International Max Planck Research School on Science and Technology of

Nanostructures
Max-Planck-Institut für Mikrostrukturphysik
Martin-Luther-Universität Halle-Wittenberg

Lecture script

Introduction to the Theory of

Nanostructures

Jürgen Henk

July 14, 2006

Contents

1 Introduction 4
1.1 What are nanostructures? . . . . . . . . . . . . . . . . . . . . . . 4
1.2 What makes nanostructures unique and interesting? . . . . . . . 6
1.3 Historical remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Scope of the lecture . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Review of quantum mechanics 10

2.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Dirac notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Free particle in one dimension . . . . . . . . . . . . . . . . . . . . 20
2.5 Density of states . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.6 Quantum well . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Tunneling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3 Review of solid-state theory 40

3.1 Lattice structures . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2 Electronic structure . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Surfaces and interfaces . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4 Defects and disorder . . . . . . . . . . . . . . . . . . . . . . . . . 71
3.5 From the many-electron to the single-electron problem . . . . . . 75

4 Nanoscale probes 80
4.1 Low-energy electron diffraction . . . . . . . . . . . . . . . . . . . 80
4.2 Photoelectron spectroscopy . . . . . . . . . . . . . . . . . . . . . 91
4.3 Other scattering techniques . . . . . . . . . . . . . . . . . . . . . 97
4.4 Scanning tunneling spectroscopy . . . . . . . . . . . . . . . . . . 102

5 Electronic transport in nanostructures 109

5.1 Spin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.2 Important quantities in mesoscopic transport . . . . . . . . . . . 117
5.3 Transport in two-dimensional electron gases . . . . . . . . . . . . 119
5.3.1 Landau quantization . . . . . . . . . . . . . . . . . . . . . 119
5.3.2 Quantum Hall effect . . . . . . . . . . . . . . . . . . . . . 126
5.3.3 Spin-orbit coupling in a two-dimensional electron gas . . . 133
5.4 Transport regimes . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.4.1 Diffusive transport . . . . . . . . . . . . . . . . . . . . . . 140
5.4.2 Ballistic transport . . . . . . . . . . . . . . . . . . . . . . 150
5.4.3 Single-electron tunneling . . . . . . . . . . . . . . . . . . . 155

2
6 Magnetic nanostructures 164
6.1 Tunnel magnetoresistance . . . . . . . . . . . . . . . . . . . . . . 164
6.2 Giant magnetoresistance . . . . . . . . . . . . . . . . . . . . . . . 185
6.3 Spin effects in the Coulomb blockade . . . . . . . . . . . . . . . . 227
6.4 Magnetism at the nanoscale . . . . . . . . . . . . . . . . . . . . . 232
6.4.1 Exchange coupling in magnetic multilayers . . . . . . . . 232
6.4.2 Magnetocrystalline anisotropy of ultrathin films . . . . . . 251
6.5 Elements of solid-state magnetism . . . . . . . . . . . . . . . . . 267
6.5.1 Hund’s rules . . . . . . . . . . . . . . . . . . . . . . . . . 267
6.5.2 Landau diamagnetism . . . . . . . . . . . . . . . . . . . . 270
6.5.3 Pauli paramagnetism . . . . . . . . . . . . . . . . . . . . . 276
6.5.4 Band ferromagnetism . . . . . . . . . . . . . . . . . . . . 278
6.5.5 Quantum-well induced ferromagnetism in thin films . . . 282
6.5.6 Other magnetisms . . . . . . . . . . . . . . . . . . . . . . 285

Bibliography 290

3
1 Introduction
This lecture provides an introduction to the basic theory of nanostructures,
aiming at graduate students.

1.1 What are nanostructures?

The word ‘nano’ stems from the Greek ‘nanos’ (or Latin ‘nanus’), meaning
‘dwarf’. Nano refers to the 10−9 power, or one billionth. In the term ‘nano-
structure’ it refers to a meter, or a nanometer (short: nm), which is on the
scale of atomic diameters (Fig. 1.1). For comparison, a human hair is about
100 000 nm thick. Nanoscience is the study of atoms, molecules, and objects,
the size of which is on the nanometer scale (1 nm – 100 nm).

1 nm = 10−9 m = 10−3 µm = 10 Å. (1.1)

Figure 1.1. Comparison of sizes typical for nanotechnology. One nanometer is one
billionth of a meter.

Nanostructures typically refer to systems with length scale of about 1 nm to

about 100 nm in at least one dimension. The striking characteristics of these
systems is that electrons—which govern essentially its properties—are confined
in the nanoscale dimension(s) but are free to move in the other dimension(s).

4
Hence, one can classify nanostructures according to the number of nanoscale
dimensions.

Quantum well Electrons are confined in one dimension (1D) but allowed to
move in the other two dimensions. Quantum wells can be realized at
interfaces between different semiconductors (see the chapter on 2DEG)
or in ultrathin films on a substrate. Quantum wells can be regarded as
two-dimensional (2D) electronic systems (see Fig. 1.2).

Figure 1.2. Quantum well. In the transmission electron microscope (TEM) picture
the Si quantum well appears as bright area, surrounded by SiN (dark).

Quantum wire Electrons are confined in two dimensions (2D). Therefore, quan-
tum wires form 1D electronic systems. They can be realized by polymer
chains, carbon nanotubes, and atomic point contacts (see Fig. 1.3).

Figure 1.3. Quantum wire. A chain of atoms is formed between two semi-infinite
leads.

Quantum dot Electrons are confined in all spatial dimensions. Quantum dots
are zero-dimensional electronic systems (0D), sometimes called ‘artifi-
cal atoms’. They can be realized by clusters and nanocrystallites (see
Fig. 1.4).

5
 Figure 1.4. Quantum dots (nanocrystallites on a surface).

In all the labels above, the term ‘quantum’ is used to characterize the systems,
giving a hint on the quantum nature of the electronic structure (see Section 1.2).
As we will see, electronic states become quantized due to the reduced dimension
(with respect to the spatially unrestricted system).
Nanostructures have unique properties which neither show up in individual
atoms or molecules (which might form the nanostructure) nor in the macro-
scopic material.1 Lying between the ‘microscopic’ (atoms) and the ‘macro-
scopic’ world (bulk) they are termed ‘mesoscopic’ systems.

1.2 What makes nanostructures unique and interesting?

For bulk materials, the intrinsic properties are independent of their sizes. For
example, the conductivity of a copper wire of 1 m length does not change if its
length is reduced to, say, one-half. In this sense, we can speak of the ‘conductiv-
ity of Cu’, meaning that of bulk Cu. However, if the length-reduction process
is repeated again and again it is obvious that the conductivity must change
since at least the electronic properties of Cu atoms differ essentially from those
of bulk Cu. Typically, intrinsic properties begin to change significantly if a
dimension is reduced to the nanometer scale.2 At this length scale, ‘quantum
effects’ show up. The following phenomena critically affect the properties of
nanostructures.

1
The spatially unrestricted system is often called the ‘bulk’.
2
Note that one can speak of the ‘conductivity’ of Cu but not of the conductivity of a
nanostructure. Instead one needs to speak of its ‘conductance’ because this property
depends on the particular realization of the nanostructure. In other words, conductivity
is a property of a material, conductance that of a (nano)system.

6
Quantum confinement If electrons are confined in the nanoscale dimension,
their energy and momentum become quantized (discrete values). In addi-
tion, the dimensionality of the electronic states is reduced. One example
of quantum confinement are so-called quantum-well states (see Fig. 1.5).
Another example are surface-state electrons which become localized in a

Figure 1.5. Quantum confinement due to Pt nanowires on a surfaces (orange, indi-

cated by arrows). The surface electrons become caught between the Pt wires. Hence,
the scanning tunneling microscope (STM) image shows distinct oscillations instead of
structureless areas (violet).

‘cage’ built from adatoms on a surface (‘quantum corral’; see Fig. 1.6).

Figure 1.6. Quantum confinement in a quantum corral. In the STM image, the
surface-state electrons are localized in the circular area made-up from adatom (peaks),
showing concentric charge-density oscillations.

Quantum coherence Specific phase relations of the electronic wavefunctions

are preserved in a nanostructure, giving rise to interference effects if the
phase-coherence length is larger than the nanoscale dimension (Fig. 1.7).

7
 However, defects in the nanostructure could reduce the phase-coherence
length, forcing one to consider both phase-coherent and -decoherent pro-
cesses. This is in contrast to the limiting cases: in an atom, for example,
all processes are coherent, in a bulk material all are decoherent (due to
the per definitionem infinite extension of the system).

Figure 1.7. Quantum coherence in the current density flowing through a nanocon-
striction (STM image). The central part shows a theoretical result, the outer parts the
experimental results.

Surface and interface effects A significant fraction of atoms in a nanostruc-

ture is located near its boundaries. Since the properties of these atoms
may differ from that of the interior atoms, the properties of the nano-
structure may be considerably determined by these surface or interface
atoms.

These factors play roles to various degrees of importance. The mixture of mi-
croscopic (atomic), mesoscopic, and macroscopic character make the properties
of nanostructures vary dramatically. It is therefore obvious that the process of
fabrication essentially determines the observed properties, making it difficult
to extract specific information from the experimental samples and to compare
theoretical with experimental results. In turn, these features make nanoscience
challenging for the scientific community and allow for new insights and break-
throughs.

1.3 Historical remarks

The take-off of nanoscience happened in the early 1990ies, being a natural conse-
quence of the developments of science and technology. [1] [2] At that time, physi-
cists were able to produce experimentally nanostructures and to describe the-
oretically small systems (atoms, molecules, surfaces) and large systems (bulk).
Considering the former, one could either follow a bottom-up approach, assem-
bling nanostructures from individual atoms or molecules (e. g. by epitaxy), or
a top-down approach, fabricating nanostructures from macroscopic materials
(e. g. by lithography). [3–5]
The development of nanoscience was essentially put forward by inventions in
fabrication and characterization. In particular, the latter is important because

8
the samples have to be characterized as best as possible (Note that the observed
properties can vary drastically from sample to sample). Nanoscale probes,
besides other well-established methods (e. g. photoelectron spectroscopy) have
become conventional tools in the laboratory. Just to name a few, scanning
tunneling spectroscopy, atomic force microscopy, etc.
Focusing in the former on the technological side, one should not forget the
interest in fundamental aspects of nanoscience. Nanoscale probes allow the
investigation of basic effects in an unparalleled manner (e. g. Kondo effect of
adatoms on a surface investigated by means of an STM), [6] pushing forward
theoretical efforts. On the other hand, due to the increasing computer power
one can accurately describe mesoscopic systems theoretically in very detail.
However, for a basic understanding of the effects one needs a fundamental
theory of nanostructures. [7, 8] In this lecture, the important topic of electronic
transport in magnetic nanostructures will be sketched. [9] [10]

1.4 Scope of the lecture

This lecture is designed for undergraduate students to comprehend a theoretical
foundation of nanostructures. Having different scientific backgrounds, the very
basics of quantum mechanics and solid-state theory will be given. For a deeper
understanding, however, further studies in these areas are desirable but cannot
be provided in the course of this lecture. So, it should be your task to replace
believing by knowing (Fig. 1.8)

Figure 1.8. Believe? No, no. Know!

This goal can be accomplished by solving selected problems which are given
in this script. The intention of the problems is ‘learning by doing’. Some
problems open the view to more advanced topics, not relevant for this lecture
but possibly of interest for the curious reader.
Keywords are set in bold sans-serif.

9
2 Review of quantum mechanics
This chapter is neither intended to provide a thorough introduction to quantum
mechanics (QM) nor will it discuss the philosophical implications of QM. For
these purposes one should consult standard literature, for example Ref. [11].

2.1 Foundations

Particle-wave dualism
According to L. de Broglie quantum-mechanical objects can be regarded either
as waves or as particles.p[12–14] For example, for photons one has from the
relativistic energy E = (pc)2 + (mc2 )2 (p momentum, c speed of light) that
E = pc, since the mass m = 0. Then, from c = νλ and the Einstein relation
E = hν [15] it follows
c hc hc h
λ= = = = . (2.1)
ν hν E p
For a matter wave,
h h
p= = k = ~k, (2.2)
λ 2π
with the wavenumber k = 2π/λ and ~ ≡ h/2π.1 In three dimensions, one obtains
p = ~k. (2.3)
For electrons, the wavelength is in the range of a few Ångström (1 Å =
10−10 m = 0.1 nm). The wave nature of particles shows up as interference
and diffraction effects.2 For example, electron waves are used in the description
of low-energy-electron diffraction (LEED) at surfaces (pioneering experiments
by Davisson and Germer [16, 17]).
Uncertainty principle
The uncertainty principle (W. Heisenberg [18]) states that it is impossible to
measure simultaneously the momentum px and the position x of a particle
exactly. In contrast, the errors ∆px and ∆x obey
∆px ∆x ≥ ~. (2.4)
This statement is in agreement with the wave nature of particles. Similar
relations hold for quantities, the product of which has the dimension of an
action, for example energy E and time t (∆E ∆t ≥ ~).3
1
Planck’s constant h = 4.1357 10−15 eV s; ~ = 6.5822 10−16 eV s.
2
Does the mathematical description of electrons as waves imply that they are waves?
3
An even more strict relation goes back to Heisenberg, ∆px ∆x ≥ ~/2. [19]

10
 A wave packet is a superposition of waves with different momenta px . For
example, a single continuous wave has a defined momentum but is spread over
the entire space. In contrast, a ‘bang’ consists of all momenta but is sharply
localized in space. The uncertainty relation, eq. (2.4), is depicted for a wave
packet in Fig. 2.1.

Figure 2.1. Uncertainty principle.

The preceding suggests to formulate QM in terms of waves, in particular

by a wavefunction that completely describes the properties of the object. The
wavefunction is the solution of a wave equation, the Schrödinger equation.

2.2 Schrödinger equation

A wavefunction Ψ(r, t) is obtained by solving the time-dependent Schrödinger
equation [20–23]
 2 
p̂ ∂
+ V (r, t) Ψ(r, t) = i~ Ψ(r, t). (2.5)
2m ∂t

Here, the momentum operator p̂ is given in configurational-space representation

by

p̂ = −i~∇. (2.6)

Problem 1. In what respect differs a wave equation from the Schrödinger equation?
A wave equation in 1D can be written in the form

∂2 1 ∂2
Ψ(x, t) = Ψ(x, t),
∂x2 c2 ∂t2
where c is the propagation velocity. In 3D this reads

1 ∂2
∆Ψ(r, t) = Ψ(r, t),
c2 ∂t2
with ∆ = ∇ · ∇. Or written in compact form, 2Ψ(r, t) = 0, with the d’Alembert operator
2 ≡ ∆ − 1/c2 ∂t2 . 2

11
 If the potential does not depend on time [V (r, t) = V (r)], one has4
−iEt/~
Ψ(r, t) = e Ψ(r), (2.7)

where Ψ(r) obeys the time-independent Schrödinger equation

 2 
p̂
+ V (r) Ψ(r) = EΨ(r). (2.8)
2m

E is the (constant) energy of the particle. If V is real, the Hamilton operator

(or Hamiltonian for short)

p̂2
Ĥ = + V (r) (2.9)
2m
is hermitian.5 Hence, the eigenvalues E (energies) are real. Observables, that
are quantities the expectation value of which one can measure in a experiment,
are represented by hermitian operators (see Table. 2.1).

Observable Operator
Position r r
Momentum p̂ −i~∇
~2
Kinetic energy T̂ − 2m ∇2
Potential energy V V (r)
~2
Total energy Ĥ − 2m ∇2 + V (r)

Table 2.1. Translation between observables and operators in configurational-space

representation.

Problem 2. Show that the momentum operator p̂ is hermitian. For simplicity, assume the
one-dimensional case p̂x = −i∂x and Ψ(x).
It follows with integration by parts
Z Z
?
(p̂x Ψ(x)) Ψ(x) dx = (−i∂x Ψ(x))? Ψ(x) dx
Z
= i (∂x Ψ(x)? )Ψ(x) dx
Z
+∞
= iΨ(x) Ψ(x) |−∞ −i Ψ(x)? ∂x Ψ(x) dx
?

Z
= Ψ(x)? p̂x Ψ(x) dx.

Here we have used that a square-integrable wavefunction tends to zero for x → ±∞. 2
The (absolute) square of the wavefunction, i. e. the probability density,

ρ(r, t) = Ψ(r, t)? Ψ(r, t) = |Ψ(r, t)|2 , (2.10)

4

∂ ∂
i~ Ψ(r, t) = i~ e−iEt/~ Ψ(r) = EΨ(r, t).
∂t ∂t

5 ?
Ψ? (ĤΨ) dV for any Ψ.
R R
An operator Ĥ is hermitian if (ĤΨ) ΨdV =

12
is interpreted as the probability to find the particle in an infinitesimal volume
element around r at time t. Obviously, the probability to find the particle
anywhere in configuration space must equal 1,
Z
ρ(r, t) dr3 = 1. (2.11)

This imposes that the wavefunction must be square-integrable, Ψ(r, t) ∈

L2 [R3 ].6 A wavefunction that obeys eq. (2.11) is said to be normalized.
Defining the probability-current density as
~
j(r, t) ≡ −i [Ψ(r, t)? ∇Ψ(r, t) − Ψ(r, t)∇Ψ(r, t)? ] (2.12a)
2m
~
= Im [Ψ(r, t)? ∇Ψ(r, t)] , (2.12b)
m
the continuity equation

∂ρ(r, t)
+ ∇ · j(r, t) = 0 (2.13)
∂t
holds. It states that the temporal change of the probability density of a given
volume equals the current flowing through the surface of that volume.
Problem 3. Write the above statement in mathematical form, starting from eq. (2.13).
Compare the result with the continuity equation in electrodynamics.
We integrate eq. (2.13) over a volume V ,
Z Z
∂ρ(r, t)
dV + ∇ · j(r, t) dV = 0.
V ∂t V

Interchanging differentiation and integration in the first term yields

Z Z
∂ρ(r, t) ∂ ∂
dV = ρ(r, t) dV = P (V, t),
V ∂t ∂t V ∂t

where P (V, t) is the probability to find the particle at time t in the volume V . For the
second term we exploit Gauss’ theorem,
Z Z
∇ · u dV = u · dS,
V S(V )

where S(V ) is the surface of the volume V . Thus,

Z Z
∇ · j(r, t) dV = j(r, t) · dS = J(S(V ), t),
V S(V )

where J(S(V ), t) is the probability current flowing through S(V ) at time t. Eventually, we
obtain
∂
P (V, t) = J(S(V ), t).
∂t
The change of probability to find the particle in the volume V at time t equals the probability
current flowing through the volume’s surface S(V ).
6
R∞
A function is square-integrable in R if −∞
|f (x)|2 dx < ∞.

13
 In electrodynamics, the continuity relation reads
∂
∇·j =− ρ,
∂t
where j is the current density (of the charged particles) and ρ is the charge density. This
statement is complete equivalent to the QM case if our QM particles are charged. 2
A measurement determines the expectation value of an observable (repre-
sented by a hermitian operator Ô),
D E Z
Ô = Ψ(r)? ÔΨ(r)dr3 . (2.14)

If the set {Ψi } comprises all eigenfunctions of Ô, i. e.

ÔΨi = oi Ψi , (2.15)

one can expand any wavefunction with respect to this set,

X
Ψ= ci Ψi . (2.16)
i

Since the eigenfunctions are orthonormal (orthogonal and normalized), that is

Z
Ψi (r)? Ψj (r) dr3 = δij , (2.17)

one has
D E X Z
Ô = ci cj Ψi (r)? ÔΨj (r)dr3
?
(2.18a)
ij
X Z
= ?
ci cj Ψi (r)? oj Ψj (r)dr3 (2.18b)
ij
X
= oi |ci |2 . (2.18c)
i

Problem 4. Prove that the eigenstates |Ψi i and |Ψj i with different eigenvalues oi and oj
are orthonormal.
We have Ô|Ψi i = oi |Ψi i and Ô|Ψj i = oj |Ψj i with oi 6= oj . Thus,

hΨi |Ô|Ψj i = hΨi |oj |Ψj i = oj hΨi |Ψj i.

Analogously, for the hermitian Ô acting on |Ψi i

hΨi |Ô|Ψj i = oi hΨi |Ψj i,

which leads to

oj hΨi |Ψj i = oi hΨi |Ψj i

which can only be fulfilled if hΨi |Ψj i = 0. 2

Problem 5. Given a set of eigenstates {|Ψi i} of an operator Ô and a state |Ψi, derive an
expression for the expansion coefficients ci in eq. (2.16).

14
 P
We have Ô|Ψi i = oi |Ψi i and |Ψi = i ci |Ψi i. From the previous problem we know
that hΨi |Ψj i = 0 if i 6= j. Thus
X
hΨj |Ψi = ci hΨj |Ψi i = cj hΨj |Ψj i.
i

Hence,
hΨj |Ψi
cj = .
hΨj |Ψj i
2
Assume that we want to measure simultaneously two observables, Ô1 and
Ô2 ,
D E Z
Ô1 Ô2 = Ψ(r)? Ô1 Ô2 Ψ(r)dr3 . (2.19)

Obviously this implies

D E D E
Ô1 Ô2 = Ô2 Ô1 , (2.20)

and, thus,

Ô1 Ô2 = Ô2 Ô1 . (2.21)

If this relation holds, the observables are said to commute, i. e., they can be
measured simultaneously. If this relation does not hold, one cannot measure
them at the same time. The commutator is defined as

[Ô1 , Ô2 ] = Ô1 Ô2 − Ô2 Ô1 . (2.22)

Problem 6. Do x̂ and p̂x commute?

Using the chain rule of differentiation, one has

p̂x x̂Ψ(x) = −i~∂x (xΨ(x))

= −i~(∂x x)Ψ(x) − i~x(∂x Ψ(x))
= −i~Ψ(x) − i~x(∂x Ψ(x)).
= −i~Ψ(x) + x̂p̂x Ψ(x)).

Thus,

p̂x x̂Ψ(x) − x̂p̂x Ψ(x)) = −i~Ψ(x).

Since this holds for any wavefunction Ψ(x), we have in general

[p̂x , x̂] = −i~.

15
2.3 Dirac notation
The preceding section heavily relies on configurational space, R3 , which, how-
ever, might not be the appropriate space for specific QM problems (just to
name one: spin). Hence, it is desirable to achieve a higher level of abstraction
by using another notation. This notation, introduced by P. A. M. Dirac, [24] is
commonly used and does not rely on a specific space. Hence, it allows one to
choose that representation which is best suited for solving the problem.
From the above it is evident that we are dealing with a Hilbert space (a
vector space with an inner product [25]), the vectors of which are represented
by the wavefunctions. To be more specific, the wavefunction Ψ(r) is the r-
representation of the abstract Hilbert-space vector |Ψi,

hr|Ψi = Ψ(r), (2.23)

denoted as ‘ket’ vector. One can also have other representations, for example
the momentum representation or p-representation (Fig. 2.2),

hp|Ψi = Ψ(p). (2.24)

The complex conjugate vector is denoted as ‘bra’ vector,

Figure 2.2. Relation of the abstract Hilbert space to the specific Hilbert spaces.

hΨ|ri = Ψ(r)? , (2.25a)

?
hΨ|pi = Ψ(p) . (2.25b)

Integration over a complete representation equals to unity, i. e.

Z
|rihr| dr3 = 1, (2.26a)
Z
|pihp| dr3 = 1. (2.26b)

16
Hence, the normalization of the wavefunction becomes
Z
1 = Ψ(r)? Ψ(r) dr3 (2.27a)
Z
= hΨ|rihr|Ψidr3 (2.27b)

= hΨ|Ψi. (2.27c)

The Schrödinger equation reads in Dirac notation

Ĥ|Ψi = E|Ψi, (2.28)

where Ĥ is the Hamilton operator which in r-representation is given by eq. (2.9).

Its eigenvectors, for example

Ĥ|Ψi i = Ei |Ψi i (2.29a)

Ĥ|Ψj i = Ej |Ψj i (2.29b)

are orthonormal, i. e.
(
1 i=j
hΨi |Ψj i = δij = . (2.30)
0 i 6= j

Hence, the product of a bra and ket forms the scalar or ‘inner’ product in this
Hilbert space, hΨ|Φi ∈ C.
Problem 7. Consider the right-hand-side eigenstate |ΨR i of Ĥ|ΨR i = λ|ΨR i and the
left-hand-side eigenstate hΨL | of hΨL |Ĥ = µhΨL |. Derive a relation between them for
hermitian Ĥ. What follows for the eigenvalues? [26]
From hΨL |Ĥ = µhΨL | it follows

µ? |ΨL i = Ĥ † |ΨL i = Ĥ|ΨL i,

where † denotes hermitian conjugation. Comparing with Ĥ|ΨR i = λ|ΨR i shows that
the r.h.s. eigenfunctions are the complex-conjugate of the l.h.s. eigenfunctions, hx|ΨR i =
hx|ΨL i = hΨL |xi? .7 Since eigenvalues of hermitian operators are real (observables!), the
sets of the r.h.s. and the l.h.s. eigenvalues are identical. 2

The δ-function
The so-called δ-function [25] is implicitly defined by
Z ∞
δ(x − a) f (x) dx = f (a). (2.31)
−∞

Hence, it is evident that it is not a usual function but a generalized function

or distribution. An often used ‘cooking recipe’ is
(
0 x 6= a
δ(x − a) = . (2.32)
∞ x=a
7
Strictly speaking, the sets of l.h.s. and r.h.s. eigenstates are one-to-one.

17
Note that the δ-function has a meaning only in integration.8 So be aware of
what you are doing! Further,
1
δ(ax) = δ(x), a 6= 0, (2.33a)
|a|
1
δ(a2 − x2 ) = [δ(x + a) + δ(x − a)] , a > 0, (2.33b)
2a
X 1 d
δ(g(x)) = d
δ(x − xn ) with g(xn ) = 0, g(xn ) 6= 0.
n | dx g(xn )|
dx
(2.33c)
Of particular importance are the following representations (Fig. 2.3),
(
1/η −η/2 < x < η/2
δ(x) = (2.34a)
0 otherwise
1 η
= lim (2.34b)
η→0 π x2 + η 2
1 sin ηx
= lim (2.34c)
η→∞ π x
1 −x2
= lim √ e /4η . (2.34d)
η→+0 2 πη

The Dirac identity reads

Figure 2.3. Representations of the δ-function. Top-hat function (1), diffraction peak
(2), Lorentzian (3), and Gaussian (4), as defined in eq. (2.34)

 
1 1
lim =P − iπ δ(x), (2.35)
η→+0 x + iη x
8
Therefore, it is a generalized function.

18
where P indicates the Cauchy principal value.9
Fourier transformation
We define the Fourier transform of Φ, Ψ = F[Φ], by
Z
1
Ψ(k) ≡ √ e−ikx Φ(x) dx. (2.36)
2π
Obviously,
Z
1
Ψ(0) = √ Φ(x) dx. (2.37)
2π

The inverse Fourier transformation, Φ = F −1 [Ψ], is given by

Z
1
Φ(x) ≡ √ eikx Ψ(k) dk. (2.38)
2π

Proof of Φ = F −1 [F[Φ]]:
Z Z
1 0
F −1 [F[Φ]] = eikx e−ikx Φ(x0 ) dx0 dk (2.39a)
2π
Z Z
1 0
= eik(x−x ) dk Φ(x0 ) dx0 (2.39b)
2π
Z
= δ(x − x0 ) Φ(x0 ) dx0 (2.39c)

= Φ(x). 2 (2.39d)

Thus, we have another representation of the δ-function, namely

Z
0 1 0
δ(x − x ) = eik(x−x ) dk. (2.40)
2π
Note that there are various definitions of Fourier transformations which differ
in the sign in the exponential function and in the placement of the pre-factor.

Figure 2.4. Examples of Fourier transformations: continuous wave (a), finite-ranged

wave (wave-train, b), wave with triangular modulation (c).
hR R∞ i
9 c−η
If f (x) diverges in x = c, then P (f ) ≡ limη→0 −∞
f (x) dx + c+η
f (x) dx is called the
Cauchy principal value.

19
 If Φ(x) is represented in configuration space, Ψ(k) is its representation in
k-space or wavenumber space. We were not forced to use different characters
(Φ and Ψ) but used them for clarity.
Using Dirac notation we have
Z
hk|Ψi ≡ hk|xihx|Φi dx = hk|Φi, (2.41)

with
1
hx|ki ≡ √ eikx . (2.42)
2π
The above proof then reads
Z Z
−1
F [F[Φ]] = hx|kihk|x0 ihx0 |Φi dx0 dk (2.43a)
Z Z
= hx|kihk|x0 i dkhx0 |Φi dx0 (2.43b)
Z
= hx|x0 ihx0 |Φi dx0 (2.43c)

= hx|Φi.  (2.43d)
Therefore, δ(x − x0 ) = hx|x0 i.

2.4 Free particle in one dimension

For a free particle in 1D, the potential V vanishes, V (x) = 0. Thus, the
Schrödinger equation Ĥ|Φi = E|Φi reads in configurational space
~2 d2
− Φ(x) = E Φ(x). (2.44)
2m dx2
With the ansatz
Φ(x) = C eikx , (2.45)
that is a plane wave with propagation along the x-axis, we arrive at
~2 2
k Φ(x) = E Φ(x). (2.46)
2m
Thus, the energy is given by
~2 2
E(k) = k , (2.47)
2m
which is called the dispersion relation (Fig. 2.5). In particular, the wavefunction
Φ(x) is an eigenfunction of the momentum operator p̂,
p̂ Φ(x) = p Φ(x), (2.48a)
d
−i~ C eikx = p C eikx , (2.48b)
dx
~k C eikx = p C eikx , (2.48c)

20
 Figure 2.5. Dispersion relation of a free particle.

with eigenvalue p = ~k [which is just the de-Broglie relation, eq. (2.2)]. Thus,

p2
E= . (2.49)
2m
Since the wavenumber k characterizes the wavefunction, the latter should be
indicated by this parameter. Note that k can be positive or negative.
Now let us normalize the wavefunctions Φk (x). It follows
Z +∞
1≡ Φk (x)? Φk (x) dx (2.50a)
−∞
Z +∞
= |Ck | 2
e−ikx eikx dx (2.50b)
−∞
Z +∞
= |Ck |2 dx, (2.50c)
−∞

which is undefined. Therefore, one has to use another normalization, in this

case the so-called k-normalization,10
Z +∞
Φk (x)? Φk0 (x) dx = δ(k − k 0 ). (2.51)
−∞

With
Z
0 1 0
δ(k − k ) = eix(k−k ) dx, (2.52)
2π
10
Other normalizations used comprise the p- and the E-normalization. See Merzbacher, p. 86.

21
cf. eq.(2.40), one obtains
1
Ck = √ . (2.53)
2π
Summarizing,
1
Φk (x) = √ eikx , (2.54a)
2π
p = ~k, (2.54b)
(~k)2
E(k) = . (2.54c)
2m

Free particle in Dirac notation

From Ĥ|Φi = E|Φi we have

p̂2
|Φi = E|Φi. (2.55)
2m
So, we have to look for an eigenfunction of the momentum operator p̂. It follows

hx|p̂|pi = p hx|pi, (2.56a)

Z
hx|p̂|x0 ihx0 |pi dx0 = p hx|pi, (2.56b)
Z
d 0
−i~δ(x − x0 ) hx |pi dx0 = p hx|pi, (2.56c)
dx
d
−i~ hx|pi = p hx|pi. (2.56d)
dx
With the ansatz

hx|pi = C eikx (2.57)

the eigenvalues are p = ~k. Thus,

p2 (~k)2
E(k) = = . (2.58)
2m 2m
√
From k- or p-normalization it follows C = 1/ 2π because of δ(k − k 0 ) = hk|k 0 i.
Extrapolation to three dimensions
In three dimensions one obtains analogously
1
Φk (r) = eik·r , (2.59a)
(2π)3/2
p = ~k, (2.59b)
(~k)2
E(k) = , (2.59c)
2m
with k = (kx , ky , kz ).

22
2.5 Density of states
The density of states n(E) dE is the number of states in the energy interval
[E, E + dE]. Since the energy depends on the wavevector k, we can write
Z
g
n(E) = δ(E − E(k)) d3 k. (2.60)
(2π)3

Note that the volume element in k-space is (dk/2π)3 .11 g is the so-called degener-
acy and equals the number of states with energy E.12 Since dE is infinitesimal,
one has

E + dE = E + ∇k E(k) · dk. (2.61)

Thus, using eq. (2.33c)

Z
g 1
n(E) = 3
dSk , (2.62)
(2π) |∇k E(k)|
where the integral is over areas in k-space with constant energy E. Equa-
tion (2.62) provides a direct relation of the density of states (DOS) with the
dispersion relation.
Example: Free electrons in 1D
It follows with eq. (2.54) and
dk
dk = dE (2.63a)
dE !
r
d 2mE
= dE (2.63b)
dE ~2
r
1 2m
= dE (2.63c)
2 E~2
that
Z +∞
g
n(E) = δ(E − E(k)) dk (2.64a)
2π −∞
Z ∞ r
g 1 2m
=2 δ(E − E(k)) dE (2.64b)
2π 0 2 E~2
r
g 2m
= (2.64c)
2π E~2
r
2m ~2
=g 2
. (2.64d)
2π~ 2mE
Thus,
 − 1
2m 2mE 2 1
n(E) = g ∝√ . (2.65)
2π~2 ~2 E
11
In d dimensions: (dk/2π)d .
12
Typically, g = 2 due to the electron’s spin.

23
Alternatively, one can use eq. (2.62) and obtains
Z
g 1
n(E) = d
dSk (2.66a)
2π dk E(k)
Z
g 1
= dSk (2.66b)
2π ~2 k
2m
Z
2m 1
=g 2
dSk (2.66c)
2π~ |k|
 1
2m 2mE − 2

=g . (2.66d)
2π~2 ~2

Free electrons in 1D, 2D, and 3D

The DOS of free electrons in one, two, and three dimensions reads
−1 √
q
2mE
∝ 1/ E 1D


 ~ 2
2m 
n(E) = g 1 ∝ const. 2D , (2.67)
2π~2 
q √
 2mE

∝ E 3D
~2

as is depicted in Fig. 2.6. Note that for the cases 2D and 1D each state gives

Figure 2.6. Density of states of free electrons confined in various dimensions (0D –
3D).

an additional contribution to the DOS, resulting in the step-like shapes. For a

quantum dot (0D), the energy levels are discrete (‘artificial atom’) and, hence,
the DOS is a sum of δ-peaks (cf. the ‘cooking recipes’ for the δ-function).

24
2.6 Quantum well
For free electrons in 3D we had

hr|Ψi ∝ exp(ik · r) ∝ hr|ki, (2.68a)

1
E(k) = k2 , (2.68b)
2
p = k, (2.68c)

in Hartree atomic units (e = m = ~ = 1, c ≈ 137).

Problem 8. Look up the length and energy units in Hartree and Rydberg atomic units.
Why are they convenient for nanoscale problems?
Hartree atomic units are defined by setting e = ~ = m = 1, leading for the speed of
light to c ≈ 137.036.13 The unit of length is the Bohr radius,
~2
1 a0 = = 0.529177 Å,
me2
the energy unit is the Hartree,
~2
1H = = 27.2116 eV.
ma20

In these units, [energy] = [length−2 ]. One Rydberg is one half of a Hartree,

~2
1 Ry = = 13.6058 eV.
2ma20
Hartree or Rydberg atomic units are convenient because they can be regarded as the
natural unit system for atomic problems. Since properties of nanostructures are determined
by their atomic constituents, these systems are often used. 2
Now we confine these electrons in 1D by means of a potential wall. Thus,
it is convenient to use adapted coordinates, r = (x, y, z) = (ρ, z), ρ = (x, y).
Analogously, k = (kk , kz ). For the potential V , we assume with a > 0

+∞ z < −a

hr|V i = 0 −a ≤ z ≤ a . (2.69)

+∞ a < z


Due to the walls of infinite height, the electrons become confined in z, so we

are dealing with a quantum well.
The Hamiltonian can be decomposed as [27]

Ĥ = Ĥρ + Ĥz , (2.70)

with
1 d2 d2
 
Ĥρ = − + , (2.71a)
2 dx2 dy 2
1 d2
Ĥz = − + V (z). (2.71b)
2 dz 2
13 2
Note that the fine structure constant is α = e /~c. Hence, c = 1/α.

25
Hence, the wavefunction is expressed as (product ansatz )

hr|Ψi = hρ|φi hz|ψi, (2.72)

where

Ĥρ |φi = Eρ |φi, (2.73a)

Ĥz |ψi = Ez |ψi. (2.73b)

Thus, the Schrödinger equation Ĥ|Ψi = E|Ψi gives

 
Ĥρ + Ĥz hρ|φi hz|ψi = Ĥρ hρ|φi hz|ψi + Ĥz hρ|φi hz|ψi (2.74a)
= hz|ψiĤρ hρ|φi + hρ|φiĤz hz|ψi (2.74b)
= hz|ψiEρ hρ|φi + hρ|φiEz hz|ψi (2.74c)
= (Eρ + Ez ) hρ|φihz|ψi. (2.74d)
P
If the total Hamiltonian Ĥ is a sum of independent operators, Ĥ = i ĥi , the
total wavefunction Ψ is a product of the eigenfunctions ψi of these independent
operators, Ψ = Πi ψi . Therefore,

E = Eρ + Ez . (2.75)

From the previous section, we already know that

1
Eρ = k2k . (2.76)
2
Thus, it remains to determine Ez .
For −a ≤ z ≤ a (named domain II here), we have a free electron (V = 0).
Hence,

hz|ψi = A exp(ikz z) + B exp(−ikz z). (2.77)

Note
√ that the ansatz consists of two independent exponentials, with kz =
2Ez > 0.
For −∞ < z < −a (domain I) we first assume that the potential V is larger
than the energy (Ez < U ) and then take the limit U → ∞. Thus,

hz|ψi = C exp(κz), (2.78)

p
with κ = 2(U − Ez ) > 0. The other independent solution, hz|ψi ∝
C exp(−κz), cannot be normalized in this domain and, thus, does not show
up. For U → ∞, we have κ → ∞, leaving hz|ψi = 0 in I. The same argument
holds for a < z < ∞ (domain III), for hz|ψi = D exp(−κz).
To determine A and B, eq.(2.77), we exploit the continuity of the wavefunc-
tion in the entire domain, i. e.

h−a|ψi = 0, (2.79a)
ha|ψi = 0. (2.79b)

26
These boundary conditions impose

0 = A exp(ikz a) + B exp(−ikz a), (2.80a)

0 = A exp(−ikz a) + B exp(ikz a), (2.80b)

which has non-trivial solutions for A and B only if the determinant of the
coefficients vanishes. Therefore,

0 = exp(2ikz a) − exp(−2ikz a). (2.81)

With Euler’s equation,

exp(iα) = cos α + i sin α, (2.82)

we arrive at

0 = sin(2kz a), (2.83)

or

2kz a = nπ, n 6= 0 integer. (2.84)

Eventually,14
π
kz(n) = n, n 6= 0 integer. (2.85)
2a
The boundary conditions on the wavefunction impose a restriction on the
wavenumbers of the electrons in the quantum well: the energy becomes quan-
tized (discrete; Fig. 2.7),

1 2 π 2 n2
Ez(n) = kz(n) = . (2.86)
2 2 a2
n is called a quantum number. It characterizes the associated solution of the
Schrödinger equation.
Problem 9. The energy of the ground state (n = 1) is larger than zero. Why? What
happens for the energy if the width of the quantum well is increased (a → ∞)? Make a
figure.
First, the ground-state energy cannot be negative because the wavefunction must be
identically zero in this case (normalization!). This means that there is no particle.
The particle is strictly confined to the quantum well. Thus, the uncertainty ∆z in position
is less than 2a, the width of the well. Heisenberg’s uncertainty relation gives ∆z ∆pz ≥ ~,
(1)
leading to ∆pz > ~/∆z. With pz = ~kz = ~π/2a we have ∆pz ∆z = π~ ≥ ~. If
the energy of the particle tends to zero, also its momentum must decrease, consequently
violating the uncertainty relation. Hence, the energy must be positive.
If the width of the quantum well is increased (a → ∞), the energy levels decrease with
a−2 . For a well with barriers of finite height, the energy levels are lower than for a well
with infinitely high barriers, due to increased ∆z (Fig. 2.8). Figure 2.9 shows the energy
of levels of quantum-well states that are confined to a Cu film grown on a Co substrate. 2
14
The case n = 0 has to be excluded since the wavefunction vanishes in all domains and,
hence, the probability density is zero (no electron).

27
Figure 2.7. Energies of a quantum well with extension from 0 to Lx and infinitely
high walls. The quantum numbers n are given on the right of each level.

The potential is symmetric with respect to reflection at the xy-plane (M̂ :

z → −z). Hence, the wavefunction must be even or odd with respect to this
symmetry transformation.15 In other words,

hz|ψi → M̂ hz|ψi = h−z|ψi = ±hz|ψi. (2.87)

This implies either A = B (case +, n odd) or A = −B (case −, n even). Thus,

(
cos n πz


(n) 2a n odd
hz|ψ i ∝ πz

(2.88)
sin n 2a n even

(Fig. 2.10). The normalization follows from

Z +a
hψ|zihz|ψi dz = 1. (2.89)
−a

Problem 10. Determine the coefficients A and B.

According to equation (2.88), the wavefunctions for even n read
 πz 
hz|Ψ(n) i = A(n) sin n .
2a
The pre-factor follows from the normalization,
Z +a  πz 
1= |A(n) |2 sin2 n dz.
−a 2a

Using
Z
1 1
sin2 (bz) dz = z− sin(2bz)
2 4b
15
This follows from the fact that upon performing two subsequent reflections the wavefunction
must be unchanged. Explicitly, M̂ |ψi = m|ψi; M̂ 2 |ψi = m2 |ψi = |ψi; m2 = 1; m = +1
(even) or m = −1 (odd). m is called the parity of the wavefunction.

28
Figure 2.8. Energy levels of an electron confined to quantum well versus thickness
a. The latter is given in layers, a = N d, d being the interlayer distance. Empty circles
indicate levels for a quantum well with infinitely high barriers. Solid lines connect
levels with the same quantum number n. Dashed lines connect levels with identical
n − N . Filled circles indicate levels for a well with finite barriers. [27]

[28, integral Nr. 275], it follows with b = nπ/2a

1
A(n) = √ .
a
The same results is obtained for odd n using [28, integral Nr. 314]. 2
In a quantum well for which the potential in domains I and III is finite, the
wavefunctions consist of decaying exponentials (cf. the above argumentation of
letting U → ∞) to be matched continuously to the oscillating wavefunction
in II (Fig. 2.11). In this case, the number of bound states is finite (here: 3;
Ez < U = 300 meV). For Ez > U , the electrons are free to move towards
±∞ (propagating states). Bound states can be defined as those states the
probability density of which vanishes for z → ±∞. For propagating states this
does not hold.
Problem 11. Solve the quantum-well problem for a finite potential in domains I and III.
According to the argument given on page 28 we can make the ansatz

κz
A e
 z < −a
hz|ψi = B cos(kz) −a ≤ z < a .
 −κz

Ae z≤a
p √
for even wavefunctions. Here, κ = 2(U − E) > 0 and k = 2E, U being the potential
in domains I and III. The continuity of the wavefunction at z = −a gives
A e−κa = B cos(ka),
the continuity of its z-derivative yields
κ A e−κa = k B sin(ka).

29
Figure 2.9. Energy levels of quantum-well states in Cu films on Co(100), as obtained
by photoelectron spectroscopy for kk = 0. [27]

Figure 2.10. Energies, wavefunctions (left), and probability densities (right) of a

quantum well with extension 10 nm and walls of infinite height. The five lowest energy
levels are shown. Note the symmetry and the number of nodes of the wavefunctions.

Note these two conditions are equivalent to demanding continuity of the logarithmic deriva-
tive of the wavefunction.16 By division one obtains the transcendental equation
κ = k tan(ka),
which can be solved numerically (e. g. by fix-point methods). An approximate solution is
obtained from the Taylor series of tan(ka),
1
tan(x) ≈ x + x3 + · · · .
3
For odd wavefunctions, one arrives at

κz
A e
 z < −a
hz|ψi = B sin(kz) −a ≤ z < a ,
−A e−κz

z≤a


16
The logarithmic derivative of a function f (x) is given (ln f (x))0 = f 0 (x)/f (x), with the
prime indicating differentiation with respect to x.

30
Figure 2.11. Energies, wavefunctions (left), and probability densities (right) of a
quantum well with extension 10 nm and walls of finite height (300 meV). The first
three lowest energy levels are shown (bound states).

which results in

κ = k cot(ka).

2
Note that the solution of a QM problem requires to (i) solve the Schrödinger
equation and (ii) to impose appropriate boundary conditions.
Quantum wire
Following the ideas presented in the preceding, we now turn to a quantum
wire, i. e. electrons confined in two dimensions (here: x and y) but allowed to
propagate in one dimension (here: along the z-axis). With the dimensions lx
and ly of the ‘box’, this leads to the energy

E = Ex + Ey + Ez , (2.90)

with
π 2 n2
Ex(n) = , (2.91a)
2 lx2
π 2 m2
Ey(m) = , (2.91b)
2 ly2
1
Ez = kz2 . (2.91c)
2
The wavefunction is characterized by the three quantum numbers n, m, and
kz and consists of a product of the associated partial wavefunctions, i. e. wave-
functions of the form given by eq. (2.88) for hx|φi and hy|ψi and a plane wave
for hz|χi. The total wavefunction is a product of these partial waves.
Quantum dot
An electron in a box with dimensions lx , ly , and lz has the energy

E = Ex + Ey + Ez , (2.92)

31
with
π 2 n2
Ex(n) = , (2.93a)
2 lx2
π 2 m2
Ey(m) = , (2.93b)
2 ly2
π 2 o2
Ez(o) = . (2.93c)
2 lz2
For lx = ly = lz , the state with the lowest energy, the so-called ground-state
(n = m = o = 1), is not degenerate. The second level, however, is threefold
degenerate because the combinations (n = m = 1, o = 2), (o = n = 1, m = 2),
and (m = o = 1, n = 2) lead to the same energy.
Problem 12. Consider a 2D quantum dot with equal length lx and ly . Sketch the
probability density of the ground state and the first excited state.
The wavefunction of the ground state is given by (lx = ly = 2a; n = m = 1).
1  πx   πy 
Ψ(x, y)(1,1) = cos cos .
a 2a 2a
The associated probability density is shown in Fig. 2.12.

Figure 2.12. Probability density of the ground state of a 2D quantum dot. Blue
(yellow) indicates small (large) density. The square quantum dot extends from −1 to
+1 in both x- and y-direction

The first excited state is two-fold degenerate, that is (n, m) = (1, 2) and (2,1) have
identical energies. One wavefunction reads
1  πx   πy 
Ψ(x, y)(1,2) = cos sin ,
a 2a a
the other is obtained by interchanging x and y. The associated density is shown in Fig. 2.13.
2

32
Figure 2.13. Probability density of the excited state with n = 1 and m = 2 of a 2D
quantum dot. Colors as in Fig. 2.12.

2.7 Tunneling
In the preceding sections we have come across a ‘particle in a box’. Now we turn
to the problem of scattering, in particular to the transport through a tunnel
barrier. For this, so-called scattering boundary conditions have to be imposed.
Further, we can decompose the 3D problem in to a 2D (free electrons) and a 1D
problem (tunneling electrons), in analogy to the quantum-well problem. The
2D problem was already solved, the 1D problem is being solved now. [11]
In potential scattering, one is concerned with an incoming particle that is
scattered at a potential into an outgoing particle. Incoming means that the
particle is impinging onto a potential V (r) from infinity (r → ∞) and is outgo-
ing into infinity (r → ∞). Hence, incoming and outgoing states are eigenstates
of the Hamiltonian at r → ∞. Hence, the scattering potential must vanish for
r → ∞.17
As for the quantum well (section 2.6), three domains are introduced. The
potential is given by (U > 0)

0 z < −a

hz|V i = U −a ≤ z ≤ a . (2.94)

0 a<z


We consider the case 0 < E < U . In domain II, the wavefunction can be
written as

hz|ψi = C e+κz + D e−κz , −a ≤ z ≤ a, (2.95)

17
Strictly speaking, the potential must be regular, that is limr→0 r2 V (r) = 0 [29, p. 47].

33
 p
with κ = 2(U − E) > 0. The solutions in domains I and III are already
known, giving

ikz −ikz z < −a
A e + B e

hz|ψi = C e−κz + D eκz , −a ≤ z ≤ a , (2.96)

F e + G e−ikz
 ikz
a<z
√
with k = 2E, E > 0. The above ansatz is the most general. The total
wavefunction comprises two incident waves—one incident from the left (with
amplitude A), the other incident from the right (G)—and two outgoing waves
(B and F ).
The continuity of the total wavefunction and of its derivative with respect to
z at z = −a gives

A e−ika + B eika = C eκa + D e−κa , (2.97a)

iκ
A e−ika − B eika = C eκa − D e−κa .


(2.97b)
k
Or in matrix form,
iκ
eκa+ika iκ
e−κa+ika
  

 
A 1 1+ k
1− k
C
= . (2.98)
B 2 1− iκ
k eκa−ika 1+ iκ
k e−κa−ika D

Similarly, we find at z = a

1 − ik 1 + ik
  
κa+ika
κa−ika   
C 1 κ e κ
e F
= −κa+ika −κa−ika . (2.99)
2 1 + ik 1 − ik


D κ e κ e G

Combining these results, we arrive at

    
A M11 M12 F
= , (2.100)
B M21 M22 G

with
iη
cosh(2κa) + i2 sinh(2κa) e2ika


sinh(2κa)
2
M≡
− iη2 sinh(2κa) cosh(2κa) − i2 sinh(2κa) e−2ika

(2.101)

and
κ k
≡ − , (2.102a)
k κ
κ k
η≡ + . (2.102b)
k κ
An equivalent representation relates the incoming waves with the outgoing ones,
    
F S11 S12 A
= . (2.103)
B S21 S22 G

34
S is called the scattering matrix. It is unitary (S † = S −1 ) and typically used
for 3D problems.
Problem 13. Express the scattering matrix S in terms of the matrix M .
Equation (2.100) gives
A = M11 F + M12 G,
B = M21 F + M22 G.
Hence, from the first equation follows
−1 −1 −1
F = M11 (A − M12 G) = M11 A − M11 M12 G.

Inserting into the second equation yields

−1 −1 −1
B = M21 M11 (A − M12 G) + M22 G = M21 M11 A + [M22 − M21 M11 M12 ]G.

Comparison with the definition of S gives then

−1
S11 = M11 ,
−1
S12 = −M11 M12 ,
−1
S21 = M21 M11 ,
−1
S22 = M22 − M21 M11 M12 .

2
Due to the independence of time (stationary problem), the current density is
conserved [cf. the continuity equation (2.13)]. In particular that on the left of
the barrier equals that on the right,

j = k |A|2 − |B|2 = k |F |2 − |G|2 ,

(2.104)

using eq. (2.12).

Problem 14. Compute the current density, eq. (2.12), in each of the three domains.
In domain I the wavefunction reads

hz|ψi = A eikz + B e−ikz .

The current density j is then

d
j = Im [hψ|zi hz|ψi]
dz
= Im [ik(A? e−ikz + B ? eikz )(A eikz − B e−ikz )]
= Im [ik(|A|2 − |B|2 − A? B e−2ikz + AB ? e2ikz )]
= Im [ik(|A|2 − |B|2 + 2iIm [AB ? e2ikz ])]
= Im [ik(|A|2 − |B|2 )]
= k(|A|2 − |B|2 ).
The total current density is the sum of those flowing to the right (with amplitude |A|2 )
and flowing to the left (|B|2 ).
For domain III an analogous calculation gives

j = k(|F |2 − |G|2 ).

In domain II one has with the wavefunction

hz|ψi = C e−κz + D eκz

35
for the current density
j = κ Im [(C ? e−κz + D? eκz )(−C e−κz + D eκz )]
= κ Im [|D|2 e2κz − |C|2 e−2κz + C ? D − D? C]
= 2κ Im (C ? D).
Since the current density is conserved (that is identical in all domains), these results
establish another relation between the coefficients A, . . . , F . 2
Now we impose scattering boundary conditions which mean that a particle
is incident from one side (here: from the left, with amplitude A) and reflected
(with amplitude B). The transmitted wave is outgoing to the right (with am-
plitude F ). Hence, G = 0 since there is no incoming wave from the right (for a
sketch see Fig. 2.14).

Figure 2.14. Total wavefunction of an electron incident from the left and scattered
from a rectangular potential barrier with height U and width L (schematic).

The transmission T , defined as the outgoing current density per incoming

current density, then reads with eq. (2.100)
k|F |2
T ≡ (2.105a)
k|A|2
= |M11 |−2 (2.105b)
−2
i
= cosh(2κa) + sinh(2κa) (2.105c)
2
1
=
2 . (2.105d)
[cosh(2κa)] + 4 k − κk [sinh(2κa)]2
2 1 κ

Analogously, the reflection R reads

k|B|2
R≡ (2.106a)
k|A|2
2
M21
= (2.106b)
M11
k 2
κ 1
[sinh(2κa)]2


k + κ
4
=
2 . (2.106c)
[cosh(2κa)]2 + 14 κk − κk [sinh(2κa)]2

36
The current conservation implies R + T = 1. Further note that the case E > U
can be obtained by replacing κ by iκ.
The transmission T is depicted in Fig. 2.15. The striking contrast to a clas-

Figure 2.15. Transmission T versus energy E for electrons scattering from a rectan-
gular potential barrier with height U = 1 eV. The red curve corresponds to a barrier
width of 1 nm, the blue curve to 0.5 nm.

sical particle is that for E < U the QM particle is transmitted to the other
side of the barrier (with probability T ). The classical particle, however, is to-
tally reflected. In a pictorial description, the QM particle ‘tunnels’ through the
barrier. Hence, this effect is called ‘tunnel effect’ (Fig. 2.16).

Figure 2.16. Tunnel effect. Particle versus wave picture (left) and classical versus
quantum-mechanical scattering (right).

Problem 15. Why must the scattering matrix S be unitary?

A matrix S is unitary if S † = S −1 . Thus, 1 = S · S −1 = S · S † . Or explicitly,
  ? ?
  
S11 S12 S11 S21 1 0
? ? = .
S21 S22 S12 S22 0 1

The unitarity of S gives several relations between its elements. With the determinant
S11 S22 − S12 S21 = 1, one has
 ? ?
  
S11 S21 S22 −S12
? ? = ,
S12 S22 −S21 S11

37
which establishes
1 = |S11 |2 + |S12 |2 ,
1 = |S21 |2 + |S22 |2 ,
? ?
0 = S11 S21 + S12 S22 .
Considering the first equation, one has

1 = |M11 |−2 + |M11 |−2 |M12 |2 = T + R,

that is conservation of probability for an incoming particle from the left.

The inverse of the matrix M is
 
1 M22 −M12
M −1 = ,
d −M21 M11
with the determinant d ≡ M11 M22 − M12 M22 . Define the transmission and reflection for
a particle incoming from the right by (in analogy to T and R)
|d|2
T̃ ≡ ,
|M11 |2
|M12 |2
R̃ ≡ .
|M11 |2

Then 1 = T̃ + R̃ gives

|M11 |2 = |d|2 + |M12 |2 .

With M11 S22 = d and −M11 S12 = M12 is follows

|M11 |2 = |M11 |2 |S22 |2 + |M11 |2 |S12 |2 ,

and thus

|S22 |2 + |S21 |2 = 1,

which is the second equation (using |S21 |2 = |S12 |2 ). Thus, also probability conservation
for a particle incoming from the right holds.
The third equation can be used to show

T̃ R̃ = T R.

2
For U < 0 a classical particle is not reflected. In contrast, a QM particle is
partially reflected even for 0 < E (Fig. 2.17). In both cases, U < 0 and U > 0,
the transmission shows resonances.
Problem 16. Why do resonances show up? What might be their origin?
There are no resonances, that are maxima of the transmission
p (T = 1), for 0 < E < U .
For E > U , the wavenumber in domain II is then k 0 = 2(E − U ). It can be shown
that the maxima show up at 2k 0 a = nπ, n integer.
The de-Broglie wavelength of the particle in domain II is
2π 4a
λ= 0
= .
k n
Thus, twice the width of the potential barrier (2a) is an integer number of the wavelength.
If this is the case, the wave in domain II that is reflected at the two interfaces at z = ±a

38
Figure 2.17. Transmission T versus energy E for electrons scattering from a rectan-
gular potential barrier with height U = −1 eV. Lines as in Fig. 2.15.

is in phase, resulting in a reinforcement (constructive interference) and consequently in

perfect transmission. Resonances prove again the wave nature of quantum-mechanical
particles. 2
Problem 17. Find examples for the tunnel effect.
Just to name a few: scanning tunneling microscope, field emission, alpha decay of nuclei.
2

2.8 Summary
This brief review should remind the audience to the following key aspects of
quantum mechanics: Hamilton operator, Schrödinger equation, wavefunction,
state, Hilbert space, eigenvalue, expectation value, boundary condition, quanti-
zation, symmetry, normalization, degeneracy, quantum number, Dirac notation,
representation, wave-particle dualism, tunnel effect.

39
3 Review of solid-state theory
In this chapter the basics of solid-state theory are given. For more complete
description, see for example [30]. First, an introduction to lattices is given,
then the electronic structure is discussed by means of the single-particle picture.
Subsequently, the many-electron problem is addressed briefly.

3.1 Lattice structures

Consider three linearly independent vectors a1 , a2 , and a3 .1 From these we
can construct points R in R3 by linear combination,

R = n1 a1 + n2 a2 + n3 a3 , n1 , n2 , n3 integer. (3.1)

Each R is uniquely specified by the triple (n1 , n2 , n3 ). The set L = {R} of

these vectors is called a Bravais lattice. Obviously, any lattice vector R can
be written as a sum of two lattice vectors, R = R0 + R00 , with R, R0 , R00 ∈ L.
Further, if R ∈ L, then −R ∈ L. Thus, (L, +) is a group.2
The simple cubic lattice (sc) has basis vectors
     
a 0 0
a1 = 0 , a2 = a , a3 = 0 , (3.2)
0 0 a

as shown in Fig. 3.1. Here, a is the lattice constant. The face-centered cubic

Figure 3.1. Three cubic lattices: simple cubic (left, sc), face-centered cubic (center,
fcc), and body-centered cubic (right, bcc).

1
Linearly independent means that the volume spanned by these vectors is nonzero, a1 · (a2 ×
a3 ) 6= 0.
2
A group (L, +) is defined as a set (L) and an operation (+) acting on this set with (i)
∀a, b ∈ L : c = a+b ∈ L, (ii) ∃0 ∈ L : a+0 = a ∀a ∈ L, and (iii) ∀a ∈ L ∃−a : a+(−a) = 0.
See T. Inui, Y. Tanabe, and Y. Onodera, Group Theory and Its Applications in Physics
(Springer, Berlin, 1990).

40
lattice (fcc) has
     
1 0 1
a  a  a 
a1 = 1 , a2 = 1 , a3 = 0 , (3.3)
2 2 2
0 1 1

whereas for the body-centered cubic lattice (bcc)

     
−1 1 1
a  a  a 
a1 = 1 , a2 = −1 , a3 = 1 , (3.4)
2 2 2
1 1 −1

The defining vectors a1 , a2 , and a3 are the basis vectors of the lattice.
They span the unit cell. The entire space can be covered by translating the
unit cell by a lattice vector R (‘stacking of unit cells’). Thus, the unit cells in
conjunction with the lattice vectors (regarded as translation vectors) provide a
complete and disjunct coverage of R3 .
Both the bcc and the fcc lattices can be regarded as a sc lattice with a basis.
For the bcc case, the fundamental lattice is sc but attached to each lattice site
R are atomic positions given by
   
0 1
a 
τ1 = 0 ,τ2 =
  1 . (3.5)
2
0 1

Evidently, two atoms belong to each sc unit cell. In the fcc case, four atoms
positioned at
       
0 1 1 0
a a a
τ 1 = 0 , τ 2 = 1 , τ 3 = 0 , τ 4 = 1 (3.6)
2 2 2
0 0 1 1

belong to each sc cell. Note that the definition of lattice and basis vectors may
differ from publication to publication.
Problem 18. Make sketches of the sc, bcc, and the fcc lattices (pen and paper). 2
Besides the translation group, there is the point group which for a fixed
lattice point R comprises all rotations and reflections that leave the lattice
invariant. Whereas the above introduced primitive unit cell is not invariant
under point-group operations, another kind of unit cells—the Wigner-Seitz
cells—are. These are constructed as follows.

1. Choose any lattice site as the origin.

2. Starting at the origin draw vectors to all neighboring lattice points.

3. Construct a plane perpendicular to and passing through the midpoint of

each vector.

4. The area enclosed by these planes is the Wigner-Seitz cell.

The Wigner-Seitz cells have the following characteristics:

41
 • Wigner-Seitz cells associated with all lattice points are identical in size,
shape and orientation as follows with the translational symmetry of the
lattice.

• When stacked the Wigner-Seitz cells fill all space.

• The Wigner-Seitz cell is a polyhedron.

• The Wigner-Seitz cell has the full point symmetry of its lattice point.

Figure 3.2. Wigner-Seitz cells of the body-centered (left) and the simple cubic (right)
lattices.

Note that a crystal is specified by the lattice and the basis.

2D lattices
For 2D systems, translational invariance in one dimension is lost, as compared
to 3D systems. However, each layer has the same lattice being spanned by a1
and a2 . For example, a square lattice has (Fig. 3.3)
   
a 0
a1 = , a2 = , (3.7)
0 a

a hexagonal lattice
   
a √1 a √−1
a1 = , a2 = . (3.8)
2 3 2 3

Problem 19. Sketch a square lattice and try to find all symmetry operations of this lattice.
Without translations, construct the multiplication table (point group). Try to identify the
point group in textbooks on group theory. [31]
The point group of the square lattice consists of the following elements (operations):
1 (identity operation), C4 (clockwise rotation about π/2), C2 (rotation about π), C4−1
(rotation about −π/2), mx (reflection at the x axis), my (reflection at the y axis), mx=y
(reflection at the x = y axis), mx=−y (reflection at the x = −y axis).
This point group is called C4v in Schoenflies notation or 4mm [31, Table B.4, page 367].
It is the point group of (001) surfaces of bcc and fcc crystals. The multiplication table is
given in Table 3.1. Each row and each column contains each element once.

