are purely local, then (2)
3 Ik {f } = 0 at every point r in space.
Z 3
dk
d3r
f (r, k, t)
(2)3
(1.16)
d3 r
d3k
F (r, k) f (r, k, t) .
(2)3
(1.17)
F
=
t
d3k
)
d r
F (r, k)
(rf
(kf ) + Ik {f }
(2)3
r
k
3
Z
=
d r
d3k
(2)3
F dr F dk
f + F Ik {f } .
r dt
k dt
(1.18)
Note that the relaxation time approximation violates all such conservation laws. Within the relaxation
time approximation, there are no collisional invariants.
10
(1.19)
which says that F (t) changes only as a result of collisions. If F is a collisional invariant,
then F = 0. This is the case when F = 1, in which case F is the total number of particles,
or when F = (k), in which case F is the total energy.
1.3.2
Local Equilibrium
f (k) =
exp
(k)
kB T
1
+1
(1.20)
f (r, k, t) =
exp
(k) (r, t)
kB T (r, t)
1
+1
(1.21)
This is, however, not a solution to the full Boltzmann equation due to the streaming terms
r r + k k in the convective derivative. These, though, act on longer time scales than
those responsible for the establishment of local equilibrium. To obtain a solution, we write
f (r, k, t) = f 0 (r, k, t) + f (r, k, t)
(1.22)
and solve for the deviation f (r, k, t). We will assume = (r) and T = T (r) are timeindependent. We first compute the differential of f 0 ,
f 0
df 0 = kB T
d
kB T
(
)
f 0
d
( ) dT
d
= kB T
kB T
kB T 2
kB T
(
)
f 0
T
=
dr +
dr
dk ,
(1.23)
r
T r
k
from which we read off
(
f 0
r
f 0
k
= ~v
T
+
r
T r
f 0
.
)
f 0
(1.24)
(1.25)
11
We thereby obtain
f
1
e
f
f 0
E+ vB
= Ik {f 0 + f }
+ v f
+ v eE +
T
t
~
c
k
T
(1.26)
where E = ( /e) is the gradient of the electrochemical potential; well henceforth
refer to E as the electric field. Eqn (1.26) is a nonlinear integrodifferential equation in f ,
with the nonlinearity coming from the collision integral. (In some cases, such as impurity
scattering, the collision integral may be a linear functional.) We will solve a linearized
version of this equation, assuming the system is always close to a state of local equilibrium.
Note that the inhomogeneous term in (1.26) involves the electric field and the temperature
gradient T . This means that f is proportional to these quantities, and if they are small
then f is small. The gradient of f is then of second order in smallness, since the external
fields /e and T are assumed to be slowly varying in space. To lowest order in smallness,
then, we obtain the following linearized Boltzmann equation:
f
e
f
f 0
vB
+ v eE +
T
= L f
t
~c
k
T
(1.27)
where L f is the linearized collision integral; L is a linear operator acting on f (we suppress
denoting the k dependence of L). Note that we have not assumed that B is small. Indeed
later on we will derive expressions for high B transport coefficients.
1.4
1.4.1
ev E = Ik {f 0 + f } .
t
(1.28)
f f0
f
= ,
(1.29)
where , which may be k-dependent, is the relaxation time. In the absence of any fields
or temperature and electrochemical potential gradients, the Boltzmann equation becomes
f = f / , with the solution f (t) = f (0) exp(t/ ). The distribution thereby relaxes to
the equilibrium one on the scale of .
12
((k))
and obtain
f (k, t) =
e E v(k) ((k)) f 0 it
e
.
1 i ((k))
(1.30)
(1.31)
The equilibrium distribution f 0 (k) results in zero current, since f 0 (k) = f 0 (k). Thus,
the current density is given by the expression
Z 3
dk
f v
j (r, t) = 2e
(2)3
it
= 2e E e
f 0
(1.32)
In the above calculation, the factor of two arises from summing over spin polarizations. The
conductivity tensor is defined by the linear relation j () = () E (). We have thus
derived an expression for the conductivity tensor,
Z 3
f 0
d k ((k)) v (k) v (k)
2
() = 2e
(1.33)
(2)3
1 i ((k))
Note that the conductivity is a property of the Fermi surface. For kB T F , we have
f 0 / (F (k)) and the above integral is over the Fermi surface alone. Explicitly,
we change variables to energy and coordinates along a constant energy surface, writing
d3k =
d dS
d dS
=
,
|/k|
~|v|
(1.34)
where dS is the differential area on the constant energy surface (k) = , and v(k) =
~1 k (k) is the velocity. For T TF , then,
Z
e2
(F )
v (k) v (k)
dSF
() = 3
.
(1.35)
4 ~ 1 i (F )
|v(k)|
For free electrons in a parabolic band, we write (k) = ~2 k2 /2m , so v (k) = ~k /m . To
further simplify matters, let us assume that is constant, or at least very slowly varying in
the vicinity of the Fermi surface. We find
Z
f 0
e2
2
() =
d g()
,
(1.36)
3m 1 i
d3k
( (k)) .
(2)3
(1.37)
13
about the dependence in g(), and not the details of its prefactor:
Z
Z
f 0
=
d f 0 () ( g())
d g()
Z
= 23 d g() f 0 () = 32 n ,
(1.39)
g() =
where n = N/V is the electron number density for the conduction band. The final result
for the conductivity tensor is
() =
ne2
m 1 i
(1.40)
This is called the Drude model of electrical conduction in metals. The dissipative part of
the conductivity is Re . Writing = and separating into real and imaginary parts
= 0 + i 00 , we have
ne2
1
0 () =
.
(1.41)
m 1 + 2 2
The peak at = 0 is known as the Drude peak.
Heres an elementary derivation of this result. Let p(t) be the momentum of an electron,
and solve the equation of motion
dp
p
= e E eit
(1.42)
dt
to obtain
e E
e E it
e
+ p(0) +
et/ .
(1.43)
p(t) =
1 i
1 i
The second term above is a transient solution to the homogeneous equation p + p/ = 0.
At long times, then, the current j = nep/m is
j(t) =
ne2
E eit .
m (1 i )
(1.44)
14
1.4.2
What happens when an electromagnetic wave is incident on a metal? Inside the metal we
have Maxwells equations,
4
1 D
4 i
H =
j+
=
ik B =
E
(1.45)
c
c t
c
c
1 B
i
E =
=
ik E = B
(1.46)
c t
c
E =B =0
=
ik E = ik B = 0 ,
(1.47)
where weve assumed = = 1 inside the metal, ignoring polarization due to virtual
interband transitions (i.e. from core electrons). Hence,
2 4i
+ 2 ()
c2
c
2 p2 i
2
() 2 ,
= 2 + 2
c
c 1 i
c
k2 =
where p =
function,
(1.48)
(1.49)
p
4ne2 /m is the plasma frequency for the conduction band. The dielectric
() = 1 +
p2 i
4i()
=1+ 2
1 i
(1.50)
15
p
(), leading to the electromagnetic
In the vacuum
Consider a wave normally incident upon a metallic surface normal to z.
(z < 0), we write
eiz/c eit + E2 x
eiz/c eit
E(r, t) = E1 x
(1.51)
c
B(r, t) =
E = E1 y eiz/c eit E2 y eiz/c eit
i
(1.52)
(1.53)
c
(1.54)
E = N E3 y eiN z/c eit
B(r, t) =
i
gives E1 + E2 = E3 . Continuity of H n
gives E1 E2 = N E3 . Thus,
Continuity of E n
E2
1N
=
E1
1+N
E3
2
=
E1
1+N
(1.55)
(1.56)
(1.57)
(1 i )
!1/2
1 + i p2
2 2
= R 1
.
N ()
p
2
N 2 () = 1 +
(1.58)
p2
i p2
+
2 3
(1.59)
16
Figure 1.3: Frequency-dependent absorption of hcp cobalt by J. Weaver et al., Phys. Rev.
B 19, 3850 (1979).
which is almost purely real and negative. Hence N is almost purely imaginary and
R 1. (To lowest nontrivial order, R = 1 2/p .) Still high reflectivity.
1 p :
Here we have
p2
p
= R =
2
2
and R 1 the metal is transparent at frequencies large compared to p .
N 2 () 1
1.4.3
(1.60)
In our analysis of the electrodynamics of metals, we assumed that the dielectric constant
due to all the filled bands was simply = 1. This is not quite right. We should instead
have written
2 4i()
+
c2
c2
(
)
2
p i
() = 1 + 2
,
1 i
k2 =
(1.61)
(1.62)
17
4ne2
m
1/2
(1.63)
where n is the conduction electron density. Note that ( ) = , although again this
is only true for smaller than the gap to neighboring bands. It turns out that for insulators
one can write
2
pv
(1.64)
' 1 + 2
g
p
where pv = 4nv e2 /me , with nv the number density of valence electrons, and g is
the energy gap between valence and conduction bands. In semiconductors such as Si and
Ge, g 4 eV, while pv 16 eV, hence 17, which is in rough agreement with the
experimental values of 12 for Si and 16 for Ge. In metals, the band gaps generally are
considerably larger.
There are some important differences to consider in comparing semiconductors and metals:
18
(1.65)
5
2
In high purity semiconductors the mobility = e /m >
10 cm /vs the low temperature
15 1
scattering time is typically 1011 s. Thus, for >
3 10 s in the optical range, we
1 2
R=
.
1 +
(1.66)
Taking = 10, one obtains R = 0.27, which is high enough so that polished Si wafers
appear shiny.
1.4.4
At high frequencies, when 1, our expression for the conductivity, eqn. (1.33), yields
ie2
() =
12 3 ~
Z
f 0
d
dS v(k) ,
(1.67)
ine2
(1.68)
mopt
1
12 3 ~n
Z
f 0
d
dS v(k) .
(1.69)
Note that at high frequencies () is purely imaginary. What does this mean? If
E(t) = E cos(t) =
1
2
E eit + e+it
(1.70)
19
then
j(t) =
=
1
2
E () eit + () e+it
ne2
mopt
E sin(t) ,
(1.71)
where we have invoked () = (). The current is therefore 90 out of phase with the
voltage, and the average over a cycle hj(t) E(t)i = 0. Recall that we found metals to be
transparent for p 1 .
At zero temperature, the optical mass is given by
Z
1
1
=
dSF v(k) .
3
12 ~n
mopt
(1.72)
1
dSF v(k) ,
(1.73)
from which one can define the thermodynamic effective mass mth , appealing to the low
temperature form of the specific heat,
cV =
m
2 2
kB T g(F ) th c0V ,
3
me
(1.74)
where
me kB2 T
(3 2 n)1/3
3~2
is the specific heat for a free electron gas of density n. Thus,
Z
1
~
mth =
dSF v(k)
1/3
2
4(3 n)
c0V
1.5
1.5.1
(1.75)
(1.76)
Let us go beyond the relaxation time approximation and calculate the scattering time
from first principles. We will concern ourselves with scattering of electrons from crystalline
impurities. We begin with Fermis Golden Rule6 ,
Ik {f } =
2 X
0 2
k U k
(fk0 fk ) ((k) (k0 )) ,
~ 0
(1.77)
Well treat the scattering of each spin species separately. We assume no spin-flip scattering takes place.
20
Metal
Li
Na
K
Rb
Cs
Cu
Ag
Au
mopt /me
thy expt
1.45 1.57
1.00 1.13
1.02 1.16
1.08 1.16
1.29 1.19
-
mth /me
thy expt
1.64 2.23
1.00 1.27
1.07 1.26
1.18 1.36
1.75 1.79
1.46 1.38
1.00 1.00
1.09 1.08
Table 1.1: Optical and thermodynamic effective masses of monovalent metals. (Taken from
Smith and Jensen).
U(r) =
U (r Rj )
(1.78)
j=1
Nimp
0 2
X i(kk0 )R 2
j
(k k0 )|2
k U k = V 2 |U
e
,
(1.79)
j=1
(1.80)
is the Fourier transform of the impurity potential. Note that we are assuming a single species
of impurities; the method can be generalized to account for different impurity species.
To make progress, we assume the impurity positions are random and uncorrelated, and we
average over them. Using
NX
2
imp
iqR
j = N
e
imp + Nimp (Nimp 1) q,0 ,
(1.81)
2 Nimp
(k k0 )|2 + Nimp (Nimp 1) |U
(0)|2 kk0 .
k0 U k =
|U
2
V
V2
(1.82)
j=1
we obtain
21
f = f 0 + f , we have
Ik {f } =
2nimp
~
d3k 0
~2 k2 ~2 k0 2
0 2
|
U
(k
k
)|
(2)3
2m
2m
!
(fk0 fk ) ,
(1.83)
where nimp = Nimp /V is the number density of impurities. Note that we are assuming a
parabolic band. We next make the Ansatz
f 0
(1.84)
fk = ((k)) e E v(k)
(k)
and solve for ((k)). The (time-independent) Boltzmann equation is
f 0
2
e E v(k)
=
n
eE
~ imp
d3k 0
~2 k2 ~2 k0 2
0 2
|U (k k )|
(2)3
2m
2m
!
f 0
f 0
((k )) v(k )
((k)) v(k)
.
(k0 )
(k)
0
(1.85)
Due to the isotropy of the problem, we must have ((k)) is a function only of the magnitude
of k. We then obtain7
Z
Z
0
nimp
~k
0 02
0 |U
(k k0 )|2 (k k ) ~ (k k0 ) ,
d
k
=
((k))
dk
k
m
4 2 ~
~2 k/m m
(1.86)
whence
1
(F )
m kF nimp
4 2 ~3
0 |U (kF k
kF k
0 )|2 (1 k
k
0) .
dk
(1.87)
(q) is a function
If the impurity potential U (r) itself is isotropic, then its Fourier transform U
2 1
2
2
0
0
m
2~2
2
(k k0 )|2 ,
|U
(1.88)
Z
= 2nimp vF d F () (1 cos ) sin
(1.89)
where vF = ~kF /m is the Fermi velocity8 . The mean free path is defined by ` = vF .
7
8
We assume that the Fermi surface is contained within the first Brillouin zone.
The subscript on F () is to remind us that the cross section depends on kF as well as .
22
Notice the factor (1 cos ) in the integrand of (1.89). This tells us that forward scattering
( = 0) doesnt contribute to the scattering rate, which justifies our neglect of the second
term in eqn. (1.82). Why should be utterly insensitive to forward scattering? Because
(F ) is the transport lifetime, and forward scattering does not degrade the current. Therefore, ( = 0) does not contribute to the transport scattering rate 1 (F ). Oftentimes
one sees reference in the literature to a single particle lifetime as well, which is given by
the same expression but without this factor:
(
1
sp
tr1
)
= 2nimp vF
d F ()
1
(1 cos )
sin
(1.90)
Note that sp = (nimp vF F,tot )1 , where F,tot is the total scattering cross section at energy
F , a formula familiar from elementary kinetic theory.
To derive the single particle lifetime, one can examine the linearized time-dependent Boltzmann equation with E = 0,
Z
fk
0 ( 0 ) (fk0 fk ) ,
(1.91)
= nimp vF dk
kk
t
where v = ~k/m is the velocity, and where the kernel is kk0 = cos1 (k k0 ). We now
expand in spherical harmonics, writing
X
Y (k
0) ,
L YLM (k)
(1.92)
(kk0 ) tot
LM
L,M
where as before
tot
Z
= 2 d sin () .
(1.93)
Expanding
fk (t) =
,
ALM (t) YLM (k)
(1.94)
L,M
(1.95)
(1.96)
which depend only on the total angular momentum quantum number L. These rates describe the relaxation of nonuniform distributions, when fk (t = 0) is proportional to some
1
spherical harmonic YLM (k). Note that L=0
= 0, which reflects the fact that the total
particle number is a collisional invariant. The single particle lifetime is identified as
1
1
sp
L
= nimp vF tot ,
(1.97)
1.5.2
23
(1.99)
1/2
Debye-H
uckel Screening : This mechanism is typical of ionic solutions, although it may
also be of relevance in solids with ultra-low Fermi energies. From classical statistical
mechanics, the local variation in electron number density induced by a potential (r)
is
ne(r)
n(r) = n ee(r)/kB T n
,
(1.100)
kB T
where we assume the potential is weak on the scale of kB T /e. Poissons equation now
gives us
2 = 4e n
=
4ne2
2
DH .
kB T
(1.101)
(1.102)
Ze2 r/
e
r
2
(q) = 4Ze .
U
q 2 + 2
(1.103)
24
F () =
Ze2
4F
!2
(1.104)
sin2 12 + (2kF )2
and the divergence at small angle is cut off. The transport lifetime for screened Coulomb
scattering is therefore given by
1
(F )
= 2nimp vF
2 Z
Ze2
d sin (1 cos )
4F
= 2nimp vF
Ze2
sin2 12 + (2kF )2
2
2F
!2
ln(1 + )
1 +
,
(1.105)
with
=
4 2 2
~2 k
kF = F2 = kF aB .
m e
(1.106)
Here aB = ~2 /m e2 is the effective Bohr radius (restoring the factor). The resistivity
is therefore given by
=
nimp
m
h
= Z 2 2 aB
F (kF aB ) ,
2
ne
e
n
(1.107)
where
1
F () = 3
ln(1 + )
1 +
.
(1.108)
A, we have
With h/e2 = 25, 813 and aB aB = 0.529
= 1.37 104 cm Z 2
Impurity
Ion
Be
Mg
B
Al
In
per %
(-cm)
0.64
0.60
1.4
1.2
1.2
nimp
n
Impurity
Ion
Si
Ge
Sn
As
Sb
F (kF aB ) .
per %
(-cm)
3.2
3.7
2.8
6.5
5.4
(1.109)
25
1.6
1.6.1
Since filled bands carry no current, we have that the current density from band n is
Z 3
Z 3
dk
dk
jn (r, t) = 2e
f
(r,
k,
t)
v
(k)
=
+2e
fn (r, k, t) vn (k) ,
(1.110)
n
n
(2)3
(2)3
26
Figure 1.6: Two states: A = ek hk 0 and B = ek hk 0 . Which state carries
more current? What is the crystal momentum of each state?
It is instructive to consider the exercise of fig. 1.6. The two states to be analyzed are
(1.111)
A = c,k v,k 0 = ek hk 0
(1.112)
B = c,k v,k 0 = ek hk 0 ,
is the creation operator for electrons in the conduction band, and hk v,k
where ek c,k
is the creation operator for
(and hence the destruction operator for electrons) in the
holes
valence band. The state 0 has all states below the top of the valence
band filled, and
all states above the bottom of the conduction band empty. The state 0 is the same state,
but represented now as a vacuum for conduction electrons and valence holes. The current
density in each state is given by j = e(vh ve )/V , where V is the volume (i.e. length) of the
system. The dispersions resemble c,v 21 Eg ~2 k 2 /2m , where Eg is the energy gap.
State A :
The electron velocity is ve = ~k/m ; the hole velocity is vh = ~k/m . Hence,
the total current density is j 2e~k/m V and the total crystal momentum is
P = pe + ph = ~k ~k = 0.
State B :
The electron velocity is ve = ~k/m ; the hole velocity is vh = ~(k)/m . The
27
c0
1 2 c 1
k k
2 ~ m
(1.113)
.
(1.114)
The velocity is
1
(1.115)
= ~ m1
k ,
~ k
where the + sign is used in conjunction with mc and the sign with mv . We compute the
acceleration a = r via the chain rule,
v (k) =
v dk
k dt
1
1
= e m E + (v B)
c
1
F = m a = e E + (v B) .
c
a =
(1.116)
(1.117)
Thus, the hole wavepacket accelerates as if it has charge +e but a positive effective mass.
Finally, what form does the Boltzmann equation take for holes? Starting with the Boltzmann equation for electrons,
f
f
f
+ r
+ k
= Ik {f } ,
t
r
k
(1.118)
(1.119)
This then is the Boltzmann equation for the hole distribution f. Recall that we can expand
the collision integral functional as
Ik {f 0 + f } = L f + . . .
(1.120)
where L is a linear operator, and the higher order terms are formally of order (f )2 . Note
that the zeroth order term Ik {f 0 } vanishes due to the fact that f 0 represents a local
equilibrium. Thus, writing f = f0 + f
Ik {1 f} = Ik {1 f0 f} = L f + . . .
(1.121)
28
vB
v eE +
T
= L f
t
~c
k
T
(1.122)
which is of precisely the same form as the electron case in eqn. (1.27). Note that the local
equilibrium distribution for holes is given by
f0 (r, k, t) =
1.7
1.7.1
exp
(r, t) (k)
kB T (r, t)
1
+1
(1.123)
In the presence of an external magnetic field B, the linearized Boltzmann equation takes
the form9
f
f 0
e
f
ev E
vB
= L f .
(1.124)
t
~c
k
We will obtain an explicit solution within the relaxation time approximation L f = f /
and the effective mass approximation,
(k) = 12 ~2 m1
k k
v = ~ m1
k ,
(1.125)
where the top sign applies for electrons and the bottom sign for holes. With E(t) = E eit ,
we try a solution of the form
f (k, t) = k A() eit f (k) eit
(1.126)
f 0
v B (k A ) = e v E
,
~c
k
(1.127)
A
v B
(k A ) = v B A + k
k
k
A
= v B A + ~ k v
= v B A ,
9
(1.128)
29
Defining
e
(1.130)
( 1 i) m B ,
c
we obtain the solution
0
f
f = e v m 1
E
.
(1.131)
From this, we can compute the current density and the conductivity tensor. The electrical
current density is
Z 3
dk
v f
j = 2e
(2)3
= +2e E
d3k
v v m 1
()
(2)3
f 0
,
(1.132)
where we allow for an energy-dependent relaxation time (). Note that () is energydependent due to its dependence on . The conductivity is then
(Z
)
3k
0
d
f
(, B) = 2~2 e2 m1
k k
1
(1.133)
()
(2)3
Z
f 0
2 2
1
= e d g() ()
,
3
(1.134)
where the chemical potential is measured with respect to the band edge. Thus,
(, B) = ne2 h1
i ,
(1.135)
0
d g() f
1
()
0
d g() f
(1.136)
(electrons)
(holes)
(1.137)
30
i
1
1 h 1
e
= 2 = 2 ( i)m B
.
ne
ne
c
(1.138)
(1.139)
We will assume that B is directed along one of the principal axes of the effective mass
y,
and z,
in which case
tensor m , which we define to be x,
1
( i) mx
eB/c
0
1
eB/c
( 1 i) my
0
(, B) = 2
ne
0
0
( 1 i) mz
(1.140)
where mx,y,z are the eigenvalues of m and B lies along the eigenvector z.
Note that
mx
(1 i )
ne2
is independent of B. Hence, the magnetoresistance,
xx (, B) =
(1.141)
(1.142)
vanishes: xx (B) = 0. While this is true for a single parabolic band, deviations from
parabolicity and contributions from other bands can lead to a nonzero magnetoresistance.
The conductivity tensor is the matrix inverse of . Using the familiar equality
1
1
d b
a b
=
,
c d
ad bc c a
(1.143)
we obtain
(1i )/mx
(1i )2 +(c )2
(, B) = ne2 c / mx my
(1i )2 +(c )2
0
where
c
c /
mx my
(1i )2 +(c )2
(1i )/my
(1i )2 +(c )2
eB
,
m c
0
1
(1i )mz
(1.144)
(1.145)
with m
31
p
mx my , is the cyclotron frequency. Thus,
ne2
1 i
2
mx 1 + (c 2 ) 2 2i
ne2
1
zz (, B) =
.
mz 1 i
xx (, B) =
(1.146)
(1.147)
Note that xx,yy are field-dependent, unlike the corresponding components of the resistivity
tensor.
1.7.2
(1.148)
we find that eB/m c = 1.75 1011 Hz in a field of B = 1 kG. In metals, the disorder is
such that even at low temperatures c typically is small. In semiconductors, however, the
smallness of m and the relatively high purity (sometimes spectacularly so) mean that c
can get as large as 103 at modest fields. This allows for a measurement of the effective mass
tensor using the technique of cyclotron resonance.
The absorption of electromagnetic radiation is proportional to the dissipative (i.e. real) part
of the diagonal elements of (), which is given by
0
xx
(, B) =
ne2
1 + (2 + 1)s2
,
mx 1 + 2(2 + 1)s2 + (2 1)2 s4
(1.149)
0 (B)
where = B/B , with B = m c /e, and s = . For fixed , the conductivity xx
is then peaked at B = B . When 1 and c 1, B approaches B , where
0 (, B ) = ne2 /2m . By measuring B one can extract the quantity m = eB /c.
xx
Varying the direction of the magnetic field, the entire effective mass tensor may be determined.
For finite , we can differentiate the above expression to obtain the location of the cyclotron
resonance peak. One finds B = (1 + )1/2 B , with
=
(2s2 + 1) +
p
(2s2 + 1)2 1
2
s
(1.150)
1
1
+ 6 + O(s8 ) .
4
4s
8s
As depicted in fig. 1.7, the resonance peakshifts to the left of B for finite values of .
The peak collapses to B = 0 when 1/ 3 = 0.577.
32
Figure 1.7: Theoretical cyclotron resonance peaks as a function of B/B for different values
of .
1.7.3
For a semiconductor with both electrons and holes present a situation not uncommon to
metals either (e.g. Aluminum) each band contributes to the conductivity. The individual
band conductivities are additive because the electron and hole conduction processes occur
in parallel , exactly as we would deduce from eqn. (1.4). Thus,
() =
(n)
() ,
(1.151)
n
(n)
where is the conductivity tensor for band n, which may be computed in either the
electron or hole picture (whichever is more convenient). We assume here that the two
distributions fc and fv evolve according to independent linearized Boltzmann equations,
i.e. there is no interband scattering to account for.
(n)
0 1
(1 ic )mc
B
1 0
c =
I+
nc e2 c
nc ec
0 0
33
we have
B = B z,
c c 0
0
0 = c c 0
0
0
0 c
v v 0
0 1 0
B
(1 iv )mv
1 0 0 = v v
I
v =
0 ,
nv e2 v
nv ec
0 0 0
0
0
v
(1.153)
(1.154)
where
c =
v =
(1 ic )mc
c =
nc e2 c
(1 iv )mv
v =
nv e2 v
(1.155)
nc ec
B
nv ec
(1.156)
v
c
c
v
+
+
2
2
2
2
2
2
2
2
+
c +c
v +v
c +c
v v
,
c
v
c
v
2+
+
2
2 + 2
2 + 2
2 + 2
v
B2
c
v
xx (B) xx (0)
nc ec
nv ec
=
2
xx (0)
B2
(c + v )2 + (c v )2 1 + 1
nc ec
(1.157)
(1.158)
nv ec
where
c = c1 =
v = v1 =
nc e2 c
mc
nv e2 v
mv
(1.159)
1 ic
1
1 iv
(1.160)
Note that the magnetoresistance is positive within the two band model, and that it saturates
in the high field limit:
2
c
v
c
v
xx (B ) xx (0)
nc ec
nv ec
(1.161)
=
2 .
xx (0)
(c v )2 1 + 1
nc ec
nv ec
34
(1.162)
and is independent of B.
In an intrinsic semiconductor, nc = nv exp(Eg /2kB T ), and xx (B)/xx (0) is finite
even as T 0. In the extrinsic (i.e. doped) case, one of the densities (say, nc in a p-type
material) vanishes much more rapidly than the other, and the magnetoresistance vanishes
with the ratio nc /nv .
1.7.4
In the high field limit, one may neglect the collision integral entirely, and write (at = 0)
f 0
e
f
vB
=0.
(1.163)
~c
dk
in which case the
Well consider the case of electrons, and take E = E y and B = B z,
solution is
~cE
f 0
f =
kx
.
(1.164)
B
Note that kx is not a smooth single-valued function over the Brillouin-zone due to Bloch
periodicity. This treatment, then, will make sense only if the derivative f 0 / confines k
to a closed orbit within the first Brillouin zone. In this case, we have
Z 3
E
dk
f 0
jx = 2ec
k
(1.165)
x
B (2)3
kx
e v E
E
= 2ec
B
d3k
f 0
k
.
x
(2)3
kx
(1.166)
Now we may integrate by parts, if we assume that f 0 vanishes on the boundary of the
Brillouin zone. We obtain
Z 3
2ecE
dk
nec
f0 =
E .
(1.167)
jx =
B
(2)3
B
We conclude that
nec
,
(1.168)
B
independent of the details of the band structure. Open orbits trajectories along Fermi
surfaces which cross Brillouin zone boundaries and return in another zone post a subtler
problem, and generally lead to a finite, non-saturating magnetoresistance.
xy = yx =
d3k
f0
nec
k
=+
E
x
3
(2)
kx
B
(1.169)
35
Figure 1.8: Nobel Prize winning magnetotransport data in a clean two-dimensional electron
gas at a GaAs-AlGaAs inversion layer, from D. C. Tsui, H. L. Stormer, and A. C. Gossard,
Phys. Rev. Lett. 48, 1559 (1982). xy and xx are shown versus magnetic field for a set of
four temperatures. The Landau level filling factor is = nhc/eB. At T = 4.2 K, the Hall
resistivity obeys xy = B/nec (n = 1.3 1011 cm2 ). At lower temperatures, quantized
plateaus appear in xy (B) in units of h/e2 .
and xy = +nec/B, where n is the hole density.
We define the Hall coefficient RH = xy /B and the Hall number
zH
1
nion ecRH
(1.170)
where nion is the ion density. For high fields, the off-diagonal elements of both and
are negligible, and xy = xy . Hence RH = 1/nec, and zH = n/nion . The high
field Hall coefficient is used to determine both the carrier density as well as the sign of the
charge carriers; zH is a measure of valency.
In Al, the high field Hall coefficient saturates at zH = 1. Why is zH negative? As it turns
out, aluminum has both electron and hole bands. Its valence is 3; two electrons go into a
filled band, leaving one valence electron to split between the electron and hole bands. Thus
36
(1.171)
(1.172)
where weve invoked ne + n0e = nion , since precisely one electron from each ion is shared
between the two partially filled bands. Thus, xy = nion ec/B = nec/3B and zH = 1. At
lower fields, zH = +3 is observed, which is what one would expect from the free electron
model. Interband scattering, which is suppressed at high fields, leads to this result.
1.8
1.8.1
Thermal Transport
Boltzmann Theory
Consider a small region of solid with a fixed volume V . The first law of thermodynamics
applied to this region gives T S = E N . Dividing by V gives
dq T ds = d dn ,
(1.173)
where s is the entropy density, is energy density, and n the number density. This can be
directly recast as the following relation among current densities:
jq = T js = j jn ,
(1.174)
37
Figure 1.10: Fermi surfaces for electron (pink) and hole (gold) bands in Aluminum.
where jn = j/(e) is the number current density, j is the energy current density,
Z
j = 2
d3k
v f ,
(2)3
(1.175)
and js is the entropy current density. Accordingly, the thermal (heat) current density jq is
defined as
jq T js = j + j
e
Z 3
dk
=2
( ) v f .
(2)3
(1.176)
(1.177)
f 0
T
.
f = () v eE +
T
(1.178)
We now consider both the electrical current j as well as the thermal current density jq .
One readily obtains
Z
j=
2 e
Z
jq = 2
d3k
v f
(2)3
3
dk
(2)3
L11 E L12 T
(1.179)
( ) v f L21 E L22 T
(1.180)
38
L
21 = T L12
L
22
Z
Z
v v
f 0
e2
d ()
dS
= 3
4 ~
|v|
Z
Z
0
e
f
v v
= 3
d () ( )
dS
4 ~
|v|
Z
Z
0
f
v v
1
d () ( )2
dS
= 3
.
4 ~ T
|v|
(1.181)
(1.182)
(1.183)
1
3
4 ~
d () ( )
f 0
Z
dS
v v
|v|
(1.184)
1
J
.
T 2
(1.185)
L
21 = T L12 = e J1
L
22 =
The linear relations in eqn. (1.180) may be recast in the following form:
E = j + QT
(1.186)
jq = u j T ,
(1.187)
Q = L1
11 L12
(1.188)
u = L21 L1
11
= L22 L21 L1
11 L12 ,
(1.189)
1
J 1 J1
eT 0
1
=
J2 J1 J01 J1 ,
T
Q=
(1.190)
(1.191)
The names and physical interpretation of these four transport coefficients is as follows:
is the resistivity: E = j under the condition of zero thermal gradient (i.e. T = 0).
Q is the thermopower: E = Q T under the condition of zero electrical current (i.e.
j = 0). Q is also called the Seebeck coefficient.
u is the Peltier coefficient: jq = uj when T = 0.
is the thermal conductivity: jq = T when j = 0 .
39
Figure 1.11: A thermocouple is a junction formed of two dissimilar metals. With no electrical current passing, an electric field is generated in the presence of a temperature gradient,
resulting in a voltage V = VA VB .
One practical way to measure the thermopower is to form a junction between two dissimilar
metals, A and B. The junction is held at temperature T1 and the other ends of the metals
are held at temperature T0 . One then measures a voltage difference between the free ends
of the metals this is known as the Seebeck effect. Integrating the electric field from the
free end of A to the free end of B gives
ZB
VA VB =
E dl = (QB QA )(T1 T0 ) .
(1.192)
What one measures here is really the difference in thermopowers of the two metals. For an
absolute measurement of QA , replace B by a superconductor (Q = 0 for a superconductor).
A device which converts a temperature gradient into an emf is known as a thermocouple.
The Peltier effect has practical applications in refrigeration technology. Suppose an electrical
current I is passed through a junction between two dissimilar metals, A and B. Due to
the difference in Peltier coefficients, there will be a net heat current into the junction of
W = (uA uB ) I. Note that this is proportional to I, rather than the familiar I 2 result
from Joule heating. The sign of W depends on the direction of the current. If a second
junction is added, to make an ABA configuration, then heat absorbed at the first junction
will be liberated at the second. 10
10
To create a refrigerator, stick the cold junction inside a thermally insulated box and the hot junction
outside the box.
40
Figure 1.12: A sketch of a Peltier effect refrigerator. An electrical current I is passed through
a junction between two dissimilar metals. If the dotted line represents the boundary of a
thermally well-insulated body, then the body cools when uB > uA , in order to maintain a
heat current balance at the junction.
1.8.2
We begin with the continuity equations for charge density and energy density :
+j =0
t
(1.193)
+ j = j E ,
t
(1.194)
where E is the electric field11 . Now we invoke local thermodynamic equilibrium and write
n
T
=
+
t
n t
T t
=
T
+ cV
,
e t
t
(1.195)
where n is the electron number density (n = /e) and cV is the specific heat. We may
now write
cV
11
Note that it is
T
=
+
t
t
e t
= j E j j
e
= j E jq .
E j and not E j which is the source term in the energy continuity equation.
(1.196)
41
1.8.3
We will henceforth assume that sufficient crystalline symmetry exists (e.g. cubic symmetry) to render all the transport coefficients multiples of the identity matrix. Under such
conditions, we may write Jn = Jn with
1
Jn =
12 3 ~
Z
f 0
d () ( )
dS |v| .
(1.198)
(1.199)
= H() +
where D kB T
2
(k T )2 H 00 () + . . .
6 B
(1.200)
kB T e()/kB T + 1 e()/kB T + 1
(1.201)
Z
H( + x kB T )
exD
=
dx
H()
(ex + 1)(ex + 1)
(ex + 1)(ex + 1)
=
"
#
X
exD
= 2i
Res
H()
,
(ex + 1)(ex + 1)
=
n=0
12
(1.202)
x=(2n+1)i
Remember that physically the fixed quantities are temperature and total carrier number density (or
charge density, in the case of electron and hole bands), and not temperature and chemical potential. An
equation of state relating n, , and T is then inverted to obtain (n, T ), so that all results ultimately may
be expressed in terms of n and T .
42
e
1 4
(1 e )(1 e )
2 + 12
+ ...
(
)
1
D
1
= 2 + 12
21 D2 + O() e(2n+1)iD .
(1.203)
e(2n+1)iD H()
n=0
= D csc(D) H()
(1.204)
(1.205)
(1.206)
J1 =
0 2 (kB T )2
+ ...
e2 2
(1.207)
J2 =
0 2
(k T )2 + . . . ,
e2 3 B
(1.208)
(1.209)
2
=
(k /e)2 = 2.45 108 V2 K2 ,
T
3 B
(1.210)
Q=
2 kB2 T
2 e F
is known as the Wiedemann-Franz law. Note also that our result for the thermopower
is unambiguously negative. In actuality, several nearly free electron metals have positive
low-temperature thermopowers (Cs and Li, for example). What went wrong? We have
neglected electron-phonon scattering!
43
Figure 1.13: QT product for p-type and n-type Ge, from T. H. Geballe and J. W. Hull,
Phys. Rev. 94, 1134 (1954). Samples 7, 9, E, and F are distinguished by different doping
properties, or by their resistivities at T = 300 K: 21.5 -cm (7), 34.5 -cm (9), 18.5 -cm
(E), and 46.0 -cm (F).
1.8.4
Onsager Relations
(1.211)
where the {Fk } are generalized forces and the {Ji } are generalized currents. Moreover,
to each force Fi corresponds a unique conjugate current Ji , such that the rate of internal
entropy production is
X
S
S =
Fi Ji = Fi =
.
(1.212)
Ji
i
(1.213)
(1.214)
44
The Onsager relations have some remarkable consequences. For example, they require, for
B = 0, that the thermal conductivity tensor ij of any crystal must be symmetric, independent of the crystal structure. In general,this result does not follow from considerations of
crystalline symmetry. It also requires that for every off-diagonal transport phenomenon,
e.g. the Seebeck effect, there exists a distinct corresponding phenomenon, e.g. the Peltier
effect.
For the transport coefficients studied, Onsager reciprocity means that in the presence of an
external magnetic field,
(B) = (B)
(1.215)
(B) = (B)
(1.216)
u (B) = T Q (B) .
(1.217)
Lets consider an isotropic system in a weak magnetic field, and expand the transport
coefficients to first order in B:
(B) = + B
(1.218)
(1.219)
(1.220)
(1.221)
(B) = + $ B
Q (B) = Q + B
u (B) = u + B .
Onsager reciprocity requires u = T Q and = T . We can now write
E = j + j B + QT + T B
(1.222)
jq = u j + j B T $ T B .
(1.223)
T
y
= jy = 0)
T
y
T
y .
= 0)
1.9
1.9.1
45
Electron-Phonon Scattering
Introductory Remarks
We begin our discussion by recalling some elementary facts about phonons in solids:
In a crystal with r atoms per unit cell, there are 3(r 1) optical modes and 3 acoustic
modes, the latter guaranteed by the breaking of the three generators of space translations. We write the phonon dispersion as = (q), where {1, . . . , 3r} labels the
If j labels an acoustic mode, j (q) = cj (q)
q as q 0.
phonon branch, and q .
Phonons are bosonic particles with zero chemical potential. The equilibrium phonon
distribution is
1
n0q =
.
(1.224)
exp(~ (q)/kB T ) 1
The maximum phonon frequency is roughly given by the Debye frequency D . The
Debye temperature D = ~D 100 K 1000 K in most solids.
At high temperatures, equipartition gives h(Ri )2 i kB T , hence the effective scattering
1
2
cross-section tot increases as T , and >
1/nion vF tot T . From = m /ne , then,
we deduce that the high temperature resistivity should be linear in temperature due to
phonon scattering: (T ) T . Of course, when the mean free path ` = vF becomes as
small as the Fermi wavelength F , the entire notion of coherent quasiparticle transport
becomes problematic, and rather than continuing to grow we expect that the resistivity
should saturate: (T ) h/kF e2 , known as the Ioffe-Regel limit. For kF = 108 cm1 ,
this takes the value 260 cm.
1.9.2
Electron-Phonon Interaction
Let Ri = Ri0 + Ri denote the position of the ith ion, and let U (r) = Ze2 exp(r/TF )/r
be the electron-ion interaction. Expanding in terms of the ionic displacements Ri ,
X
X
Helion =
U (r Ri0 )
Ri U (r Ri0 ) ,
(1.225)
i
where i runs from 1 to Nion 13 . The deviation Ri may be expanded in terms of the vibrational normal modes of the lattice, i.e. the phonons, as
1 X
Ri =
Nion q
13
~
2 (q)
!1/2
0
(1.226)
46
e (q) e (q) = M 1 ,
(1.228)
where M is the ionic mass. The number of unit cells in the crystal is Nion = V /, where
is the Wigner-Seitz
cell volume. Again, we approximate Bloch states by plane waves
k (r) = exp(ik r)/ V , in which case
2
0
i
0
0 4Ze (k k )
k0 U (r Ri0 ) k = ei(kk )Ri
.
2
V
(k k0 )2 + TF
(1.229)
mod G
(1.230)
i=1
so that
1 X
Helph =
g (k, k0 ) (aq + aq ) k
k0 k0 ,k+q+G
V kk0
q G
(1.231)
47
with
g (k, k + q + G) = i
!1/2
4Ze2
(q + G) e (q) .
(q + G)2 + 2
TF
2 (q)
(1.232)
M and hence,
Restricting our attention to the longitudinal phonon, we have eL (q) = q/
for small q = k0 k,
gL (k, k + q) = i
~
2M
1/2
(1.233)
where cL is the longitudinal phonon velocity. Thus, for small q we that the electronlongitudinal phonon coupling gL (k, k + q) gq satisfies
|gq |2 = elph
~cL q
g(F )
(1.234)
where g(F ) is the electronic density of states, and where the dimensionless electron-phonon
coupling constant is
elph
2Z m
=
=
3 M
2M c2L g(F )
Z2
F
kB s
!2
,
(1.235)
with s ~cL kF /kB . Table 1.3 lists s , the Debye temperature D , and the electron-phonon
coupling elph for various metals.
EXERCISE: Derive eqn. (1.235).
Metal
Na
K
Cu
Ag
s
220
150
490
340
D
150
100
315
215
elph
0.47
0.25
0.16
0.12
Metal
Au
Be
Al
In
s
310
1940
910
300
D
170
1000
394
129
elph
0.08
0.59
0.90
1.05
Table 1.3: Electron-phonon interaction parameters for some metals. Temperatures are in
Kelvins.
14
G 6= 0 Umklapp processes.
48
1.9.3
Earlier we had quoted the result for the electron-phonon collision integral,
n
2 X
Ik {f, n} =
|g (k, k0 )|2 (1 fk ) fk0 (1 + nq, ) (k + ~q k0 )
~V 0
k ,
+(1 fk ) fk0 nq (k ~q k0 )
fk (1 fk0 ) (1 + nq ) (k ~q k0 )
o
fk (1 fk0 ) nq (k + ~q k0 ) q,k0 k mod
.
(1.236)
The four terms inside the curly brackets correspond, respectively, to cases (a) through (d)
in fig. 1.1. The (1 + n) factors in the phonon emission terms arise from both spontaneous
as well as stimulated emission processes. There is no spontaneous absorption.
EXERCISE: Verify that in equilibrium Ik {f 0 , n0 } = 0.
In principle we should also write down a Boltzmann equation for the phonon distribution
nq and solve the two coupled sets of equations. The electronic contribution to the phonon
collision integral is written as Jq {f, n}, with
nq
4 2 X
=
g
(1 + nq ) fk+q (1 fk )
Jq {f, n}
t coll ~V q
k
nq fk (1 fk+q ) (k+q k ~q ) .
(1.237)
Here, we will assume that the phonons are always in equilibrium, and take nq = n0q .
Phonon equilibrium can be achieved via anharmonic effects (i.e. phonon-phonon scattering),
or by scattering of phonons from impurities or crystalline defects. At low temperatures,
impurity scattering
A
1
2
3
= B T
(1.238)
anharmonic phonon scattering
()
C/L
boundary scattering (L = crystal size)
where A, B, and C are constants.
We now linearize Ik {f }, and obtain
i
2 X 2 h
0
L f =
gq
(1 fk0 + n0q )fk+q (fk+q
+ n0q )fk (k+q k ~q )
~V
q
h
i
0
0
0
0
(1 fk+q + nq )fk (fk + nq )fk+q (k+q k + ~q ) .
(1.239)
This integral operator must be inverted in order to solve for fk in
f 0
L f = e v E
.
(1.240)
49
(1.242)
k
dk
F
3
2
8 ~ m kF / 2 ~2 ~kF
~ 2
= elph
(2kB s ~) .
kB s
2 F () =
(1.243)
Z
n
o
d 2 F () f 0 ( + ~) f 0 ( ~) + 2n0 () + 1 .
(1.244)
elph 2 |F |3
2elph 2
3
~ (kB s ) it | F | > 2kB s .
(1.245)
Note that (F ) = , unlike the case of impurity scattering. This is because at T = 0 there
are no phonons! For T 6= 0, the divergence is cut off, and one obtains
2elph kB T 3
1
2s
=
G
(1.246)
()
~
2s
T
Zy
4 (3) if y =
2
x
G(y) = dx
(1.247)
=
2 sinh x
1
0
if y 1 ,
4y
50
and so
7(3)
2~
kB T 3
2s
elph
1
=
()
2 k T
elph
~ B
if T s
(1.248)
if T s
Chapter 2
Mesoscopia
2.1
References
2.2
Introduction
Current nanofabrication technology affords us the remarkable opportunity to study condensed matter systems on an unprecedented small scale. For example, small electron boxes
known as quantum dots have been fabricated, with characteristic size ranging from 10 nm
to 1 m; the smallest quantum dots can hold as few as one single electron, while larger
dots can hold thousands. In systems such as these, one can probe discrete energy level
spectra associated with quantization in a finite volume. Oftentimes systems are so small
that Blochs theorem and the theoretical apparatus of Boltzmann transport are of dubious
utility.
2.3
52
CHAPTER 2. MESOSCOPIA
(2.1)
Here, N () is the density of states in the leads per spin degree of freedom, and corresponding
to motion in a given direction (right or left but not both); v() is the velocity, and f (L,R )
are the respective Fermi distributions. T () is the transmission probability that electrons of
energy emerging from the left reservoir will end up in right reservoir; R0 () is the reflection
probability that electrons emerging from right reservoir will return to the right reservoir.
The three terms on the right hand side of (2.1) correspond, respectively, to: (i) electrons
emerging from L which make it through the wire and are deposited in R, (ii) electrons
emerging from R which fail to swim upstream to L and are instead reflected back into
R, and (iii) all electrons emerging from reservoir R (note this contribution is of opposite
sign). The transmission and reflection probabilities are obtained by solving for the quantum
mechanical scattering due to the disordered region. If the incoming flux amplitudes from
the left and right sides are i and i0 , respectively, and the outgoing flux amplitudes on those
sides o0 and o, linearity of the Schrodinger equation requires that
0
r t0
i
o
.
(2.2)
;
S=
=S 0
t r0
i
o
The matrix S is known as the scattering matrix (or S-matrix, for short). The S-matrix elements r, t, etc. are reflection and transmission amplitudes. The reflection and transmission
probabilities are given by
R = |r|2
2
T = |t|
T 0 = |t0 |2
0
0 2
R = |r | .
(2.3)
(2.4)
Going back to (2.1), let us assume that we are close to equilibrium, so the difference R L
in chemical potentials is slight. We may then expand
f ( R ) = f ( L ) + f 0 ( L ) (L R ) + . . .
(2.5)
Z
e
f 0
= (R L ) d
T () ,
h
(2.6)
valid to lowest order in (R L ). We have invoked here a very simple, very important
result for the one-dimensional density of states. Considering only states moving in a definite
direction (left or right) and with a definite spin polarization (up or down), we have
N () d =
dk
2
N () =
1 dk
1
=
2 d
hv()
(2.7)
53
e
( L ) T (F ) ,
h R
(2.8)
where T (F ) is the transmission probability at the Fermi energy. The chemical potential
varies with voltage according to (V ) = (0) eV , hence the conductance G = I/V is
found to be
G=
=
e2
T (F )
h
2e2
T (F )
h
(2.9)
(2.10)
h
e2
Gi =
e2 T (F )
e2 T (F )
=
,
h 1 T (F )
h R(F )
(2.11)
where Gi is the intrinsic conductance of the wire, per spin channel. Now we see that when
T (F ) 1 the intrinsic conductance diverges: Gi . When T 1, Gi G = (e2 /h)T .
This result (2.11) is known as the Landauer Formula.
To derive this result in a more systematic way, let us assume that the disordered segment
is connected to the left and right reservoirs by perfect leads, and that the leads are not in
54
CHAPTER 2. MESOSCOPIA
(2.14)
(2.15)
where the factor of two accounts for both directions of motion. To lowest order, then, we
obtain
2(L
L ) = (L R ) T 0
2(R
R ) = (R L ) T
=
=
L = L + 21 T 0 (R L )
1
2
R = R + T (L R )
(2.16)
(2.17)
and therefore
(
L
R ) = 1 21 T 12 T 0 (L R )
= (1 T ) (L R ) ,
(2.18)
where the last equality follows from unitarity (S S = SS = 1). There are two experimental
configurations to consider:
Two probe measurement Here the current leads are also used as voltage leads. The
voltage difference is V = (R L )/e and the measured conductance is G2probe =
(e2 /h) T (F ).
Four probe measurement Separate leads are used for current and voltage probes.
The observed voltage difference is V = (
R
L )/e and the measured conductance
2
is G4probe = (e /h) T (F )/R(F ).
2.3.1
Perhaps the simplest scattering problem is one-dimensional scattering from a potential step,
V (x) = V0 (x). The potential is piecewise constant, hence the wavefunction is piecewise a
plane wave:
x<0:
x>0:
ik0 x
0 ik0 x
+I e
(2.19)
,
(2.20)
55
~2 k 0 2
~2 k 2
=
+ V0 .
(2.21)
2m
2m
The requirement that (x) and its derivative 0 (x) be continuous at x = 0 gives us two
equations which relate the four wavefunction amplitudes:
E=
I + O0 = O + I 0
0
(2.22)
0
k(I O ) = k (O I ) .
As emphasized earlier, the S-matrix acts on flux amplitudes. We have
i
o
I
O
0
= v
= v
,
,
o0
i0
O0
I0
(2.23)
(2.24)
with v = ~k/m and v 0 = ~k 0 /m. One easily finds the S-matrix, defined in eqn. 2.2, is given
by
2
1
t
r0
1+
1+
=
S=
,
(2.25)
2
1
r
t0
1+
1+
q
where v 0 /v = k 0 /k = 1 VE0 , where E = F is the Fermi energy. The two- and
four-terminal conductances are then given by
e2 2 e2
4
|t| =
h
h (1 + )2
e2 |t|2
e2
4
=
=
.
h |r|2
h (1 )2
G2probe =
(2.26)
G4probe
(2.27)
Both are maximized when the transmission probability T = |t|2 = 1 is largest, which occurs
for = 1, i.e. k 0 = k.
56
CHAPTER 2. MESOSCOPIA
Figure 2.2: Dimensionless two-terminal conductance g versus k 0 /k for the potential step.
The conductance is maximized when k 0 = k.
2.4
Multichannel Systems
The single channel scenario described above is obtained as a limit of a more general multichannel case, in which there are transverse degrees of freedom (due e.g. to finite crosssectional area of the wire) as well. We will identify the transverse states by labels i. Within
the perfect leads, the longitudinal and transverse energies are decoupled, and we may write
= i + k (k) ,
(2.28)
where k (k) is the one-dimensional dispersion due to motion along the wire (e.g. k (k) =
~2 k 2 /2m , k (k) = 2t cos ka, etc.). k is the component of the wavevector along the axis of
the wire. We assume that the transverse dimensions are finite, so fixing the Fermi energy F
in turn fixes the total number of transverse channels, Nc , which contribute to the transport:
Nc () =
( i )
(continuum)
(2.29)
(tight binding) .
(2.30)
2t | i |
Equivalently, an electron with energy in transverse state i has wavevector ki which satisfies
k (ki ) = i .
(2.31)
57
Nc is the number of real positive roots of (2.31). Typically Nc kFd1 A, where A is the
cross-sectional area and kF is the Fermi wavevector. The velocity vi is
1 k (k)
1
=
vi () =
.
(2.32)
~ k ki
~ k
k=ki
The density of states Ni () (per unit spin, per direction) for electrons in the ith transverse
channel is
Z
dk
1 dk
Ni () =
v(k) i k (k) =
,
(2.33)
2
2 d
k k=ki
Nc n
o
X
left (x , z) =
Ij e+ikj z + Oi0 eikj z Lj (x )
(2.34)
j=1
R
Nc n
o
X
Oa e+ika z + Ia0 eika z Ra (x ) .
right (x , z) =
(2.35)
a=1
Here, we have assumed a general situation in which the number of transverse channels NcL,R
may differ between the left and right lead. The quantities {Ij , Oj0 , Oa , Ia0 } are wave function
amplitudes. The S-matrix, on the other hand, acts on flux amplitudes {ij , o0j , oa , i0a }, which
are related to the wavefunction amplitudes as follows:
1/2
Ii
1/2
Oi0
ii = vi
o0i = vi
oa = va1/2 Oa
i0a = va1/2 Ia0 .