42
 Figure 3.3. Two-dimensional lattices: square (left) and hexagonal (right) lattice.

Figure 3.4. Construction of the Wigner-Seitz cell of the 2D hexagonal lattice (cf. the
instruction on p. 41).

1 C4 C2 C4−1 mx my mx=y mx=−y

1 1 C4 C2 C4−1 mx my mx=y mx=−y
C4 C4 C2 C4−1 1 mx=y mx=−y my mx
C2 C2 C4−1 1 C4 my mx mx=−y mx=y
C4−1 C4−1 1 C4 C2 mx=−y mx=y mx my
mx mx mx=−y my mx=y 1 C2 C4−1 C4
my my mx=y mx mx=−y C2 1 C4 C4−1
mx=y mx=y mx mx=−y my C4 C4−1 1 C2
mx=−y mx=−y my mx=y mx C4−1 C4 C2 1

Table 3.1. Multiplication table of 4mm.

The rotations form an abelian subgroup.3 Two subsequent reflections are equivalent to
a single rotation. Two subsequent reflection do not commute. 2
Problem 20. Construct the Wigner-Seitz cells of the square and the hexagonal lattice
(pen and paper). 2

Reciprocal lattice
For each lattice in configurations space, one can define a reciprocal lattice L?
(or dual lattice) which is very important in solid-state theory. The basis vectors
b1 , b2 , and b3 of the reciprocal lattice obey

ai · bj = 2πδij , i, j, = 1, 2, 3. (3.9)

3
A subgroup is a subset of group elements which itself forms a group. Abelian means that
the operations commute, leading to a symmetric multiplication table.

43
A solution of the above equations is given by
aj × ak
bi = 2π , i, j, k cyclic. (3.10)
a1 · (a2 × a3 )

Obviously, the volume of the reciprocal unit cell obeys V (b1 , b2 , b3 ) =

(2π)3/V (a1 ,a2 ,a3 ). The reciprocal lattice is a Bravais lattice. Further, the re-
ciprocal lattice of the fcc lattice is bcc and vice versa (Fig. 3.5), whereas as the
sc lattice is ‘self-dual’ (except for a change of scale). The reciprocal lattice of a
reciprocal lattice is the original lattice.

Figure 3.5. Wigner-Seitz cells of the fcc (left) and the bcc (right) lattice.

An analogous definition holds for the 2D case. Examples are given in Fig. 3.6.

Figure 3.6. Two-dimensional reciprocal lattices: square (left) and hexagonal (right)
lattice.

Problem 21. Derive an analogous solution of eq. (3.9) for the 2D case. What is the
meaning of the reciprocal lattice vectors bi in both the 2D and the 3D case?
The area A spanned by a1 and a2 is given by

A = a1x a2y − a1y a2x .

44
A vector normal to a1 is obtained as (−a1y , a1x ). Hence, we can define
2π
b1 ≡ (a2y , −a2x ),
A
2π
b2 ≡ (a1y , −a1x ).
A
A vector product in 3D has the meaning of a directed area.4 It is a pseudovector because
it does not change sign upon space inversion. Thus, the bi can be viewed as normalized
areas; they do change sign under inversion. In 2D, the bi change also sign. 2
The importance of the reciprocal lattice can be evidenced by considering the
phase factor

eiG·R = 1, (3.11)

since with G ≡ m1 b1 + m2 b2 + m3 b3 the scalar product G · R is an integer

multiple of 2π.
The Wigner-Seitz cell of the reciprocal lattice is called the (first) Brillouin
zone (BZ). The center of the BZ is typically called Γ. The names of other
points, see Fig. 3.7, are also commonly used.

Figure 3.7. First Brillouin zone of the fcc lattice.

Lattice planes are typically specified by their Miller indices h, k, l (integer): A

plane with indices h, k, l is normal to the reciprocal lattice vector hb1 +kb2 +lb3 .

Problem 22. Construct the lines with Miller indices (0 1), (1 1), (2 1), and (3 1) for the
square lattice (pen and paper). How are they related to the respective (h k 0) planes of the
sc lattice? 2

4
The meaning of vectors and their operations is much more clear using Geometric Algebra.
[32, 33]

45
 Figure 3.8. Selected lattice planes and their Miller indices of the sc lattice.

3.2 Electronic structure

Translation operator
An essential property of lattices is their translational invariance. Hence, before
investigating their electronic structure in more detail, we look at the translation
operator T̂a . It shifts a wavefunction by a given vector a.5

T̂a hr|ψi = hr + a|ψi. (3.12)

Now we determine its eigenvalues λa ,

T̂a hr|ψi = λa hr|ψi. (3.13)

Two subsequent translations, by a and b, are identical to a single translation

by a + b. Hence,

T̂a+b = T̂a T̂b , (3.14)

and

λa+b = λa λb . (3.15)

The latter equation is just the functional equation of the exponential function.6
With b = −a, equation (3.15) gives
1
λa = , (3.16)
λ−a
since λ0 = 1.
5
a must not be a lattice vector.
6
That is exp(a + b) = exp(a) exp(b).

46
 From the normalization of an eigenfunction hr|ψi of T̂a we obtain
Z
1= hr|ψi? hr|ψi dr3 (3.17a)
Z
= hr + a|ψi? hr + a|ψi dr3 (3.17b)
Z
= (T̂a hr|ψi)? T̂a hr|ψi dr3 (3.17c)
Z
2
= |λa | hr|ψi? hr|ψi dr3 . (3.17d)

Thus, |λa |2 = 1. Combining all the above results, we can express the eigenvalues
as

λa = eik·a . (3.18)

Remembering the eigenfunctions of the momentum operator (free particles; see

page 20 and equation (2.59)),
1
hr|pi = eik·r , (3.19)
(2π)3/2
with p = ~k, one sees immediately that these are eigenfunctions of the trans-
lation operator,
1
T̂a hr|pi = eik·(r+a) (3.20a)
(2π)3/2
1
= eik·a eik·r (3.20b)
(2π)3/2
1
= λa eik·r (3.20c)
(2π)3/2
= λa hr|pi. (3.20d)

The above results are a consequence of Noether’s theorem which states that
a symmetry of the Hamilton operator7 corresponds to a conserved quantity. In
the present case, the Hamiltonian (of the free particle) is translational invariant
for any vector a. And hence, the momentum p is a ‘good quantum number’
that is, it is conserved. Another example is a time-invariant Hamiltonian (that
of the time-independent Schrödinger equation) which leads to the conservation
of energy. Or rotational invariance of an atom leads to conservation of angular
momentum, giving rise to ‘good quantum numbers’ l and m. Or for a symmetric
quantum well (page 28), the parity (‘even’ or ‘odd’) of its wavefunctions is
conserved.
The eigenfunctions of the respective operator can be labeled by the respective
quantum number, as is done in the above example, for the eigenfunctions of the
momentum operator p̂, hr|pi.
A lattice, however, is not fully translational invariant, that is, it cannot be
shifted by any vector a but only by lattice vectors R. As we will see in the
7
The proof of the theorem uses Lagrangians.

47
following, this implies conservation of momentum up to a reciprocal lattice
vector G.
Bloch states
If the potential V of a crystal is periodic, hr|V i = hr + R|V i for all R ∈ L, it
is said to be translationally invariant. Or in terms of the translation operator
T̂R : T̂R hr|V i = hr + R|V i = hr|V i for all R ∈ L. Since the kinetic-energy
operator p̂2 /(2m) is (fully) translational invariant, also the Hamiltonian Ĥ is
translationally invariant. The set {T̂R |R ∈ L} forms an abelian group, the
translation group of the lattice.
We are now looking for wavefunctions that are eigenfunctions of all T̂R . From
the preceding, it is evident that these look like

hr|Ψk i ∝ eik·r ,

with eigenvalues exp(ik · R) for T̂R . But one can multiply these with any func-
tion hr|uk i that is translational invariant for all R ∈ L, T̂R hr|uk i = hr|uk i∀R ∈
L. Hence,

hr|Ψk i = eik·r hr|uk i. (3.21)

Note that in the preceding section on fully translational invariance hr|uk i = 1.

Now consider the wavefunction

hr|Ψk+G i = ei(k+G)·r hr|uk+G i, (3.22)

for a reciprocal lattice vector G ∈ L? . For any R ∈ L it follows

T̂R hr|Ψk+G i = ei(k+G)·(r+R) hr + R|uk+G i (3.23a)

i(k+G)·R i(k+G)·r
=e e hr|uk+G i (3.23b)
= eik·R ei(k+G)·r hr|uk+G i (3.23c)
ik·R
=e hr|Ψk+G i. (3.23d)

Thus, hr|Ψk+G i has the same eigenvalue for T̂R as hr|Ψk i, namely exp(ik · R).
In other words, these wavefunctions are degenerate. But all eigenstates of an
abelian group are single degenerate.8 Hence, hr|Ψk+G i = hr|Ψk i.
These findings imply that the momentum p = ~k of an electron in a crystal is
conserved except for reciprocal lattice vectors, ~G. Thus, one has to distinguish
the momentum and the so-called crystal momentum. The former is unique,
the latter is not.
In summary, the eigenstates of the Hamiltonian [cf. eq. (2.9)] fulfill Floquet’s
theorem: for all translation vectors R ∈ L one has

hr + R|Ψi = λR hr|Ψi. (3.24)

Normalization [Ψ ∈ L2 (R3 )] requires |λR | = 1 ∀ R ∈ L. Thus, we can write

hr + R|Ψi = eik·R hr|Ψi, (3.25)

8
The representations of an abelian group have dimension one. [31]

48
which is called the Bloch theorem. Hence, any electronic state can be char-
acterized by a reciprocal vector k. A state which fulfills Bloch’s theorem is
called a Bloch state. It can be written as the product of a phase factor and a
lattice-periodic part hr|uk i,

hr|Ψk i = eik·r hr|uk i, (3.26)

with hr|uk i = hr + R|uk i ∀ R ∈ L. The Bloch theorem implies that

|Ψk i = |Ψk+G i (3.27)

for any lattice vector G of the reciprocal lattice. [34]

Nearly-free electrons
A periodic potential can be expanded into a Fourier series,
X
hr|V i = VG eiG·r , (3.28)
G∈L?

with Fourier coefficients VG . Hence, it appears natural to expand the n-th

eigenfunction of Ĥ also into plane waves,
(n)
X (n)
hr|Ψk i ≡ ψG (k) hr|k + Gi. (3.29)
G

Inserting this ansatz into the Schrödinger equation (in Hartree atomic units)
(n) (n)
Ĥ|Ψk i = En |Ψk i (3.30)

yields with for simplicity dropped k and n

!
1 2 X X 0 X 0
− ∇ + VG eiG·r ψG0 ei(k+G )·r = E ψG0 ei(k+G )·r . (3.31)
2 0 0
G G G

We now multiply with

00
e−i(k+G )·r

and integrate over the whole space. Then the right-hand side reads
Z Z
−i(k+G00 )·r i(k+G0 )·r 0 00
X X
E ψG0 e e 3
dr = E ψG0 ei(G −G )·r dr3
G0 G0
(3.32a)
X
3
= (2π) E ψG0 δG0 ,G00 (3.32b)
0
G
3
= (2π) EψG00 , (3.32c)

where we have used eq. (2.52). The same calculation yields for the kinetic-
energy part
1
(2π)3 (k + G00 )2 ψG00 . (3.33)
2

49
The potential part gives
Z Z
X 00 0 X 0 00
VG ψG0 e−i(k+G )·r ei(k+G+G )·r dr3 = VG ψG0 ei(G+G −G )·r dr3
GG0 GG0
(3.34a)
X
= (2π)3 VG ψG0 δG+G0 ,G00
GG0
(3.34b)
X
= (2π)3 VG00 −G0 ψG0 .
0
G
(3.34c)
Thus, the Schrödinger equation reads
1 X
(k + G00 )2 ψG00 + VG00 −G0 ψG0 = EψG00 . (3.35)
2 0
G
It can be cast into a matrix equation (secular equation),
H(k)ψn (k) = E (n) (k)ψn (k). (3.36)
Here, ψn (k) comprises all Fourier coefficients of the wavefunction. The elements
of the Hamilton matrix read in Hartree atomic units
1
HGG0 (k) = (k + G)2 δG,G0 + VG−G0 . (3.37)
2
The above equation involves in principle an infinite number of reciprocal lattice
vectors and, hence, cannot be solved in general. However, typically a few recip-
rocal lattice vectors are sufficient to describe the potential accurately enough.
Problem 23. For a real potential V , what follows its Fourier coefficients? What follows
further if V is centro-symmetric, that is hr|V i = h−r|V i?
?
For a real potential one has hr|V i = hV |ri, or V (r) = V (r) , for all r. Thus,
X X X
VG? e−iG·r = VG eiG·r = V−G e−iG·r .
G∈L? G∈L? −G∈L?

Since the exponential functions are linearly independent, the above equality is fulfilled for
each term. Hence,
VG? = V−G .
From hr|V i = h−r|V i it follows with
X X X
VG e−iG·r = VG eiG·r = V−G e−iG·r
G∈L? G∈L? −G∈L?

that
VG = V−G .
In summary,
?
VG = V−G = V−G .
The potential can then be expressed as (using Euler’s equation)
X
hr|V i = 2 VG cos(G · r).
G∈L?
2

50
The empty lattice. For hr|V i ≡ 0 an artificial lattice periodicity is introduced,
the so-called empty lattice. The Bloch states are single plane waves,

|Ψ(n) i = |k + Gn i, (3.38)

where the band index n is one-to-one to G.9 The dispersion relation is given
by
1
E (n) (k) = (k + Gn )2 , (3.39)
2
i. e. that of free electrons but shifted by −Gn in k-space. Note that
E (0) (−Gn /2) is degenerate with E (n) (−Gn /2).
As an example, we consider the 1D case and restrict the sum over G to
−g, 0, +g,

hz|V i = V−g e−igz + V0 + Vg eigz (3.40)

where g is the smallest nonzero lattice vector. Since the potential is real and
periodic, one has Vg = V−g ? real.10 Thus, with the convenient choice V0 = 0

1
hz|V i = Vg cos(gz). (3.41)
2
The secular equation reads explicitly
1 2
   
2 (k − g) Vg 0 ψ−g ψ−g
1 2
 Vg 2k Vg   ψ0  = E(k)  ψ0  . (3.42)
1 2
0 Vg 2 (k + g) ψg ψg

At k = ±g/2 band gaps show up, with width of about |2Vg |. The potential
is considered weak if the Fourier coefficients are small compared to the band
width. Here, the band width is the energy range which is occupied by a band.
Hence, the name nearly free electrons.
Problem 24. Set-up and solve the Schrödinger equation (3.42) for the two reciprocal
lattice vectors 0 and g (2 × 2 matrix). Determine the width of the lowest band and the
band gap.
The Schrödinger equations reads
1 2    
2k Vg ψ0
= E(k)
ψ0
,
Vg 12 (k + g)2 ψg ψg
or
k 2 − 2E(k)
  
2Vg ψ0
= 0.
2Vg (k + g)2 − 2E(k) ψg

The determinant of this matrix must vanish. Hence,

0 = [k 2 − 2E(k)][(k + g)2 − 2E(k)] − 4Vg2 .

9
That means that there is a one-to-one correspondence between each band index and each
reciprocal lattice vector.
10
This holds also for the coefficients in eq. (3.28).

51
And further
1 1
0 = E(k)2 − E(k)[k 2 + (k + g)2 ] + k 2 (k + g)2 − Vg2 .
2 4
From this follows
r
1 2 2 1 2 1
E(k) = [k + (k + g) ] ± [k + (k + g)2 ]2 − k 2 (k + g)2 + Vg2 .
4 16 4
Eventually,
1 2 1q 2
E(k) = [k + (k + g)2 ] ± [k − (k + g)2 ]2 + 4Vg2 .
4 4
The resulting band structure is shown in Fig. 3.9 For nonzero Vg a band gaps opens at

Figure 3.9. Band structure in the two-band model for Vg = 0 (black, solid) and
Vg = 0.2 (blue, dashed). g = 1.

k = −g/2. q
The lower band has the minimum energy at k = 0 with value [g 2 − g 4 + 4Vg2 ]/4. Its
maximum is g 2 /8 − |Vg |/2 at k = −g/2. The band width is given by twice the square root
at k = −g/2, that is |Vg |. 2
Note that band gaps are produced by the potential (Fig. 3.10). The Fourier
coefficients can be obtained by fitting the NFE band structure to those of
sophisticated methods or to experimental data.11 A typical application involves
optical data of zinc-blende semiconductors.
11
In the latter case, the NFE approach is called an empirical local pseudopotential method
(EPM).

52
Figure 3.10. Origin of band gaps. (left) Dispersion relation of a free electron (a single
parabola). (center-left) Empty lattice, restricted to two reciprocal lattice vectors. Note
the degeneracy at a single k-point. (center right) The degeneracy is lifted due to the
potential. (right) Same, but in the extended zone scheme.

Since E(k) = E(k + G) (repeated-zone scheme; Fig. 3.11) it is sufficient to

restrict oneself to the first BZ (reduced-zone scheme). Sometimes the extended-
zone scheme is used.

Figure 3.11. Zone schemes.

Problem 25. Show that E(k) = E(k + G) if all reciprocal lattice vectors are considered.
What is the effect of a finite set of G on the band structure?
Bloch states are eigenstates of the Hamiltonian,

Ĥ|Ψk i = E(k)|Ψk i.

Since

|Ψk i = |Ψk+G i,

it follows with

Ĥ|Ψk+G i = E(k + G)|Ψk+G i = E(k + G)|Ψk i,

which proofs the statement.

If a finite subset of reciprocal lattice vectors G is taken into account, the above statement
will no longer hold. The band structure will be described well only in the ‘interior’ of the
subset. To be more explicit, consider the subset {−2G, −G, 0, G, 2G}. The bands are well
described for k ≈ 0 but not at k ≈ 2G. 2

Tight-binding model
In the NFE model the potential is considered weak, the band structures are
similar to that of free electrons. An opposite approach would consider a solid

53
as made up from an ensemble of atoms which interact weakly. The electrons
remain tightly bound to the lattice sites, in contrast to nearly-free electrons.
We start from atomic orbitals hr − R|φn i where the index n comprises the
set of atomic quantum numbers (principal number, angular momentum, etc.;
Fig. 3.12). Hence, the tight-binding method is sometimes called linear combi-
nation of atomic orbitals method (LCAO).12 The atom is located at the lattice
site R.13 A Bloch function is constructed as
X
hr|Ψn i ≡ eik·R hr − R|φn i. (3.43)
R

Figure 3.12. Sketches of atomic orbitals: p-orbitals (l = 1; top) and d-orbitals (l = 2;

bottom).

Problem 26. Prove that this state fulfills the Bloch theorem.

12
Sometimes different types of orbitals are used, for example, Gaussian orbitals. The common
feature, however, is that the basis orbitals are localized (located at a lattice site) and that
the basis set does not depend on the energy.
13
For simplicity, a lattice without basis is considered.

54
 It follows
X 0 0
hr + R0 |Ψn i = eik·(R+R −R ) hr − R − R0 |φn i
R
0 X 0
= e−ik·R eik·(R+R ) hr − R − R0 |φn i
R
−ik·R0
X
= e eik·R hr − R|φn i
R
−ik·R0
= e hr|Ψn i.

2
The crystal potential V is given by a sum over atomic potentials W ,
X
hr|V i = hr − R|W i. (3.44)
R

Inserting these into the Schrödinger equation gives the following types of matrix
elements, for which without loss of generality one atomic orbital is located at
site 0.
Overlap integrals
Snm (k, R) ≡ eik·R hφn0 |φmR i. (3.45)
Atomic orbitals located at the same site are orthogonal, whereas those
located at different sites are not. In other words, the atomic orbitals form
a non-orthogonal basis. However, one can transform this basis into an
orthogonal one. [35] In particular in empirical tight-binding approaches,
orthogonality is assumed a priori.
Two-center integrals The two-center integrals are like overlap-integrals but
involve the Hamilton operator
p̂2
Hnm (k, R) ≡ eik·R hφn0 | + W (0)|φmR i. (3.46)
2
The site-diagonal two-center integrals (R = 0) give approximately the
atomic levels.
Three-center integrals Whereas the two-center integrals involve a potential of
one site of the orbitals (0 or R), the three-center integrals comprise a
potential at R0 6= 0, R,
eik·R hφn0 |W (R0 )|φmR i. (3.47)
Since the potential decreases rapidly in space, these integrals are typically
small and, hence, are often neglected.
In principle, one should consider the infinite set of overlap and two-center
integrals, that is, between the origin 0 and all lattice points R. Thus, the
Hamiltonian matrix comprises elements of the form
X
Hnm (k) = Hnm (k, R). (3.48)
R

55
However, due to the decreasing potential, one restricts the R-sum to nearest-
neighbor shells (1NN, 2NN, 3NN, . . . ). The Schrödinger equation then takes
the form of a generalized eigenvalue problem,
H(k) c(k) = E(k) S(k) c(k), (3.49)
where the Hamilton matrix H(k) consists of all orbital matrix elements Hnm (k)
and S(k) is the overlap matrix. The eigenfunctions |Φi (k)i with eigenvalues
Ei (k) are expressed by the Bloch functions,
X
|Φi (k)i = c(i)
n (k)|Ψn (k)i, (3.50)
n

with coefficients ci (k).

Problem 27. Derive eq. (3.49).
Consider the right-hand side. A Bloch function is given by
X
hr|Ψn i = eik·R hr − R|φn i.
R

Hence,
X X
|Φi (k)i = c(i)
n (k) eik·R hr − R|φn i.
n R

Multiplying with hφm |r − 0i and integrating over all space gives

X X X X
hφm0 |Φi (k)i = c(i)
n (k) eik·R hφm0|φnR i = cn(i) (k) Snm (k, R).
n R n R

Defining
X
Snm (k) ≡ Smn (k, R),
R

we arrive at
X
hφm0 |Φi (k)i = Smn (k)c(i)
n (k).
n

Analogously, the left-hand side reads

X
hφm0 |Ĥ|Φi (k)i = Hmn (k)cn(i) (k),
n

which together with the first results establishes in matrix form eq. (3.49). 2

Empirical tight-binding method. Instead of calculating the Hamilton and

overlap matrix elements one can regard them as parameters, the values of which
are determined by fitting the band structure to known data (e. g. band gaps,
critical points, band widths). For example, a matrix element of the Hamiltonian
then can be written as
X
Hnm (k) = nn δnm + t1NN (R)eik·R , (3.51)
R∈1NN

considering only the shell of nearest neighbors (1NN). The terms t1NN (R) can
be further decomposed with regard to the angular-momentum decomposition of
the involved orbitals. This introduces new parameters labeled with the bonding
type (σ, π, δ, etc. [36]; Fig. 3.13).

56
 Figure 3.13. Types of overlap integrals, (ssσ), (ppσ), and (ppπ).

1D TB model. As an example, we solve the 1D TB-model with one orbital

per site.14 The Hamiltonian can be written in the orbital basis as
X X
Ĥ = |nihn| + (|n − 1ithn| + |n + 1ithn|) , (3.52)
n n

where n labels the sites, being located at na (a lattice constant). Note that the
Dirac equation emphasizes that Ĥ is an operator.  are the on-site energies, t
the so-called hopping matrix elements. The first t-term describes hopping from
site n to site n − 1, the second hopping from n to n + 1 (Fig. 3.14).

Figure 3.14. Tight-binding model for a linear chain of atoms.

In matrix notation one arrives as15

 
..
. t
 
 t  t 
 
H=  t  t .
 (3.53)

 t  t  
..
t .

For the wavefunction we make the ansatz

X
|Ψi = eikma |mi. (3.54)
m
14
See, for example, Refs. [27, 37, 38]
15
The tridiagonal form of the matrix is often used, to speed up numerical computations.

57
and arrive at the dispersion relation

E(k) =  + 2t cos(ka). (3.55)

The first BZ is [−π/a, π/a].

Problem 28. Using the orthonormality of the orbitals, hn|mi = δnm , derive eq. (3.55)
from eqs. (3.52) and (3.54).
It follows
X X
Ĥ|Ψi = |nihn|mieikma + (|n − 1ithn|mi + |n + 1ithn|mi) eikma
nm nm
X X
ikna
|n − 1iteikna + |n + 1iteikna


= |nie +
n n
X X 
ikna
= |nie + |niteik(n+1)a + |niteik(n−1)a
n n
ika −ika

X
=  + te + te |nieikna
n
= ( + 2t cos(ka)) |Ψi.

Hence, E(k) =  + 2t cos(ka). 2

Problem 29. Write down a Hamiltonian with 2NN hopping terms t0 in both operator and
matrix form. How would the dispersion relation be changed? Make a sketch of E(k).
In operator form it follows
X X X
Ĥ = |nihn| + (|n − 1ithn| + |n + 1ithn|) + (|n − 2it0 hn| + |n + 2it0 hn|) ,
n n n

which results to the matrix form

 
.. 0
. t t
0
 
 t
 0  t t

0 

H= t t  t t .
0


 t t  t 

0 . .
t t .

The band structure is given by

E(k) =  + 2t cos(ka) + 2t0 cos(2ka).

The 2NN give rise to a modulation of the 1NN band structure with half the wavelength
(Fig. 3.15). 2

Chain of finite length. Restricting the matrix in eq. (3.53) to a finite dimen-
sion N , one has discrete eigenvalues

E(ki ) =  + 2t cos(ki a), (3.56)

with
πi
ki = , i = 1, . . . N. (3.57)
a(N + 1)
As in a quantum well for free electrons, the eigenenergies become quantized
(Fig. 3.16).

58
Figure 3.15. TB band structure without (blue, dashed) and with (black, solid) 2NN
interaction ( = 0, t = 1, t0 = 0.5, a = 1).

Figure 3.16. Tight-binding model for a finite linear chain of atoms. (a) Energies of
chains with n sites. (b) and (c) Density of states (DOS) and dispersion for an infinite
chain.

Problem 30. Compare the quantized wavenumbers for a finite chain in the 1D-TB model
with those of a 1D free-electron quantum well, eq. (2.85).
For the TB chain we had
πi
ki = , i = 1, . . . N,
a(N + 1)
for the free-electron quantum well with infinite walls
π
kz(n) = n, n 6= 0 integer.
2a
Or in terms of the wavelength,
2π 2a(N + 1)
λi = =
ki i
and
4a
λ(n) = .
n

59
Note that the latter is just the path length for constructive interference (cf. the origin of
resonances in tunneling). Hence, in both cases eigenstates show up at wavelength that fit
perfectly into the chain or well, leading to constructive interference. 2

Resolvent and Green function. The resolvent of the Hamiltonian is implicitly

defined as

1 = (z − Ĥ)G(z) = G(z)(z − Ĥ), (3.58)

for any complex number z. Note that G(z) is not the inverse of (z − Ĥ)
because G(z) is singular if z is an eigenvalue of Ĥ. The Green function is the
representation of the resolvent in configuration space,

G(x, x0 ; z) = hx|G(z)|x0 i. (3.59)

Its spectral representation is given by

X hx|ψi ihψi |x0 i
hx|G(z)|x0 i = , (3.60)
z − Ei
i

|ψi i being an eigenstate of Ĥ with eigenvalue Ei .

Problem 31. Show that the spectral representation fulfills eq. (3.58).
|ψi i is an eigenstate of the Hamiltonian Ĥ with energy Ei , Ĥ|ψi i = Ei |ψi i. Thus,
X hx|(z − Ĥ)|ψi ihψi |x0 i
hx|(z − Ĥ)G(z)|x0 i =
i
z − Ei
X hx|(z − Ei )|ψi ihψi |x0 i
=
i
z − Ei
X hx|ψi ihψi |x0 i
= (z − Ei )
i
z − Ei
X
= hx|ψi ihψi |x0 i
i
= hx|x0 i = hx|1|x0 i = δ(x − x0 ).

Here, we have used that the eigenstates of Ĥ form a complete basis of the Hilbert space,
that is
X
|ψi ihψi | = 1.
i

A basis is complete if any state vector in the Hilbert space can be expressed in terms of
the basis.
In the preceding we have used the spatial representation. But we have also could proved
the equality in the abstract Hilbert space, as follows
X |ψi ihψi | X |ψi ihψi |
(z − Ĥ)G(z) = (z − Ĥ) = (z − Ei )
i
z − Ei i
z − Ei
X
= |ψi ihψi | = 1.
i

60
 The Green function allows to compute the expectation value of any operator
(cf. eq. (2.14)),
D E 1
Ô = − Im Tr (ÔG). (3.61)
π
The spectral function A(x, x0 ; z) is defined as the side limit of the imaginary
part of the Green function,
1
A(x, x0 ; E) ≡ − lim Im G(x, x0 ; E + iη). (3.62)
π η→+0
From the spectral function one can compute the density of states (DOS),
Z
N (E) = Tr A(x, x; E) = A(x, x; E) dx, (3.63)

(Tr is called operator trace) and the charge or probability density,

Z EF
ρ(x) = A(x, x, ; E) dE. (3.64)
−∞

At band edges, the DOS can become infinite (Fig. 3.16). These singularities
are called Van-Hove singularities.
Problem 32. Derive an expression for the spectral function using the spectral representa-
tion of the Green function, eq. (3.60). Hint: Dirac identity, eq. (2.35). What is the effect
of a nonzero imaginary part of the (complex) energy z? Express the spectral function
A(x, x; E) in terms of Green functions for E ± iη.
First notice that
X  hx|ψi ihψi |x0 i ?
0 ?
G(x, x ; z) =
i
z − Ei
X hx0 |ψi ihψi |xi
= ?−E
= G(x, x0 ; z ? ).
i
z i

The side limits G± of the Green function are defined as

G± (x, x0 ; E) ≡ lim G± (x, x0 ; E ± iη).
η→+0

Thus,
G+ (x, x0 ; E) = G− (x, x0 ; E)? ,
0
+
2i Im G (x, x ; E) = G+ (x, x0 ; E) − G− (x, x0 ; E).
We now use the Dirac identity
 
1 1
lim =P − iπ δ(x),
η→+0 x + iη x
to show that
1
A(x, x0 ; E) = − Im G+ (x, x0 ; E)
π
1 X 1
= − Im hx|ψi ihψi |x0 i lim
π η→+0 E + iη − Ei
i
1 X
= − Im hx|ψi ihψi |x0 i[−iπδ(E − Ei )]
π i
X
= hx|ψi ihψi |x0 iδ(E − Ei ).
i

61
The principal value in the Dirac identity gives no contribution. With this result one has
further
1  +
A(x, x0 ; E) = − G (x, x0 ; E) − G− (x, x0 ; E) .

2πi
The density of states N (E) is given by
XZ
N (E) = Tr A(x, x; E) = hx|ψi ihψi |xi dxδ(E − Ei )
i
X X
= hψi |ψi iδ(E − Ei ) = δ(E − Ei ).
i i

With the representation of the δ-function we have

Xη 1
N (E) = lim 2 2
.
η→+0
i
π (E − E i) + η

So, instead taking the limit η → +0, one computes typically

Xη 1
N (E + iη) ≡
i
π (E − Ei )2 + η 2

for finite η. The δ-peaks at the eigenenergies become smeared out by a Lorentzian distri-
bution. 2
In equation (3.64), EF is the so-called Fermi energy. Electronic states with
energies E ≤ EF are occupied, those with E > EF are not (see page 77).

Typical features of band structures. The notion band structure E(k) implies
translational invariance in directions along the wavevector k is defined. This
means further that we are considering a system of infinite extension in these
directions (i. e., no confinement). In other words, band structure means bulk
system.
For confined electrons, e. g. as in quantum well or in an atom, the energies
are quantized. There is no continuous dispersion relation E(k). Hence, we
should not speak of a band structure but of the electronic structure in general.
From the TB model it is obvious that the band width is determined by the
hopping energy t. The latter can be attributed to the overlap of orbitals that are
located at neighboring sites.16 Further, the overlap depends on the character
of the orbitals. For example, s- and p-orbitals are comparably extended and
lead, thus, to a large overlap, and consequently to a strong dispersion (large
band width). Typical bands that are mainly derived from these orbital (the
‘sp-bands’) exhibit an almost parabolic dispersion, like for free electrons (cf.
the band with about −0.1 Ryd minimum energy in Fig. 3.17).
Bands derived from d-orbitals show much less dispersion (‘flat bands’) due to
stronger localization (smaller spatial extension) of these orbitals as compared to
that of s- and p-orbitals (cf. the bands in the energy from 0.2 Ryd to 0.5 Ryd).
In a specific energy range, bands of different character can hybridize, that is
the associated Bloch state possess a mixed character (‘s − p − d’ for example).17
16
Confer the Extended Hückel Theory (EHT). [39]
17
Hybridization can show up only between Bloch states that belong to the same representa-
tion.

62
Figure 3.17. Band structure of Cu. The numbers denote the representation of the
associated Bloch states. The horizontal line indicates the Fermi energy. The Brillouin
zone is depicted in Fig. 3.7.

3.3 Surfaces and interfaces

A surface or an interface breaks the 3D translational invariance of the 3D lattice.
Assuming without loss of generality that the 2D translational invariance in
the xy-plane is maintained, only the in-plane wavevector kk = (kx , ky ) is a
‘good quantum number’. For the z-direction, usually kz cannot characterize
the wavefunctions. So Bloch states are indexed by kk = (kx , ky ) only.
Superlattices.
Sometimes the periodicity of the surface differs from that of the substrate: the
surface is reconstructed. For example, the sites of atoms deposited on a surface
form a 2D lattice, the unit cell of which is larger than that of the substrate.
These atoms are said to form a superlattice or an overlayer. The basis vectors
A1 and A2 of the superstructure can be related to that of the substrate a1 and
a2 by

A1 = n11 a1 + n12 a2 (3.65)

A2 = n21 a1 + n22 a2 . (3.66)

Depending on the values nij , three categories can be distinguished.

Simply related All nij are integer.

Coincident Rational nij are present.

Incoherent Irrational nij are present. There is no common periodicity of sub-

strate and superstructure.

Instead of using the coefficients nij for characterizing the superstructure, one
often uses length relations between the basis vector sets and an azimuthal angle,
accompanied by the Miller indices
√ √of the unreconstructed surfaces. Examples
would be Pt(110)-1×2 and ( 3 × 3)R30 ◦ -Bi/Ag(111).

63
Figure 3.18. Overlayers on fcc(001) surfaces (blue atoms): unreconstructed 1×1
surface (left), c(2×2) (center), and p(2×2) (right). The right inset shows also a c(4×2)
structure in the right half.

Two types of overlayer on an fcc(001) surface are depicted in Fig. 3.18. The
p-type stands for ‘parallel’, the c-type for ‘centered’.
Problem 33. Make a 2D sketch of the fcc(001) surface (top view). Identify unit cells of
the 1×1, p(2×2), and the c(2×2) superstructures. How many sites belong the unit cells
of reconstructed surfaces?
The surface unit cells can be identified from Fig. 3.18. The number of atoms per unit
cell is 1, 2, and 4 for the 1×1, c(2×2), and the p(2×2)√case,√respectively. 2
Problem 34. Identify unit cells of the 1×1 and the ( 3 × 3)R30 ◦ superstructures of
fcc(111) (Fig. 3.19). The point group is 3m in the unreconstructed case. What is the
point group of the reconstructed surface?
The unit cells are depicted in the lecture slides. The point group is also 3m in the
reconstructed case. 2

√ √
Figure 3.19. Top view of the ( 3 × 3)R30 ◦ -Bi/Ag(111) surface, with top-layer
atoms large, second-layer atoms medium-sized, third-layer atoms small. Bi in blue, Ag
in red.

Complex band structure. The 2D translational invariance allows only for

Bloch states compatible with this periodicity. Hence, Bloch states can be writ-
ten as
hρ; z|ψkk i = eikk ·ρ hρ; z|ukk i, (3.67)

64
with

hρ + Rk ; z|ukk i = hρ; z|ukk i, (3.68)

kk = (kx , ky ), and ρ = (x, y).

For the 3D system we have ruled out those Bloch states with |λ| = 6 1, due to
normalization. The occurrence of a boundary (surface or interface), however,
allows to normalize the states which decay in direction apart from the boundary.
Writing λ as exp(ik⊥ z), with complex wavenumber k⊥ , implies
(
>0 z>0
Im k⊥ (3.69)
<0 z<0

if the boundary is located at z = 0. This analytic continuation of the real

bandstructure (real wavenumbers) is called the complex bandstructure. [40]
The important point is that there exist decaying solutions of the Schrödinger
equations in regions where no Bloch states exist, that is in band gaps. This
feature provides a direct link to surface or interface states since these are also
located in a band gap.
Problem 35. The dispersion relation of an infinite 1D chain is given by E(k) = +2t cos ka,
eq. (3.55). Sketch the complex band structure for Re k = 0 and Re k = π/a [41].
The analytic continuation of the dispersion relation is obtained by allowing for complex
wavenumbers k, k = Re k + i Im k = kr + iki . This gives

E(k) =  + t eika + e−ika

 
=  + t ei(kr +iki )a + e−i(kr +iki )a
=  + t eikr a e−ki a + e−ikr a eki a .

For kr = 0 one has

E(k) =  + t e−ki a + eki a =  + 2t cosh(ki a).

For kr = π/a one has

E(k) =  − 2t cosh(ki a).

The cosinus hyperbolicus is shown in Fig. 3.20. 2

Surface states
The existence of a surface can lead to electronic states, being localized at the
surface. [42] In the following two basic types of surface state formation are
discussed.

1D NFE model. We start with the crystal potential [43]

V (z) = −V0 + 2Vg cos gz, (3.70)

with the smallest reciprocal lattice vector g = 2π/a. For the wavefunction, we
make the ansatz

hz|Ψi = αeikz + βei(k−g)z , (3.71)

65
 Figure 3.20. Cosinus hyperbolicus.

which leads to the secular equation

 2  
k − 2V0 − 2E 2Vg α
= 0. (3.72)
2Vg (k − g)2 − 2V0 − 2E β

Here, the ansatz of two plane waves is sufficient since we are interested in the
band-gap range, that is at k ≈ g/2. For the energy one readily obtains
r
g 2 κ2 g 2 κ2
E(k) = −V0 + + ± + Vg2 , (3.73)
8 2 4
with k = g/2 + κ. Hence, κ is the deviation from the BZ boundary. The
wavefunction reads

hz|Φk i = eiκz cos(gz/2 + δ), (3.74)

with exp(2iδ) = 2(E+V0 )−k2/2Vg . Note that κ can be both real and imaginary.
Problem 36. Confirm the above results. Consider the symmetry of the wavefunction upon
reflection at z = 0. How does it depend on the sign of Vg ?
The secular equation gives with E 0 ≡ E + V0

0 = [k 2 − 2E 0 ][(k − g)2 − 2E 0 ] − 4Vg2

0 = 4E 0 − 2E 0 [k 2 + (k − g)2 ] + k 2 (k − g)2 − 4Vg2
0 = 4E 0 − 2E 0 [(κ + g/2)2 + (κ − g/2)2 ] + [(κ + g/2)(κ − g/2)]2 − 4Vg2
0 = 4E 0 − 2E 0 [2κ2 + g 2 /2] + [κ2 − g 2 /4]2 − 4Vg2
1
0 = E 0 − E 0 [κ2 + g 2 /4] + [κ2 − g 2 /4]2 − Vg2 .
4
Thus,
r
0 1 2 2 1 2 1
E = [κ + g /4] ± [κ + g 2 /4]2 − [κ2 − g 2 /4]2 + Vg2
2 4 4
r
2 2
κ g 1 2 2
= + ± κ g + Vg2 .
2 8 4

66
For Vg = 0 the two bands are degenerate at k = g/2 (i. e. at κ = 0) with E 0 = g 2 /8. For
nonzero Vg one has at k = g/2
g2
E(g/2) = −V0 + ± |Vg |,
8
implying a band gap of size 2|Vg |. Further,
2E + 2V0 − k 2 |Vg |
e2iδ = =± = ±1,
2Vg Vg
which implies δ = nπ (case +1) or δ = (n + 1/2)π (case −1). The wavefunction can then
be expressed as
(
cos(gz/2) +1
hz|Φk i ∝ ,
sin(gz/2) −1

that is even or odd under reflection at z = 0. 2

Problem 37. Sketch E(κ2 ), allowing for negative κ2 .
Real bands are obtained either for real or imaginary wavenumber κ.18 The complex band
structure for k ≈ g/2 is shown in Fig. 3.21. Note that bands for imaginary wavenumbers
(blue) bridge the bands for real wavenumbers (black). 2

Figure 3.21. Complex band structure (schematic). Bands for real wavenumbers are
shown in black, those for imaginary wavenumbers in blue.

The infinite crystal is now cut at z = a/2, with the bulk potential for z < a/2
and zero potential for z > a/2 (Fig. 3.22). Further, we consider the bulk-band
gap range, that is Re κ = 0 (κ imaginary). Changing κ to iκ, the wavefunction
is given by
(
eκz cos(gz/2 + δ) z < a/2
hz|Φk i ∝ , (3.75)
e−qz z > a/2

with q 2 = V0 − E. Note that the wavefunction increases with z in z < a/2 (cf.
the ‘complex band structure’) and decreases for z > a/2. Both solution would
not be allowed in an infinite (bulk) crystal.
18
Remember that the bands must be real because the Hamiltonian is hermitian.

67
Figure 3.22. One-dimensional model for surface-state formation. The bulk potential
(left) is abruptly replaced by a constant potential (right). The dashed line depicts a
surface-state wavefunction.

If the logarithmic derivative of the wavefunction19 can be made continuous

at z = a/2, an electronic state exists which is located at the surface (Fig. 3.23).
Note that the only ingredient for the surface-state formation is the truncation of

Figure 3.23. Wavefunction matching at z = a/2 for Vg < 0 (a) and Vg > 0 (b). The
labels 1, 2, and 3 indicate increases energies starting from the bottom of the band gap.

the bulk crystal. This type of surface state is called Shockley surface state. [44]
At last, two explanations of important energies. The vacuum level is the
potential energy in the vacuum infinitely far away from the surface. In the
above example, it is assumed zero. The work function is the energetic difference
between the vacuum level and the Fermi energy. Typically it is in the order of
4 eV to 5.5 eV.
The surface potential is, of course, smooth (and not abrupt as in the example)
and behaves like V ∝ 1/4z . Like for an hydrogen atom, Rydberg-like series of
electronic surface states can show up, typically above the Fermi energy but
below the vacuum level. Since these are a consequence of the image potential
at the solid-vacuum boundary (and not solely of the truncation), they are called
image states.20

19 d 0
That is dz ln Ψ = Ψ /Ψ.
20
Remember the method of image charges in electrostatics.

68
1D TB model. As example, we consider now the occurrence of surfaces states
in the 1D TB model. By cutting the infinite Hamiltonian matrix (i. e. replacing
one half-space by vacuum) one arrives at
 
s ts
 ts  t 
 
H=
 t  t .

(3.76)

 t  t  
.
t ..
where s and ts are allowed to differ from their respective bulk values. This,
still infinite, problem can be solved exactly using recursion or renormalization
methods, thereby exploiting the tridiagonal structure of the Hamilton matrix.
[45–48]
The resulting site-resolved spectral density is shown in Fig. 3.24 (left). Note

Figure 3.24. Site-resolved spectral density of 1D TB models. Left: semi-infinite

chain without modification of the surface parameters. Hence, no surface states ap-
pear. [47] Right: antiferromagnetic semi-infinite chain with modification of the surface
parameters. Surface states appear for different values of Ys (layer S). [49]

that n − 1 maxima show up at the n-th layer. The concentration of spectral

weight at the band center  in the surface layer can be regarded as band narrow-
ing, i. e. an effective reduction of the band width (2t in the bulk). No surface
states show up because ts = t and s = .

69
 For an antiferromagnetic semi-infinite chain (i. e. a chain with alternating
orientation of the magnetization) surface states show up when increasing or
decreasing the surface-layer magnetization, as compared to the bulk-layer mag-
netization (right in Fig. 3.24).
In contrast to Shockley surface states, the crystal potential at the surface
(here expressed in terms of s and ts ) must deviate considerably of the respective
bulk quantity in order to form a Tamm surface state.
As an example, Fig. 3.25 shows the surface-electronic structure of the relaxed
GaAs(110) surface. In the band-gap regions, a lot of surface state show up, typ-
ically close to band edges. Some surface states can be traced into the projected

Figure 3.25. Surface states on GaAs(110). The hashed areas depict the range of
bulk bands (projected bandstructure). Theoretical and experimental surface states are
shown as solid and dashed lines, respectively. A5 is the famous dangling-bond surface
state. The two-dimensional Brillouin zone is depicted in the fundamental bang gap.

bulk-band structure, see e. g. A5 in the vicinity of Γ. Here, the surface state hy-
bridizes with Bloch states, maintaining its large spectral weight at the surface.
In turn, these electronic states can be viewed as bulk states with an increased
spectral weight at the surface. Thus, they are called surface resonances.
Interface states
An interface is the boundary region between two different materials. Hence, the
surface can be regarded as a special interface, namely the boundary between
a semi-infinite solid and the semi-infinite vacuum region. In the vacuum re-
gion, electronic states with energies below the vacuum level decay exponentially

70
whereas states with energies above the vacuum level behave asymptotically as
free electrons.

3.4 Defects and disorder

So far we have considered perfect crystals. In reality, however, solids are not per-
fect. There can be defects, like unoccupied lattice sites, impurities (occupation
of a lattice site by a ‘wrong’ atomic species), or anti-site defects (interchange
of atomic species in a bi-atomic lattice). Defects break the translational sym-
metry of the lattice and, thus, Bloch’s theorem does not hold. Consequently,
electronic states localized at the defect site can show up, as for surfaces and in-
terfaces.21 Further, Bloch states are scattered diffuse at the defect, that is, the
wavevector k has changed (Fig. 3.26). Defects can be impurities or vacancies.

Figure 3.26. Scattering of a Bloch state at a single defect (point scatterer). The
incoming plane wave is turned into an outgoing spherical wave.

If multiple defects show up, one is concerned with disorder. Disorder com-
prises structural disorder, for example random orientation of molecules in an
embedding matrix, or substitutional disorder, that is random occupation of lat-
tice sites by different species. Magnetic disorder shows up as random orientation
of magnetic moments.
Substitutional disorder, which we want to discuss in the following, can be
described as random occupation of lattice sites by different atomic species.
There are several ways to treat disorder theoretically, for example supercell
and mean-field approaches.
Supercell approach
We now discuss the supercell approach for a substitutional binary alloy Ac B1−c ,
made up of atomic species A and B with concentration c ∈ [0, 1].
Consider a lattice with its primitive cell spanned by a1 , a2 , and a3 . We now
introduce an artificially enlarged primitive cell spanned by Ai = Ni ai , with
i = 1, 2, 3 and Ni positive integer. This enlarged cell is called a supercell. For

21
A famous example is Anderson localization.

71
Figure 3.27. Ordered and randomly disordered InGaAs crystal (In red, Ga blue, As
white).

the reciprocal basis vectors, one has (i, j, k cyclic)

Aj × Ak
B i = 2π (3.77a)
A1 · (A2 × A3 )
Nj aj × Nk ak
= 2π (3.77b)
N1 N2 N3 a1 · (a2 × a3 )
Ni Nj Nk aj × ak
= 2π (3.77c)
Ni N1 N2 N3 a1 · (a2 × a3 )
1
= bi . (3.77d)
Ni
The unit cell of the reciprocal lattice associated to the supercell lattice is smaller
than that of the original lattice.
Up to now the original lattice and the supercell lattice have identical elec-
tronic structures because both contain the same atoms, say of species A. Espe-
cially they are both ordered (no disorder). Hence, to treat disorder in the
alloy, we replace randomly A atoms by B atoms such that their numbers
NA and NB in the supercell are compatible with the given concentration c,
c = NA /(NA + NB ). This particular arrangement of A and B atoms is called a
configuration. It can be viewed as a snapshot of the real alloy in a finite spatial
area (namely in the supercell). Evidently, the computed properties hold only for
this particular configuration. To obtain the properties of the real alloy, one has
to average over all configurations that are compatible with the concentration c.
Note that the supercell lattice with A and B atoms is ordered, since it obeys
the translational invariance for any vector R = n1 A1 +n2 A2 +n3 A3 , ni integer.
Thus, we can describe the electronic structure in terms of Bloch states |Ψi of

72
the supercell lattice.
In the preceding, it was mentioned that disorder leads to diffusive scatter-
ing. The supercell approach allows to investigate both coherent and decoherent
scattering. Therefore we consider the transmission through a disordered film.
This is equivalent to the tunneling problem (Section 2.7) but with Bloch states
of ordered electrodes (instead of free particles) and a disordered barrier (instead
of the potential step).
We start with the Bloch states in the ideal electrodes. For the supercell lat-
tice these are labeled by a wavevector K in first Brillouin zone of that lattice
(described by reciprocal vectors B i ), |Ψ(K)i. The Bloch states of the origi-
nal lattice are named |ψ(k)i, with k in the first Brillouin zone of that lattice
(described by reciprocal vectors bi ).
Since the electronic structure is identical in all Brillouin zones, we can use
the first Brillouin zone of the supercell lattice. Writing
K = αB 1 + βB 2 + γB 3 , (3.78)
with α, β, γ ∈ [−1/2, +1/2], k can be expressed as
k =K +G (3.79a)
= αB 1 + βB 2 + γB 3 + n1 B 1 + n2 B 2 + n3 B 3 (3.79b)
α + n1 β + n2 γ + n3
= b1 + b2 + b3 . (3.79c)
N1 N2 N3
Thus, k can be viewed as wavevector in the first Brillouin zone of the supercell
lattice that is shifted by a reciprocal lattice vector G, namely by n1 B 1 +n2 B 2 +
n3 B 3 . Knowing the Bloch states |Ψ(K)i of the supercell lattice we can map
them to those of the original lattice, |ψ(k)i. This is called an umklapp process.
Or in other words, the Bloch states of the original lattice are mapped into the
first Brillouin zone of the supercell lattice.
Now consider scattering at a disordered film, the latter being described within
the supercell approach (Fig. 3.28). In the picture of the supercell approach, the
transmission of an incoming Bloch state is coherent because of the 2D trans-
lational invariance. Mathematically this means that the in-plane components
of the wavevector, K k are conserved in the transmission process. The normal
component K⊥ is not conserved since the translational invariance in that di-
rection is broken by the film. Thus, the incoming Bloch state |Ψ0 (K)i can
be scattered into outgoing Bloch states |Ψn (K)i that have identical K k but
different K⊥ .
Mapping now the supercell-lattice Bloch states to those of the original lat-
tice, the scattering comprises now two terms. The coherent contribution is the
transmission from the incoming Bloch state |ψ0 (k)i into outgoing Bloch states
|ψn (k)i with identical kk . These states are umklapped by the same vector G.
The diffusive (or decoherent) contribution comprises transmission into Bloch
states with different kk (due to a different umklapp vector).
Mean-field approaches
In a mean-field approach, the disordered medium is replaced by an effective
medium that is translationally invariant. We will consider here binary substitu-

73
Figure 3.28. Transmission through an disordered interface (box) between ordered Cr
and Fe electrodes. The incoming Bloch state has wavevector (kk , kz ), the outgoing has
(k0k , kz0 ). The bottom panel shows the in-plane view of the randomly occupied supercell
at the interface layers.

tional alloys Ac B1−c , made up of species A and B with concentration c ∈ [0, 1].
The simplest approach is the virtual crystal approximation (VCA). Here, the
potentials of both species are replaced by their average,

Veff (r) = cVA (r) + (1 − c)VB (r). (3.80)

This approximation works rather well if the electronic structure of the species
do not differ to much, for example, for iso-electronic constituents. It does not
work well in the opposite case, for example for Ni and Cu. In Cu, the d-shell is
filled and the electronic structure at the Fermi energy consists of sp-states. Ni,
however, is an open d shell and the electronic structure at the Fermi energy is
made up by d-states.
The today probably most often used approach is the coherent potential ap-
proximation (CPA). Here, the effective medium is constructed by the condition
that there is no scattering in the effective medium (so-called CPA condition;
Fig. 3.29). The scattering at a single defect of species A (yellow, left) in the
effective medium (green) weighted by its concentration cA = c and the scatter-
ing at a single defect of species B (blue, right) in the effective medium (green)
weighted by its concentration cB = 1 − c should have no effect (bottom).

74
 Figure 3.29. Coherent potential approximation (CPA).

3.5 From the many-electron to the single-electron

problem
So far we have assumed that electrons in the solid do not interact with each
other which, due to the mutual Coulomb interaction, is not true. In other words,
we worked within the single-particle or independent-particle approximation.
Suppose we have an ensemble of Ne electrons (with mass m and Nc ionic
cores (with mass M ). The total Hamiltonian can be written as

Ĥ = Ĥe−e + Ĥc−c + Ĥe−c + Ĥext . (3.81)

The electron-electron Hamiltonian Ĥe−e is given by

X p̂2 0
k 1X e2
Ĥe−e = + , (3.82)
2me 2 0 |r k − r k0 |
k k,k

where the electrons are labeled by k. The term k = k 0 in the Coulomb inter-
action is excluded (second term). For the ions we have

X P̂ 2 0
i 1X
Ĥc−c = + Vc (Ri − Ri0 ), (3.83)
2M 2 0
i i,i

and for the electron-ion interaction

X
Ĥe−c = Ve−c (r k − Ri ). (3.84)
k,i

An external potential (perturbation) is described by Ĥext .

75
 Since the mass m of the electrons is much smaller than the ion mass M , the
motion of the electrons can be approximated as being independent of the ionic
motion. This is the adiabatic or Born-Oppenheimer approximation. Thus,
the problem remains to describe the electronic motion for a fixed geometric
arrangement of ions. Hence, the position Ri of the ions appear as parameters
and are neglected for clarity.
Since the total problem decomposes into two independent ones (electrons and
ions), the total wavefunction is the product of the electronic wave function and
the ionic wavefunction (cf. the quantum-well problem, page 26).
The Schrödinger equation for the electrons then reads
X p̂2 0
k
X 1X e2
Ĥ = + V (r k ) + , (3.85)
2me 2 0 |r k − r k0 |
k k k,k

where V (r k ) comprises the potentials from the ions and the external potential.
Without the Coulomb interaction (third term), one would have a sum of single-
electron Hamiltonians (first and second terms) which can be solved by a product
ansatz,

hr 1 , . . . , r Ne |Φi = hr 1 |φi · · · hr Ne |φi. (3.86)

This leads to the Hartree equation for the single-electron wavefunctions (atomic
Hartree units)
 
Z 0
|hr |φk i| 2
− 1 ∆ + V (r) +
X
dr03  hr|φj i = Ej hr|φj i. (3.87)
2 |r − r 0 |
k6=j

The solution for the j-th electron depends on the density of all the other elec-
trons. The integral is called the Hartree term.
For Fermions, the total wavefunction must change sign when interchanging
two electrons,

h. . . , r i , . . . , r j , . . . |Φi = −h. . . , r j , . . . , r i , . . . |Φi (3.88)

for i 6= j. Thus, the above product ansatz is not sufficient. An ansatz which
fulfills this requirement is a Slater determinant,

hr 1 |φ1 i ··· hr 1 |φNe i

1 .. ..
hr 1 , . . . , r Ne |Φi = √ . . . (3.89)
Ne !
hr Ne |φ1 i · · · hr Ne |φNe i

This results in the Hartree-Fock equation

 
1 X Z |hr 0 |φk i|2
− ∆ + V (r) + dr03  hr|φj i
2 |r − r 0 |
k6=j
(3.90)
X Z hφk |r 0 ihr 0 |φj i
− dr03 hr|φk i = Ej hr|φj i.
|r − r 0 |
k6=j

76
The second integral is the exchange interaction between electrons with parallel
spin.
The Hartree-Fock equation can be solved for a few electrons but in a solid the
number of electrons is in the order of 1023 . Thus, we need another approach.
Instead working with wavefunctions exclusively, Hohenberg and Kohn showed
that the electron density can be taken as the relevant quantity [50]. Thus, the
large number of wavefunctions was replaced by a single quantity, namely the
density. They showed that the ground state of the N electron system is a
functional of the electron density n(r) and that the energy functional E[n]
is minimal for the ground-state density (Hohenberg-Kohn theorems). This
allowed to formulate a single-electron theory in which an electron moves in an
effective potential (density-functional theory) [51]

Veff = VH + Vext + Vxc , (3.91)

that is the sum of Hartree, external and exchange-correlation potential. The

single-particle wavefunction fulfills the Kohn-Sham equations [52, 53]

(T̂ + Veff [n])|φk i = Ek |φk i (3.92a)

X
n(r) = |hr|φk i|2 , (3.92b)
k

to be solved self-consistently. The exchange-correlation potential is, however,

not known and must be approximated. The most important approximation
is the so-called local density approximation (LDA) in which one takes the
exchange-correlation potential of an electron gas (no ion and external poten-
tials).22 It works well if the electron density is smooth. There exist sev-
eral extension of the LDA; [54] just to name a few, self-interaction correc-
tion (SIC), [56] generalized gradient approximation (GGA), [57] and extensions
based on the Hubbard model [58, 59] (LDA+U, dynamical mean field theory
(DMFT) [60]). Today, DFT-LDA calculations are standard.
Fermi-Dirac distribution
In equilibrium, Fermions occupy states with energy E with a probability given
by the Fermi-Dirac distribution,
1
f (E, T ) ≡ (E−µ)/k T
. (3.93)
e B +1
Here, µ is the chemical potential, that is that energy for which the density of
occupied states with larger energies equals the density of unoccupied (‘empty’)
states with smaller energy. T is the temperature and kB Boltzmann’s constant.
The Fermi energy EF is the chemical potential for T = 0 K.
Problem 38. How large, in eV, is kB T at room temperature (300 K)?
With kB = 8.617 10−5 eV/K one has an energy of 0.023538 eV at room temperature.
2

22
The exchange-correlation potential of the electron gas is computed with sophisticated meth-
ods and parameterized.

77
 Figure 3.30. Fermi-Dirac distribution for several temperatures.

The electron density n(T ) is given by

Z +∞
n(T ) = N (E) f (E, T ) dE, (3.94)
−∞

where N (E) is the density of states (at T = 0 K). The effect of the Fermi-Dirac
distribution on the electron density is sketched in Fig. 3.31.

Figure 3.31. Effect of the Fermi-Dirac distribution on the density of states. (top)
Density of states. (center) Fermi-Dirac distribution. Product of DOS and FD distri-
bution (green; bottom).

78
 For metals EF lies within (at least) a band, whereas for semiconductors or
insulators it resides in the fundamental band gap (which separates the occupied
from the empty electronic states).

79
4 Nanoscale probes
This chapter gives an introduction to methods which probe properties of nano-
structures. Properties of nanostructures can be cast in three classes.
Geometric properties Nanostructures comprise a great deal of atoms that are
located at a surface or interface. Evidently, their positions are likely not
those which can be extrapolated from the bulk properties of the consti-
tuting material.
Following the bottom-up-approach (p. 8), nanostructures can be build
by assembling atomic structures, for example by arranging atoms on a
surface. To do so, their geometric positions must known.
Methods that probe the geometry should preferably be sensitive in all
three spatial dimensions, that is laterally and vertically.
Electronic structure As we have learned in the introduction, electronic trans-
port in nanostructures differs from that in the corresponding bulk ma-
terial. Hence, electronic probes must be sensitive to the surface or the
interface electronic structure, sometimes even to buried structures.
Magnetic properties Nanostructures often contain magnetic parts, in order to
exploit the spin degree of freedom of the electrons (‘spintronics’ or ‘mag-
netoelectronics’). The magnetic structure at an interface can differ con-
siderably from that in the interior in both magnitude and orientation of
the magnetization.
Obviously, all three classes of properties depend on each other: the geometric
arrangement of atoms adsorbed on a surface depends on the electronic proper-
ties of that surface. Further, magnetic properties can be viewed as electronic
properties of spin-up and spin-down electrons.
In order to probe specific properties of a sample, one applies a spectroscopy.
In a spectroscopic method, an external perturbation is applied to the sam-
ple and the change of an observable in dependence on external parameters is
recorded. The sample is said to respond to the perturbation. Typical per-
turbations comprise incident electrons or electromagnetic radiation. Measured
quantities are an outgoing electron current or emitted light. Or in a transport
measurement, one records the current for an applied bias voltage.
In the following a brief overview on a few nanoscale probes is given.

4.1 Low-energy electron diffraction

In low-energy electron diffraction (LEED) a beam of electrons with low kinetic
energy impinges onto a surface (typically 20 eV to 500 eV). The currents of the

80
elastically reflected beams is recorded and gives information on the morphology
of the sample surface.
Three ingredients make LEED important for the determination of surface
geometric structures. [61–63]

Slow electrons The de-Broglie wavelengths of electrons with kinetic energies

of 20 eV or 500 eV are 2.7 Å or 0.6 Å, respectively. This is in the range of
lattice spacings (about 2.5 Å) and, hence, interference effects upon diffrac-
tion at the lattice of the sample should show up.

Mean-free path The mean-free path (MFP) of electrons with energies from
20 eV to 500 eV is about 6 Å for inelastic scattering processes, [64] that is
less than the thickness of three monolayers at the surface. This implies
that electrons that penetrate deeper into the sample than the MFP are
likely to be scattered inelastically. In LEED, however, only elastically
scattered electrons are detected. By this means, LEED becomes very
sensitive to the surface geometry.

Figure 4.1. Mean free path in solids. [64]

Cross section The cross-section of elastically back-scattered electrons is very

large. This implies that the scattering properties are given by a few
surface layers, whereas deeper layers do not contribute. Also by this
means, LEED becomes very surface-sensitive.

In the 3D case, the wavevector ki of the incident and the wavevector kf of the
elastically scattered electron differ by a reciprocal lattice vector g, kf − ki = g.
Only in this case, the scattered electron waves interfere constructively and,
hence, a nonzero intensity can be measured. Note that a change of the kinetic
energy changes the de-Broglie wavelength and thus the interference condition.
Problem 39. Derive the above condition for constructive interference.
The incident wave is given by

ψi (r) ∝ eiki ·r ,

81
while the outgoing can be written as
X
ψf (r) ∝ eikf ·(r+R) ,
R

where the sum runs over all lattice sites R. The total wavefunction is then
X
ψ(r) ∝ ei(kf −ki )·(r+R) .
R

Remembering the representation of the δ-function, we arrive at

ψ(r) ∝ ei(kf −ki )·r δ(kf − ki + g),

which establishes the condition for constructive interference. 2

The surface breaks the translational symmetry normal to the surface. Hence,
only conditions for the in-plane (coplanar with the surface) components of the
wavevectors remain valid: kkf − kki = g k . Since there is no condition for
the normal components,1 there is nonzero intensity at an kinetic energy. In a
graphical representation, this implies that a 3D reciprocal-lattice vector must
lie within the sphere around ki and kf (Fig. 4.2).

Figure 4.2. Condition for constructive interference in LEED.

The above condition for the in-plane components establishes so-called LEED
spots, that is intensity shows up on the 2D reciprocal lattice shifted by kki .
Since each 2D g can be written as g = nb1 + mb2 , the LEED spots can be
labeled (n, m) (Fig. 4.2). LEED spots are also called ‘beams’.

1
In order to leave the crystal, outgoing electrons must have k⊥f > 0.

82
Kinematical theory of LEED
In order to compute the intensities of the reflected beams one can use as an
approximation the so-called kinematical theory that takes into account only
single-scattering events. For a better description, however, a dynamical theory
has to be used. The latter include all multiple-scattering events.
Evidently, the reflected intensities depend on the geometric arrangement of
the scatterers at the topmost layers of the sample. Hence, we decompose the
sample into a substrate (in which the scatterers sit on bulk sites) and the surface
layer (with scatterers at different positions). The incoming wavefunction can
be written as

Ψi = Ψ0 eik0 ·r . (4.1)

An outgoing wave scattered at R0 and detected at R is given by

eik·R 0
ΨR0 = Ψ0 f (k − k0 ; R0 ) ei(k−k0 )·R . (4.2)
R
The first term is a spherical wave. The second term is the structure factor which
in a kinematical theory depends only on the difference of the wavevectors. It
describes the scattering amplitude for site R0 . The third term takes care of the
phase difference relative to the origin of the coordinate system. In practice, the
detector is positioned very far from the sample. Thus, the spherical wave can
be replaced by a plane wave in very good approximation.
The total outgoing wave is given by the superposition of the waves reflected
at the substrate and at the surface layer, Ψf = Ψfsub + Ψfsurf . Then,
X
Ψfsub ∝ fsub (k − k0 ) ei(k−k0 )·(n1 a1 +n2 a2 +n3 a3 ) . (4.3)
n1 n2 n3

Hence, the wavefunction is the product of a single-site scattering term and a

geometrical factor, a typical decomposition in scattering theory. Defining
X
Ssubj (k − k0 ) ≡ ei(k−k0 )·nj aj , (4.4)
nj

the intensity from the substrate at the detector is given by

3
Y
Isub ∝ |Ψfsub |2 ∝ |fsub (k − k0 )| 2
|Ssubj (k − k0 )|2 . (4.5)
j=1

Now we perform the summation in the geometrical factors, with the con-
venient choice of a1 and a2 lying within the surface plane.2 To describe the
surface sensitivity, we introduce an empirical attenuation factor µ, such that
X
Ssub3 (k − k0 ) = ei(k−k0 )·jaj −jµ , (4.6)
j

2
The basis vectors can always by chosen this way.

83
yielding (geometrical series)
−1
|Ssub3 (k − k0 )|2 = 1 − 2e−µ cos[(k − k0 ) · a3 ] + e−2µ

. (4.7)

This term is maximal if (k − k0 ) · a3 = 2πh3 , with h3 integer. This condition is

called the third Laue condition. Analogously, one has the other Laue conditions
(k−k0 )·a1 = 2πh1 and (k−k0 )·a2 = 2πh2 from Ssub1 (k−k0 ) and Ssub2 (k−k0 ).
Problem 40. Derive the first and the second Laue condition.
With
X
Ssub1 (k − k0 ) = ei(k−k0 )·n1 a1
n1

we can immediately use the δ-function, giving (k − k0 ) · a1 = 2πh1 .

Or we use the result for Ssub3 and making the limit µ → 0,
−1
|Ssub1 (k − k0 )|2 = lim 1 − 2e−µ cos[(k − k0 ) · a1 ] + e−2µ

.
µ→0

This results in
1 1
|Ssub1 (k − k0 )|2 =
2 1 − cos[(k − k0 ) · a1 ]
which becomes infinite if (k − k0 ) · a1 = 2πh1 . Remembering the ‘cooking recipes’ for the
δ-function, we have established the von Laue condition. Of course, the latter ‘derivation’
is mathematically questionable.
An analogous calculation can be performed for Ssub1 . 2
If the surface atoms are shifted by τ from their bulk positions, the surface
structure factor reads

fsurf (k − k0 ; τ ) = fsurf (k − k0 ) ei(k−k0 )·τ . (4.8)

Hence, a displacement gives rise to interference effects.

The wavefunction reflected from the surface layer reads
X
Ψfsurf ∝ fsurf (k − k0 ; τ ) ei(k−k0 )·(n1 a1 +n2 a2 +n3 a3 ) . (4.9)
n1 n2 n3

Note that in principle the 2D lattice at the surface can differ from that of the
bulk. Here, we consider relaxation only,3 especially in the topmost layer. The
sums over n1 and n2 establish again the first and second Laue condition. The
sum over n3 comprised only a single term, due to the monolayer surface; hence,
there is no third Laue condition for a single layer.
The total intensity at the detector is then given by
?
Itot ∝ |Ψfsub + Ψfsurf |2 ∝ Isub + Isurf + 2Re (Ψfsub Ψfsurf ). (4.10)

The displacement τ shows up as modulation of the total intensity. Thus, by

comparing measured intensities with calculated one, one can (iteratively) de-
termine τ . The error is, for dynamical theories, in the range of a few per cent
of the lattice spacing.
3
Relaxation means a displacement of atoms from their bulk positions, keeping the 2D unit
cell unchanged. Reconstruction means a change of the 2D periodicity and, hence, a change
of the 2D unit cell as compared to that in the bulk.

84
 The LEED image is a direct image of the reciprocal lattice, as is clear from
Fig. 4.3, hence allowing to determine the 2D periodicity of the sample. The
intensity in each LEED spot provides information on the atomic arrangement
at the surface.4

Figure 4.3. LEED apparatus (left) and reciprocal space (right). An electron beam
with wavevector ki impinges onto the sample, giving rise to reflected beams with kf
made visible on a fluorescent screen. The right column gives the index of the LEED
beams.

Using a spin-polarized incident electron beam and recording the spin polariza-
tion of the outgoing beams one achieves information on magnetic surfaces. But
also the spin-orbit interaction leads to spin-polarized reflection (similar to Mott
scattering). This method is known as spin-polarized LEED (SPLEED). [65]
Up to now we were concerned with ordered surfaces. But in reality, surfaces
can show steps, adatoms, impurities, and defects. All these perturbations re-
lax the Laue conditions. Hence, the LEED beams become ‘smeared out’ in a
characteristic way. The intensity profile when scanned through the LEED beam
thus provides information on the surface morphology as compared to the perfect
surface. This method is known as spot-profile analysis-LEED (SPA-LEED).
In a SPA-LEED experiment, both the energy (and hence k⊥ ) and the in-plane
wavevector kkf are scanned in order to determine the spot profile. The phase
shift S between an outgoing wave that is reflected from the bare surface and a
step (Fig. 4.4) is given by
k⊥ d
S= , (4.11)
2π
where d is the interlayer distance. Thus, constructive interference shows up
for S integer, destructive interference for S half-integer. In a reciprocal-space
representation, the intensity shows up as shown in Fig. 4.5. SPA-LEED pro-
vides information on surface roughness, terrace width, and mean island size.
Another technique is low-energy electron microscopy (LEEM) that provides
direct images of the sample surface using electron beams.
4
Instead of recording the intensities of each beam, one can also determine the current which
flows through the grounded sample. This is known as target current spectroscopy (TCS)
and is typically applied to very low electron energies (less than, say, 30 eV).

85
Figure 4.4. Phase shift S between two waves reflected at the surface and a step of
one monolayer height. λel is the de-Broglie wavelength of the electron.

Figure 4.5. Spot profiles for a perfect surface (left) and a surface with monatomic
islands (left). S is the phase shift defined in eq. (4.11).

As is evident from the above, LEED relies heavily on phase-coherent pro-

cesses. In particular the first two Laue conditions are necessary, which are only
valid if a considerably large part of the surface is probed. Thus, LEED provides
mainly information on the surface on a large scale. It has no atomic resolution.
Surface structure determination by LEED
The 2D periodicity of the sample can be determined by LEED images, via the
von Laue conditions. Here, the position of the reflected current is imaged. For
simple 2D lattices, for example that of a bcc(001) surface, this can be readily
achieved (Fig. 4.6). Note that the distance between LEED spots decreases
with increasing energy of the incident beam. For large superstructures, such as
Si(111)-7×7, the LEED image becomes rather complicated.
In the analysis of LEED images one has to be careful since large intensities
show up only if the third von-Laue condition is fulfilled (cf. Fig. 4.8).
Problem 41. Why is the third von-Laue condition not strictly fulfilled in Fig. 4.8? Give
possible reasons.
First, the electron in the solid are not free electrons. Hence, band-structure effects
can show up, resulting in additional maxima or suppression of maxima. Second, multiple

86
Figure 4.6. LEED images of W(100) at 45 eV (left) and 145 eV (right). The LEED
spots are indicated in the right picture.

Figure 4.7. LEED image of Si(111)-7×7 at 81 eV (top) and corresponding structural

model (bottom).

scattering. Third, the structure factor can depend on the direction of the wavevectors, not
only on their difference. Fourth, relaxation of the surface layers. Fifth, imperfections. 2
The LEED images provide information only over the periodicity, not on the
positions of the atoms in the 2D unit cell. If the surface has an overlayer (a
reconstructed layer), the reciprocal lattice differs from that of the bulk system
(substrate). What will the LEED pattern look like? One would guess that the
LEED pattern is just the sum of the two reciprocal lattices but this is only
partly true: due to multiple scattering one does not only get the spots of both
reciprocal lattices (truncated bulk and surface/overlayer) but also all possible
combinations between them. If we adopt the point of view that the lattice of
the surface is made up by the adsorbate and the first few layers of the substrate,

87
Figure 4.8. Intensity of the (0,0) spot from Ni(100) versus of electron kinetic energy.
The arrows indicate the positions where maxima would be expected if the third von
Laue condition would be valid.

Figure 4.9. The LEED pattern shows the sum of the reciprocal lattices from substrate
and overlayer plus all possible combinations between them. For a simple overlayer
structure as in (a), this combination does not lead to any new spots. For a coincidence
structure (b) it does (grey spots). The arrows indicate the size of the surface unit cell
as a whole. When this unit cell is taken to calculate the reciprocal lattice, the ‘extra’
spots appear in a natural way.

these additional spots enter in a natural way into the reciprocal lattice. This is
illustrated in Fig. 4.9.
A lot about the surface structure can be learned simply by the inspection of
the LEED pattern without considering the quantitative intensity behaviour of
the spots. Basically, we see the reciprocal lattice of the surface and from this
we can construct models for the real lattice. There are, however, several effects
complicating this analysis.
The first question which arises when inspecting a LEED pattern of a clean
surface is if the surface is reconstructed or not. This can be found out by
comparing the position of the spots to the positions one would expect for the
(1×1) unreconstructed surface. The simplest way is to estimate the energy at
which the (1×1) spots would first appear on the fluorescent screen and compare
this to the measured energies. Once the (1×1) spots are identified one can
describe any superstructure referring to them.
Now consider the LEED patterns of simple overlayer structures and coinci-

88
dence structures. Since we know which spots are the original (1×1) substrate
spots we can now deduce the reciprocal lattice of the superstructure from the
LEED pattern. Fig. 4.10 gives a few examples. From such LEED patterns the
surface periodicity and the point-group of the surface may be deduced.

Figure 4.10. Three examples for overlayer structures and the LEED patterns pro-
duced by them. (a) a (4×2) structure, (b) a c(4×2) structure. In the LEED patterns
the open circles are the (1×1) spots. The (1×1) unit cell in reciprocal space is also
given.

As we have seen, the inspection of the LEED pattern gives information about
the surface periodicities and to some degree also about the surface symmetry.
But there are still many important things one would like to know. One is
the site of the adsorbed atoms. If we shift the whole overlayer, the diffraction
pattern will be unchanged. Also, one could put more adsorbate atoms into the
same unit cell and still keep the pattern the same. Clearly, a more quantitative
analysis of LEED is needed.
This is achieved by analyzing the I(V ) curves. The I(V ) curves for a par-
ticular model structure can be calculated by a computer program. Then they
are compared to the measured I(V ) curves. If the agreement is not good, the
model structure is changed and the structure is re-calculated. This process is
repeated until eventually a good agreement between experiment and theory is
achieved. The degree of agreement is quantified by a so-called reliability or
R-factor. The lower the R-factor, the better the agreement.
In a LEED calculation the crystal is described by a so-called muffin-tin po-
tential. This consists of spherical potentials for the ion cores and a constant
potential everywhere else. The spherical potentials are characteristic for the
element of the scatterer and depend somewhat also on its environment (but not
very much). They can be described by a set of scattering phase shifts. This

89
choice of potential has the advantage that reduces the problem basically to
scattering from spherical potentials which can be treated very efficiently. The
program must now explicitly solve the Schrödinger equation in the muffin-tin
potential including all possibilities of multiple scattering.
There are two additional effects which also have to be taken into considera-
tion by the program. The first is the inelastic scattering of the electrons. This
is handled by making the constant part of the potential in the solid complex.
The imaginary part corresponds to the energy-dependent mean free path of the
electrons and takes care of the inelastic scattering. The other effect is finite
temperature. It reduces the scattering coherence in otherwise periodic struc-
tures and thereby reduces the intensity in the I(V ) curves. Finite temperature
is taken into account by temperature-dependent scattering phase shifts.
The level of agreement which can be obtained between experiment and theory
is remarkable, at least for many metal surfaces, and gives a high confidence into
the LEED technique. Fig. 4.11 gives an example.

Figure 4.11. Agreement between measured and calculated LEED I(V ) data for
different spots in case of the Al(111) surface. The full line is experimental data, the
dashed line calculation.

The possibility of high-quality calculations also means that the structural

90
parameters are determined very precisely by LEED. The atomic positions are
given within a tenth of an Ångström or even better. But there are also some
problems associated with the analysis approach employed for LEED. The suc-
cess depends on the researcher’s ability to come up with the right structural
model which can then be refined in a trial and error iterative analysis. This is
not too difficult for unreconstructed metal surfaces where the truncated bulk
can be taken as a starting model. But it is a major problem for semiconductor
surfaces where reconstructions with large surface unit cells are possible as we
shall see below. Such large unit cells have the further disadvantage that they
are extremely expensive in terms of computer time needed for the calculations.
Many atoms in the unit cell mean many structural parameters which one has to
get all right in order to obtain satisfactory agreement. A great danger is that
there are cases where the agreement between theory and experiment is rather
good but the model structure is not the right one. If one wants to have confi-
dence in the result, it is important to have a very good agreement, or, in other
words, one has to get everything right before one knows that one has anything
right.

4.2 Photoelectron spectroscopy

The most prominent tool for investigations of the electronic structure is pho-
toelectron spectroscopy, or photoemission. [66, 67] While the origins of LEED
go back to the 1930ies, those of photoemission are in the late 19th century.
A first highlight was the Nobel prize for Einstein in 1921 for his work on the
photo-electric effect. [15] Photoemission as a spectroscopic method was first
used by Siegbahn, exploiting x-rays as incident radiation. [68] Today, photons
with energies in the vacuum-ultraviolet range, supplied by storage rings, are
used for investigating the band structure
A photoemission experiment is characterized by a number of parameters
(Fig. 4.12).
Incident light: incidence direction, photon energy ~ω, polarization

Photoelectrons: detection direction, kinetic energy, spin polarization.

Light with photon energy ω impinges onto a surface (grey; black dots: surface
atoms). The components of its polarization are As and Ap . The detection
direction of the photoelectrons is denoted e− (here: in the yz-plane).
In an energy diagram, an electron (filled circle) is excited into a state with
energy larger than the vacuum energy Evac , leaving behind a so-called hole
(empty circle) below the Fermi energy EF . The electrons, that is either from
the valence-band regime (grey area) or from the core-level regime (horizontal
lines) leaves the sample and is subsequently detected.
From the energy-space scheme is becomes clear that the photoemission can be
presented by three steps (‘three-step model’ [69]). In the beginning, electronic
states are occupied up to the Fermi energy EF . (i) The incident photon with
energy ω excites an electron into an unoccupied state. (ii) The electron is moved
to the surface. (iii) It can leave the sample if its energy is larger than the vacuum

91
 Figure 4.12. Photoelectron spectroscopy.

level Evac . The remaining sample contains one electron less. The minimum
photon energy for detection is given by the work function Φ = Evac − EF ,
typically about 5 eV.
A simple theoretical description is based on Fermi’s golden rule. The tran-
sition probability wf i from the initial state |Ψi i with energy Ei into the final
state |Φ(kk , Ef )i with energy Ef = Ei + ω and momentum kk is given by
2
wf i = hΦ(kk , Ef )|∆|Ψi i δ(Ef − Ei + ω). (4.12)
The transition is mediated by the dipole operator ∆. The photocurrent j of
electrons with kinetic energy Ekin and kk is obtained by summing over all initial
states,
p X
j ∼ Ekin wf i (4.13)
i

The detection angles ϑe and ϕe as well as the kinetic energy determine the
in-plane momentum kk ,
 
p cos ϕe
kk = 2Ekin sin ϑe . (4.14)
sin ϕe
Today, angle-resolved (or kk -resolved photoelectron spectroscopy (ARPES))
with VUV light is the most successful method for investigating the electronic
structure in the valence-band regime [67, 70–72]. A question remains how to
interpret the experimental results.
Single-particle theory of photoemission
The theoretical basis for ARPES as spectroscopic method needs several features
to be taken into account [73–75].
Electronic structure The electronic structure comprises both occupied and un-
occupied states for the system truncated by the surface. Peak positions
(i. e. energy and wavevector) in the spectra are essentially determined by
band structure.

92
Transition matrix elements The intensities are determined by the transition
matrix elements which involve the wavefunctions of the initial and the
final states as well as the dipole operator. In specific cases (depending
on the symmetry of the surface, light polarization, etc), transitions are
forbidden, giving rise to so-called selection rules.

Surface effects These comprise not only the electronic structure at the surface,
as compared to that in the bulk (cf. the section on surface states) but also
inelastic processes (as in LEED) which make ARPES surface sensitive.
The starting point of photoemission theory is the Green function formalism
for non-equilibrium systems [76, 77]. A semi-infinite solid is illuminated by
monochromatic light. Its Hamiltonian then reads
 2
1 1
H(t) = p + A(r, t) + V (r), (4.15)
2 c
with the potential V (r). The light is described by the vector potential A.
Perturbation theory then yields the current density j at the detector position
R

j ∼ − 2(E + ω)hΦ?0 |G− (E + ω) ∆ Im G+ (E) ∆† G+ (E + ω)|Φ?0 i, (4.16)

p

where |Φ?0 i is a plane wave.

This expression can be interpreted as follows (Fig. 4.13). The photoelectron
states |Φ?0 i with energy E + ω is propagated by the retarded Green function G+
from the detector into the solid. Subsequently, the dipole operator de-excites
the electron into the initial states with energy E described by Im G+ ). Then,
it is excited again into the outgoing state hΦ?0 |G− via ∆. From this it is clear

Figure 4.13. Feynman diagram of photoemission.

the the photocurrent contains information on both the electronic structure of

the occupied and the unoccupied states.
The photoelectron is subject to inelastic processes, for example, scattering at
other electrons, at phonons and at defects. This leads to a finite escape depth

93
(as in LEED), making ARPES surface sensitive. Inelastic processes show up as
self-energy, that is a complex potential that is added to the crystal potential
V (r) [78]. Its real part shifts the energies of the electronic states, implying
that maxima in the photoemission spectra appear shifted with respect to those
expected from electronic-structure calculations. Its imaginary part describes
finite lifetimes and gives to rise to peak broadening. In the spectral density
A(E) = −Im Tr G(r, r; E)/π the self-energy Σ shows up as

δ(E − Em )
 real energy E
X Γ 1
A(E) = 2 2 complex energy E + iΓ .
 π (E−Em ) +Γ
m  Im Σ(Em )
 2 2
[E−Em −Re Σ(Em )] +[Im Σ(Em )]
general case
(4.17)

There is rather close relation of ARPES and LEED because the final state in
ARPES (that of the outgoing photoelectron) is that of a time-reversed LEED
experiment. This state obeys
Z
3
Φ(R, E + ω) = exp(ik · R) + G+ (R, r 0 ; E + ω) V (r 0 ) exp(ik · r 0 ) dr0 ,
(4.18)
p
[k = 2(E + ω)] and corresponds to the superposition of an incoming plane
wave hR|Φ0 i = exp(ik · R) and outgoing partial waves, the latter given by the
integral. The Green function acts as a propagator, here from the interior of the
sample to the detector.
In spin-resolved or spin-polarized ARPES (SPARPES) also the spin degree
of freedom of the electrons is measured. Hence, the spin-dependent electronic
structure of magnetic samples can be investigated in detail [79]. Magnetic
dichroism is the change of the photocurrent upon reversing the magnetization
orientation of a magnetic sample and is due to spin-orbit coupling [80]. Us-
ing electronic states from the valence bands does not provide element specific
results. Therefore, one employs core levels as initials states using x-rays as
electromagnetic radiation (XPS). Images of surface can be obtained using the
photoelectron microscope (PEEM), this way combing high resolution with the
ability detect element-selective (as in XPS).
Examples
A state-of-the-art ARPES experiment is shown in Fig. 4.14, using a 2D electron
detector allowing for simultaneously resolving the energy and the wavevector.
This gives rise to the energy-distribution and momentum-distribution curves,
conveniently shown in a single diagram.
The wavevector-resolved electronic structure can be obtained by scanning
both energy and in-plane wavevector. The occupied electronic states, with
energy less than the Fermi energy, are probed by conventional ARPES. The
unoccupied electronic states, with energy larger than the Fermi energy, are
probed by inverse photoelectron spectroscopy (ARIPES). Here, an incoming
beam of electrons is de-excited into a state above the Fermi energy while photons

94
 Figure 4.14. Set-up of a photoemission experiment.

Figure 4.15. Photoemission from Cu. Confer the Cu band structure shown in
Fig. 3.17. Large PE intensity shows up in bright areas.

leave the sample and are detected. This allows band-mapping of energy ranges
on both sides of the Fermi level, as is shown for Si(100)-1×2-K in Fig. 4.16.
Problem 42. How can one prove experimentally that an observed feature in the ARPES
is due to a surface state?
First, a surface state must lie in a band gap of the bulk system. Second, it must show the
2D periodicity of the surface. Third, it is sensitive to adsorbates. One covers the surface
with a small amount of adatoms that suppress the surface state but leaves the bulk states
unaffected. Comparing then the PE spectra for the clean and the uncovered surface gives
hints on the surface state. For example, bulk-related features should decrease in intensity
but leave the shape (almost) unaffected. The surface-state feature should also decrease

95
Figure 4.16. Band-structure determination for Si(100)-1×2-K. Left: Photoemission
intensities for ARPES and ARIPES shown for various polar angles. Right: Band
structure with PE maxima (dots).

in intensity but due to the modification of the surface-barrier potential its energy might
change. 2
When discussing surface states, image potential states were also briefly
sketched. These states have an energy between the Fermi energy and the vac-
uum level (Fig. 4.17) and can, thus, be probed by ARIPES. An alternative
is two-photon photoemission (2PPE), in which two ultra-short photon pulses
are used to excited the electron. In a first step, the electron is excited into the
image-potential state and, in the second step, into the outgoing photoelectron
state. By changing the delay between the two photon pulses, the lifetime of the
image-potential state can be obtained (Fig. 4.18).
Problem 43. Why is the lifetime of image-potential states finite? Give reasons.
For example, scattering at phonons and defects, disorder. In the many-electron picture,
the state |N − 1, ei is an excited state is an excited state, not the ground state |N i. Here,
|N − 1, ei denotes the N -electron state of N − 1 electrons with energy less than the Fermi
energy and one electron in the image-potential state. 2
Using photons with high energy, for example x-rays, the core levels can be
probed, a method known as x-ray photoelectron spectroscopy (XPS). It be-
came fashionable in the investigation of magnetic samples. in x-ray magnetic
circular dichroism (XMCD), the intensity depends on the orientation of the
magnetization using circularly polarized light, allowing to conclude on the mag-
netic moment in an element-specific manner. This effect is due to the spin-orbit
interaction, coupling the angular momentum and the spin of the electron. The
transition matrix elements depend then on the relative orientation of helicity

96
 Figure 4.17. Formation of image-potential states at a metal surface.

(of the incident radiation) and magnetization.

Problem 44. Can you imagine why spin-orbit coupling is necessary for magnetic dichroism?
Without the spin part and the space-angular-momentum part of a wavefunction are not
coupled, meaning that any direction can serve as spin quantization axis. The Spin-orbit
interaction couples these parts and fixed the spin-quantization direction. 2
Element-specific probing is due to the difference of the binding energy of the
core-levels of the atomic species. This allows for detailed investigations of the
magnetic structure of alloys, in particular if combined with a local probe such
as PEEM (Fig. 4.20).

4.3 Other scattering techniques

Both LEED and ARPES are essentially determined by scattering of the elec-
trons at the atomic positions. However, there is vast amount of other scattering
techniques that will be briefly sketched in the following.
EXAFS
The techniques of Extended X-ray Absorption Fine Structure (EXAFS) have
been made possible by the construction of synchrotron radiation (SR) sources.
SR has a continuous energy spectrum. In combination with a monochromator it

97
Figure 4.18. 2PPE for Cu(100). Left: intensities for various delay times. Right:
Band structure with the dispersion of the n = 1 and n = 2 image-potential states.

provides a tunable x-ray source. The (S)EXAFS measurements involve scanning

the photon energy around the absorption edges of the atoms in a material or on
a surface. The fine structure in the absorption cross section gives information
about the neighbors of the emitting atoms. The main advantages of these x-ray
absorption techniques are that they work for materials where long-range order
is not present and that at least the nearest neighbor distances can be obtained
with rather high precision.
Fig. 4.21 shows an example for an EXAFS spectrum. Such a spectrum can
be taken by exposing a thin film of material to x-rays and simply measuring
the transmission through the film. The x-ray absorption of Cu is plotted vs
the photon energy. As the energy reaches the K-edge the absorption increases
steeply because of the the possibility to excite the K-electrons. Above the edge
the absorption shows a slow decrease due to the matrix element. On this slow
decrease rapid oscillations in the cross section are superimposed. These are the
EXAFS oscillations.
The physical origin of these oscillations is quite easy to understand. The
absorption cross section is given by Fermi’s golden rule and thus by a matrix
element as in photoemission. The initial state is simply the localized core state.
The final state is the extended state of the outgoing electron wave, including
all the multiple scattering processes. One can now think about the EXAFS
oscillations in the following way. At low kinetic energies, from zero to a few
hundred electron volts, the cross section for the back-scattering of the electrons
from the neighbor atoms is rather high. These back-scattered waves have to be
added coherently to the outgoing wave and this directly influences the final state
at the emitter and thus the matrix element. The interference from outgoing and
back-scattered waves changes with a periodicity given by the nearest neighbor
distance. The effect is illustrated in Fig. 4.22 This simple picture works only for

98
Figure 4.19. XMCD. Left: Set-up, right: x-ray absorption for two different helicities
of the light (σ ± ) and difference spectrum (‘dichroic signal’).

electron energies which are not too close to the edge. Below kinetic energies of
50 eV or so the oscillations contain resonant absorption from the valence states.
Photoelectron diffraction
The principle in photoelectron diffraction (PED) is given in the Fig. 4.23. One
measures the intensity of a core-level line as a function of energy and emission
angle. The electron can reach the detector on a direct path or it can be scattered
at the atoms surrounding the emitter and then reach the detector. The final
state is given by interference between the direct and the scattered components.
This interference depends on the path-length difference and on the scattering
phase shifts. This, in turn, depends on the electron wavelength (kinetic energy),
on the position of the detector and on the position of the emitter with respect
to the scatterers.
By varying the position of the detector or the kinetic energy of the emitted
electrons the interference conditions are changed. Normally the photoemission
intensity is recorded as a function of kinetic energy or emission angle. The
modulation in the intensity which results from the change in the interference
conditions can be used to extract geometrical information.
The technique is related to surface EXAFS (SEXAFS)and LEED. SEXAFS
is similar to PED in that the electron source is also a core level from an atom at
the surface. Here the interference between the outgoing and the backscattered
waves is of major importance. The backscattering changes the final state inten-
sity at the core and this in turn changes the absorption cross section. SEXAFS
is very similar to PED with the difference that the emitter itself is the detector.
In a simple picture, it can be viewed as an angle-integrated PED experiment.
The PED modulations do, of course, also contain the SEXAFS part, i. e. the
modulations in the absolute cross section. This is, however, not a problem

99
 Figure 4.20. Element-specific XMCD for a Co/Ni/Cu trilayer system.

Figure 4.21. X-ray absorption of Cu in the vicinity of the K-edge.

because the SEXAFS modulations are much weaker (1-3

The big difference between these two techniques and LEED is that the elec-
tron source in the latter is not in the system itself but in the far-field. It is not
possible to pick an atom in the adsorbate layer and make it special by consid-
ering just the core-level intensity from that atom. All the atoms are of equal
importance as scatterers. The information about the adsorbate layer relative to
the substrate is only contained the in the wave-part scattered by the adsorbate
and its interference with the wave-part scattered by the substrate. In PED and
SEXAFS the position of the adsorbate relative to the substrate is contained in
all scattering pathways. A further difference is that LEED is a true diffraction
technique which relies on the long-range order in the adsorbate layer. PED
and SEXAFS only probe the local structure around the adsorbate. This is a

100
Figure 4.22. Schematic illustration of the interference leading to the EXAFS oscilla-
tions.

Figure 4.23. Principle of photoelectron diffraction.

great advantage for disordered systems. Multiple scattering is very important

in LEED, less important in PED and least important in SEXAFS. The reason is
the 1/r fall-off of the spherical wave from the electron source and, in SEXAFS,
also from the electron scatterer. PED contains the 1/r effect once, SEXAFS
twice. Finally, SEXAFS and PED have chemical sensitivity: it is possible to
pick out on specific atom in the adsorbate. Indeed, if there are two, say, carbon
atoms in a different chemical environment it is even possible to pick one of them
as an emitter because of the chemical shift in the XPS lines (only in PED).
Let us illustrate this by a simple example, the adsorption of the acetate
species on Cu(110). A O 1s diffraction spectrum is compared to theoretical
calculations for different adsorption sites and distances from the surfaces. The
best agreement is found for a bridge site (see Fig. 4.24).
The big advantage of PED with respect to LEED lies in the dependence of
local order only. However, if a molecule is adsorbed on more than one site at the

101
Figure 4.24. O 1s diffraction data to determine the adsorption site of the acetate
species on Cu(110). The R-factor has the same meaning as in a quantitative LEED
analysis.

same time more information is needed to obtain the right structural parameters.
In such a case many modulation functions are measured and compared to theory.
A further advantage of PED is its chemical sensitivity. The positions of O
and C in the above example can be determined separately considering the O
1s and C 1s emission. Moreover, the two C atoms are in a chemically different
environment. This means that the binding energies are also different (chemical
shift). Hence the PED from both C atoms can be measured independently and
the modulations are quite different. This is illustrated in Fig. 4.25.

Figure 4.25. Modulation functions for the two chemically different C atoms in the
acetate species on Cu(110).

4.4 Scanning tunneling spectroscopy

The techniques discussed so far did not show atomic resolution but provided
an average of the surface. In view of nanostructures, it is also clearly desirable

102
to have a real-space microscopic technique which can image the structure of
surfaces on a truly atomic scale. Field emission microscopy is a possibility with
a rather limited range of possible applications. First the advent of scanning
tunneling microscopy has made a real space atomic scale view on most surfaces
possible. But the section should start with a clear warning. STM does not
measure the structure of surfaces but, as we shall see below, the electronic
structure. It is essential to keep this in mind, in particular when working with
semiconductors and insulators.
Basic principle and experimental set-up.
Scanning tunneling microscopy is based on the quantum mechanical effect of
tunneling illustrated in Fig. 4.26 The wavefunctions at the Fermi level expo-
nentially leak out of the metal with an inverse decay length of
1√
κ≈ 2mΦ, (4.19)
~
where m is the mass and Φ is the (local) workfunction. If now two metals
are brought in close contact and a small voltage is applied between them, a
tunneling current can be measured which is
I ∝ e−2κd , (4.20)
where d is the distance between the conductors. The important message here,
and the reason why STM works, is the exponential dependence of the tunnel
current on the distance between the conductors. We will come back to this
several times in the following.
Figure 4.27 shows the principle setup for an STM. It consists of a sharp tip,
very close to the sample, which can be moved with high precision using three
mutually orthogonal piezoelectric transducers (PET). A small voltage is applied
between the tip and the sample and the current is measured. Typical values
for the tunneling voltage are from a few mV to several V or so and for the
current from 0.5 nA to 5 nA. The tip-sample distance is a few Ångström. The
tunnel current depends very strongly on this distance. A change of 1 Å causes
a change in the tunneling current by a factor of ten.
Most STM topography studies are performed in the so-called constant current
mode. The tip is scanned across the surface by the X and Y PETs. The Z PET
is in a feedback loop which applies a correction voltage to the Z voltage in
order to maintain a constant tunneling current as the XY position of the tip
is changed. Since the piezo extension is proportional to the applied voltage,
this correction voltage is a direct measure for the change in Z the tip has to
perform in order to follow the contours of the sample and it can be used as
an “image” of the sample when displayed as a function of X and Y voltages.
The exponential decay of the tunneling current with distance from the sample
is crucial for this operation: even if we can only keep the current stable by
10 percent or so, this will still mean that we have a very high precision in the
Z measurement because a small uncertainty in the current means virtually no
uncertainty in the distance.
While the principle of STM operation is simple, its practical realization faces
some formidable difficulties. The first is that the tip has to be brought at a

103
Figure 4.26. (a) Exponential leakage of the wavefunctions from a conductor into the
vacuum. (b) Application of a voltage and tunneling between two conductors because
of the overlap of the wavefunction tails. Φ is the workfunction.

distance of a few Ångström from the surface and has to be stabilized there with
sub-Ångström stability. This process has to be performed in the presence of
mechanical vibrations and thermal drift. The vibrational problem is solved by
vibrational insulation, like suspensions by springs or the use of a support frame
which has a resonance frequency very different from the usual noise frequencies
of the environment. The only way to get completely rid of the thermal drift
problem is to stabilize the whole microscope and sample at the same (low)
temperature. The most stable STMs today are working in a UHV vessel which
is placed inside a dewar filled with liquid Helium. The next practical problem is
to make an atomically sharp tip. Different techniques are used like cutting and
etching and most tips are made off W or Ir. Again, the exponential decay of the
tunneling current helps: one can hope that there is one single atom sticking out
a little further than all the others on an otherwise rather blunt tip. Then most
of the tunneling will happen through this atom and it will be possible to obtain
images with atomic resolution. Unfortunately, it is not possible to prepare a
tip a controlled and reproducible way. In particular, one has no control about
the chemical nature of the outermost atom of the tip but the images depend a
lot on this.

104
 Figure 4.27. Schematic construction principle of an STM.

Basic theory of operation

The calculation of STM images is rather difficult. For the sample one can
make reasonable total energy calculations for a perfect surface. For the tip
such calculations are not so easy because of the unknown shape. Even if one
could calculate the wavefunctions for each system, sample and tip, separately
bringing them together would render the problem complicated because now
one has a new system which totally lacks the translational invariance which
was needed to solve the surface problem. The last hurdle is circumvented by
Bardeen’s suggestion to basically use the unperturbed wavefunctions of tip and
sample and to consider just the tunneling between them. Tersoff and Hamann
have developed a simple and useful STM theory based on this approximation.
They furthermore assume that the tip can be approximated by an s-wave state
(i. e. that its end has spherical symmetry) and that temperature and tunneling
voltage are very low. This results in

I ∝ Ntip (EF )R2 e−2κd ρsamp (r tip ; EF ), (4.21)

where R is the radius of the tip, Ntip (EF ) is the tip density of states at the
Fermi level, r tip is the center position of the tip and
X
ρsamp (r; E) = |Ψi (r; En )|2 δ(E − En ) (4.22)
i

is the local density of sample states. The only z dependence in these equations
is in the wavefunctions such that we recover our exponential decay into the
vacuum, as we must. These equations give a simple interpretation for constant-
current mode STM images. If we scan across the surface at constant current,
adjusting just the distance of the tip over the surface, we basically follow the

105
contours of sample density of states at the Fermi-level. Note, however, that this
statement is only correct if we assume that there are no lateral changes in the
workfunction of the sample.
In practice, the applied voltage is not always very small. Especially for semi-
conducting materials a small voltage can be impossible because there are no
carriers in the fundamental bandgap which can be involved in the tunneling.
The voltage can be chosen to be both negative or positive. This means that
STM can look at both, occupied and unoccupied states of the sample depending
on the bias voltage.
We will give a better theoretical description of the STM in the forthcoming
sections. Problems in STM theory are the following.

Loss of 2D translation invariance Usual methods for electronic-structure cal-

culations rely on translational invariance. The STM tip breaks the 2D
translational invariance of the otherwise planar tunnel junction. Hence,
one uses either a supercell approach, by artificially increasing the size of
the 2D unit cell each containing a STM tip, or applies a so-called real-
space approach that does not rely on any spatial symmetry.

Tip-sample interaction Approaching the STM tip close to the surface can
change the electronic structure of the sample in the vicinity of the apex.
For very close tip-sample distances, even new electronic states can be
induced (‘tip-induced states’). Or, the atomic positions of the sample
change with respect to that of the surface (without tip). Thus, the theory
should describe the interacting system, that is sample and tip. In simpler
theories, like the Tersoff-Hamann or the transfer-Hamiltonian approaches,
the tip-sample interaction is neglected. Here, the electronic structures of
sample and tip are computed separately.

Figure 4.28. Tip sample interaction. The electronic structure of the sample is influ-
enced by the STM tip (right), as compared to the sample without tip (left).r

Finite bias voltage (non-equilibrium problem) Usual electronic-structure cal-

culations apply to equilibrium conditions. The applied bias voltage, how-
ever, results in a current flowing through the tunnel junction, that is a
non-equilibrium system in a stationary state.

106
A few examples
The STM can provide atomic resolution, as is demonstrated for the recon-
structed Si(111)-7×7 surface (Fig. 4.29). Here, a sharp STM tip is necessary

Figure 4.29. STM image of the Si(111)-7×7 surface.

(Fig. 4.30).

Figure 4.30. STM tip.

Instead of the topography one can also look at the local electronic and mag-
netic structure. For example, spin resolution can be obtained by use of a mag-
netic tip (Fig. 4.31). The tunnel current then depends on the relative orientation
of the magnetization of sample and tip (Fig. 4.32).
By fixing the spatial position of the tip and ramping the applied voltage,
one can record the electronic structure, a method called scanning tunneling
spectroscopy (STS).

107
 Figure 4.31. Spin-resolved STM.

Figure 4.32. Spin polarized in-plane contrast of Fe-islands of different thickness

deposited on W(110); since the W(110) is covered with a monolayer of Fe, also the
area between the islands exhibits a contrast.

Figure 4.33. Top: topography with atomic resolution on graphite. Bottom: spec-
troscopy on the band gap of a superconductor.

108
5 Electronic transport in
nanostructures
Electronic transport in nanostructures often depends on the spin of the elec-
trons. Hence, we will first introduce a basic description of spin.

5.1 Spin
The angular-momentum operator l̂ is defined as

l̂ ≡ r × p̂. (5.1)

Its components obey

[ˆli , ˆlj ] = ˆli ˆlj − ˆlj ˆli = i~ˆlk , i, j, k cyclic. (5.2)

Problem 45. Prove the above rule using the commutation rule for spatial components and
momentum, page 15.
With [p̂x , x] = −i~ and [p̂x , y] = 0 is follows

[ˆlx , ˆly ] = ˆlx ˆly − ˆly ˆlx

= (y p̂z − z p̂y )(z p̂x − xp̂z ) − (z p̂x − xp̂z )(y p̂z − z p̂y )
= y p̂z z p̂x − y p̂z xp̂z − z p̂y z p̂x + z p̂y xp̂z
−z p̂x y p̂z + z p̂x z p̂y + xp̂z y p̂z − xp̂z z p̂y
= y p̂z z p̂x + z p̂y xp̂z − z p̂x y p̂z − xp̂z z p̂y
= y(−ih+̄z p̂z )p̂x + zxp̂y p̂z − zy p̂x p̂z − x(−i~ + z p̂z )p̂y
= −ihȳ p̂x + yz p̂z p̂x + zxp̂y p̂z − zy p̂x p̂z + i~xp̂y − xz p̂z p̂y
= −ihȳ p̂x + i~xp̂y
= i~ˆlz .

And analogously for the other components. 2

Spin is an observable that obeys the commutation rules of an angular mo-
mentum [81],

ŝx ŝy − ŝy ŝx = i~ŝz , (5.3)

and equivalent for cyclic permutations of x, y, and z. From this it follows

immediately, that measuring one component of the spin makes it impossible to
measure another component simultaneously (because they do not commute).
Defining
~
s≡ σ, (5.4)
2

109
one obtains a matrix representation, namely the Pauli matrices,
     
0 1 0 −i 1 0
σx ≡ , σy ≡ , σz ≡ . (5.5)
1 0 i 0 0 −1
These matrices have the following properties.

σi2 = 1, (5.6a)
σi σj = iσk , i, j, k cyclic, (5.6b)
σi σj = −σj σi i 6= j, (5.6c)
det(σi ) = −1, (5.6d)
tr(σi ) = 0, (5.6e)
σi σj = iijk σk + δij , (5.6f)

where ijk is the Levi-Civita symbol.1 The 2 × 2 unit matrix and the hermitian
and unitary Pauli matrices form a basis for the vector space of 2 × 2 complex
matrices.
The eigenspinors χ± of σz are

σz |χ+ i = |χ+ i, (5.7a)

− −
σz |χ i = −|χ i, (5.7b)

with
   
+ 1 − 0
|χ i = , |χ i = . (5.8)
0 1

For spinors, see Refs. [82] and [83].

Problem 46. Compute the eigenspinors of σx and σy .
For σx follows for the secular equation
  
−λ 1 α
= 0.
1 −λ β

The characteristic polynomial reads 0 = λ2 − 1 Hence, λ = ±1. For λ = +1 follows

  
−1 1 α
= 0,
1 −1 β
which is solved by
 
1 1
|χ+ i = √ .
2 1
For λ = −1 follows
  
1 1 α
=0
1 1 β
1
The Levi-Civita symbol is given by
8
<1
> even permutation of i, j, k
ijk = −1 odd permutation of i, j, k .
>
0 otherwise
:

110
which is solved by
 
1 1
|χ− i = √ .
2 −1

For σy follows for the secular equation

  
−λ −i α
= 0.
i −λ β

The characteristic polynomial reads also 0 = λ2 − 1 Hence, λ = ±1. For λ = +1 follows

  
−1 −i α
= 0,
i −1 β

which is solved by
 
1 1
|χ+ i = √ .
2 i

For λ = −1 follows
  
1 −i α
=0
i 1 β

which is solved by
 
1 1
|χ− i = √ .
2 −i
2
A general state is the given by
 
a1
|χi = = a1 |χ+ i + a2 |χ− i (5.9)
a2

and

hχ| = a?1 a?2 = a?1 hχ+ | + a?2 hχ− |.

(5.10)

Hence, normalization requires

 

a1
?
1 = hχ|χi = a1 a2 ? = |a1 |2 + |a2 |2 . (5.11)
a2

Problem 47. Show that ŝ2 has the eigenvalue s(s + 1)~2 = 3~ /4.
2

It follows with the matrix representation that

~2 2 ~2 2 ~2 3~2
ŝ2 = σ = (σx + σy2 + σz2 ) = (1 + 1 + 1) = 1.
4 4 4 4
2
Thus, ŝ2 is represented by 3~4 times the 2 × 2 unit matrix. Since the latter has eigenvalues
2
1, that of ŝ2 is 3~4 . 2
2
Problem 48. Show that ŝ commutes with ŝx , ŝy , and ŝz .
With the above result,
3~2
ŝ2 = 1,
4

111
and noticing that the 2 × 2 unit matrix commutes with all 2 × 2 matrices, it is evident that
ŝ2 commutes with ŝx , ŝy , and ŝz .
Or, using

[A + B, C] = [A, C] + [B, C]

and

[AB, C] = [A, C]B + A[B, C],

the long way

[ŝ2 , ŝx ] = [ŝ2x + ŝ2y + ŝ2z , ŝx ]

= [ŝ2y + ŝ2z , ŝx ]
= [ŝ2y , ŝx ] + [ŝ2z , ŝx ]
= ŝy [ŝy , ŝx ] + [ŝy , ŝx ]ŝy + ŝz [ŝz , ŝx ] + [ŝz , ŝx ]ŝz
= −i~ŝy ŝz − i~ŝz ŝy + i~ŝz ŝy + i~ŝy ŝz
= 0.

2
Now consider the statement, “the spin is along the z-direction”. In a vectorial
representation, the z-component of the spin vector is along the x-axis (with
length ~/2). Thus, the spin vector (with length 3~2/4) lies on a cone. The x-
and y-components are unknown.
From the above it follows that an electron beam in which all spin directions
are equally likely and an electron beam in which half of the spins are parallel
or antiparallel aligned to some arbitrary reference direction cannot be distin-
guished. This is because a measurement of any spin component gives half of
the spins parallel or antiparallel to the specified direction.
Let e = (ex , ey , ez ) the unit vector, the direction of which is determined by
the angles ϑ and ϕ,

ex = sin ϑ cos ϕ, ey = sin ϑ sin ϕ, ez = cos ϑ. (5.12)

We are now calculating the eigenvalues of

(σ · e)|χi = λ|χi. (5.13)

It follows with
           
a a2 a1 −ia2 a1 a1
σx 1 = , σy = , σz = (5.14)
a2 a1 a2 ia1 a2 −a2

that
a1 cos ϑ + a2 sin ϑ e−iϕ
 
(σ · e)χ = = λχ. (5.15)
a1 sin ϑ eiϕ − a2 cos ϑ

Hence, from the characteristic polynomial

− cos2 ϑ + λ2 − sin2 ϑ = 0, (5.16)

112
or

λ = ±1. (5.17)

Thus, we can write

ϑ ϑ
a1 = cos , a2 = sin eiϕ for λ = +1, (5.18a)
2 2
ϑ ϑ
a1 = sin , a2 = − cos eiϕ for λ = −1. (5.18b)
2 2
In the special case ϑ = 0 (or π) and ϕ = 0, that is e along the z-axis, yields the
eigenspinors of σz , namely χ± . For ϑ = π/2 and ϕ = 0 (spin along the x-axis),
gives
   
1 1 1 1
√ , √ . (5.19)
2 1 2 −1

Compare this result with your solution of the above problem.

Adding spin states with opposite spin (coherently) does not produce cancel-
lation of spin but a spin in another direction. For example,
     
1 1 1 0 1 1
√ +√ =√ . (5.20)
2 0 2 1 2 1

This is analogy to the (coherent) superposition of left-handed and right-handed

circularly polarized light which gives not unpolarized but linearly polarized
light.
The polarization P is defined as the spin expectation value,

P ≡ hσi = hχ|σ|χi. (5.21)

With the above definitions we have

Px = sin ϑ cos ϕ, (5.22a)

Py = sin ϑ sin ϕ, (5.22b)
Pz = cos ϑ, (5.22c)

as expected. Further |P | = 1. For an unnormalized spinor,

hχ|σ|χi
P ≡ . (5.23)
hχ|χi

We define further the density matrix % by

% ≡ |χihχ|. (5.24)

Or explicitly,

|a1 |2 a1 a?2
 
%= . (5.25)
a2 a?1 |a2 |2

113
From this follows that (for each component)

tr(ρσ)
P = . (5.26)
tr(ρ)
Problem 49. Prove that Px = tr(%σx ) for a normalized spinor.
It follows
|a1 |2 a1 a?2
  
0 1
tr(%σx ) = tr
a2 a?1 |a2 |2 1 0
? 2
 
a a |a1 |
= tr 1 22
|a2 | a2 a?1
= a1 a?2 + a2 a?1
= 2Re (a1 a?2 ).

For λ = +1, we have a1 = cos(ϑ/2) and a2 = sin(ϑ/2) exp(iϕ). Thus,

tr(%σx ) = 2Re cos(ϑ/2) sin(ϑ/2)eiϕ

= 2 cos(ϑ/2) sin(ϑ/2) cos(ϕ)
= sin(ϑ) cos(ϕ).

For λ = −1, we have a1 = sin(ϑ/2) and a2 = − cos(ϑ/2) exp(iϕ). Thus,

tr(%σx ) = − sin(ϑ) cos(ϕ),

that is the spin is aligned with −e. Comparison with eq. (5.22a) establishes the statement.
2
The density matrix can be expressed in terms of polarization,
 
1 1 + Pz Px − iPy 1
%= = (1 + P · σ) . (5.27)
2 Px + iPy 1 − Pz 2

Consider a partially polarized beam of electrons, with the n-th electron in

state |χ(n) i. With all states being normalized, it follows
X X
P = w(n) P (n) = w(n) hχ(n) |σ|χ(n) i. (5.28)
n n

The statistical weight of state |χ(n) i is w(n) . If N (n) electrons are in state |χ(n) i,
then
N (n)
w(n) = P (n)
. (5.29)
nN

The density matrix reads

X
%= w(n) |χ(n) ihχ(n) |, (5.30)
n

from which follows

X X
P = w(n) hχ(n) |σ|χ(n) i = w(n) P (n) . (5.31)
n n

114
 If there are only two states, namely |χ↑ i and |χ↓ i, then Px = Py = 0. For
the z-component we have
X
Pz = w(n) Pz(n) (5.32)
n
= w Pz↑ + w↓ Pz↓
↑
(5.33)
N↑ N↓
= 1+ ↑ (−1) (5.34)
N↑ + N↓ N + N↓
N↑ − N↓
= ↑ . (5.35)
N + N↓
Problem 50. In an ensemble of 100 electrons there are 80 with spin in the +z-direction
and 20 with spin in −z-direction. How large is the polarization of the ensemble? How many
electrons of the ensemble can be regarded as unpolarized and as completely polarized?
With N ↑ = 80 and N ↓ = 20, the polarization is
N↑ − N↓ 80 − 20 60
P = ↑ ↓
= = = 0.6.
N +N 80 + 20 100
Hence, 60 % of the ensemble is totally polarized (spin-up), 40 % is unpolarized (that is an
incoherent superposition of 20 spin-up and 20 spin-down electrons). 2
Problem 51. Consider the coherent superposition of two electron beams,

hz|Ψi ≡ α↑ eik↑ z |χ↑ i + α↓ eik↓ z |χ↓ i.

Compute and sketch P (z) of |Ψi.

The spinor |Ψi has components
a1 = α↑ eik↑ z ,
a2 = α↓ eik↓ z .
The density matrix
|a1 |2 a1 a?2
 
%=
a2 a?1 |a2 |2
thus reads
|α↑ |2 α↑ α↓ ? ei(k↑ −k↓ )z
 
%= .
α↑ α↓ e−i(k↑ −k↓ )z
?
|α↓ |2
The z-component of the spin polarization is given by
tr(%σz )
Pz =
tr(%)
|α↑ |2 − |α↓ |2
= ,
|α↑ |2 + |α↓ |2
that is constant in space. For Px it follows
tr(%σx )
Px =
tr(%)
Re (a1 a?2 )
= 2
|α↑ |2 + |α↓ |2
2
= Re α↑ α↓ ? ei(k↑ −k↓ )z .
|α↑ |2 + |α↓ |2

115
Assuming for simplicity real coefficients α↑ and α↓ , it follows further
2α↑ α↓
Px = cos[(k↑ − k↓ )z].
|α↑ |2 + |α↓ |2
For Py one obtains

tr(%σy )
Py =
tr(%)
Im (a1 a?2 )
= −2
|α↑ |2 + |α↓ |2
2α↑ α↓
= sin[(k↑ − k↓ )z].
|α↑ |2 + |α↓ |2
The length of the spin polarization is given by

P = Px2 + Py2 + Pz2

1
4α↑ α↓ cos2 [(k↑ − k↓ )z] + 4α↑2 α↓2 sin2 [(k↑ − k↓ )z] + (α↑2 − α↓2 )2
 2 2
= 2 2 2
(α↑ + α↓ )
1  2 2
= 2 2 2
4α↑ α↓ + α↑4 − 2α↑2 α↓2 + α↓4
(α↑ + α↓ )
1  4
= 2 2 2
α↑ + 2α↑2 α↓2 + α↓4
(α↑ + α↓ )
= 1.

The superposition is fully polarized.

The spin polarization is shown in Fig. fig:spinsuperpos. It precesses about the z-direction,
an effect being named spin motion. It shows up for example in spin-resolved LEED and
ARPES from ultrathin magnetic films. [84] 2

Figure 5.1. Spin polarization of two coherently superposed plane waves (α↑ = 0.8,
α↓ = 0.2, k↑ = 1.2 Bohr−1 , k↓ = 0.8 Bohr−1 ).

116
5.2 Important quantities in mesoscopic transport
Quantum effects in mesoscopic structures show up at characteristic scales in
space, time, and energy. These will be introduced in this section.

Fermi wavelength λF The Fermi wavelength is the de-Broglie wavelength of

electrons at the Fermi level EF . It is, as usual, related to the Fermi
wavenumber kF by λF = 2π/kF . Size quantization effects show typically
up on the length scales comparable to λF .
For free electrons in d dimensions, λF depends on the electron density n.
One obtains

3/2 π 1/3 d = 3


2

 q 3n
λF = 2π
n d=2. (5.36)

4

d=1
n

Problem 52. For free electrons, there is a single band but in a real solid several bands
cross the Fermi level. Can one speak in the latter case of the Fermi wavelength?
Discuss the case of Cu.
Cu is usually taken as a material the band structure of which is close to that of free
electrons (Fig. 3.17, cf. in particular in particular the strong dispersive sp-band).
Also for Cu, a single band crosses the Fermi level. If there would be a single Fermi
wavelength or a single Fermi wavevector, the Fermi surface (FS) would be a sphere.
The Fermi surface is the set of wavevectors k at which a band crosses (or touches)
the Fermi level. The FS of Cu is shown in Fig. 5.2. Note in particular the band gap
at the L points of the BZ.

Figure 5.2. Fermi surfaces of K (left), Cu (center), and Fe (right).

A much more free-electron like material is for example K, with its single s-valence
electron (Fig. 5.2, left). In contrast, the FS of Fe is much more complicated because
of its ferromagnetism and the Fermi-level crossings of many d-bands (Fig. 5.2, right).
2

Elastic scattering time and length The quantum scattering time τq is the
average time between subsequent elastic scattering events. It is related to
the quantum scattering length lq by

lq = vF τq , (5.37)

117
 where vF denotes the Fermi velocity of the electrons at the Fermi level.
From p = m? v and E = p2/2m? is follows
r
2EF
vF = . (5.38)
m?
Hence, lq is average distance between two elastic scattering events.
The resistivity of a device depends on the scattering times and length
between strong scattering events. The elastic mean free path le is defined
as

le = vF τ, (5.39)

where τ is the Drude scattering time, i. e. the average time between two
strong scattering events.
In typical two-dimensional electron gases (2DEGs) one finds le ≈ 8 µm. If
the scattering potential is weak, then τq/τ  1. In a 2DEG one typically
finds τq ≈ 0.1τ .
In the above, we have introduced quietly the effective mass m? of the
electrons. In transport, one is mainly interested in a simple descrip-
tion of band electrons with energy close to the Fermi level. Hence, one
describes the band structure E(k) in terms of free-electron parabolae
E(k) = ~2 k 2 /2m but with an adjusted mass. At an extremal point of the
band structure, we expand the dispersion relation into a Taylor series,
1X ∂2E
E(k) ≈ E0 + ki kj . (5.40)
2 ∂ki ∂kj
ij

The effective-mass tensor is then given by

1 ∂2E
 
1
≡ . (5.41)
m? ij ~2 ∂ki ∂kj

A further approximation is to assume an isotropic dispersion, as for free

electrons. In this case, the above tensor becomes diagonal with equal
elements
1 ∂2E
 
1
≡ , (5.42)
m? ~2 ∂k 2
known as the effective mass.
Diffusion constant The diffusion constant D originates from gradients in the
electron density n. It shows up in the diffusion equation (in 1D)
dn d2 n
= D 2. (5.43)
dt dx
The Einstein relation gives
kB T kB T
D= τ= µ, (5.44)
m? e

118
 with the mobility µ of the electrons [see eq. (5.159)]. Further,
1
D = vF2 τ. (5.45)
2
In a 2DEG one finds typically D = 0.1 m2 /s.

Phase-coherence length Inelastic scattering events can result in dephasing, for

example spin-flip scattering at magnetic impurities or electron-phonon
scattering. Elastic scattering does no cause dephasing. For transport
phenomena that rely on interference of wavefunction, one can define a
dephasing time.
The phase-coherence length lφ is the distance electrons travel before their
phase is randomized. In ballistic transport, for example, the extension of
the sample is much less than the phase-coherence length. The phase-
coherence time in mesoscopic samples is about 1 ps.

Magnetic length An external magnetic field sets a length scale, namely the
spatial extension of wavefunctions in that magnetic field. It is given by
the magnetic length
r
~
lB = (5.46)
eB
which corresponds to the width of the ground state of a quantizing mag-
netic field with strength B.
The cyclotron radius is the radius of the circle electrons follow in a mag-
2.
netic field, rc = kF lB

Further quantities There are several more quantities that can be used to clas-
sify transport regimes and samples. Just to name a few, the electron-
electron scattering time, the thermal length, and the localization length.

5.3 Transport in two-dimensional electron gases

5.3.1 Landau quantization
Transport experiments on mesoscopic samples are often performed in an exter-
nal magnetic field. The latter allows to tune non-destructively for example the
magnetic length or the effective mass of the electrons.
2DEG in a perpendicular magnetic field
The Schrödinger equation of a free two-dimensional electron gas (2DEG) in a
magnetic field reads

(p̂ + eA)2
 
+ V (z) Φ(r) = EΦ(r). (5.47)
2m?

The electrons are confined in the xy plane. Hence, V (z) can be taken as a
constant.

119
 Figure 5.3. Formation of a 2DEG at the interface of two semiconductors.

In electrodynamics, the electric and magnetic fields can be expressed in terms

of a scalar potential φ and a vector potential A,

E = −∇φ, (5.48a)
B = rot A = ∇ × A. (5.48b)

Both fields are invariant under gauge transformations, that is

A → A − ∇ψ, (5.49a)
∂ψ
φ→φ+ , (5.49b)
∂t
give identical fields.
In the present case, the magnetic field should be homogeneous and along the
z-axis, B(r) = Bez . This gives for the vector potential

0 = ∂y Az − ∂z Ay , (5.50a)
0 = ∂z Ax − ∂x Az , (5.50b)
B = ∂x Ay − ∂y Ax , (5.50c)

which is for example fulfilled by

A = (−By, 0, 0) (5.51)

(Landau gauge) and by

1
A = (−By, Bx, 0). (5.52)
2

120
Here, we will choose the Landau gauge. Thus, the Hamiltonian reads
1  2 2

Ĥ = (p̂ x − eBy) + p̂ y (5.53a)
2m?
1 
(−i~∂x − eBy)2 − ~2 ∂y2

= ?
(5.53b)
2m
1  2 2 2 2 2 2 2

= −~ ∂x + 2i~eBy∂ x + e B y − ~ ∂y (5.53c)
2m?
~2 ~eB e2 B 2 2 ~2 2
= − ? ∂x2 + i ? y∂x + y − ∂ . (5.53d)
2m m 2m ? 2m? y
For the wavefunction we make the ansatz

Φ(x, y) = Ψ(y) eikx x . (5.54)

Inserting into the Schrödinger equation yields

~2 2 e2 B 2 2 ~2 2
 
~eB
− ? ∂x + i ? y∂x + y − ∂ Ψ(y) eikx x = EΨ(y) eikx x
2m m 2m? 2m? y
(5.55a)
 2 2 2 2 2

~ kx ~kx eB e B 2 ~
− y+ y − ∂ 2 Ψ(y) = EΨ(y). (5.55b)
2m ? m ? 2m ? 2m? y
With the cyclotron frequency
eB
ωc ≡ (5.56)
m?
we have
 2 2
~2 2

~ kx 1 ? 2 2
− ~k ω
x c y + m ω c y − ∂ Ψ(y) = EΨ(y). (5.57)
2m? 2 2m? y
Substituting
~kx
v≡y− (5.58)
m? ω c
we arrive at
~2 2 1 ? 2 2
 
− ? ∂v + m ωc v u(v) = Eu(v), (5.59)
2m 2
i e. at the Schrödinger equation of an harmonic oscillator.
The eigenenergies of an harmonic oscillator are given by [11]
 
1
En = ~ω n + , n = 0, 1, . . . (5.60)
2

The ground-state energy is E0 = ~ω/2 (n = 0). This is a very significant

physical result because it tells us that the energy of a system described by a
harmonic-oscillator potential cannot have zero energy. Physical systems such
as atoms in a solid lattice or in polyatomic molecules in a gas cannot have zero

121
energy even at absolute zero temperature. The energy of the ground vibrational
state is often referred to a ‘zero-point vibration’. The zero-point energy is
sufficient to prevent liquid He4 from freezing at atmospheric pressure, no matter
how low the temperature.
The eigenfunctions involve Hermite polynomials, the first four read
 α 1/4 2
H0 (y) = e−y /2 , (5.61a)
π
 α 1/4 √ 2
H1 (y) = 2y e−y /2 , (5.61b)
π
 α 1/4 1 2
H2 (y) = √ (2y 2 − 1) e−y /2 , (5.61c)
π 2
 α 1/4 1 2
H3 (y) = √ y(2y 2 − 3) e−y /2 , (5.61d)
π 2
with α = mω/~ (Fig. 5.4).