(2.36)
(2.37)
rN L N L
c
c
tN R N L
c
t0N L N R
c
c
0
rN
R N R
c
!
(2.38)
z }| {
0
o
r t0
i
=
.
t r0
i0
o
Unitarity of S means that S S = SS = I, where
r t0
r
t
S=
=
S = 0 0 ,
t r0
t
r
(2.39)
(2.40)
58
CHAPTER 2. MESOSCOPIA
rik rjk
+ t0ic t0
jc = ij
0
0
tak tbk + rac
rbc
0
rik tak + t0ic rac
rki
rkj + tci tcj = ij
0
0 0
t0
ka tkb + rca rcb
0
0
rki
tka + tci rca
= ab
=0
(2.41)
= ab
(2.42)
=0,
(2.43)
r r + t t
= IN L N L
(2.44)
t t + r0 r0 = t0 t0 + r0 r0 = IN R N R
(2.45)
r t + t0 r0 =
= ON L N R
(2.46)
= ON R N L .
(2.47)
r t0 + t r0
t r + r0 t0 =
t0 r + r0 t
Ri =
Nc
X
rik rik
Ta =
Nc
X
k=1
Nc
X
(2.48)
0 0
rac
rac ,
(2.49)
k=1
Ti0
tak tak
t0ic t0
ic
Ra0
Nc
X
c=1
c=1
Ra0 + Ta = 1
(2.50)
for all i {1, . . . , NcL } and a {1, . . . , NcR }. Unitarity of the S-matrix preserves particle
flux:
|i|2 |i0 |2 = |o|2 |o0 |2 ,
(2.51)
Nc
X
j=1
|ij | +
Nc
X
a=1
|i0a |2
Nc
X
|oa | +
a=1
Nc
X
j=1
|o0j |2 .
(2.52)
59
Let us now compute the current in the right lead flowing past the imaginary surface
Ta ()
R
I = e
Nc Z
X
(
d Na () va ()
z
}|
{
NcL
X
tai ()2 f ( L )
a=1
i=1
0 ()
Ra
"
+
z
}|
{
#
)
NcR
X
2
0
r () 1 f ( R )
ab
b=1
NcR
Z
e
f 0 X
= (R L ) d
Ta () .
h
(2.53)
a=1
G2probe
e2
=
=
h
R L
eI
NcR
f 0 X
Ta () .
d
(2.54)
a=1
e2
Tr tt
h
(2.55)
where
L
Nc Nc
X
X 2
tai .
Tr tt = Tr t t =
(2.56)
i=1 a=1
nL =
Nc Z
X
0
d Ni () 1 + Ri () f ( L ) + Ti () f ( R )
(2.57)
i=1
=2
XZ
d Ni () f (
L )
(2.58)
L = L 21 T 0 (L R )
(2.59)
(2.60)
60
CHAPTER 2. MESOSCOPIA
nR =
Nc Z
X
0
d Na () Ta () f ( L ) + 1 + Ra () f ( R )
(2.61)
a=1
=2
XZ
d Na () f (
R )
(2.62)
R = R + 12 T (L R )
where
(2.63)
P 1
a va Ta
T P
1 .
a va
(2.64)
(We have assumed zero temperature throughout.) The difference in lead chemical potentials
is thus
(2.65)
(
L
R ) = 1 12 T 21 T 0 (L R ) .
Hence, we obtain the 4-probe conductance,
G4probe =
2.4.1
P
T
e2
a a P
P
1
1 P 1
1 P
1
h 1 1
0
v
v
T
v
T
v
i i
a a
i i i
a a a
2
2
(2.66)
The transfer matrix S acts on incoming flux amplitudes to give outgoing flux amplitudes.
This linear relation may be recast as one which instead relates flux amplitudes in the right
lead to those in the left lead, i.e.
M
z }| {
0
o
r t0
i
=
o
t r0
i0
z
}|
!{
o
M11 M12
i
=
.
0
i0
o
M21 M22
(2.67)
M is known as the transfer matrix. Note that each of the blocks of M is of dimension
NcR NcL , and M itself is a rectangular 2NcR 2NcL matrix. The individual blocks of M
are readily determined:
o0 = r i + t0 i0
0 0
o = ti + r i
i0 = t01 r i + t01 o0
0 01
o = (t r t
(2.68)
0 01 0
r) i + r t
o ,
(2.69)
so we conclude
M11 = t1
01
M21 = t
M12 = r0 t01
r
M22 = t
01
(2.70)
(2.71)
WARNING: None of this makes any sense if NcL 6= NcR ! The reason is that it is problematic to take the inverse of a rectangular matrix such as t or t0 , as was blithely done
61
Figure 2.3: Two quantum scatterers in series. The right side data for scatterer #1 become
the left side data for scatterer #2.
above in (2.68,2.69). We therefore must assume NcR = NcL = Nc , and that the scatterers are
separated by identical perfect regions. Practically, this imposes no limitations at all, since
the width of the perfect regions can be taken to be arbitrarily small.
EXERCISE: Show that M11 = t r0 t01 r = t1 .
The virtue of transfer matrices is that they are multiplicative. Consider, for example, two
disordered regions connected by a region of perfect conductor. The outgoing flux o from
the first region becomes the incoming flux i for the second, as depicted in fig. 2.3. Thus,
if M1 is the transfer matrix for scatterer #1, and M2 is the transfer matrix for scatterer
#2, the transfer matrix for the two scatterers in succession is M = M2 M1 :
M = M2 M1
z
11
}| 11
{
12
12
12
11
o2
M2 M2
i2
M2 M2
M1 M1
i1
.
=
=
21 M22
21 M22
0
22
0
o
M21
M
o
M
M
i02
2
2
1
1
1
2
2
2
(2.72)
(2.73)
Unitarity of the S-matrix means that the transfer matrix is pseudo-unitary in that it satisfies
!
I
O
Nc Nc
Nc Nc
M M =
where =
.
(2.74)
O
I
Nc Nc
Nc Nc
(2.75)
62
CHAPTER 2. MESOSCOPIA
1
4
tt 0
.
0 t0 t0
(2.76)
= M =
M11 M21
M12 M22
!
.
(2.77)
2.4.2
(2.79)
(2.80)
from which we conclude that p() = 0 implies p(1 ) = 0, and the eigenvalues of M M
come in (, 1 ) pairs. We can therefore write Pichards formula as
G2probe =
Nc
e2 X
4
,
1
h
i=1 i + i + 2
(2.81)
where without loss of generality we assume i 1 for each i {1, . . . , Nc }. We define the
ith localization length i through
2L
2L
i exp
=
i =
,
(2.82)
i
ln i
63
Nc
e2 X
2
h
1 + cosh(2L/i )
(2.83)
i=1
Suppose now that L is finite, and furthermore that 1 > 2L > N . Channels for which
c
2L j give cosh(2L/j ) 1, and therefore contribute a quantum of conductance e2 /h
to G. These channels are called open. Conversely, when 2L j , we have cosh(2L/j )
1
2
2L/j to the
2 exp(2L/j ) 1, and these closed channels each contribute Gj = (2e /h)e
conductance, a negligible amount. Thus,
G(L) '
e2 open
N
h c
Ncopen
Nc
X
(j 2L) .
(2.85)
j=1
Of course, Ncclosed = Nc Ncopen , although there is no precise definition for open vs. closed
for channels with j 2L. This discussion naturally leads us to the following classification
scheme:
When L > 1 , the system is in the localized regime. The conductance vanishes exponentially with L according to G(L) (4e2 /h) exp(2L/), where () = 1 () is the
localization length. In the localized regime, there are no open channels: Ncopen = 0.
When Ncopen = ` Nc /L, where ` is the elastic scattering length, one is in the Ohmic
regime. In the Ohmic regime, for a d-dimensional system of length L and ((d 1)dimensional) cross-sectional area A,
e2 d1 A
e2 ` d1
kF A =
k
`
.
(2.86)
hL
h F
L
Note that G is proportional to the cross sectional area A and inversely proportional
to the length L, which is the proper Ohmic behavior: G = A/L, where
GOhmic
is the conductivity.
e2 d1
k
`
h F
(2.87)
64
CHAPTER 2. MESOSCOPIA
When L < N , all the channels are open: Ncopen = Nc . The conductance is
c
G(L) =
e2 d1
e2
Nc
k
A.
h
h F
(2.88)
()
,
1 + cosh
(2.89)
where
() =
Nc
X
( i )
(2.90)
i=1
R
is the density of values. This distribution is normalized so that 0 d () = Nc . Spectral properties of the {i } thus determine the statistics of the conductance. For example,
averaging over disorder realizations gives
Z
hgi = 2 d
()
.
1 + cosh
(2.91)
2.4.3
() ( 0 )
.
(1 + cosh )(1 + cosh 0 )
(2.92)
Let us consider the case of two scatterers in series. For simplicity, we will assume that
Nc = 1, in which case the transfer matrix for a single scatterer may be written as
1/t r /t
M=
.
(2.93)
r/t0
1/t0
A pristine segment of wire of length L has a diagonal transfer matrix
i
e
0
N =
,
0 ei
(2.94)
65
where = kL. Thus, the composite transfer matrix for two scatterers joined by a length L
of pristine wire is M = M2 N M1 , i.e.
i
1/t2
r2 /t2
e
0
1/t1
r1 /t1
M=
.
(2.95)
r2 /t02
1/t02
r1 /t01
1/t01
0 ei
In fact, the inclusion of the transfer matrix N is redundant; the phases ei can be completely absorbed via a redefinition of {t1 , t01 , r1 , r10 }.
Extracting the upper left element of M gives
1
ei ei r10 r2
=
,
t
t1 t2
(2.96)
T1 T2
1 + R1 R2 2 R1 R2 cos
R
R1 + R2 2 R1 R2 cos
R=
=
T
T1 T2
p
= R1 + R2 + 2 R1 R2 2 R1 R2 (1 + R1 )(1 + R2 ) cos .
(2.97)
(2.98)
(2.99)
The first two terms correspond to Ohms law. The final term is unfamiliar and leads to
a divergence of resistivity as a function of length. To see this, imagine that that R2 =
% dL is small, and solve (2.99) iteratively. We then obtain a differential equation for the
dimensionless resistance R(L):
dR = (1 + 2R) % dL
=
R(L) = 21 e2%L 1 .
(2.100)
In fact, the distribution PL (R) is extremely broad, and it is more appropriate to average
the quantity ln(1 + R). Using
Z2
p
d
ln(a b cos ) = ln 21 a + 12 a2 b2
2
(2.101)
with
a = 1 + R1 + R2 + 2 R1 R2
p
b = 2 R1 R2 (1 + R1 )(1 + R2 ) ,
(2.102)
(2.103)
66
CHAPTER 2. MESOSCOPIA
(2.104)
x(L) ln 1 + R(L) ,
(2.105)
(2.106)
x(L) 1 2%L
e
=2 e
+1 .
(2.107)
Note that hex i 6= ehxi . The quantity x(L) is an appropriately self-averaging quantity in
that its root mean square fluctuations are small compared to its average, i.e. it obeys the
central limit theorem. On the other hand, R(L) is not self-averaging, i.e. it is not normally
distributed.
Abelian Multiplicative Random Processes
Let p(x) be a distribution on the nonnegative real numbers, normalized according to
Z
dx p(x) = 1 ,
(2.108)
and define
X
N
Y
xi
Y ln X =
i=1
N
X
ln xi .
(2.109)
i=1
Z
=
i=1
d iY
e
2
N
Z
dx p(x) ei ln x
iN
d iY h
e
1 ihln xi 21 2 hln2 xi + O( 3 )
2
d i(Y N hln xi) 1 N 2 (hln2 xihln xi2 )+O(3 )
e
e 2
2
1
2N 2
n
o
1 + O(N 1 ) ,
(2.110)
(2.111)
2 /2N 2
e(Y N )
with
= hln xi
and
f (x)
67
Z
dx p(x) f (x) .
(2.112)
(2.113)
2 /2 2
(2.114)
hX k i
2
= eN k(k1)/2 ,
k
hXi
(2.115)
(2.116)
2.4.4
The case of parallel quantum resistors is more difficult than that of series resistors. The
reason for this is that the conduction path for parallel resistances is multiply connected, i.e.
electrons can get from start to finish by traveling through either resistor #1 or resistor #2.
Consider electrons with wavevector k > 0 moving along a line. The wavefunction is
(x) = I eikx + O0 eikx ,
hence the transfer matrix M for a length L of pristine wire is
ikL
e
0
M(L) =
.
0
eikL
(2.117)
(2.118)
68
CHAPTER 2. MESOSCOPIA
Now lets bend our wire of length L into a ring. We therefore identify the points x = 0 and
x = L = 2R, where R is the radius. In order for the wavefunction to be single-valued we
must have
h
i I
M(L) I
=0,
(2.119)
O0
and in order to have a nontrivial solution (i.e. I and O0 not both zero), we must demand
det (M I) = 0, which says cos kL = 1, i.e. k = 2n/L with integer n. The energy is then
quantized: n = k (k = 2n/L).
Next, consider the influence of a vector potential on the transfer matrix. Let us assume the
vector potential A along the direction of motion is nonzero over an interval from x = 0 to
x = d. The Hamiltonian is given by the Peierls substitution,
e
A(x) .
(2.120)
H = k ix + ~c
Note that we can write
H = (x) k (ix ) (x)
( Zx
)
ie
0
0
(x) = exp
dx A(x ) .
~c
(2.121)
(2.122)
(2.123)
(2.124)
with
e
=
~c
Zd
dx A(x) .
(2.125)
We are free to choose any gauge we like for A(x). The only constraint is that the gaugeinvariant content, which is encoded in he magnetic fluxes through every closed loop C,
I
C = A dl ,
(2.126)
C
must be preserved. On a ring, there is one flux to speak of, and we define the dimensionless
flux = e/~c = 2/0 , where 0 = hc/e = 4.137107 Gcm2 is the Dirac flux quantum.
In a field of B = 1 kG, a single Dirac quantum is enclosed by a ring of radius R = 0.11 m.
It is convenient to choose a gauge in which A vanishes everywhere along our loop except for
69
Figure 2.4: Energy versus dimensionless magnetic flux for free electrons on a ring. The
degeneracies are lifted in the presence of a crystalline potential.
a vanishingly small region, in which all the accrued vector potential piles up in a -function
of strength . The transfer matrix for this infinitesimal region is then
i
e
0
= ei I .
(2.127)
M() =
0 ei
If k > 0 corresponds to clockwise motion around the ring, then the phase accrued is ,
which explains the sign of in the above equation.
For a pristine ring, then, combining the two transfer matrices gives
ikL i
i
ikL
e e
0
e
0
e
0
=
,
M=
0
eikL ei
0
eikL
0 ei
(2.128)
(right-movers)
kL = 2n +
(left-movers) .
Note that different n values are allowed for right- and left-moving branches since by assumption k > 0. We can simplify matters if we simply write (x) = A eikx with k unrestricted
in sign, in which case k = (2n )/L with n chosen from the entire set of integers. The
allowed energies for free electrons are then
n () =
which are plotted in fig. 2.4.
2 2 ~2
2
n
,
2
mL
2
(2.129)
70
CHAPTER 2. MESOSCOPIA
Figure 2.5: Scattering problem for a ring enclosing a flux . The square and triangular
blocks represent scattering regions and are describes by 2 2 and 3 3 S-matrices, respectively. The dot-dash line represents a cut across which the phase information due to the
enclosed flux is accrued discontinuously.
Now let us add in some scatterers. This problem was first considered in a beautiful paper
by B
uttiker, Imry, and Azbel, Phys. Rev. A 30, 1982 (1984). Consider the ring geometry
depicted in fig. 2.5. We want to compute the S-matrix for the ring. We now know
how to describe the individual quantum resistors #1 and #2 in terms of S-matrices (or,
equivalently, M-matrices). Assuming there is no magnetic field penetrating the wire (or
that the wire itself
thin), we have S = S t for each scatterer. In this case,
is infinitesimally
0
i
we have t = t = T e . We know |r|2 = |r0 |2 = 1 |t|2 , but in general r and r0 may have
different phases. The most general 2 2 transfer matrix, under conditions of time-reversal
symmetry, depends on three parameters, which may be taken to be the overall transmission
probability T and two phases:
i
i
1
e
1
T
e
M(T, , ) =
.
(2.130)
1 T ei
ei
T
We can include the effect of free-particle propagation in the transfer matrix M by multiplying M on the left and the right by a free propagation transfer matrix of the form
i/4
e
0
N =
,
(2.131)
0
ei/4
where = kL = 2kR is the phase accrued by a particle of wavevector k freely propagating
once around the ring. N is the transfer matrix corresponding to one quarter turn around
the ring. One easily finds
M N M(T, , ) N = M(T, + 12 , ).
(2.132)
For pedagogical reasons, we will explicitly account for the phases due to free propagation,
and write
0
0
1
1
2
2
f
= N M1 N
,
= N M2 N
,
(2.133)
0
0
1
1
2
2
71
1 0
0 1
(2.134)
0 1
1 0
.
(2.135)
We now have to model the connections between the ring and the leads, which lie at the
confluence of three segments. Accordingly, these regions are described by 3 3 S-matrices.
The constraints S = S (unitarity) and S = S t (time-reversal symmetry) reduce the number
of independent real parameters in S from 18 to 5. Further assuming that the scattering
is symmetric with respect to the ring branches brings this number down to 3, and finally
assuming S is real reduces the dimension of the space of allowed S-matrices to one. Under
these conditions, the most general 3 3 S-matrix may be written
(a + b)
a
b
(2.136)
b
a
where
(a + b)2 + 2 = 1
(2.137)
a2 + b2 + = 1 .
(2.138)
The parameter , which may be taken as a measure of the coupling between the ring and
the leads ( = 0 means ring and leads are decouped) is restricted to the range 0 12 .
There are four solutions for each allowed value of :
a = 12 1 2 1 , b = 12 1 2 + 1
(2.139)
and
a = 12
1 2 + 1
b = 12
1 2 1 .
(2.140)
We choose the first pair, since it corresponds to the case |b| = 1 when = 0, i.e. perfect
transmission through the junction. We choose the top sign in (2.139).
We therefore have at the left contact,
0
(aL + bL )
L
L
o
i
0
2 =
,
aL
bL
2
L
0
1
1
L
bL
aL
(2.141)
(aR + bR )
R
R
i
o
e
0
e1 =
aR
bR 1 ,
R
20
2
R
bR
aR
(2.142)
72
CHAPTER 2. MESOSCOPIA
Figure 2.6: Two-probe conductance G(, ) of a model ring with two scatterers. The
enclosed magnetic flux is ~c/e, and = 2kR (ring radius R). G vs. curves for various
values of are shown.
where accounting for the vector potential gives us
e1 = ei 1
e10 = ei 10 .
(2.143)
We set i = 1 and i0 = 0, so that the transmission and reflection amplitudes are obtained
from t = o and r = o0 .
From (2.141), we can derive the relation
QL
z
1
1
=
10
bL
}|
!{
2
L bL aL
aL
20
L aL
+
.
2
1
aL
1
bL
b2
(2.144)
z
0
1
2
=
2
bR
}|
!{
2
aR
1
i
R aR
e
.
0
aR
1
1
b2
(2.145)
The matrices QL and QR resemble transfer matrices. However, they are not pseudo-unitary:
Q Q 6= . This is because some of the flux can leak out along the leads. Indeed, when
73
Figure 2.7: Two-probe conductance G(, ) of a model ring with two scatterers. The
enclosed magnetic flux is ~c/e, and = 2kR (ring radius R). G vs. curves for various
values of are shown. The coupling between leads and ring is one tenth as great as in
fig. 2.6, and accordingly the resonances are much narrower. Note that the resonances at
= 0.45 and = 0.85 are almost completely suppressed.
= 0, we have b = 1 and a = 0, hence Q = 1, which is pseudo-unitary (i.e. flux preserving).
Combining these results with those in (2.133), we obtain the solution
n
o1
1
bL aL
L
i
f
= I e QL N M2 N QR N M1 N
.
(2.146)
10
1
bL
From this result, using (2.133), all the flux amplitudes can be obtained.
We can define the effective ring transfer matrix P as
f N QR N M N ,
P ei QL N M
2
1
(2.147)
which has the following simple interpretation. Reading from right to left, we first move 14 turn clockwise around the ring (N ). Then we encounter scatterer #1 (M1 ). After another
quarter turn (N ), we encounter the right T-junction (QR ). Then its yet another quarter
turn (N ) until scatterer #2 (M2 ), and one last quarter turn (N ) brings us to the left
T-junction (QL ), by which point we have completed one revolution. As the transfer matrix
acts on both right-moving and left-moving flux amplitudes, it accounts for both clockwise
as well as counterclockwise motion around the ring. The quantity
1
IP
= 1 + P + P2 + P3 + . . . ,
(2.148)
then sums up over all possible integer windings around the ring. In order to properly account
for the effects of the ring, an infinite number of terms must be considered; these may be
74
CHAPTER 2. MESOSCOPIA
Figure 2.8: Two-probe conductance G(, ) of a model ring with two scatterers. The
enclosed magnetic flux is ~c/e, and = 2kR (ring radius R). G vs. curves for various
values of are shown.
resummed into the matrix inverse in (2.147). The situation is analogous to what happens
when an electromagnetic wave reflects off a thin dielectric slab. At the top interface, the
wave can reflect. However, it can also refract, entering the slab, where it may undergo an
arbitrary number of internal reflections before exiting.
The transmission coefficient t is just the outgoing flux amplitude: t = o. We have from
(2.142,2.133) that
R 1 ei + 2
R
1
=
bR aR 1
0
bR
1
L R
1
1
1
1
=
Y12
+ Y22
Y11
Y21
bL bR
t=
(2.149)
where
f N QR ei N 1 M1 N 1 .
Y = QL N M
2
1
(2.150)
2.5
75
The conductance of a disordered metal is a function of the strength and location of the
individual scatterers. We now ask, how does the conductance fluctuate when the position
or strength of a scatterer or a group of scatterers is changed. From the experimental point
of view, this seems a strange question to ask, since we generally do not have direct control
over the position of individual scatterers within a bulk system. However, we can imagine
changing some external parameter, such as the magnetic field B or the chemical potential
(via the density n). Using computer modeling, we can even live the dream of altering the
position of a single scatterer to investigate its effect on the overall conductance. Navely, we
would expect there to be very little difference in the conductance if we were to, say, vary the
position of a single scatterer by a distance `, or if we were to change the magnetic field by
B = 0 /A, where A is the cross sectional area of the system. Remarkably, though, what
is found both experimentally and numerically is that the conductance exhibits fluctuations
with varying field B, chemical potential , or impurity configuration (in computer models).
The root-mean-square magnitude of these fluctuations for a given sample is the same as
that between different samples, and is on the order G e2 /h. These universal conductance
fluctuations (UCF) are independent on the degree of disorder, the sample size, the spatial
dimensions, so long as the inelastic mean free path (or phase breaking length) satisfies
L > L, i.e. the system is mesoscopic.
Theoretically the phenomenon of UCF has a firm basis in diagrammatic perturbation theory.
Here we shall content ourselves with understanding the phenomenon on a more qualitative
level, following the beautiful discussion of P. A. Lee in Physica 140A, 169 (1986). We begin
with the multichannel Landauer formula,
Nc
e2
e2 X
G=
Tr tt =
|taj |2 .
h
h
(2.151)
a,j=1
The transmission amplitudes taj can be represented as a quantum mechanical sum over
paths ,
X
taj =
Aaj () ,
(2.152)
where Aaj () is the probability amplitude for Feynman path to connect channels j and a.
The sum is over all such Feynman paths, and ultimately we must project onto a subspace of
definite energy this is, in Lees own words, a heuristic argument. Now assume that the
Aaj () are independent complex random variables. The fluctuations in |taj |2 are computed
from
X
|taj |4 =
Aaj (1 ) Aaj (2 ) Aaj (3 ) Aaj (4 )
1 ,2
3 ,4
DX
E2
X
Aaj ()2 +
Aaj ()4
=2
2 n
= 2 |taj |2 1 + O(M 1 ) ,
(2.153)
76
CHAPTER 2. MESOSCOPIA
where M is the (extremely large) number of paths in the sum over . As the O(M 1 ) term
is utterly negligible, we conclude
2
|taj |4 |taj |2
=1.
(2.154)
2
|taj |2
Now since
var(G) = hG2 i hGi2
o
e4 X n
= 2
|taj |2 |ta0 j 0 |2 |taj |2 |taj |2
,
h a,a0
(2.155)
(2.156)
j,j 0
we also need to know about the correlation between |taj |2 and |ta0 j 0 |2 . The simplest assumption is to assume they are uncorrelated unless a = a0 and j = j 0 , i.e.
2
|taj |2 |ta0 j 0 |2 |taj |2 |taj |2 = |taj |4 |taj |2
aa0 jj 0 ,
(2.157)
in which case the conductance is a given by a sum of Nc2 independent real
random
variables,
each of which has a standard deviation equal to its mean, i.e. equal to |taj |2 . According
to the central limit theorem, then, the rms fluctuations of G are given by
G =
e2
var(G) = Nc |taj |2 .
h
(2.158)
Further assuming that we are in the Ohmic regime where G = Ld2 , with (e2 /h) kFd1 `,
and Nc (kF L)d1 , we finally conclude
1 `
|taj |2
Nc L
e2 `
.
h L
(2.159)
This result is much smaller than the correct value of G (e2 /h).
To reiterate the argument in terms of the dimensionless conductance g,
`
g
=
|taj |2 = 2
L
Nc
a,j
o
X n
var(g) =
|taj |2 |ta0 j 0 |2 |taj |2 |taj |2
g=
|taj |2 ' Nc
(2.160)
a,a0
j,j 0
X n
2 o
|taj |4 |taj |2
aj
g 2
X
2
g2
=
=
|taj |2 Nc2
Nc2
Nc2
(2.161)
aj
p
g
`
var(g)
=
Nc
L
(WRONG!)