Figure 5.4. Wavefunctions (left) and probability density (right) of an harmonic os-
cillator. Only the four lowest levels are shown

If we confine the 2DEG in the x-direction with dimension L, then kx becomes

quantized with
nπ
kx,n = . (5.62)
L
Hence, the harmonic oscillators are located at
~kx ~π
yn = =n ? . (5.63)
m? ω c m ωc L
Its energy is
 
1
Ej = ~ωc j − , j = 1, 2, . . . (5.64)
2

122
(note the minus-sign) and is called the j-th Landau level (LL). Note that the
degeneracy of each LL is given by the number of allowed wavenumbers in x-
direction. The density of states of an ideal 2DEG in a perpendicular magnetic
field reads
gs gs eB
D(E) = 2 δ(E − Ej ) = δ(E − Ej ), (5.65)
2πlB h

where lB is the magnetic length, eq. (5.46) and gs is the spin degeneracy. In real
systems, the δ-peaks are smeared out (Gaussian distribution) due to potential
fluctuations.
The LL spacing is ~ωc . Hence, taking into account the degeneracy, one has

~ωc ν = gs EF (5.66)

for the occupied (filled) LLs, where ν is the filling factor,

gs EF
ν= . (5.67)
~ωc
Partially filled LL result in an non-integer filling factor (Fig. 5.5).

Figure 5.5. Density of states of a 2DEG for zero magnetic field (B = 0; left) and with
increasing magnetic field (a-c). The B-field gives rise to Landau levels, shown here as
broadened δ-peaks. The level spacing ~ωc increases with field strength, shifting LLs
through the Fermi energy (dashed line). In case c, the filling factor ν is non-integer, in
contrast to cases a and b.

Due to the applied magnetic field, the levels become split due to the Zeeman
effect. The magnetic field (along the z-axis) couples to the magnetic dipole
moment of the electron. The latter is related to the spin,
eB
Ĥz = ŝz . (5.68)
2m? c
Hence, the spin degeneracy is lost and the splitting is proportional to the applied
field.2
2
Note the magnetic moment ms is opposite to the spin.

123
Figure 5.6. Zeeman effect. A magnetic field along the z-direction (red arrow) lifts
the spin degeneracy, the level spacing increases with field strength. The levels are pure
spin-up (ms = −1/2) and pure spin-down (ms = +1/2) states. The level spacing can
be obtained by absorption, as sketched.

Experimental observation. Landau levels can be observed using a STM. The

tip of the STM lowers the potential at the surface by about 30 meV in an area
of about 40 nm extension. Hence, the 2DEG at the surface becomes confined
in this quantum dot (Fig. 5.7). In spectroscopy mode, the differential conduc-

Figure 5.7. Formation of a quantum dot for a 2DEG at an InAs surface by the tip
electrode of a STM.

tance dI/dV shows a maximum at the energy position of a quantum-dot level

(Fig. 5.8). Switching on the magnetic field (here: 6 T) gives rise to Landau lev-
els which are further split by the Zeeman effect. Indications that the structures
are indeed due to Landau levels are the equidistant spacing and the dependence
on field strength (Fig. 5.9).

124
Figure 5.8. Observation of energy levels in a quantum dot at an InAs surface. The
QD is induced by the STM tip. No magnetic field applied

Chemical potential in strong magnetic fields.

Assuming a fixed magnetic field B and degeneracy gs = 2, the integrated density
of states in each Landau level is
gs 1
DLL = 2 = 2. (5.69)
2πlB πlB
At zero temperature, the Fermi level µ = EF is independent of the electron
density n as long as there are empty states in the highest occupied LL. In this
case, EF = (j − 1/2)~ωc .
At nj = jDLL all LL are either full or empty, and EF = j~ωc . The 2DEG
is an insulator because the density of states at the Fermi level is zero. Thus, a
2DEG in a magnetic field goes through a series of metal-insulator transitions
upon increasing n. In our simplified model, the insulating phases are just points,
due to the δ-peaks in the density of states. In real sample, however, these peaks
are broadened.
Problem 53. What can be reasons for the broadening?
The main reason are fluctuations of the potential V (z) in which the 2DEG is located. 2
Similar phase transitions are found upon varying B while the electron density
is fixed. Here, the insulating points appear at
π~n 1
Bj = . (5.70)
e j
At these points the Fermi energy varies from
~eBj
EF (Bj − δB) = (j − 1/2) (5.71)
m?

125
Figure 5.9. Observation of Landau levels in a quantum dot at an InAs surface.
Each LL is split further by the Zeeman effect into spin sublevels (indicated by vertical
arrows). The QD is induced by the STM tip. The applied magnetic field has a strength
of 6 T.

via
~eBj
EF (Bj ) = j = EF (B = 0) (5.72)
m?
to
~eBj
EF (Bj + δB) = (j + 1/2). (5.73)
m?

5.3.2 Quantum Hall effect

Before we discuss the quantum Hall effect (QHE), the conventional Hall effect
is sketched.
Hall effect
In the conventional Hall effect, a current of charged particles, for example
electrons, is passed through a conductor to which a magnetic field is applied
(Fig. 5.10). The magnetic field acts via the Lorentz force
Z
F = [ρ(r)E(r) + j(r) × B(r)] dr3 (5.74)
V

on the particles. Here, ρ is the charge density and j the current density. As-
suming a homogenous sample and homogenous fields, the Lorentz force reads

F = qE + I × B, (5.75)

126
 Figure 5.10. Conventional Hall effect.

where q is the charge and I the current. The current can be expressed as qv,
that is charge multiplied by velocity, resulting in

F = q(E + v × B). (5.76)

Here, we are interested in the magnetic contribution F m to the force,

F m ≡ qv × B, (5.77)

which produces a force perpendicular to the current and the magnetic field
(‘right-hand rule’). According to Fig. 5.10, positively charged particles are
refracted to the right and accumulate at the right edge of the sample while
negatively charged particles are refracted to the left and accumulate at the left
edge of the sample. This charge accumulation builds up an electric force that
results in a voltage VH measured between the right and the left edge of the
sample. The occurrence of this voltage is the Hall effect.
For the magnetic field along the z-axis, the Lorentz force is aligned along the
y-direction,

Fy = −qvd Bz , (5.78)

where vd is the drift velocity of the carriers, vd = µEx (µ is the mobility); see
eq. (5.161). Thus,

Fy = −qµEx Bz . (5.79)

The force due to the charge accumulation and the Lorentz force must cancel in
equilibrium, so that

qEy = −qµEx Bz , (5.80)

Ey = −µEx Bz . (5.81)

The Hall voltage is then given by VH = Ey dy , dy being the extension of the

sample in y-direction.

127
 The Hall coefficient is defined as
Ey −µEx Bz −µ 1
RH ≡ = = =− , (5.82)
Bz jx Bz σEx σ qn
where we have used Ohm’s law, j = σE. n is the carrier concentration of the
sample. Hence,
IB IB
VH = = RH . (5.83)
ned d
where n is the carrier density in the sample, e the electric charge, and d the
thickness of the sample. The Hall effect allows to determine the carrier concen-
tration.
One very important feature of the Hall effect is that it differentiates between
positive charges moving in one direction and negative charges moving in the
opposite. The Hall effect offered the first real proof that electric currents in
metals are carried by moving electrons, not by protons. The Hall effect also
showed that in some substances (especially semiconductors), it is more appro-
priate to think of the current as positive ‘holes’ moving rather than negative
electrons.
Problem 54. Can there be a Hall effect without an external magnetic field? What kind of
samples can give rise to such an effect?
The key property of the conventional Hall effect is the external magnetic field. To
observe a Hall effect without external field, one needs an internal magnetic field which
shows up in a magnetic sample. This effect is called anomalous Hall effect (AHE). It
depends on the direction of the magnetization with respect to the current induced by the
external electric field (voltage). Thus, one needs to take spin-orbit coupling into account
to describe the AHE.
If one records the Hall resistivity versus external magnetic field (as for the conventional
Hall effect), the is a finite value at zero field, shown in Fig. 5.11 in the diluted magnetic
semiconductor (DMS) InMnAs. The gate voltage is used to vary the carrier concentration
and consequently the magnetization in the sample. 2

The quantum Hall effect

In a two-dimensional metal or semiconductor (2DEG), the Hall effect is also
observed, but at low temperatures a series of steps appears in the Hall resis-
tance as a function of magnetic field, instead of the monotonic increase in the
conventional Hall effect. These steps occur at precise values of resistance which
are the same no matter what sample is investigated (Fig. 5.12). The resistance
is quantized in units of h/e2 divided by an integer. This is the quantum Hall
effect (QHE), for the observation of which K. von Klitzing won the Nobel prize
in 1985.
The QHE can be seen at liquid helium temperatures, but in the mK regime
the plateaus in ρxy are much wider. Also included is the diagonal component
ρxx of resistivity, which shows regions of zero resistance corresponding to each
QHE plateau. In the above figure the plateau index is, from top right, 1, 2,
3, 4, 6, 8, . . . . Odd integers correspond to the Fermi energy being in a spin
gap and even integers to an orbital Landau level (LL) gap (Fig. 5.9). As the
Zeeman splitting is small compared to LL gaps, the odd integer plateaux are
only seen at high magnetic fields. Important points to note are:

128
Figure 5.11. Magnetization hysteresis loops determined by measurements of the
anomalous Hall effect at a constant temperature of 22.5 K for various gate voltages in
a field-effect transistor structure with (In,Mn)As channel.

• The value of resistance only depends on the fundamental constants of

physics: e the electric charge and h Planck’s constant.

• It is accurate to 10−9 .

• The IQHE can be used as resistance standard, h/e2 = 25812.807 Ω.

Origin of the IQHE. The resistivity is quantized with

1 h
ρxy (B) = , (5.84)
ν e2
and is universal, meaning that is is observed in all 2DEGs. There is also the
fractional QHE (FQHE) which shows up in extremely high-quality samples.
There, ν is a rational number.

129
Figure 5.12. Integer quantum Hall effect. The figure shows the integer quantum Hall
effect (IQHE) in a GaAs-GaAlAs heterojunction, recorded at 30 mK.

Figure 5.13. Integer quantum Hall effect (schematic).

Suppose that the magnetic field is tuned to an insulating point [eq. (5.70)],

π~n 1 hn 1
Bj = = . (5.85)
e j 2e j
The Hall resistivity is then
Bj h 1
ρxy,j = − =− 2 . (5.86)
en 2e j

130
 Figure 5.14. Integer and fractional quantum Hall effect.

If there is no diffusive scattering, the classical equation of motion (Newton)

reads

m? ∂t2 r = −e(E + ∂t r × B), (5.87)

where r is the position of the electron in the xy-plane. With E = (0, Ey , 0) and
B = (0, 0, B), the solutions are given by

Ey Ey
x(t) = sin(ωc t) − t + x0 , (5.88a)
ωc B B
Ey
y(t) = − cos(ωc t) + y0 . (5.88b)
ωc B
The electronic motion appears as superposition of two motions. That of the
guiding center is given by

X(t) = −vL t + x0 , (5.89a)

Y (t) = y0 , (5.89b)

while that relative to the guiding center is given by

xr (t) = rc sin(ωc t), (5.90a)

yr (t) = −rc cos(ωc t). (5.90b)

Here, vL is the drift velocity due to the Lorentz force and rc the cyclotron radius
(Fig. 5.15c).
Averaged over time, the circular motion around the guiding center gives no
contribution to the conductivity. Hence, only the contribution of the guiding
center remains. From Ohm’s law we have

jx = σxx Ex + σxy Ey = −ne∂t x, (5.91a)

jy = σyx Ex + σyy Ey . (5.91b)

131
Thus,
ne
σxy = − (5.92a)
B
σyy = 0. (5.92b)
Likewise, we find σxx = 0 and σxy = −σxy for an electric field in x-direction.
Filling now a Landau level, that is increasing the density at the Fermi level
by δn = DLL L, the Hall conductance increases by δσxy = −2e2 /h. In summary,
a jump in the resistivity or conductivity shows up each time a LL crosses the
Fermi energy upon increasing the magnetic field.

The effect of disorder. Disorder increases the insulating regions from points to
extended intervals while the extended metallic regions are considerably reduced
(Fig. 5.15). This allows the observation of the QHE.

Figure 5.15. Effect of disorder on the integer QHE.

Disorder results in fluctuations of the potential, a so-called potential land-

scape is formed (a). The first LL follows the energy of the conduction band
bottom, resulting in the peak in the density of states (right). At energies E1
and E3 the electronic states are localized, at E3 at the potential maxima, at
E1 in the potential minima (b). At energy E2 , the localized states merge and
can travel (‘percolate’) across the sample. This results in the center of the DOS
peak due to extended states and in the nonzero ρxx . Note that edge states (b)
and localized states give no contribution to the Hall conductivity.

132
5.3.3 Spin-orbit coupling in a two-dimensional electron gas
Another effect in a 2DEG is the effect of spin-orbit coupling (SOC) on its
electronic structure. It leads to splitting of the free-electron like states and to
a complete spin polarization. [85, 86]
Spin-orbit interaction
For an electron in a magnetic field H, the Hamiltonian is given by

Ĥ = −µ · H, (5.93)

cf. eq. (5.68). Here, µ is the magnetic moment of the electron. If there is an
additional electric field E, the electron is subject to the additional field
v
H1 = ×E (5.94)
c
in the rest frame of the laboratory. v is the electron’s velocity. This field leads
to a precession of the spin axis around H 1 with the Larmor frequency
eH 1
ω1 = , (5.95)
mc
Fig. 5.16. The acceleration a of the electron by the electric field leads to an

Figure 5.16. Larmor precession.

additional precession
1 e
ω2 = 2
v×a=− v × E. (5.96)
2c 2mc2
The total precession is then given by
1 e
ω ≡ ω1 + ω2 = ω1 = v × E. (5.97)
2 2mc2
With eE = −∇V , we arrive at the Hamiltonian for spin-orbit coupling
~
ĤSO = ~ω · ŝ = ŝ · (v × ∇V ). (5.98)
2mc2

133
 In a central potential one has
1 dV
∇V = r. (5.99)
r dr
With v = p̂/m it follows
~ 1 dV
ĤSO = ŝ · (r × p̂). (5.100)
2m2 c2 r dr
With the angular-momentum operator l̂ = r × p̂, we arrive at
~ 1 dV
ĤSO = ŝ · l̂. (5.101)
2m2 c2 r dr
In this expression, the interaction of spin and angular momentum is obvious.
Problem 55. In what cases is spin-orbit coupling large (compared to what?) and cannot
be neglected?
From equation (5.101) it is clear that SOC becomes important if the gradient of the
potential is large. For a core level it should be compared to the binding energy of that
level. For example, the Fe-2p levels are split due to SOC into 2p1/2 and 2p3/2 , where
the subscript gives the total angular momentum j = l + s. These sublevels have binding
energies of 719.9 eV and 706.8 eV, with a spin-orbit splitting of about 13.1 eV. Note that
these sublevels are clearly distinguished in XMCD as the LII and LIII edges. For the 3p
levels these sublevels cannot be clearly distinguished [87] and show up in ARPES as broad
maxima at 52.7 eV binding energy.
In the valence band regime the SO splitting is even less but gives rise to magnetocrys-
talline anisotropy (MCA). For bulk solids it is in the order of a few µeV, at surfaces of a
few meV. 2
Problem 56. The above derivation does not explicitly rely on special relativity. Why is
spin-orbit coupling be regarded as a relativistic effect?
In the above derivation one distinguishes between the laboratory frame and the rest
frame of the electron. The magnetic field in the latter (which is due to the electric field in
the former) is an effect of Maxwell’s equations, the latter being Lorentz invariant. Here,
special relativity comes into play. Hence, spin-orbit coupling is best described in a relativistic
theory, being provided by Dirac theory [88, 89]. In the Dirac equation which replaces the
Schrödinger equation spin-orbit coupling and magnetism are treated on equal footing. 2
Problem 57. Electrons in a solid are much slower than the speed of light. Why comes
spin-orbit coupling into play?
Spin-orbit coupling is automatically taken into account in a relativistic theory, even if the
particles are slow (compared to the light velocity). Starting from the Schrödinger equation,
SOC has to be treated non-relativistically, as in Section 5.1. 2

Spin-orbit interaction in a 2DEG

The nonrelativistic Hamilton operator for electrons with spin-orbit interaction
is given by [81, 88]
1 1
Ĥ = p̂2 + V (r) + 2 σ · [∇V (r) × p], (5.102)
2 2c
with σ the vector of Pauli matrices and p̂ = −i∇ the momentum operator
(in Hartree atomic units, e = m = ~ = 1, c ≈ 137). In order to form a two-
dimensional electron gas (2DEG), the potential V has to confine the electrons to

134
a plane (say z = 0) and allow for free motion within that plane [i. e., V = V (z)].
Assuming without loss of generality V (0) = 0, the Hamiltonian for the 2DEG
then reads
1 ∂z V (z)|z=0
Ĥ = − (∂x2 + ∂y2 ) + i (σx ∂y − σy ∂x ) . (5.103)
2 2c2
The ansatz ψk for the eigenfunctions consists of a plane-wave spinor,

ψk (r) = (µk χ+ + νk χ− )ei(kx x+ky y) , (5.104)

with k = (kx , ky ) and r = (x, y). The Pauli spinors χ± are quantized with
respect to the z-axis (σz χ± = ±χ± ). With σx χ± = χ∓ and σy χ± = ±iχ∓ , the
Schrödinger equation Hψk = E(k)ψk yields the condition

(ky + ikx )νk2 = (ky − ikx )µ2k (5.105)

for the coefficients µk and νk .

Problem 58. Derive the above relation.
With γ ≡ ∂z V (z)|z=0 /(2c2 ) and

σx (µk χ+ + νk χ− ) = (µk χ− + νk χ+ ),
σy (µk χ+ + νk χ− ) = i(µk χ− − νk χ+ ),

the l. h. s. of the Schrödinger equation Ĥψk (r) = Eψk (r) reads

1 2
k (µk χ+ + νk χ− ) + iγ iky (µk χ− + νk χ+ ) + i2 kx (µk χ− − νk χ+ ) ,
 
2
1 2
k (µk χ+ + νk χ− ) + iγ iky (µk χ− + νk χ+ ) − kx (µk χ− − νk χ+ ) ,
 
2
1 2
k (µk χ+ + νk χ− ) − γ ky (µk χ− + νk χ+ ) + ikx (µk χ− − νk χ+ ) ,
 
2
whereas the r. h. s. equals

E(µk χ+ + νk χ− ).

Comparing now the spin-up part (χ+ ) results in

1 2
k µk − γ (ky − ikx ) νk = Eµk .
2
The spin-down part gives
1 2
k νk − γ (ky + ikx ) µk = Eνk .
2
Multiplication of the first equation by νk and of the second by µk and subtracting yields

γ (ky − ikx ) νk2 = γ (ky + ikx ) µ2k ,

from which follows eq. (5.105). 2

Normalization of ψk results in two sets of coefficients, (indexed by ±),
s
1 1 ky − ikx
µk± = √ , νk± = ∓ √ for k 6= 0. (5.106)
2 2 ky + ikx

135
The eigenvalues of ψk± are given by
1
E± (k) = k 2 ± γ|k|, (5.107)
2
with γ = ∂z V (z)|z=0 /(2c2 ), Fig. 5.17. Hence, the strength of the spin-orbit
induced splitting, E+ (k) − E− (k) = 2γ, is proportional to the structural asym-
†
metry given by ∂z V . The spin polarization P ± (k) = hψk± |σ|ψk± i is complete
and perpendicular to k,
ky kx
P±x (k) = ∓ , P±y (k) = ± , P±z (k) = 0. (5.108)
|k| |k|

Hence, P + (P − ) rotates anti-clockwise (clockwise) around the z-axis.

Figure 5.17. Rashba-Bychkov effect in a 2DEG.

Note that the sign of γ determines the order of the spin-orbit split states in
reciprocal space, that is, which one is the inner and which one is the outer in
the momentum distribution. But since it does not affect the spin polarization,
the latter—for instance obtained from experiment—is a measure for the sign of
γ.
The spin splitting of the 2DEG was found by Rashba and Bychkov, thus
named today Rashba-Bychkov effect [90, 91].
Problem 59. For the Rashba-Bychkov effect one needs an asymmetric confinement of the
2DEG. How can this be achieved experimentally?
Asymmetric confinement is obtained at an interface of different materials. For example,
in a semiconductor heterostructure (GaAs/GaAlAs) or at a metal surface). In addition, the
SOC strength can be tuned by an external electric field normal to the xy-plane. 2

Rashba-Bychkov effect in a semiconductor heterostructure. In the field of

spin electronics, or spintronics, the quantum mechanical property of the electron
spin is used for switching purpose instead of its charge. Potential advantages of

136
this concept are lower power consumption as well as faster switching. A typical
example of a spinelectronic device is the spin transistor proposed by Datta and
Das [92]. Here, ferromagnetic (FM) source and drain contacts are used as spin
injector and analyzer. The spin orientation is controlled by the Rashba effect
by means of a gate electrode, Fig. 5.18 (left).

Figure 5.18. Left: Spin field-effect transistor. Right: band profile of a semiconduc-
tor quantum well. The asymmetric confinement of the 2DEG results in the Rashba-
Bychkov effect.

The Rashba-Bychkov effect can originate from the macroscopic electric field
in a semiconductor quantum well. In Fig. 5.18 (right), a typical conduction band
profile of a semiconductor quantum well is depicted. Due to the band offsets at
the interface of two different materials the electrons are confined in a quantum
well, and a 2DEG is formed. If the potential well is asymmetric, the electrons
are moving in an effective electric field E. In the reference system of the electron
this electrical field transforms into a magnetic field B. Depending of the spin
orientation and the corresponding magnetic moment an energy lowering or an
energy increase occurs, respectively. For applications it is essential, that the
strength of the Rashba effect and thus the spin splitting can be controlled by
means of a gate electrode, Fig. 5.18 (left).
The Rashba effect is found to be very pronounced in two-dimensional elec-
tron gases with a low band-gap channel layer. In this respect a 2DEG in an
InGaAs/InP heterostructures is an almost ideal system. As can be seen in
Fig. 5.19 (left), the band profile (red curve) is asymmetric so that the electrons
are propagating in an effective electric field. In order to enhance the electron
mobility, the concept of modulation doping is used. Here, the dopant atoms are
separated by an undoped InP spacer layer from the two-dimensional channel in
order to suppress scattering processes.
The strength of the Rashba spin-orbit coupling can be extracted from mag-
netotransport measurements. Shubnikov-de Haas oscillations are observed if
the longitudinal resistance is measured as a function of a magnetic field.3 The
oscillation frequency in 1/B is directly related to the electron concentration
3
The Shubnikov-de Haas effect is the oscillatory behaviour of the conductivity of a magnetic
sample in a magnetic field. The de Haas-van Alphen effect is the oscillatory behaviour of
the magnetic susceptibility. The periodicity in the de Haas-van Alphen effect measures
the extremal cross-section area S in k-space of the Fermi surface, the cross section taken
perpendicular to B: ∆(1/B) = 2πe/~cS.

137
Figure 5.19. Left: Semiconductor heterostructure for investigating the Rashba-
Bychkov effect. Right: Shubnikov-de Haas oscillations. The beating pattern is related
to the Rashba-Bychkov effect.

in the 2DEG. As mentioned above, the presence of the Rashba effect results
in a spin splitting and thus to an effective separation in two different two-
dimensional systems with different electron concentrations. The superposition
of these slightly different oscillation frequencies results in a beating pattern
of the Shubnikov-de Haas oscillations. By analyzing the beating pattern the
strength of the Rashba-Bychkov effect can be determined (Fig. 5.19, right).

Rashba-Bychkov effect at metal surfaces. As for semiconductor heterostruc-

tures, the potential at a metal surface can result in a sizable Rashba-Bychkov
effect of surface states. While Shubnikov-de Haas measurements provide rather
indirect connection to the Rashba-Bychkov effect, in articular it has no spin
resolution, surface states can be investigated in much more detail by means of
spin- and angle-resolved photoelectron spectroscopy (ARPES, Sect. 4.2).
At the (111) surface of noble metals, free-electron like surface state exist [85].
In particular the surface state at Au(111) shows a sizable splitting, as is shown
in Fig. 5.20 [86]. One clearly resolves the splitting into an inner and into an
outer surface state (cf. Fig. 5.17). The spin polarization is obtained by spin-
resolved photoelectron spectroscopy (SPARPES).
Problem 60. The spin polarization obtained in the model calculation, eq. (5.108), is
complete, within the xy-plane, and perpendicular to the wavevector. For Au(111) the
spin polarization of the photoelectrons is considerably less and shows a threefold rotational
symmetry, although the surface-state dispersion looks close to that of the model. Give
possible reasons for the deviations from the model behaviour.
The threefold rotational symmetry of the spin polarization is due to the surface symmetry.
It allows also for a nonzero Pz . However, the deviation from the model is very small. The
in-plane component is about ±97 % while the z-component is about ±1.4 % for the surface
state.
There are two main reasons for the much smaller spin polarization of the photoelectrons
than that of the surface state. First, a finite lifetime of the photoelectrons makes the
intensities of inner and the outer overlap, leading to a reduction of the spin polarization.
Second, spin-orbit coupling itself gives rise to spin-polarized photoelectrons [80], even if
the initial states are not spin-polarized (this effect depends on the experimental set-up).
Hence, the surface-state polarization can be reduced. 2
The Rashba-Bychkov effect was also found at other metal surface, even for

138
Figure 5.20. (a) Dispersion of the split surface state on Au(111) as obtained by
density-functional calculations. The green area indicates the bulk-band region. The
Fermi energy is at 0 eV. (b) Momentum distribution at the Fermi energy, as obtained
from panel a. (c) and (d) Experimental dispersion and momentum distribution, as
obtained by angle-resolved photoelectron spectroscopy.

not free-electron like surface states.

5.4 Transport regimes

The transport in nanostructures can be cast in two major regimes, the diffusive
regime and the ballistic regime. Both are theoretically described by different
means, which will be discussed in the following.
First, we recall main results of solid-state theory.
• Neither the electronic states (Bloch states) of fully occupied nor empty
bands can carry an electric current. In other words, only partially filled
bands can carry a current, that are those bands that cross the Fermi level.
• The resistance of a perfect crystal with at least one partially filled band
vanishes. The electrons obey the semiclassical equations of motion
1
v(k) = ∇k E(k), (5.109a)
~
dk e
= − (E + v(k) × B). (5.109b)
dt ~
Resistance is due to scattering at imperfections, that is at phonons, im-
purities, lattice dislocations, surface, and interfaces.4
4
As we will see later, there is a intrinsic upper limit for the conductance that is due to the
finite number of Bloch states that can carry the current (page 153).

139
Figure 5.21. Top: Components of the spin polarization obtained by photoemission
calculations (cf. the momentum distribution in Fig. 5.20b). (a) Tangential, (b) radial,
and (c) perpendicular component of the spin polarization. Note the threefold rotational
symmetry. Black segments indicate the positions of the split surface states. Bottom:
Experimental counterparts of the tangential (d) and the perpendicular component (e)
of the photoelectron spin polarization. The surface states are shown schematically in
the center.

• Electron-electron scattering does not change the total momentum of the

electron gas (except in exceptional cases) and, hence, does not produce
resistance.

• For small bias voltages, only a small fraction of the electrons with energy
around the Fermi level contributes to the current. These can then be
described in the effective-mass approximation.

5.4.1 Diffusive transport

In diffusive transport the extension of the sample is much larger than the mean
free path (page 81), Fig. 5.22. Hence, scattering can be viewed as an internal
homogenous friction of the electron system. Diffusive transport is the a steady
state established by the applied bias voltage (electric field) and the friction.

The Boltzmann equation. In the ground state, the electronic states are oc-
cupied according to the Fermi-Dirac (FD) distribution [eq. (3.93)]
1
f (k) = . (5.110)
1 + e[E(k)−µ]/kB T
Both external fields and scattering will lead to a deviation from the FD dis-
tribution, giving rise to a electron distribution function φ(k, r, t) that depends

140
Figure 5.22. Transport regimes. (a) In the diffusive regime the extension L of the
sample is much larger than the mean free path l, L  l. (b) In the quasi-ballistic
regime the scattering rate is much less than in the diffusive regime. (c) In the ballistic
regime there is no scattering, meaning L  l.

on the wavevector, the spatial position and on time. As in classical mechanics,

momentum and spatial coordinate make up the phase space, {k, r}. Thus,
2
φ(k, r, t) dk 3 dr3 (5.111)
(2π)3

is the number of electrons in the phase-space volume dk 3 dr3 at time t. Here,

the spin degeneracy is taken into account by the factor of 2.
We now consider the evolution of the distribution function but will consider
only an external electric field E for simplicity. Within the time interval dt,
an electrons moves in phase space from (k, r) to (k + δk, r + δr). From the
semiclassical motion, eq. (5.109), we obtain
e
δk = − E dt, (5.112a)
~
δr = v(k) dt, (5.112b)

resulting in
e
(k + δk, r + δr) = (k − E dt, r + v(k) dt). (5.113)
~
This result is only valid if the electron is not scattered out of the phase-space
volume under consideration or another electron is scattered into the phase-space

141
Figure 5.23. Temporal evolution of electrons in a phase-space volume (oval). Starting
at time t, electrons (red circles) are scattered out of the volume and electrons are
scattered into the volume, the latter propagating in phase space to its position at time
t + dt.

volume at time t. Consequently, there is an additional contribution δφ which

takes into account the scattering events,
 
∂φ(k, r, t)
δφ = dt. (5.114)
∂t scatter

Therefore, the number of electrons in (k, r) at time t is changed into that at

time t + dt, expressed in terms of the distribution function
e
φ(k − E dt, r + v(k) dt, t + dt) dk 3 dr3 = φ(k, r, t) dk 3 dr3
~  
∂φ(k, r, t)
+ dk 3 dr3 dt.
∂t scatter
(5.115)

Liouville’s theorem states that the volume element in phase space cannot change
with time. Hence,
e
φ(k − E dt, r + v(k) dt, t + dt) = φ(k, r, t)
~ 
∂φ(k, r, t)
 (5.116)
+ dt.
∂t scatter

Now we expand the left-hand side into a Taylor series up first order (infinitesimal
time step!),
e
φ(k − E dt, r + v(k) dt, t + dt) ≈ φ(k, r, t)
~
(5.117)
e ∂φ(k, r, t)
− E · ∇k φ(k, r, t) dt + v(k) · ∇r φ(k, r, t) dt + dt.
~ ∂t

142
Inserting the above result into eq. (5.116), we arrive at the general Boltzmann
equation
 
e ∂φ(k, r, t) ∂φ(k, r, t)
v(k) · ∇r φ(k, r, t) − E · ∇k φ(k, r, t) + = .
~ ∂t ∂t scatter
(5.118)

The right-hand side, the scattering term, can in principle be calculated for
all considered scattering mechanism, each weighted with the corresponding oc-
cupation probability and with the probability to find the final state empty.
However, these probabilities are just the distributions functions φ(k, r, t) and
1 − φ(k, r, t), respectively. Thus, the general Boltzmann equations is a com-
plicated integro-differential equation which hardly cannot be solved rigorously.
Therefore, we have to made further approximations.
Assume that the system is large enough to be considered homogenous. Thus,
we can drop the spatial coordinate r. Switching off the external field at a time
t0 will result in an exponential approach (i. e., a relaxation) of the distribution
function φ(k, t) to the FD distribution,

φ(k, t) − f (k) = [φ(k, t0 ) − f (k)] e−t/τ . (5.119)

Here, τ is the relaxation time. In this picture, the relaxation-time approxima-

tion, all different scattering mechanism are comprised into a single one and it
is assumed that the latter generates an average relaxation time τ . The relax-
ation will take place exclusively via scattering. The functional equation of the
exponential function,
∂g 1
= g, (5.120)
∂t τ
which is solved by g(t) ∝ exp(−t/τ ), leads to
∂ 1
[φ(k, t) − f (k)] = − [φ(k, t) − f (k)], (5.121a)
∂t τ
or by indicating explicitly the origin of the relaxation,
 
∂φ(k, t) 1
= − [φ(k, t) − f (k)]. (5.122)
∂t scatter τ
The general Boltzmann equation then reduces to
e ∂φ(k, t) 1
− E · ∇k φ(k, t) + = − [φ(k, t) − f (k)]. (5.123)
~ ∂t τ
In the stationary state, there is no time dependence. Hence we arrive at
e 1
− E · ∇k φ(k) = − [φ(k) − f (k)]. (5.124)
~ τ
Rearranging yields eventually
eτ
φ(k) = f (k) + E · ∇k φ(k). (5.125)
~

143
 We now consider further a small external electric field E. In this case, the
deviation of φ(k) from the FD distribution should be linear in E. This allows
us to write

∇k φ(k) ≈ ∇k f (k), (5.126)

resulting in
eτ
φ(k) = f (k) + E · ∇k f (k). (5.127)
~
This is a Taylor series expansion of φ(k) up to first order. It is a good approx-
imation for f (k + eτ E/~), provided eτ E/~ is small.
Finally we have solved the general Boltzmann equation under several assump-
tion, resulting in the simplified Boltzmann equation

φ(k) = f (k + eτ E/~). (5.128)

The electric field E shifts the Fermi surface in reciprocal space by eτ E/~.
The electrons become accelerated by the electric field and are scattered into
unoccupied electronic states (energy larger than the Fermi level) via both elastic
and inelastic processes. Only electrons close to the edge of the Fermi see, that
is at the Fermi surface, contribute to the current. Electronic states with energy
much less than the Fermi energy do not contribute because a current with
momentum ~k is canceled by a current with momentum −~k. This implies
that fully occupied bands do not contribute to the current (cf. insulators).
In the above approximation, the Fermi velocity vF is much larger than the
drift velocity vd (Fig. 5.24). Without electric field, vd equals zero and there is
no net current. For example, the currents with vx = −vF and vx = vF cancel.
With an electric field applied, e. g. in x-direction, the entire Fermi sphere is
shifted by vd (which is also along x). Explicitly, vx = −vF → −vF + vd and
vx = vF → vF + vd , stating that there is no cancellation of velocities.
From Figure 5.24 it is evident that only electronic states close to the Fermi
surface (v ≈ vF ) contribute to the current. Those with small velocities do not.
It is further clear the formerly empty states become populated (cf. the light
green area at the right) while formerly occupied states become depopulated
(dark green area at the left).
Problem 61. Sketch the above consideration in the band structure E(k). What is more
appropriate, the reduced or the repeated zone scheme?
The left panel of Figure 5.25 shows the distribution of electrons (red circles) in equilibrium
(no external electric field). The bands are occupied to the Fermi level (horizontal blue
line). The extension of the Fermi surface is indicated by the horizontal green line. At
zero temperature, the distribution function f (k) equals 1 (0) if E(k) < EF (E(k) > EF ).
There are as many states with positive group velocity occupied as states with negative
group velocity. Therefore, the net velocity vanishes.
The distribution function is shifted along k due to an external electric field, eq. (5.128)
(right panel). Note the horizontal displacement of the green line. Hence, formerly empty
states become occupied, as is seen on the right side (k > 0). In turn, formerly occupied
states become empty, as on the left side (k < 0). This is visualized by shifting the
electrons to the right (cf. the arrows). Now there is an imbalance of states with positive
and negative group velocity. In the present case there are more states with positive group

144
 Figure 5.24. Displacement of the Fermi sphere by an electric field in x-direction.

velocity occupied than with negative group velocity. Thus, there is net drift velocity to
the right (k > 0). Note further that the distribution function φ equals 1 in the k-range
indicated by the green line for occupied states.
The balance of fully occupied or of fully empty bands is not affected by the external
field. Hence, they do not contribute to the electric current. In other words, only electronic
states close to the Fermi level can carry an electric current.
From the above it is clear that electrons close to a boundary of the Brillouin zone are
shifted across that boundary. Hence, the repeated zone scheme is more appropriate. 2

Conductivity via Boltzmann equation. Ohm’s law relates the current density
j with the external electric field E via the conductivity tensor σ,
j = σE, (5.129)
or in components,
X
ji = σkl El , i = x, y, z. (5.130)
k

Here, we assume that σ is a scalar, implying that j k E and being isotropic.

In order to compute j, one has to sum over all electronic states with wavevec-
tor k,
Z
j = j(k) dk 3 . (5.131)

For the contribution of wavevector k one has “charge times velocity times
weight”,
j(k) = −ev(k)φ(k). (5.132)

145
Figure 5.25. Distribution of electrons (red circles) for a system without (left) and
with (right) external electric field, as shown in the band structure E(k). The blue line
indicates the Fermi level. Grey areas depict band gaps.

In effective-mass approximation this reads

~k
j(k) = −e φ(k). (5.133)
m?
And with equation (5.127),
~k h eτ i
j(k) = −e ? f (k) + E · ∇k f (k) . (5.134)
m ~
This gives for the current density
Z Z
2 e~ 3 2 e~ eτ
j=− 3 ?
kf (k) dk − 3
k E · ∇k f (k) dk 3 , (5.135)
(2π) m (2π) m? ~
in which each states occupies a volume of (2π)3 in reciprocal space. The spin
degeneracy gives the additional factor of 2.
Problem 62. The first term,
Z
2 e~
− kf (k) dk 3
(2π)3 m?
vanishes. Give reasons.
Since the dispersion E(k) is an even function in k, E(k) = E(−k), the Fermi-Dirac
distribution f (k) is also even. But the components of k are odd and, hence, the integral
vanishes for each of the k-components. 2
Thus, we are left with
2 e2 τ
Z
j=− kE · ∇k f (k) dk 3 . (5.136)
(2π)3 m?
It follows
∂f (k) ∂f (k) ~2 k 2 ~2 ∂f (k)
∇k f (k) = ∇k E(k) = ∇k = k. (5.137)
∂E ∂E 2m? m? ∂E

146
Hence,

2 e2 ~2 τ
Z
∂f (k) 3
j=− k(E · k) dk . (5.138)
(2π)3 m? 2 ∂E

For low temperatures, the Fermi-Dirac distribution (page 77) becomes a

Heaviside function Θ,

f (k(E)) = 1 − Θ(k(E) − kF ), (5.139)

implying that the integral is over the Fermi surface only (E = EF ),

f (k(E)) = 1 − Θ(E − EF ). (5.140)

The Heaviside function is defined as

(
0 x<0
Θ(x) ≡ . (5.141)
1 x≥0

Thus,
Z x
Θ(x) = δ(x0 ) dx0 (5.142)
−∞

and (formally)5

∂
δ(x) = Θ(x). (5.143)
∂x
Equation (2.33c),
X 1 d
δ(g(x)) = d
δ(x − xn ) with g(xn ) = 0, g(xn ) 6= 0,
n | dx g(xn )| dx

gives
m?
δ(E − EF ) = δ(k − kF ), (5.144)
~2 k
5
To be more mathematically accurate, we have to use generalized functions [25]. Then with
eq. (5.141) for any test function (square-integrable) Φ
Z +∞ Z +∞
(Θ, Φ) = Θ(x) Φ(x) dx = Φ(x) dx.
−∞ 0

The derivative of a generalized function T obeys

„ « „ «
dT dΦ
, Φ ≡ − T, .
dx dx
Thus,
„ « „ « Z +∞ Z +∞
dΘ dΦ dΦ
,Φ = − Θ, =− dx = Φ(0) = δ(x) Φ(x) dx = (δ, Φ),
dx dx 0 dx −∞

which establishes eq. (5.143).

147
using

~2 k 2
E= . (5.145)
2m?
With this follows
∂f (k) m?
= −δ(E − EF ) = − 2 δ(k(E) − kF ). (5.146)
∂E ~ k
Hence, the current density reads

2 e2 τ
Z
1
j= 3 ?
k(E · k) δ(k(E) − kF ) dk 3 . (5.147)
(2π) m k

Because of the δ-function, a surface integral remains. With

dk 3 = k 2 sin ϑ dk dϑ dϕ (5.148)

we have
π 2π
2 e2 τ
Z Z
j= kF kF (E · kF ) sin ϑ dϕ dϑ. (5.149)
(2π)3 m? 0 0

For convenience, we assume that the electric field is along the z-axis, E =
(0, 0, Ez ). Hence,
Z π Z 2π
2 e2 τ 3
j= Ez kF cos2 ϑ sin ϑ dϕ dϑ (5.150a)
(2π)3 m? 0 0
Z π
2 e2 τ 3
= E k
z F cos2 ϑ sin ϑ dϑ. (5.150b)
(2π)2 m? 0
2 e2 τ 1 π
=− 2 ?
Ez kF3 cos3 ϑ 0 (5.150c)
(2π) m 3
2
2 2 e τ 3
= k Ez , (5.150d)
3 (2π)2 m? F

where one could have asked the Bronstein-Semendjajew,

Z
1
sin x cos2 x dx = − cos3 x (5.151)
3

(integral no. 357 with n = 2 and a = 1 [28]). The conductivity is therefore

2 2 e2 τ 3 1 e2 τ 3
σ= kF = k . (5.152)
3 (2π)2 m? 3π 2 m? F

The electron density n is defined as the number N of electrons per volume

V,
N
n≡ . (5.153)
V

148
If each electron occupies a sphere with radius rs , then
1 4π 3
= r , (5.154)
n 3 s
or
 1/3
3
rs = . (5.155)
4πn

For Cu, rs = 1.41 Å. In reciprocal space, the level density is V /(2π)3 . Hence,

4πkF3 V
N =2 (5.156)
3 (2π)3

where the factor 2 is again due to spin degeneracy. From this follows

N 4πkF3 1 1
n= =2 3
= 2 kF3 . (5.157)
V 3 (2π) 3π

The conductivity is thus given by [eq. (5.152)]

e2 τ
σ= n. (5.158)
m?
Defining the mobility by
eτ
µ≡ , (5.159)
m?
then

σ = eµn. (5.160)

The drift velocity (for electrons, charge = −e) is defined by

j
vd ≡ − = −µE (5.161)
en
and is the average effective velocity of the electron system (Fig. 5.24).
Problem 63. From Figure 5.24 we learned that only electronic states close to the Fermi
surface contribute to the current. But equation (5.158) seems to imply that all electrons
contribute (electron density n). Can you solve this apparent contradiction?
One has to be aware that the present result is obtained in the effective-mass approxima-
tion, that is, the Fermi surface is a sphere. Hence, an increase of the (total) electron density
has two effects. First, the electron density at the Fermi level is also increased. Second,
the Fermi velocity is increased, too. Both effects lead to a conductivity proportional to the
total density n, although only electronic states at the Fermi level contribute to the current.
2

149
5.4.2 Ballistic transport
Let us recall the various length scales related with electronic transport: the
spatial extension l of the system, the Fermi wavelength λF , the elastic scattering
length lel , and the phase coherence length lφ . For diffusive transport, one has
l  lMFP  λF . For ballistic transport, however, one has both lφ  l and
lmfp  l. These requirements allow a formulation of the transport within the
framework of the Landauer-Büttiker theory. In particular, scattering at defects
is not likely, for example in defect-free samples of small spatial extension.
A second point considers the difference between conductivity and conduc-
tance. As was sketched in the introduction (page 6), we can speak of the con-
ductivity of, say, Cu but not on that of a tunnel junction. In the latter we have
to speak of its conductance. Conductivity is material-specific, conductance is
device-specific.
In the following we will consider a magnetic tunnel junction (MTJ) which is
build from two ferromagnetic electrodes L and R, both being separated by a
tunnel barrier S (Fig. 5.26). This is a problem already treated in Section 2.7.

Figure 5.26. Magnetic tunnel junction, consisting of electrodes L (cyan) and R

(magenta), separated by a barrier S (yellow) (a) A in L incident scattering channel
(dashed) is scattered at both interfaces and results in a reflected (outgoing in L) and a
transmitted (outgoing in R) scattering channel (solid). (b) Magnetic configurations for
the tunnel junction. In P configuration the magnetizations (arrows) in both electrodes
are parallel, in AP configuration they are antiparallel.

Characteristic quantities are current-voltage characteristics (I(V ) character-

istics) and the tunnel magnetoresistance (TMR). The TMR is the change of the
current flowing through the device upon reversal of the magnetization direction
in one of the electrodes. Typically one measures or computed the conductance
for the P (parallel) and the AP (antiparallel) configuration.

Ballistic transport viewed as transmission – the Landauer-Büttiker theory.

The conductance G = I/∆V can be viewed as a global quantity because it de-
scribes the ration of total current I and voltage drop ∆V between the entrance
and exit of the charge carriers into the device. In contrast, the conductivity
σ can be regarded as a local material-specific quantity because of the concept
of a charge flow (current density) mediated by a local electric field. For short:
conductance describes a system (reservoirs, electrodes and tunnel junction),

150
conductivity describes a material. Material-specific properties determine obvi-
ously the conductance.
The Landauer-Büttiker (LB) theory can be viewed as ‘conductance by trans-
mission’ [93]. The transmission modes are characterized by internal degrees of
freedom, for example wavevector k, and by external parameters, as for exam-
ple applied voltage and dimension of the sample. The transmission modes are
named scattering channels. The conductance G depends on the assumption
that are made for the reservoirs and their coupling to the electrodes (Fig. 5.27).

Figure 5.27. Tunnel junction (schematic). Top: The scattering region S (yellow)
is connected with ideal leads L (left, green) and R (right, cyan). These are supplied
by reservoirs Lres and Rres (circles) with electrons. The arrows represent a scattering
channel incoming from Lres that is on one hand reflected at S and absorbed in Lres
and on the other hand is transmitted into a scattering channel in R and absorbed
in Rres . Bottom: Chemical potentials in the tunnel junction. Lres emits electrons
with energies up to the quasi-Fermi level µLres while Rres emits up to µRres . The
chemical potentials of L and R are denoted µL and µR . The z-axis defines the positive
propagation direction of the electrons.

In the LB theory the following assumptions are made [94]. (i) The reservoirs
Lres and Rres feed only electrons with energies below the chemical potentials
µLres and µRres into the ideal electrodes. (ii) The incoming scattering channels
are emitted incoherently. This way, interference effects between different chan-
nels can be neglected. (iii) Outgoing electrons are completely absorbed in the
reservoirs and thermalized.
A scattering channel iL incoming from the left reservoir Lres is scattered into
an outgoing channel jR in the right electrode R with probability Tji++ . The
−+
reflection probability is Rji , where j indicates a channel in L. The superscripts
denote the propagation direction of the electrons after and before the scattering

151
process. Analogously we have the scattering probabilities for channels incoming
from the right.
For simplicity
P −+ we define the
P total transmission and reflection probabilities
−+ ++ ++
Rj ≡ i Rji and Tj ≡ i Tji into the outgoing channel j in L and R,
respectively.
In the following we consider the case of small bias voltage, that is we as-
sume that the linear-response theory is valid. The current of scattering
channel jL that is injected from reservoir Lres into the system is given by
evjL (∂njL /∂E)(µL − µR ). The density of states of this channel with positive
group velocity vjL is
∂njL 1
= . (5.162)
∂E hvjL
Thus, the current in channel jL is independent of the velocity: IjL = e(µL −
µR )/h. The left reservoir supplies the same current to all channels.
Problem 64. Prove equation (5.162) for free electrons.
The dispersion relation for a free electron is
~2 k 2
E(k) = .
2m
The group velocity v then reads
1 ~
v≡ ∇k E(k) = k.
~ m
Hence, it follows with
r
~ ~ 2mE
v= k=
m m ~2
that the density of states is given by
r −1 r −1
∂n 1 1 m 2mE m 2mE
= = = .
∂E hv 2π~ ~ ~2 2π~2 ~2
This result agrees with eq. (2.67) for the 1D case,
r −1
2m 2mE
N (E) = g ,
2π~2 ~2
except for the factors g and 2. The spin degeneracy g = 2 is included as an internal
degree of freedom in the scattering channel. Hence, in a nonmagnetic lead, there are two
scattering channels that differ only in the spin orientation. The factor 2 is explained by the
fact that the density of states contains all scattering channels, that are those propagating
to the left or to the right. In the LB theory, however, only channels of one propagation
direction are considered.
In summary, ∂n/∂E is the density of states of scattering channels of a single spin orien-
tation and of a single propagation direction. 2
ScatteringPchannel iR is fed by all incoming channels jL . Thus, its current
++
is given by j Tij IjL . The total current is obtained by summing over all
outgoing channels iR ,
e X e X
Itot = (µL − µR ) Ti++ = (µL − µR ) (1 − Ri−+ ), (5.163)
h h
i i

152
 P ++ P −+
where i Ti = i (1 − Ri ) holds because of current conservation (cf.
Sect. 2.7). By dividing the total current by the bias voltage V (eV = µL − µR )
we obtain for the conductance
e2 X ++ e2 X ++
G= Ti = Tij , (5.164)
h h
i ij

that is the sum over all transmission probabilities. The conductance quantum
is G0 = e2 /h.
Note that in the above derivation we could associate to each electrode a
unique chemical potential. This is the case if the voltage drop is localized to
the barrier. The latter is guaranteed by the electronic screening in the metallic
leads.
The maximum conductance is given by the number of scattering channels6
and does not diverge. The ‘problem’ of the finite conductance for perfect trans-
mission is solved by the four-point conductance. Here, the system is extended by
two further contacts at the ideal leads, resulting in µLres 6= µL and µRres 6= µR
(Fig. 5.27). This leads to
P ++
e2 T
G= PNL −+ i i 1 PNR ++ , (5.165)
h 1+ 1 i=1 (R i /viL ) − (T
i=1 i /viR )
gL gR

with gN = i viN , N = L, R. In the case of very small transmissions Ti++

P
the four-point conductance is reduced to the two-point result [eq. (5.164)]. The
group velocities are important only for very thin or conducting barriers S. We
note in passing that a derivation of the LB results within the linear-response
framework of Kubo results in the four-point conductance.
Problem 65. For ideal transmission, Tij++ = 1, the ballistic conductance is finite . But for
diffusive transport we have learned that the resistance is zero for a perfect crystal, page 139.
Can you resolve this contradiction?
Defects in the system result in scattering of electrons out of and into the phase-space
volume (Fig. 5.23). If there are no defects, the number of electrons in that volume is
constant (independent of time). However, their distribution function φ is not that of
equilibrium, that is the Fermi-Dirac distribution f , due to the applied electric field. The
defects lead to relaxation of the non-equilibrium distribution to the equilibrium distribution,
in an average relaxation time τ . Without defects there is no relaxation and, thus τ = ∞.
This implies that the conductivity is infinity because if once accelerated by the electric field,
the electrons stay in this propagating state. There is no ‘internal friction’.
The last point becomes more evident when considering the group velocity, page 152. It
depends only on the dispersion relation, a property of the ideal system.
On the other side, the LB theory relies on uniquely defined chemical potentials in the
reservoirs and in the ideal electrodes. This implies that the voltage drop must happen in
a finite region around the spacer (tunnel barrier). If the system consists only of a single
ideal lead (no barrier), this assumption does not hold and a more careful treatment of
the chemical potentials is needed. This results in the so-called four-point conductance,
eq. (5.165), which takes into account the difference of the chemical potentials of the
6
The conductance of a lead itself, the Sharvin conductance [95] is finite because for this
++
particular ‘tunnel junction’ Tji = δjR iL . The transmission equals then the number of
channels.

153
reservoirs and the electrodes. For ideal transmission (Ti++ = 1, Ri−+ = 0) the conductance
is infinite, as for diffusive transport without defects. 2
The above derivation relies on the linear-response theory, that is for very
small bias voltages. To compute current-voltage characteristics one deliberately
integrates over the ‘energy window of tunneling’,
Z µR
I(V ) = G(E) dE, (5.166)
µL

where the bias voltage is V = (µR − µL )/e.

Problem 66. In the Tersoff-Hamann model, page 105, the current is proportional to
the charge density of the sample. Hence, surface states show up in the current-voltage
characteristics, in both theory and experiment. Do surface states show up in the Landauer-
Büttiker theory? If not, why not? Can you resolve this apparent disagreement between TH
and LB theory?
The TH model does not take into account the scattering boundary conditions, in contrast
to the LB theory. Hence, all electronic states at the sample surface contribute to the tunnel
current, in particular surface states.
The scattering channels in the LB theory are the propagating eigenstates of the ideal
electrodes, that are the Bloch states. These are transmitted through the barrier into Bloch
states of the other lead. Since surface states are orthogonal to Bloch states,7 there must be
a mechanism which scatters Bloch states into and out of surface states. Then surface states
(or any other localized states) would show up in the conductance. Possible mechanisms
are inelastic processes (scattering at phonons, magnons, etc.). Further, the 2D periodicity
parallel to the barrier must be broken, as is in a STM set-up by the STM tip itself of by
interface roughness. In a perfect tunnel junction with 2D translational invariance, surface
states do not contribute to the conductance. 2

Quantum point contact. A quantum point contact (QPC; atomic point con-
tact) is a constriction in which the number of scattering channels can be tuned
by an external parameter. Figure 5.28 shows a QPC in which the reservoirs are
2DEGs. The constriction can be regarded as an ideal lead the DOS of which
can be tuned by a gate voltage. Hence, the scattering channels are transmit-
ted between the reservoirs with probability 1, and the conductance equals the
number of channels (Sharvin conductance).
By increasing the gate voltage, the conductance increases in steps of twice
the conductance quantum (Fig. 5.29), due to spin degeneracy (factor of 2). The
conductance is quantized (conductance quantization).
An alternative QPC is made by the break junction technique. Here, a
junction is bent in order to form an ultrathin wire.

Nanowire. Another type of systems in which the conductance is quantized are

nanowires. Here, the tunable parameter is the diameter of the wire. In analogy
to quantum-well states (Sect. 2.6), the electronic states become quantized in
transverse direction but can move freely along the wire. The ticker the wire the
more states (scattering channels) are occupied and contribute to the current

7
Note that they lie in a band gap, page 65.

154
 Figure 5.28. Quantum point contact.

passed through the wire. As for a QPC, the leads act as reservoirs and the
nanowire itself is taken as ideal lead (Sharvin conductance).
Because the diameter is difficult to control, one makes a series of samples and
records the conductance for each. The resulting histogram (Fig. 5.32) shows
prominent maxima at integer multiples of the conductance quantum. However,
there seem to be significant deviations from perfect transmission, in particular
for thicker wires.

5.4.3 Single-electron tunneling

In the preceding we were concerned with an electron current flowing through a
tunnel junction. The system is in a steady-state. But what might happen if the
current is so small that the discrete nature of charge, namely single electrons,
comes into play?
With decreasing size of electronic devices, classical concepts of electronic
transport become questionable because of quantum-size effects. In particular,
in electronic devices of nanometer size, significant effects due to single electrons
can show up, e.g., charging effects which govern the electrical properties of a
device [96]. A paradigm is single-electron transport: the transfer of single elec-
trons between two leads of a single-electron transistor was shown by Geerlings et
al, allowing single electrons to hop via a small conductive island [97]. Although
this phenomenon was already observed in granular metallic materials more than
three decades ago [96, 98, 99], the use of modern lithographic processes, needed
for the controlled fabrication of devices, allows to study the transport of single
electrons in full detail [100–102].
Resonant tunneling
In the section on tunneling 2.7 we had investigated the transmission of an
electron through a single barrier. Extending the system by a second barrier can
lead to resonant tunneling (Fig. 5.33).
In regions A and E we assume, as in the former, free electrons, whereas in B
and D we have decaying wavefunctions. In region C, quantum-well states can
show up (Section 2.6). The transmission through the entire device is increased

155
Figure 5.29. Conductance quantization in a quantum point contact. The tempera-
ture, given for each data set in Kelvin, must be less than the step spacing to observe
sharp steps (kB T ≤ eVg ).

if the energy of the incoming electron equals that of the quantum-well state,
E0 .8 Note that E0 can be tuned by gate voltage, that is the potential in region
C can be rigidly shifted up- or downward.
Resonant tunneling can be used to build so-called resonant tunneling devices,
for example diodes (Fig. 5.34).
Coulomb blockade
First, us discuss the specific role of Coulomb interaction in a mesoscopic system.
Consider a system with a dot created by a split-gate system. If one transfers
the charge q from the source to the dot the change in the energy of the system
is
q2
∆E = qVG + . (5.167)
2C

8
The transmission can be obtained by the usual method of wavefunction matching, cf.
page 68.

156
Figure 5.30. Quantum point contact, made by etching and break junction technique.

Figure 5.31. Formation of a nanowire, with the initial step shown at the top, the
final step at the bottom. The nanowire is indicated by red atoms.

Here the first term is the energy due to the source of the gate voltage VG whereas
the second one is the energy of Coulomb repulsion at the dot. We describe it by
the effective capacitance C to take into account polarization of the electrodes.
The graph of this function is the parabola with the minimum at

q = q0 − CVG . (5.168)

So it can be tuned by the gate voltage VG . The charge is transferred by electrons,

that is with charge −e (q = −ne). Then, the energy as a function of the number
n of electrons at the grain is

n2 e2
∆E(n) = −neVG + . (5.169)
2C

157
 Figure 5.32. Conductance histogram for Na nanowires.

Figure 5.33. Double barrier, resulting in resonant tunneling.

Hence,

(n + 1)2 e2 n2 e2
∆E(n + 1) − ∆E(n) = −(n + 1)eVG + + neVG −
2C 2C
(5.170a)
e2 e2
= −eVG + n + (5.170b)
C 2C
e2
≈ −eVG + n . (5.170c)
C
Thus, at specific values of VG , namely at
e
VGn = n , (5.171)
C
the difference vanishes. This means that only at these VG values resonant
transfer is possible. Otherwise one has ‘to pay for the transfer’ that means that
only inelastic processes can contribute. As a result, at kB T ≤ e2 /2C the linear
conductance is exponentially small if the above condition (5.171) is met. This
phenomenon is called the Coulomb blockade of conductance.

158
Figure 5.34. Current-voltage characteristics of a double barrier resonant tunnel diode
(center and right). For small voltages the current increases linearly (a) until the bias
voltage equals the energy of a quantum-well state (b). The current drops (c) and
increases again (d), as for field emission.

As a result of the Coulomb blockade, electron tunnel one-by-one, and the

conductance versus gate voltage dependence is a set of sharp peaks (Fig. 5.35).
Or the current-voltage characteristics shows a step-shaped increase, known as
Coulomb staircase. That fact allows one to create a so-called single-electron
transistor (SET) which is the most sensitive electrometer. Such a device (as was
recently demonstrated) can work at room temperature provided the capacitance
(dot size) is small enough.
Problem. Why is n in eq. (5.171) not the total number of electrons on the dot? 2

The ‘orthodox theory’ of Coulomb blockade For simplicity, let us ignore the
discrete character of the energy spectrum of the dot (cf. the section on quantum
wells etc.) and assume that its state is fully characterized by the number n of
excess electrons with respect to an electrically neutral situation. To calculate
the energy of the systems let us employ the equivalent circuit shown in Fig. 5.36.

The left (emitter) and right (collector) tunnel junctions are modeled by par-

159
Figure 5.35. Current-voltage characteristics showing the Coulomb staircase (top)
and differential conductance (Coulomb blockade, bottom).

Figure 5.36. Equivalent circuit for a single-electron transistor. The gate voltage Vg
is coupled to the dot (grain) via the gate capacitance Cg . The voltages Ve and Vc of
emitter and collector are taken with respect to ground.

tial resistances and capacitances. The charge conservation requires that

−ne = Qe + Qc + Qg = Ce (Ve − U ) + Cc (Vc − U ) + Cg (Vg − U ) (5.172)
where U is the potential of the grain. The effective charge of the grain is hence
X X
Q = CU = ne + Ci Vi , C ≡ Ci . (5.173)
i=e,c,g i

This charge consists of four contributions, the charge of excess electrons and
the charges induced by the electrodes. Thus, the electrostatic energy of the

160
grain is
!2
Q2 (ne)2 ne X 1 X
En = = + Ci Vi + Ci Vi . (5.174)
2C 2C C 2C
i i

The last term is not important because it is n-independent. In the stationary

state, the currents through both junctions are the same. Here, we shall con-
centrate on this case. In the non-stationary situation, an electric charge can be
accumulated at the grain, and the currents are different. To organize a trans-
port of one electron one has to transfer it first from emitter to grain and then
from grain to collector. The energy cost for the first transition,
(2n + 1)e2 e X
En+1 − En = + Ci Vi (5.175)
2C C
i

must be less than the voltage drop eVe . In this way we come to the criterion

En − En+1 + eVe ≥ 0. (5.176)

Similarly, to organize the transport from the grain to the collector one needs

En+1 − En − eVc ≥ 0. (5.177)

These inequalities provide the relations between Ve , Vc , and Vg to make the

current flow. For simplicity let us consider a symmetric system, where

Ge = Gc = G, (5.178a)
C
Ce = Cc ≈ (Cg  C), (5.178b)
2
Vb
Ve = −Vc = , (5.178c)
2
where Vb is bias voltage. Then we get the criterion
|e| Cg
Vb ≥ (2n + 1) − 2 Vg . (5.179)
C C
We observe that there is a threshold voltage which is necessary to exceed to or-
ganize transport. This is a manifestation of Coulomb blockade. It is important
that the threshold depends linearly on the gate voltage which makes it possible
to create a transistor. Of course, the above considerations are applicable at
zero temperature.

The single-electron transistor. The simplest device in which the effect of

Coulomb blockade can be observed is the single-electron transistor (SET). It
consists of two tunnel junctions sharing one common electrode with a low self-
capacitance, known as the island. The electrical potential of the island can be
tuned by a third electrode (the gate), capacitively coupled to the island.
In the blocking state no accessible energy levels are within tunneling range
of the electron (red; see top left of Fig. 5.37) on the source contact. All energy
levels on the island electrode with lower energies are occupied.

161
Figure 5.37. Single-electron transistor (SET). Left: Atomic-force microscopy image
of an SET. The red region is the island of size 100 × 200 nm. Right: energy diagram
of an SET (see text).

If a positive voltage is applied to the gate electrode, the energy levels of

the island electrode are lowered. The electron (green, 1.) can tunnel onto
the island (2.), occupying a previously vacant energy level. From there it can
tunnel onto the drain electrode (3.) where it is inelastically scattered and
reaches the drain electrode Fermi level (4.). Therefore, the conductance of this
device should considerably increase if the tunnel process is allowed. Otherwise
the conductance is small due to the blockade state. Plotting the conductance
versus the gate voltage one observes equidistant sharp maxima (Fig. 5.38).
The energy levels of the island electrode are evenly spaced with a separation
of ∆E. ∆E is the energy needed for the subsequent electron to tunnel to the
island, acting as a self-capacitance C. The lower C, the bigger ∆E. It is crucial
for ∆E to be larger than the energy of thermal fluctuations kB T , otherwise an
electron from the source electrode can always be thermally excited onto an
unoccupied level of the island electrode, and no blocking can be observed.