(2.162)
77
What went wrong? The problem lies in the assumption that the contributions Aaj () are
independent for different paths . The reason is that in disordered systems there are certain
preferred channels within the bulk along which the conduction paths run. Different paths
will often coincide along these channels. A crude analogy: whether youre driving from
La Jolla to Burbank, or from El Cajon to Malibu, eventually youre going to get on the 405
freeway anyone driving from the San Diego area to the Los Angeles area must necessarily
travel along one of a handful of high-volume paths. The same is not true of reflection,
though! Those same two hypothetical drivers executing local out-and-back trips from home
will in general travel along completely different, hence uncorrelated, routes. Accordingly,
let us compute var(Nc g), which is identical to var(g), but is given in terms of a sum over
reflection coefficients. We will see that making the same assumptions as we did in the case
of the transmission coefficients produces the desired result. We need only provide a sketch
of the argument:
X
Nc g
`
Nc g =
|rij |2 ' Nc 1
=
|rij |2 =
,
(2.163)
L
Nc2
i,j
so
var(Nc g) =
X n
o
|rij |2 |ri0 j 0 |2 |rij |2 |ri0 j 0 |2
i,i0
j,j 0
X n
2 o
|rij |4 |rij |2
ij
N g 2
X
2
` 2
c
=
1
=
|rij |2 Nc2
Nc2
L
(2.164)
ij
p
p
`
var(Nc g) = var(g) = 1
L
(2.165)
Each path will have a time-reversed mate T for which, in the absence of external magnetic
fields,
Aij () = Aij ( T ) .
(2.167)
This is because the action functional,
Zt2 n
o
e
S[r(t)] = dt 21 mr 2 V (r) A(r) r
c
t1
(2.168)
78
CHAPTER 2. MESOSCOPIA
satisfies
S[r(t)] = S[r(t)]
if B = 0 .
(2.169)
There is, therefore, an extra negative contribution to the conductance G arising from phase
coherence of time-reversed paths. In the presence of an external magnetic field, the path
and its time-reversed mate T have a relative phase = 4 /0 , where is the
magnetic flux enclosed by the path . A magnetic field, then, tends to destroy the phase
coherence between time-reversed paths, and hence we expect a positive magnetoconductance
(i.e. negative magnetoresistance) in mesoscopic disordered metals.
Conductance Fluctuations in Metallic Rings
The conductance of a ring must be periodic under + n0 for any integer n rings
with flux differing by an integer number of Dirac quanta are gauge-equivalent, provided no
magnetic field penetrates the ring itself. The conductance as a function of the enclosed flux
must be of the form
G() = Gcl +
X
m=1
Gm cos
2m
0
+ m
(2.170)
where Gcl is the classical (Boltzmann) conductance of the ring. The second harmonic Gm=2
is usually detectable and is in many cases much larger than the m = 1 term. The origin
of the m = 2 term, which is periodic under + 12 0 , lies in the interference between
time-reversed paths of winding number 1. The m = 1 fundamental is easily suppressed,
e.g. by placing several rings in series.
2.5.1
Weak Localization
2
` +
A(r) +
C(r, r 0 ) = (r r 0 ) ,
(2.172)
~c
(2.173)
79
L (r) = (r)
X (r) (r 0 )
C(r, r 0 ) =
,
(2.175)
(2.176)
which resembles a Greens function from quantum mechanics, where the energy parameter
is identified with / . There is accordingly a path integral representation for C(r, r 0 ):
Z
C(r1 , r2 ) =
ds eis /
( Zs
)
1 r 2 2e
r
Dr[u] exp i du
.
A(r)
4`2 u
~c
u
r (0)=r1
(2.177)
r (s)=r2
Notice that there is no potential term V (r) in the action of (2.177). The effect of the static
random potential here is to provide a step length ` for the propagator. According to the
AAS result (2.171), the corrections to the conductivity involve paths which begin and end
at the same point r in space. The charge e = 2e which appears in the Cooperon action
arises from adding up contributions due to time-reversed paths, as we saw earlier.
Lets try to compute , which is called the weak localization correction to the conductivity.
In the absence of an external field, the eigenvalues of L are simply k2 `2 , where k = 2n /L.
We then have
Z d
dk
1
C(r, r) =
.
(2.178)
2
2
d
(2) k ` +
A cutoff is needed in dimensions d 2 in order to render the integral convergent in the
ultraviolet. This cutoff is the inverse step size for diffusion: `1 . This gives, for ,
`1
e2
ln( / )
d
h
p
` /
(d = 3)
(d = 2)
(2.179)
(d = 1) .
2` 2
`B
(2.180)
p
~c/eB is the magnetic length (for a charge q = e electron). Each of these
80
CHAPTER 2. MESOSCOPIA
` /4`
1
2e2 `2 BX
(B) =
2
1
4`2
h `B
n=0 (n + 2 ) `2 +
B
` /4`
e2 BX
1
=
1
2h
n=0 n + 2 +
`2B
L2
(2.181)
+
,
2 h
2 4`2
2 4L2
(2.182)
where
(z) =
1 d(z)
1
1
= ln z +
+ ...
(z) dz
2z 12z 2
(2.183)
1 e2 L 2
,
6 h `B
(2.184)
which is positive, as previously discussed. The magnetic field suppresses phase coherence
between time-reversed paths, and thereby promotes diffusion by suppressing the resonant
backscattering contributions to . At large values of the field, the behavior is logarithmic.
Generally, we can write
1 e2 B
(B) (0) =
f
,
(2.185)
2 h
B
where
2B
`2
B2
B
4L
and
(
f (x) = ln x +
2.6
1
2
1
x
B =
0
8L2
x2
24
ln x 1.96351 . . .
(2.186)
as x 0
as x .
(2.187)
Anderson Localization
In 1958, P. W. Anderson proposed that static disorder could lead to localization of electronic
eigenstates in a solid. Until this time, it was generally believed that disorder gave rise to
an elastic scattering length ` and a diffusion constant D = vF `. The diffusion constant is
related to the electrical conductivity through the Einstein relation: = 12 e2 D(F )N (F ). If
the states at the Fermi level are localized, then D(F ) = 0.
81
~2 2
+ V (r) ,
2m
(2.188)
where V (r) is chosen from an ensemble of random functions. Physically, V (r) is bounded
and smooth, although often theorists often study uncorrelated white noise potentials where
the ensemble is described by the distribution functional
Z
n
o
1
P [V (r)] = exp
ddr V 2 (r) ,
(2.189)
2
for which
V (r) V (r 0 ) = (r r 0 ) .
A tight binding version of this model would resemble
X
X
i ci ci ,
H = t
ci cj + cj ci +
hiji
(2.190)
(2.191)
where the single site energies {i } are independently distributed according to some function
p(). The first term can be diagonalized by Fourier transform:
X
X
(2.192)
H0 = t
ci cj + cj ci = zt
k ck ck ,
hiji
P
where k = z 1 eik , where is a nearest neighbor direct lattice vector and z is the
lattice coordination number; the bandwidth is 2zt. What happens when we add the random
potential term to (2.192)? Suppose the width of the distribution p() is W , e.g. p() =
W 1 ( 41 W 2 2 ), with W zt. We expect that the band edges shift from zt to
(zt + 21 W ). The density of states must vanish for || > zt + 12 W ; the regions in the vicinity
of (zt + 21 W ) are known as Lifshitz tails.
Aside from the formation of Lifshitz tails, the density of states doesnt change much. What
does change is the character of the eigenfunctions. Suppose we can find a region of contiguous sites all of which have energies i 12 W . Then we could form an approximate
eigenstate by concentrating the wavefunction in this region of sites, setting its phase to
be constant throughout. This is an example of a localized state. We can think of such a
state as a particle in a box the electron binds itself to local fluctuations in the potential.
Outside this region, the wavefunction decays, typically exponentially. Scattering states are
then extended states, and are associated with average configurations of the {i }. The
typical spatial extent of the localized states is given by the localization length (). The
localization length diverges at the mobility edges as
() | c | .
(2.193)
There is no signature of the mobility edge in the density of states itself N () is completely
smooth through the mobility edge c .
82
CHAPTER 2. MESOSCOPIA
Figure 2.9: Schematic picture of density of states in a disordered system showing mobility
edges (dashed lines) and localized states in band tails.
2.6.1
One way of characterizing the localization properties of quantum mechanical eigenstates (of
H) is to compute the participation ratio Q. The inverse participation ratio for the state
is given by
X
X
2 2
4
1
(discrete)
(2.194)
(i)
Q =
(i)
i
Z
=
4
d x (x)
d
Z
2
d x (x)
d
2
(continuous)
(2.195)
k
P
Consider the discrete case. If is localized on a single site, then we have i (i) = 1
2
1 ,
for all k, i.e. Q = 1. But if is spread evenly over N sites, then Q1
= N/N = N
and Q = N . Hence, Q tells us approximately how many states i participate in the
state . The dependence of Q on the system size (linear dimension) L can be used as
a diagnostic:
localized = Q L0
(2.196)
extended = Q L ,
(2.197)
where > 0.
Another way to distinguish extended from localized states is to examine their sensitivity to
boundary conditions:
(x1 , . . . , x + L , . . . , xd ) = ei (x1 , . . . , x , . . . , xd ) ,
(2.198)
83
where {1, . . . , d}. For periodic boundary conditions = 0, while antiperiodic boundary
conditions have = . For plane wave states, this changes the allowed wavevectors, so
that k k + /L. Thus, for an extended state,
ext = pbc apbc '
k L1 .
k
(2.199)
(2.200)
(2.201)
(2.202)
2.6.2
(2.204)
Pioneering work in numerical studies of the localization transition was performed by MacKinnon and Kramer in the early 1980s. They computed the localization length M (W/t, E)
for systems of dimension M d1 N , from the formula
d1
1
M
M
2 E
1 D X
, E = lim
ln
G1i,N j (E) ,
N 2N
t
W
(2.205)
i,j=1
where G(E) = (E + i H)1 is the Greens function, and i, j label transverse sites.
The average h i is over disorder configurations. It is computationally very convenient
to compute the localization length in this manner, rather than from exact diagonalization,
because the Greens function can be computed recursively. For details, see A. MacKinnon
and B. Kramer, Phys. Rev. Lett. 47, 1546 (1981). Note also that it is h|G|2 i which is
computed, rather than hGi. The reason for this is that the Greens function itself carries
a complex phase which when averaged over disorder configurations results in a decay of
hGR,R0 i on the scale of the elastic mean free path `.
84
CHAPTER 2. MESOSCOPIA
Figure 2.10: Mock-up of typical raw data from numerical study of localization length for
d = 2 (left panel) and d = 3 (right panel) systems. For d = 2, (W/t) is finite and
monotonically decreasing with increasing W/t. For d = 3, there is a critical value of
W/t. For W/t < (W/t)c , M (W/t) diverges as M ; this is the extended phase. For
W/t > (W/t)c , M (W/t) remains finite as M ; this is the localized phase.
MacKinnon and Kramer computed the localization length by employing the Ansatz of finite
size scaling. They assumed that
M
W .
, E = M f
,E M ,
t
t
W
(2.206)
where f (x) is a universal scaling function which depends only on the dimension d. MK
examined the band center, at E = 0; for d > 2 this is the last region to localize as W/t is
increased from zero. A mock-up of typical raw data is shown in fig. 2.10.
In the d = 2 case, all states are localized. Accordingly, M /M 0 as M , and M (W/t)
decreases with increasing W/t. In the d = 3 case, states at the band center are extended
for weak disorder. As W/t increases, M (W/t) decreases, but with (W/t) still divergent.
At the critical point, (W/t)c , this behavior changes. The band center states localize, and
(W/t) is finite for W/t > (W/t)c . If one rescales and plots M /M versus /M , the
scaling function f (x) is revealed. This is shown in fig. 2.11, which is from the paper by
MacKinnon and Kramer. Note that there is only one branch to the scaling function for
d = 2, but two branches for d = 3. MacKinnon and Kramer found (W/t)c ' 16.5 for a
disordered tight binding model on a simple cubic lattice.
85
2.6.3
(2.208)
d ln g
,
d ln L
(2.209)
which describes the change of g when we vary the size of the system. We now know the
limiting values of (g) for small and large g:
metallic (g 1)
localized (g 1)
(g) = d 2
2L
(g) =
= ln g + const.
(2.210)
(2.211)
86
CHAPTER 2. MESOSCOPIA
Figure 2.12: Sketch of the -function for the localization problem for d = 1, 2, 3. A critical
point exists at g = gc for the d = 3 case.
If we assume that (g) is a smooth monotonic function of g, we arrive at the picture in fig.
2.12. Note that in d = 1, we can compute (g) exactly, using the Landauer formula,
g=
T
,
1T
(2.212)
(2.213)
It should be stressed that the very existence of a -function is hardly clear. If it does exist,
it says that the conductance of a system of size L is uniquely determined by its conductance
at some other length scale, typically chosen to be microscopic, e.g. L0 = `. Integrating the
-function, we obtain an integral equation to be solved implicitly for g(L):
ln
L
L0
lnZg(L)
d ln g
.
(g)
(2.214)
ln g(L0 )
A priori it seems more likely, though, that as L is increased the changes to the conductance
may depend on more than g alone. E.g. the differential change dg might depend on the
entire distribution function P (W ) for the disorder.
87
1
(g) ' ln(g/gc )
ln g
(2.215)
ln g =
ln gc .
1
(2.216)
(2.217)
ln(g/gc )
ln(g0 /gc )
ln g0
ln
L
L+
Zln g
=
d ln g = ln(g/g+ ) ,
(2.219)
ln g+
L
(g0 gc )1/ ,
L0
(2.220)
e2 g
e2 A+ gc
=
(g0 gc )1/ .
h L
h
L0
(2.221)
If instead we start with g0 = gc g and integrate out to large negative values of ln g, then
!
ln
Z g
d ln g
ln(gc /g )
L
1
ln
=
= ln
(2.222)
L0
ln(g/gc )
ln(gc /g0 )
ln g0
ln
L
L
Zln g
=
ln g
d ln g
= ln
ln g
ln g
ln g
,
(2.223)
88
CHAPTER 2. MESOSCOPIA
which says
g(L) = e2L/
2L0
(gc g0 )1/ ,
=
A
with
A =
ln(1/g )
gc ln(gc /g )
1/ .
(2.224)
(2.225)
(2.226)
On the metallic side of the transition, when g0 > gc , we can identify a localization length
through
g gc / ,
(2.227)
which says
=
L0
(g gc )1/ .
A+ 0
(2.228)
Finally, since g0 is determined by the value of the Fermi energy F . we can define the critical
energy, or mobility edge c , through
in which case
g(L0 , c ) = gc ,
(2.229)
g(L0 , )
g g(L0 , F ) gc =
F c .
=c
(2.230)
Thus, g (F c ).
Integrating the -function: d = 2
In two dimensions, there is no fixed point. In the Ohmic limit g 1, we have
c
(g) = + O(g 2 ) ,
g
(2.231)
L
L0
Zln g
d ln g
g g0
=
=
+ ...
(g)
c
(2.232)
ln g0
and
g(L) = kF ` c ln(L/`) ,
(2.233)
where we have used the Drude result g = kF `, valid for L0 = `. We now see that the
localization length is the value of L for which the correction term is on the same order
as g0 : = ` exp(kF `/c). A first principles treatment yields c = 2 . The metallic regime in
d = 2 is often called the weak localization regime.
89
2 + dimensions
At or below d = 2 dimensions, there is no mobility edge and all eigenstates are localized.
d = 2 is the lower critical dimension for the localization transition. Consider now the
problem in d = 2 + dimensions. One has
(g) =
c
+ O(g 2 ) .
g
(2.234)
The critical conductance lies at gc = c/. For 0+ , this is large enough that higher order
terms in the expansion of the -function can safely be ignored in the metallic limit. An
analysis similar to that for d = 3 now yields
h
L
e2
g > gc
g(L) =
g < gc
g(L) = e2L/ ,
(2.235)
(2.236)
with
=
e2
L (g0 gc )
h 0
(2.237)
2L0
(gc g0 ) .
A
(2.238)
(2.239)
with = 1 + O(). Close to the transition on the metallic side, the conductivity vanishes
as
() | c |s .
(2.240)
The relation s = (d2), which follows from the above treatment, may be used to relate the
localization length and conductivity critical exponents. (In d = 3, MacKinnon and Kramer
obtained = s ' 1.2.)
2.6.4
Finite Temperature
h
3
` L
(
)
e2
2
L
d=2 (L) =
kF ` ln
h
`
o
e2 n
d=1 (L) =
4` 2(L `) .
h
(2.241)
(2.242)
(2.243)
90
CHAPTER 2. MESOSCOPIA
Clearly the d = 1 result must break down for even microscopic L >
3`. The above results
are computed using the -function
(g) = (d 2)
cd
+ O(g 2 ) ,
g
(2.244)
where the coefficients cd are computed from perturbation theory in the disorder.
q
At finite temperature, the cutoff becomes min(L, L ), where L = D is the inelastic
scattering length and D = vF ` is the diffusion constant. Suppose that (T ) T p as
T 0, so that L = a (T /T0 )p/2 , where T0 is some characteristic temperature (e.g. the
Debye temperature, if the inelastic mechanism is electron-phonon scattering). Then, for
L > L,
(
)
2 e2 1 1 T p/2
d=3 (T ) = d=3 2
h ` a T0
(
)
a p T
2 e2
B
ln
ln
d=2 (T ) = d=2
h
`
2
T0
)
(
e2
T0 p/2
B
d=1 (T ) = d=1 2
` ,
a
h
T
B
(2.245)
(2.246)
(2.247)
where dB is the Boltzmann conductivity. Note that (T ) decreases with decreasing temperature, unlike the classic low T result for metals, where (T ) = 0 + AT 2 . I.e. usually
(T ) increases as T increases due to a concomitant decrease in transport scattering time .
Weak localization physics, though, has the opposite effect, as the enhanced backscattering
is suppressed as T increases and L decreases. The result is that (T ) starts to decrease as
T is lowered from high temperatures, but turns around at low T and starts increasing again.
This behavior was first observed in 1979 by Dolan and Osheroff, who studied thin metallic
PdAu films, observing a logarithmic increase in d=2 (T ) at the lowest temperatures.
Chapter 3
(3.1)
where is the damping rate ( > 0) and 0 is the natural frequency in the absence of
damping1 . We adopt the following convention for the Fourier transform of a function H(t):
Z
d
H() eit
2
(3.2)
dt H(t) e+it .
(3.3)
H(t) =
H()
=
(02 2i 2 ) x()
= f() ,
(3.4)
with the solution
x()
=
02
f()
()
f()
2i 2
(3.5)
where ()
=
with
02
1
1
=
,
2
2i
( + )( )
(3.6)
q
= i 02 2 .
(3.7)
91
92
f() eit
d
+ xh (t)
2 02 2i 2
(3.8)
(3.9)
Since Im( ) < 0, xh (t) is a transient which decays in time. The coefficients A may
be chosen to satisfy initial conditions on x(0) and x(0),
(s)
(3.10)
d
()
eis ,
2
(3.11)
is analytic
Claim: The response is causal , i.e. (t t0 ) = 0 when t < t0 , provided that ()
in the upper half plane of the variable .
Proof: Consider eqn. (3.11). Of ()
e
2
LHP
=
i.e.
(s) =
iei+ s
+
iei s
+
p
s
e
2 2 (s)
sin
2 2
0
e 2
02
p
sinh
2 02 (s)
(3.12)
if 02 > 2
(3.13)
if 02 < 2 ,
where (s) is the step function: (s 0) = 1, (s < 0) = 0. Causality simply means that
events occuring after the time t cannot influence the state of the system at t. Note that, in
general, (t) describes the time-dependent response to a -function impulse at t = 0.
3.1.1
93
Energy Dissipation
How much work is done by the force f (t)? Since the power applied is P (t) = f (t) x(t),
we
have
Z
P (t) =
d
(i) ()
f() eit
2
d
f () e+it
2
(3.14)
E =
dt P (t) =
d
f()2 .
(i) ()
(3.15)
Separating ()
=
0 () + i
00 () ,
(3.16)
00 () = 2
(0 2 )2 + 4 2 2
0 () =
(3.17)
(3.18)
2
d
00 () f() .
2
(3.19)
() is the dissipative part. When experimentalists measure a lineshape, they usually are
referring to features in
00 (), which describes the absorption rate as a function of driving
frequency.
3.2
Kramers-Kronig Relations
Let (z) be a complex function of the complex variable z which is analytic in the upper
half plane. Then the following integral must vanish,
I
dz (z)
=0,
(3.20)
2i z
C
(3.21)
94
d
()
2i + i
P
d 0
00
() + i ()
i( )
2i
(3.22)
where P stands for principal part. Taking the real and imaginary parts of this equation
reveals the Kramers-Kronig relations:
0 () = P
d 00 ()
(3.23)
00
() = P
d 0 ()
.
(3.24)
The Kramers-Kronig relations are valid for any function (z) which is analytic in the upper
half plane.
If (z) is analytic everywhere off the Im(z) = 0 axis, we may write
Z
(z) =
d 00 ()
.
z
(3.25)
(3.26)
95
.
+ 2
(3.27)
i/(z + i)
d 1
(z) =
2
=
z + 2
i/(z i)
if Im(z) > 0
(3.28)
if Im(z) < 0 .
Note that (z) is separately analytic in the UHP and the LHP, but that there is a branch
cut along the Re(z) axis, where ( i) = i/( i).
EXERCISE: Show that eqn. (3.26) is satisfied for () = /( 2 + 2 ).
If we analytically continue (z) from the UHP into the LHP, we find a pole and no branch
cut:
i
.
(3.29)
(z)
=
z + i
The pole lies in the LHP at z = i.
3.3
Here, the {Qi } are a set of Hermitian operators, and the {i (t)} are fields or potentials.
Some examples:
M B(t)
magnetic moment magnetic field
R
H1 (t) =
d3r %(r) (r, t)
density scalar potential
1R 3
c d r j(r) A(r, t) electromagnetic current vector potential
We now ask, what is hQi (t)i? We assume that the lowest order response is linear, i.e.
Z
hQi (t)i = dt0 ij (t t0 ) j (t0 ) + O(k l ) .
(3.31)
Note that we assume that the O(0 ) term vanishes, which can be assured with a judicious
choice of the {Qi }2 . We also assume that the responses are all causal, i.e. ij (t t0 ) = 0 for
2
96
t < t0 . To compute ij (t t0 ), we will use first order perturbation theory to obtain hQi (t)i
and then functionally differentiate with respect to j (t0 ):
Q
(t)
i
ij (t t0 ) =
.
(3.32)
j (t0 )
The first step is to establish the result,
(t) = T exp
Zt
dt0 H0 + H1 (t0 ) (t0 ) ,
(3.33)
t0
where T is the time ordering operator, which places earlier times to the right. This is easily
derived starting with the Schr
odinger equation,
i~
d
(t) = H(t) (t) ,
dt
(3.34)
(3.35)
(3.36)
hence
(t2 ) = U (t2 , t1 ) (t1 )
U (t2 , t1 ) = T exp
(3.37)
Zt2
dt H(t) .
(3.38)
t1
(3.39)
(3.40)
97
where
Qj (t) eiH0 t/~ Qj eiH0 t/~
(3.41)
(t) Qi (t)
0
j (t )
(t)
(t)
=
Qi (t) + (t) Qi
j (t0 )
j (t0 )
n i
=
(t0 ) eiH0 t0 /~ Qj (t0 ) e+iH0 t/~ Qi (t)
~
o
i
+
(t) Qi eiH0 t/~ Qj (t0 ) e+iH0 t0 /~ (t0 )
(t t0 )
~
i
Qi (t), Qj (t0 ) (t t0 ) ,
(3.42)
=
~
were averages are with respect to the wavefunction exp(iH0 t0 /~) (t0 ) , with
t0 , or, at finite temperature, with respect to a Boltzmann-weighted distribution of
such states. To reiterate,
ij (t t0 ) =
ij (t t0 ) =
i
Qi (t), Qj (t0 ) (t t0 )
~
(3.43)
3.3.1
Spectral Representation
We now derive an expression for the response functions in terms of the spectral properties
of the
H0 . We stress that H0 may describe a fully interacting system. Write
Hamiltonian
H0 n = ~n n , in which case
i
ij () =
~
dt eit Qi (t), Qj (0)
0
Z
i
1 X ~m n
dt eit
e
m Qi n n Qj m e+i(m n )t
=
~
Z m,n
0
o
m Qj n n Qi m e+i(n m )t ,
(3.44)
where = 1/kB T and Z is the partition function. Regularizing the integrals at t with
exp(t) with = 0+ , we use
Z
dt ei(+i)t =
0
i
+ i
(3.45)
98
ij ( + i) =
e
~Z m,n
)
m Qj n n Qi m
m Qi n n Qj m
m + n + i
+ m n + i
(3.46)
We will refer to this as
ij (); formally
ij () has poles or a branch cut (for continuous
spectra) along the Re() axis. Diagrammatic perturbation theory does not give us
ij (),
but rather the time-ordered response function,
i
T Qi (t) Qj (t0 )
~
i
=
Qi (t) Qj (t0 ) (t t0 ) +
Qj (t0 ) Qi (t) (t0 t) .
~
~
Tij (t t0 )
(3.47)
Tij ( + i) =
e
~Z m,n
)
m Qj n n Qi m
m Qi n n Qj m
m + n i
+ m n + i
(3.48)
The difference between
ij () and
Tij () is thus only in the sign of the infinitesimal i
term in one of the denominators.
Let us now define the real and imaginary parts of the product of expectations values encountered above:
m Qi n n Qj m Amn (ij) + iBmn (ij) .
(3.49)
That is4 ,
1
1
m Qi n n Qj m +
m Qj n n Qi m
2
2
1
1
m Qi n n Qj m
m Qj n n Qi m .
Bmn (ij) =
2i
2i
Amn (ij) =
(3.50)
(3.51)
Note that Amn (ij) is separately symmetric under interchange of either m and n, or of i and
j, whereas Bmn (ij) is separately antisymmetric under these operations:
Amn (ij) = +Anm (ij) = Anm (ji) = +Amn (ji)
(3.52)
(3.53)
e
( n + m ) ,
~Z m,n
Bmn (ij)
%B
ij ()
3
4
(3.54)
99
which satisfy
A
%A
ij () = +%ji ()
%B
ij ()
%B
ji ()
~ A
%A
%ij ()
ij () = +e
%B
ij ()
= e
%B
ij ()
(3.55)
.
(3.56)
0ij ()
00ij () = P d
2
%A () (1 e~ ) %B
0ij ()
ij () = +
2 2 ij
(3.57)
2
%B () + (1 e~ ) %A
00ij ().
ij () =
2 2 ij
(3.58)
0ijT () = P d
00ijT () = P d
2
%A () (1 + e~ ) %B
ij ()
2 ij
(3.59)
2
%B () + (1 + e~ ) %A
ij () .
2 2 ij
(3.60)
Hence, knowledge of either the retarded or the time-ordered response functions is sufficient
to determine the full behavior of the other:
0T
0ij () +
0ji () =
ij () +
0jiT ()
0
0T
ij ()
0ji () =
ij ()
0jiT () tanh( 12 ~)
00
00 T
ij () +
00ji () =
ij () +
00jiT () tanh( 21 ~)
00
00 T
ij ()
00ji () =
ij ()
00jiT () .
3.3.2
(3.61)
(3.62)
(3.63)
(3.64)
Energy Dissipation
The work done on the system must be positive! he rate at which work is done by the
external fields is the power dissipated,
d
1
= (t)
Qi (t) i (t) ,
(t) =
t
P =
(3.65)
100
where we have invoked the Feynman-Hellman theorem. The total energy dissipated is thus
a functional of the external fields {i (t)}:
Z
Z Z
W = dt P (t) = dt dt0 ij (t t0 ) i (t) j (t0 )
Z
=
d
(i) i ()
ij () j () .
2
(3.66)
Since the {Qi } are Hermitian observables, the {i (t)} must be real fields, in which case
i () = j (), whence
Z
d
(i)
ij ()
ji () i () j ()
W =
4
d
Mij () i () j ()
2
(3.67)
where
Mij () 12 (i)
ij ()
ji ()
B
()
+
i%
()
.
= 1 e~ %A
ij
ij
(3.68)
3.3.3
Correlation Functions
Sij (t) Qi (t) Qj (t0 ) ,
which has the spectral representation
h
i
B
Sij () = 2~ %A
ij () + i%ij ()
2 X ~m
=
e
m Qi n n Qj n ( n + m ) .
Z m,n
(3.69)
(3.70)
Note that
Sij () = e~ Sij
()
Sji () = Sij
() .
(3.71)
and that
ij ()
ji () =
i
1 e~ Sij ()
~
(3.72)
3.4. EXAMPLE: S =
1
2
101
(3.73)
3.3.4
(3.74)
Continuous Systems
The indices i and j could contain spatial information as well. Typically we will separate
out spatial degrees of freedom, and write
Sij (r r 0 , t t0 ) = Qi (r, t) Qj (r 0 , t0 ) ,
(3.75)
where we have assumed space and time translation invariance. The Fourier transform is
defined as
) =
S(k,
Z
d r dt eikr S(r, t)
3
(3.76)
=
t) Q(k,
0) .
dt e+it Q(k,
V
(3.77)
3.4
Example: S =
Consider a S =
1
2
1
2
(3.78)
S n n S
S n n S
2 X
() =
~ n
+ n + i
+ n + i
S +
+ S
S +
+ S
2
=
,
~
+ B0 + i
B0 + i
(3.79)
(3.80)
102
(3.81)
(3.82)
or, equivalently,
1
1
xx () =
= +yy ()
+ B0 + i B0 + i
1
1
2
i
= yx () .
+
xy () = 4 ~
+ B0 + i B0 + i
2
1
4 ~
3.4.1
(3.83)
(3.84)
Bloch Equations
(3.85)
= 0. In equilibrium, we have M = M0 z,
(3.86)
My
M y = Mz Hx Mx Hz
T2
(3.87)
M z M0
M z = Mx Hy My Hx
.
T1
(3.88)
These are known as the Bloch equations. Mathematically, they are a set of coupled linear,
first order, time-dependent, inhomogeneous equations. These may be recast in the form
M + R M = ,
1
1
with R (t) = T
H (t), = T
M0 , and
T2 0 0
T = 0 T2 0 .
0 0 T1
(3.89)
(3.90)
3.4. EXAMPLE: S =
1
2
103
(3.91)
Zt
0
0
U (t) = T exp dt R(t ) ,
(3.92)
(3.96)
(3.97)
(3.98)
(3.99)
2 H0 M0
= yy ()
2 H02 + T22 2 2iT21
(3.100)
xy () =
(i T21 ) M0
= yx () .
2 H02 + T22 2 2iT21
(3.101)
(3.102)
( 2 H02 + T22 2 ) 2 H0 M0
2
2 H02 + T22 2 + 4 T22 2
(3.103)
00xx () =
2 H0 M0 T21
,
2
2 H02 + T22 2 + 4 T22 2
(3.104)
104
3.5
Electromagnetic Response
Consider an interacting system consisting of electrons of charge e in the presence of a timevarying electromagnetic field. The electromagnetic field is given in terms of the 4-potential
A = (A0 , A):
E = A0
1 A
c t
B =A .
(3.105)
(3.106)
N
X
(x xi )
(3.108)
i=1
N
o
e Xn
pi (x xi ) + (x xi ) pi
j (x)
2m
(3.109)
(3.110)
i=1
1 0 0 0
0 1 0 0
(3.111)
g = g =
0 0 1 0 ,
0 0 0 1
so that
j = (j 0 , j 1 , j 2 , j 3 ) (j 0 , j)
(3.112)
j = g j = (j 0 , j 1 , j 2 , j 3 )
(3.113)
j A = j g A = j 0 A0 + j A j A
(3.114)
The quantity jp (x) is known as the paramagnetic current density. The physical current
density j (x) also contains a diamagnetic contribution:
H
= jp (x) + jd (x)
A (x)
e2
e
j d (x) =
n(x)A(x) = 2 j0p (x) A(x)
mc
mc
j0d (x) = 0 .
j (x) = c
(3.115)
(3.116)
(3.117)
c
3 0
j (x, t) =
d x dt K (xt; x0 t0 ) A (x0 , t0 ) ,
4
valid to first order in the external 4-potential A . From
Z
Z
p
i
3 0
d x dt0 jp (x, t), jp (x0 , t0 ) (t t0 ) A (x0 , t0 )
j (x, t) =
~c
d
e
105
(3.118)
(3.119)
(3.120)
we conclude
K (xt; x0 t0 ) =
E
4 D p
p 0 0
j
(x,
t),
j
(x
,
t
)
(t t0 )
i~c2
+
(3.121)
4e
p
j0 (x, t) (x x0 ) (t t0 ) (1 0 ) .
2
mc
p
The first
term is sometimes
known as the paramagnetic response kernel, K
(x; x0 ) =
p
p
(4i/i~c2 ) j (x), j (x0 ) (tt0 ) is not directly calculable by perturbation
theory. Rather,
p,T
(x; x0 ) = (4/i~c2 ) T jp (x) jp (x0 ) ,
one obtains the time-ordered response function K
where x (ct, x).
Z
Z
1X
(3.122)
+
d3x d3x0 (x) 0 (x0 ) v(x x0 ) 0 (x0 ) (x) ,
2 0
,
where v(x x0 ) is a two-body interaction, e.g. e2 /|x x0 |, and U (x) is the external scalar
potential. Expanding in powers of A ,
Z
Z
1
e2 X
3
p
H(A ) = H(0)
d3x (x) (x) A2 (x) ,
(3.123)
d x j (x) A (x) +
c
2mc2
where the paramagnetic current density jp (x) is defined by
X
j0p (x) = c e
(x) (x)
(3.124)
ie~ X
j (x) =
(x) (x) (x) (x) .
2m
p
(3.125)
106
3.5.1
c
j (q) = K (q) A (q) .
4
Note our convention on Fourier transforms:
Z 4
dk
H(k) e+ikx
H(x) =
(2)4
Z
(3.126)
(3.127)
(3.128)
(3.129)
where is an arbitrary scalar function. Since the physical current must be unchanged
by a gauge transformation, we conclude that K (q) q = 0. We also have the continuity
equation, j = 0, the Fourier space version of which says q j (q) = 0, which in turn
requires q K (q) = 0. Therefore,
X
q K (q) =
K (q) q = 0
(3.130)
In fact, the above conditions are identical owing to the reciprocity relations,
Re K (q) = +Re K (q)
(3.131)
Im K (q) = Im K (q) ,
(3.132)
3.5.2
(3.133)
A Sum Rule
c
ji (q) = Kij (q) Aj (q)
4
c
c j
= Kij (q)
E (q)
4
i
ij (q) E j (q) .
(3.134)
ic2
Kij (q, ) .
4
(3.135)
107
If, in the 0 limit, the conductivity is to remain finite, then we must have
Z
Z D
E
ie2 n
ij ,
d x dt jip (x, t), jjp (0, 0) e+it =
m
3
(3.136)
where n is the electron number density. This relation is spontaneously violated in a superconductor, where () 1 as 0.
3.5.3
In other words, the three components of K0j (q) are in fact completely determined by Kk (q)
and q itself. When = 0,
0 = q 0 K00 + q i Ki0 =
c
K00 q 2 Kk ,
c
(3.140)
which says
K00 (q, ) =
c2 2
q Kk (q, )
2
(3.141)
Thus, of the 10 freedoms of the symmetric 4 4 tensor K (q), there are only two independent ones the functions Kk (q) and K (q).
3.5.4
Neutral Systems
In neutral systems, we define the number density and number current density as
n(x) =
N
X
(x xi )
(3.142)
i=1
N
j(x) =
o
1 Xn
pi (x xi ) + (x xi ) pi .
2m
i=1
(3.143)
108
(3.144)
(3.145)
We define the longitudinal and transverse susceptibilities for homogeneous systems according to
ij (q, ) = k (q, ) qi qj + (q, ) (ij qi qj ) .
(3.146)
From the continuity equation,
j+
follows the relation
k (q, ) =
n
=0
t
n
2
+ 2 (q, ) .
m q
(3.147)
(3.148)
(3.149)
(3.150)
and therefore
o
4e2 n n
(q,
)
k
c2
m
n
o
2
4e n
(q, ) .
K (q, ) = 2
c
m
Kk (q, ) =
3.5.5
(3.151)
(3.152)
c
j(q, ) = K (q, ) A(q, ) .
(3.153)
4
This leads directly to the Meissner effect whenever limq0 K (q, 0) is finite. To see this,
we write
B = ( A) 2 A
4
1 E
=
j+
c
c t
4
c
1 2A
=
K (i, i t ) A 2 2 ,
c
4
c t
(3.154)
109
which yields
1 2
A = K (i, i t )A .
c2 t2
(3.155)
(3.156)
mc2
4e2 2L
(3.157)
= n m lim (q, 0) .
q0
(3.158)
ij (q, t) =
(3.159)
(3.160)
)t
(3.161)
(3.162)
Using
[AB, CD] = A [B, C] D + AC [B, D] + C [A, D] B + [A, C] DB ,
(3.163)
we obtain
~2 X
i(
)t
ji (q, t), jj (q, 0) =
(2ki +qi )(2kj +qj ) e k k+q n0 (k )n0 (k+q ) , (3.164)
2
4m
k
1
e~ e
(3.165)
110
Thus,
n0 (k+q ) n0 (k )
~ X
ij (q, ) =
(2ki + qi )(2kj + qj )
4m2 V
+ k k+q + i
k
1
1
~n0
qi qj
=
4m2 + q + i q + i
Z 3
0
0
~
d k n (k+q/2 ) n (kq/2 )
+ 2
ki kj ,
m
(2)3 +
+
i
kq/2
k+q/2
(3.166)
(3.167)
ij (q 0, 0) =
n0
n0
qi qj + ij ,
m
m
(3.168)
n
m
(q 0, 0) =
n0
,
m
(3.169)
where n = n0 + n0 is the total boson number density. The superfluid density, according to
(3.158), is ns = n0 (T ).
In fact, the ideal Bose gas is not a superfluid. Its excitation spectrum is too soft - any
superflow is unstable toward decay into single particle excitations.
3.6
Density-Density Correlations
PN
i=1 (r ri ),
and
(3.170)
3 0
d r dt0 (r r 0 , t t0 ) U (r 0 , t0 )
(q, ) ,
h n(q,
)i = (q, ) U
(3.171)
(3.172)
111
where
(q, ) =
1 X ~m
e
~VZ m,n
1
=
~
S(q, ) =
2
(
q n
m n
m + n + i
Z
d S(q, )
2 )
q n
mn
+ m n + i
1
1
+ + i + i
(3.173)
2
2 X ~m
e
m
n
q n ( n + m ) .
VZ m,n
(3.174)
Note that
q =
n
N
X
eiqri ,
(3.175)
i=1
q = n
q . S(q, ) is known as the dynamic structure factor . In a scattering exand that n
periment, where an incident probe (e.g. a neutron) interacts with the system via a potential
U (r R), where R is the probe particle position, Fermis Golden Rule says that the rate at
which the incident particle deposits momentum ~q and energy ~ into the system is given
by
I(q, ) =
=
2
2 X ~m
e
m; p H1 n; p ~q ( n + m )
~Z m,n
1 2
U (q) S(q, ) .
~
(3.176)
(q)2 is called the form factor. In neutron scattering, the on-shell conThe quantity U
dition requires that the incident energy and momentum p are related via the ballistic
dispersion = p2 /2mn . Similarly, the final energy and momentum are related, hence
~ =
p2
2mn
~ =
(p ~q)2
2mn
~ =
~q p
mn
~2 q 2
2mn
(3.177)
Hence, for fixed momentum transfer ~q, can be varied by changing the incident momentum
p.
Another case of interest is the response of a system
to a foreign object moving with trajectory
R(t) = V t. In this case, U (r, t) = U r R(t) , and
(q, ) =
U
Z
d r dt eiqr eit U (r V t)
3
(q)
= 2 ( q V ) U
(3.178)
hn(q, )i = 2 ( q V ) (q, ) .
(3.179)
so that
112
3.6.1
Sum Rules
1 X ~m
q n n [H, n
q ] m
e
m n
~VZ m,n
1
1
q [H, n
q ] =
q , [H, n
q ] ,
n
n
(3.180)
~V
2~V
where the last equality is guaranteed by q q symmetry. Now if the potential is velocity
independent, i.e. if
N
~2 X 2
H=
(3.181)
i + V (r1 , . . . , rN ) ,
2m
i=1
P
iqri we obtain
q = N
then with n
i=1 e
=
q ] =
[H, n
=
q ] =
q , [H, n
n
~2 X 2 iqri
i , e
2m
i=1
N
X
~2
q
2im
~2
2im
i eiqri + eiqri i
i=1
N X
N h
X
i=1 j=1
2 2
= N ~ q /m .
We have derived the f -sum rule:
Z
(3.182)
d
N ~q 2
S(q, ) =
.
2
2mV
(3.183)
(3.184)
Note that this integral, which is the first moment of the structure factor, is independent of
the potential !
n times
z
}|
{
Z
D
h
iE
d n
1
q H, H, [H, n
q ]
S(q, ) =
n
.
(3.185)
2
~V
Moments with n > 1 in general do depend on the potential. The n = 0 moment gives
Z
d n
1
q n
q
S(q)
S(q, ) =
n
2
~V
Z
1
=
d3r hn(r) n(0)i eiqr ,
(3.186)
~
which is the Fourier transform of the density-density correlation function.
113
1 V
1 n
= 2
.
V n T
n T
(3.187)
hni = (0, 0)
1
T =
lim
~n2 q0
d S(q, )
.
(3.188)
3.7
The dynamic structure factor S(q, ) tells us about the spectrum of density fluctuations.
P
q = i eiqri increases the wavevector by q. At T = 0, in order for
The density operator n
q G to be nonzero (where G is the ground state, i.e. the filled Fermi sphere), the
nn
state n must correspond to a particle-hole excitation. For a given q, the maximum excitation
frequency is obtained by taking an electron just inside the Fermi sphere, with wavevector
k = kF q and transferring it to a state outside the Fermi sphere with wavevector k + q. For
|q| < 2kF , the minimum excitation frequency is zero one can always form particle-hole
excitations with states adjacent to the Fermi sphere. For |q| > 2kF , the minimum excitation
frequency is obtained by taking an electron just inside the Fermi sphere with wavevector
k = kF q to an unfilled state outside the Fermi sphere with wavevector k + q. These cases
are depicted graphically in fig. 3.2.
We therefore have
max (q) =
min (q) =
~q 2 ~kF q
+
2m
m
~q2
2m
~kF q
m
(3.189)
if q 2kF
(3.190)
if q > 2kF .
This is depicted in fig. 3.3. Outside of the region bounded by min (q) and max (q), there
are no single pair excitations. It is of course easy to create multiple pair excitations with
arbitrary energy and momentum, as depicted in fig. 3.4. However, these multipair states
do not couple to the ground state G through a single application of the density operator
q , hence they have zero oscillator strength: n n
q G = 0 for any multipair state n .
n
114
Figure 3.2: Minimum and maximum frequency particle-hole excitations in the free electron
gas at T = 0. (a) To construct a maximum frequency excitation for a given q, create a hole
just inside the Fermi sphere at k = kF q and an electron at k0 = k + q. (b) For |q| < 2kF the
minumum excitation frequency is zero. (c) For |q| > 2kF , the minimum excitation frequency
is obtained by placing a hole at k = kF q and an electron at k0 = k + q.
3.7.1
Explicit T = 0 Calculation
We start with
S(r, t) = hn(r, t) n(0, 0)
Z 3 Z 3 0
dk
d k ikr X
ikri (t) ik0 rj
e
.
e
e
=
(2)3 (2)3
(3.191)
(3.192)
i,j
The time evolution of the operator ri (t) is given by ri (t) = ri + pi t/m, where pi = i~i .
Using the result
1
eA+B = eA eB e 2 [A,B] ,
(3.193)
which is valid when [A, [A, B]] = [B, [A, B]] = 0, we have
eikri (t) = ei~k
2 t/2m
(3.194)
hence
Z
S(r, t) =
Z 3 0
d3k
d k i~k2 t/2m ikr X
ikri ikpi t/m ik0 rj
e
e
e
e
e
.
(2)3 (2)3
i,j
(3.195)
115
Figure 3.3: Minimum and maximum excitation frequency in units of F /~ versus wavevector q in units of kF . Outside the hatched areas, there are no single pair excitations.
We now break the sum up into diagonal (i = j) and off-diagonal (i 6= j) terms.
For the diagonal terms, with i = j, we have
XX
1
(kF q) (kF q 0 )
N (N 1) q
q0
Z 3 Z 3
d ri d rj i~kqt/m ikri ik0 rj
e
e
e
V
V
i(qq 0 k)ri i(q 0 q+k0 )rj
e
e
XX
(2)6
(kF q) (kF q 0 )
N (N 1)V 2
q
q0
n
o
ei~kqt/m (k) (k0 ) (k k0 ) (k + q 0 q) . (3.197)
116
d3q
(kF q) ei~kqt/m ,
(2)3
(3.198)
Soffdiag
d3k ikr
e
(2)3
and hence
d3q
~k2 ~k q
S(k, ) = n (2) (k) () +
(k
q)
2
F
(2)3
2m
m
2
~k
~k q
(kF |k q|) 2 +
(3.200)
2m
m
Z 3
dq
~k2 ~k q
4 2
= (2) n (k)() +
(k
q)
(|k
+
q|
k
)
.
F
F
(2)3
2m
m
2
117
ZkF
Z1
p
~k 2 ~kq
2
dq q dx k 2 + q 2 + 2kqx kF
x
2m
m
1
m
=
2~k
kF
Z
0
m
=
4~k
r
Z1
2m
k
m
2
dq q
q +
dx x +
kF
~
2q ~kq
1
2
kF
Z
0
k m 2
2m
2
du u +
.
kF u
~
2
~k
(3.201)
2m k m 2
kF u max kF
,
.
~
2
~k
2
(3.202)
2m k m 2
~
2
~k
~kF k ~k 2
,
m
2m
(3.203)
(3.204)
2m k m 2
~
2
~k
~kF k ~k 2
.
m
2m
(3.205)
(3.206)
~k
~k
|k 2kF |
|k + 2kF | .
2m
2m
(3.207)
(3.208)
118
Figure 3.5: The dynamic structure factor S(k, ) for the electron gas at various values of
k/kF .
Putting it all together,
mkF
2 ~2 2v k
kF
F
S(k, ) = mk
1
2
2
4k
~
vF k
if 0 < vF k
k
2kF
2
if vF k
~k2
2m
if vF k +
~k2
2m
vF k +
~k2
2m
~k2
2m
(3.209)
1
S(k) =
n n =
V k k
d
S(k, ) .
2
(3.210)
The result is
S(k) =
3k
4kF
2
Vn
k3
3
16kF
if 0 < k 2kF
if k 2kF
if k = 0 ,
(3.211)
3.8
3.8.1
119
= e
n(k)
ext (k, t) .
(2)3 k2
(3.212)
n(q,
) =
4e
(q, ) ext (q, ) .
q2
(3.213)
(3.214)
E = 4 ext e hni .
(3.215)
In Fourier space,
iq D(q, ) = 4 ext (q, )
(3.216)
(3.217)
1
iq E(q, )
n(q,
)
=
=1
(q, )
iq D(q, )
ext (q, )
Zn
(3.218)
4e2
(q, ) .
q2
(3.219)
4e2
(q, 0) = 1 .
q0 q 2
(3.220)
=1
A system is said to exhibit perfect screening if
(q 0, = 0) =
lim
2
1 X
2n
nn
q 0 ,
2
2
~V n n ( + i)
(3.221)
120
(3.222)
(3.223)
we find
or
n(q,
)
1 1 (q, )
4i
(q, ) =
= (q, ) 1 .
1 (q)
(3.224)
Thus, we arrive at
1
4e2
= 1 2 (q, )
(q, )
q
(q, ) = 1 +
4i
(q, )
(3.225)
Taken together, these two equations allow us to relate the conductivity and the charge
response function,
i
e2 (q, )
(q, ) = 2
.
(3.226)
q 1 4e2 2 (q, )
q
3.8.2
Imagine a time-independent, slowly varying electrical potential (r). We may define the
local chemical potential
e(r) as
e(r) e(r) ,
(3.227)
where is the bulk chemical potential. The local chemical potential is related to the local
density by local thermodynamics. At T = 0,
2/3
~2 2
~2 2
kF (r) =
3 n + 3 2 n(r)
2m
2m
2
~
2 n(r)
2 2/3
=
(3 n)
1+
+ ... ,
2m
3 n
e(r)
(3.228)
3en
(r) .
2
(3.229)
This makes sense a positive potential induces an increase in the local electron number
density. In Fourier space,
h n(q,
= 0)i =
3en
(q, = 0) .
2
(3.230)
121
6ne2
(3.231)
0) ,
(q,
(3.232)
6ne2
,
(3.233)
we have
0) = 4 ext (q, 0) ,
(q,
2
q 2 + qTF
(3.234)
hence
e h n(q,
0)i =
2
qTF
ext (q, 0)
2
q 2 + qTF
(q, 0) = 1 +
2
qTF
q2
(3.235)
TF =
1/6
rs aB ' 0.800 rs aB ,
12
(3.236)
where rs is the dimensionless free electron sphere radius, in units of the Bohr radius aB =
~2 /me2 = 0.529
A, defined by 34 (rs aB )3 n = 1, hence rs n1/3 . Small rs corresponds to
high density. Since Thomas-Fermi theory is a statistical theory, it can only be valid if there
1/3
are many particles within a sphere of radius TF , i.e. 43 3TF n > 1, or rs <
(/12) ' 0.640.
TF theory is applicable only in the high density limit.
In the presence of a -function external charge density ext (r) = Ze (r), we have ext (q, 0) =
Ze and
h n(q,
0)i =
2
ZqTF
2
q 2 + qTF
hn(r)i =
Z er/TF
4r
(3.237)
Note the decay on the scale of TF . Note also the perfect screening:
0, = 0)i = ext (q 0, = 0) = Ze .
e h n(q
3.8.3
(3.238)
We have
1 (q, ) = 1
4e2
(q, )
q2
(3.239)
122
and, at T = 0,
1
1 X
2
1
q 0
,
(q, ) =
j n
~V
+ j0 + i j0 + i
j
(3.240)
q =
n
P
iqr
ie i
(1st quantized)
(3.241)
k k+q
nd
(2
quantized:
{k , k 0 }
= kk0 ) .
(3.242)
Invoking the f -sum rule, the above integral is n~q 2 /2m, hence
(q, ) =
nq 2
,
m 2
and
1 (q, ) = 1 +
p2
,
2
(3.243)
(3.244)
where
r
p
4ne2
m
(3.245)
3.8.4
The electron charge appears nowhere in the free electron gas response function 0 (q, ).
An interacting electron gas certainly does know about electron charge, since the Coulomb
repulsion between electrons is part of the Hamiltonian. The idea behind the RPA is to
obtain an approximation to the interacting (q, ) from the noninteracting 0 (q, ) by
self-consistently adjusting the charge so that the perturbing charge density is not ext (r),
but rather ext (r, t) e hn(r, t)i. Thus, we write
4e2 RPA
(q, ) ext (q, )
q2
n
o
4e2
e h n(q,
)i =
(3.246)
(3.247)
123
Figure 3.6: Perturbation expansion for RPA susceptibility bubble. Each bare bubble contributes a factor 0 (q, ) and each wavy interaction line v(q). The infinite series can be
summed, yielding eqn. 3.249.
which gives
RPA (q, ) =
0 (q, )
1+
4e2 0
(q, )
q2
(3.248)
0 (q, )
.