162
Figure 5.38. Conductance of a quantum dot (‘island’) versus gate voltage.

163
6 Magnetic nanostructures
Nanostructures often comprise magnetic subsystems, such as magnetic elec-
trodes or magnetic spacers. This allows to combine conventional electronics
(which exploits the electronic charge) with the magnetic moment of the elec-
trons, resulting in so-called spintronics or magnetoelectronics [103–108].
In this chapter we are going to enhance the methods introduced in the chap-
ter of electronic transport (Chapter 5) in order to obtain theories for spin-
dependent transport.

6.1 Tunnel magnetoresistance

A magnetic tunnel junction (MTJ) consists of two magnetic electrodes that are
separated by an insulating spacer. Hence, for electronic transport, the electrons
have to tunnel through the barrier. Typical electrodes are made of Fe and Co,
spacers of MgO or Al2 O3 .
Transmission electron microscope (TEM) images of a Fe/MgO/Fe MTJ are
shown in Fig. 6.1. [110] In the atomically resolved picture (b) lattice dislocations
are indicated by circles, giving evidence that the sample is of high quality.
Since the number of defects is small, the transport is expected to be ballistic,
not diffusive. Note that the theory of ballistic transport is applicable only for
small and defect-free samples. It is important to notice at this point that TEM
does not provide detailed information on the interface structure (substitutional
disorder, page 71, occupation probabilities of sites, relaxation, etc). As it turned
out, surface x-ray diffraction (SXRD) measurements gave string evidence that
the layers at the Fe/MgO interfaces are disordered [111, 112].
Obviously, the current passed through the device depends on the relative
orientation of the magnetizations in the ferromagnetic electrodes. Typically one
distinguishes two basic configurations. In the parallel configuration (P) the lead
magnetizations are aligned parallel whereas in the antiparallel configuration
(AP) the lead magnetizations are aligned antiparallel (Fig. 6.2).
The orientation can be switched by an external magnetic field, provided that
the magnetization of one of the electrodes is ‘pinned’ by an adjacent antiferro-
magnetic layer. The pinned electrode can be viewed as the magnetically ‘hard’
one, the other as ‘soft’. For large external fields, the magnetizations are ori-
ented parallel while for smaller fields they are aligned antiparallel (Fig. 6.3, left).
Hence, the resistance of the MTJ shows a hysteresis if plotted versus external
field H (Fig. 6.3, right). The hysteresis of the resistance of a Fe/MgO/Fe MTJ
is shown in Fig. 6.4a.
The resistance of the AP configuration is typically much larger than that of

164
Figure 6.1. TEM images of a single-crystal MTJ with the
Fe(001)/MgO(001)(1.8 nm)/Fe(001) structure. b is a magnification of a. The
vertical and horizontal directions respectively correspond to the MgO[001] (Fe[001])
axis and MgO[100] (Fe[110]) axis. Lattice dislocations are circled. The lattice spacing
of MgO is 0.221 nm along the [001] axis and 0.208 nm along the [100] axis. The lattice
of the top Fe electrode is slightly expanded along the [110] axis.

Figure 6.2. Magnetic tunnel junction in parallel (left) and antiparallel (right) config-
uration.

the P configuration, cf. Fig. 6.4b. The magnetoresistance ratio is defined as

R(H)
MR ≡ , (6.1)
R(0)
where R(H) is the field-dependent resistance. In the following, we are inter-
ested in the tunnel magnetoresistance (TMR), that is the change of the tunnel
current upon reversal of the relative magnetization orientation,
I(P) − I(AP)
δ≡ , (6.2)
I(P) + I(AP)
in the field-free case (H = 0). Defined as an asymmetry, the TMR is in [−1, +1].
There are other definitions, for example the optimistic one,
I(P) − I(AP)
δ≡ , (6.3)
I(AP)

165
Figure 6.3. Magnetization (left) and resistance (right) of an MTJ versus strength of
the external magnetic field.

and the pessimistic one,

I(P) − I(AP)
δ≡ . (6.4)
I(P)
Since typically I(P)  I(AP), the former definition results in larger values than
the latter.
Besides the dependence on the external magnetic field, the TMR depends also
on the thickness of the spacer. For Fe/MgO/Fe, some samples show distinct
oscillations in the TMR (Fig. 6.4c). A typical finding is that the TMR increases
significantly upon increasing the number of spacer layers, up to a saturation.
This behaviour is also found in the calculated TMR for realistic samples.
Another important dependence is that on the bias voltage (Fig. 6.5a). For
samples with oxide spacers, the TMR shows a distinct maximum at zero bias
and drops (almost symmetrically) with increasing bias voltage (in absolute val-
ues). This effect is called zero bias anomaly.
In summary, a complete theory of the TMR needs to take into account the
following aspects.
Structure To describe realistic samples, the geometric, electronic, and magnetic
structure of the MTJ must be accounted for as good as possible, that is
on a DFT basis.

Disorder Tunnel junctions are not perfect, although the sample preparation
has considerably improved. Hence, the transport through the device is
not fully ballistic but also diffusive (quasi-ballistic), as noted in the intro-
duction (Sect. 1.2 and Fig. 5.22). Theories for ballistic transport must be
extended to treat diffusive transport, too. This can be achieved by either
a supercell approach or by using the coherent potential approximation
(pages 71 and 74).

Bias voltage Theories typically assume the linear-response regime, that is they
are valid for small bias voltages (page 152). Nonzero bias voltages can be
treated by non-equilibrium Green functions (Keldysh formalism).

166
Figure 6.4. Tunnel magnetoresistance of Fe(001)/MgO(001)/Fe(001) junctions. a,
Magnetoresistance curves (measured at a bias voltage of 10 mV) at T = 293 K and
20 K (MgO thickness tMgO = 2.3 nm). The resistance-area product R A plotted here is
the tunnel resistance for a 1 × 1µm area. Arrows indicate magnetization configurations
of the top and bottom Fe electrodes (Fig. 6.1). The MR ratio is 180 % at 293 K and
247 % at 20 K. b, R A at T = 20 K (measured at a bias voltage of 10 mV) versus tMgO .
Open and filled circles represent parallel and antiparallel magnetic configurations. The
scale of the vertical axis is logarithmic. c, MR ratio at T = 293 K and 20 K (measured
at a bias voltage of 10 mV) versus tMgO .

External magnetic field Experiments also consider the dependence on the ex-
ternal magnetic field that is not treated in theory. Strictly speaking,
theories are valid for the TMR in remanent samples.
A presentation of sophisticated theories of the TMR is beyond the scope of
the lecture. Instead, we introduce to two popular models, the Jullière-Maekawa-
Gafvert model and the Slonczewski model.
The Jullière-Maekawa-Gafvert model
The probably most simple model for ballistic transport through a tunnel junc-
tion was proposed by Jullière [113]. It completely neglects the properties of the
tunnel barrier and neglects further detailed properties of the electrodes. The
conductance is just given by the number of spin-up and -down electrons in the
leads or, following Maekawa and Gafvert [114], by the density of states at the
Fermi level.

167
Figure 6.5. Bias voltage dependence of the TMR effect. Relation between bias voltage
V and the normalized MR ratio at room temperature of Fe(001)/MgO(001)/Fe(001)
tunnel junctions with various MgO thicknesses tMgO . The direction of bias voltage is
defined with respect to the top electrode.

The conductance G is proportional to the density of states of the electrode

that supplies electrons (source) and proportional to that of the electrode which
absorbs electrons (drain). For small bias voltages the tunnel energy is the
common Fermi level EF of the leads. If the spin is conserved during the tunnel
process (no spin-orbit coupling) one has for the P configuration G(P) ∝ L↑ R↑ +
L↓ R↓ , where L and R are the density of states in the left and in the right
electrode, respectively. In the AP configuration the orientation of the magnetic
moments reverses in one of the electrodes, say in the right one. Because now the
↑-electrons are the minority electrons and the ↓-electrons are majority electrons,
one has G(AP) ∝ L↑ R↓ + L↓ R↑ . For transition metals the majority DOS at
EF is smaller than the minority DOS (cf. for example Fig. 6.48), implying
G(P) > G(AP) (Fig. 6.6).
The tunnel magnetoresistance δ is defined as the asymmetry of the conduc-
tances,
G(P) − G(AP)
δ≡ , (6.5)
G(P) + G(AP)
as is clear from I = GV and eq. (6.2). It can be expressed as

δ = PL PR , (6.6)

with the spin polarizations PL and PR of the electrodes,

N↑ − N↓
PN = , (6.7)
N↑ + N↓
N = L, R [eq. (5.35)]. This result implies that the conductance depends solely
on the spin polarization of the electrodes, a statement that is not supported by

168
Figure 6.6. Jullière model for the tunnel magnetoresistance. (a) In P-configuration
electrons tunnel from the left lead L to the right lead R. The conductance is propor-
tional to the product of the spin-resolved densities of states (cyan and green) at the
Fermi energy EF . The spin orientation is indicated by the vertical arrows. (b) In the
AP configuration, majority and minority electrons change roles in the right electrode.

experiments. However, the TMR is energy-dependent. So it could explain the

zero-bias anomaly (ZBA, see below) as a density-of-states effect.
Problem 67. Derive equation (6.6). Derive similar expressions for the optimistic and the
pessimistic definitions.
From equation (6.5) follows
L↑ R↑ + L↓ R↓ − L↑ R↓ − L↓ R↑
δ =
L↑ R↑ + L↓ R↓ + L↑ R↓ + L↓ R↑
L↑ (R↑ − R↓ ) − L↓ (R↑ − R↓ )
=
L↑ (R↑ + R↓ ) + L↓ (R↑ + R↓ )
(L↑ − L↓ )(R↑ − R↓ )
=
(L↑ + L↓ )(R↑ + R↓ )
= PL P R .

For the optimistic TMR we first consider

(L↑ − L↓ )(R↑ − R↓ )
1 − PL P R = 1−
(L↑ + L↓ )(R↑ + R↓ )
(L↑ + L↓ )(R↑ + R↓ ) (L↑ − L↓ )(R↑ − R↓ )
= −
(L↑ + L↓ )(R↑ + R↓ ) (L↑ + L↓ )(R↑ + R↓ )
L↑ R↑ + L↓ R↓ + L↑ R↓ + L↓ R↑ L↑ R↑ + L↓ R↓ − L↑ R↓ − L↓ R↑
= −
(L↑ + L↓ )(R↑ + R↓ ) (L↑ + L↓ )(R↑ + R↓ )
L↑ R↓ + L↓ R↑
= 2 .
(L↑ + L↓ )(R↑ + R↓ )
Hence,
1
G(AP) = (1 − PL PR )(L↑ + L↓ )(R↑ + R↓ ).
2

169
With this one derives
L↑ R↑ + L↓ R↓ − L↑ R↓ − L↓ R↑
δ =
L↑ R↓ + L↓ R↑
(L↑ − L↓ )(R↑ − R↓ )
=
L↑ R↓ + L↓ R↑
(L↑ − L↓ )(R↑ − R↓ )
= 2
(1 − PL PR )(L↑ + L↓ )(R↑ + R↓ )
PL P R
= 2 .
1 − P L PR
For the pessimistic TMR it follows
L↑ R↑ + L↓ R↓
1 + PL P R = 2 .
(L↑ + L↓ )(R↑ + R↓ )
Thus,
1
G(P) = (1 + PL PR )(L↑ + L↓ )(R↑ + R↓ ),
2
and eventually
P L PR
δ=2 .
1 + PL P R
2
Although the Jullière model accounted more or less correctly for the MR
ratio, the connection of the tunnel current to the spin polarization remained
somewhat unclear. In fact, the model was used to conclude from the TMR
on the spin polarization, leading to inconsistent results. Since properties of the
tunnel barrier do not enter the above model, Maekawa and Gafvert replaced the
spin polarization of the electrodes by that at the interfaces [114]. But like the
Tersoff-Hamann model, a description of transport in terms of the (spin-resolved)
density of states is rather poor. The validity of the Jullière model was checked
by comparing its TMR to that obtained within Landauer-Büttiker theory for
realistic MTJs, [115] giving strong support for the LB theory. Therefore, we
turn now to a theory inspired by the work of Landauer-Büttiker.
Tunneling of free electrons — the Slonczewski model
The tunnel current should mediated essentially by delocalized electrons,
whereas localized electrons should not contribute significantly (Fig. 6.7).1 This
suggests to propose a model for free electrons. In case of small bias voltages,
the tunnel barrier can be further taken as a step-shaped barrier. This model
was suggested by Slonczewski [116] and will be sketched for the 1D case.
The tunnel energy E is less than the barrier height W (0 < E < W ). The
electronic states are chosen for scattering boundary conditions (Section 2.7).
The ↑-wavefunctionsp in the left electrode consist of an incoming plane wave
with amplitude 1/ kL↑ and a reflected wave with amplitude R↑ ,
1
ΨL↑ (x) = p exp(ikL↑ x) + R↑ exp(−ikL↑ x). (6.8)
kL↑
1
Confer for example the band structure of Cu, Fig. 3.17. The strongly dispersive sp-band
can be regarded as due to delocalized electrons whereas the almost flat d-bands are due to
localized electrons. For Fe and Co one finds similar but spin-split bands.

170
Figure 6.7. Electronic bands in bulk fcc Ni in the [110] direction for the majority-spin
(a) and minority-spin (b) electrons. The heavy curves show the free-electron-like bands
which dominate tunneling. k ↑ and k ↓ are the Fermi wavevectors that determine the
spin polarization of the tunnel current.

The incoming partial wave is normalized to unit current. Within the barrier
there are exponentially increasing and decreasing plane waves,

ΨS↑ (x) = A↑ exp(κ↑ x) + B↑ exp(−κ↑ x), (6.9)

whereas is in the right electrode there is only an outgoing partial wave,

ΨR↑ (x) = T↑ exp(ikR↑ x). (6.10)

An analogous ansatz is made for the ↓-wavefunction. The wavenumbers are

√
k↑ = 2E, (6.11a)
p
k↓ = 2(E − V0 ), (6.11b)
p
κ = 2(W − E), (6.11c)

with the exchange splitting V0 > 0 in the electrodes (Fig. 6.8). The eight
coefficients (R, A, B and T , each for the two spin orientations ↑ and ↓) are
obtained from the continuity requirement of the total wavefunctions at the two
interfaces (left lead–step and step–right lead).
The resulting TMR

δ = PL PR ∆ (6.12)

171
Figure 6.8. Slonczewski model for the tunnel magnetoresistance. An electron incom-
ing in the left lead (1/k 1/2 ) is reflected at the step barrier (R) and transmitted into the
right lead (T ). The constant potentials in the leads are spin-dependent (black arrows
in R for P-configuration, blue arrows for AP-configuration).

complies with Jullière’s result, that is with the product of the spin polarizations
kN ↓ − kN ↑
PN = , (6.13)
kN ↑ + kN ↓

N = L, R, of the electrodes. But it is modified by a transmission factor

κ2 − kL↑ kL↓ κ2 − kR↑ kR↓

∆= . (6.14)
κ2 + kL↑ kL↓ κ2 + kR↑ kR↓

It comprises the electronic structures of the electrodes and of the barrier via
the wavenumbers. Here, properties of the tunnel barrier influence the TMR
significantly.
Problem 68. Derive equation (6.13).
Without loss of generality we can choose the spin-quantization axis as the z-axis (no
spin-orbit coupling!). The incident spin-up wavefunction is then given by
1
Ψ↑ (x) = p eik↑ x χ+ ,
k↑

the spin-down wavefunction by

1
Ψ↓ (x) = p eik↓ x χ− ,
k↓

The total wavefunction (coherent superposition) reads

Ψ(x) = Ψ↑ (x) + Ψ↓ (x).

172
 The z-component of the spin polarization is given by

Ψ† (x)σz Ψ(x)
Pz (x) =
Ψ† (x)Ψ(x)
(Ψ†↑ (x) + Ψ†↓ (x))σz (Ψ↑ (x) + Ψ↓ (x))
=
(Ψ†↑ (x) + Ψ†↓ (x))(Ψ↑ (x) + Ψ↓ (x))
(Ψ†↑ (x) + Ψ†↓ (x))(Ψ↑ (x) − Ψ↓ (x))
=
(Ψ†↑ (x) + Ψ†↓ (x))(Ψ↑ (x) + Ψ↓ (x))
Ψ†↑ (x)Ψ↑ (x) − Ψ†↓ (x)Ψ↓ (x)
=
Ψ†↑ (x)Ψ↑ (x) + Ψ†↓ (x)Ψ↓ (x)
1 1
k↑ − k↓
= 1 1
k↑ + k↓
k↓ − k↑
= ,
k↑ + k↓

which establishes eq. (6.13).

At a given energy, k↓ < k↑ , resulting in a negative spin polarization. This is consistent
√
with the density of states which decreases with 1/ E in the 1D case. 2
The above result gives rise to an effective spin polarization, that is the spin
polarization of the electrodes (which is given by k ↑ and k ↓ ) influenced by the
tunnel barrier (which depends on κ), see Fig. 6.9. In the limit of a high barrier
it tends to unity, reducing Slonczewski’s result for TMR to Jullière’s formula.
However, if the barrier is not very high and the decay constant is comparable
to or less than the wavevectors of electrons in the ferromagnetic metals, the
magnitude of the TMR decreases with decreasing step height and even changes
sign for sufficiently low barriers, which is demonstrated in Fig. 6.9.

Figure 6.9. Spin polarization of the tunnel conductance as a function of the normal-
ized potential barrier height for various values of k ↑ /k ↓ .

Problem 69. Show that Slonczewski’s result for the TMR reduces to that of Jullière’s for
increasing barrier height.

173
 To obtain Jullière’s result for the TMR, δ = PL PR , from Slonczewski’s result, δ =
PL PR ∆, one has to show that ∆ → 1 for increasing barrier height.
Increasing barrier height means κ → ∞. It follows

κ2 − kL↑ kL↓ κ2 − kR↑ kR↓

∆ =
κ2 + kL↑ kL↓ κ2 + kR↑ kR↓
kL↑ kL↓ kR↑ kR↓
1− κ2 1− κ2
= kL↑ kL↓ kR↑ kR↓
1+ 1+
κ2 κ2
1−0 1−0
→
1+0 1+0
= 1.

2
The functional form of the TMR remains unchanged if the model is extended
to the 3D case. After integration over all in-plane wavevectors, however, this
form is lost. The Slonczewski model describes the ‘conductance as transmis-
sion’, as in the Landauer-Büttiker theory (Sect.5.4.2). In fact, it is the LB
theory specialized for free electrons.
Problem 70. For the Slonczewski model, discuss how the transmission behaves as a
function of kk .
The electrons move freely parallel to the barrier, i. e. along ρ = (x, y). Hence, the total
energy can be written as

E = Eρ + Ez

(Sect. 2.6), with

1 2
Eρ = k .
2 k
Since the total energy E is given (fixed), the effective energy for a given kk is

1
Ez = E − k2k .
2
Thus, with increasing |kk | the effective energy decreases, until Ez = 0. Therefore, the
transmission decreases, too (Fig. 2.15).
Further, the effective barrier width increases with |kk | (E). The length l of the path
from one side of the barrier to the other is given by
d
l= ,
cos ϑ
where d is the nominal barrier width and ϑ is the angle of the wavevector (kk , κ) with the
z axis. For kk = 0, l = d. For ϑ → π/2 it follows l → ∞, leading to a stronger decrease
of the wavefunction and consequently to less transmission. 2

Examples
In the following different aspects of the spin-dependent transport in a MTJ will
be considered.
The first observation of a TMR goes back to Jullière [113], as is shown in
Fig. 6.10. Note the the effect shows up at a rather low temperature. Further, the
bias-voltage range was very small, as compared to today’s experiments. It took

174
Figure 6.10. The original demonstration of the TMR effect. The relative conductance
change due to an applied magnetic field versus applied bias in a Fe/Ge/Co junction at
4.2 K.

Figure 6.11. The first observation of reproducible, large room temperature magne-
toresistance in a CoFe/Al2 O3 /Co MTJ. The arrows indicate the relative magnetization
orientation in the CoFe and Co layers.

up to twenty years to observe a reproducible TMR at room temperature [117]

(Fig. 6.11).
As pointed out in the introduction of this section, one of the electrodes is
adjacent to a ‘pinning layer’ in order to make that electrode the magnetically
‘hard’ one. The unpinned electrode remains magnetically ‘soft’. Attaching a
ferromagnetic pinning layer results in symmetric hysteresis loops in the TMR
dependence on the external magnetic field (Fig. 6.12, left). If the pinning layer
is an antiferromagnet, the hysteresis loop becomes shifted (Fig. 6.12, right),
due to exchange bias (page 285).
If the applied magnetic field is less than the coercive field of one of the
electrodes but larger than that of the other, the magnetization of the mag-

175
Figure 6.12. Magnetoresistance versus magnetic field for a hard-soft MTJ (a) and an
exchange-biased MTJ (b), both at 10 K. Vertical arrows refer to sweep direction. Both
curves are taken at V = 0.

netically soft electrode can be rotated by the external field. This allows to
investigate the angular dependence of the resistance (Fig. 6.13). [118] At an

Figure 6.13. The angular dependence of the magnetoresistance of a CoFe/Al2 O3/Co

junction measured in a magnetic field lower than the coercive field of one electrode but
higher than the coercive field of the other electrode.

relative angle of 0 ◦ the resistance should be small (P configuration), at angles

of ±180 ◦ it should be large (AP configuration), as is evident from the figure
(see also Ref. [119]). Thus, the TMR is anisotropic [tunnel anisotropic mag-
netoresistance (TAMR)]. Theory predicts a cos-dependence [120], as is found
experimentally.
Problem 71. Looking at Fig. 6.13 in more detail, one finds significant deviations from a
perfect cos-dependence. What could be the origin?
The angular dependence is obtained for an external magnetic field less than the coercive
field (Fig. 6.14) of one electrode (say, the bottom electrode) but larger than that of the other
(the top electrode). In this case, the magnetizations in the individual magnetic domains

176
Figure 6.14. Magnetization versus external magnetic field. The insets sketch the
magnetic structure at several points of the hysteresis.

of the bottom electrode (Fig. 6.15) are not fully aligned with the external magnetic field
but the net magnetization is pinned by the pinning layer. Hence, it does not align with
the external field it the latter is rotated about the sample normal. nonzero. The top
electrode, however, is almost completely saturated (Fig. 6.14), because not pinned, and
its magnetization follows the rotation of the external field. However, both electrodes are
not saturated, that is the tunnel junction is not in perfect P configuration if the external
magnetic field is along the magnetization in the pinning layer.

Figure 6.15. Magnetization versus external magnetic field. The insets sketch the
magnetic structure at several points of the hysteresis.

In summary, the net magnetization of the top electrode is misaligned with the external
magnetic field which shows up as a shift of the minimum in Fig. 6.13 by a few degree to
the left and as deviation from a perfect cos-shape. 2
In the preceding it was motivated that the TMR is positive. However, there
are examples for which the TMR is negative, that is the current for the AP

177
configuration is larger than for the P configuration. Figure 6.16b shows the bias
dependence of the TMR of a MTJ with Co and the perovskite La0.7 Sr0.3 MnO3
(LSMO) electrodes. The TMR does not show the typical peak at zero bias (cf.

Figure 6.16. (a) A schematic diagram of the spin-polarized densities of states of

LSMO (as derived from photoemission) and the Co(100) surface (calculated). (b)
TMR ratio versus applied bias for a Co/SrTiO3 /LSMO junction at 5 K. Inverse TMR
is observed for V < 0.8 V, while normal TMR is observed for V > 0.8 V, indicating that
the Co/SrTiO3 spin polarization is negative for V < 0.8 V. Inset: TMR ratio versus
applied bias for a Co/Al2 O3 /SrTiO3 /LSMO junction. In this case, the polarizations
of Co/Al2 O3 and LSMO are both positive, and a normal (positive) TMR is seen.

the inset as well as Figs. 6.5 and 6.10). Instead it displays a small TMR at V >
0.8 V and, more striking, a sizable negative TMR at V < 0.8 V. This finding
is explained within the Jullière-Maekawa-Gafvert model, keeping in mind its
limitations. For V = 0, the spin polarization of Co is negative (at the Fermi
level, right in Fig. 6.16), resulting in a negative TMR. Considering only the spin
polarization in the electrodes, as in the strict Jullière model, this could not
explain the positive TMR found for Co/Al2 O3 /SrTiO3 /LSMO, since barrier
properties do not enter. Hence, it was concluded that the spin polarization at
the Co/Al2 O3 changes sign with respect to that of Co/SrTiO3 . Hence, the spin
polarization at the interfaces has to be considered, as in the Maekawa-Gafvert
model.
In Figure 6.4c a distinct oscillation in the spacer-thickness dependence of the
TMR in Fe/MgO/Fe MTJs was found, the origin of which appears unresolved
up to now. Since the period is about twice the MgO interlayer spacing, it might
possibly be related to the geometric structure. The stacking sequence in the
MgO spacer might lead then to an even/odd effect.
Introducing a conducting spacer in addition to the insulating one, results
in two main effects. First, the TMR decreases significantly upon increasing
the Cu spacer thickness (Fig. 6.17). Second, it displays significant oscillations
with a period of 11.4 Å. This value agrees well with that found for the interlayer
exchange coupling mediated by Cu spacers (about 11.0 Å), suggesting to explain
both findings by the same mechanism.

178
Figure 6.17. TMR at 2 K and a bias of 10 mV as a function of Cu interlayer thickness
for Co(001)/Cu(001)/Al2 O3 /Ni80 Fe20 junctions. The period of the oscillation observed,
11.4 Å, is in agreement with the Fermi surface of Cu (Fig. 5.2).

The spin-dependent reflectivity is illustrated in Fig. 6.18a, where it is assumed

that electrons with their spins parallel (antiparallel) to the magnetizations M 1
and M 2 are weakly (strongly) reflected at these interfaces. This spin-dependent
reflection at the interfaces gives rise to spin-dependent quantum well states that
cause an oscillatory polarization in the interlayer. As a result there are spin-
dependent interference effects like the formation of standing electron waves for
certain interlayer thicknesses (a). Due to the similarity of the arrangement with
an optical Fabry-Perot interferometer this is sometimes also called the Fabry-
Perot model of oscillatory coupling. Oscillations in the coupling observed as
the thickness of a magnetic layer or other overlayer is varied are explained by
additional spin-dependent interference effects. The predominant contribution
to the coupling is from electrons with wavevectors, Qi , that are critical span-
ning vectors of the Fermi surface of the interlayer material, i. e., vectors in the
direction perpendicular to the interface that connect two sheets of the Fermi
surface parallel to each other. For the Fermi surface shown in panel b, there are
two such vectors in the [100] and one vector in the [111] direction. The periods
of the oscillatory coupling are given by λ = 2π/Qi and thus are determined
solely by the electronic properties of the interlayer material.
In the tunneling through a step-shaped barrier, the transmission takes places
via decaying states in the barrier. In that case, there is only a single pair of
states, one decaying to the left, the other decaying to the right. When tunneling
through an insulator, however, the transmission takes place via several decay-
ing states within the fundamental band gap of the insulator (complex band
structure). The question is what states have the largest transmission.
Figure 6.19 shows the decay of Bloch states through a thick ZnSe spacer at
kk = 0, the center of the 2BZ. The most delocalized states, namely the sp-

179
Figure 6.18. Illustration of spin-dependent reflectivity at nonmagnetic/magnetic
interfaces for the explanation of oscillatory coupling (a). Panel (b) shows a cross section
of a Fermi surface with critical spanning vectors in the [100] and [111] directions.

states have the least decay constant. The localized d-states, however, decay
much stronger. This finding gives support to the Slonczewski model which
considers the transmission of delocalized states (free electrons).

Figure 6.19. The calculated layer-dependent DOS for the majority (a) and minority
(b) Bloch states at kk = 0 for the Fe/ZnSe/Fe junction. Different decay rates and
injection efficiencies for the states of different characters and symmetries are seen.

Disorder at an interface results in incoherent scattering, that is, the in-plane

wavevector kk is not conserved. Hence, we are concerned with both ballistic and
diffusive transport. Both transports regimes can be treated within a supercell
approach (page 71), within the virtual crystal approximation (page 74) or within
the coherent potential approximation (page 74).
As an example we discuss the transmission through an Fe/Cr interface [121].
The effect of disorder is shown in Fig. 6.20. Without disorder, the diffusive

180
Figure 6.20. Spin-resolved transmission through an Fe/Cr interface versus degree
disorder for several theoretical treatments of the disorder. The disorder is limited to
one layer.

contribution vanishes, due to the 2D translational invariance (a, majority). It

becomes maximal at 50 % disorder. While the CPA, averaged t-matrix approx-
imation (ATA) and the supercell approach give similar results for the specular
(coherent) transmission, the two former do not account for the diffusive contri-
bution because vertex corrections are not considered in this calculation. The
VCA deviates considerably from the other approximations.
Increasing the interface thickness, diffusive transmission becomes increasingly
important (Fig. 6.21). The specular part, however, decreases rapidly. This find-
ing emphasizes that ballistic transport requires small and defect-free samples.

Problem 72. Why can one apply a theory for ballistic transport, the Landauer-Büttiker
theory, to diffusive scattering at a disordered interface? Can one use the general or simplified
Boltzmann equation instead?
The point is in what part of the sample disorder shows up. The LB theory is valid for
ideal (defect-free) leads connected to a scattering region of finite size, so that scattering
boundary condition can be applied. Further assumptions on the scattering region are not
made. Hence, there can be either specular or diffusive scattering (or both). Therefore, the
LB theory is valid for such systems.
In principle, the general Boltzmann equation can also be applied to tunnel junctions

181
Figure 6.21. Majority-spin transmission through an Fe/Cr interface versus interface
thickness for several theoretical treatments of the disorder.

with a disordered scattering region. The simplified Boltzmann equation, however, cannot
be applied since it was derived for a homogenous sample. Hence, the disorder must be
present in the entire sample, which is not the case. 2

Fe/MgO/Fe
As an example worth to be looked into more detail, we turn to the case of
Fe/MgO/Fe tunnel junctions. The band structures of Fe and MgO are depicted
in Fig. 6.22 for kk = 0.
Problem 73. Find out the sp- and the d-bands of Fe.
The spin-split sp-bands of Fe have the largest dispersion (Fig. 6.22). Their band max-
imum is at about 10 eV at the N point. They hybridize with the d-band range which
shows up between −4.5 eV and 0 eV for spin-up (blue) and between −[3]eV and 2.5 eV
for spin-down (red). The sp-bands cross the −5 eV-line at about 0.4 of the Γ–∆–N line,
dispersing downward to reach the Γ point. 2
The fundamental band gap of MgO is about 7.2 eV in experiment but much
less in theory due to the use of the LSDA.2 The Bloch states of the Fe elec-
trodes are transmitted through the MgO spacer via decaying states within the
fundamental band gap of MgO (cf. the complex band structure).As for ZnSe,
the sp-states decay least and, thus, contribute most to the conductance. The
energy position of the fundamental gap is roughly centered about the Fermi
level of Fe. Note that its energy dependence determines the decay constants.
The ballistic transmission of a Fe/Fe junction was computed within the
Landauer-Büttiker formalism using multiple-scattering theory (layer-KKR in
DFT-LSDA). In P configuration the transmission Tij++ (E, kk ) equals δij , that
is perfect transmission of a Bloch state (scattering channel) with band index i

2
The LSDA treats electronic correlations well for localized electrons because it is not free
of electronic self-interaction. The self-interaction correction (SIC) improves the band gap
significantly.

182
Figure 6.22. Band structures of Fe (a, left) and MgO (b, right) along the [001]
direction, kk = 0.

at energy E and wavevector kk . Hence, the conductance

Z X
G= Tij++ (E, kk ) dk 2 (6.15)
2BZ ij

is the Sharvin conductance and is limited by the number of scattering channels.

The transmission
X X
Tij++ (EF , kk ) = δij = N (EF , kk )
ij ij

is shown in Fig. 6.23, where N (EF , kk ) is the number of Bloch states at (EF , kk ).
It gives a projection of the Fermi surface of Fe onto the (001) plane (Fig. 6.24).

Problem 74. Identify the structures in the transmission, Fig. 6.23, in the Fermi surface,
Fig. 6.24.
First of all, one has to notice that the transmission is given by the number of bands
with positive group velocity. Hence, not all bands in the Fermi surface show up in the
transmission map.
For spin-down, the purple octahedron at the BZ center gives rise to the transmission 2.
The yellow areas at the corner increase the transmission to 4.

183
 Figure 6.23. Spin-resolved transmission of an Fe/Fe junction in P configuration.

For spin-up the large surface gives a transmission of 2. The green lobes increase the
transmission to 3. Note that green lobes on opposite sides of the BZ are perpendicular to
each other. 2
The interface geometry was first assumed as ideal, that is the Fe(001) surface
was ‘glued’ to the MgO(001) surface without relaxation and reconstruction.
Surface x-ray diffraction, however, shows that a partially occupied FeO interface
can be formed [112]. Depending on the growth conditions, only one interface
(asymmetric case) or both interfaces (symmetric case) of a MTJ can be of FeO
type.
Problem 75. Sketch the ideal Fe/MgO interface in [001] orientation. Fe has a bcc lattice,
MgO crystallizes in the rock-salt structure.
See Figure 6.26. 2
One important issue in perfect MTJs is resonant transmission (as for single-
electron tunneling), in which case the conductance is drastically enhanced. This
‘hand shake’ results in so-called hot spots in the wavevector-resolved transmis-
sion, that are small areas in the 2BZ in which the transmission is very large
compared to that in other areas. Small perturbations, such as a bias voltage or
disorder, destroy the ‘hand shake and the conductance is significantly reduced.

184
 Figure 6.24. Spin-resolved Fermi surfaces of Fe (spin-up left, spin down right).

Figure 6.25. Structure of Fe/MgO interfaces with FeO layer, as obtained from SXRD
experiments.

With increasing MgO thickness, the conductance decreases due to the decay
of electronic states in the barrier. Since the transmission takes place via different
pairs of Bloch states for P and AP configuration, the TMR also depends strongly
on the MgO thickness.

6.2 Giant magnetoresistance

The second important magnetoresistive effect in nanostructures, besides the
tunnel magnetoresistance (TMR), is the giant magnetoresistance (GMR).
[122–124] For the TMR, we applied theories for ballistic transport. For GMR,
we are going to apply diffusive transport theories, such as the Boltzmann equa-

185
Figure 6.26. Structures of Fe/MgO interfaces. Left: ideal structure with bcc Fe(001)
(red) and MgO(001) (O blue, Mg green), the latter rotated by 45 ◦ about the [001]
axis. Right: as for the ideal structure but with FeO layer.

tion.
The GMR was independently discovered by Fert and Grünberg [125]. In
magnetic multilayer system, for example a trilayer, the resistance depends on
the magnetic configuration of the magnetic subsystems (Fig. 6.29). For parallel
alignment (P) of the magnetizations, the resistance is low whereas for antipar-
allel alignment (AP) it is large. This finding is also observed in magnetic tunnel
junctions but in GMR devices the spacer is conducting.3
For a trilayer system, one of the magnetic electrodes can be pinned by a
pinning layer (exchange bias) and the magnetic configuration can switched by
an external magnetic field. Such a system is called a spin valve (Fig. 6.30c).
In a multilayer system, the above approach does not work. However, this can
be achieved due to the effect of antiferromagnetic interlayer coupling which is a
particular case of interlayer exchange coupling (IEC). The interlayer exchange
coupling is mediated by the itinerant electrons in the conducting spacer layer
and is an analogue of the Ruderman-Kittel-Kasuya-Yosida (RKKY) interaction
between localized magnetic moments in a nonmagnetic host metal. The inter-
layer exchange coupling oscillates between ferromagnetic and antiferromagnetic
as a function of the thickness of the nonmagnetic layer. By choosing an appro-
priate thickness of the nonmagnetic layer it is, therefore, possible to create an
antiparallel configuration of the ferromagnetic layers and then reorient (align)
the moments by an applied magnetic field.
There two major geometries in which the GMR is investigated (Fig. 6.31).

3
In TMR devices, the spacer is insulating.

186
Figure 6.27. Wavevector-resolved transmissions of symmetric and asymmetric
Fe/MgO/Fe tunnel junctions showing ‘hot spots. The MgO spacer is 4 ML thick.

In current-in-plane (CIP) geometry, the electric current flows parallel to the

layered structure whereas in current-perpendicular-to-plane (CPP) geometry
it flows normal to the layers (as in a TMR device). The commonly used set-up
is the CIP geometry.
Origin of the GMR
GMR can be qualitatively understood using the Mott model, which was intro-
duced as early as 1936 to explain the sudden increase in resistivity of ferro-
magnetic metals as they are heated above the Curie temperature (that is the
magnetic ordering temperature of a ferromagnet).
There are two main points proposed by Mott. First, the electrical conduc-
tivity in ferromagnetic metals can be described in terms of a spin-up and a
spin-down current (‘conducting channels’). In order to establish these two in-
dependent channels, the probability of spin-flip scattering processes must be
small (as compared to the probability of spin-conserving scattering processes)
as is normally the case in metals. Second, in ferromagnetic metals the scatter-

187
Figure 6.28. TMR of a symmetric and a asymmetric Fe/MgO/Fe tunnel junction
versus MgO-spacer thickness. Numbers in brackets give the optimistic TMR.

ing rates of the spin-up and spin-down electrons are quite different, whatever
the nature of the scattering centers is.
According to Mott, the electric current is primarily carried by electrons from
the sp-bands due to their small effective mass and high mobility (cf. the Slon-
czewski model). The d bands play an important role in providing final states for
the scattering of the sp electrons. In ferromagnets, the d bands are exchange-
split, so that the density of states is not the same for the spin-up and spin-down
electrons at the Fermi energy (cf. Fig. 6.16a for Co). The probability of scat-
tering into these states is proportional to their density, so that the scattering
rates are spin-dependent. In summary, spin-up and spin-down electrons are
scattered at the same defects but the scattering rates become spin-dependent
due to the spin-dependent DOS of the d bands.
Following Mott s arguments the CIP-GMR is explained in magnetic multilay-
ers (Fig. 6.32). We assume that the scattering is strong for electrons with spin
antiparallel to the magnetization direction, and is weak for electrons with spin
parallel to the magnetization direction. This is supposed to reflect the asym-
metry in the density of states at the Fermi level, in accordance with Mott s
second argument.
For the parallel-aligned magnetic layers (the top panel in Fig.6.32a), the spin-
up electrons pass through the structure almost without scattering, because their
spin is parallel to the magnetization of the layers. On the contrary, the spin-
down electrons are scattered strongly within both ferromagnetic layers, because
their spin is antiparallel to the magnetization of the layers. Since conduction

188
Figure 6.29. Schematic representation of the GMR effect. (a) Change in the re-
sistance of the magnetic multilayer as a function of applied magnetic field. (b) The
magnetization configurations (indicated by the arrows) of the multilayer (trilayer) at
various magnetic fields: the magnetizations are aligned antiparallel at zero field; the
magnetizations are aligned parallel when the external magnetic field H is larger than
the saturation field HS . (c) The magnetization curve for the entire multilayer system.

occurs in parallel for the two spin channels, the total resistivity of the multilayer
is determined mainly by the highly-conductive spin-up electrons and appears
to be low.
The resistance for P alignment is then
1 1 1 1 1
= + = + , (6.16)
RP R↑ + R↑ R↓ + R↓ 2R↑ 2R↓

that is
2R↑ R↓
RP = . (6.17)
R↑ + R↓

In a multilayer system, the unit cell perpendicular to the layers comprises four
layers, two magnetic (in P configuration) and two nonmagnetic (conducting
spacer). Hence, the resistance of such a four-layer stack is given by resistivities

189
Figure 6.30. Various structures in which GMR can be observed: magnetic multilayer
(a), pseudo spin valve (b), spin valve (c) and granular thin film (d). Note that the layer
thickness is of the order of a few nanometers, whereas the lateral dimensions can vary
from micrometers to centimeters. In the magnetic multilayer (a) the ferromagnetic
layers (FM) are separated by nonmagnetic (NM) spacer layers. Due to antiferromag-
netic interlayer exchange coupling they are aligned antiparallel at zero magnetic field
as is indicated by the dashed and solid arrows. At the saturation field the magnetic
moments are aligned parallel (the solid arrows). In the pseudo spin valve (b) the GMR
structure combines hard and soft magnetic layers. Due to different coercivities, the
switching of the ferromagnetic layers occurs at different magnetic fields providing a
change in the relative orientation of the magnetizations. In the spin valve (c) the top
ferromagnetic layer is pinned by the attached antiferromagnetic (AF) layer. The bot-
tom ferromagnetic layer is free to rotate by the applied magnetic field. In the granular
material (d) magnetic precipitates are embedded in the nonmagnetic metallic material.
In the absence of the field the magnetic moments of the granules are randomly oriented.
The magnetic field aligns the moments in a certain direction.

ρ and the respective thicknesses dFM and dNM ,

R↑ = ρNM dNM + ρ↑ dFM , (6.18a)
R↓ = ρNM dNM + ρ↓ dFM . (6.18b)
The interface resistance is omitted for simplicity. For a stack of N unit cells,
the total resistance is N RP .
For the antiparallel-aligned multilayer (the top panel in Fig. 6.32b), both
the spin-up and spin-down electrons are scattered strongly within one of the
ferromagnetic layers, because within the one of the layers the spin is antiparallel
to the magnetization direction. Therefore, in this case the total resistivity of
the multilayer is high.
The AP resistance one has
1 1 1 2
= + = , (6.19)
RAP R↑ + R↓ R↓ + R↑ R↑ + R↓
that is
R↑ + R↓
RAP = . (6.20)
2

190
Figure 6.31. (a) CIP and CIP geometries for GMR investigations. (b) CPP set-up
for a multilayer system.

Thus, for a stack of N unit cells, the total resistance is N RAP .

The GMR is typically defined as4
N RAP − N RP
δ≡ . (6.21)
N RP
Inserting the above results, we obtain
(R↓ − R↑ )2
δ= , (6.22)
4R↑ R↓
which does not depend on N .
Problem 76. Derive eq. (6.22).
It follows
N RAP − N RP RAP
δ = = −1
N RP RP
R↑ + R↓ R↑ + R↓ (R↑ + R↓ )2 4R↑ R↓
= −1= −
2 2R↑ R↓ 4R↑ R↓ 4R↑ R↓
(R↑ − R↓ )2
=
4R↑ R↓
2
Assuming that the resistance of the spacer is much less than that of the
ferromagnet, we have
R↑ ≈ ρ↑ dFM , (6.23a)
R↓ ≈ ρ↓ dFM . (6.23b)
Hence,
(ρ↓ − ρ↑ )2
δ≈ . (6.24)
4ρ↑ ρ↓
4
Confer the optimistic definition of the TMR.

191
Figure 6.32. Schematic illustration of electron transport in a multilayer for parallel
(a) and antiparallel (b) magnetizations of the successive ferromagnetic layers. The
magnetization directions are indicated by the arrows. The solid lines are individual
electron trajectories within the two spin channels. It is assumed that the mean free
path is much longer than the layer thicknesses and the net electric current flows in
the plane of the layers (CIP). Bottom panels show the resistor network within the
two-current series resistor model. For the parallel-aligned multilayer (a), the spin-up
electrons pass through the structure almost without scattering, whereas the spin-down
electrons are scattered strongly within both ferromagnetic layers. Since conduction
occurs in parallel for the two spin channels, the total resistivity of the multilayer is low.
For the antiparallel-aligned multilayer (b), both the spin-up and spin-down electrons
are scattered strongly within one of the ferromagnetic layers, and the total resistivity
of the multilayer is high.

Defining the spin asymmetry

ρ↓
α≡ , (6.25)
ρ↑

is follows
(α − 1)2
δ≈ . (6.26)
4α
Obviously, a large GMR requires large asymmetries (α  1 or α  1). The
GMR vanishes for α = 1.
The most common GMR devices are Co/Cu and Fe/Cr multilayers. Estimat-
ing the spin-dependent scattering rates by the density of states at the Fermi

192
level, one finds α = 7 (Co) and α = 3 (Fe). The respective GMR values are

(7 − 1)2
δCo/Cu ≈ ≈ 130 %, (6.27a)
4·7
(3 − 1)2
δFe/Cr ≈ ≈ 30 %. (6.27b)
4·3
While the former value is rather close to experimentally observed ones (120 %),
the latter is far too small (experiment: 220 %).
Problem 77. Give possible reasons for the apparent discrepancies between the model and
experimental data.
The scattering rates in the model are solely given by the spin-dependent density of
states at the Fermi level. But not all states serve identically as final states. Also sp-
states contribute and the scattering rates depend on the detailed band structure, leading to
wavevector-dependent scattering rates. A detailed account for the scattering mechanisms
is not provided. Further, interface scattering is not considered. [126] 2
Taking into account the spacer, one obtains

(α − 1)2
δ=   , (6.28)
4 α + p ddNM
FM
1 + p dNM
dFM

where p ≡ ρNM /ρ↑ .

Problem 78. Derive eq. (6.28).
The GMR is given by

(R↓ − R↑ )2
δ= ,
4R↑ R↓

with

R↑ = ρNM dNM + ρ↑ dFM ,

R↓ = ρNM dNM + ρ↓ dFM .

It follows
1 (ρNM dNM + ρ↓ dFM − ρNM dNM + ρ↑ dFM )2
δ =
4 (ρNM dNM + ρ↓ dFM )(ρNM dNM + ρ↑ dFM )
1 (ρ↓ dFM − ρ↑ dFM )2
=
4 (ρNM dNM + ρ↓ dFM )(ρNM dNM + ρ↑ dFM )
1 (ρ↓ − ρ↑ )2
=
4 (ρNM ddNM + ρ↓ )(ρNM ddNM + ρ↑ )
FM FM

1 (α − 1)2 ρ2↑
=
4 (pρ↑ ddNM + αρ↑ )(pρ↑ ddNM + ρ↑ )
FM FM

1 (α − 1)2
= .
4 (p dNM + α)(p ddNM + 1)
d
FM FM

2
Evidently, the GMR increases with decreasing p dNM /dFM . So, a low spacer-
resistance is required for a large GMR.

193
 For large spacer thicknesses, the GMR decreases with dNM as

(α − 1)2 1
δ= ∝ . (6.29)
d2NM

p p2
4 α + (α + 1) dFM dNM + d2FM
d2NM

Although a decrease is observed experimentally, the actual dependence is dif-

ferent. This suggests again that the resistor model does not account for all
important ingredients.
As for the TMR (Fig. 6.16), also an inverse GMR was observed, that is the
resistance for AP alignment is less than that for P alignment. This findings can
be accounted for by a GMR device with two different ferromagnets. It is easy
to show that the GMR is given by

(α(1) − 1)(α(2) − 1)
δ≈ , (6.30)
α(1) (1 + q) + α(2) (1 + q −1 )

where α(1) and α(2) are the spin asymmetries of the two ferromagnets and
(1) (2)
q ≡ ρ↑ /ρ↑ . Obviously it is sufficient for a negative TMR to have either
α(1) < 1 and α(2) > 1 or vice versa.
Problem 79. Derive eq. (6.30).
It follows
1 1 1
= + ,
RP R1↑ + R2↑ R1↓ + R2↓
(R1↑ + R2↑ )(R1↓ + R2↓ )
RP =
R1↑ + R2↑ + R1↓ + R2↓
and
1 1 1
= + ,
RAP R1↑ + R2↓ R1↓ + R2↑
(R1↑ + R2↓ )(R1↓ + R2↑ )
RAP = .
R1↑ + R2↓ + R1↓ + R2↑
The GMR then reads
RAP − RP (R1↑ + R2↓ )(R1↓ + R2↑ ) − (R1↑ + R2↑ )(R1↓ + R2↓ )
δ = =
RP (R1↑ + R2↑ )(R1↓ + R2↓ )
R1↑ R2↑ − R1↑ R2↓ + R1↓ R2↓ − R1↓ R2↑
=
(R1↑ + R2↑ )(R1↓ + R2↓ )
R1↑ (R2↑ − R2↓ ) − R1↓ (R2↑ − R2↓ ) (R1↓ − R1↑ )(R2↓ − R2↑ )
= =
(R1↑ + R2↑ )(R1↓ + R2↓ ) (R1↑ + R2↑ )(R1↓ + R2↓ )
R
1↓ R
R1↑ R2↑ ( R1↑ − 1)( R2↓
2↑
− 1)
=
(R1↑ + R2↑ )(R1↓ + R2↓ )
(α1 − 1)(α2 − 1)
= 1
R1↑ R2↑ (R 1↑ 2↓ + R1↓ R1↑ + R2↓ R1↑ + R2↓ R2↑ )
R
(α1 − 1)(α2 − 1) (α1 − 1)(α2 − 1)
= R1↓ R1↓ R2↓ R2↓
= R1↓ R1↑ R2↓ R2↑
R1↑ + R2↑ + R2↑ + R1↑ R1↑ (1 + R2↑ ) + R2↑ (1 + R1↑ )
(α1 − 1)(α2 − 1)
=
α1 (1 + q) + α2 (1 + q −1 )

194
 2
The above resistor model can also account for the interface resistance by
adding additional resistors. By this means, interface resistances were obtained
from experimental data.
Experimental findings
Before introducing a more sophisticated theory for the GMR, a brief survey
over experimental findings is given.
Giant magnetoresistance was discovered in 1988 by the group of Albert Fert
on Fe/Cr magnetic multilayers1 and the group of Peter Grünberg on Fe/Cr/Fe
trilayers. In both cases the samples were grown using MBE and had [001] orien-
tation of the layers. The Cr spacer layers were about 1 nm thick, so that the Fe
layers were coupled antiferromagnetically providing an antiparallel alignment
(AP) of their magnetizations at zero applied magnetic field. As the applied field
is increased, the magnetic moments of the ferromagnetic layers progressively ro-
tate towards the field, leading to a decrease in the resistance of the multilayer
(trilayer). At saturation the magnetizations end up in a configuration of par-
allel alignment with the lowest value of the resistance. Figure 6.33 shows the
variation in the resistance of the Fe/Cr multilayer measured by Baibich et al.
The highest magnitude of GMR in these experiments was found of 79 % at
T = 4.2 K. The GMR effect was ascribed to the spin-dependent transmission
of the conduction electrons between the Fe and Cr layers.

Figure 6.33. Normalized resistance versus applied magnetic field for several antifer-
romagnetically coupled Fe/Cr multilayers at 4.2 K. Arrows indicate the saturation field
HS , which is required to overcome the antiferromagnetic interlayer coupling between
the Fe layers and align their magnetizations parallel.

195
 In 1990 a significant step towards the industrial application of GMR was made
by Parkin et al. who demonstrated that GMR can be observed in multilayers de-
posited by sputtering rather than the much slower MBE growth process. They
succeeded in obtaining similar GMR values on sputtered polycrystalline Fe/Cr
multilayers4 and later found a sizeable GMR of 120 % on Co/Cu multilayers.
In the magnetic multilayers the successive ferromagnetic layers are exchange-
coupled through a nonmagnetic spacer layer. Parkin et al. found that the sign
of the coupling oscillates between ferromagnetic and antiferromagnetic with
increasing thickness of the spacer layer. The magnitude of GMR is also oscil-
lating from a finite value to zero as the spacer thickness increases, as shown in
Fig. 6.34 for the Fe/Cr multilayer. These oscillations in GMR reflect the oscil-
lations in the interlayer coupling. Sizeable values of GMR are observed when
the coupling is antiferromagnetic, since this provides an antiparallel alignment
of the magnetizations in the successive ferromagnetic layers at zero magnetic
field, as for dCr = 1 nm and dCr = 2.5 nm. No GMR (a much diminished GMR)
is observed when the coupling is ferromagnetic which prevents the change in
the relative alignment of the magnetizations as the applied field is varied, as for
dCr = 1.8 nm. Hence, antiferromagnetic coupling is not a necessary condition
for GMR to occur. All that is necessary is that the magnetic moments of the
layers are not locked by the ferromagnetic coupling, but can be reoriented by
an applied magnetic field.

Figure 6.34. Saturation magnetoresistance at 4.2 K versus Cr thickness for

Si(111)/Cr(10 nm)/[Fe(2 nm)/Cr(t)]N Cr(5 nm) multilayers deposited at various tem-
peratures: triangles and squares −40 ◦ C (N = 30); circles −125 ◦ C (N = 20).

Dependence on the thickness of the nonmagnetic layer. When consider-

ing the dependence of GMR on the nonmagnetic layer thickness in magnetic
multilayers and spin valves one should compare the resistances of the perfectly
parallel (P) and antiparallel (AP) magnetic configurations. The presence of the

196
interlayer exchange coupling leads to oscillations in GMR. This oscillatory con-
tribution to GMR reflects the extent of antiparallel alignment, which is achieved
at zero magnetic field, rather than an intrinsic variation in GMR. Spin valves
are in this sense better for studying the spacer thickness dependence of GMR
than magnetic multilayers. This is due to the pinned ferromagnetic layer, which
keeps the direction of its magnetization and helps to maintain an antiparallel
alignment of the magnetizations in a certain field interval, provided that the
ferromagnetic interlayer coupling is not stronger than the exchange-bias field.
However, at small spacer thicknesses the magnetic layers may become strongly
coupled ferromagnetically due to the presence of pinholes in the nonmagnetic
film, leading to a decreased GMR ratio.
The dependence of GMR on the nonmagnetic layer thickness in spin valves
was studied by Dieny et al. Figure 6.35 shows the variation of GMR as a func-
tion of the thickness of the nonmagnetic layer (NM) in spin valve structures. As
is seen from the figure, the value of GMR decreases monotonically with increas-
ing nonmagnetic layer thickness. This decrease can be qualitatively ascribed to
two factors.
• With increasing spacer thickness the probability of scattering increases as
the conduction electrons traverse the spacer layer, which reduces the flow
of electrons between the ferromagnetic layers and consequently reduces
GMR.

• The increasing thickness of the nonmagnetic layer enhances the shunting

current within the spacer, which also reduces GMR.
These two contributions to GMR can be phenomenologically described by the
expression

∆R e−dNM /lNM
∝ . (6.31)
R 1 + dNM /d0
The exponential factor represents the probability that an electron is not scat-
tered within the NM layer. The factor in the denominator describes the shunting
effect due to the NM layer. The parameter lNM is related to the mean free path
of the conduction electrons in the spacer layer. One expects that lNM will be less
than the mean free path in the spacer layer, λNM , due to the fact that electrons
which most effectively contribute to GMR have out-of-plane velocities. Dieny
et al. proposed that for systems of practical interest lNM is approximately equal
to half of the mean free path. The parameter d0 is an effective thickness, which
depends on the conductance of the system in the absence of the NM layer.
Although the above formula is a phenomenological expression, it contains a
significant part of the physics involved. As we will see, within a Boltzmann
approach to free electrons the exponential is replaced by more complicated
exponential integrals over various incidence angles of the conduction electrons
with respect to the plane of the layers. Nevertheless, the typical variation of
GMR versus nonmagnetic layer thickness remains qualitatively the same.
It was found that the Cu and Au thickness dependence of GMR, can be fitted
well by using decay lengths of lCu = 6 nm and lAu = 5 nm, respectively. These

197
Figure 6.35. Magnetoresistance at room temperature
versus thickness of the noble-metal layer in spin valves
Si/Co(70 nm)/NM(dNM )/Ni80 Fe20 (5 nm)/Fe50 Mn50 (8 nm)/NM(1.5 nm) with NM
= Cu and Au. The solid lines represent fits according to resistor model.

decay lengths are determined by scattering in the spacer, due to phonons, grain
boundaries, and other defects, and are correlated with the mean free path λNM .
The smaller value found for Au is consistent with the higher resistivity of Au,
deduced from measurements on sputtered samples: λAu = 8.5 nm for Au versus
λCu = 11.5 nm for Cu.
As is evident from Fig. 6.35, the GMR values are larger for the Cu spacer layer
than for the Au spacer layer. This fact was ascribed to a lower transmission
through the ferromagnetic/noble-metal interfaces for Au than for Cu, which
reduces the intensity of the flow of electrons that continuously escape from each
ferromagnetic layer across the interfaces. The low GMR values for Au may also
reflect the higher spin-orbit scattering expected of the heavier element, which
leads to spin-flip scattering in the spacer layer. The effect of the microstructure
may also be important: the large lattice mismatch between ferromagnetic layers
and Au may result in misfit dislocations and be an additional cause of the lower
GMR.
GMR in magnetic multilayers versus thickness of the nonmagnetic spacer
layer behaves in a similar fashion as in spin valves. Figure 6.36 displays values of
GMR in Co/Cu multilayers measured at relatively large Cu thicknesses, so that
the interlayer exchange coupling is small.5 Therefore, the GMR results from
the random arrangement of magnetic domains in successive magnetic layers.
Parkin et al. found that at T = 4.2 K the GMR decays approximately as
5
Note that the interlayer exchange coupling decreases with increasing Cu thickness much
faster than the GMR, such that the exchange coupling fields become much weaker than
the saturation fields.

198
1/dCu (a). As was explained above, this behavior is the direct consequence of
the shunting of the electric current due to increasing thickness of the spacer
layers. At room temperature the scattering within the spacer layers diminishes
the flow of electrons from one magnetic layer to neighboring magnetic layers
and therefore reduces the magnitude of the GMR. Such scattering is related to
volume scattering within the interior of the spacer layers due to electron-phonon
interactions.

Figure 6.36. Saturation magnetoresistance versus Cu spacer

layer thickness for several series of multilayers of the form,
Si(111)/Ru(5 nm)/[Co(1.1 nm)/Cu(dCu )]6 /Ru(1.5 nm). Data are shown for tem-
peratures of 4.2 K (a) and 295 K (b). The actual curves shown in the figure have the
form of ∆R/R = 0.28+55.4/(1.3+dCu ) and ∆R/R = 28.9/(0.43+dCu ) exp(−dCu /31.8)
at 4.2 K and 295 K, respectively. dCu is given in nm .

Dependence on the thickness of the magnetic layer. A typical variation of

the magnitude of GMR versus the thickness of the free ferromagnetic layer in
a spin valve versus the thickness of the ferromagnetic free layer is plotted in
Fig. 6.37. As is evident from the figure, the three curves have very similar shapes
characterized by a broad maximum between 6 nm and 10nm. The position of
the maximum depends on the location of the spin-dependent scattering centers.
In the case of interfacial spin-dependent scattering, the maximum is located at
smaller thicknesses than for bulk spin-dependent scattering. The appearance

199
of the maximum is explained by the following arguments. The decrease in
GMR at large magnetic layer thickness is due to the increasing shunting of the
current in the inner part of the ferromagnetic layers. The decrease in GMR at
low thickness is due to the scattering at the outer boundaries (substrate, buffer
layer or capping layer). This scattering significantly affects the GMR when the
thickness of the ferromagnetic layer becomes smaller than the longer of the two
mean-free paths associated with the spin-up and spin-down electrons.

Figure 6.37. Magnetoresistance in FM(dFM )/Cu(2.2 nm)/Ni80 Fe20 (5 nm)/Fe50 Mn50 (8 nm)/Cu(1.5 nm)
spin valve versus thickness of the ferromagnetic free layer FM = Co, Ni80 Fe20 , and Ni
at room temperature. The solid lines represent fits according to the resistor model.

Phenomenologically, the variation of spin-valve MR with the thickness of the

ferromagnetic layers can be fairly well represented by
∆R 1 − e−dFM /lFM
∝ . (6.32)
R 1 + dFM /d0
The numerator describes the variation of the scattering rates of the electrons
with thickness dFM . It characterizes the angle-averaged probability for an elec-
tron with the longest mean free path to be scattered within the ferromagnetic
layer before being scattered diffusively at the outer boundary of the spin valve.
This factor is responsible for the decrease of the GMR at low thicknesses dFM :
if the ferromagnetic layers are too thin the contrast between the spin-dependent
mean free paths decreases due to the stronger diffuse scattering of the electrons
with the longer mean free path at the outer boundaries. lFM is therefore related
to the longest mean free path in the ferromagnetic layer. The denominator de-
scribes the shunting of the current within the ferromagnetic layers, so that d0
is an effective thickness which represents the shunting of the current in the rest
of the structure, i. e. in all layers except the ferromagnetic layer the thickness
of which is varied.

200
 Since the magnitude of GMR is related to the asymmetry in the scattering
rates within the two conduction channels, it was expected that modifying the
spin-dependent scattering by introducing appropriate impurities either at the
interfaces or in the bulk of the ferromagnetic layers would enhance the GMR.
A number of attempts have, therefore, been made to find a correlation between
the magnitude of the scattering asymmetries in bulk magnetic alloys and the
magnitude of GMR in magnetic multilayers.
Parkin demonstrated that inserting a very thin layer of Co results in a dra-
matic increase in GMR. Figure 6.38a shows the room temperature resistance
response to the applied magnetic field in a spin valve and in the same spin
valve with 0.25 nm thick Co layers added at each NiFe/Cu interface. As is seen,
the GMR value increases by a factor of two, demonstrating the strong effect
of the inserted Co layer. The positive effect of the Co layers on GMR in the
permalloy-based spin valves was found to be strongly localized at the interfaces.
By varying the distance of the Co layer from the interface, d, no significant in-
crease in the GMR was found for d > 0.5 nm (Fig. 6.38b). Contrary to inserting
Co layers at the interfaces of permalloy-based spin valves, adding a permalloy
layer at the Co/Cu interfaces reduces the value of GMR (Fig. 6.38d).

Figure 6.38. Effect of a thin layer inserted at the interfaces in spin valves. (a) Resis-
tance versus magnetic field for Si/NiFe(5.3)/Cu(3.2)/NiFe(2.2)/FeMn(9)/Cu(1) spin
valve without (open circles) and with (filled circles) 0.25 nm thick Co layers added at
each NiFe/Cu interface. Dependence of the saturation magnetoresistance on (b) Co in-
terface layer thickness, dCo , in Si/NiFe(5.3-dCo )/Co(dCo )/Cu(3.2)/Co(dCo )/NiFe(2.2-
dCo )/FeMn(9)/Cu(1) spin valves, (c) distance d of a 0.5 nm thick Co layer from the
NiFe/Cu interfaces in Si/NiFe(4.9-d)/Co(5)/NiFe(d)/Cu(3)/NiFe(d)/Co(5)/NiFe(1.8-
d)/FeMn(9)/Cu(1) spin valves, and (d) NiFe interface layer thickness, dNiFe ,
in Si/Co(5.7-dNiFe )/NiFe(dNiFe )/Cu(2.4)/FeNi(dNiFe )/Co(2.9-dNiFe )/FeMn(10)/Cu(1)
spin valves. Note that NiFe stands for permalloy and the layer thicknesses are given in
nm. Experiments are performed at room temperature.