1 + v(q) 0 (q, )
(3.249)
(3.250)
(3.251)
4e2 0
(q, ) .
q2
(3.252)
124
5. Explicitly,
"
(
2
( ~q 2 /2m)2
1 kF
qTF
vF q ~q 2 /2m
1
+
ln
Re (q, ) = 1 + 2
+ vF q ~q 2 /2m
q
2 4q
(vF q)2
#)
( ~q 2 /2m)2
vF q + ~q 2 /2m
ln
+ 1
(3.253)
+ vF q + ~q 2 /2m
(vF q)2
q2
qTF
if 0 vF q ~q 2 /2m
2vF q
2
2
2
RPA
Im (q, ) = kF 1 (~q /2m) qTF
if vF q ~q 2 /2m vF q + ~q 2 /2m
4q
q2
(vF q)2
0
if > vF q + ~q 2 /2m
(3.254)
RPA
6. Note that
RPA (q, ) = 1
p2
,
2
(3.255)
2
qTF
,
q2
(3.256)
q2
(q, 0) = 1 + TF
q2
)
q + 2kF
1 kF
q2
,
+
1 2 ln
2 2q
4kF
2 2kF
(3.257)
which is real and which has a singularity at q = 2kF . This means that the long-distance
behavior of hn(r)i must oscillate. For a local charge perturbation, ext (r) = Ze (r),
we have
Z
Z
1
hn(r)i = 2
dq q sin(qr) 1
,
(3.258)
2 r
(q, 0)
0
Z cos(2kF r)
,
r3
(3.259)
3.8.5
125
Plasmons
The RPA response function diverges when v(q) 0 (q, ) = 1. For a given value of q, this
occurs for a specific value (or for a discrete set of values) of , i.e. it defines a dispersion
relation = (q). The poles of RPA and are identified with elementary excitations of the
electron gas known as plasmons.
To find the plasmon dispersion, we first derive a result for 0 (q, ), starting with
i
t), n(q,
0) i
(3.260)
0 (q, t) =
h n(q,
~V
h
i
X
X
i
=
h
k, k+q, ,
k0 ,0 k0 q,0 i ei((k)(k+q))t/~ (t) ,
(3.261)
~V
0 0
k
k ,
where (k) is the noninteracting electron dispersion. For a free electron gas, (k) =
~2 k2 /2m. Next, using
AB, CD = A B, C D A, C BD + CA B, D C A, D B
(3.262)
we obtain
0 (q, t) =
i X
(fk fk+q ) ei((k)(k+q))t/~ (t) ,
~V
(3.263)
and therefore
0
(q, ) = 2
fk+q fk
d3k
(2)3 ~ (k + q) + (k) + i
(3.264)
Here,
fk =
1
e((k))/kB T
+1
(3.265)
(3.267)
(3.268)
Recall that the particle-hole continuum frequencies are bounded by min (q) and max (q),
which are given in eqs. 3.190 and 3.189. Eventually the plasmon penetrates the particle-hole
continuum, at which point it becomes heavily damped since it can decay into particle-hole
excitations.
126
Chapter 4
Magnetism
4.1
References
4.2
Introduction
Magnetism arises from two sources. One is the classical magnetic moment due to a current
density j:
Z
1
m=
d3r r j .
(4.1)
2c
The other is the intrinsic spin S of a quantum-mechanical particle (typically the electron):
m = g S/~
q~
= magneton,
2mc
(4.2)
where g is the g-factor (duh!). For the electron, q = e and = B , where B = e~/2mc
is the Bohr magneton.
The Hamiltonian for a single electron is
2
e~
~
~2
( 2 )2
2
+ V (r) +
H +
V
+
+ . . . , (4.3)
2m
2mc
4m2 c2
8mc2
8m3 c2
where = p + ec A. Where did this come from? From the Dirac equation,
2
mc + V
c
i~
=
= E .
(4.4)
c mc2 + V
t
H=
127
128
CHAPTER 4. MAGNETISM
The wavefunction is a four-component Dirac spinor. Since mc2 is the largest term for our
applications, the upper two components of are essentially the positive energy components.
However, the Dirac Hamiltonian mixes the upper two and lower two components of . One
can unmix them by making a canonical transformation,
H H0 ei H ei ,
(4.5)
g =2 1+
+ O(2 ) ,
(4.6)
2
where = e2 /~c 1/137 is the fine structure constant.1
There are two terms in (4.3) which involve the electrons spin:
e~
H
2mc
~
=
V p + ec A
2
2
4m c
HZ =
(Zeeman term)
(4.7)
Hso
(spin-orbit interaction) .
(4.8)
B =
(4.9)
(4.10)
6
So on the scale of electron volts, laboratory scale fields (H <
10 G) are rather small. (And
2000 times smaller for nucleons!).
1 F
,
V H
M (r, t)
2F
0
0
(r, t | r , t ) =
=
,
H (r 0 , t0 )
H (r, t) H (r 0 , t0 )
(4.11)
(4.12)
where F is replaced by a suitable generating function in the nonequilibrium case. Note that
M has the dimensions of H.
1
Note that with n = e~/2mp c for the nuclear magneton, gp = 2.793 and gn = 1.913. These results
immediately suggest that there is composite structure to the nucleons, i.e. quarks.
4.2.1
129
It is amusing to note that classical statistical mechanics cannot account for orbital magnetism. This is because the partition function is independent of the vector potential, which
may be seen by simply shifting the origin of integration for the momentum p:
Z N N
d r d p H({pi q A(ri ),ri })
H
c
Z(A) = Tr e
=
(4.13)
e
(2~)dN
Z N N
d r d p H({pi ,ri })
e
= Z(A = 0) .
(4.14)
=
(2~)dN
Thus, the free energy must be independent of A and hence independent of H = A, and
M = F/H = 0. This inescapable result is known as the Bohr-von Leeuwen theorem.
Of course, classical statistical mechanics can describe magnetism due to intrinsic spin, e.g.
P
Y Z d
i J P
j g H i
i
hiji i
ZHeisenberg (H) =
e
e
4
i
P
X J P
hiji i j eg H
i i .
ZIsing (H) =
e
(4.15)
(4.16)
{i }
Theories of magnetism generally fall into two broad classes: localized and itinerant. In the
localized picture, we imagine a set of individual local moments mi localized at different
points in space (typically, though not exclusively, on lattice sites). In the itinerant picture,
we focus on delocalized Bloch states which also carry electron spin.
4.3
4.3.1
1
e 2
1
p + A + V (r) + gB H s/~ +
s V p + ec A .
2
2
2m
c
2m c
(4.17)
For a single atom or ion in a crystal, let us initially neglect effects due to its neighbors. In
that case the potential V (r) may be taken to be spherically symmetric, so with l = r p,
the first term in the spin-orbit part of the Hamiltonian becomes
Hso =
1
1 1 V
s V p =
sl ,
2
2
2m c
2m2 c2 r r
(4.18)
(4.19)
130
CHAPTER 4. MAGNETISM
1
1 1 V
p2
+ V (r) + B (l + 2s) H +
ls
2m
~
2m2 c2 r r
e2
B rV 0 (r)
2
+
(H
r)
+
2s H r(H r) .
2
2
8mc
~ 4mc
(4.20)
(4.21)
The last term is usually negligible because rV 0 (r) is on the scale of electron volts, while
mc2 = 511 keV.2 The (H r)2 breaks the rotational symmetry of an isolated ion, so in
principal we cannot describe states by total angular momentum J. However, this effect is of
order H 2 , so if we only desire energies to order H 2 , we neednt perturb the wavefunctions
themselves with this term, i.e. we
simply2 treat
it within first
order perturbation theory,
can
e2
leading to an energy shift 8mc2 (H r) in state n .
4.3.2
V
=
(r) ,
8m2 c2
2m2 c2
(4.22)
which is centered at the nucleus. This leads to an energy shift for s-wave states,
Eswave =
2
2
Ze2 ~2
= Z 2 a3B (0)2 e ,
(0)
2m2 c2
2
aB
(4.23)
1
~2
e2
137
is the fine structure constant and aB = me
where = ~c
2 0.529 A is the Bohr
radius. For large Z atoms and ions, the Darwin term contributes a significant contribution
to the total energy.
4.3.3
The full N -electron atomic Hamiltonian, for nuclear charge Ze, is then
"
#
N
N
N
X
X
X
p2i
Ze2
e2
+
+
(ri ) li si
H=
2m
ri
|ri rj |
i<j
i=1
i=1
(
)
N
2
X
e
B
2
+
(li + 2si ) H +
(H ri )
,
~
8mc2
(4.24)
i=1
where li = ri pi and
Ze2 1
Z
(r) =
= 2
2m2 c2 r3
~
2
e2
~c
2
e2 aB 3
.
2aB r
(4.25)
131
i li
and S =
i si ,
respectively.
The full many-electron atom is too difficult a problem to solve exactly. Generally progress
is made by using the Hartree-Fock method to reduce the many-body problem to an effective
one-body problem. One starts with the interacting Hamiltonian
#
"
N
N
X
X
p2i
Ze2
e2
H=
+
,
(4.26)
2m
ri
|ri rj |
i=1
i<j
and treats Hso as a perturbation, and writes the best possible single Slater determinant
state:
h
i
... (r1 , . . . , rN ) = A 1 (r1 ) N (rN ) ,
(4.27)
1
where A is the antisymmetrizer, and i (r) is a single particle wavefunction. In secondquantized notation, the Hamiltonian is
X
X
0
(4.28)
H=
Tij i
j +
Vijkl
i j
0 k 0 l ,
ij
ijkl
0
where
~2 2 Ze2
d3r i (r)
j (r)
2m
|r|
Z
Z
e2
3
1
0 (r 0 ) l (r) .
= 2 d r d3r0 i (r) j0 (r 0 )
|r r 0 | k
Tij =
0
Vijkl
(4.29)
(4.30)
(4.31)
ij 0
OCC
XZ
j6=i
j (r 0 ) i (r 0 )
d3r0
|r r 0 |
j (r) ,
(4.33)
which is a set of N coupled integro-differential equations. Multiplying by i (r) and integrating, we find
OCC
X
0
0
i = Tii + 2
Vijji
Vijij
0 .
(4.34)
j 0
132
CHAPTER 4. MAGNETISM
It is a good approximation to assume that the Hartree-Fock wavefunctions i (r) are spherically symmetric, i.e.
i (r) = Rnl (r) Ylm (, ) ,
(4.35)
independent of . We can then classify the single particle states by the quantum numbers
n {1, 2, . . .}, l {0, 1, . . . , n 1}, ml {l, . . . , +l}, and ms = 21 . The essential physics
introduced by the Hartree-Fock method is that of screening. Close to the origin, a given
electron senses a potential Ze2 /r due to the unscreened nucleus. Farther away, though,
the nuclear charge is screened by the core electrons, and the potential decays faster than
1/r. (Within the Thomas-Fermi approximation, the potential at long distances decays as
Ce2 a3B /r4 , where C ' 100 is a numerical factor, independent of Z.) Whereas states of
different l and identical n are degenerate for the noninteracting hydrogenic atom, when the
nuclear potential is screened, states of different l are no longer degenerate. Smaller l means
smaller energy, since these states are localized closer to the nucleus, where the potential
is large and negative and relatively unscreened. Hence, for a given n, the smaller l states
fill up first. For a given l and n there are (2s + 1) (2l + 1) = 4l + 2 states, labeled by
the angular momentum and spin polarization quantum numbers ml and ms ; this group of
orbitals is called a shell .
4.3.4
Based on the energetics derived from Hartree-Fock3 , we can start to build up the Periodic
Table. (Here I follow the pellucid discussion in G. Bayms Lectures on Quantum Mechanics,
chapter 20.) Start with the lowest energy states, the 1s orbitals. Due to their lower angular
momentum and concomitantly lower energy, the 2s states get filled before the 2p states.
Filling the 1s, 2s, and 2p shells brings us to Ne, whose configuration is (1s)2 (2s)2 (2p)6 .
Next comes the 3s and 3p shells, which hold eight more electrons, and bring us to Ar:
1s2 2s2 2p6 3s2 3p6 = [Ne] 3s2 3p6 , where the symbol [N e] denotes the electronic configuration
of neon. At this point, things start to get interesting. The 4s orbitals preempt the 3d
orbitals, or at least most of the time. As we see from table 4.1, there are two anomalies
The 3d transition metal series ([Ar] core additions)
Element (AZ )
Sc21
Ti22
V23
Cr24
Mn25
Configuration 4s2 3d1 4s2 3d2 4s2 3d3 4s1 3d5
4s2 3d5
Element (AZ )
Fe26
Co27
Ni28
Cu29
Zn30
2
6
2
7
2
8
1
10
Configuration 4s 3d
4s 3d
4s 3d
4s 3d
4s2 3d10
Table 4.1: Electronic configuration of 3d-series metals.
in the otherwise orderly filling of the 3d shell. Chromiums configuration is [Ar] 4s1 3d5
rather than the expected [Ar] 4s2 3d4 , and coppers is [Ar] 4s1 3d10 and not [Ar] 4s2 3d9 . In
3
Hartree-Fock theory tends to overestimate ground state atomic energies by on the order of 1 eV per
pair of electrons. The reason is that electron-electron correlations are not adequately represented in the
Hartree-Fock many-body wavefunctions, which are single Slater determinants.
133
reality, the ground state is not a single Slater determinant and involves linear combinations
of different configurations. But the largest weights are for Cr and Cu configurations with
only one 4s electron. Zinc terminates the 3d series, after which we get orderly filling of the
4p orbitals.
Row five reiterates row four, with the filling of the 5s, 4d, and 5p shells. In row six, the
lanthanide (4f) series interpolates between 6s and 5d (see also table 4.2), and the actinide
(5f) series interpolates in row seven between 7s and 6d.
Shell:
Termination:
Shell:
Termination:
1s
He2
4d
Cd46
2s
Be4
5p
Xe54
2p
Ne10
6s
Ba56
3s
Mg12
4f
Lu71
3p
Ar18
5d
Hg80
4s
Ca20
6p
Rn86
3d
Zn30
7s
Ra88
4p
Kr36
5f/6d
No102
5s
Sr38
Table 4.2: Rough order in which shells of the Periodic Table are filled.
4.3.5
The electronic configuration does not uniquely specify a ground state. Consider, for example, carbon, whose configuration is 1s2 2s2 2p2 . The filled 1s and 2s shells are inert. However,
there are 62 = 15 possible ways to put two electrons in the 2p shell. It is convenient to
label these states by total L, S, and J quantum numbers, where J = L + S is the total
angular momentum. It is standard to abbreviate each such multiplet with the label 2S+1 LJ ,
where L = S, P, D, F, H, etc.. For carbon, the largest L value we can get is L = 2, which
requires S = 0 and hence J = L = 2. This 5-fold degenerate multiplet is then abbreviated
1 D . But we can also add together two l = 1 states to get total angular momentum L = 1
2
as well. The corresponding spatial wavefunction is antisymmetric, hence S = 1 in order to
achieve an antisymmetric spin wavefunction. Since |L S| J |L + S| we have J = 0,
J = 1, or J = 2 corresponding to multiplets 3 P0 , 3 P1 , and 3 P2 , with degeneracy 1, 3, and
5, respectively. The final state has J = L = S = 0: 1 S0 . The Hilbert space is then spanned
by two J = 0 singlets, one J = 1 triplet, and two J = 2 quintuplets: 0 0 1 2 2. That
makes 15 states. Which of these is the ground state?
The ordering of the multiplets is determined by the famous Hunds rules:
1. The LS multiplet with the largest S has the lowest energy.
2. If the largest value of S is associated with several multiplets, the multiplet with the
largest L has the lowest energy.
3. If an incomplete shell is not more than half-filled, then the lowest energy state has
J = |L S|. If the shell is more than half-filled, then J = L + S.
134
CHAPTER 4. MAGNETISM
Hunds rules are largely empirical, but are supported by detailed atomic quantum manybody calculations. Basically, rule #1 prefers large S because this makes the spin part of the
wavefunction maximally symmetric, which means that the spatial part is maximally antisymmetric. Electrons, which repel each other, prefer to exist in a spatially antisymmetric
state. As for rule #2, large L expands the electron cloud somewhat, which also keeps the
electrons away from each other.
Lets practice Hunds rules on a couple of ions:
Mn4+ : The electronic configuration [Ar] 4s2 3d3 has an unfilled 3d shell with three
electrons. First maximize S by polarizing all spins parallel (up, say), yielding S = 32 .
Next maximize L consistent with Pauli exclusion, which says L = 2 + 1 + 0 = 3.
Finally, since the shell is less than half-filled, J = |L S| = 32 , and the ground state
term is 4 F3/2 .
Fe2+ : The electronic configuration [Ar] 4s2 3d6 has an unfilled 3d shell with six electrons, or four holes. First maximize S by making the spins of the holes parallel,
yielding S = 2. Next, maximize L consistent with Pauli exclusion, which says
L = 2 + 1 + 0 + (1) = 2 (adding Lz for the four holes). Finally, the shell is
more than half-filled, which means J = L + S = 4. The ground state term is 5 D4 .
Nd3+ : The electronic configuration [Xe] 6s2 4f3 has an unfilled 4f shell with three
electrons. First maximize S by making the electron spins parallel, yielding S = 23 .
Next, maximize L consistent with Pauli exclusion: L = 3 + 2 + 1 = 6. Finally, since
the shell is less than half-filled, we have J = |L S| = 92 . The ground state term is
4I
9/2 .
4.3.6
135
Spin-Orbit Interaction
N
X
(ri ) li si .
(4.36)
i=1
This commutes with J 2 , L2 , and S 2 , so we can still classify eigenstates according to total
J, L, and S. The Wigner-Eckart theorem then guarantees that within a given J multiplet,
we can replace any tensor operator transforming as
X
J
R TJM R =
DM
(4.37)
M 0 (, , ) TJM 0 ,
M0
J J 0 J 00
.
(4.38)
T
J
JM TJ 00 M 00 J 0 M 0 = C
00 J
J
M M 0 M 00
In other words, if two tensor operators have the same rank, their matrix elements are
proportional. Both Hso and L S are products of rank L = 1, S = 1 tensor operators, hence
we may replace
= L S ,
Hso H
(4.39)
so
where = (N,
L, S) must be computed from, say, the expectation value of Hso in the
state JLSJ . This requires detailed knowledge of the atomic many-body wavefunctions.
However, once is known, the multiplet splittings are easily obtained:
so = 1 J 2 L2 S 2 )
H
2
= 12 ~2 J(J + 1) L(L + 1) S(S + 1) .
(4.40)
Thus,
E(N, L, S, J) E(N, L, S, J 1) = J ~2 .
(4.41)
If we replace (ri ) by its average, then we can find by the following argument. If the
last shell is not more than half filled, then by Hunds first rule, the spins are all parallel.
Thus S = 12 N and si = S/N , whence = hi/2S. Finding hi is somewhat tricky. For
Z 1 r/aB 1, one can use the WKB method to obtain (r = aB /Z) Z, whence
(4.42)
Z 2 2 ~2 Ry ,
(4.43)
hi
Ze2
~c
2
me4
~4
and
where = e2 /~c ' 1/137. For heavy atoms, Z 1 and the energy is on the order of that
for the outer electrons in the atom.
136
CHAPTER 4. MAGNETISM
For shells which are more than half filled, we treat the problem in terms of the holes relative
to the filled shell case. Since filled shells are inert,
Hso =
Nh
X
lj sj ,
(4.44)
j=1
where Nh = 4l + 2 N . lj and sj are the orbital and spin angular momenta of the holes;
P
P
L = j lj and S = j sj . We then conclude = hi/2S. Thus, we arrive at Hunds
third rule, which says
4.3.7
N 2L + 1
( half-filled)
>0
J = |L S|
(4.45)
N > 2L + 1
(> half-filled)
<0
J = |L + S| .
(4.46)
Consider an ion with a single d electron (e.g. Cr3+ ) or a single d hole (e.g. Cu2+ ) in a cubic
or octahedral environment. The 5-fold degeneracy of the d levels is lifted by the crystal
electric field. Suppose the atomic environment is octahedral, with anions at the vertices of
the octahedron (typically O2 ions). In order to minimize the Coulomb repulsion between
the d electron and the neighboring anions, the dx2 y2 and d3x2 r2 orbitals are energetically
disfavored, and this doublet lies at higher energy than the {dxy , dxz , dyz } triplet.
The crystal field potential is crudely estimated as
(nbrs)
VCF =
V (r R) ,
(4.47)
where the sum is over neighboring ions, and V is the atomic potential.
The angular dependence of the cubic crystal field states may be written as follows:
dx2 y2 (
r) =
1 Y (
r)
2 2,2
1 Y
(
r)
2 2,2
d3z 2 r2 (
r ) = Y2,0 (
r)
dxy (
r) =
dxz (
r) =
dyz (
r) =
i Y
(
r ) i2 Y2,2 (
r)
2 2,2
1
1
Y (
r ) + 2 Y2,1 (
r)
2 2,1
i Y
(
r ) i2 Y2,1 (
r)
2 2,1
(4.48)
Note that all of these wavefunctions are real . This means that the expectation value of
Lz , and hence of general L , must vanish in any of these states. This is related to the
phenomenon of orbital quenching, discussed below.
If the internal Hunds rule exchange energy JH which enforces maximizing S is large compared with the ground state crystal field splitting , then Hunds first rule is unaffected.
137
4.4
To compute the susceptibility, we will need to know magnetic energies to order H 2 . This
can be computed via perturbation theory. Treating the H = 0 Hamiltonian as H0 , we have
Z
ion
1
e2
X
2
n
(H
r
)
n
En (H) = En (0) + B H n L + 2S n +
i
~
8mc2
0 i=1
n0
X
n L + 2S n
n
L
+
2S
1
+ 2 2B H H
+ O(H 3 ) , (4.49)
0
~
E
E
n
n
0
n 6=n
where Zion is the number of electrons on the ion or atom in question. Since the (H ri )2
Larmor term is already second order in the field, its contribution can be evaluated in
138
CHAPTER 4. MAGNETISM
Figure 4.3: The splitting of one-electron states in different crystal field environments.
first order perturbation theory, i.e. by taking its expectation value in the state n . The
(L+2S)H term, which is linear in the field, is treated in second order perturbation theory.
4.4.1
If the ground state G is a singlet with J G = L G = S G = 0, corresponding to a
filled shell configuration, then the only contribution to the ground state energy shift is from
the Larmor term,
E0 (H) =
Zion
e2 H 2
X
2
G ,
G
r
i
12mc2
(4.50)
i=1
(4.51)
i=1
where n = N/V is the density of ions or atoms in question. The sum is over all the electrons
in the ion or atom. Defining the mean square ionic radius as
Zion
1
X
hr i
G
ri2 G ,
Zion
2
(4.52)
i=1
we obtain
ne2
=
Z hr 2 i = 61 Zion na3B
6mc2 ion
e2
~c
2
hr2 i
.
a2B
(4.53)
139
(4.54)
5
Typically, h(r/aB )2 i 1. Note that with na3B ' 0.1, we have || <
10 and M = H is
much smaller than H itself.
Molar Susceptibilities of Noble Gas Atoms and Alkali and Halide Ions
Atom or
Molar
Atom or
Molar
Atom or
Molar
Ion
Susceptibility Atom or Ion Susceptibility Atom or Ion Susceptibility
He
-1.9
Li+
-0.7
+
F
-9.4
Ne
-7.2
Na
-6.1
Cl
-24.2
Ar
-19.4
K+
-14.6
Br
-34.5
Kr
-28
Rb+
-22.0
I
-50.6
Xe
-43
Cs+
-35.1
Table 4.3: Molar susceptibilities, in units of 106 cm3 /mol, of noble gas atoms and alkali
and halide ions. (See R. Kubo and R. Nagamiya, eds., Solid State Physics, McGrow-Hill,
1969, p. 439.)
4.4.2
There are two cases to consider here. The first is when J = 0, which occurs whenever the
last shell is one electron short of being half-fille. Examples include Eu3+ (4f6 ), Cr2+ (3d4 ),
Mn3+ (3d4 ), etc. In this case, the first order term vanishes in E0 , and we have
2
z
Zion
z G
n
L
+
2S
2
X
X
ne
2
2
=
G
r
G
+
2
n
.
(4.55)
i
B
6mc2
En E0
i=1
n6=0
The second term is positive, favoring alignment of M with H. This is called van Vleck
paramagnetism, and competes with the Larmor diamagnetism.
The second possibility is J > 0, which occurs in all cases except filled shells and shells which
are one electron short of being half-filled. In this case, the first order term is usually dominant. We label the states by the eigenvalues of the commuting observables {J 2 , J z , L2 , S 2 }.
From the Wigner-Eckart theorem, we know that
JLSJz L + 2S JLSJz0 = gL (J, L, S) JLSJz J JLSJz0 ,
(4.56)
where
gL (J, L, S) =
3
2
S(S + 1) L(L + 1)
2J(J + 1)
(4.57)
140
CHAPTER 4. MAGNETISM
Figure 4.4: Reduced magnetization curves for three paramagnetic salts and comparison with
Brillouin theory predictions. L(x) = BJ (x) = ctnh (x) x1 is the Langevin function.
is known as the Lande g-factor. Thus, the effective Hamiltonian is
Heff = gL B J H/~ .
(4.58)
N F
= nJ BJ (JH/kB T ) ,
V H
1
2J
x
1
2J
ctnh (x/2J) .
(4.60)
(4.61)
(4.62)
(4.63)
p2
.
3kB T
(4.64)
141
p
J(J + 1) .
(4.65)
One finds that the theory works well in the case of rare earth ions in solids. There, the 4f
electrons of the rare earths are localized in the vicinity of the nucleus, and do not hybridize
significantly with orbitals from neighboring ions. In transition metal compounds, however,
Calculated and Measured Magneton Numbers of Rare Earth Ions
Electronic
Ground State magneton magneton
Ion
Configuration Term (2S+1) LJ
ptheory
pexpt
La3+
Ce3+
Pr3+
Nd3+
Pm3+
Sm3+
Eu3+
Gd3+
Tb3+
Dy3+
Ho3+
Er3+
Tm3+
Yb3+
Lu3+
[Xe] 4f 0
[Xe] 4f 1
[Xe] 4f 2
[Xe] 4f 3
[Xe] 4f 4
[Xe] 4f 5
[Xe] 4f 6
[Xe] 4f 7
[Xe] 4f 8
[Xe] 4f 9
[Xe] 4f 10
[Xe] 4f 11
[Xe] 4f 12
[Xe] 4f 13
[Xe] 4f 14
1S
0
2F
5/2
3H
4
4I
9/2
5I
4
6H
5/2
7F
0
8S
7/2
7F
6
6H
15/2
5I
8
4I
15/2
3H
6
2F
7/2
1S
0
0.00
2.54
3.58
3.62
2.68
0.84
0.00
7.94
9.72
10.63
10.60
9.59
7.57
4.54
0.00
<0
2.4
3.5
3.5
1.5
3.4
8.0
9.5
10.6
10.4
9.5
7.3
4.5
<0
Table 4.4: Calculated and measured effective magneton numbers p for rare earth ions.