The results of these experiments are explained by the dominant contribu-

tion to GMR from spin-dependent scattering at the interfaces. In general, the
magnetic moments at interfaces (especially rough interfaces) can be very dif-

201
ferent from those in the bulk. The nonmagnetic layers at the interfaces are
detrimental to the GMR. These layers are a source of strong spin-independent
scattering. Misoriented spins also reduce the GMR due to spin mixing and
spin-flip scattering.

Temperature dependence. Many experiments found that the GMR decreases

with increasing temperature. Typically, the magnitude of the GMR is a factor
of two or three smaller at room temperature than at liquid helium temperature.
The major factor, which contributes to the temperature variation of the
GMR, is inelastic scattering by phonons. Although electron-phonon scatter-
ing conserves spin, it enhances the saturation resistivity of the multilayer, ef-
fecting negatively the GMR ratio. In addition, it shortens the mean free path
in the spacer layer, which prevents the flow of electrons between successive
ferromagnetic layers and hence reduces the GMR. Scattering by phonons in
ferromagnetic metals is spin-dependent due to the spin-dependent density of
states at the Fermi energy. If other spin-dependent scattering processes are
important (e. g., spin-dependent scattering at the interfaces due to interface
roughness), which are characterized by a different spin asymmetry, the contri-
bution from electron-phonon scattering necessarily changes the GMR. Another
non-trivial mechanism reducing the GMR is related to interband transitions
that are driven by the applied electric field in the presence of spin-independent
scattering potentials.
Another factor, which might influence the temperature variation of GMR,
is electron-magnon scattering. Contrary to phonons, scattering by magnons is
associated with spin-flip processes which intermix the majority- and minority-
spin current channels. At high temperatures when spin fluctuations become
sizeable the electron-magnon scattering would inevitably suppress the GMR.
However, whether this effect is important at room temperature and below for
GMR in magnetic multilayers based on 3d ferromagnets, which are characterized
by very high Curie temperatures, still remains unclear.
The temperature variation of the GMR in Co/Cu multilayers deposited on
grooved substrates is shown in Fig. 6.39a, so that the geometry in these exper-
iments was similar to CPP GMR. The results were analyzed in terms of spin-
dependent bulk and interface resistances within the two-current series resistor
model. It was found that the spin-dependent interface resistances are weakly
temperature-dependent (a, b), which implies that up to room temperature the
major contribution to the interface resistances comes from elastic scattering.
On the other hand, as can be seen from (d), the bulk resistivities of Co and Cu
increase by more than a factor of two between 4.2 K and 300 K, reflecting the
sizeable contribution from inelastic scattering. However, it was found that the
spin asymmetry of the resistivity in the Co layers, αCo , is almost independent
of temperature (c). The results of these experiments demonstrate, firstly, that
phonons, rather than magnons, influence the spin-dependent resistance of the
Co/Cu multilayer. Scattering by magnons would inevitably reduce the spin
asymmetry of the bulk and interface resistances with increasing temperature
due to spin-flip processes. Secondly, the spin asymmetry is about the same

202
for elastic and inelastic scattering in the Co layers. This is evidence that the
scattering spin asymmetry is mainly determined by intrinsic properties of bulk
Co, namely by its band structure.

Figure 6.39. The temperature dependence of the spin-dependent parameters for

Co/Cu multilayers, as determined from the experiments on grooved substrates using
the two-current series resistor model: (a) the spin-asymmetry parameter, γ, of the
Co/Cu interface, (b) the resistances of the Co/Cu interface, (c) the spin-asymmetry
parameter of the bulk Co, β, (d) the resistivity of Co and Cu layers. γ is defined as
γ = (αCo/Cu − 1)/(αCo/Cu + 1), where αCo/Cu is the ratio of the interface resistances
AR↓Co/Cu and AR↑Co/Cu . β is defined as β = (αCo − 1)/(αCo + 1), where αCo is the ratio
of the bulk Co resistivities ρ↓Co and ρ↑Co .

Angular dependence. We have so far considered GMR that arises from par-
allel and antiparallel magnetizations of the successive ferromagnetic layers. In
this section the variation of the magnetoresistance as a function of the angle
between the magnetizations Θ is discussed. In spin valves which comprise a
free and a pinned magnetic layer a continuous change of the angle Θ can be
obtained by applying a rotating field, which rotates the magnetization of the
free layer, but keeps the direction of the magnetization of the pinned layer fixed.
It was found that there are two components contributing to the magnetoresis-
tance: the anisotropic magnetoresistance (AMR), which varies as the cosine
squared of the angle between the rotating magnetization and the sensing cur-
rent, and the giant magnetoresistance. By subtracting the contribution from
the AMR, it was found that the GMR varies linearly with cos Θ and can be
phenomenologically described by
1 − cos Θ
R(Θ) = RP + (RAP − RP ) , (6.33)
2

203
where RP and RAP are the resistances of the spin valve for the parallel and
antiparallel magnetizations respectively. Such a linear variation of the resistance
with cos Θ was also observed in Fe/Cr multilayers.

Figure 6.40. Normalized resistance and conductance versus the cosine of the relative
angle between the magnetizations of the soft permalloy layers and hard layers composed
of Co clusters for [Co(0.4 nm)/Ag(4 nm)/NiFe(4 nm)/Ag(4 nm)]15 multilayer.

Semiclassical theory
The resistor model, being introduced in the preceding, is too simple to describe
correctly CIP GMR in magnetic multilayers and in spin valves. This model is
based on the assumption that the mean free path (MFP) is large for both spin
channels as compared to layer thicknesses. This approximation is not justified
for real layered systems because the MFP within one of the spin channels is
comparable to or even less than the layer thickness. In addition, the resistor
model is unable to predict the asymptotic behavior of GMR for large layer
thicknesses.
The above suggests to describe the GMR by means of the semiclassical Boltz-
mann theory of transport (Sect. 5.4.1). This theory considers electron transport
using classical dynamics (Lorentz force). Nevertheless, it includes many aspects
of quantum mechanics. For example, within this approach quantum-mechanical
statistics is used, and scattering can be calculated quantum-mechanically using
realistic band structures.
Although discussed in Sect. 5.4.1, we recall the main aspects here, but need
to consider the electronic spin in more detail in order to describe the GMR.
The distribution function φ(r, k, t) is defined as the number of electrons with
given position r and wavevector k at time t. We assume that the two spin
states of the electrons are uncoupled and, therefore, the distribution function
can be considered independently for the spin-up and spin-down channels. The
Boltzmann transport equation is obtained by balancing the change in the distri-
bution function caused by the applied electric field and the scattering processes

204
that act to bring it back towards equilibrium,
 
dφ(r, k, t) ∂φ(r, k, t)
= −ṙ · ∇r φ(r, k, t) − k̇ · ∇k φ(r, k, t) + .
dt ∂t scatter
(6.34)

The first term in this equation describes the electron drift due to their velocity,
the second term reflects the acceleration of the electrons due to the applied
electric field, and the scattering term describes scattering of the electrons by
imperfections, such as defects or impurities. It can be written in terms of the
probability Pkk0 for an electron to scatter between momentum k and k0 ,
 
∂φ(r, k, t) X
= Pkk0 [1 − φ(r, k, t)] φ(r, k0 , t)
∂t scatter 0 (6.35)
k
− Pk0 k 1 − φ(r, k0 , t) φ(r, k, t) .
 

The right-hand term describes ‘scattering-out’ processes, in which an electron

from an occupied state of momentum k scatters into unoccupied states k0 .
The left-hand term describes ‘scattering-in’ processes, in which electrons from
occupied states of momentum k0 scatter into an unoccupied state k.
We are interested in a steady state solution (time-independent), so that
dφ/dt = 0. Thus,
 
∂φ(r, k, t)
ṙ · ∇r φ(r, k) + k̇ · ∇k φ(r, k) = . (6.36)
∂t scatter

Taking into account further the principle of microscopic reversibility, that is

Pk0 k = Pkk0 , the scattering term reads
 
∂φ(r, k, t) X
Pkk0 φ(r, k0 ) − φ(r, k) .
 
= (6.37)
∂t scatter 0
k

Assuming a uniform external electric field E, one obtains with

e
δk = − E dt, (6.38a)
~
δr = v(k) dt, (6.38b)

[eq. (5.112)] that

e X
Pkk0 φ(r, k0 ) − φ(r, k) .
 
v(k) · ∇r φ(r, k) − E · ∇k φ(r, k) =
~ 0
k
(6.39)

Here, v(k) is the group velocity of the electrons (page 152).

Aiming at a linear response theory, that is considering a weak external
electric field (small bias voltage), it is convenient to represent the distribution
function as

φ(r, k) = f (k) + g(r, k), (6.40)

205
where g(r, k) is the deviation of the distribution function φ(r, k) from the
equilibrium Fermi-Dirac distribution
1
f (k) = , (6.41)
1 + e(E(k) EF )/kB T

due to the applied electric field. With

∇r φ(r, k) = ∇r [f (k) + g(r, k)] = ∇r g(r, k) (6.42)

and

∇k φ(r, k) = ∇k [f (k) + g(r, k)] (6.43a)

df (k)
= ∇k E(k) + ∇k g(r, k) (6.43b)
dE(k)
df (k)
= ~v(k) + ∇k g(r, k). (6.43c)
dE(k)
Note that ∇k g(r, k) is a deviation of second order in the electric field E.
Retaining only the lowest order contributions in E, one obtains then
df (k) X
Pkk0 g(r, k0 ) − g(r, k) .
 
v(k) · ∇r g(r, k) − eE · v(k) =
dE(k) 0
k
(6.44)
This is a general representation of the linearized Boltzmann kinetic equation
for the description of the electric current, the density of which is given by
e X
j(r) = − v(k)g(r, k) (6.45)
Ω
k

[cf. eq. (5.131)]. Ω is the volume of the system.

The evaluation of eq. (6.44) is not easy to perform because of the scattering-in
term Pkk0 g(r, k0 ) which links the values of the distribution function at various
momenta. Thus, one neglects this term and obtains
df (k) X
v(k) · ∇r g(r, k) − eE · v(k) =− Pkk0 g(r, k). (6.46)
dE(k) 0
k

In general, neglecting the scattering-in term is not a trivial approximation and

has to be justified.
The Boltzmann equation can be considerably simplified using the relaxation-
time approximation,
df (k) g(r, k)
v(k) · ∇r g(r, k) − eE · v(k) =− , (6.47)
dE(k) τ (k)
where τ (k) is the relaxation time for an electron to scatter out of momentum
state k. It is defined by
1 X
≡ Pkk0 . (6.48)
τ (k) 0
k

206
 For a bulk homogeneous system it is straightforward within the relaxation-
time approximation to derive the expression for the conductivity tensor σµν ,
defined by
X
jµ = σµν Eν , µ, ν = x, y, z (6.49)
ν

(page 145). In this case, ∇r g(r, k) = 0 from which follows

df (k) g(k)
eE · v(k) = . (6.50)
dE(k) τ (k)

Solving for g(k) gives

df (k)
g(k) = eτ (k) E · v(k). (6.51)
dE(k)

Taking the zero-temperature limit,

df (k)
= −δ(E(k) − EF ) (6.52)
dE(k)

[cf. eq. (5.146)],

g(k) = −eτ (k) δ(E(k) − EF ) E · v(k). (6.53)

With eq. (6.45), one has

e X e2 X
j(r) = − v(k)g(r, k) = τ (k)v(k) [v(k) · E] δ(E(k) − EF ).
Ω Ω
k k
(6.54)

Comparing with eq. (6.49) yields

e2 X
σµν = τ (k)vµ (k)vν (k) δ(E(k) − EF ). (6.55)
Ω
k

In the case of films and multilayers (CIP-GMR device) which are assumed to
be homogeneous in the xy-plane (parallel to the layers) but inhomogeneous in
the z-direction (perpendicular to the planes; due to the interfaces and bound-
aries), the distribution function g(z, k) depends on z but is independent of x
and y. In this case the solution of the Boltzmann equation takes the form

df (k) h ∓z i
g ± (z, k) = eτ (k) E · v(k) 1 − A± (k)e τ (k)|vz (k) . (6.56)
dE(k)

Here, the signs ± refer to whether the z-component of the velocity is positive
or negative. The coefficients A± (k) are determined from the boundary condi-
tions at the interfaces and at the outer boundaries in terms of reflection and
transmission probabilities and will be considered below.

207
 Note that the solution of the Boltzmann equation takes the form of eq. (6.56)
only for the CIP geometry which is considered here. In this case, the current and
the external field can be assumed to be uniform within the planes of the multi-
layer system. For the CPP geometry, however, the electric field is position- and
spin-dependent because the magnetic layers are inhomogeneous in the direction
of the electric current and, therefore, eq. (6.56) does not hold.
Up to this point, the band structure E(k) of the system under consideration
was not specified. The above derivations are valid for the multiband electronic
structure (assuming that the band index is included in k). Now we are going to
simplify things more, by considering free electrons. Thus, the electronic struc-
ture of transition metals (Fe, Co, Ni) is significantly simplified by neglecting
contribution from the d bands and their strong hybridization with the sp bands.
We now derive the conductivity for a given spin orientation. The total current
is obtained by integrating the current density,
Z Z
3 e X
J = j(r) dr = − v(k) g(r, k) dr3 (6.57)
Ω Ω
k

At zero temperature we have to integrate eq. (6.45) over the Fermi surface
(due to the δ-function). With the group velocity v = ~k/m,
Z
e~
j ± (r) = − k g ± (r, k) dk 3 (6.58)
Ωm
and with a k-independent relaxation time
e~ h ∓z i
g ± (z, k) = − τ δ(k − kF ) E · k 1 − A± (k)e τ |vz (k)| . (6.59)
m
The multilayer system comprises N layers, each with a mean free path

λi = τi vF , i = 1, . . . , N. (6.60)

Defining the thicknesses di and resistivities

P ρi of the metal layers, the total
thickness of the multilayer is d ≡ i di . The spin-dependent conductance is
then given by two parts. The first part,
X τi n i d i 1 X di
G1 = e2 = , (6.61)
mi d ρi
i i

describes as if the current would flow in parallel through all the layers. Finite-
size effects are described by the second term,
X 3λi Z 1  di

2 +
G2 = − (1 − µ) µAi (µ) 1 − e λi µ dµ
4ρi 0
i
Z 0    (6.62)
di
2 −
+ (1 − µ) µAi (µ) 1 − e λi µ
dµ ,
−1

where µ refers to the cosine of the momentum perpendicular to the interfaces.

208
 The coefficients A±
i can be found using the boundary conditions. If the film is
placed at 0 < z < d, the distribution function at z = 0 must have no electrons
with vz > 0 other than those specularly reflected from the surface because
there are no electrons outside the film. Therefore, defining a fraction p of the
electrons which are specularly reflected, the boundary condition at z = 0 is

g + (0, vz ) = pg − (0, vz ) (6.63)

On the opposite side of the film, at z = d, the electrons with vz < 0 could only
be those which are specularly reflected from the boundary:

g − (d, vz ) = pg + (d, vz ). (6.64)

In the case p = 1, which corresponds to perfect reflection from the bound-

aries, the conductivity of the film is identical to the conductivity of an infinite
homogeneous metal. If p < 1, a fraction of electrons, (1 − p), is scattered dif-
fusively. The case of p = 0 corresponds to perfectly diffuse scattering, when all
the reflected electrons lose memory of their velocity. In the presence of diffuse
scattering the conductivity of a thin film decreases with the film thickness. The
effect is significant if the film thickness is reduced down to become comparable
to the mean free path, d ≈ λ = vF τ . For thicker films, the current density is
close to the bulk value in the center of the film, but lowered in the vicinity of
the boundaries at distances of the order of λ.
In the case of layered structures, apart from the probability p for specular
reflection at the outer boundaries, additional boundary conditions at the in-
terfaces are required. These boundary conditions can be imposed by assuming
that the electrons are coherently transmitted with probability T , coherently
reflected with probability R, or diffusely scattered with probability D at the
interface. The electrons which are diffusely scattered are assumed to be sim-
ply lost so that the transmission and reflection probabilities are related by the
expression D = 1 − T − R. Thus, at the interface between layers i and i + 1,
+ −
gi+1 = T gi+ + Rgi+1 (6.65)

and

gi− = T gi+1
−
+ Rgi+ . (6.66)

All the coefficients can be computed recursively, allowing to obtain the spin-
dependent conductances. From the total conductances for P and AP configu-
ration, follows the GMR.
According to the semiclassical free-electron model the variation of GMR as
a function of the thickness of the magnetic spacer layer is different depend-
ing on whether the bulk or interface spin-dependent scattering is dominant.
In the case of bulk scattering, the GMR ratio exhibits a maximum at a cer-
tain thickness. This can be seen from the calculated results presented by the
solid lines in Fig. 6.41, which display the magnitude of GMR as a function of
the ferromagnetic layer thickness dFM for multilayers with various number of

209
Figure 6.41. Magnetoresistance versus ferromagnetic layer thickness in
(FM/NM)N FM multilayers for bulk (the solid lines) and interface (the dashed line)
scattering as calculated using the semiclassical free-electron model. λNM = 20 nm,
dNM = 20 nm, and diffuse outer boundary scattering are assumed in the calculation.
The other parameters are set as follows: N is varied, λ↑FM = 12 nm, λ↓FM = 6 nm,
T ↑ = T ↓ = 1 for bulk spin-dependent scattering and N = ∞, λ↑FM = λ↓FM = 0.6 nm,
T ↑ = 1, T ↓ = 0.1 for interface spin-dependent scattering.

FM/NM bilayers. In this calculation, bulk spin-dependent scattering is intro-

duced through the spin-dependent mean free paths in the ferromagnetic layers,
and diffuse scattering is assumed at the outer boundaries. As is evident from
this figure, the position of the GMR maximum shifts towards lower FM layer
thicknesses with increasing the number of bilayers N . The presence of the
GMR maximum and its shift with N is a direct consequence of the bulk spin-
dependent scattering and the diffuse scattering at the outer boundaries. In the
case of a small number of bilayers, e.g., for the trilayer structure with N = 1,
the position of the GMR maximum is related to the long mean free path in the
FM layer, so that dFM ≈ λ↑FM . At this FM layer thickness the spin-up elec-
trons are able to contribute to the conduction before being diffusely scattered
at the outer boundaries. On the other hand, in the case of a large number of
the FM/NM bilayers, the position of the GMR maximum is determined by the
shorter of the two mean free paths, so that dFM ≈ λ↓FM . This is because for
large N the scattering at the outer boundaries becomes unimportant so that
the maximum value of GMR will be obtained when the ferromagnetic layer
thickness is sufficient to provide scattering of the spin-down electrons.
In the case of interface spin-dependent scattering, the GMR decreases mono-
tonically as a function of the FM layer thickness. This can be seen from the
dashed line in Fig. 6.41, which is calculated for the multilayer with an infinite

210
number of repetitions N by introducing spin-dependent transmission coefficients
at the FM/NM interfaces. The decrease of GMR reflects the fact that the bulk
scattering in the ferromagnetic layers is assumed to be spin-independent and,
therefore, increasing the FM layer thickness enhances the relative contribution
of this type of scattering. When dFM becomes much longer that the mean free
path in the FM layer, the GMR is inversely proportional to dFM , ≈ λ↑FM /dFM .
This dependence can be explained by the argument that only those electrons
which leave the FM region of thickness λFM adjacent to the interface have
a sufficiently high probability not to be scattered within this FM layer and,
therefore, reach the opposite interface. The rest of the FM layer is inactive and
serves only as a shunt, which reduces the GMR inversely proportional to dFM .
The semiclassical model predicts that with increasing number of FM/NM
bilayers within a multilayer the value of GMR increases until it reaches satura-
tion, which is similar to that found experimentally. This tendency can also be
seen in Fig. 6.41, according to which at a fixed value of the FM layer thickness
the increment of the GMR growth decreases for larger N . This behaviour of
GMR versus N is due to the diffuse scattering at the outer boundaries of the
multilayer. If the longest mean free path is larger than the total thickness of
the multilayer, then the diffuse outer-boundary scattering reduces the conduc-
tivity of the ‘good’ spin channel and hence effects the GMR negatively. The
magnetoresistance ratio becomes independent of the number of bilayers when
the total thickness of the multilayer is much larger than the longest mean free
path.
Increasing the specular scattering at the outer boundaries strongly enhances
GMR in FM/NM/FM trilayers, provided that the FM layers are not too thick.
This effect is evident from Fig. 6.42, which shows the calculated magnetoresis-
tance as a function of the top FM layer thickness in the spin valve with varied
probability of specular scattering p at the top outer boundary of the trilayer.
Bulk spin-dependent scattering in the ferromagnets and specular reflection at
the bottom outer boundary simulating an antiferromagnetic NiO pinning layer
are assumed in this calculation. The enhancement of the GMR with increasing
specularity p is due to the stronger specular scattering from the top surface,
which unlike diffuse scattering reflects electrons back allowing them to cross the
spin valve many times, thereby increasing spin-filtering effects. Note that the
optimum thickness of the FM layer at which the maximum GMR is observed
also depends on p and decreases with top surface specularity p, which is similar
to what was found for the multilayers with an increasing number of FM/NM
bilayers. This is not surprising because a multilayer with an infinite number of
periods can be simulated by considering a bilayer with specular scattering at
the outer boundaries in which the FM layer thickness is taken as half the actual
one. At larger FM layer thicknesses the GMR ratio for the specular top surface
becomes lower than the GMR ratio for the diffuse-scattering top surface. As is
seen from the insert in Fig. 6.42, the crossover occurs at dFM ≈ 10 nm. This
is a consequence of the current shunting which becomes dominant over GMR
enhancement. Indeed, when the FM layer thickness is larger than the longest
mean free path within this layer the specular-reflected electrons are not able to
reach the spacer layer and instead contribute to the shunting current.

211
Figure 6.42. The effect of specular scattering at the top outer boundary of a
FM/NM/FM trilayer as calculated from the semiclassical free-electron model. The
magnitude of the GMR is plotted as a function of the top FM layer thickness for various
probabilities of specular reflection p at the top outer boundary. Bulk spin dependent
scattering, T ↑ = T ↓ = 1 and λ↑FM /λ↓FM = 10, and specular reflection at the bottom
outer boundary, pbottom = 1, are assumed in the calculation. The other parameters of
the model are set as follows: dFM−bottom = 2 nm, dNM = 2 nm, ρFM = 15 µΩcm, and
ρNM = 6 µΩcm.

The semiclassical free-electron model predicts a linear variation in the conduc-

tance as a function of cos Θ, where Θ is the angle between the magnetizations
of the two ferromagnetic layers in a spin valve. When the magnetic moments of
the FM layers are not aligned, the momentum transfer between the spin-up and
spin-down conduction channels is determined by the transmission coefficients
which are given by

T↑↑ = T↓↓ = cos2 (Θ/2), (6.67a)

2
T↑↓ = T↓↑ = sin (Θ/2), (6.67b)

Using these expressions and assuming diffuse scattering at the outer boundaries
and no specular reflection at the interfaces, it can be shown that the conduc-
tance of the trilayer varies in a linear fashion with cos Θ,
1 − cos Θ
G = GAP + (GP − GAP ) . (6.68)
2
We see that the semiclassical free-electron model predicts correctly a num-
ber of important features of GMR which are observed experimentally. For

212
example, it qualitatively explains the variation of GMR versus ferromagnetic
and nonmagnetic layer thickness, the effect of specular/diffuse scattering at
the outer boundaries, the enhancement of GMR with the increasing number of
repetitions within a multilayer, and the angular variation of the conductance
in spin valves. The great advantage of this model is the ease of application to
a particular layered system, which allows understanding qualitative trends in
the transport properties. However, as was mentioned above the semiclassical
free-electron model ignores the realistic band structure of the multilayer and,
therefore, can not be applied for a quantitative description of GMR. Although
much experimental data can be fitted well using the semiclassical free-electron
model, the parameters, which are extracted from the fitting, should be treated
with caution.
A qualitative failure of the semiclassical free-electron model to describe con-
sistently in-situ conductance experiments in NiO/Co/Cu/Co spin valves was
demonstrated (Fig. 6.43). As is evident from the figure, addition of about 1
monolayer of Co to a NiO/Co/Cu surface causes the net film conductance to
decrease. The reverse case of Cu on NiO/Co shows a strong positive curva-
ture of the conductance, indicating a reduction of the conductivity in Cu near
the interface with Co. Detailed microstructural characterization using in-situ
Auger electron spectroscopy and ex-situ X-ray diffraction measurements indi-
cated that the defect concentration does not vary noticeably as a function of
thickness. These microstructural measurements suggest that the bulk scattering
parameters ρ and λ should be considered to be constant within each layer, and
that the surface scattering parameter p does not change between the layers.
Under these constraints, it appears to be impossible to fit even qualitatively
the highly asymmetric scattering behavior measured during the formation of
Co/Cu versus Cu/Co interfaces. As can be seen from Fig. 6.43b, depending on
the choice of the interfacial transmissivity parameter T the thickness-dependent
conductance either do not display any conductance step at the interfaces or
display a step at both Co/Cu and Cu/Co interfaces, neither being observed
experimentally. As we will see in section 17, calculations incorporating a realis-
tic band structure resolve the observed inconsistency between the free-electron
model and the experiments.
Some of the band structure effects can be captured within an extended free-
electron model. One can introduce layer- and spin-dependent effective masses
and the relevant band fillings. Unfortunately, the number of free parameters in
such a phenomenological model is so large that the analysis of the experimental
data in terms of this model becomes uncontrolled.
Quantum mechanical theory — Kubo formalism
In addition to the lack of an accurate description of the electronic structure, the
semiclassical free-electron model suffers from the inability to describe quantum
effects, which become important at small film thicknesses. The confinement
of electrons in a thin film leads to a discretization of the energy levels. The
corresponding quantum effects become observable when the typical spacing δE
between the energy levels near the Fermi energy becomes larger than the level
broadening ~/τ arising from various scattering mechanisms. Since δE ∝ ~vF /d,

213
Figure 6.43. Experimental (a) and calculated (b) thickness-dependent conductance
of NiO/Co(2 nm)/ Cu(dCu )/Co(4 nm) spin valves. (a): Conductance is measured in-
situ during the deposition of the spin valves with various thickness of the Cu layers:
dCu = 5.8 nm (A), 1.1 nm (B), 1.6 nm (C), 2.3 nm (D). The position of the interface
with the bottom Co layer is the same for all the samples and is marked by the vertical
line. Note the strong deviations from linearity in the vicinity of the interfaces: a drop in
film conductance for Co on Cu and positive curvature for Cu on Co. (b): Calculations
are performed using a free-electron semiclassical model using parameters which provide
a best fit of the experimental data. Interface scattering (T < 1) must be introduced
to produce the conductance drop observed during deposition of Co on Cu, producing
a complementary drop for Cu on Co which is not observed.

where d is the film thickness, the condition for observing quantum-size effects
is d < λ, where λ is the mean free path.
The failure of the semiclassical theory becomes apparent if one considers
the conductivity of a thin film with diffusively reflecting surfaces. In the limit
when the mean free path becomes much longer than the film thickness the
conductivity tends to infinity, implying that in the absence of bulk scattering
the scattering by the rough surfaces induces no dissipation of electrical current.
This unphysical result is a direct consequence of ignoring quantum size effects
within the semiclassical theory.
In order to resolve this deficiency of the semiclassical model a quantum-
mechanical approach to electronic transport is required. There are several
different quantum-mechanical formulations of transport theory which include
those of Kubo, Landauer, and Keldysh. The Kubo (linear response) formalism
considers the electronic transport in a disordered metallic system as a linear
response to an applied electric field. The Landauer formalism describes the
conductance from the point of view of the transmission of electrons through
a conductor and is applicable to mesoscopic transport. The Keldysh (non-
equilibrium Green function) formalism is conceptually more complicated than
the Kubo and Landauer formalisms, but is more general because it provides
a description of the quantum transport in the presence of dissipative interac-
tions. All these theories can be used for calculations of GMR within a quantum-
mechanical approach. In the present review we outline basic principles of the
Kubo theory, which is the most widely used for the treatment of the GMR.
The starting point of the Kubo formalism is the density matrix. The den-
sity matrix is the quantum-mechanical operator, which describes the statistical

214
properties of a quantum-mechanical system. It is the analogue of the distribu-
tion function within the semiclassical theory (Boltzmann).
The density matrix is defined by
X
ρ≡ |αipα hα|, (6.69)
α

where pα is the statistical weight of state |αi.6 The P

states |αi are normalized but
not necessarily orthogonal. Evidently, pα ≥ 0 and α pα = 1. The expectation
value of an operator  can be obtained by

hÂi = Tr ρÂ. (6.70)

In fact, ρ is a hermitian operator but loosely called ‘matrix’.

Problem 80. Compute Tr ρ.
It follows
X XZ
Tr ρ = Tr |αipα hα| = hr|αipα hα|ri dr3
α α
XZ X X
= pα hα|rihr|αi dr3 = pα hα|αi = pα = 1.
α α α

2
Problem 81. Show that ρ is positive definite.
An operator  is positive definite if for any |ui

hu|Â|ui ≥ 0.

It follows
X X
hu|ρ|ui = hu|αipα hα|ui = |hu|αi|2 pα ≥ 0,
α α

as a sum of positive numbers. 2

Problem 82. Show that the density matrix does not depend on the choice of the basis in
eq. (6.69).
Consider two basis sets, {|αi} and {|βi}. These are related by the matrices S and T ,
X
|αi = Sαβ |βi,
β
X
|βi = Tβα |αi.
α

From this follows

X X XX X
|αi = Sαβ Tβα0 |α0 i = Sαβ Tβα0 |α0 i = δαα0 |α0 i,
β α0 α0 β α0

which establishes ST = 1 in matrix notation. Thus S = T −1 , or explicitly

X
|βi = (S −1 )βα |αi.
α

6 P
If for example there Nα particles in state |αi, the weight would be pα = Nα / β Nβ .

215
 Because both basis sets must be complete, that is
X X
1= |αihα| = |βihβ|,
α β

one arrives at
 †
X XX X
|αihα| = Sαβ |βi  Sαβ 0 |β 0 i
α α β β0
!
XX X X
= |βiSαβ (Sαβ 0 )? hβ 0 | = |βi Sαβ (Sαβ 0 )? hβ 0 |
α β,β 0 β,β 0 α
X X
= |βiδβ,β 0 hβ 0 | = |βihβ|,
β,β 0 β

which requires (Sαβ 0 )? = (S −1 )β 0 α . Or in matrix notation, S † = S −1 . In summary, S is

unitary.
The statistical weights for the basis {β} are
X
pβ = hβ|ρ|βi = pα |hα|βi|2 .
α

Obviously, pβ ≥ 0 and
X X X X
pβ = pα hα|βihβ|αi = pα hα|αi = pα = 1.
β αβ α α

Thence,
X X X
ρ = |βipβ hβ| = |βihβ|αipα hα|βihβ| = |αipα hα|.
β αβ αβ

2
The time-dependent density matrix ρt satisfies the quantum-mechanical equa-
tion of motion
dρt h i
i~ = Ĥt , ρt , (6.71)
dt

where Ĥt is the time-dependent Hamiltonian of the system. [19] This equation
describes the evolution of the system affected by a time-dependent perturbation
U (t), due to the applied electric field. We assume that the electric field takes
the form E exp(t), so that it is uniform in space, is applied at t = −∞, and
grows adiabatically to its value E at t = 0. The latter is taken into account by
an infinitesimal positive , so that the limit  → 0 should be taken in the final
result. The single-electron Hamiltonian of the system can then be represented
by

Ĥt = Ĥ + U (t) = Ĥ + eE · ret , (6.72)

where Ĥ is the time-independent Hamiltonian of the unperturbed system.

Equation (6.71) is the quantum-mechanical analogue of the semiclassical
equation. It describes the time evolution of the system. Initially, at t = −∞,

216
the system is at equilibrium and is characterized by the unperturbed density
matrix ρ, according to the Fermi-Dirac distribution
1
ρ= . (6.73)
1 + e[Ĥ−EF ]/kB T
Due to the applied electric field, the system adiabatically follows the perturba-
tion.
Within the Kubo formalism we are looking for the solution of the eq. (6.71) to
first order with respect to the applied electric field. We can, therefore, represent
the density matrix as

ρt = ρ + δρ(t), (6.74)

where δρ(t) is a small time-dependent deviation from the equilibrium Fermi-

Dirac distribution due to the applied electric field. The linearized equation for
the density matrix then takes the form

dδρ(t) h i
i~ = Ĥ, δρt + [U (t), ρ] . (6.75)
dt
Now we rewrite this operator equation in terms of matrix elements by introduc-
ing a basis of eigenstates α of the unperturbed Hamiltonian Ĥ. The equilibrium
density matrix ρ has the same eigenstates as ρ so that
1
ραβ = δαβ = δαβ f (Eα ). (6.76)
1+ e[Eα −EF ]/kB T
Here,
1
f (Eα ) ≡ (6.77)
1 + e[Eα −EF ]/kB T

is the Fermi-Dirac distribution function and Eα is an eigenvalue of Ĥ.

Problem 83. Derive eq. (6.76).
The equilibrium density matrix is given by
1
ρ= .
1+ e[Ĥ−EF ]/kB T

For any eigenstate |αi of Ĥ with energy Eα , Ĥ|αi = Eα |αi, one has

g(Ĥ)|αi = g(Eα )|αi,

where g is an analytic function.7 Thus,

hβ|g(Ĥ)|αi = g(Eα )hβ|αi = g(Eα )δβα .

Therefore,

ραβ = hα|ρ|βi = hα|f (Ĥ)|βi = f (Eα )δαβ .

7
An analytic function can be expanded into a power series. A proof is then based on the
fact that Ĥ n |αi = Eα Ĥ n−1 |αi = Eαn |αi.

217
 2
The operator r which enters the term [U (t), ρ] through eq. (6.75) is non-
diagonal in the α representation. It is convenient to represent the matrix ele-
ments of r in terms of the matrix elements of the velocity operator v using the
relation

i~v = [r, Ĥ], (6.78)

where we can use Ĥ instead of Ĥt by neglecting high order terms. Taking this
into account we find
f (Eα ) − f (Eβ ) t
[U (t), ρ]αβ = i~eE · v αβ e (6.79)
Eα − Eβ

Problem 84. Derive eq. (6.78).

The time dependence of expectation value of an operator  is given by

∂Ψ?
Z Z Z
dhÂi ? ∂Ψ ∂ Â
= ÂΨ dτ + Ψ Â dτ + Ψ? Ψ dτ.
dt ∂t ∂t ∂t
The last term consider a possible explicit time dependence of the operator. With the
time-dependent Schrödinger equation
∂Ψ
i~ = ĤΨ,
∂t
it follows
Z Z Z
dhÂi ∂ Â
i~ = − Ψ? Ĥ ÂΨ dτ + Ψ? ÂĤΨ dτ + i~ Ψ? Ψ dτ
dt ∂t
Z Z
∂ Â
= Ψ? (ÂĤ − Ĥ Â)Ψ dτ + i~ Ψ? Ψ dτ
∂t
Z Z
∂ Â
= Ψ? [Â, Ĥ]Ψ dτ + i~ Ψ? Ψ dτ
∂t
Or
dhÂi ∂ Â
i~ = h[Â, Ĥ]i + i~h i.
dt ∂t
Choosing r as operator, this establishes
Z
dhri
i~hvi = i~ = Ψ? [r, Ĥ]Ψ dτ,
dt
since r does not depend explicitly on time. In general one arrives at

i~v = [r, Ĥ].

2
Problem 85. Derive eq. (6.79).
The perturbation is given by

U (t) = eE · r et .

218
It follows
[U (t), ρ]αβ = hα|U (t)ρ − ρU (t)|βi = hα|U (t)ρ|βi − hα|ρU (t)|βi
= ehα|E · rf (Ĥ)|βi et − ehα|f (Ĥ)E · r|βi et
 
= eE · hα|rf (Ĥ)|βi − hα|f (Ĥ)r|βi et
= eE · [hα|r|βif (Eβ ) − f (Eα )hα|r|βi] et
= eE · hα|r|βi [f (Eβ ) − f (Eα )] et .
Further, with
hα|[Ĥ, r]|βi = hα|Ĥr|βi − hα|r Ĥ|βi = Eα hα|r|βi − hα|r|βiEβ
= (Eα − Eβ )hα|r|βi

and −i~v = [Ĥ, r] follows

(Eα − Eβ )hα|r|βi = −i~hα|v|βi = −i~v αβ .
Thus,
v αβ
hα|r|βi = −i~ .
Eα − Eβ
In summary, one obtains
[U (t), ρ]αβ = eE · hα|r|βi [f (Eβ ) − f (Eα )] et
f (Eα ) − f (Eβ ) t
= i~eE · v αβ e .
Eα − Eβ
2
The time dependence of δρ(t) is determined by the perturbation, so that

δρ(t) = δρ et . (6.80)

Thus, the solution for the density matrix at t = 0 reads

i~eE · v αβ f (Eα ) − f (Eβ )
δραβ = . (6.81)
Eβ − Eα − i Eα − Eβ
This solution can now be used for calculating the electrical current. For a
spatially homogeneous system the current-density operator is defined by
e
j = − v, (6.82)
Ω
where Ω is the total volume of the system. We need to calculate the expectation
value of j which is determined by8

hji = Tr (jδρ), (6.83)

where we have used the fact that Tr (jρ) = 0, as in equilibrium there is no net
current in the system. Using the definition of the conductivity and taking the
limit  → 0 we obtain
e2 π~ X µ ν
σ µν = vαβ vβα δ(Eα − Eβ )(−f 0 (Eα )), (6.84)
Ω
αβ
8
Note the similarity with the expectation value of the spin polarization, page 113.

219
where we have taken into account that
i
Re lim = π δ(x) (6.85)
→0 x + i

[Dirac identity, eq. (2.35)] and used the fact that as Eα → Eβ , which is required
by the δ-function,

f (Eα ) − f (Eβ )
→ f 0 (Eα ). (6.86)
Eα − Eβ

The last limit is the usual derivative. Using the identity

Z +∞
δ(Eα − Eβ ) = δ(E − Eα ) δ(E − Eβ ) dE (6.87)
−∞

and taking the zero-temperature limit at which −f 0 (Eα ) = δ(E − EF ) [cf.

Heaviside step function, eq. (5.141)], the conductivity can neatly be represented
by

e2 π~ h µ i
σ µν = Tr v δ(EF − Ĥ)v ν δ(EF − Ĥ) . (6.88)
Ω
The above formula for the conductivity would generally depend on the par-
ticular type of the disorder responsible for the scattering. In order to obtain a
result that is independent of the particular disorder configuration but depends
only on average characteristics (e. g., defect density or impurity concentration),
one has to perform a configurational average of this expression. In this form
the above expression is known as the Kubo-Greenwood formula.
In many cases it is convenient to carry out the configurational averaging using
techniques that were developed for Green functions. Therefore it is useful to
rewrite the Kubo-Greenwood formula in the following form

e2 ~
σ µν = hTr [v µ Im G(EF )v ν Im G(EF )]i . (6.89)
πΩ
Here, the angular brackets stand for the configurational average and we have
introduced the Green function, being defined by
1
G(E) = lim . (6.90)
→0 E − Ĥ + i

Strictly speaking, this is the sidelimit of the resolvent of Ĥ. Further, we have
taken into account that δ(E − Ĥ) = Im G(E)/π (Dirac identity).
There are several general techniques for configurational averaging in disor-
dered homogeneous systems. Here, we briefly summarize the main results.
We assume that the total Hamiltonian of the system can be represented by
Ĥ = Ĥ0 + V , where Ĥ0 describes the undisturbed periodic system and V is
scattering potential due to defects or impurities. By configurational averag-
ing, this system which is characterized by the random nonperiodic potential is

220
replaced by an effective medium which possesses translational invariance (mean-
field theory). The configurational averaging leads to a renormalization of the
Green function so that
1
hG(E)i = . (6.91)
E − Ĥ0 − Σ(E)

where Σ is the self-energy. The above equation can be considered as the def-
inition of Σ, which is an energy-dependent non-Hermitian operator. Its real
part shifts the energy levels of the undisturbed system, whereas its imaginary
part characterizes the broadening of the levels due to the finite scattering life-
time. Im Σ(EF ) determines, therefore, the relaxation time τ , which has been
introduced within the semiclassical theory.
It follows from equation (6.89) that the conductivity tensor requires an aver-
age over the product of two Greens functions,

σ ∝ hG(E)G(E)i. (6.92)

In general, performing this averaging explicitly is a complicated problem. This

is the reason why very often the conductivity is approximated by the product
of the average of the Green functions,

σ ∝ hG(E)i hG(E)i. (6.93)

By making this approximation one ignores the contribution from the vertex cor-
rections in the linear response formalism. This approximation is equivalent to
neglecting the scattering-in term in the semiclassical theory, which allows the in-
troduction of the relaxation-time approximation. Similar to the relaxation-time
approximation, the neglect of the vertex corrections in quantum-mechanical
linear-response theory is a non-trivial approximation and has to be justified.
We note that the above derivation of the formula for conductivity is valid
for a homogeneous system. In this case the current and applied field can be
assumed to be uniform. In a general case of an inhomogeneous system the
current density is determined by the non-local conductivity according to
XZ
µ
j (r) = σ µν (r, r 0 )E ν (r 0 ) dr3 . (6.94)
ν

The electric field E(r) in this equation is the local electrostatic field, which
arises from the application of the potential difference across the sample. Note
that this is an internal field, which is not the same as the field applied externally.
For an inhomogeneous magnetic systems the local internal field is position- and
spin-dependent. If the rate at which the electrons are scattered varies from one
region to another, then electrical conduction will lead to a spatial redistribution
of charge. This charge redistribution in an inhomogeneous media results in a
nonuniform internal electric field.
In magnetic systems electron conduction is spin-dependent. Spin-polarized
electric currents in inhomogeneous media lead to a spatial redistribution of
spin as well as charge. This phenomenon is known as spin accumulation or

221
current-driven magnetization. The internal electric field may, therefore, be
different for different spins. In magnetic multilayers the effect of a position-
and spin-dependent internal field is very important for perpendicular transport
because these systems are inhomogeneous in the direction of electric current.
For parallel transport, however, the internal electric field is a constant, because
these layered systems are homogeneous in the plane of the layers. The Kubo-
Greenwood formula can, therefore, be used for the treatment of CIP GMR.
Using the Kubo formalism the conductivity of free electrons scattered by a
spin-dependent potential
X
V (R) = (va + ja ma · σa ) (r − r a ) (6.95)
a

which is produced by random-point scatterers. Here, r a is a random position of

the scatterer, m is a unit vector in the direction of magnetization, σ is vector of
the Pauli matrices, va and ja characterize the strength of the spin-independent
and spin-dependent contribution to the scattering potential respectively, which
may be different for different layers and interfaces. Performing the configu-
rational averaging in momentum space, one derives a simple formula for the
local conductivity, which can be expressed in terms of the local spin-dependent
scattering rate ∆((z),

ne2
σ(z) = (6.96)
n∆(z)

where z is a coordinate perpendicular to the layers. The local scattering rate

∆(z) is determined by the appropriate average of the z-dependent scattering
rates within the different layers. The total CIP conductivity of the multilayer
can be found by integrating over the multilayer thickness L,
Z
1
σ= σ(z) dz. (6.97)
L
In the limit of the mean free path being long compared to the layer thickness,
λ  d, this model reduces to the series-resistor model.
The above quantum model for GMR assumes that the origin of GMR lies
in spin-dependent scattering potentials at the interfaces or in the bulk metals.
By using the free-electron approximation the variation of the intrinsic spin-
dependent electronic potential of the layered structures is ignored.
First-principles models
First-principle models (DFT+LSDA) for the electronic structure in conjunction
with the quantum-mechanical formulation of the electronic transport become
a powerful tool to study the conductivity and GMR in magnetic layered sys-
tems. On one hand, the major advantage of the first-principle approach is the
possibility to treat the electronic structure in a self-consistent way taking into
account the charge and spin transfer at the interfaces. On the other hand, the
quantum-mechanical formulation of the transport theory is necessary for the
realistic modeling of GMR. [127–129]

222
 The principal challenge to the ab-initio simulations lies in a realistic modeling
of the scattering. Assuming that interdiffusion at the interfaces is the only
mechanism of scattering in magnetic multilayers results in highly overestimated
values of GMR. A reliable approach must include other contributions to the
resistivity, such as from intrinsic structural defects.
Unfortunately, proper first-principle treatment of all the possible mechanisms
of scattering in layered structures is very complicated. In a simplified approach,
one uses the Kubo-Greenwood formula for the conductivity and averages the
two Greens functions independently, thereby neglecting the vertex corrections.
Further, one assumes that the effect of this averaging is that each atomic poten-
tial acquires an imaginary term that describes the scattering rate in its vicinity.
µν
We introduce a non-local site-dependent conductivity σij which determines
the electrical current at site (layer) i, j i , related to the local electric field at
site (layer) j, E j ,

jiµ =
X µν
σij Ejν , (6.98)
j,ν

where it is assumed that j i is the average of the current density over the atomic
cell at that site and the local field E j is a constant over each atomic cell. If the
electric field is applied parallel to the layers (CIP geometry). The local fields
are uniform by symmetry and equal to the average applied field. This allows
obtaining the overall conductivity by summating over all layers according to
1X µν
σ µν = di σij , (6.99)
d
ij

where di is the thickness of layer i and d is the total film thickness.

The non-local conductivity can be found according to the Kubo-Greenwood
formula
e2 ~
Z Z
µν
σij = v µ Im hG(r, r 0 )iv ν Im hG(r 0 , r)i dr3 dr03 , (6.100)
πΩi Ωi Ωj

where the integration is performed over the atomic cells at sites i and j, Ωi
and Ωj being the volumes of the corresponding cells. The angular brackets
denote the configurational average, which is introduced as a phenomenological
imaginary part of the self-energy in the Greens function G(r, r 0 ).
The non-local layer-dependent conductivity is useful for a better understand-
ing of the mechanism of magnetoresistance. An example of the calculation of
the non-local conductivity for a Co/Cu/Co trilayer is shown in Fig. 6.44. In
this calculation the phenomenological lifetimes are constrained to provide typi-
cal experimental resistivities for sputtered Cu and Co films. It is assumed that
the scattering rate for the minority carriers in the bulk Co layers is 7 times that
for the majority carries, reflecting the difference in the density of states in bulk
Co. Due to intermixing at the interfaces the scattering rate of the minority
spins at the interfacial layer is set 12 times that for the minority spins, the
factor of 12 being based on coherent potential approximation calculations of

223
the resistivity due to Cu impurities in Co and Co impurities in Cu. In addition,
strong spin-independent scattering at the outer boundaries of the trilayer is
included.

Figure 6.44. Non-local layer-dependent conductivities for a Co7 /Cu10 /Co7 (111) tri-
layer: (a) majority channel for the parallel (P) alignment of magnetizations, (b) minor-
ity channel for the P alignment, (c) spin channel, which is locally majority in the left
side Co layer, for the antiparallel (AP) alignment, and (d) giant magnetoconductance,
which is defined as the difference between the non-local conductivities for the P and
AP alignments. The atomic layers are labeled by indices i and j.

As is obvious from (a), the conductivity is strongly non-local, which is re-

flected in large values of σij for i 6= j especially within the Cu layers due to
a long mean free path in Cu. A comparison of (a) with (b) shows that the
conductivities are ‘more local’ for the minority carriers than for the majority
carriers in Co because of a higher DOS and consequently a shorter mean free
path for the former. As is seen from (a–c), for both parallel and antiparallel
alignments the conductivities have a tendency to be confined within the Cu
or Co layers displaying a decrease at the interfaces. This ‘difficulty’ in passing
between layers results from the significant probability for reflection at the inter-
faces as a consequence of the potential steps created by the potential mismatch
between Co and Cu.
As is evident from (d), a large contribution to GMR comes from the Cu spacer
layer, which is a consequence of electron channeling. Another contribution to
GMR comes from the Co layers. This contribution arises from electrons that are
accelerated in one Co layer, travel through the Cu layer to the other Co layer,
and contribute to the current there. These show up in (d) as the contributions
in the near and far corners, which correspond to i and j being in the Co layers
on opposite sides of the Cu spacer layer.
Figure 6.45 shows the results of the calculations for transmission across
Cu/Co(001) interface. The majority Fermi surface in Co is similar to the Fermi
surface in Cu, but is smaller. This similarity leads to the almost complete
transmission of the majority-spin electrons from Co into Cu (the left bottom
panel). The smaller size of the majority Co Fermi surface leads to the complete
reflection for the states in Cu with the largest momenta parallel to the inter-

224
face (the right top panel). These electrons contribute to channeling within the
Cu layer in the Co/Cu/Co trilayer when magnetizations of the Co layers are
parallel. Although the transmission of the minority-spin electrons is not simply
characterized (due to the complicated nature of the minority Fermi surfaces in
Co), it is seen from the right top panel that the same Cu states transmit much
better into the minority Co states. Channeling by these electrons can give,
therefore, a contribution to GMR if the scattering rate in Cu is much smaller
than it is in Co.

Figure 6.45. Spin-dependent transmission probabilities for the Co/Cu(001) inter-

face. The transmission probabilities are shown for various points on the Fermi surface
projected onto the interface Brillouin zone. For minority Co, each sheet of the Fermi
surface is shown in a fraction of the Brillouin zone.

All the models considered above assume electrical transport in parallel within
the two spin conduction channels, the total conductivity being the sum of the
spin conductivities. Strictly speaking this is not the case due to the spin-orbit
interaction. The spin-orbit interaction is a relativistic effect, which couples
the electronic and the spin degrees of freedom and therefore makes the two-
current model inapplicable. Although the spin-orbit interaction is relatively
small in the transition 3d metals and does not effect strongly the electronic
structure of these metals, it can affect transport properties of bulk alloys, for
example, avoiding the short circuits that can arise in either spin channel. It
is questionable, however, whether this effect is important for the description of
GMR in 3d-metal multilayers, because the short circuit effects, which appear

225
if the intermixing at the interfaces is the only mechanism of scattering, are
avoided due to structural disorder and phonon scattering in real systems. The
spin-orbit interaction might be important when heavy elements like gold or
platinum are used as a spacer layer or introduced as impurities.
Summary and conclusions
Giant magnetoresistance is the large change in electrical resistance of metallic
layered systems when the magnetizations of the ferromagnetic layers are reori-
ented relative to one another under the application of an external magnetic field.
This reorientation of the magnetic moments alters both the electronic structure
and the scattering of the conduction electrons in these systems, which causes
the change in the resistance. Various types of magnetic layered structures have
been found which show sizable values of GMR. Highest values are obtained in
magnetic multilayer structures, such as Fe/Cr and Co/Cu, which remain attrac-
tive from the point of view of studying the fundamental physics involved. The
exchange-biased spin valves show a combination of properties that make these
systems more useful for applications in low-field sensors, such as read heads for
magnetic recording.
The discovery of GMR has stimulated significant progress in the theory of
electronic transport in magnetic layered structures. First theories were based
on simplified band structures, such as free-electron models or single-band tight-
binding models. These models are physically transparent and, though simple,
capture some important physics of GMR. For example, a free-electron theory
based on the semiclassical description of electronic transport provided an un-
derstanding of the thickness dependence of CIP GMR in magnetic multilayers
and spin valves. However, these simple band models have no predictive power
for the quantitative description of GMR and, therefore, incorporating an ac-
curate band structure of the multilayer within the theory of GMR is crucial.
Another important ingredient for the predictive modeling of the conductivity
and magnetoresistance in real metallic layered structures is using a quantum-
mechanical theory of transport. The semiclassical Boltzmann transport theory
can only be applied in systems with sufficiently low defect/impurity density. It
breaks down in magnetic multilayers of practical interest because the subband
energy splitting is comparable with the life-time broadening due to scatter-
ing. In these circumstances the quantum-mechanical theory within an accurate
multiband treatment of the electronic structure is the only way to make defini-
tive statements about the origins of GMR.
The principal challenge for first-principle modeling lies in the realistic descrip-
tion of the defect scattering. A reliable approach must include contribution to
the resistivity from intrinsic structural defects, which are always present in
experimental conditions. However, proper first-principle treatment of all the
existing defects in the multilayers is very complicated and, therefore, reliable
simplified models become of great importance. The presence of structural de-
fects may also be decisive for predicting realistic values of the scattering spin
asymmetries of impurities, since some results obtained in the dilute limit in
the absence of structural defects disagree strongly with experiment. Including
misaligned magnetic moments at the interfaces in the transport theory is an

226
important issue, which has not yet been considered in detail. Progress in devel-
oping a fully relativistic theory of transport may help to elucidate the influence
of spin-orbit scattering on GMR and to calculate the spin-diffusion length in
those cases when it originates from the spin-orbit interaction at some impuri-
ties. Considering the effect of the electron-phonon scattering is important for
a description of the temperature dependence of GMR. Little has been done to
include this mechanism of scattering in the theory to date. The problem is,
however, very complicated—it is already difficult to calculate the temperature
dependence of the resistivity of bulk non-magnetic metals.
Recent experimental and theoretical results indicate that the mechanism of
GMR within the CPP geometry has not yet been fully understood. It has been
believed for a long time that the only relevant scale which governs CPP GMR is
the spin-diffusion length: once the overall thickness of a magnetic multilayer is
shorter than the spin-diffusion length, the series-resistor model becomes valid,
in which there are no relevant lengths except the layer thicknesses. It ap-
pears, however, that if the layer thicknesses are less than or comparable to the
mean free path the interface resistance depends on the layer thicknesses making
the series-resistor model inapplicable. The mean free path becomes, therefore,
an additional scale, which controls CPP GMR. The application of the series-
resistor model is not in general valid for analyzing experimental results, even
when the spin-diffusion length is infinite. On the other hand, deviations from
the series-resistor model do not necessarily imply that spin-flip scattering pro-
cesses cause these deviations. Thus, one has to re-examine the parameters
obtained from experimental data that have been modeled by the series-resistor
model. Comparison of experimental results on CPP GMR with calculations
based on accurate multiband models is greatly desirable.
Spin valves containing an antiferromagnetic oxide layer, such as NiO, for
exchange biasing and a noble metal overlayer are of increasing interest. This is
due to the enhanced values of GMR which can be obtained in these spin valves
and which makes them attractive for applications. These enhanced values are
explained by the specular scattering of the electrons from the ferromagnet/oxide
and ferromagnet/noble metal interfaces.
Finally, although magnetoelectronics applications based on magnetotransport
phenomena distinct from GMR, such as tunneling magnetoresistance (TMR)
or spin injection into semiconductors, have recently started to attract more
and more attention, the physics of GMR in metallic layered structures is so
multifaceted that it will undoubtedly remain the subject of great interest in the
future.

6.3 Spin effects in the Coulomb blockade

After having introduced the Coulomb blockade (CB) in Sect. 5.4.3, we now
discuss spin effects in the CB involving magnetic quantum dots. It is important
to note that the so-called orthodox theory (page 159) does not work here, as
we will see.
The Coulomb staircase was observed by many research groups and is well

227
understood. A straightforward way to study this effect is using scanning tun-
neling microscopy (STM) of metallic islands grown on a thin insulating film
which itself is deposited on a metal substrate [130]. In this double-barrier con-
figuration, the metallic substrate and the metallic tip of the STM are the two
leads. The metallic island is the nanoparticle, being electrically decoupled from
the substrate and the STM tip by the insulator film and by the vacuum, re-
spectively. The striking advantage of STM is to combine imaging (Section 4.4)
with controlled recording of the I(U ) characteristics of individual nanoscale
islands [130].
Here, we discuss the Coulomb staircase in small Fe islands (as quantum dots)
and illuminate the strong influence of the electron’s spin on the transport prop-
erties. The I(U ) curves show characteristic kinks, i. e., the signature of the
Coulomb staircase (indicated by arrows in Fig. 6.46). The change of the elec-
trostatic potential Ug upon charging an Fe island with a single electron can be
directly obtained from the I(U ) curve. It corresponds to the separation of (the
first) Coulomb stairs in the current [Ug = (2.6 ± 0.1) eV for the island indicated
in Fig. 6.46b]. This is the Coulomb blockade.
As an alternative to local I(U ) curves, the current at U = −3.8 V was plotted
versus the position of the STM tip. Here, the metallic Fe islands show up as
black in Fig. 6.46c, as they display a negative tunnel current while the MgO
insulator shows a vanishing current. These local current images were used to
determine the lateral extension of the Fe islands with higher precision than by
topographic images (Fig. 6.46a).
Having determined the area A of an island, its capacitance C can be approx-
imated by that of a plate capacitor,
A
C = MgO 0 , (6.101)
d
where 0 and d are the electrical permittivity constant and the MgO spacer
thickness, respectively. Since the electrostatic energy EC is proportional to
e2 /C, plotting the experimentally observed Ug = EC /e versus 1/C should result
in a straight line through the origin with slope e.
The experimental gap Ug is displayed as function of the inverse estimated
island capacitance for Fe and Pd islands in Fig. 6.47. The Coulomb gap of
nonmagnetic Pd islands shows the expected behavior: a linear fit of Ug versus
1/C (dashed line) goes within the error margin through the origin. Its slope
is close to the expected one (dotted line). In fact, the slope is slightly less,
which might be explained by the assumptions, i. e. neglect of the island-tip
capacitance, neglect of edge effects in the substrate-island capacitance or due
to a slightly different dielectric constant of an MgO film with respect to that of
MgO bulk.
Within the error bars, the slope for the Fe islands (solid line in Fig. 6.47)
is identical to that of the Pd islands. This finding provides evidence that the
MgO spacer decouples electronically both the Fe and the Pd islands from the
Fe substrate [131, 132]. This implies that the electrons have to tunnel from the
Fe substrate to the Fe island, as in double-step potential.
The most striking difference with respect to the Pd result, however, is that

228
Figure 6.46. (a) Topography of Fe islands (≈0.1 ML Fe) on a 3.5 ML MgO film on
Fe(001) taken with I = 1 nA and U = 3.8 V. The inset shows a line profile. (b)
Magnified image of several Fe islands. (c) Map of the local tunnel current at −3.8 V.
(d) I(U ) characteristics measured with the STM tip on-top of the bare MgO insulator
(dotted) and an Fe island on MgO (solid). The inset shows I(U ) curves of a larger
island.

the linear fit does not hit the origin. In fact, it shows a considerable offset of
about 1 V. Note that data for small 1/C in Fig. 6.47 are associated with large
islands, the relative size error of which is small. This proves that the offset is
not due to errors in the determination of the island sizes.
Summarizing at this point, we find a striking difference in the Coulomb block-
ade of magnetic Fe islands compared to that of nonmagnetic Pd islands, namely
an offset of the gap plotted versus inverse capacitance for Fe islands (Fig. 6.47).
Since the slopes of the Ug (1/C) curve of Fe and Pd islands are identical, this
unexpected behavior of the Fe islands suggests an explanation in terms of mag-

229
Figure 6.47. Experimental gap Ug for Fe and Pd islands on MgO films on Fe(001)
versus the inverse estimated island capacitance and linear fits. For comparison, the
estimated relation (ed)/(0 A) ≈ e/C is shown as a dotted line.

netic properties, that is a spin effect.

In the following, we restrict ourselves without loss of generality to the case of
islands charged by a single additional electron (n = 0 → n = 1). Without CB,
a bias voltage U > 0 opens an ‘energy window’ for tunneling in which occupied
electronic states in the source lead tunnel into unoccupied ones in the drain
lead [cf. eq. (5.166)].
In the drain lead, this energy window extends from the Fermi level EF to
EF + eU . The current I from the source to the drain lead is given by I = GU ,
G being the conductance of the device.
In the CB regime, the tunneling electrons have to overcome the electrostatic
energy EC of the island. Hence, the energy window becomes effectively reduced
in size, namely
(
G(U − Ug ) U > Ug
I= . (6.102)
0 U ≤ Ug

The findings for the Fe islands can be modeled by an additional term U0 =

1 V. Assuming that U0 is independent of the island size, the gap of Fe islands
can be expressed as
e e d
Ug (C) = + U0 = + U0 . (6.103)
C MgO 0 A

For Pd islands, note that U0 = 0 V implies that Ug → 0 V for 1/C → 0

(Fig. 6.47). The spin-dependent case of Coulomb staircase can be treated simply
within a two-current model, i. e., the total current I is the sum of the currents
of spin-up (↑) and spin-down (↓) electrons, I = I↑ + I↓ .

230
 While the Fe substrate shows a significant density of states (DOS) at the
Fermi level in both spin channels (Fig. 6.48a), the DOS of the Fe island reveals
prominent spectral weight only in the ↓-channel (Fig. 6.48b). Considerable
weight in the ↑-channels shows up at energies less than −1 eV. At energies larger
than −1 eV, the ↑-DOS is almost constant and comparably small. Although it
is not equal to zero, this property can be regarded as half-metallic or as close to
being half-metallic. In contrast to the Fe film, the Pd film displays significant
spectral weight in both spin channels at energies around EF .

Figure 6.48. Density of states of Fe and Pd films on MgO/Fe(001). (a) Spin-resolved

density of states (DOS) of Fe bulk. (b) and (c) As (a), but for the surface layers of
2 ML-films of Fe and Pd deposited on 4 ML Mg)/Fe(001), respectively. The energy
shift E0 = 1 eV discussed in the text is indicated by a horizontal arrow in (b).

This difference has a direct impact on the electron transport. The Fe islands
are quasi half-metallic, i.e. a band gap for ↑-electrons shows up at the Fermi
level but not for ↓-electrons. Hence, to have I↑ nonvanishing, ↑-electrons have
to overcome the binding energy E0 and the electrostatic energy, i. e.,
I↑ = G↑ (U − U0 − e/C) for U > U0 + e/C, (6.104)

231
with U0 = E0 /e. In contrast, the ↓-electrons have to overcome only the elec-
trostatic potential:

I↓ = G↓ (U − e/C) for U > e/C. (6.105)

This leads to a complete CB for U < e/C (i. e., both ↑- and ↓-currents are
inhibited). One would observe a first kink at the onset of the ↓-channel (at
e/C) and a second kink at the onset of the ↑-channel in the I(U ) spectrum
(at e/C + U0 ). Further, if G↓ is small compared to G↑ , one would observe
mainly the kink at U = e/C + U0 , which results in an offset by U0 in the
U (1/C) data. Indeed, ab-initio calculations have shown that G↓  G↑ due to
selective transmission of ∆1 electrons of Fe through the MgO film [133–135].
These theoretical predictions were confirmed by the observation of giant tunnel
magnetoresistances in this system [136]. Concluding, the large asymmetry in
the spin-resolved conductances together with the significantly shifted energetic
position of the ∆↓1 -band (with respect to the ∆↑1 -band) explains the observed
offset of 1 eV in Ug .