(From N. W. Ashcroft and N. D. Mermin, Solid State Physics.) The discrepancy in the
cases of Sm and Eu is due to the existence of low-lying multiplets above the ground state.
one finds poor agreement except in the case of S states (L = 0). This is because crystal
field effects quench the orbital angular momentum, effectively rendering L = 0. Indeed, as
shown in Table 4.5, the theory can be rescued if one ignores the ground state terms obtained
by Hunds rules, and instead takes L = 0 and J = S, yielding gL = 2.
4.5
4.5.1
In a metal, the conduction electrons are delocalized and described by Block states. If we
ignore the orbital effects of the magnetic field, we can easily compute the susceptibility at
low fields and temperatures. At T = 0 and H = 0, and electrons fill respective Fermi
142
CHAPTER 4. MAGNETISM
[Ar] 3d1
[Ar] 3d1
[Ar] 3d2
[Ar] 3d3
[Ar] 3d3
[Ar] 3d3
[Ar] 3d4
[Ar] 3d4
[Ar] 3d5
[Ar] 3d5
[Ar] 3d6
[Ar] 3d7
[Ar] 3d8
[Ar] 3d9
2D
3/2
2D
3/2
3F
2
4F
3/2
4F
3/2
4F
3/2
5D
0
5D
0
6S
5/2
6S
5/2
5D
4
4F
9/2
3F
4
2D
5/2
1.55
1.55
1.63
0.77
0.77
0.77
0.00
0.00
5.92
5.92
6.70
6.54
5.59
3.55
1.73
1.73
2.83
3.87
3.87
3.87
4.90
4.90
5.92
5.92
4.90
3.87
2.83
1.73
1.8
2.8
3.8
3.7
4.0
4.8
5.0
5.9
5.9
5.4
4.8
3.2
1.9
Table 4.5: Calculated and measured effective magneton numbers p for transition metal
ions. (From N. W. Ashcroft and N. D. Mermin, Solid State Physics.) Due to the orbital
quenching, the angular momentum is effectively L = 0.
seas out to wavevector kF . In an external field H, the Zeeman interaction splits the energies
of the different polarization states:
HZ = B H .
(4.66)
(4.67)
N N
1 HZ
= B
.
V H
V
(4.68)
Now since the energies of the and electrons are shifted by B H, the change in their
number is
(4.69)
N = N = B H 12 g(F ) V ,
where g(F ) is the density of states per unit volume (including both spin species), at the
Fermi energy. Putting this all together, we find
M = 2B g(F ) H P H ,
(4.70)
143
where P = 2B g(F ) is the Pauli susceptibility. The Pauli susceptibility is positive, and
hence is paramagnetic.
Using the formula for the density of states,
g(F ) =
we find
P =
1 m
4 2 m
m kF
,
2 ~2
e2
~c
(4.71)
2
(kF aB ) .
(4.72)
Using e2 /~c ' 1/137.036 and assuming kF aB 1, we find P 106 , which is comparable
in magnitude (though opposite in sign) from the Larmor susceptibility of closed shells.
4.5.2
Landau Diamagnetism
Next, we investigate the orbital contribution. We assume a parabolic band, in which case
H=
2
1
e
p
+
A
+ B H .
c
2m
(4.73)
Appealing to the familiar results of a quantized charged particle in a uniform magnetic field,
the energy levels are given by
(n, kz , ) = (n + 12 ) ~c + B H +
~2 kz2
,
2m
(4.74)
where c = eH/m c is the cyclotron frequency. Note that B H = mm 21 ~c . The threedimensional density of states is a convolution of the two-dimensional density of states,
g2d () =
1 X
(n + 12 ) ~c ,
2
2`
(4.75)
n=0
p
where ` = ~c/eH is the magnetic length, and the one-dimensional density of states,
Thus,
1 dk
m 1
.
g1d () =
=
d
2 ~
(4.76)
m 1 X X n
g() =
.
n
2 ~ 2`2 n=0 =1
(4.77)
(4.78)
144
CHAPTER 4. MAGNETISM
m eH X X
(T, V, , H) = V kB T
F ( n~c ) ,
8 2 ~2 c n=0 =1
(4.79)
21 (1 + )~c ,
(4.80)
Z
n
o
d
F () = ln 1 + e()/kB T .
(4.81)
with = m /m,
and
Z
f (n) = dx f (x) + 21 f (0)
n=0
1
12
f 0 (0) +
1
720
f 000 (0) + . . . ,
(4.82)
which gives
V kB T m 3/2 X
=
2 2 2 ~3 =
( Z
d F () + 21 ~c F ( ) +
)
2 0
1
12 (~c ) F ( )
+ ...
(4.83)
(
)
2
= 1 + 21 1 31 2 (B H)2
+ O(H 4 ) (T, V, , 0) .
(4.84)
2
Thus,
1
M = 1
V
1
32
2
B H
,
2 H=0
2
(4.85)
(4.86)
The quantity P = 2B (n/) is simply the finite temperature Pauli susceptibility. The
orbital contribution is negative, i.e. diamagnetic. Thus, = P + L , where
L = 13 (m/m )2 P
(4.87)
145
4.6
4.6.1
This is none other than the famous Hubbard model , which has served as a kind of Rosetta
stone for interacting electron systems. The first term describes hopping of electrons along
the links of some regular lattice (hiji denotes a link between sites i and j). The second
term describes the local (on-site) repulsion of electrons. This is a single orbital model, so
the repulsion exists when one tries to put two electrons in the orbital, with opposite spin
polarization. Typically the Hubbard U parameter is on the order of electron volts. The last
term is the Zeeman interaction of the electron spins with an external magnetic field. Orbital
effects can be modeled by associating a phase exp(iAij ) to the hopping matrix element t
between sites i and j, where the directed sum of Aij around a plaquette yields the total
magnetic flux through the plaquette in units of 0 = hc/e. We will ignore orbital effects
here. Note that the interaction term is short-ranged, whereas the Coulomb interaction falls
off as 1/|Ri Rj |. The Hubbard model is thus unrealistic, although screening effects in
metals do effectively render the interaction to be short-ranged.
Within the Hubbard model, the interaction term is local and written as U n n on any given
site. This term favors a local moment. This is because the chemical potential will fix the
146
CHAPTER 4. MAGNETISM
mean value of the total occupancy n + n , in which case it always pays to maximize the
difference |n n |.
4.6.2
There are no general methods available to solve for even the ground state of an interacting
many-body Hamiltonian. Well solve this problem using a mean field theory due to Stoner.
The idea is to write the occupancy ni as a sum of average and fluctuating terms:
ni = hni i + ni .
(4.90)
Here, hni i is the thermodynamic average; the above equation may then be taken as a definition of the fluctuating piece, ni . We assume that the average is site-independent. This
is a significant assumption, for while we understand why each site should favor developing
a moment, it is not clear that all these local moments should want to line up parallel to
each other. Indeed, on a bipartite lattice, it is possible that the individual local moments
on neighboring sites will be antiparallel, corresponding to an antiferromagnetic order of the
pins. Our mean field theory will be one for ferromagnetic states.
We now write the interaction term as
(flucts )2
z }| {
ni ni = hn i hn i + hn i ni + hn i ni + ni ni
= hn i hn i + hn i ni + hn i ni + O (n)2
(4.91)
= 14 (m2 n2 ) + 12 n (ni + ni ) + 12 m (ni ni ) + O (n)2 ,
where n and m are the average occupancy per spin and average spin polarization, each per
unit cell:
n = hn i + hn i
(4.92)
m = hn i hn i ,
(4.93)
X
+ B H + 12 U m
ci ci + 41 Nsites U (m2 n2 ) ,
(4.94)
147
with the background. Second, the effective magnetic field has been shifted by an amount
1
2 U m/B , so the effective field is
Heff = H +
Um
.
2B
(4.95)
The bare single particle dispersions are given by (k) = t(k) + B H, where
X
t(k) =
t(R) eikR ,
(4.96)
and tij = t(Ri Rj ). Including the mean field effects, the effective single particle dispersions
become
e (k) = t(k) 21 U n + B H + 21 U m .
(4.97)
We now solve the mean field theory, by obtaining the free energy per site, (n, T, H). First,
note that = + 2n, where = /Nsites is the Landau, or grand canonical, free energy
per site. This follows from the general relation = F N ; note that the total electron
number is N = nNsites , since n is the electron number per unit cell (including both spin
species). If g() is the density of states per unit cell (rather than per unit volume), then we
have4
Z
(
+)/kB T
(
)/kB T
2
2
1
1
+ ln 1 + e
= 4 U (m + n ) +
n 2 kB T d g() ln 1 + e
(4.98)
where
U n and
From this free energy we derive two self-consistent
equations for and m. The first comes from demanding that be a function of n and not
of , i.e. / = 0, which leads to
B H + 21 U m.
Z
n
o
n=
d g() f (
) + f ( +
) ,
1
2
(4.99)
1
where f (y) = exp(y/kB T ) + 1
is the Fermi function. The second equation comes from
minimizing f with respect to average moment m:
Z
n
o
m=
d g() f (
) f ( +
) .
1
2
(4.100)
Here, we will solve the first equation, eq. 4.99, and use the results to generate a Landau
expansion of the free energy in powers of m2 . We assume that is small, in which case
we may write
Z
n
n = d g() f (
) + 12 2 f 00 (
) +
1
24
4 f 0000 (
) + . . .
(4.101)
148
CHAPTER 4. MAGNETISM
We write
() =
0 +
and expand in
. Since n is fixed in our (canonical) ensemble,
we have
Z
n = d g() f
0 ,
(4.102)
which defines
0 (n, T ).5 The remaining terms in the
expansion of eqn. 4.101 must sum
to zero. This yields
D(
0 )
+ 21 2 D0 (
0 ) + 12 (
)2 D0 (
0 ) + 21 D00 (
0 ) 2
+
where
1
24
D000 (
0 ) 4 + O(6 ) = 0 ,
(4.103)
Z
D() = d g() f 0 ( )
(4.104)
is the thermally averaged bare density of states at energy . Note that the k th derivative is
D
(k)
Z
() = d g (k) () f 0 ( ) .
(4.105)
Solving for
, we obtain
= 12 a1 2
1
24
3a31 6a1 a2 + a3 4 + O(6 ) ,
(4.106)
where
ak
D(k) (
0 )
.
D(
0 )
(4.107)
2
1
4Un
Z
+ n
0 d () f
0 ,
(4.109)
where g() = 0 (), so () is the integrated bare density of states per unit cell in the
absence of any magnetic field (including both spin species).
5
The Gibbs-Duhem relation guarantees that such an equation of state exists, relating any three intensive
thermodynamic quantities.
149
U 0
Hm + . . . ,
2B
(4.110)
where 0 = 2B D(
0 ) is the Pauli susceptibility, and
a=
1
2U
1
2 U D)
b=
1
32
(D0 )2 1 00
3D
D
!
U4
(4.111)
U 0
= 0 H +
m.
H
2B
(4.112)
U 0
H=0,
2B
H
.
B 1 21 U D
(4.113)
(4.114)
(4.115)
2
.
D(
0 )
(4.116)
U
Uc
The denominator of increases the susceptibility above the bare Pauli value 0 , and is
referred to as I kid you not the Stoner enhancement.
It is worth emphasizing that the magnetization per unit cell is given by
M =
1 H
= B m .
Nsites H
(4.117)
This is an operator identity and is valid for any value of m, and not only small m.
When H = 0 we can still get a magnetic moment, provided U > Uc . This is a consequence
of the simple Landau theory we have derived. Solving for m when H = 0 gives m = 0 when
U < Uc and
1/2 p
U
m(U ) =
U Uc ,
(4.118)
2b Uc
when U > Uc , and assuming b > 0. Thus we have the usual mean field order parameter
exponent of = 21 .
150
CHAPTER 4. MAGNETISM
4.6.3
Antiferromagnetic Solution
In addition to ferromagnetism, there may be other ordered states which solve the mean field
theory. One such example is antiferromagnetism. On a bipartite lattice, the antiferromagnetic mean field theory is obtained from
hni i = 12 n + 12 eiQRi m ,
(4.119)
where Q = (/a, /a, . . . , /a) is the antiferromagnetic ordering wavevector. The grand
canonical Hamiltonian is then
X
X
KMF = 12
tij ci cj + cj ci 12 U n
ci ci
i,j,
1
2Um
ci ci
iQRi
2
1
4 Nsites U (m
n2 )
1
2
X
k
ck,
!
(k) + 1 U n
1
c
U
m
k,
2
2
ck+Q,
1
(k + Q) + 12 U n
ck+Q,
2 Um
+ 41 Nsites U (m2 n2 ) ,
(4.120)
where (k) = t(k), as before. On a bipartite lattice, with nearest neighbor hopping only,
we have (k + Q) = (k). The above matrix is diagonalized by a unitary transformation,
yielding the eigenvalues
= (k)
p
= 2 + 2 ,
(4.121)
(4.122)
2
1
4 U (m
+n )+
n
Z
(
)/kB T
(
+ )/kB T
d g() ln 1 + e
+ ln 1 + e
.
1
2 kB T
(4.123)
The mean field equations are then
Z
n
o
d g() f
+ f
n=
1
2
1
=
U
1
2
o
g() n
f
f
.
2
(4.124)
(4.125)
4.6.4
151
Let us compare the mean field theories for the ferromagnetic and antiferromagnetic states
at T = 0 and H = 0.. Due to particle-hole symmetry, we may assume 0 n 1 without
loss of generality. (The solutions repeat themselves under n 2 n.) For the paramagnet,
we have
Z
(4.126)
n = d g()
2
1
4Un
Z
+ d g()
(4.127)
1
2Un
with
=
is the renormalized Fermi energy and g() is the density of states per
unit cell in the absence of any explicit (H) or implicit (m) symmetry breaking, including
both spin polarizations.
For the ferromagnet,
n=
1
2
Z+
d g() +
4
=
U
1
2
d g()
(4.128)
Z+
d g()
(4.129)
2
1
4Un
+
U
Z+
d g() +
d g()
(4.130)
(4.131)
Z
= 2 d g()
(
> 0)
(4.132)
Z
2
g()
= d
U
2 + 2
(4.133)
2
1
4Un
Z
p
2
+
d g() 2 + 2
U
0
(4.134)
152
CHAPTER 4. MAGNETISM
Figure 4.5: Mean field phase diagram of the Hubbard model, including paramagnetic (P),
ferromagnetic (F), and antiferromagnetic (A) phases. Left panel: results using a semicircular density of states function of half-bandwidth W . Right panel: results using a twodimensional square lattice density of states with nearest neighbor hopping t, from J. E.
Hirsch, Phys. Rev. B 31, 4403 (1985). The phase boundary between F and A phases is
first order.
p
where 0 =
2 2 and = 21 U m as before. Note that |
| for these solutions.
Exactly at half-filling, we have n = 1 and
= 0. We then set 0 = 0.
The paramagnet to ferromagnet transition may befirst or second order, depending on the
P ), where
P (n) is the paramagnetic
details of g(). If second order, it occurs at UcF = 1 g(
solution for
. The paramagnet to antiferromagnet transition is always second order in
this mean field theory, since
the RHS of eqn. (4.133) is a monotonic function of . This
R
transition occurs at UcA = 2
d g() 1 . Note that UcA 0 logarithmically for n 1,
since
P = 0 at half-filling.
For large U , the ferromagnetic solution always has the lowest energy, and therefore if UcA <
UcF , there will be a first-order antiferromagnet to ferromagnet transition at some value
U > UcF . In fig. 4.5, I plot the phase diagram obtained
by solving the mean field equations
assuming a semicircular density of states g() = 2 W 2 W 2 2 . Also shown is the phase
diagram for the d = 2 square lattice Hubbard model obtained by J. Hirsch (1985).
How well does Stoner theory describe the physics of the Hubbard model? Quantum Monte
Carlo calculations by J. Hirsch (1985) showed that the actual phase diagram of the d =
2 square lattice Hubbard Model exhibits no ferromagnetism for any n up to U = 10.
153
Furthermore, the antiferromagnetic phase is entirely confined to the vertical line n = 1. For
n 6= 1 and 0 U 10, the system is a paramagnet6 . In order to achieve a ferromagnetic
solution, it appears necessary to introduce geometric frustration, either by including a nextnearest-neighbor hopping amplitude t0 or by defining the model on non-bipartite lattices.
Numerical work by M. Ulmke (1997) showed the existence of a ferromagnetic phase at T = 0
on the FCC lattice Hubbard model for U = 6 and n [0.15, 0.87] (approximately).
4.7
While it is true that electronshave magnetic dipole moments, the corresponding dipoledipole interactions in solids are usually negligible. This is easily seen by estimating the
energy scale of the dipole-dipole interaction:
Edd =
m1 m2 3(m1 n)(m
2 n)
,
3
|r1 r2 |
(4.135)
4.7.1
e2
R4 R3 c c 0 c
V = 21
R1 R2
R1 R2 R3 0 cR4 ,
0
|r r |
0
R ,R
1
(4.137)
R3 ,R4
where we have assumed a single energy band. The Coulomb matrix element is
Z Z
2
e2
R4 R3 = d3r d3r0 (rR1 ) (r 0 R2 ) e
R1 R2
(r 0 R3 ) (rR4 ) .
0
|r r |
|r r 0 |
(4.138)
Due to overlap factors, the matrix element will be small unless R2 = R3 and R1 = R4 , in
which case we obtain the direct Coulomb interaction,
e2
R R0
V (R R0 ) = R R0
|r r 0 |
Z Z
2 e2
(r 0 R0 )2 .
= d3r d3r0 (r R)
0
|r r |
6
(4.139)
A theorem due to Nagaoka establishes that the ground state is ferromagnetic for the case of a single
hole in the U = system on bipartite lattices.
154
CHAPTER 4. MAGNETISM
V (0) nR nR +
1
2
V (R R0 ) nR nR0 ,
(4.140)
R6=R0
where nR nR + nR . The first term is the on-site Hubbard repulsion; one abbreviates
U V (0).
A second class of matrix elements can be identified: those with R1 = R3 R and R2 =
R4 R0 , with R 6= R0 . These are the so-called exchange integrals:
e2 0
R R
J(R R0 ) = R R0
|r r 0 |
Z Z
e2
(r 0 R) (r R0 )
= d3r d3r0 (r R) (r 0 R0 )
|r r 0 |
Z Z
e2
= d3r d3r0 (r) (r + R R0 )
(r 0 + R R0 ) (r 0 ) .
|r r 0 |
Note that J(R R0 ) is real. The exchange part of V is then
X
Vexchange = 12
J(R R0 ) cR cR0 cR0 0 cR0
(4.141)
(4.142)
R6=R0
0
14
J(R R0 ) nR nR0 + R R0 .
(4.143)
R6=R0
The nR nR0 piece can be lumped with the direct density-density interaction. What is new
is the Heisenberg interaction,
X
VHeis =
J(R R0 ) SR SR0 .
(4.144)
R6=R0
J(R R0 ) is usually positive, and this gives us an explanation of Hunds first rule, which
says to maximize S. This raises an interesting point, because we know that the ground
state spatial wavefunction for the general two-body Hamiltonian
H=
~2
21 + 22 + V |r1 r2 |
2m
(4.145)
is nodeless. Thus, for fermions, the ground state spin wavefunction is an antisymmetric
singlet state, corresponding to S = 0. Yet the V3+ ion, with electronic configuration
[Ar] 3d2 , has a triplet S = 1 ground state, according to Hunds first rule. Why dont the
two 3d electrons have a singlet ground state, as the no nodes theorem would seem to
155
imply? The answer must have to do with the presence of the core electrons. Two electrons
in the 1s shell do have a singlet ground state indeed that is the only possibility. But
the two 3d electrons in V3+ are not independent, but must be orthogonalized to the core
states. This in effect projects out certain parts of the wavefunction, rendering the no nodes
theorem inapplicable.
4.7.2
p21
e2
p2
e2
+ 2
2m |r1 Ra | 2m |r2 Rb |
+
e2
e2
e2
e2
+
.
|Ra Rb | |r1 Rb | |r2 Ra | |r1 r2 |
(4.146)
1 +
2
=
1
S
2
(4.147)
A symmetric spin wavefunction requires an antisymmetric spatial one, and vice versa.
Despite the fact that H does not explicitly depend on spin, the effective low-energy Hamiltonian for this system is
Heff = K + JS1 S2 .
(4.149)
The singlet-triplet splitting is E = ES=0 ES=1 = J, so if J > 0 the ground state is
the singlet, and if J < 0 the ground state is the three-fold degenerate triplet.
The one-electron 1s eigenfunction (r) satisfies the following eigenvalue equation:
~2 2 e2
(r) = 0 (r) (r) .
(4.150)
2m
r
In the Heitler-London approach, we write the two-electron wavefunction as a linear combination
(r1 , r2 ) = I (r1 , r2 ) + II (r1 , r2 ) ,
(4.151)
156
CHAPTER 4. MAGNETISM
with
I (r1 , r2 ) = (r1 Ra ) (r2 Rb ) a (r1 ) b (r2 )
(4.152)
(4.153)
Assuming the atomic orbital (r) to be normalized, we define the following integrals:
Z
= d3r a (r) b (r)
(4.154)
Z
Z
2 e2
e2
e2
e2
3
3
(4.155)
X = d r1 d r2 I (r1 , r2 )
+
Y = d r1 d r2 I (r1 , r2 ) II (r1 , r2 )
+
,
(4.156)
Rab r12 r1b r2a
with r1a = r1 Ra , etc. The expectation value of H in the state is
H = (||2 + ||2 ) (20 + X) + ( + ) (2||2 + Y ) ,
(4.157)
= ||2 + ||2 + ||2 ( + ) .
(4.158)
We now minimize hHi subject to the condition that be normalized, using a Lagrange
multiplier E to impose the normalization. Extremizing with respect to and yields
!
1
||2
20 + X
20 ||2 + Y
,
(4.159)
=E
2
2
||
1
20 || + Y
20 + X
and extremizing with respect to E yields the normalization condition
||2 + ||2 + ||2 ( + ) = 1 .
(4.160)
Y X||2
.
1 ||4
(4.162)
157
If J > 0, the triplet lies higher than the singlet, which says the ground state is antiferromagnetic. If J < 0, the triplet lies lower, and the ground state is ferromagnetic. The energy
difference is largely determined by the Y integral:
2
Z
Z
e
e2
3
3
Y = d r1 d r2 (r1 ) (r2 )
+
Rab r12
Z
e2
2 d3r a (r)
b (r) ,
(4.163)
|r Rb |
with (r) = a (r) b (r). The first term is positive definite for the Coulomb interaction.
The second term competes with the first if the overlap is considerable. The moral of the
story now emerges:
weak overlap = ferromagnetism (J < 0)
strong overlap = antiferromagnetism (J > 0) .
One finds that the H2 molecule is indeed bound in the singlet state the total energy has
a minimum as a function of the separation |Ra Rb |. In the triplet state, the molecule is
unbound.
4.7.3
4.7.4
Herrings approach
Conyers Herring was the first to elucidate the failure of Heitler-London theory at large
separations. He also showed how to properly derive a Heisenberg model for a lattice of
158
CHAPTER 4. MAGNETISM
N
Y
(ri Ri ) .
(4.165)
i=1
JP H E0 Pr .
(4.166)
The effective Hamiltonian is then
Heff = E0 +
JP Pr .
(4.167)
P SN
The spin permutation operators P may be written in terms of the Pauli spin matrices,
once we note that the two-cycle (ij) may be written
P(ij) =
1
2
+ 21 i j .
(4.169)
(4.170)
i6=j
1 + i k 1 + j k
h
i
1
= 4 1 + i j + j k + i k + ii j k .
4.8
(4.171)
(4.172)
Here, even and odd refer to the number of 2-cycles into which a given permutation is decomposed.
159
and write
Si = mi + Si ,
(4.173)
(4.174)
The last term is quadratic in the fluctuations, and as an approximation we ignore it. This
results in the following mean field Hamiltonian,
X
X
X
HMF = + 12
Jij mi mj
Hi +
Jij mj Si
i,j
= E0
Hieff
Si ,
(4.175)
Jij mi mj
(4.176)
(4.177)
where
E0 =
1
2
i,j
Hieff = Hi + 1
Jij mj .
Note how the effective field Hieff is a sum of the external field Hi and the internal field
P
Hiint = 1 j Jij mj . Self-consistency now requires that
Tr Si exp Hieff Si /kB T
mi =
Tr exp Hieff Si /kB T
,
(4.178)
where Tr means to sum or integrate over all local degrees of freedom (for site i). The free
energy is then
X
X
F {mi } = 12
Jij mi mj kB T
ln Tr exp Hieff Si /kB T .
(4.179)
i,j
(4.180)
(4.181)
160
CHAPTER 4. MAGNETISM
(4.182)
and
F =
1
2
Jij mi mj kB T
i,j
ln 1 + 2 cosh Hi +
Jij mj
(4.183)
mi = cos i =
(4.184)
R2
di exp Hieff cos i /kB T
0
P
I1 Hi + j Jij mj
,
=
P
I0 Hi + j Jij mj
(4.185)
(4.186)
O(3) model with Si = (sin i cos i , sin i sin i , cos i ). Suppose that mi points in the
direction of Hieff . Then
2
mi = cos i =
R2
di sin i cos i exp Hieff cos i /kB T
0
(4.187)
R2
2 di sin i exp Hieff cos i /kB T
0
k T
= ctnh Hieff /kB T B eff
Hi
X
= ctnh Hi +
Jij mj
j
(4.188)
Hi +
kB T
P
j
Jij mj
(4.189)
1
2
X
i,j
Jij mi mj kB T
X
i
ln
)
P
4 sinh Hi + j Jij mj
P
.
Hi + j Jij mj
(4.190)
4.8.1
161
Ferromagnets
Ising Model Let us assume that the system orders ferromagnetically, with mi = m on all
sites. Then, defining
X
J(R) eiqR ,
(4.191)
J(q)
=
R
(4.192)
m = tanh H + J(0)m
,
(4.193)
m/kB T kB T .
m = ctnh J(0)
m
J(0)
(4.194)
m/kB T , then,
With x J(0)
kB T
1
x x3
x = ctnh x =
+ ... ,
x
3 45
J(0)
(4.195)
hence Tc = J(0)/3k
B.