6.4 Magnetism at the nanoscale

Up to now, we have considered electronic transport in magnetic nanostructures,
taking the magnetic structure for granted. However, magnetic nanostructures
show a lot of phenomena that are not present in the respective bulk material.
Before discussing these effects, the interlayer exchange coupling will be discussed
because we have come across this phenomenon in GMR devices (spin valves and
multilayers). Note that at zero external magnetic field a GMR device is typically
in an antiparallel (AP) configuration.

6.4.1 Exchange coupling in magnetic multilayers

Varying the thickness of a layer separating two magnetic layers can give rise to
dramatic oscillations in the magnetic coupling between them. [137–139] A mag-
netic multilayer consists of a sequence of thin films grown on top of each other,
each film ranging in thickness from tenths of nanometers to tens of nanometers.
Typically, two thin films of ferromagnetic material are separated from each
other by a thin film of a nonmagnetic material, referred to as a spacer layer
(Fig. 6.49). The magnetization directions of the ferromagnetic layers are cou-
pled to each other through an exchange interaction. The sign of this coupling
oscillates as the thickness of the spacer layer is varied, with the best multilayer
samples showing up to thirty periods of oscillation.
There are a lot of theoretical studies devoted to the theory of IEC, a few
cited in this section. [140–144]

Oscillations of physical quantities. In metals, the sharp cut-off in occupancy

at the Fermi surface (Fermi-Dirac distribution!) causes oscillatory phenomena.
The oscillation of the induced magnetization as a function of applied mag-
netic field, called the de Haas-van Alphen effect, is a well known example.

232
Figure 6.49. Typical magnetic multilayer. Here two Fe layers are separated by an Au
spacer layer. An Au capping is grown on top to protect the multilayer from corrosion.
The multilayer was grown on a GaAs substrate with a buffer layer of Ag to promote
better growth of the Fe and Au films.

Many other properties oscillate as well, for example the conductivity, known
as Shubnikov-de Haas effect (page 137). These oscillation arise from field
induced oscillations in the density of states. The cross-sectional areas of the
Fermi surface, a geometrical property, determine the oscillation periods. This
is the origin of the integer quantum Hall effect, in which Landau levels are
shifted through the Fermi level by the external magnetic field (Sect. 5.3.2).
Spatially localized disturbances in metals lead to another type of oscillation.
The oscillation of the electron density near surfaces of metals, known as Friedel
oscillations, is an example. Another example is the oscillation in the inter-
action between two localized magnetic impurities in a metallic host. As the
separation between the impurities is increased, the interaction between them
oscillates between favoring parallel alignment and antiparallel alignment of the
magnetic moments. This coupling of the moments is known as the Ruderman-
Kittel-Kasuya-Yosida (RKKY) interaction. A different geometrical property
of the Fermi surface determines the spatial periods of the oscillations between
the localized disturbances, in this case, a vector spanning the Fermi surface
from one side to the other (‘nesting vector’). Interlayer exchange coupling is a
particularly dramatic instance of this type of oscillation.

Excursion: Friedel oscillations. One of the fundamental concepts of solid-

state physics is screening, that is, the ability of a material to rearrange its
‘electron sea’ in order to minimize the disturbance of an impurity or defect.
Friedel showed that screening of a point-charged impurity in the bulk results in
long-range oscillatory behavior in the charge density. For simple free-electron-
like metals, these oscillations have a wave vector of 2kF as a consequence of the
sharp Fermi surface in a metal. At a metal surface, this screening phenomenon

233
results in an oscillation in the charge density perpendicular to the surface as well
as a lateral disturbance from impurities or steps. Friedel oscillations constitute
the mechanism for long-range interaction in metals.
Friedel solved the screening problem by deriving a sum rule on the scattering
phase shifts necessitated by charge neutrality. A more intuitive approach is to
use dielectric response theory, where the static dielectric constant (q) is used
to screen the impurity or defect potential V0 (r). The 3D screened potential is
given by
Z
V0q iq·r 3
Vsc (r) = e dq (6.106)
(q)

where V0q is the Fourier transform (FT) of V0 (r). The induced potential is
given by

Vind (r) = Vsc (r) − V0 (r), (6.107)

and the induced charge density, ρind (r), can be determined from Vind (r) with
Poissons equation,

div Vind (r) = −4πeρind (r). (6.108)

The relation between the static dielectric constant and the susceptibility for a
3D impurity is
χ(q)
(q) = 1 − 4π , (6.109)
q2
for a 2D impurity the relation is
χ(q)
(q) = 1 − 4π . (6.110)
q
Here, χ(q) is the electric susceptibility,

ρind (q) = χ(q)V (q). (6.111)

This changes the functional form of the induced charge density. If the Lindhard
response function

6πe2 n
(q) ≈ 1 + (6.112)
EF q 2
is used, the 3D asymptotic limit for large r gives an induced charge density
from a point impurity of the form
cos(2kF r + Φ)
ρ(r) = A . (6.113)
r3
where A is amplitude and Φ is a phase shift associated with the scattering poten-
tial. The long-range oscillations are caused by the singularity in the derivative
of the Lindhard dielectric function at q = 2kF . Both the Friedel scattering

234
Figure 6.50. A quantum corral of 60 Fe atoms assembled and viewed on Cu(001) by
STM at 4 K.

Figure 6.51. (A) Constant-current STM image (40 Å by 40 Å) of Be(0001). (B) The
2D FT of the image in (A). (C) The 2D Brillouin zone of Be(0001) in which the circles
(shaded region) correspond to the surface states (projected bulk bands) at EF . The
reciprocal space unit cell with the corresponding 2kF ‘ring’ is also shown.

approach and the Lindhard dielectric response approach are based on single-
particle concepts.
RKKY interaction
In this section we will discuss the interlayer exchange interaction within a
quantum-well model. Before doing so, we are going to sketch the RKKY inter-
action.
Indirect exchange couples magnetic moments over relatively large distances.
It is the dominant exchange interaction in metals where there is little or no di-
rect overlap between neighboring magnetic electrons. It therefore acts through
an intermediary which in metals are the conduction electrons (itinerant elec-
trons). This type of exchange was first proposed by Ruderman and Kittel and
later extended by Kasuya and Yosida to give the theory now generally know as
the Ruderman-Kittel-Kasuya-Yosida interaction (RKKY interaction).
The interaction is characterized by a coupling coefficient j given by
j2
j(Rl − Rl0 ) = 9π F (2kF |Rl − Rl0 |). (6.114)
EF
where kF is the radius of the Fermi surface, Rl is the lattice position of the
magnetic moment, EF is the Fermi energy and
x cos x − sin x
F (x) = . (6.115)
x4

235
The RKKY exchange coefficient j, oscillates from positive to negative as the
separation of the ions changes and has the damped oscillatory nature (Fig. 6.52).
Therefore, depending upon the separation between a pair of ions their magnetic
coupling can be ferromagnetic or antiferromagnetic. A magnetic ion induces a
spin polarisation in the conduction electrons in its neighborhood. This spin
polarisation in the itinerant electrons is felt by the moments of other magnetic
ions within range, leading to an indirect coupling.

Figure 6.52. Variation of the indirect exchange coupling constant j, of a free electron
gas in the neighborhood of a point magnetic moment at the origin r = 0.

In rare-earth metals, the spin-polarized electrons in the 4f shell are screened

by the 5s and 5p electrons, direct exchange is rather weak and insignificant and
indirect exchange via the conduction electrons gives rise to magnetic order in
these materials.
Quantum well model
Model for transition-metal ferromagnetism. Developing a simple model for
interlayer exchange coupling (IEC) requires starting with a simple model for
the electronic structure of ferromagnets. [146, 147] Unfortunately the electronic
structure of transition metal ferromagnets is quite complicated. Ferromag-
netism in transition metals is driven by atomic-like exchange and correlation
effects in the partially filled d-electron shells. The atomic-like effects suggest
a localized description of this part of the electronic structure. However, the d
orbitals are strongly hybridized with d orbitals on neighboring atoms and also
with the sp orbitals on neighboring atoms. The strong hybridization suggests
an itinerant description of the electronic structure.
Simplifying the model requires approximations that favor one aspect of the
physics over the other. Here, the itinerant aspects are favored because the

236
interlayer exchange coupling depends strongly on the properties of the electrons
at the Fermi surface and a realistic description of the Fermi surface requires
treating the itinerant nature of the d electrons.
A material becomes ferromagnetic when it is energetically favorable for a ma-
jority of electrons to align their spins parallel to one another. The electrons then
split into those with spins parallel to the majority of the other spins (majority)
and those antiparallel (minority). Spin-orbit coupling mixes these states, but
is typically weak enough to ignore to a first approximation. Ignoring the spin-
orbit coupling leads to a description of ferromagnets in terms of two separate
sets of electrons, majority and minority, which have different properties. This
description also holds in the nonmagnetic layers in magnetic multilayers pro-
vided the magnetizations in the different ferromagnetic layers are all collinear
with each other. When the magnetizations are not collinear, spin currents flow
in the nonmagnetic layers and exert torques on the ferromagnetic layers, as is
described below.
Figure 6.53 shows the band structure for Co in the face-centered cubic (fcc)
structure calculated using the LSDA, highlighting the importance of the itin-
erant aspects of the electronic structure. The d bands have a width in energy,
approximately 5 eV, that is large compared to the exchange splitting between
the majority and minority bands, about 2 eV. In addition, the hybridization
between the d bands and the sp band is large enough to lead to a gap of about
3 eV where the bands would cross if they were not hybridized. Finally, images
of the spin-resolved Fermi surface for majority and minority electrons are shown
on the right. Clearly, these are quite different, and the two sets of electrons
have different properties.
There are two simplified models for the electronic structure of ferromagnets
that have been used extensively in studies of magnetic multilayers. Each empha-
sizes different aspects. Both include free-electron-like bands. In the spin-split
free-electron model, the free-electron-like bands for majority and minority elec-
trons are shifted in energy relative to each other by a constant exchange shift
(‘exchange splitting’). This model ignores the d electrons completely, but the
Fermi surfaces for majority and minority electrons are different.
The other model, called the s–d model, emphasizes the atomic-like aspects
of the d orbitals by ignoring the d–d hybridization and treating these orbitals as
completely localized. This model was originally used to treat isolated magnetic
impurities in a nonmagnetic host. For ferromagnets and magnetic multilayers,
the model for isolated impurities is generalized to treat a dense set of such
impurities. In this model, the sp electrons are weakly hybridized with the d
electrons.
Spin-polarized quantum well states
An interface between two materials acts on the electrons like a potential step.
Electrons that strike it have a transmission probability reduced from 1. For a
free electron going down a simple potential step of height V the transmission

237
Figure 6.53. Band structure and Fermi surface of face-centered cubic (fcc) Co. The
red (blue) curves give the majority (minority) bands along two high symmetry direc-
tions through the Brillouin zone center, Γ. The dotted black curve shows the sp band
if it were not hybridized with the d bands. The bars to the right of the bands show
the width of the d bands and the shift between the majority and minority bands. The
red and blue arrows in the band structure plots give the width of the gap due to the
hybridization between the sp and d bands of the same symmetry along the chosen di-
rections. Note that Co crystallizes in the hexagonal-closed-packed (hcp) structure but
in fcc if grown on Cu(001), for example.

probability is
 2
q 2k
Tstep = . (6.116)
k k+q

Here, the incident wave vector is

r
2mE
k= (6.117)
~2
and the transmitted wave vector is
r
2m(E + V )
q= . (6.118)
~2
The first factor accounts for the change in velocity on going over the step. The
transmission probability is plotted in Fig. 6.54.
Problem 86. Derive eq. (6.116).
The wavefunction om the left-hand side reads

ψ(z) = eikz + R e−ikz

the outgoing is given by

ψ(z) = T eiqz .

238
 Figure 6.54. Electron transmission probabilities for a step and a quantum well.

Wavefunction matching requires that the logarithmic derivatives are identical at the match-
ing position, see page 30. This gives at z = 0

ik (1 − R)
= iq,
1+R
and hence
k−q
R= .
k+q

Current conservation requires 1 − R2 = T 2 . Thus,

2
(k − q)2 (k + q)2 − (k − q)2

2 4kq q 2k
T =1− = = = ,
(k + q)2 (k + q)2 (k + q)2 k k+q

which is just Tstep . 2

If another step is introduced, the electron undergoes multiple reflection inside
the quantum well that is formed (see Fig. 6.54b). If the steps are separated by
a distance D, the transmission probability is
2
4 eiqD kq
Tqw = . (6.119)
(k + q)2 − e2iqD (k − q)2

Note that the transmission probability goes to 1 whenever 2qD = 2nπ, where n
is an integer. At a fixed thickness D, there is a series of transmission resonances
at the energies
 nπ 2
En = 2m~2 − V, (6.120)
D
for integer n. At a fixed energy E, there are resonances for
r
2n 2m(E − V )
D= with q = . (6.121)
2πq ~2
These resonances are separated by a fixed increment in thickness. At the trans-
mission resonances, the electrons have an increased probability to be in the

239
quantum well, in other words, there is a peak in the density of states in the
quantum well at the energy of the resonance.
These peaks in the density of states are seen in magnetic multilayers using
photoemission and inverse photoemission (see Sect. 4.2). Photoemission is a
technique in which photons of a particular energy, generally ultraviolet light
or x-rays, are incident on a surface. When the photons are absorbed by the
material, they excite electrons which can leave the surface and be measured.
The energy of the electron in the solid can be inferred from the photon en-
ergy and the energy of the photoemitted electron. Peaks are observed in the
photoemission spectrum where there is a large density of states in the material.
Problem 87. Is the last statement always true? If not, why?
No, it isn’t. The photocurrent depends strongly on the transition matrix elements, see
page 92. Hence, the current can be small although the density of states of the initial states
is large, depending on the photoemission set-up (light incidence direction, light polarization,
etc.). Only the reversed statement holds: zero DOS implies vanishing photocurrent. 2
Photoemission probes the near surface region because the escape depth of
the photoemitted electrons is on the order of a nanometer. In order to see the
density of states peaks in the non-magnetic spacer layer, the top magnetic layer
needs to be stripped off (or never deposited in the first place). In other words,
the quantum well states studied in photoemission are not exactly the same
quantum well states that are present in the complete magnetic multilayer (cov-
ered layer). Nonetheless, there is good correspondence between these states and
the related states in magnetic multilayers. Figure 6.55 shows the photoemission
intensity as a function of energy and spacer layer thickness. This figure shows
the fixed spacing between peaks as a function of thickness and the quadratic
variation of the peaks as a function of energy. There are some differences be-
tween what would be expected for a free electron model and what is observed.

Figure 6.55. Photoemission from a Cu overlayer on Co. Yellow indicates high pho-
toemission intensity and red low intensity.

To understand how the free electron model generalizes to real materials it is

240
instructive to rewrite the transmission probability in terms of the transmission
probability for a step and the reflection amplitude at an isolated step,
k−q
R= . (6.122)
k+q
Problem 88. Derive eq. (6.122).
See problem 86. 2
This substitution emphasizes the contribution made by multiple reflection
inside the quantum well. One round trip through the well has the amplitude

ei2qD R2 (6.123)

from reflecting from each step and propagating both directions through the
well. The transmission probability is
∞ 2
2 n
X
2 iqD i2qD


Tqw = T e e R (6.124)
n=0

The second form shows explicitly the coherent multiple scattering in the well.
Problem 89. Using transmission and reflection amplitudes for the steps, derive the above
equation.
The total transmission can be obtained as follows. In zero-th order, the incident wave
is transmitted at the left step, propagates to the right step, and is transmitted to the right
region. This gives a transmission of

T0 = T eiqD T.

In first order, it is additionally reflected at the right step and makes one round-trip, leading
to

T1 = T eiqD ReiqD ReiqD T = T eiqD (ReiqD )2 T.

For two round-trips one obtains analogously

T2 = T eiqD (ReiqD )4 T.

Defining

Tn ≡ T eiqD (ReiqD )2n T,

the total transmission is given by

∞
X ∞
X ∞
X
Tn = T 2 eiqD (ReiqD )2n = T 2 eiqD (R2 e2iqD )n ,
n=0 n=0 n=0

which establishes the result. 2

Summing up the geometrical series gives
2
eiqD
Tqw = Tstep . (6.125)
1 − ei2qD R2
The basic physics of quantum well states in real materials is captured by re-
placing the wavevector for propagating through the spacer layer, q, by the

241
appropriate value from the real band structure and by replacing the reflection
amplitude and transmission probability by the values calculated for a realistic
interface.
If one of the materials is ferromagnetic, the potential step for the majority
and minority electrons is of different height. In a multilayer with two magnetic
layers there are four possible quantum wells formed depending on the relative
alignment of the magnetizations, see Fig. 6.56. The quantum well states for

Figure 6.56. Quantum wells used to compute interlayer exchange coupling. On the
left, the two panels give typical band structures for free electron models of interlayer
exchange coupling. On the right, the four panels give the quantum wells for spin-
up and spin-down electrons for parallel (P) and antiparallel (AP) alignment of the
magnetization. The shaded regions designate the occupied states.

each of these are different because the potential steps, and hence reflection
probability, is different for each quantum well. However, at a particular energy,
like the Fermi energy, the quantum well states in all of the wells have the
same periodicity as a function of the thickness of the spacer layer, because the
periodicity only depends on the wave length of the electron in the spacer layer
at that energy.
Interlayer Exchange Coupling
The interlayer exchange coupling can be described in terms of an energy that
depends on the magnetization directions of the two layers, mi ,

E = −JA m1 · m2 . (6.126)

where A is the area of the two films, and the exchange integral J < 0 gives
antiferromagnetic coupling favoring antiparallel alignment. The form of the
coupling is called bilinear. For interlayer exchange coupling, the strength of
the exchange interaction is determined by the difference in energy between the
quantum well with parallel magnetizations and that with antiparallel magne-
tizations. Computing the exchange interaction simply requires computing the
energies of the quantum wells given in Fig. 6.56. For large spacer layer thick-

242
nesses, the result is
~vF h i
J≈ Re (R↑ R↑ + R↓ R↓ − R↑ R↓ − R↓ R↑ ) ei2kF D + O(D−2 ),
2πD
(6.127)
or
~vF
J≈ |R↑ − R↓ |2 cos(2kF D + φ). (6.128)
2πD
Here, kF is the Fermi wave vector in the spacer layer, vF = ~kF /m is the
Fermi velocity, m is the is the electron mass, and R↑ and R↓ are the reflection
amplitudes for a majority and minority electron to reflect from the interface.
Expanding the square of the reflection amplitudes gives four terms, one for each
of the quantum wells in Fig. 6.56.
Problem 90. Derive equation (6.127).
The density of states (per unit area and unit width of the QW) is given by
2 dk
,
π dE
including the spin degeneracy. The effect of interference in the QW shows up as in increase
of the DOS (constructive interference) or as a decrease (destructive interference). Thus,
the DOS is modified in first order by

∆n(E) ≈ cos(2kD + 2φ),

where φ is the phase shift of the reflection, R = |R| exp(iφ). Further, the effect on the
DOS should be proportional to the well width D. Therefore, by summing up all round-trips
one has
∞
2D dk X
∆n(E) ≈ |R|2n cos[n(2kD + 2φ)].
π dE n=1

With Euler’s equation

cos x = Re [exp(ix)] = Im [i exp(ix)],

we have
∞ ∞
! !
2D dk X
2n 2inkD 2inφ 2D dk X
2n 2inkD
∆n(E) ≈ Re |R| e e = Re R e .
π dE n=1
π dE n=1

Summing up the geometrical series gives

R2 e2ikD
 
2D dk
∆n(E) ≈ Re .
π dE 1 − R2 e2ikD
The integrated DOS is defined by
Z E
N (E) ≡ n(E 0 ) dE 0 ,
−∞

or
dN (E)
n(E) = .
dE

243
Both equations hold also for ∆n(E) and ∆N (E). From the latter equation it follows
immediately
∞
!
1 X R2n 2inkD
∆N (E) = Re −i e ,
π n=1
n

where we have used

dN dN dk
= .
dE dk dE
Using
∞
X xn
ln(1 − x) = ,
n=1
n

we arrive at
1
∆N (E) = − Re i ln 1 − R2 e2ikD .


π
The finite size of the well changes the energy by
Z EF
∆E = (E 0 − EF ) ∆n(E 0 ) dE 0
−∞

Integrating by parts,
Z Z
u0 v = uv − uv 0 ,

results in
Z EF Z EF
1
∆N (E 0 ) dE 0 = ln 1 − R2 e2ikD dE 0 .


∆E = − Re i
−∞ π −∞

Thus, for the parallel magnetic configuration,

Z EF
1
ln 1 − R↑ R↑ e2ikD + ln 1 − R↓ R↓ e2ikD dE 0



∆EP = Re i
π −∞
Z EF
1
ln 1 − R↑ R↑ e2ikD 1 − R↓ R↓ e2ikD dE 0 .



= Re i
π −∞

In the limit of weak confinement, one obtains

Z EF
1
ln 1 − R↑ R↑ e2ikD − R↓ R↓ e2ikD dE 0


∆EP ≈ Re i
π −∞
Z EF
1
≈ Re i (R↑ R↑ + R↓ R↓ ) e2ikD dE 0 ,
π −∞

where the first-order terms were kept. For large D, the exponential factor oscillates rapidly,
leading to cancellation of contributions for small energies. Thus, the predominant contri-
bution is from the Fermi level. With
dE ~2 k
dE = dk = dk = ~v dk
dk m

244
and taking values at EF we obtain
~vF
Re (R↑ R↑ + R↓ R↓ ) e2ikF D
 
∆EP ≈
2πD
Analogously for the antiparallel configuration,
~vF
Re (R↑ R↓ + R↓ R↑ ) e2ikF D .
 
∆EAP ≈
2πD
The IEC energy then reads
~vF
Re (R↑ R↑ + R↓ R↓ − R↑ R↓ − R↓ R↑ ) e2ikF D .
 
J = EP − EAP =
2πD
2
The exchange coupling oscillates in sign with a period π/kF , the oscillation
decays like D−1 , and the amplitude of the oscillation is determined by the spin
dependence of the reflection amplitudes.
Physically these oscillations arise because the quantum well resonances cross
through the Fermi level as the thickness of the spacer layer is increased
(Fig. 6.55). Each time the thickness increases by π/kF another resonance crosses
through the Fermi level. The resonances for each of the quantum wells in
Fig. 6.56 are all different, but they all have the same period because the period
is determined by the Fermi wavevector in the spacer layer, and is independent
of the properties of the magnetic material.
Critical spanning vectors
The expression for the interlayer exchange coupling, eq. (6.128), is based on a 1D
model (cf. the Slonczewski model for the tunnel magnetoresistance, page 170).
Generalizing to realistic 3D systems is straightforward if the growth of the
multilayer is coherent. For coherent growth, the in-plane lattice of all of the
layers is the same, so the in-plane momentum (wavevector) is conserved when
the electrons scatter from the interfaces (2D translational invariance). In this
case, electrons with different values of in-plane wavevector do not couple, so
the quantum well states for different in-plane wavevectors are independent of
each other. Including the full three dimensions of the multilayer then simply
requires integrating the one-dimensional result over the in-plane wavevector kk ,

dk 2
Z
J ~ 2
vF (kk )Re R↑ (kk ) − R↓ (kk ) ei2kz (kk )D + O(D−3 ).

≈
A 2πD (2π)2
(6.129)

This integral is simple in the limit of large D; there is a contribution when-

ever the spacer-layer spanning vector, 2kz (kk ), is constant as a function of kk .
The vector connecting one sheet of the Fermi surface to another at this in-plane
wavevector is called a critical spanning vector. In general, the oscillating con-
tributions from different parts of the Fermi surface are out of phase and tend to
cancel each other.9 However, when the spacer-layer spanning vector is constant
9
This is similar for the electron diffraction from a surface, giving rise to LEED spots
(Sect. 4.1.)

245
as a function of kk , the contributions from finite region of the Fermi surface near
the critical point are in phase. This region gives a contribution to the integral
that oscillates in thickness with a period determined by the critical spanning
vector. As the thickness D becomes larger, the region of the Fermi surface
that contributes in phase gets smaller so that the amplitude of the oscillation
decreases with an additional power of D.
For free electrons, the critical spanning vector is qc = 2kF , the vector that
goes from one side of the sphere to the other (Fig. 6.57). For a free electron
model, the interlayer exchange coupling is

Figure 6.57. Slices through a free-electron and the Cu Fermi surfaces. For the
spherical free-electron Fermi surface, there is one critical spanning vector qc = 2kF . For
the Cu Fermi surface, there are two inequivalent critical spanning vectors, qL which is
closely related to the free electron critical spanning vector, and qS which occurs in the
necks of the Fermi surface. The Cu Fermi surface is shown within the Brillouin zone
appropriate for Cu(001) multilayers. Indicated in gray are parts of the Fermi surface in
the extended zone scheme. The critical spanning vector qL0 is equivalent to qL because
of the discrete nature of the lattice.

J ~vF kF
≈− Re [R↑ (0) − R↓ (0)]2 sin(2kF D) + O(D−3 ). (6.130)
A 16π 2 D2
The in-plane wavevector dependence of the Fermi velocity and the reflection
amplitudes is ignored to lowest order. The oscillation period only depends
on the geometry of the spacer-layer Fermi surface, but the strength of the
oscillation depends on the Fermi surface through vF and kF and also the details
of the electronic structure of the ferromagnet through R↑ and R↓ .
The generalization to realistic materials is illustrated for Cu in Fig. 6.57. For
a free-electron model, the only critical spanning vector is one of length 2kF
connecting one side of the Fermi sphere to the other. However, for realistic
materials, the Fermi surface is more complicated than a sphere. In general
there are several critical spanning vectors and there is a contribution to the
coupling from each. For thick spacer layers, the coupling is
X ~v (α) κ(α) 
(α) (α) 2 iq (α) D iχ(α)
 
J≈ ⊥
2 D2
Re R↑ − R↓ e ⊥ e + O(D−3 ). (6.131)
α
16π

246
 (α)
For each critical point (indicated by α), q⊥ is the critical spanning vector,
(α)
v⊥ is the Fermi velocity, κ(α) is the average radius of curvature of the Fermi
surface, and exp(iχ(α) ) is a phase that depends on whether the stationary point
is a minimum, maximum or saddle point. Equation (6.131) is known as the
asymptotic formula and is valid for thick spacer layers. For thinner spacer
layers, the O(D−3 ) represents preasymptotic corrections. [148]
Coupling strength
While the critical spanning vectors and other properties of the Fermi surface
can be extracted from experiment, the reflection amplitudes cannot. For real
multilayers, these can be calculated, [149] but the resulting coupling strengths
do not agree nearly as well with measured values as the periods do. Some of
the disagreement is due to experimental difficulties, but some disagreement is
due to theoretical difficulties.
The asymptotic form is an approximation to the difference in energy between
the total energies of multilayers with parallel magnetizations and antiparallel
magnetizations. Unfortunately, due to the complexities of the electron-electron
interaction, it is currently impossible to compute these total energies without
approximation. The best available approximation, as described in the beginning
of this section, is the local spin density approximation (LSDA). This approx-
imation works quite well for magnetic multilayers but with one caveat. The
band structure is only approximate, so the Fermi surface deviates from its ac-
tual shape. Since the oscillation periods of the interlayer exchange coupling
depend on the critical spanning vectors of the spacer layer Fermi surface, the
periods computed using the LSDA will be wrong. This means that after a few
oscillations, the calculated coupling is out of phase with the experimental cou-
pling. The sign of the coupling may even be wrong. A direct comparison is
then misleading because even if the physics is essentially correct, the agreement
might be quite poor (Fig. 6.58). There, the agreement is made even worse by
the effect of disorder on the measured results.
In Figure 6.58, the calculated and measured periods of the oscillations dis-
agree; the periods extracted from the experiment are 2.48 ± 0.05]M L and
8.6 ± 0.3 ML and the critical spanning vectors of the theoretical Fermi sur-
face would give periods of 2.65 ML and 8.03 ML. One way to compensate for
the errors in the periods is to fit both theory and experiment to the asymptotic
form, and compare the envelopes.
Measurement of Interlayer Exchange Coupling
Growth and Disorder Growing magnetic multilayers to compare measure-
ments with calculations is quite difficult. Calculations are only tractable for
systems that are close to ideal, requiring growth of systems equivalently close
to ideal. The first requirement is that the lattices of the different materials need
to be compatible. For example, when Co grows on Cu it grows pseudomorphi-
cally, that is, it adopts the fcc structure of Cu, with a lattice constant that is
very close to that of Cu.
Unfortunately, these are the only two pairs of materials with such similar
crystal structures. The only other pairs that can be grown sufficiently well are

247
Figure 6.58. Calculated and measured coupling strengths for Fe/Au/Fe(100) multi-
layers. The red (blue) symbols are the measured (calculated) coupling strengths. The
red curve is the best fit to the experimental data when the measured thickness fluc-
tuations are taken into account. The blue curve is a linear interpolation between the
coupling strengths calculated for complete layers.

Au or Ag and Fe, but only in the (001) interface orientation. It is somewhat

surprising that these systems can be grown well because Fe has a body-centered
cubic (bcc) structure while Au and Ag have an fcc structure. In addition, the
lattice constants are very different. However, it turns out that if the Fe lattice is
rotated by 45 ◦ around the interface normal, there is very good in-plane lattice
match for each of these pairs. If the starting substrates are sufficiently flat, very
good growth can be achieved. However, the presence of steps leads to extended
defects through the layer because the Au/Fe growth is not pseudomorphic and
the layer thicknesses are quite different.
Even when the lattice mismatch is close to zero, real multilayers are still
not perfect. The starting substrate is never perfectly flat and the growth is
never perfectly layer by layer, so there are always variations in the thickness of
the layers, typically called thickness fluctuations. Frequently, the lateral length
scales of the thickness fluctuations are in an intermediate regime. The flat ter-
races are large enough that the ideal coupling is reasonably well defined over
each terrace. On the other hand, the terraces are small enough that the mag-
netizations do not vary significantly on that length scale. In this intermediate
regime, the comparison between calculated and measured coupling strengths is
improved by averaging over measured variations of the spacer layer thickness.
If the coupling for an ideal thickness of n layers of is J(n) and the probability
of having a thickness nd for a nominal deposited thickness of D is P (n, D), the
effective coupling strength is
X
J(D) = P (n, D) J(n). (6.132)
n

248
 One of the key advances that allowed accurate determination of the oscilla-
tion periods was the use of wedge-shaped spacer layers (Fig. 6.59). Growing
wedge samples simply involves moving a shutter between the sample and the
evaporator during growth to expose different parts of the sample to different
total fluxes. Such samples ensure that all thicknesses are grown under the
same conditions because they are grown simultaneously on the same substrate.
Wedge samples also make it easier to accurately determine the thickness of the
samples.

Figure 6.59. Interlayer exchange coupling in a wedge shaped Fe/Au/Fe trilayer

measured by SEMPA and RHEED. A schematic view of the wedge-shaped sample
is shown at the top of the figure. The approximate dimensions give an indication
of the very small slope of the wedge. Immediately below is a SEMPA image of the
magnetization of the Fe overlayer. White and black indicate parallel and antiparallel
alignment to the substrate, and hence ferromagnetic and antiferromagnetic coupling.
Below that is a line scan (cyan) through the image and then a measurement of the
RHEED intensity (green) along the wedge. The oscillations are used to determine
the thickness of the spacer layer along the wedge. The wedge is slightly curved. The
RHEED and the magnetization curves have been corrected for this curvature, but the
image has not, hence the variation of the lines connecting the image with the line scan.

Biquadratic coupling
In almost all multilayers, there is a contribution to the coupling that favors
perpendicular alignment of the magnetizations. In many multilayers, this con-
tribution dominates, leading to actual perpendicular alignment of the magne-
tizations. Phenomenologically, this alignment can be explained by a coupling

249
of the form
E
= −J2 (m1 · m2 )2 , (6.133)
A
called biquadratic in contrast to the bilinear coupling discussed above. It is
called biquadratic because it is quadratic in both of the magnetization direc-
tions. The fact that all measured values of J2 are negative, favoring perpen-
dicular orientation of the two magnetizations, shows that biquadratic coupling
does not have an intrinsic origin similar to the bilinear coupling, but has an
extrinsic origin due to disorder.
Thickness fluctuations lead to variations of the coupling strength on different
terraces. To lowest order, the intralayer exchange coupling forces the magneti-
zations in each layer to be uniform so that the bilinear coupling gets averaged
over the growth front [cf. eq. (6.132]. To next order, the magnetization can
fluctuate around its average direction. Over each terrace, the magnetization
fluctuates in the direction that lowers the energy. The fluctuations lower the
energy the most when the two magnetizations are perpendicular to each other.
This is the origin of the effective interaction favoring perpendicular alignment
between the magnetizations.
Consider the simple model shown in Fig. 6.60, in which the spacer layer
consists of parallel strips of width L with alternating thicknesses and hence
coupling strengths Jn and Jn+1 . The relative angle of the magnetizations is

Figure 6.60. Thickness variations and biquadratic coupling. The thickness of the
spacer layer varies periodically between n and n + 1 layers, with ferromagnetic coupling
for thicknesses n + 1 and antiferromagnetic for thicknesses n. The heavy arrows show
the local rotation in the magnetization direction into the direction that minimizes the
coupling for each terrace.

Θ = Θ0 + δΘ sin(πx/L), (6.134)

where δΘ is the size of the fluctuations. Over the region from 0 to L, where
the coupling is Jn , the energy change due to the fluctuations is proportional to
Jn sin(Θ0 ) δΘ. Over the region from L to 2L, the sine function changes sign and
the energy change due to the fluctuations is proportional to −Jn+1 sin(Θ0 ) δΘ.
The net coupling energy per area due to the fluctuations is proportional to

250
−∆J sin(Θ0 ) δΘ, where ∆J = Jn − Jn+1 . Fluctuations in the correct direction
lower the energy of the system. The energy gain is balanced by the cost in in-
tralayer exchange energy because the magnetization now varies spatially. Since
the intralayer exchange coupling depends on the square of the gradient of the
magnetization, for this simple model, it is proportional to (Aex /L) δΘ2 , where
Aex is the strength of the exchange interaction. Combining the changes due to
the fluctuations for the interlayer exchange coupling and the intralayer exchange
and finding the minimum with respect to the amplitude of the fluctuations gives
L
δΘ ∝ − sin(Θ0 ) ∆J . (6.135)
Aex
For this fluctuation amplitude, the change in the energy per area due to the
fluctuations gives the strength of the biquadratic coupling

(∆J)2 L
J2 ∝ − . (6.136)
Aex
While equation (6.136) is quite simple, it qualitatively describes the features of
more realistic situations. In real systems, a characteristic length scale L of the
arbitrary-shaped terraces replaces the width of the stripes. As this length scale
increases, the biquadratic coupling strength increases because the fluctuations
can get larger. Also, realistic growth fronts generally consist of more than two
thicknesses, which introduces an effective ∆J. The coupling increases as the
difference in the coupling for the different terraces get larger. The differences
tend to be largest when the coupling is oscillating rapidly, that is when there
is short period coupling. The coupling gets weaker as the intralayer exchange
interaction increases because exchange suppresses the fluctuations in magneti-
zation direction that lower the energy.
Summary
In magnetic multilayers, the magnetizations of two adjacent ferromagnetic lay-
ers separated by a nonmagnetic spacer layer are coupled by the electrons in the
spacer layer. This coupling oscillates in sign as a function of the thickness of
the spacer layer. Extensive research done on these systems has led to a simple
model for this coupling and remarkable agreement between predictions of the
model and measurements of the coupling. The model predicts that the periods
of the coupling are determined by geometrical properties of the Fermi surface
belonging to the spacer layer material. The oscillatory coupling is an instance
of oscillations in metals caused by the existence of a Fermi surface.

6.4.2 Magnetocrystalline anisotropy of ultrathin films

The magnetocrystalline anisotropy (MCA) is the orientation of the magnetiza-
tion in a solid along the crystallographic axes. The magnetization M will take
the configuration with the least total energy. In case of a ferromagnetic system
at zero temperature (Fig. 6.61), this means that the local magnetic moments
are aligned along the ‘easy axis’. But there are also noncollinear configurations
possible, for example spin waves.

251
Figure 6.61. Orientation of local magnetic moments in a paramagnetic and in a
ferromagnetic solid.

Problem 91. A local magnetic moment can be regarded as the spin moment integrated
over a unit cell or an atomic sphere. Does one expect strong deviations of the orientation
of the magnetization density m(r) from its average value?
The origin of deviations from the averaged magnetization direction are due to spin-orbit
coupling. As will be discussed in Sect. 6.4.2, the effect of SOC on the electronic structure
is minute for Fe, Co, and Ni. Therefore, one can expect slight deviations. 2
The magnetocrystalline anisotropy requires to couple the electronic degree
of freedom with the spin degrees of freedom of the electrons. Hence, spin-
orbit coupling (SOC) is necessary for the effect. Spin-orbit coupling removes
degeneracies in the band structure E(k), in dependence on the orientation of
the magnetization in different region of the Brillouin zone (BZ). In large parts
of the BZ there are no significant changes in the band structure. Hence, to
compute the MCA one needs very fine k-sampling. Note that the energy gain
per atom in bulk system in in the order of µeV.
The effect of exchange splitting and spin-orbit coupling is illuminated in
Fig. 6.62.
Problem 92. Find differences in the four band structure provided by Fig. 6.62
See the slides of Lecture 13. Small band gaps show up in the center of the Γ–Σ–N line.
Note that large blue circle might hide smaller red circles, so spin mixing shows up in both
spin channels but is not visualized perfectly. Further, the size of the circles represents the
contribution of the respective spin channel. 2
In thin films, the energy gain per atom is larger (in the order of meV) as com-
pared to bulk systems due to the lower symmetry of the system. Of particular
interest is the perpendicular anisotropy, that is the magnetization orientation
along the surface normal, in contrast to in-plane anisotropy, the magnetization
orientation within the surface plane. The energy gain of orientation of the easy

252
Figure 6.62. Band structures of Fe along the [110] direction. Top row: without spin
orbit coupling, bottom row: with spin-orbit coupling; left column: without exchange
splitting, right column: with exchange splitting. Spin-up and -down channels are shown
in blue and red, respectively.

axis within the surface plane is orders of magnitude less.

The magnetocrystalline anisotropy energy (MAE) is parameterized by con-
stants Ki , by expressing the magnetization direction in spherical coordinates
(ϑ, ϕ). For a tetragonally distorted system, the free-energy density reads

E(ϑ, ϕ) = − K2⊥ cos2 (ϑ) − K2k cos(2ϕ) sin2 (ϑ)

1 1 (6.137)
− K4⊥ cos4 (ϑ) − K4k [3 + cos(4ϕ)] sin4 (ϑ) + · · · .
2 8
Note that the Ki are not uniquely defined [150].

253
 Experimental access to the constants Ki is provided by ferromagnetic res-
onance (FMR). Another method is x-ray magnetic dichroism (XMCD) from
core levels (page 96) in which the orbital and spin moments can be obtained
via so-called sum rules. The orbital moments corresponds to the anisotropy
energy [151, 152]. In a bulk solid the orbital moments are quenched, due to the
almost spherical charge distribution, resulting in a small anisotropy energy. In
thin films, the larger orbital moment leads to an increased MAE.
With increasing film thickness d, the volume contribution KiV to Ki is con-
(surf) (if)
stant. The interface contributions, KiS of the surface and KiS of the film-
substrate interface, decrease with
(surf) (if)
KiS + KiS
Ki (d) = KiV + . (6.138)
d
For large film thicknesses, the magnetostatic dipole contribution exceeds, so
that the perpendicular anisotropy for thin films becomes in-plane at a critical
film thickness. That is the case for Fe and Co films grown on Cu(001). In con-
trast, Ni films on Cu(001) show a complementary behaviour. The are in-plane
magnetized for thin films (d < 7 sML), but display a transition to perpendicular
anisotropy For thicker films (d > 35 sML) the dipole contribution forces again
in-plane anisotropy. Each transition is called a spin reorientation transition
(SRT). In order to understand this unusual behaviour, one needs to compute
the total energy for Ni films on Cu(001). Typically these are based on the
magnetic force theorem.

Magnetic force theorem. For the determination of the MAE, the force theo-
rem is essential. It allows to compute the MAE in the framework of the local
spin-density approximation (LSDA) to density functional theory (DFT), see
Sect. 3.5. The force theorem states that the MCA is given by the difference of
the relativistic band energies (computed with SOC) computed for two orienta-
tions of the magnetization using the same scalar-relativistic potentials.
Starting point of the force theorem are potentials that are determined self-
consistently by a spin-polarized scalar-relativistic DFT calculation [153–155].
The associated charge and magnetization densities ρ0 (r) and m0 (r) give the
total energies

Esr [ρ0 , m0 ] = Tsr [ρ0 , m0 ] + U [ρ0 ] + Exc [ρ0 , |m0 |] + Eext [ρ0 ] (6.139)

in the framework of the LSDA. If spin-orbit coupling is added to the scalar-

relativistic Hamiltonian and performs a self-consistent calculation (now fully
relativistic), the total energy reads

E[ρ, m] = T [ρ, m] + U [ρ] + Exc [ρ, |m|] + Eext [ρ]. (6.140)

for the charge and magnetization densities ρ(r) and m(r). Thus, the energy
difference due to SOC is
∆E = E[ρ, m] − Esr [ρ0 , m0 ]
(6.141)
= (E[ρ, m] − E[ρ0 , m0 ]) + (E[ρ0 , m0 ] − Esr [ρ0 , m0 ]) .

254
It can be shown that the expression in the first bracket is of fourth order in the
SOC strength and therefore can be neglected with respect to terms of second
order. The remaining second bracket is just the difference of the band energies,
occ X
X
∆E ≈ E[ρ0 , m0 ] − Esr [ρ0 , m0 ] = [i (m, k) − sr
i (m, k)] , (6.142)
i k

computed for the scalar-relativistic self-consistent potential with [i (m, k)] and
without [sr
i (m, k)] SOC. The band contribution to the MAE for two orienta-
tions of the magnetization, M 1 and M 2 with magnetization densities m1 and
m2 , is then
occ X
X
Ebnd = [i (m1 , k) − i (m2 , k)] . (6.143)
i k

In contrast to calculations for the bulk MCA the Fermi energy must not be
corrected because it is fixed by the nonmagnetic substrate. In case of bulk Ni,
the change of the Fermi levels is about 10−6 eV and is hence in the range of
the MAE [153]. For thin Ni films, however, the band energy is three orders of
magnitude larger.
The second contribution to the MAE is the magnetostatic energy Edip of the
magnetization density m(r),
Z Z
1 1
Edip = 2 ×
c |r − r 0 |3
 (6.144)
[(r − r 0 ) · m(r)] [(r − r 0 ) · m(r 0 )]

0
× m(r) · m(r) − 3 drdr .
|r − r 0 |2

Performing a multi-pole expansion of m(r) results in

1 X 1
c2 |R + τ − τ 0 |3
Rτ τ 0
[(R + τ − τ 0 ) · mτ )] [(R + τ − τ 0 ) · mτ 0 )]
 
mτ · mτ 0 − 3
|R + τ − τ 0 |2

for the interatomic contribution of the moment for angular momentum l = 0.

Here, the sum is over all unit cells (R) and over all atoms within each unit
cell (τ and τ 0 ). Note that the above sum is conditionally convergent and is
solved using Ewald methods. Due to the long-ranged dipole-dipole interaction,
the MAE depends on the shape of the sample and is therefore called Shape
anisotropy. For very thin films, the magnetic moments are forced in-plane due
to the dipole contribution.
The main task of numerical computations of the MAE is the evaluation of
the band energy. The latter can be expressed in terms of the layer-resolved
density of states Nl (),
XZ EF
Ebnd (M ) =  Nl (; M ) d. (6.145)
l −∞

255
Here, the sum runs over all magnetic layers l.10 The energy integration is
conveniently performed in the complex energy plane as contour integration,
requiring only a small number of energy points [156].
Because the essential effect of SOC with respect to the MAE is the lifting of
degeneracies in tiny parts of the BZ, a fine k mesh is needed [157]. The layer-
resolved DOS is given by the layer-diagonal Green function Gll (layer index
l),
Z
1
Nl (; M ) = − lim Im Tr Gll (kk ,  + iη; M ) dkk . (6.146)
π η→0+ 2BZ

Superposition principle. In order to interpret the MAE calculations, it ap-

pears convenient to resolve the individual contributions to the band energy
with respect to layers. The resulting band energy profiles can be viewed as
sum over a surface contribution [Ebnd,L (M , l)] and an interface contribution
[Ebnd,R (M , l)] from which the volume contribution [Ebnd,B (M )] is subtracted
(Fig. 6.63). Explicitly,

Ebnd (M , l) ≈ Ebnd,L (M , l) + Ebnd,R (M , l) − Ebnd,B (M ). (6.147)

This approximation can be derived by Green functions for the constituting

subsystems in which SOC is treated as a perturbation [156], as is sketched now.
Unperturbed bulk Ni is described by a spin-dependent nonrelativistic Green
function GσB , σ = ± being the spin index. We construct a Ni film on a Cu
substrate by replacing an infinite Ni crystal [bulk (B)] by Cu (R) on the right-
hand side and by vacuum (L) on the left-hand side of the Ni layers which
constitute the film. The film layers are then influenced by spin-dependent
perturbations VLσ and VRσ , at the two boundaries. The film GF Gσ thus has to
fulfill the Dyson equation

Gσ = GσB + GσB (VLσ + VRσ )Gσ . (6.148)

Its series expansion can be written as

∞
X
Gσ = GσB [(VLσ + VRσ )GσB ]j (6.149a)
j=0

= GσB + δGσL + δGσR + δGσQW . (6.149b)

The second and third terms, δGσL and δGσR , describe perturbations due to scat-
tering solely on the left or right boundary, respectively, and thus comprise only
expressions with VLσ or VRσ , respectively. The last term, δGσQW , collects all
expressions with both VLσ and VRσ and thus results from multiple scattering be-
tween both interfaces, i. e., due to scattering in the quantum well (QW) formed
by the two interfaces. The latter may give rise to spin-polarized QW states in
the Ni film.
10
Note that the magnetic film induces magnetic moments in the otherwise nonmagnetic sub-
strate.

256
 If δGσQW is neglected, the properties of the film can be regarded as a super-
position of perturbations originating from the two independent interfaces. In
particular, the magnetization profile of the film M (l), l being the layer index,
can be regarded as a superposition of the magnetization profiles originating
from the Ni/Cu interface, M R (l), and from the Ni surface, M L (l),

M (l) ≈ M L (l) + M R (l) − M B , (6.150)

where M B is the magnetic moment of bulk Ni. Since the contribution from the
QW term usually gives rise only to small constant shifts of the magnetization
profile, the superposition provides a good estimate of the magnetization profiles
in films. By definition, quantum-size effects show up as the difference between
the superimposed and the true magnetization profile of the film.

Figure 6.63. Construction of a Ni film on Cu(001). Starting from the Ni bulk solid
(light grey; Bulk, B), two interfaces of the Ni film (Film) are produced by replacing
semi-infinite partial solids: by vacuum (white ) and by Cu (dark grey). The first gives
the surface system (Surface, L), the other the interface system (Interface, R). The
mathematical symbols ≈, +, and − refer to eq. (6.147).

The above approximative decomposition is not limited to magnetization pro-

files. Indeed, the response to additional perturbations can be estimated by a
superposition of independent interface perturbations as well. This way we are,
for example, able to analyze different contributions to the MAE in films, as will
be discussed in the following.
Regarding spin-orbit coupling as a perturbation of the Hamiltonian, the
Green function will depend on the magnetization direction, for example per-
pendicular (M (⊥ )) or parallel (M (k) ) to the interfaces. Further, the spin σ
is no longer a good quantum number. In matrix representation, the resulting
Green functions can be written as 2 × 2 block matrices where each block has
the size of the unperturbed GF matrix. We denote the corresponding SOC per-
turbation operator as Θ(M ), emphasizing its dependence on the direction of
the magnetic moments. The GF of the perturbed Ni film, G(M ), is implicitly
given by

G(M ) = G + G Θ(M ) G(M ) = G + δG(M ). (6.151)

257
The Green function G is diagonal in spin space (no spin-orbit coupling) and
constructed from G± of the film with G±
QW being neglected. Here, δG(M ) is
given by the Dyson series

δG(M ) = G Θ(M ) G + G Θ(M ) GΘ(M ) G + · · · (6.152a)

=GB Θ(M ) GB + GB Θ(M ) GB Θ(M ) GB + · · ·
+δGL Θ(M ) δGL + δGL Θ(M ) δGL Θ(M ) δGL + · · · (6.152b)
+δGR Θ(M ) δGR + δGR Θ(M ) δGR Θ(M ) δGR + · · · .

Eventually, we obtain the grouping of terms

δG(M ) = δGB (M ) + δGL (M ) + δGR (M ) + δGQW (M ), (6.153)

where δGL (M ) and δGR (M ) include contributions solely from scattering at the
left-hand and right-hand side perturbations, respectively. The term δGB (M )
represents the spin-orbit perturbation of bulk Ni, and δGQW (M ) includes mul-
tiple scattering at both interfaces. If we neglect the latter, the band energy can
be expressed as

Ebnd (M , l) ≈ Ebnd,L (M , l) + Ebnd,R (M , l) − Ebnd,B (M ), (6.154)

with
Z EF
1
Ebnd,B (M ) = − Im Tr E δGB (M ) dE (6.155)
π −∞

and
Z EF
1
Ebnd,B (M ) = − Im Tr E [δGB (M ) + δGD (M )] dE, (6.156)
π −∞

where D = L, R. This is simply a superposition of three band-energy profiles

which result from two independent interfaces on the left- and right-hand sides,
Ebnd,L (l) and Ebnd,R (l), in conjunction with the subtraction of the bulk contri-
bution Ebnd,B . It provides a picture derived from individual contributions from
different kinds of interfaces and surfaces. Again, finite-size effects can be ob-
tained from the difference between the superimposed and the true band energy
for the thin film. Only a small number of constituents has to be regarded in
order to understand the magnetic anisotropy of layered structures. Thus, the
MAE of thin films may be regarded as the sum of a surface, an interface part
and a bulk contribution if multiple-scattering effects are neglected. However, all
three contributions will be of importance and may give rise to a broad spectrum
of possible effects on the MAE of thin films.

Magnetic moments at an interface. Layered resolved magnetic interface mo-

ments are hard to determine experimentally. A few measurements have been
done by means of the element specific XMCD for atoms in ultrathin overlay-
ers, which may give information on the interface magnetism of buried layers.

258
Also measurements of the total magnetization of layered systems can be used
in order to estimate magnetic effects of the interfaces.
An outermost interface atom A of an A/B interface is in its nearest-
neighboring shell surrounded by nA of A atoms and nB number of B atoms.
We now define an averaged interface moment mifAB (x) as

mifAB (x) = xmifA (x) + (1 − x)mifB (x), (6.157)

where
nA
x= (6.158)
nA + nB

and mifA and mifB are the magnetic moments of the outermost A and B interface
atoms, respectively. This averaged interface moment can be interpreted as the
magnetic moment of a highly correlated (ordered) bulk alloy if we neglect the
fact that the outermost B atom at the interface actually has nA of B atoms
in its nearest-neighboring shell. Thus, if more long-range ordering effects on
the individual magnetic moments in the alloy are small, there should be a good
agreement between the averaged interface moment and the magnetization of the
corresponding bulk alloys. In Figure 6.64, the averaged interface magnetization
mifAB (x) for the interfaces is shown as a function of the average valence charge

Ñ = xNA + (1 − x)NB (6.159)

where NA and NB are the number of valence electrons of the constituent ma-
terials, i. e., a kind of Slater-Pauling curve for the interface magnetization.

The similarity between this graph and the well-known Slater-Pauling curve
for the 3d alloys clearly elucidate the relation between the magnetic interface
moments and the corresponding bulk alloys. The dashed lines display some of
the experimental data for the magnetic moments for bulk alloys. If the magnetic
moment of the individual atomic types does not change in the alloy or at the
interface as a function of mixing concentration, or interface orientation, the
magnetization curves in the Slater-Pauling graph will be straight lines between
the magnetic moments of the constituent pure bulk materials. Deviations from
such a linear interpolation therefore indicate changes of the magnetic moments
compared to the bulk for the individual atomic types in the alloy and at the
interface.
For example, the increased interface moment of Fe at the Fe/Co interfaces
can be explained in terms of the increased total magnetization of the Fex Co1−x
alloy for small concentrations of Co in Fe. This increase is mainly due to
an increased Fe moment since the Co moment is saturated, in full agreement
with what is found at the interface. Furthermore, the linear behavior of the
magnetization curve between Co and Ni for the experimental Cox Ni1−x alloy
indicates that there is no large deviation of the interface magnetization in the
Co/Ni interface compared to bulk Co and Ni, which is in complete agreement
with the calculated magnetic interface moments.

259
Figure 6.64. The averaged interface moment mifAB (Ñ ) as a function of valence charge
Ñ for the different magnetic interfaces (Slater-Pauling curve). The dashed lines show
the experimental moments for FeV, FeCo, CoNi, NiCu, and NiCr alloys. The dot-
dashed line shows the moment for the NiV alloy.

In the same way we can understand, for example, the decreased magnetic
interface moments of Ni at the Ni/V, Ni/Cr, and Ni/Cu interfaces. Thus, devi-
ations from a linear behavior of the Slater-Pauling curve between two materials
for a binary bulk alloy indicate that the corresponding interface moments are
different from bulk. If the Slater-Pauling curve is below the linear interpolation
at least one of the interface moments is less than in bulk and opposite if it is
higher. A more quantitative estimate of the magnetic interface moments might
be achieved if the atomic projected magnetic moments mA (x) and mB (1 − x)
of an Ax B1−x alloy are known through the relations

mifA = mA (x), (6.160a)

mifB = mB (1 − x). (6.160b)

where
nA
x= (6.161)
nA + nB
and nA , nB are the number of A and B atoms in the first coordination shell
around an outermost A interface atom. The reason for this is that we now
compare atomic projected moments in the bulk alloy and at the interface with
the same number of nearest neighbors and one can expect that this leads to a
better quantitative estimate of the magnetic interface moments.

Magnetic structure. Before discussing the MAE of Ni/Cu(001), the magnetic

structure of these films is briefly discussed. As is well known, an increase in
volume, as well as a reduction of the coordination number, usually favors an

260
increase of the magnetic moment. Thus, the bulk spin magnetic moment of
Ni (0.67 µB, in the fcc case with Cu lattice constant) is slightly enhanced with
respect to that of bulk Ni (0.62 µB, with Ni lattice constant).
In Figure 6.65, the spin magnetization profiles of fcc Ni films on Cu(001) are
depicted for coverages from 1 ML up to 10 ML. At 1 ML coverage the magnetic
moment is only 44 % (0.296 µB) of the Ni bulk value. For 2 ML the magnetic
moments are increased to 0.741 µB and 0.499 µB for layers S and S-1, respec-
tively. This increase is continued for the 3-ML film. The surface layer of the
latter has the largest moment of all systems considered here (0.762 µB). At
4 ML coverage the building up of a plateau region starts: the interior Ni layers
start to show nearly the moments of bulk Ni, while that of the surface layer
is enhanced and those at the interface are reduced by about 30 % for the first
interface Ni layer (adjacent to the first Cu substrate layer, S-n+1) and 3 %
for the second interface Ni layer (S-n+2). This evolution of the magnetization
profile is completed at about 7 ML coverage.
The surface magnetic structure of films with n > 7 shows a nearly complete
agreement with that of semi-infinite Ni(001). The magnetic moment of the
outermost Ni layer (S, 0.746 µB for n = ∞) compared to those of the interior
Ni layers of the film and that of bulk Ni with Cu lattice constant is enhanced
by about 12 %. This enhancement of moments is typical for surface layers and
is due to the reduction in coordination number.
The reduced Ni magnetic moment at the interface can be understood from the
behavior of binary bulk alloys Nix Cu1−x when the concentration x is adjusted
to correspond to the Ni-Ni coordination number at the fcc(001) interface. The
magnetic properties of binary 3d transition-metal alloys are often essentially
determined by the nearest-neighbor correlation, i. e., the local environment of
a particular atom. The Ni atoms at the Ni/Cu interface can be regarded as
constituents of a ‘perfectly ordered alloy’ where only 8 of 12 nearest neighbors
surrounding an interface Ni site are of the same kind. This corresponds to a local
Ni concentration of x = 2/3. According to the Slater-Pauling curve, there is a
substantial reduction of the nickel magnetic moment to approximately 0.3 µB
in the NiCu alloy.
The Cu layers next to the interface show induced magnetic moments of about
0.015 µB with an antiferromagnetic coupling to Ni. For example, this common
antiparallel spin behavior has been observed for Fe films on Au(001).
The small oscillations of the size of the magnetic moments in the interior of
thicker Ni films (n > 6) can be regarded as ‘magnetic’ Friedel oscillations due
to the spin-dependent perturbations from the surface and the Ni/Cu interface
[they were also found for Co films on Cu(001)]. In Figure 6.65, both the surface
(n = ∞) and the interface system (n = if) show oscillations which decay rapidly.
We now turn to the picture of the magnetization as a superposition of mag-
netic profiles. The superimposed profiles are compared with their exactly calcu-
lated counterparts. Except for the 1-ML film, where the superposition is appar-
ently questionable, one finds excellent agreement which means that quantum-
well contributions to the spin magnetic moment are generally very small. The
largest deviations are found for 2 ML and 4 ML but are smaller than 0.03 µB.
As mentioned above, the 3-ML film shows the largest surface magnetic moment.

261
Figure 6.65. Layer-resolved spin magnetic moments of fcc Ni films on Cu(001)
(squares, solid lines) with thicknesses n = 1, . . . , 10 as obtained by LMTO calcula-
tions. The moments of a Ni(001)/Cu(001) interface (n = if), diamonds; interface at
layers S-10 (Ni) and S-11 (Cu!) and semi-infinite Ni(001) (n = ∞, open circles)—both
with Cu lattice constant—are shown in addition. Magnetic moments obtained by su-
perposition of the surface and the interface profile are indicated by solid circles and
dotted lines. For clarity, the data sets are offset; their respective zeros are represented
by dash-dotted lines. Layers are labeled S, S-1, S-2, . . . , starting with the surface layer
S. Solid and dotted lines serve as guide to the eye.

This can easily be explained within the superposition picture: The moment of
the surface layer S is obtained from the surface layer S of the surface system
and the third Ni layer of the interface system. Both show the largest moments
of their systems and, therefore, give rise to the enhanced surface moment of
the 3-ML film. In conclusion, one can construct layer-dependent spin magnetic
profiles of films from the surface and the interface profiles with high accuracy,
the only exception being the 1-ML case.
Summarizing, the magnetization profiles exhibit no spectacular features but

262
rather the expected behavior: enhancement at the surface, reduction at the
interface, oscillations in between. Relevant for the discussion of the anisotropy,
in particular the band energy, may be the building up of a plateau, i. e., the
appearance of bulk-like layers, which starts at
unit[6]ML.

Spin reorientation transition. A mentioned above, Ni films on Cu(001) show

a peculiar behaviour with respect to the MCA. Between 3 ML and 5 ML, de-
pending on sample preparation and temperature, the anisotropy switches from
in-plane to perpendicular orientation (SRT) At larger thicknesses (35 ML to
70 ML) the transition back to in-plane anisotropy is found. Early theoretical
investigations in tetragonally distorted Ni bulk gave hints that this findings is
due to the distorted lattice structure [158]. However, bulk calculations were not
able to make statements on the critical film thickness at which the SRT takes
place. These are only possible by calculations of the MAE for semi-infinite
Ni/Cu(001) systems in which both the Ni film thickness and the distortion are
accounted for [156]. The second gradual SRT is explained by island growth
because the Ni lattice structure is not ‘pinned’ by the Cu substrate for thick
films. Because this phenomenon is hardly accessible to DFT calculations, it
was focused on the first and more important SRT [159].
The spin reorientation transition can be discussed by means of band energy
profiles for tetragonally distorted Ni films (case ‘fct’).11 Here, the interlayer
spacings of all Ni layers are reduced by 6.9 % with respect to that of bulk Cu,
in order to reproduce the atomic volume in bulk Ni. The layer-resolved band
energy differences
(l) (l) (l)
∆Ebnd = Ebnd (M (k) ) − Ebnd (M (⊥) ) (6.162)

are shown in Fig. 6.67. The magnetization orientations M (k) and M (⊥) are
aligned along the crystallographic [100]- and the [001] direction, respectively.
In particular for thicker films, the surface contribution (‘∞’ in Fig. 6.67) and
the contribution from the Ni-Cu interface (‘if’) can easily be recognized. In
comparison with the unrelaxed Ni film (case ‘fcc’, not shown) the band energy
differences are shifted to positive values, thus favoring perpendicular anisotropy.
Note that in bulk Ni this effect vanishes due to the higher symmetry.
The volume contribution of the uniaxial anisotropy is KV = 0.081 meV per
atom and agrees well with experimental data. The surface and interface contri-
butions are also shifted to positive values. The former is negative at the surface
layer S in the fcc case, hence favoring in-plane anisotropy.
The band energy differences obtained by the superposition principle agree in
general well with those from the DFT-LSDA calculations. Only for very thin
films, n ≤ 3 considerable deviations show up that can be traced back to missing
finite-size effects in the superposition profiles.
The profile for the interface case shows prominent oscillations that proceed
far into the Ni bulk (cf. in this context also page 232 ). From this it becomes

11
fct = face-centered-tetragonal.