4.8.2
Antiferromagnets
If the lattice is bipartite, then we have two order parameters: mA and mB . Suppose
Jij = J < 0 if i and j are nearest neighbors, and zero otherwise. The effective fields on
the A and B sublattices are given by
eff
HA,B
H 1 zJmB,A ,
(4.196)
Note that the internal field on the A sublattice is 1 zJmB , while the internal field
on the B sublattice is 1 zJmA . For the spin-S quantum Heisenberg model, where
S z {S, . . . , +S}, we have
sinh S + 21
,
(4.197)
Tr exp( S) =
sinh 12
eff /k T , we have
hence, with = HA,B
B
S = S BS (S)
(4.198)
162
CHAPTER 4. MAGNETISM
(4.199)
In order to best take advantage of the antiferromagnetic interaction and the external magnetic field, the ordered state is characterized by a spin flop in which mA and mB are, for
weak fields, oriented in opposite directions in a plane perpendicular to H, but each with a
small component along H.
When H = 0, the mean field equations take the form
mA = SBS zJSmB /kB T
(4.200)
(4.201)
and mB = mB n,
4.8.3
Susceptibility
For T > Tc the system is paramagnetic, and there is a linear response to an external field,
mi
Mi
2F
=
Hj
Hj
Hi Hj
2 n
o
=
Si Sj Si Sj
kB T
ij =
(4.202)
(4.203)
where {i, j} are site indices and {, } are internal spin indices. The mean field Hamiltonian
is, up to a constant,
X
HMF =
Hieff Si ,
(4.204)
i
which is a sum of single site terms. Hence, the response within HMF must be purely local
as well as isotropic. That is, for weak effective fields, using Mi = mi ,
Mi = 0 Hieff = 0 Hi + 2 0 Jij Mj ,
(4.205)
ij 2 0 Jij Mj = 0 Hi ,
(4.206)
which is equivalent to
and the mean field susceptibility is
1
0 2 J 1 .
ij =
ij
(4.207)
It is convenient to work in Fourier space, in which case the matrix structure is avoided and
one has
0
.
(4.208)
(q)
1 2 0 J(q)
163
Tr S exp(H S/kB T )
M = S =
Tr exp(H S/kB T )
,
= SB (SH/k T ) H
S
(4.209)
0 ,
(4.210)
=
0
kB T
Tr 1
where 0 = N1 Tr (S 2 )/ Tr 1, where N is the number of components of S . Thus, for the Ising
model (N = 1) we have Ising
= 2 /kB T , while for the spin-S quantum Heisenberg model
0
we have Heis
= S(S + 1) 2 /3kB T . Note that 0 T 1 ; the splitting of the degenerate
0
energy levels by the magnetic field is of little consequence at high temperatures.
In many cases one deals with single ion anisotropy terms. For example, one can add to
the Heisenberg Hamiltonian a term such as
X
2
Ha = D
Siz ,
(4.211)
i
which for D < 0 results in an easy axis anisotropy (i.e. the spins prefer to align along the
z-axis),
and for D > 0 results in an easy plane anisotropy (i.e. the spins prefer to lie in
the (x, y) plane). Since this term is already the sum of single site Hamiltonians, there is no
need to subject it to a mean field treatment. One then obtains the mean field Hamiltonian
X
X
2
HMF = D
Siz
Hieff Si .
(4.212)
i
In this case, 0 is now anisotropic in spin space. The general formula for 0 is
0 =
2
S S
kB T
(4.213)
where the thermodynamic average is taken with respect to the single site Hamiltonian.8
One then has
1
I 2 J(q)
0
,
(4.214)
0 (q) = 0
4.8.4
164
CHAPTER 4. MAGNETISM
We now ask what distribution P minimizes the free energy F = hHi T S. Working in an
eigenbasis of H, we have
F =
P E + kB T
P ln P .
(4.216)
(4.217)
and demand dF /dP = 0 for all as well as dF /d = 0. This results in the Boltzmann
distribution,
X
1
,
Z=
eEl /kB T .
(4.218)
Peq = eE /kB T
Z
l
Peq
with = 1/kB T , is too cumbersome to work with. We replace this with a variational
single-site distribution,
P =
Pi (i ) =
N
Y
j=1
1
2 (1
Pi (i )
(4.220)
+ mi ) i ,+1 + 12 (1 mi ) i ,1 .
(4.221)
(
s(m) = kB
1 + m 1 + m 1 m 1 m
ln
+
ln
2
2
2
2
)
(4.223)
165
(4.224)
hence from
H = 12
Jij i j
i,j
Hi i ,
(4.225)
(
+ kB T
(4.226)
X
i
1 + mi 1 + mi 1 mi 1 mi
ln
+
ln
2
2
2
2
Varying with respect to each mi , we obtain the coupled nonlinear mean field equations,
h X
i
mi = tanh
Jij mj + Hi /kB T .
(4.227)
j
For uniform magnetization (mi = m i), the free energy per site is
)
(
1 + m 1 m 1 m
F
1
+
m
m2 Hm + kB T
= 12 J(0)
ln
+
ln
N
2
2
2
2
1
1
= 12 kB T J(0)
m2 Hm + 12
kB T m4 + 30
kB T m6 + . . .
(4.228)
To compute the correlations, we may use the expression
o
2 n
ij (T ) =
hi j i hi i hj i
kB T
mi
2F
Mi
=
=
.
=
Hj
Hj
Hi Hj
(4.229)
(4.230)
Thus, there are two ways to compute the susceptibility. One is to evaluate the spin-spin
correlation function, as in (4.229). The other is to differentiate the magnetization to obtain
the response function, as in (4.230). The equality between the two called the fluctuationdissipation theorem is in fact only valid for the equilibrium Boltzmann distribution Peq .
Which side of the equation should we use in our variational mean field theory? It is more
accurate to use the response function. To roughly see this, let us write P = P eq + P , with
P small in some sense. The free energy is given by
F
eq
+ O (P )2 .
F [P ] = F [P ] + P
(4.231)
P
eq
P =P
Our variational treatment guarantees that the second term vanishes, since we extremize F
with respect to P . Thus, in some sense, the error in F is only of order (P )2 . If we compute
the correlation function using hAi = Tr (P A), where A is any operator, then the error will
be linear in P . So it is better to use the response function than the correlation function.
EXERCISE: Articulate the correspondence between this variational version of mean field
theory and the neglect of fluctuations approach we derived earlier.
166
CHAPTER 4. MAGNETISM
4.9
Magnetic Ordering
=
0
.
1 2 0 J(Q)
+ 2 0 J(Q)
J(q)
(4.232)
= 1.
2 0 (Tc ) J(Q)
(4.233)
J(q)
= J(Q)
1 (q Q)2 R2 + . . . ,
(4.234)
we have
(q)
where
2 (T ) =
0 /R2
2 (T ) + (q Q)2
00 (Tc )
0 (Tc )
(4.235)
R2 (T Tc ) .
(4.236)
Thus, (T ) (T Tc )1/2 . The real space susceptibility (Ri Rj ) oscillates with wavevector Q and decays on the scale of the correlation length (T ).
Ferromagnet: Jij = +J > 0 if i and j are nearest neighbors; otherwise Jij = 0. On a
hypercubic lattice (d dimensions, 2d nearest neighbors), we then have
X
J(q)
=J
eiq = 2J cos(q1 a) + cos(q2 a) + . . . + cos(qd a) .
(4.237)
(T
1 2
3 S(S + 1)/kB
P
TC ) + TC d1 d=1 1
cos(q a)
.
(4.238)
2 S(S + 1)
3kB (T TC )
(4.239)
167
J(q)
= J
eiq = 2J cos(q1 a) + cos(q2 a) + . . . + cos(qd a) .
(4.240)
The ordering wavevector is Q = (/a, . . . , /a), at the zone corner, where J(Q)
=
2
2dJ. For the spin-S Heisenberg model, then, TN = 3 d S(S + 1) J/kB , and the susceptibility is
2 S(S + 1)/3kB
(q)
=
(4.241)
.
P
(T TN ) + TN d1 d=1 1 + cos(q a)
The uniform susceptibility = (q
= 0) is then
(T ) =
2 S(S + 1)
3kB (T + TN )
(4.242)
which does not diverge. Indeed, plotting 1 (T ) versus T , one obtains an intercept
along the T -axis at T = TN . This is one crude way of estimating the Neel temperature. What does diverge is the staggered susceptibility stag (Q,
T ), i.e. the
susceptibility at the ordering wavevector:
stag (T ) =
2 S(S + 1)
3kB (T TN )
(4.243)
3
1
2, 2
4 3 1
b1 =
, 2
a 3 2
4
b2 = 0, 1 ,
a 3
(4.244)
(4.245)
where a is the lattice constant. The six nearest neighbor vectors are then
n
o
a1 , a2 , a2 a1 , a1 , a2 , a1 a2 ,
(4.246)
J(q)
= 2J cos(2x1 ) + cos(2x2 ) + cos(2x1 2x2 ) .
(4.247)
We suspect that this should be maximized somewhere along the perimeter of the
168
CHAPTER 4. MAGNETISM
+J1 > 0
Jij = J2 < 0
(4.248)
Then
J(q)
= 2J1 cos(qx a) + cos(qy a) + cos(qz a)
(4.249)
4J2 cos(qx a) cos(qy a) + cos(qx a) cos(qz a) + cos(qy a) cos(qz a) .
The ordering wavevector is then
(
J1
+ y + z)
(x
a1 cos1 4J
2
Q=
0
if J1 < 4J2
(4.250)
if J1 4J2 .
Thus, for J1 < 4J2 the order is incommensurate with the lattice. The maximum value
of J(q)
is
( 2
3J1
if J1 < 4J2
J(Q)
= 4J2
(4.251)
6(J1 2J2 ) if J1 4J2 ,
hence incommensurate order sets in at TI = S(S + 1)J12 /4kB J2 . The uniform susceptibility is
2 S(S + 1)/3kB
.
(0)
=
(4.252)
2
T 8 TI JJ21 1 2J
J1
Thus,
(
(T ) '
4.9.1
C
T +T
C
T T
(4.253)
H=
zX
i<j
inter
}|
{
z
}|
{
X
X
k
Jij Si Sj
Jij Si Sj
Hi Si .
i<j
(4.254)
Here, Jij only connects sites within the same plane (quasi-2d) or chain (quasi-1d), while
Jij only connects sites in different planes/chains. We assume that we have an adequate
theory for isolated plains/chains, and we effect a mean field decomposition on the interplane/interchain term:
(fluct)2
z }| {
Si Sj = hSi i hSj i + hSi i Sj + hSj i Si + Si Sj ,
(4.255)
169
(4.256)
(q,
) =
k (q , )
k
k
(q , )
1 2 J (q )
k
(4.257)
k (q , ) is assumed known.
where
k
4.9.2
Quasi-1D Chains
Consider a ferromagnet on a cubic lattice where the exchange interaction along the zdirection (k) is much larger than that in the (x, y) plane (). Treating the in-plane interactions via mean field theory, we have
1D (qz )
1 2 J (q )
(4.258)
n
o
J (q ) = 2J cos(qx a) + cos(qy a) .
(4.259)
(q
, qz ) =
with
1D (qz )
(4.261)
Thus,
2 X
1D (qz ) =
tanh|n| (Jk /kB T ) einqz c
kB T n=
2
1
kB T cosh(2J /kB T ) sinh(2J /kB T ) cos(qz c)
k
k
(4.262)
2 2 1
1
,
2
ckB T + qz2
(4.263)
where c is the lattice spacing along the chains, and where the last approximation is valid
for q 0 and . The correlation length in this limit is given by
(T ) '
c
exp(2Jk /kB T ) .
2
(4.264)
170
CHAPTER 4. MAGNETISM
Note that (T ) diverges only at T = 0. This is consistent with the well-known fact that
the lower critical dimension for systems with discrete global symmetries and short-ranged
interactions is d = 1. That is to say that there is no spontaneous breaking of any discrete
symmetry in one-dimension (with the proviso of sufficiently short-ranged interactions). For
continuous symmetries the lower critical dimension is d = 2, which is the content of the
Hohenberg-Mermin-Wagner (HMW) theorem.
Accounting for the residual interchain interactions via mean field theory, we obtain the
anisotropic (in space) susceptibility
(q
, qz ) =
1D (qz )
.
1 2 2J cos(qx a) + cos(qy a)
1D (qz )
(4.265)
has a
pole. The equation for this pole is
4 2 J 1D = 1
4J
= exp(2Jk /kB Tc ) .
kB Tc
(4.266)
1
(4.267)
2Jk
ln Jk /2J
+ ... .
(4.268)
Thus, Tc > 0 for all finite J , with Tc going to zero rather slowly as J 0.
Similar physics is present in the antiferromagnetic insulator phase of the cuprate superconductors. The antiferromagnetic (staggered) susceptibility of the two-dimensional Heisenberg
1 exp(Jk T ), where is a dimensionless measure
model diverges as T 0 as stag
B
2D J
of quantum fluctuations. As in the d = 1 Ising case, there is no phase transition at any
finite temperature, in this case owing to the HMW theorem. However, when the quasi-2D
layers are weakly coupled with antiferromagnetic coupling J 0 (the base structure is a cubic
perovskite), three-dimensional Neel ordering sets in at the antiferromagnetic wavevector
Q = (/a, /a, /c) at a critical temperature TN J/kB ln(J/J 0 ).
4.10
[S + , S ] = 2 S z .
(4.269)
171
The Holstein-Primakoff transformation (1940) maps the spin algebra onto that of a single
bosonic oscillator:
S + = a (2S a a)1/2
(4.270)
S = (2S a a)1/2 a
(4.271)
S =a aS .
(4.272)
The state | S z = S i is the vacuum | 0 i in the boson picture. The highest weight state,
| S z = +S i corresponds to the state | 2S i in the boson picture, i.e. an occupancy of n = 2S
bosons.
EXERCISE: Verify that the bosonic representation of the spin operators in (4.272) satisfies
the SU(2) commutation relations of quantum spin.
What does it mean to take the square root of an operator like 2S a a? Simple! Just
evaluate it in a basis diagonal in a a, i.e. the number basis:
a a | n i = n | n i
(4.273)
Note that physical boson states are restricted to n {0, 1, . . . , 2S}. What about states with
n > 2S? The nice thing here is that we neednt worry about them at all, because S + , S ,
and S z do not connect states with 0 n 2S to states with n > 2S. For example, when
applying the spin raising operator S + to the highest weight state | S z = +S i, in boson
language we have
S + S z = +S = a (2S a a)1/2 n = 2S = 0 ,
(4.274)
as required.
While the HP transformation is exact, it really doesnt buy us anything unless we start
making some approximations and derive a systematic expansion in spin wave interactions.
4.10.1
Consider the classical ground state | F i = | i in which all spins are pointing down,
with S z = S. In the boson language, the occupancy at each site is zero. This is in fact
an eigenstate of the Heisenberg Hamiltonian
X
H=
Jij Si Sj
(4.275)
i<j
P
with eigenvalue E0 = S 2 i<j Jij . If all the interactions are ferromagnetic, i.e. Jij >
0 (i, j), then this state clearly is the ground state. We now express the Heisenberg
interaction Si Sj in terms of the boson creation and annihilation operators. To this end,
we perform a Taylor expansion of the radical,
1 a a
1 a a 2
1/2
+ ... ,
(4.276)
(2S a a) = 2S 1
2 2S
8 2S
172
CHAPTER 4. MAGNETISM
so that
Si+ Sj + 21 Si Sj+ + Siz Sjz
(4.277)
aj aj
ai ai
+ ...
1
+ . . . aj
(4.278)
= S ai 1
4S
4S
aj aj
ai ai
+ . . . ai aj 1
+ . . . + (ai ai S) (aj aj S)
+S 1
4S
4S
n
= S 2 + S ai aj + aj ai ai ai aj aj + ai ai aj aj 14 ai ai ai aj
o
(4.279)
14 ai aj aj aj 14 aj ai ai ai 14 aj aj aj ai + O(1/S) .
Si Sj =
1
2
Note that a systematic expansion in powers of 1/S can be performed. The Heisenberg
Hamiltonian now becomes
classical ground
state energy O(S 2 )
spin-wave
z
}|
zX
}|
{ interactions
z }| {
X {
H = S 2
Jij + S
Jij ai ai + aj aj ai aj aj ai + O(S 0 ) .
i<j
(4.280)
i<j
We assume our sites are elements of a Bravais lattice, and we Fourier transform:
1 X +iqRi
e
aq
ai =
N q
1 X iqRi
aq =
ai
e
N i
1 X iqRi
ai =
e
aq
N q
1 X +iqRi
aq =
ai .
e
N i
(4.281)
(4.282)
Note that the canonical commutation relations are preserved by this transformation:
[ai , aj ] = ij
(4.283)
(4.284)
X
J(q)
J(0)
aq aq ,
(4.285)
~q = S J(0)
hX
i
= 16 S
J(R) R2 q 2 + O(q 4 ) .
R
(4.286)
(4.287)
4.10.2
173
1/2
1/2
hSi+ Sj i = ai 2S ai ai
2S aj aj
aj
= 2S
hai aj i +
Z d
dk
= 2S
O(S 0 )
eik(Rj Ri )
(4.288)
(4.289)
(4.290)
(4.291)
(4.292)
(4.293)
Now as k 0 the denominator above vanishes as k 2 , hence the average spin polarization
per site diverges when d 2. This establishes a poor mans version of the HMW theorem:
as infinite spin polarization is clearly absurd, there must have been something wrong with
our implicit assumption that long-ranged order persists to finite T . In d = 3 dimensions,
one finds hSiz i = S + O(T 3/2 ).
4.10.3
The case of the ferromagnet is special because the classical ground state | F i is in fact a
quantum eigenstate indeed the ground state of the ferromagnetic Heisenberg Hamiltonian.9 In the case of the Heisenberg antiferromagnet, this is no longer the case. The
ground state itself is a linear combination of classical states. What is the classical ground
state? For an antiferromagnet on a bipartite lattice,10 the classical ground state has each
sublattice maximally polarized, with the magnetization on the two sublattices oppositely
this means Siz = S is i A and Siz = +S
oriented. Choosing the axis of polarization as z,
if i B. Well call this state | N i, since it is a classical Neel state.
Let is assume that the lattice is a Bravais lattice with a two-element basis described by
basis vectors 0 and . Thus, if R is any direct lattice vector, an A sublattice site lies at R
9
Of course, |Fi is also an eigenstate the highest lying excited state of the antiferromagnetic Heisenberg
Hamiltonian.
10
A bipartite lattice is one which may be decomposed into two sublattices A and B, such that all the
neighbors of any site in A lie in B, and all the neighbors of any site in B lie in A. Examples of bipartite
lattices: square, honeycomb, simple cubic, body-centered cubic, hexagonal. Examples of lattices which are
not bipartite: triangular, Kagome, face-centered cubic.
174
CHAPTER 4. MAGNETISM
X
1
2
(4.294)
R,R0
.
Here SA (R) represents the spin on the A sublattice located at position R, while SB (R)
represents the B sublattice spin located at R + . The factor of 12 multiplying the JAA and
JBB terms avoids double-counting the AA and BB interactions. The Neel state will be the
classical ground state if JAA > 0 and JBB > 0 and JAB < 0. It may remain the ground
state even if some of the interactions are frustrating, i.e. JAA < 0, JBB < 0, and/or JAB > 0
between certain sites.
Wed like the Neel state | N i = | . . . i to be the vacuum for the Holstein-Primakoff
bosons. To accomplish this, we rotate the spin operators on the B sublattice by about
the y-axis
in the internal SU(2) space, sending S x S x , S y S y , and S z S z . In
the language of HP bosons, we have the following:
A Sublattice
B Sublattice
S + = a (2S a a)1/2
1/2
S = (2S a a)
S + = (2S b b)1/2 b
S z = a a S
(4.295)
1/2
S = b (2S b b)
(4.296)
S z = S b b
(4.297)
(4.298)
(4.299)
(4.300)
Xn
o
21 JAA (R R0 ) 12 JBB (R R0 ) + JAB (R R0 ) .
(4.301)
R,R0
X
JAA (R R0 ) aR aR0 aR aR + JBB (R R0 ) bR bR0 bR bR
R,R0
0
+ JAB (R R )
aR aR
bR bR
aR bR0
aR bR0
175
1 X ikR
aR =
e
ak
N k
1 X ik(R+)
bR =
e
bk ,
N k
(4.302)
(4.303)
which leads to
X
1 XX
0
0
JAA (R R0 ) aR aR0 =
JAA (R R0 )ei(k R kR) ak ak0
N
R,R0
k,k0 R,R0
X
=
JAB (k) ak ak
(4.304)
JAB (R R
) aR bR0
R,R0
1 XX
=
JAB (R R0 )ei
N
k,k0 R,R0
X
=
JAB (k) ak bk ,
k0 (R0 +)kR
ak bk0
(4.305)
where, assuming JAA , JBB and JAB are functions only of the magnitude of their arguments,
X
JAA (k)
JAA |R| eikR
(4.306)
R
JBB (k)
JBB |R| eikR
(4.307)
JAB |R + | eik(R+) .
(4.308)
JAB (k)
X
R
Note that JAA (k) = JAA (k) = JAA (k) (similarly for JBB ), and JAB (k) = JAB (k) .
The spin-wave Hamiltonian may now be written as
(
X
Hsw = S
JAA (0) JAA (k) JAB (0) ak ak + JBB (0) JBB (k) JAB (0) bk bk
k
(4.309)
In other words,
Hsw =
o
Xn
BB
AA
k ak ak + k bk bk + k ak bk + k ak bk
(4.310)
with
AA (0) JAA (k) JAB (0)
AA
=
S
J
k
BB
k = S JBB (0) JBB (k) JAB (0)
(4.311)
(4.312)
176
CHAPTER 4. MAGNETISM
and
k = S JAB (k) .
(4.313)
BB
Henceforth we shall assume JAA (R) = JBB (R), so AA
k = k k .
Note that the vacuum 0 for the a and b bosons is not an eigenstate of Hsw , owing to the
spin-wave pair creation term k ak bk . This can be traced back to the effect on the Neel
state of the Heisenberg interaction,
Si Sj =
1
2
(4.314)
If i A and j B, then the term Si+ Sj acts on the configuration | S , +S i and converts
it to 2S | S + 1 , S 1 i. Nevertheless, we can diagonalize Hsw by means of a canonical
(but not unitary!) transformation, known as the Bogoliubov transformation. Note that for
the spin-wave Hamiltonian couples only four operators: a , a , b , and b .
each k ,
k
k
k
k
We write the Bogoliubov transformation as
ak = uk k vk k
bk = uk k vk k
(4.315)
ak = uk k vk k
vk k
bk = uk k
(4.316)
One can readily verify that this transformation preserves the canonical bosonic commutation
relations,
ak , ak0 = bk , bk0 = k , k 0 = k , k 0 = kk0
(4.317)
provided that
uk uk vk vk = 1 .
(4.318)
k = uk bk + vk ak
(4.319)
k = uk ak + vk bk
= uk bk + vk ak .
k
(4.320)
vk = exp(ik ) sinh(k ) .
(4.321)
Well write
uk = exp(ik ) cosh(k )
(4.322)
(4.323)
sinh(k ) ak
(4.324)
.
(4.325)
177
k ak ak + bk bk = k cosh(2k ) k k + k
k + 1 k
k sinh(2k ) k k
+ k k
k ak ak + k ak bk = k sinh(2k ) k k + k
k + 1
+ k cosh(2k ) k k
+ k k ,
(4.326)
(4.327)
where we have taken k = 12 arg(k ). Up until now, k has been arbitrary. We now use
k
sinh(2k ) =
Ek
(4.329)
q
2
2k k .
(4.330)
Hsw =
X
Ek k k + k k +
Ek k .
(4.331)
(4.332)
SW + SXk2 + . . .
X
X
2
k = +S
JAB |R + | 16 S k2
JAB |R + | R + + . . .
R
R
2
SW + SY k + . . . .
(4.333)
p
The energy dispersion is linear: Ek = ~c|k|, where c = S 2W (X + Y ). Antiferromagnetic spin waves are Goldstone bosons corresponding to the broken continuous symmetry
of global spin rotation. The dispersion vanishes linearly as k 0, in contrast to the case
of ferromagnetic spin waves, where Ek vanishes quadratically.
178
CHAPTER 4. MAGNETISM
(4.334)
BZ
where v0 is the Wigner-Seitz cell volume, and the integral is over the first Brillouin zone.
The deviation S z = ha ai from the classical value hS z i = S is due to thermal and
quantum fluctuations. Note that even at T = 0, when the thermal fluctuations vanish,
there is still a reduction in sublattice magnetization due to quantum fluctuations. The Neel
state satisfies the Siz Sjz part of the Heisenberg interaction, but the full interaction prefers
neighboring spins to be arranged in singlets, which involves fluctuations about local Neel
order.
Weve seen that k ' SW and Ek ' ~c |k| as k 0. Thus, the integrand behaves as
T /k2 for the first term and as 1/|k| for the second term. The integral therefore diverges in
d 2 at finite T and in d = 1 even at T = 0. Thermal and quantum fluctuations melt the
classical ordered state.
4.10.4
E(T ) = V
(4.335)
where d = 2 d/2 /(d/2) is the area of the unit sphere in d dimensions. Thus, E(T )
d
T 1+ , leading to a low-temperature heat capacity of
CV = (2 +
1
2 d) (1
1
2 d)
k B V d
(2)d
kB T
A
d/
.
(4.336)
179
At high T , one must impose a cutoff at the edge of the Brillouin zone, where k /a, in
order not to overcount the modes. One finds
Z d
dk
E(T ) = kB T V
= N kB T ,
(4.337)
(2)d
where N is the number of unit cells. This simply is the Dulong-Petit result of kB T per
mode.
For ferromagnetic spin waves, we found = 2, hence CV T d/2 at low temperatures. As
we shall see, for antiferromagnetic spin waves, one has = 1, as in the case of acoustic
phonons, hence CV T d .
Suppose we write the long-wavelength ferromagnetic spin-wave dispersion as ~q = CJ(qa)2 ,
where a is the lattice spacing, J is the nearest neighbor exchange, and C is a dimensionless
constant. The ferromagnetic low-temperature specific heat is then
CVF
= (2 +
1
2 d) (1
1
2 d)
kB V d
(2a)d
kB T
CJ
d/2
,
(4.338)
hence CVF (T /J )d/2 , with J CJ/kB . Acoustic phonons with a k = ~c|k| dispersion
lead to a Debye heat capacity
CVD
k V d
= (2 + d) (1 + d) B d
(2a)
kB T
~c/a
d
,
(4.339)