263
Figure 6.66. Experimental proof by XPEEM of the SRT in wedge-shaped
Cu/Ni/Cu(001).

evident that approaches that restrict the surface and the interface contribution
to a small finite number of layers are not sufficient to deal with the problem.
As is evident from Fig. 6.68 the magnetostatic dipole contribution (circles in
Fig. 6.68) is of minor importance for the SRT. The dipole energy differences
(l) (l) (l)
∆Edd = Edip (M (k) ) − Edip (M (⊥) ) (6.163)
are mush smaller (in absolute values) than the band energy contributions. The
spin reorientation transition can thus be explained solely by the changes in

264
Figure 6.67. Band energy contributions to the MAE of fct Ni/Cu(001) with thick-
nesses n = 1 ML to 10 ML. The profile for semi-infinite Ni is denoted ‘∞’, that of a
Ni–Cu interface by ‘if’. Results from ab initio calculations are represented by squares
while those obtained from the superposition principle are given by filled circles. Atomic
layers are indicated by S, S−1, S−2, . . . , with S being the outermost Ni surface layer.

the band energies due to the lattice distortion. Indeed, there is a change of
sign for the fct case between 5 ML and 6 ML, that is a transition from in-
plane to perpendicular anisotropy, as is found in experiment. In contrast, the
fcc case (undistorted films) shows in-plane anisotropy even for thicker films
(∆Ebnd < 0).
The result for the distorted case is supported by calculations for 3 ML and
5 ML (‘fct LEED’), in which the Ni interlayer distances were adopted from
LEED I(V ) analyses (Sect. 4.1).
The calculations for very thin films (1 ML and 2 ML thick) show perpen-

265
Figure 6.68. Magnetic-anisotropy energies per atom of Ni films on Cu(001). Their
band-energy contributions ∆Ebnd are represented by squares, their dipole-dipole in-
teraction energy Edd by open circles. Diamonds denote band-energy contributions
obtained by superposition of the surface and the interface profiles. Solid lines refer
to cubic Ni films (fcc case), dotted lines to homogeneously tetragonal distorted Ni
films (fct case). Solid circles show band-energy data obtained for in-homogeneously
tetragonal distorted Ni films of 3 ML and 5 ML thickness (fct LEED case). Lines serve
as guide to the eye. Positive values imply perpendicular anisotropy, negative values
in-plane anisotropy.

dicular anisotropy that is not found in experiment. Possible reasons for this
disagreement are that the calculations are for zero temperature. Note that the
Curie temperature for such thin films is very small and, hence, the local mag-
netic moments could fluctuate strongly and reduce the MAE. Further, there
are both experimental and theoretical indications that Ni does not grow on-top
of Cu but diffuses below the Cu surface layer, resulting in a surface alloy and
thus a reduced MAE.
As shown in Fig. 6.67, the superposition principle allows for a qualitative
description of the band energy profiles. However, it does not allow to predict
the SRT, as is evident from Fig. 6.68. The deviations between the superposition
and the first-principles results are significant, in particular for the fct case in
which the super position results give no SRT (∆Ebnd > 0 for n ≥ 2). This
implies that finite-size effects are important for the SRT in Ni/Cu(001).
Further calculations for various polar angles ϑ of the magnetization direction
show that the important contributions to the MAE are due to K2⊥ [eq. (6.137)].
Higher-order contributions are not significant.

Concluding remarks. Film geometry, i. e., tetragonal distortion, and magnetic

properties, in particular the anisotropy, of ultrathin Ni layers on Cu(001), are
closely related.

266
Figure 6.69. Higher-order contribution to the band-energy difference δEbnd of 7 ML
fct Ni on Cu(001). ϑ is the polar tilt angle of the magnetic moments. Circles represent
results of the numerical calculations. The dotted line is a fit taking into account
only the first-order contribution; the solid line is a fit for both first- and second-order
contributions.

Thick fct films, with thicknesses larger than five monolayers, show an in-
crease of the volume contribution to the magnetic-anisotropy energy which
overrides the dipole-dipole interaction contribution. Thus, the experimentally
observed spin reorientation transition from in-plane to perpendicular anisotropy
at around six monolayers Ni can be attributed to this volume contribution. The
surface anisotropy, i.,e., perpendicular anisotropy at the outermost two to three
layers, is too small compared to the volume contribution and thus cannot ac-
count for the observed transition. Further, the interface contribution favors an
in-plane anisotropy which partially cancels the surface anisotropy.

6.5 Elements of solid-state magnetism

In this section, the basics of solid-state magnetism will be briefly discussed.

6.5.1 Hund’s rules

In atomic physics, Hund’s rules, refer to a simple set of rules used to determine
which is the term symbol that corresponds to the ground state of a multi-
electron atom. They are named in honour of Friedrich Hund who contributed
Hund’s Rule, rule two as listed here.
The four rules are:

1. Full shells and subshells do not contribute to total S, the total spin an-
gular momentum and L, the total orbital angular momentum quantum
numbers.

267
 It can be shown that for full shells and subshells both the residual elec-
trostatic term (repulsion between electrons) and the spin-orbit interaction
can only shift all the energy levels together. Thus when determining the
ordering of energy levels in general only the outer valence electrons need
to be considered, for example, the d electrons of transition metals.

2. The term with maximum multiplicity (maximum S) has the lowest energy
level.
Due to the Pauli exclusion principle, two electrons cannot share the same
set of quantum numbers within the same system. Therefore, there is
‘room’ for only two electrons in each spatial orbital. One of these elec-
trons must have (for some chosen direction z), SZ = 1/2, and the other
must have SZ = −1/2. Hund’s second rule states that the lowest energy
atomic state is the one which maximizes the sum of the S values for all
of the electrons in the system, maximizing the number of unpaired elec-
trons. This is usually the lowest energy atomic state because it forces
the unpaired electrons to reside in different spatial orbitals. Resulting in
a larger average distance between the two electrons, reducing electron-
electron repulsion energy.

3. For a given multiplicity, the term with the largest value of L has the lowest
energy.
This rule deals again with reducing the repulsion between electrons. It can
be understood from the classical picture that if all electrons are orbiting
in the same direction (higher orbital angular momentum) they meet less
often than if some of them orbit in opposite directions. In that last
case the repulsive force increases, which separates electrons. This adds
potential energy to them, so their energy level is higher.

4. For atoms with less than half-filled shells, the level with the lowest value
of J lies lowest in energy. Otherwise, if the outermost shell is more than
half-filled the term with highest value of J is the one with the lowest
energy.
This rule considers the energy shifts due to spin-orbit coupling. In the
case where the spin-orbit coupling is weak compared to the residual elec-
trostatic, where L and S are still good quantum numbers the splitting is
given by

∆E = ζ(L, S)L · S
(6.164)
= ζ(L, S)[J(J + 1) − L(L + 1) − S(S + 1)].

The value of ζ(L, S) changes from plus to minus for shells greater than
half full. The second term gives the dependence of the ground state on
the magnitude of J.
Problem 93. Show that the square of any angular momentum operator l̂ has eigenvalues
l(l + 1).

268
 We define
ˆl± ≡ ˆlx ± iˆly .

These obey

[ˆlz , ˆl± ] = ±ˆl± ,

[ˆl+ , ˆl− ] = 2ˆlz .

Further,

ˆl2 = ˆl2 + ˆl2 + ˆl2 = 1 ˆl+ ˆl− + ˆl− ˆl+ + ˆl2 .

 
x y z z
2
ˆl+ , ˆl− , and ˆlz commute with ˆl2 .
Simultaneous eigenstates of ˆl2 and ˆlz fulfill
ˆl2 |a, bi = a|a, bi and ˆlz |a, bi = b|a, bi.

Thus, from the commutation rules it follows

ˆl2 ˆl± |a, bi = aˆl± |a, bi

and
ˆlz ˆl± |a, bi = ˆl± ˆlz |a, bi ± ˆl± |a, bi = (b ± 1)ˆl± |a, bi.

Thus
ˆl± |a, bi ∝ |a, b ± 1i.

From the commutation rules follows further that

ˆl∓ ˆl± = ˆl2 − ˆlz (ˆlz + 1),

hence

ha, b|ˆl∓ ˆl± |a, bi = [a − b(b ± 1)]ha, b|a, bi.

Therefore, ˆl+ |a, bmax i = 0 and ˆl− |a, bmin i = 0 because otherwise the norm of [a − b(b ±
1)]ha, b|a, bi could be imaginary. This leads to a = bmax (bmax + 1) = bmin (bmin − 1), from
which follows bmax = −bmin . Denoting the eigenvalue of ˆlz as l, a = l(l + 1). 2
The rules deal in a simple way the usual energy interactions dictate the
groundstate term. The rules assume that the repulsion between the outer elec-
trons is very much greater than the spin-orbit interaction which is in turn
stronger than any other remaining interactions. This is referred to as the LS
coupling regime.
Hund’s rules are often used to order the excited levels. This is a common
misapplication of the rules and is generally incorrect. Hund’s rules presume LS
coupling and presume that the electrons can be considered to be in a unique
configuration. Neither is always true. For heavier elements, the j–j coupling
scheme often gives better agreement with experiment.

269
Application to Fe. Let us consider two configurations of an Fe atom. In
configurations A and B we have
(
2 6 (A) 4s ↑↓ 3d ↑↑↑↓↓↓ S = 0 Mspin = 0 µBohr
Fe4s 3d . (6.165)
(B) 4s ↑↓ 3d ↑↑↑↑↑↓ S = 2 Mspin = 4 µBohr

Hund’s first rule says, S should be maximum. This is due to a maximum

antisymmetric orbital part of the wave function which minimizes the Coulomb
energy. Hund’s second rule states that L should be maximum (“avoiding” of
electrons”) and again the Coulomb energy is minimized. And Hund’s third rule
gives that for the total angular momentum J holds

J = |L − S| n ≤ 5 λ > 0, (6.166a)
J =L+S n>5 λ < 0. (6.166b)

This minimizes the spin-orbit coupling energy λL · S.

The momenta of the 3d elements are summarized in Fig. 6.70,
using LS coupling. This results in the following spin moments,

Figure 6.70. Orbital, total and spin momentum of the 3d elements.

Sc Ti V Cr Mn Fe Co Ni Cu
config s2 d1 s2 d2 s2 d3 s1 d5 s2 d5 s2 d6 s2 d7 s2 d8 s1 d10
µspin 1 2 3 6 5 4 3 2 1
Problem 94. Find out what magnetic atoms are nonmagnetic in the bulk.
Cr and Mn are antiferromagnets, Fe, Co, and Ni are ferromagnets. The other elements
are nonmagnetic. 2

6.5.2 Landau diamagnetism

Diamagnetism is a form of magnetism which is only exhibited by a substance
in the presence of an externally applied magnetic field. It is the result of

270
changes in the orbital motion of electrons due to the application of an externally
applied magnetic field. Applying a magnetic field creates a magnetic force on
a moving electron (Lorentz force). This force changes the centripetal force on
the electron, causing it to either speed up or slow down in its orbital motion.
This changed electron speed modifies the magnetic moment of the orbital in a
direction against the external field.

Figure 6.71. A small (about 6 mm) piece of pyrolytic graphite levitating over a
permanent neodymium magnet array (5 mm cubes on a piece of steel). Note that the
poles of the magnets are aligned vertically and alternate (two with north facing up,
and two with south facing up, diagonally).

For an electron in a magnetic field the Hamiltonian reads

1
Ĥ = (p̂ + eA)2 + µB B. (6.167)
2m
In the following we will inly consider the first term (cf. Sect. 5.3.1).
The magnetic field is chosen along the z-axis, B = (0, 0, B). The Coulomb
gauge, div A = 0 gives the vector potential A = (0, Bx, 0). The Hamiltonian
then reads
1  2
p̂x + (p̂2y + eBx)2 + p̂2z .

Ĥ = (6.168)
2m

271
Because it is separable (sum of independent terms), we make the ansatz

Ψ(r) = eikz z eiky y ψx (x). (6.169)

The Schrödinger equation for ψx then reads

~2 ∂ 2 ~2 kz2
 
1 2
− + (~ky + eBx) + ψx (x) = Eψx (x). (6.170)
2m ∂x2 2m 2m
Substituting
eB
ωc ≡ (6.171a)
m
~ky
x0 = , (6.171b)
mωc
with the cyclotron frequency ωc , we arrive at
~2 ∂ 2 ~2 kz2
 
1 2 2
− + ω (x − x0 ) + ψx (x) = Eψx (x). (6.172)
2m ∂x2 2m c 2m
This is a superposition of a harmonic oscillator and free motion, the eigenener-
gies are
~2 kz2
 
1
En = n + ~ωc + . (6.173)
2 2m
The results is the magnetic field allows for free motion along the filed direction
and quantizes the energy that is associated with motion perpendicular to the
field (Fig. 6.72).

Figure 6.72. Without magnetic field, the electronic states are discrete in reciprocal
space (separated by 2π/L, left). In a magnetic field the wavevectors ‘condensate’ onto
Landau tubes (right).

For B = 0 the energy of the free electrons must be equal to the of the Landau
tube,

~2 kx2 ~2 ky2 ~2 kz2 ~2 kz2

 
1
+ + = n+ ~ωc + . (6.174)
2m 2m 2m 2 2m

272
Thus,

~2 kx2 ~2 ky2
 
1
+ = n+ ~ωc . (6.175)
2m 2m 2

Figure 6.73. Landau tubes in reciprocal space.

The area between two levels is independent of the level, n,

2mπωc
π(kx2 + ky2 )n+1 − π(kx2 + ky2 )n = . (6.176)
~
Hence, the degeneracy is
2mπωc
~ mL2 ωc
g= (2π)2
= . (6.177)
2π~
L2

Landau susceptibility. The external magnetic field results in a series of quan-

tized Landau levels. The energy is increased upon further increasing B, which
is accompanied by the presence of a magnetization. This is Landau diamag-
netism.
For different values of kz there a different highest Landau levels. We define
~2 kz2
E⊥ ≡ EF − . (6.178)
2m

273
Now we are going to calculate the mean energy of the free electron gas with and
without magnetization. To do this we consider first the motion in the xy-plane.
Without field, the electronic states are filled up to the highest E⊥ for each kz .
With field this is no longer possible due to the Landau quantization. We denote
the highest Landau level by En . Define

x ≡ E⊥ − En , (6.179)

see Fig. 6.74.

Figure 6.74. Without magnetic field, all shaded ares are filled up to E⊥ (left). With
field, right, only discrete levels are present.

The occupation of the highest block is

 
g 1 g 2x
N= [E⊥ − (En − ~ωc )] = 1+ . (6.180)
~ωc 2 2 ~ωc

Hence, the energy of the highest block is

1 1
U (B = 0; x) = N [E⊥ + (En − ~ωc )]. (6.181)
2 2
With En = E⊥ − x,
  
1 2x x ~ωc
U (B = 0; x) = 1+ E⊥ − − . (6.182)
2 ~ωc 2 4

Averaging over the block yields

Z +~ωc /2  
1 g ~ωc
hU (B = 0)i = U (B = 0; x) dx = E⊥ − . (6.183)
~ωc −~ωc /2 2 3

Problem 95. Derive the last equation.

274
 Only terms even in x contribute to the integral. These are

E⊥ ~ωc x2
− −
2 8 2~ωc
Integration then yields

(~ωc )2
 
1 ~ωc
hU (B = 0)i = E⊥ − .
~ωc 2 6

Taking into account the degeneracy, gives the desired result. 2

With magnetic field, the ‘condensation’ of the topmost block gives

U (B > 0; x) = N En . (6.184)

In analogy to the above it follows

 
g ~ωc
hU (B > 0)i = E⊥ − . (6.185)
2 6

Thus, the mean energy difference is

~ωc
h∆U i = hU (B > 0)i − hU (B = 0)i = g . (6.186)
12
Eventually, we have to account for the spin degeneracy and the number of states
in kz . This results in
kF kF L
∆U (B) = 2 2π h∆U i = g ~ωc . (6.187)
L
12π

The energy difference is positive and, hence, we are concerned with diamag-
netism.
The susceptibility

µ0 ∂ 2
χ=− ∆U (B) (6.188)
V ∂B 2
results in
µ0 e2 kF
χ=− . (6.189)
12π 2 m
Problem 96. Derive the last equation.
From eq. (6.187) follows

kF L mL2 ωc kF L eB kF e2 B 2 kF e2 B 2
∆U (B) = g ~ωc = ~ = L3 = V
12π 2π~ 12π m 24π 2 m 24π 2 m
with V ≡ L3 . Thus,

µ0 ∂ 2 kF e2 B 2 kF e2
 
χ=− V = −µ0 .
V ∂B 2 24π 2 m 12π 2 m
2

275
6.5.3 Pauli paramagnetism
Paramagnetic materials are attracted when subjected to an applied magnetic
field. The alignment of the atomic dipoles with the magnetic field tends to
strengthen it and is described by a relative magnetic permeability greater than
unity (or, equivalently, a small positive magnetic susceptibility).
Paramagnetism requires that the atoms individually have permanent dipole
moments even without an applied field, which typically implies an unpaired
electron in the atomic or molecular orbitals. In pure paramagnetism, these
atomic dipoles do not interact with one another and are randomly oriented in
the absence of an external field, resulting in zero net moment. If they do in-
teract, they can spontaneously align or anti-align, resulting in ferromagnetism
(permanent magnets) or antiferromagnetism, respectively. Paramagnetic be-
haviour can also be observed in ferromagnetic materials that are above their
Curie temperature, and in antiferromagnets above their Nel temperature.
In atoms with no permanent dipole moment, e. g. for filled electron shells, a
weak dipole moment can be induced in a direction anti-parallel to an applied
field (diamagnetism). Paramagnetic materials also exhibit diamagnetism, but
the latter effect is typically orders of magnitude weaker. Paramagnetic materials
in magnetic fields will act like magnets but when the field is removed, thermal
motion will quickly disrupt the magnetic alignment. In general paramagnetic
effects are small.
Both spin states are split by the magnetic field (Zeeman splitting), that is

~2 k 2
E= ± µB B. (6.190)
2m
Hence, the spin-resolved density of states reads
   3/2
g(E) V 2m p
= E ± µB B. (6.191)
2 ↓,↑ 4π 2 ~2

Both spin channels share a common Fermi level, implying that there are
more electrons with spin parallel to the field than with opposite orientation
(Fig. 6.75). The energy difference is 2µB B, which is a small quantity compared
to the Fermi level.12
For T = 0 K, the difference of the electron densities is

2m 3/2 EF +µB B √
  Z
V
∆n ≡ n↓ − n↑ = 2 E dE (6.192a)
4π ~2 EF −µB B

2m 3/2
 
V
≈ 2 EF 2µB B (6.192b)
4π ~2
= g(EF ) µB B. (6.192c)

The magnetization M is given by

M ≡ µB ∆n = g(EF ) µ2B B. (6.193)

12
µB = 5 · 10−5 eV/T.

276
Figure 6.75. A magnetic field B shifts the density of states g(E) by ±µB B depending
on the spin orientation.

The susceptibility then reads

∂M
χ = µ0 = µ0 g(EF ) µ2B . (6.194)
∂B
Note that the above susceptibility is independent of temperature because we
set explicitly T = 0 K. Hence, no Curie-Weiss law. Further note that Pauli
paramagnetism is small because only electrons at the Fermi level contribute to
the magnetization.

Comparison of Landau diamagnetism and Pauli paramagnetism. In order

to compare the Landau susceptibility, eq. (6.189), with the Pauli susceptibility,
eq. (6.194), we write
2
µ0 e2 kF

1 e~ mkF 1
χLandau =− 2
= − µ0 2 2
= − χPauli . (6.195)
12π m 3 2m π ~ 3
From this it would follow that all metals are paramagnetic because Pauli para-
magnetism is three times stronger than Landau diamagnetism. Experiments
contradict this statement. For example, Pt is a good paramagnet whereas Cu,

277
Ag, and Bi are diamagnets. The reason for this disagreement is the approxi-
mation of the band structure by that of free electrons.
A simple cure for the above problem is to introduce the effective mass. A
careful recalculations including the effective mass m? gives
1  m 2
χLandau = − χPauli . (6.196)
3 m?
Thus, the total susceptibility of a metal can be expressed as
 
1  m 2
χmetal = 1 − χPauli . (6.197)
3 m?

If the effective mass is small enough, a metal can be diamagnetic. Bismuth for
example, with m? = 10−2 m, is a strong diamagnet.

6.5.4 Band ferromagnetism

Ferromagnetism is the ‘normal’ form of magnetism which most people are fa-
miliar with, as exhibited in horseshoe magnets and refrigerator magnets, for
instance. It is responsible for most of the magnetic behavior encountered in ev-
eryday life. All permanent magnets are either ferromagnetic or ferrimagnetic,
as are the metals that are noticeably attracted to them.
Historically, the term ‘ferromagnet’ was used for any material that could
exhibit spontaneous magnetization: a net magnetic moment in the absence
of an external magnetic field. This general definition is still in common use.
More recently, however, different classes of spontaneous magnetization have
been identified when there is more than one magnetic ion per primitive cell of
the material, leading to a stricter definition of ‘ferromagnetism’ that is often
used to distinguish it from ferrimagnetism. In particular, a material is ‘ferro-
magnetic’ in this narrower sense only if all of its magnetic ions add a positive
contribution to the net magnetization. If some of the magnetic ions subtract
from the net magnetization (if they are partially anti-aligned), then the mate-
rial is ‘ferrimagnetic’. If the ions anti-align completely so as to have zero net
magnetization, despite the magnetic ordering, then it is an antiferromagnet. All
of these alignment effects only occur at temperatures below a certain critical
temperature, called the Curie temperature (for ferromagnets and ferrimagnets)
or the Néel temperature (for antiferromagnets).
Bulk Fe has a magnetic moment of about 2.2 µB per atom. This cannot
be understood within the picture of localized magnetic moments (atomic mo-
ments). Hence, as we will see, we are concerned with band magnetism or
itinerant magnetism that is due to spontaneous splitting of the band in the
solid (Fig. 6.76).
Figure 6.76 shows the spin-resolved density of states of the strong ferromagnet
Co. At the Fermi level, there are more charge carriers of one spin orientation.
It appears that the bands are shifted with respect to each other (‘rigid band be-
haviour’). The d-bands are almost completely occupied, so mainly sp electrons
are at EF .

278
Figure 6.76. Spin-resolved density of states of Fe, Co, and Ni (green: spin-up, red:
spin-down).

As we know from Section 3.2, an electron can be described as moving in the

mean field that is produced by all the other electrons (‘molecular field’). The
strength of this field is given by the parameter λ. Due to Pauli paramagnetism,
the molecular field results in polarization and magnetization of the electron
system that further reinforces the molecular field (‘positive feedback’).

Stoner criterion. In order to understand the effect one asks for the energetic
preference of being in a ferromagnetic state. The number of electrons that can
contribute to transport is g(EF ) δE/2. The electrons are shifted by the energy

279
δE (Fig. 6.77). Hence, the kinetic energy is increased by
δE
∆Ekin = g(EF ) δE. (6.198)
2
This is an energy cost.

Figure 6.77. Spontaneous splitting of the spin-up and spin-down energy bands with-
out external magnetic field.

However, also energy is gained by the magnetization of the molecular field.

Assuming that each electron contributes µB to the magnetization, the spin-
dependent densities of states reads
1
n↓,↑ = n [1 ± g(EF ) δE] . (6.199)
2
The magnetization is obtained by the density difference,

M = µB (n↓ − n↑ ) = µB g(EF ) δE. (6.200)

To obtain the change in potential energy, one has to integrate over the magne-
tization because the derivative of the free energy ∂F/∂B is the magnetization,

dF = −M dB = −M µ0 λ dM. (6.201)

Therefore,
Z
1
∆Epot = − M 0 µ0 λ dM 0 = − µ0 λM 2 (6.202)
2
and
1
∆Epot = − U (g(EF ) δE)2 , (6.203)
2

280
with

U ≡ µ0 µ2B λ. (6.204)

The total energy gain is therefore

1
∆Etot = ∆Ekin + ∆Epot = g(EF ) (δE)2 (1 − U g(EF )) , (6.205)
2
Spontaneous ferromagnetism shows up if ∆Etot < 0, or equivalently if

U g(EF ) > 1. (6.206)

That is the Stoner criterion. Note that is does not depend on the energy shift
δE.

Stoner amplification of the susceptibility. Even if the Stoner criterion is not

fulfilled, electronic correlations can (contained in U ) result in an effect, namely
the Stoner amplification of the susceptibility. In a qualitative picture: a
material that if ‘almost ferromagnetic’ can be magnetized easier than a material
with less correlations.
The material is applied to an external magnetic field B. The shift of the
spin-split bands is results in a magnetization given by eq. (6.193). Thus, the
change in total energy is given by
1
∆Etot = g(EF ) (δE)2 (1 − U g(EF )) − M B (6.207a)
2
1 − U g(EF ) 2
= M − M B. (6.207b)
2µ2B g(EF )

Minimization of the energy gain,

∂∆Etot
= 0, (6.208)
∂M
gives
B
M= µ2 g(EF ), (6.209)
1 − U g(EF ) B

from which follows the susceptibility

∂M µ0 µ2B g(EF ) χPauli

χ = µ0 = = . (6.210)
∂B 1 − U g(EF ) 1 − U g(EF )

That means that Pauli paramagnetism is increased by a factor [1 − U g(EF )]−1 ,

the so-called Stoner amplification. It is in particular important for the noble
metals Pd and Pt, both being ‘close to ferromagnetism’.

281
6.5.5 Quantum-well induced ferromagnetism in thin films
Both Pt and Pd are ‘close to being ferromagnetic’. To produce magnetic nano-
structures of these elements, one has to fulfill the Stoner criterion. Since the
exchange integral U is hardly to influence, an obvious choice is to increase the
density of states at the Fermi level to obtain U g(EF ) > 1.
During the course of this lecture series we have come across several times
quantum well states, the latter resulting in an oscillatory DOS at the Fermi
level. So, one can search for Pd and Pt films grown on a substrate that fulfill
[160, 161]
1
g(EF ) > . (6.211)
U
The maximum density of states in Pd films on Ag(100) is shown in Fig. 6.78
For films 3 ML and 4 ML thick as well as for films with thicknesses N larger
than 8 ML there should be a magnetic moment within the Pd film.

Figure 6.78. Maximum Pd layer-resolved state density at the Fermi level as a function
of Pd coverage.

And indeed, spin-resolved first-principles calculations are fully consistent

with this prediction (Fig. 6.79), as is evident from the magnetic profiles of
PdN /Ag(100).
In order to prove the origin of the magnetism, one needs to investigate
the electronic structure at the Fermi level of both Pd and Ag in more de-
tail (Fig. 6.80). In order to confine Pd electrons in the film, there must be

282
Figure 6.79. Calculated magnetic-moment profiles for PdN /Ag(100) as functions of
Pd coverage N .

a gap in the Ag band structure. From panel (a) it is evident that the free-
electron-like sp band of Ag crosses the Fermi level whereas the d-band range is
below −2.8 eV. Thus, there is no fundamental band gap in Ag at the specified
in-plane wavevector. Note that the sp is of the representation ∆1 , that is, its
is build from s- and pz -orbitals. [31]

Figure 6.80. Energy bands for Ag and Pd in the direction perpendicular to the Ag/Pd
interface. The vertical arrow indicates the energy range of the Pd d-band quantum well.
In the Pd panel we show by open squares as an example the topmost QW branch of
Fig. 6.81 quantized according to the number of Pd layers.

In Pd two band cross the Fermi level, the sp-band of ∆1 representation and a
d-band of ∆5 representation. The electronic states associated with the former

283
can hybridize with those of the sp-band in Ag. Thus, there is no confinement
for these electrons. But the d-electrons are confined to the Pd film because of
the different representations.
With increasing thickness N , the quantum well states disperse to larger en-
ergies, consequently crossing the Fermi level (Fig. 6.81). As is evident, there
are no QW states close to EF for N = 1, 2, 5, 6, 7, in agreement with Fig.6.78.
At all other thicknesses, there are QWS in an energy range [EF − , EF + ].
Due to the ‘positive feedback’ scenario (page 279), these states are spin-split.

Figure 6.81. Quantum-well states in PdN /Ag(001) as functions of Pd coverage N .

Open triangles correspond to minority-spin states and full triangles to majority-spin
states. The dotted lines correspond to the paramagnetic QW states. The horizon-
tally shaded area indicates the energy range  where QW states must be present at a
particular coverage in order to induce magnetism in the Pd film.

The energy range  can be estimated as follows. The Stoner exchange parame-
ter U is known to be 0.65 eV. With the largest magnetic moment of M = 0.35 µB
within a Pd film, it follows

 = M U ≈ 0.2 eV. (6.212)

Hence, in order to maintain the magnetism, the spin-up state must be occupied
while the spin-down state must be unoccupied. Therefore, the paramagnetic

284
QWS (no exchange splitting) must be closer than 0.1 eV to the Fermi level (see
Fig. 6.81).

6.5.6 Other magnetisms

In the previous we already have come across various magnetisms. Here is a
brief overview.
Antiferromagnetism
In materials that exhibit antiferromagnetism, the spins of magnetic electrons
align in a regular pattern with neighboring spins pointing in opposite direc-
tions. This is a different manifestation of magnetism. Generally, antiferromag-
netic materials exhibit antiferromagnetism at a low temperature, and become
disordered above a certain temperature; the transition temperature is called
the Néel temperature. Above the Néel temperature, the material is typically
paramagnetic.
The antiferromagnetic behaviour at low temperature usually results in dia-
magnetic properties, but can sometimes display ferrimagnetic behaviour, which
in many physically observable properties is more similar to ferromagnetic inter-
actions.
The magnetic susceptibility of an antiferromagnetic material will appear to
go through a maximum as the temperature is lowered; in contrast, that of a
paramagnet will continually increase with decreasing temperature.
Antiferromagnetic materials have low geometrical frustration, and a negative
coupling between adjacent moments.
Antiferromagnetic materials are relatively uncommon. More everyday exam-
ples include metals such as chromium, alloys such as iron manganese (FeMn),
and oxides such as nickel oxide (NiO). There are also numerous examples among
high nuclearity metal clusters. Organic molecules can also exhibit antiferromag-
netic coupling under rare circumstances.
Antiferromagnets can also couple to ferromagnetic materials through a mech-
anism known as exchange anisotropy, in which the ferromagnetic film is either
grown upon the antiferromagnet or annealed in an aligning magnetic field, caus-
ing the surface atoms of the ferromagnet to align with the surface atoms of the
antiferromagnet. This provides the ability to ‘pin’ the orientation of a ferromag-
netic film, which provides one of the main uses in so-called spin valves, which
are the basis of magnetic sensors including modern hard drive read heads.
Excursion: Exchange bias
The phenomena of exchange anisotropy and exchange bias, particularly, were
discovered in 1956 by Meiklejohn and Bean when studying Co particles sur-
rounded with antiferromagnetic oxide (CoO). They found that the field re-
quired to switch the ferromagnet from the field cooled state into the reversed
state is larger than that to rotate the ferromagnet back to its original direction
(Fig. 6.82).
A unidirectional anisotropy in the ferromagnet requires the so-called field
cooling process: The bilayer system is cooled below the Néel temperature (TN )
of the antiferromagnet, starting at a temperature T (TN < T < TC , where

285
 Figure 6.82. Hysteresis loop of an unbiased (black) and a biased film (red).

TC is the Curie temperature of the ferromagnet). Usually an external field is

applied to fix the direction of the ferromagnet during the cooling process (1 in
Fig. 6.83).

Figure 6.83. Magnetic configurations of an exchange biased film.

Below the Néel temperature the antiferromagnet chooses the state that mini-
mizes the energy due to the coupling to the ferromagnet (2). When the external
field reverses the ferromagnet, the antiferromagnetic spins do not switch if the
assumption of sufficiently high anisotropy is fulfilled (3–4). The AF spins at the
interface try to keep the ferromagnetic spins in their original direction. Conse-
quently the external field needed to reverse an exchange biased ferromagnet is
larger than for single ferromagnetic layers.
For increasing H the ferromagnet switches back at lower external fields due to

286
the codirectional AF torque (5–2). Altogether the magnetization curve exhibits
a shift along the field axis, i.e. exchange bias.
Although this model of Meiklejohn and Bean gives an intuitive insight into
the phenomena of exchange bias, it predicts bias fields several orders of magni-
tude larger than all experimentally measured values. Moreover, exchange bias
was found experimentally for uncompensated as well as for fully compensated
interfaces.
Since the introduction of giant magnetoresistive (GMR) spin-valve heads in
magnetic recording, the bias effect has been used widely in modern technology.
But the physical mechanisms leading to the hysteresis shift are still in discus-
sion. One of the most striking experimental facts is the presence of exchange
bias at fully compensated antiferromagnetic interfaces in which the net spin
averaged over a microscopic length scale is zero (Fig. 6.84, right).

Figure 6.84. Uncompensated (left) and compensated (right) magnetic structure at

the interface of an exchange biased film.

For uncompensated interfaces (Fig. 6.84, left) the antiferromagnetic spins of

one sublattice dominate at the AF/F contact surface. The coupling between the
ferromagnet and the net moment of the antiferromagnet is collinear or direct.
Obviously the initial state and the reversed state of the ferromagnetic layer are
antisymmetric resulting in the shift of the hysteresis curve.
At perfectly compensated interfaces (Fig. 6.84, right) the net spin of the
antiferromagnet averaged over a microscopic length scale is zero. A fully com-
pensated antiferromagnetic interface will lead to a 90 ◦ or spin flop coupling
between the two layers. A weak canting of the AF spins close to the interface
provide a small net moment parallel to the interface to which the ferromagnet
is able to couple. Simple models predict that AF/F layers with compensated
interfaces exhibit an enhanced coercivity but no exchange bias. However, our
simulations resulted in exchange bias fields and coercivities comparable to ex-
perimentally found data.
Ferrimagnetism
In physics, a ferrimagnetic material is one in which the magnetic moment of the
atoms on different sublattices are opposed, as in antiferromagnetism; however,
in ferrimagnetic materials, the opposing moments are unequal and a sponta-
neous magnetization remains. This happens when the sublattices consist of
different materials or ions.
Ferrimagnetic materials are like ferromagnets in that they hold a spontaneous

287
Figure 6.85. The calculated atomic structure and local DOS for majority-spin elec-
trons (top panels) and minority-spin electrons (bottom panels) for a Co/Al2 O3 /Co
tunnel junction.

magnetization below the Curie temperature, and show no magnetic order (are
paramagnetic) above this temperature. However, there is sometimes a tem-
perature below the Curie temperature at which the two sublattices have equal
moments, resulting in a net magnetic moment of zero; this is called the com-
pensation point. The compensation point is observed easily in garnets.
Ferrimagnetism is exhibited by ferrites and magnetic garnets. The oldest-
known magnetic substance, magnetite, is a ferrimagnet; it was originally clas-
sified as a ferromagnet before Néel’s discovery of ferrimagnetism and antiferro-
magnetism.
Superdiamagnetism
Superdiamagnetism (or perfect diamagnetism) is a phenomenon occurring in
certain materials at low temperatures, characterized by the complete absence
of magnetic susceptibility and the exclusion of the interior magnetic field. Su-
perdiamagnetism is a feature of superconductivity. It was identified in 1933, by
Walter Meissner and Robert Ochsenfeld (the Meissner effect).
Superdiamagnetism established that the superconductivity was a stage of
phase transition. Superconducting magnetic levitation is due to the Superdia-

288
magnetism (which repels a permanent magnet) and flux pinning, which stops
the magnet from sliding away.
Superparamagnetism
Superparamagnetic is a term used to describe materials intermediate between
paramagnetic and ferromagnetic. Examples include iron oxide particles (used
in MRI contrast agents), ferritin and haemosiderin.

289
Bibliography
[1] Y. Imry. Introduction to Mesoscopic Physics. Oxford University Press,
Oxford, 2nd edition, 2002.

[2] T. Ando, Y. Arakawa, K.Furuya, S. Komiyama, and H. Nakashima. Meso-

scopic Physics and Electronics. Springer, Berlin, 1998.

[3] J. I. Martn, J. Nogues, K. Liu, J. L. Vicent, and I. K. Schuller. Or-

dered magnetic nanostructures: Fabrication and properties. Journal of
Magnetism and Magnetic Materials, 449:430, 2003.

[4] D. Shi, B. Aktas, L. Pust, and F. Mikailov, editors. Nanostructured

Magnetic Materials and Their Application. Springer, Berlin, 2002.

[5] Th. Schäpers. Nanostructures, chapter B2, pages B2.1–B2.24. Jülich

KFZ, 1998.

[6] W.-D. Schneider. Scanning tunneling microscopy/spectroscopy of nano-

structures. physica status solidi (a), 187:125, 2001.

[7] S. Datta. Electronic Transport in Mesoscopic Systems. Cambridge Uni-

versity Press, Cambridge, 1995.

[8] S. Maekawa and T. Shinjo, editors. Spin Dependent Transport in Magnetic

Nanostructures. Taylor & Francis, London, 2002.

[9] H. S. Nalwa, editor. Magnetic Nanostructures. American Scientific Pub-

lishers, Stevenson Ranch, 2002.

[10] F. J. Himpsel, J. E. Ortega, G. J. Mankey, and R. F. Willis. Magnetic

nanostructures. Advances in Physics, 47:511, 1998.

[11] E. Merzbacher. Quantum Mechanics. John Wiley & Sons, New York, 2
edition, 1970.

[12] L. de Broglie. Sur un théoremè de M. Bohr. Comte Rendu, 179:676, 1924.

[13] L. de Broglie. Sur la dynamique du quantum de luimière et les in-

terfèrences. Comte Rendu, 179:1039, 1924.

[14] L. de Broglie. Sur la frèquence propre de l’électron. Comte Rendu,

180:498, 1925.

[15] A. Einstein. Über einen die Erzeugung und Verwandlung des Lichtes
betreffenden heuristischen Gesichtspunkt. Annalen der Physik, 17:132–
148, 1905.

290
[16] C. Davisson and L. H. Germer. Diffraction of electrons by a single layer
of atoms. Physical Review, 31:155, 1928.

[17] C. Davisson and L. H. Germer. Diffraction of electrons by a single layer of

gas atoms adsorbed upon a nickel crystal. Physical Review, 31:307, 1928.

[18] E. D. Chisolm. Generalizing the Heisenberg uncertainty relation. Amer-

ican Journal of Physics, 69:368, 2001.

[19] A. Messiah. Quantenmechanik, volume 2. W. de Gruyter, Berlin, 1979.

[20] E. Schrödinger. Quantisierung als Eigenwertproblem. Erste Mitteilung.

Annalen der Physik, 79:361, 1926.

[21] E. Schrödinger. Quantisierung als Eigenwertproblem. Zweite Mitteilung.

Annalen der Physik, 79:489, 1926.

[22] E. Schrödinger. Quantisierung als Eigenwertproblem. Dritte Mitteilung.

Annalen der Physik, 80:437, 1926.

[23] E. Schrödinger. Quantisierung als Eigenwertproblem. Vierte Mitteilung.

Annalen der Physik, 81:109, 1926.

[24] A. Pais, M. Jacob, D. I. Olive, and M. F. Atiyah. Paul Dirac. The Man
and His Work. Cambridge University Press, Cambridge, 1999.

[25] W.-H. Steeb. Hilbert Spaces, Generalized Functions and Quantum Me-
chanics. B. I. Wissenschaftsverlag, Mannheim, 1991.

[26] P. M. Morse and H. Feshbach. Methods of Theoretical Physics, volume I.

McGraw Hill, New York, 1953.

[27] J. Henk. Theory of low-energy diffraction and photoelectron spectroscopy

from ultra-thin films. In H. S. Nalwa, editor, Handbook of Thin Film
Materials, volume 2, chapter 10, page 479. Academic Press, San Diego,
2001.

[28] I. N. Bronstein and K. A. Semendjajwew. Taschenbuch der Mathematik.

Harri Deutsch, Thun, 19 edition, 1979.

[29] P. Weinberger. Electron Scattering Theory of Ordered and Disordered

Matter. Clarendon Press, Oxford, 1990.

[30] N. W. Ashcroft and N. D. Mermin. Solid State Physics. Holt-Saunders

International Editions, London, 1976.

[31] T. Inui, Y. Tanabe, and Y. Onodera. Group Theory and Its Applications
in Physics, volume 78 of Springer Series in Solid State Sciences. Springer,
Berlin, 1990.

[32] T. G. Vold. An introduction to geometric algebra with an application in

rigid body mechanics. American Journal of Physics, 61:491, 1993.

291
[33] T. G. Vold. An introduction to geometric calculus and its application to
electrodynamics. American Journal of Physics, 61:505, 1993.
[34] V. V. Nemoshkalenko and V. N. Antonov. Computational Methods in
Solid State Physics. CRC Press, Boca Raton, 1998.
[35] P.-O. Löwdin. On the Non-Orthogonality Problem Connected with the
Use of Atomic Wave Functions in the Theory of Molecules and Crystals.
The Journal of Chemical Physics, 18:365–375, 1950.
[36] J. C. Slater and G. F. Koster. Simplified LCAO Method for the Periodic
Potential Problem. Physical Review, 94:1498–1524, 1954.
[37] H. J. W. M. Hoekstra. On the Tight-Binding Model for Semi-Infinite
Lattices. Surface Science, 205:523–548, 1988.
[38] J. Henk and B. Johansson. Quantum-size effects in photoemission from
ultra-thin films: Theory and application to Cu-films on fcc-Co(001). Jour-
nal of Electron Spectroscopy and Related Phenomena, 105:187, 1999.
[39] Y. Wang, P. Nordlander, and N. H. Tolk. Extended Hückel theory for
ionic molecules and solids: An application to alkali halides. The Journal
of Chemical Physics, 89:4163–4169, 1988.
[40] V. Heine. On the general theory of surface states and scattering of elec-
trons in solids. Proceedings of the Physical Society, 81:300–310, 1963.
[41] W. A. Harrison. Tight-binding theory of surface states in metals. Physica
Scripta, 67:253, 2003.
[42] S. G. Davison and M. Stȩślicka. Basic Theory of Surface States. Claren-
don, Oxford, 1992.
[43] A. Zangwill. Physics at Surfaces. Cambridge University Press, Cam-
bridge, 1988.
[44] W. Shockley. On the surface states associated with a periodic potential.
Physical Review, 56:317, 1939.
[45] R. Haydock. The Recursive Solution of the Schrödinger Equation. Solid
State Physics, 35:215, 1980.
[46] M. P. Lopéz Sancho, J. M. Lopéz Sancho, and J. Rubio. Highly convergent
schemes for the calculation of bulk and surface Green functions. Journal
of Physics F: Metal Physics, 15:851–858, 1985.
[47] J. Henk and W. Schattke. A subroutine package for computing Green’s
functions of relaxed surfaces by the renormalization method. Computer
Physics Communications, 77:69, 1993.
[48] A. Bödicker, W. Schattke, J. Henk, and R. Feder. Interface electronic
structure by the renormalization method: Theory and application to
Sb/GaAs. Journal of Physics: Condensed Matter, 6:1927, 1994.

292
[49] S. V. Halilov, J. Henk, T. Scheunemann, and R. Feder. Surface states
and photoemission of magnetic multilayer systems. Physical Review B,
52:14 235, 1995.

[50] P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Physical Review

B, 136:864, 1964.

[51] R. G. Parr and W. Yang. Density-Functional Theory of Atoms and

Molecules. Oxford University Press, New York, 1989.

[52] W. Kohn and L. J. Sham. Self-consistent equations including exchange

and correlation effects. Physical Review, 140:A1133, 1965.

[53] L. J. Sham and W. Kohn. One-particle properties of an inhomogeneous

interacting electron gas. Physical Review, 145:561, 1966.

[54] A. I. Lichtenstein and M. I. Katsnelson. Magnetism of correlated systems:

Beyond LDA. In Baberschke et al. [55], page 75.

[55] K. Baberschke, W. Nolting, and M. Donath, editors. Band-

Ferromagnetism: Ground-State and Finite-Temperature Phenomena, vol-
ume 580 of Lecture Notes in Physics. Springer, Berlin, 2001.

[56] J. P. Perdew and A. Zunger. Self-interaction correction to density-

functional approximations for many-electron systems. Physical Review
B, 23:5048, 1981.

[57] J. P. Perdew, K. Burke, and M. Ernzerhof. Generalized gradient approx-

imation made simple. Physical Review Letters, 77:3865, 1996.

[58] H. Tasaki. The Hubbard model – an introduction and selected rigorous

results. Journal of Physics: Condensed Matter, 10:4353, 1998.

[59] A. Montorsi, editor. The Hubbard Model—A Reprint Volume. World

Scientific, Singapore, 1992.

[60] K. Held, I. A. Nekrasov, N. Blümer, V. I. Anisimov, and D. Vollhardt. Re-

alistic modeling of strongly correlated electron systems: An introduction
to the LDA+DMFT approach. International Journal of Modern Physics
B, 15:2611, 2001.

[61] J. B. Pendry. Low Energy Electron Diffraction. Academic Press, London,

1974.

[62] S. Y. Tong. Theory of low-energy electron diffraction. In S. G. Davisson,

editor, Progress in Surface Science, volume 7, page 1. Pergamon, London,
1975.

[63] M. A. van Hove and S. Y. Tong. Surface Crystallography by LEED.

Theory, Computation and Structural Results, volume 2 of Springer Series
in Chemical Physics. Springer, Berlin, 1979.

293
[64] M. P. Seah and W. A. Dench. Quantitative electron spectroscopy of
surfaces: A standard data base for electron inelastic mean free paths in
solids. Surface & Interface Analysis, 1:2, 1979.

[65] R. Feder. Spin-polarised low-energy electron diffraction. Journal of

Physics C: Solid State Physics, 14:2049, 1981.

[66] S. Hüfner. Photoelectron Spectroscopy. Springer, Berlin, 3rd edition, 2003.

[67] H.-P. Bonzel and C. Kleint. On the History of Photoemission. Progress

in Surface Science, 49:107–153, 1995.

[68] K. Siegbahn, C. Nordling, A. Fahlman, R. Nordberg, K. Hamrin, J. Hed-

man, G. Johanson, T. Bergmark, S.-E. Karlsson, I. Lindgren, and B. Lind-
berg. ESCA – Atomic, Molecular and Solid State Structure Studied by
Means of Electron Spectroscopy. Almqvist och Wiksells, Uppsala, 1967.

[69] C. N. Berglund and W. E. Spicer. Photoemission Studies of Copper and

Silver: Theory. Physical Review, 136:A1030–A1044, 1964.

[70] M. Cardona and L. Ley, editors. Photoemission in Solids I. Number 26

in Topics in Applied Physics. Springer, Berlin, 1978.

[71] A. Liebsch. Angle-resolved photoemission from adsorbates. In B. Feuer-

bacher, B. Fitton, and R. F. Willis, editors, Photoemission and the elec-
tronic properties of surfaces, page 167. Wiley, Chichester, 1978.

[72] S. V. Kevan, editor. Angle-resolved Photoemission: Theory and Current

Applications. Elsevier, Amsterdam, 1992.

[73] J. B. Pendry. Theory of photoemission. Surface Science, 57:679–705,

1976.

[74] J. F. L. Hopkinson, J. B. Pendry, and D. J. Titterington. Calculations of

photoemission spectra for surfaces of solids. Computer Physics Commu-
nications, 19:69–92, 1980.

[75] J. Braun. The theory of angle-resolved ultra-violet photoemission and

its applications to ordered materials. Reports on Progress in Physics,
59:1267, 1996.

[76] P. J. Feibelman and D. E. Eastman. Photoemission spectroscopy—

Correspondence between quantum theory and experimental phenomenol-
ogy. Physical Review B, 10:4932, 1974.

[77] C. Caroli, D. Lederer-Rozenblatt, B. Roulet, and D. Saint-James. In-

elastic effects in photoemission: Microscopic formulation and qualitative
discussion. Physical Review B, 8:4552, 1973.

[78] T. Fujikawa and L. Hedin. Theory of electron scattering from solids.

Physical Review B, 40:11 507, 1989.

294
[79] R. Feder, editor. Polarized Electrons in Surface Physics. Advanced Series
in Surface Science. World Scientific, Singapore, 1985.
[80] R. Feder and J. Henk. Magnetic dichroism and spin polarization in angle-
resolved photoemission. In H. Ebert and G. Schütz, editors, Spin-Orbit
Influenced Spectroscopies of Magnetic Solids, number 466 in Lecture Notes
in Physics, page 85. Springer, Berlin, 1996.
[81] J. Kessler. Polarized Electrons, volume 1 of Springer Series on Atoms
and Plasmas. Springer, Berlin, 2nd edition, 1985.
[82] E. Cartan. The Theory of Spinors. Hermann, Paris, 1966.
[83] R. Penrose. The mathematics of the electron’s spin. European Journal of
Physics, 18:164, 1997.
[84] J. Henk, P. Bose, Th. Michael, and P. Bruno. Spin motion of photoelec-
trons. Physical Review B, 68:052403, 2003.
[85] F. Reinert. Spin-orbit interaction in the photoemission spectra of noble
metal surface states. Journal of Physics: Condensed Matter, 15:S693,
2003.
[86] J. Henk, M. Hoesch, J. Osterwalder, A. Ernst, and P. Bruno. Spin-orbit
coupling in the L-gap surface states of Au(111): Spin-resolved photoe-
mission experiments and first-principles calculations. Journal of Physics:
Condensed Matter, 16:7581, 2004.
[87] J. Henk, A. M. N. Niklasson, and B. Johansson. Multiple-scattering the-
oretical approach to magnetic dichroism and spin polarization in angle-
resolved core-level photoemission. Physical Review B, 59:13 986, 1999.
[88] E. M. Rose. Relativistic Electron Theory. Wiley & Sons, New York, 1961.
[89] P. Strange. Relativistic Quantum Mechanics: With Applications in Con-
densed Matter and Atomic Physics. Cambridge University Press, Cam-
bridge, 1998.
[90] E. I. Rashba. Soviet Physics Solid State, 2:1109, 1960.
[91] Y. A. Bychkov and E. I. Rashba. Soviet Physics JETP Letters, 39:78,
1984.
[92] S. Datta and B. Das. Electronic analogue of the electronic modulator.
Applied Physics Letters, 56(7):665–667, 1990.
[93] R. Landauer. Spatial variation of currents and fields due to localized scat-
terers in metallic conduction. IBM Journal of Research and Development,
1:223, 1957.
[94] M. Büttiker, Y. Imri, R. Landauer, and S. Pinhas. Generalized many-
channel conductance formula with application to small rings. Physical
Review B, 31:6207, 1985.

295
 [95] Y. Y. Sharvin. A possible method for studying Fermi surfaces. Soviet
Physics JETP, 21:655, 1965.

[96] I. Giaever and H. R. Zeller. Physical Review Letters, 20:1504, 1968.

[97] L. J. Geerlings, V. F. Anderegg, P. A. M. Holweg, J. E. Mooij, H. Poth-

ier, D. Esteve, C. Urbina, and M. H. Devoret. Physical Review Letters,
64:2691, 1990.

[98] C. J. Gorter. Physica, 17:777, 1951.

[99] J. Lambe and R. C. Jaklevic. Physical Review Letters, 22:1371, 1969.

[100] J. H. F. Scott-Thomas, S. B. Field, M. A. Kastner, H. I. Smith, and D. A.

Antonidis. Physical Review Letters, 62:583, 1989.

[101] U. Meirav, M. A. Kastner, and S. J. Wind. Physical Review Letters,

65:771, 1990.

[102] D. Goldhaber-Gordon, H. Shtrikman, D D. Mahalu, Abusch-Magder,

U. Meirav, and M. A. Kastner. Nature, 391:156, 1998.

[103] G. A. Prinz. Magnetoelectronics. Science, 282:1660, 1998.

[104] D. Grundler. Spintronics. Physics World, 15(4):39, 2002.

[105] T. J. Klemmer, J. Z. Sun, A. Fert, and J. Bass, editors. Spintronics,

volume 260 of Symposium Proceedings. Materials Research Society, War-
rendale, 2002.

[106] S. Zhang, M. R. Fitzsimmons, W. Kuch, I. K. Schuller, G. Güntherodt,

J. B. Kortright, C. Broholm, T. Shinjo, A. D. Kent, and Y. Zhu, editors.
Magnetoelectronics and Magnetic Materials—Novel Phenomena and Ad-
vanced Characterization. Materials Research Society, Warrendale, 2003.

[107] I. Žutić, J. Fabian, and S. Das Sarma. Spintronics: Fundamentals and

applications. Reviews of Modern Physics, 76:323, 2004.

[108] G. A. Prinz. Magnetoelectronics. In Bland and Heinrich [109], chapter 2.

[109] B. Heinrich and J. A. C. Bland, editors. Ultrathin Magnetic Structures

IV. Applications of Nanomagnetism. Springer, Berlin, 2005.

[110] S. Yuasa, T. Nagahama, A. Fukushima, Y. Suzuki, and K. Ando. Giant

room-temperature magnetoresistance in single-crystal Fe/MgO/Fe mag-
netic tunnel junctions. Nature Materials, 3:868, 2004.

[111] H. L. Meyerheim, R. Popescu, J. Kirschner, N. Jedrecy, M. Sauvage-

Simkin, B. Heinrich, and R. Pinchaux. Geometrical and compositional
structure at metal-oxide interfaces: MgO on Fe(001). Physical Review
Letters, 87:076102, 2001.

296
[112] C. Tusche, H. L. Meyerheim, N. Jedrecy, G. Renaud, A. Ernst, J. Henk,
P. Bruno, and J. Kirschner. Oxygen-induced symmetrization and struc-
tural coherency in Fe/MgO/Fe(001) magnetic tunnel junctions. Physical
Review Letters, 95:176101, 2005.
[113] M. Jullière. Tunneling between ferromagnetic films. Physics Letters A,
54:225, 1975.
[114] S. Maekawa and U. Gafvert. Electron tunneling between ferromagnetic
films. IEEE Transactions on Magnetics, MAG-18:707, 1982.
[115] J. M. MacLaren, X.-G. Zhang, and W. H. Butler. Validity of the Jullière
model of spin-dependent tunneling. Physical Review B, 56:11 827, 1997.
[116] J. C. Slonczewski. Conductance and exchange coupling of two ferromag-
nets separated by a tunneling barrier. Physical Review B, 39:6995, 1989.
[117] J. S. Moodera, L. R. Kinder, T. M. Wong, and R. Meservey. Large
magnetoresistance ar room temperature in ferromagnetic thin film tunnel
junctions. Physical Review Letters, 74:3273, 1995.
[118] J. Barnaś, O. Baksalary, and A. Fert. Angular dependence of giant mag-
netoresistance in magnetic multilayers. Physical Review B, 56:6079, 1997.
[119] S. Khmelevskyi, K. Palotás, L. Szunyogh, and P. Weinberger. Ab initio
calculation of the anisotropic magnetoresistance in Ni1−c Fec bulk alloys.
Physical Review B, 68:012402, 2003.
[120] P. M. Tedrow and R. Meservey. Physical Review Letters, 26:192, 1971.
[121] J. Velev and W. H. Butler. Applicability of the coherent potential ap-
proximation for transmission through disordered interfaces. Journal of
Applied Physics, 97:10C517, 2005.
[122] E. Y. Tsymbal and D. G. Pettifor. Perspectives of giant magnetoresis-
tance. In H. Ehrenreich and F. Spaepen, editors, Solid State Physics,
volume 56, page 113. Academic Press, San Diego, 2001.
[123] U. Hartmann, editor. Magnetic Multilayers and Giant Magnetoresistance.
Fundamentals and Industrial Applications, volume 37 of Springer Series
in Surface Sciences. Springer, Berlin, 1999.
[124] A. Barthélémy, A. Fert, and F. Petroff. Giant magnetoresistance in mag-
netic multilayers. In K. H. J. Buschow, editor, Handbook of Magnetic
Materials, volume 12, page 1. Elsevier, Amsterdam, 1999.
[125] M. Baibich, J. M. Broto, A. Fert, F. Nguyen van Dau, F. Petroff, P. Eti-
enne, G. Creuzet, A. Friedrich, and J. Chazeles. Physical Review Letters,
61:2472–2475, 1988.
[126] P. Zahn, J. Binder, I. Mertig, R. Zeller, and P. H. Dederichs. Origin of
giant magnetoresistance: Bulk or interface scattering. Physical Review
Letters, 80:4309, 1998.

297
[127] P. Weinberger and L. Szunyogh. Interlayer exchange coupling and giant
magnetoresistance. Journal of Physics: Condensed Matter, 15:S479, 2003.

[128] P. Zahn, J. Binder, and I. Mertig. Impurity scattering and quan-

tum confinement in giant magnetoresistive systems. Physical Review B,
68:100403(R), 2003.

[129] F. Erler, P. Zahn, and I. Mertig. Finite-size effects in giant magnetoresis-

tance: An ab initio calculation. Physical Review B, 64:094408, 2001.

[130] M. Amman, R. Wilkins, E. Ben-Jacob, P. D. Maker, and R. C. Jaklevic.

Physical Review B, 43:1146, 1991.

[131] W. H. Butler, X.-G. Zhang, T. C. Schulthess, and J. M. MacLaren. Spin-

dependent tunneling conductance of Fe—MgO—Fe sandwiches. Physical
Review B, 63:054416, 2001.

[132] J. Mathon and A. Umerski. Theory of tunneling magnetoresistance of an

epitaxial Fe/MgO/Fe(001) junction. Physical Review B, 63:220403, 2001.

[133] J. M. MacLaren, W. H. Butler, and X. G. Zhang. Spin-dependent tun-

neling in epitaxial systems: Band dependence of conductance. Journal of
Applied Physics, 83:6521, 1998.

[134] Ph. Mavropoulos, O. Wunnicke, and P. H. Dederichs. Ballistic spin in-

jection and detection in Fe/semiconductor/Fe junctions. Physical Review
B, 66:024416, 2002.

[135] X.-G. Zhang and W. H. Butler. Band structure, evanescent states, and
transport in spin tunnel junctions. Journal of Physics: Condensed Matter,
15:R1603, 2003.

[136] S. Yuasa, A. Fukushima, T. Nagahama, K. Ando, and Y. Suzuki.

High tunnel magnetoresistance at room temperature in fully epitaxial
Fe/MgO/Fe tunnel junctions due to coherent spin-polarized tunneling.
Japanese Journal of Applied Physics, 43:L588, 2004.

[137] D. E. Bürgler, P. Grünberg, S. O. Demokritov, and M. T. Johnson. In-

terlayer exchange coupling in layered magnetic materials. In K. H. J.
Buschow, editor, Handbook of Magnetic Materials, volume 13, page 1.
Elsevier, Amsterdam, 2001.

[138] P. Bruno. Theory of interlayer exchange interactions in magnetic multi-

layers. Journal of Physics: Condensed Matter, 11:9403, 1999.

[139] M. D. Stiles. Interlayer exchange coupling. Journal of Magnetism and

Magnetic Materials, 200:322, 1999.

[140] P. Bruno. Theory of intrinsic and thermally induced interlayer magnetic

coupling between ferromagnetic films separated by an insulating layer.
Physical Review B, 49:13 231, 1994.

298
[141] P. Bruno. Quantum interferences and interlayer exchange coupling in
Co/Cu/Co(001). Journal of Magnetism and Magnetic Materials, 148:202,
1995.

[142] P. Bruno. Recent progress in the theory of interlayer exchange coupling.

Journal of Applied Physics, 76:6972, 1995.

[143] P. Bruno. Theory of interlayer magnetic coupling. Physical Review B,

52:411, 1995.

[144] J. C. Slonczewski. Overview of interlayer exchange theory. Journal of

Magnetism and Magnetic Materials, 150:13, 1995.

[145] J. A. C. Bland and B. Heinrich, editors. Ultrathin Magnetic Structures

III. Fundamentals of Nanomagnetism. Springer, Berlin, 2005.

[146] M. D. Stiles. Interlayer exchange coupling. In Bland and Heinrich [145],

chapter 4.

[147] Z. Q. Qiu and N. V. Smith. Quantum well states and oscillatory mag-
netic interlayer coupling. Journal of Physics: Condensed Matter, 14:R169,
2002.

[148] P. Bruno. Interlayer exchange coupling: Preasymptotic corrections. Eu-

ropean Physics Journal B, 11:83, 1999.

[149] P. H. Dederichs, K. Wildberger, and R. Zeller. Interlayer exchange cou-

pling and interface reflectivities in Fe/Cu, Co/Cu and Ni/Cu(001) layers.
Physica B, 237–238:239, 1997.

[150] K. Baberschke. Anisotropy in magnetism. In Baberschke et al. [55],

page 27.

[151] P. Bruno. Tight-binding approach to the orbital magnetic moment and

magnetocrystalline anisotropy of transition-metal monolayers. Physical
Review B, 39:865, 1989.

[152] P. Gambardella, S. Rusponi, M. Veronese, S. S. Dhesi, C. Grazioli,

A. Dallmeyer, I. Cabria, R. Zeller, P. H. Dederichs, K. Kern, C. Car-
bone, and H. Brune. Giant magnetic anisotropy of single cobalt atoms
and nanoparticles. Science, 300:1130, 2003.

[153] G. H. O. Daalderop, P. J. Kelly, and M. F. H. Schuurmans. First-

principles calculation of the magnetocrystalline anisotropy energy of iron,
cobalt, and nickel. Physical Review B, 41:11 919, 1990.

[154] X. Wang, D.-S. Wang, R. Wu, and A. J. Freeman. Validity of the force
theorem for magnetocrystalline anisotropy. Journal of Magnetism and
Magnetic Materials, 159:337, 1996.

[155] H. J. F. Hansen. Magnetic anisotropy in density-functional theory. Phys-

ical Review B, 59:4699, 1999.

299
[156] J. Henk, A. M. N. Niklasson, and B. Johansson. Magnetism and
anisotropy of ultra-thin Ni films on Cu(001). Physical Review B, 59:9332,
1999.

[157] R. A. Evarestov and V. P. Smirnow. Special Points of the Brillouin Zone

and Their Use in the Solid State Theory. physica status solidi (b), 119:9–
40, 1983.

[158] O. Hjortstam, K. Baberschke, J. M. Wills, B. Johansson, and O. Eriksson.

Magnetic anisotropy and magnetostriction in tetragonal and cubic Ni.
Physical Review B, 55:15 026, 1997.

[159] C. Uiberacker, J. Zabloudil, P. Weinberger, L. Szunyogh, and C. Som-

mers. Lattice relaxation driven reorientation transition in Nin /Cu(100).
Physical Review Letters, 82:1289, 1999.

[160] S. Mirbt, H. L. Skriver, and B. Johansson. Quantum-well driven mag-

netism in thin films. Physical Review B, 53:R13 310, 1996.

[161] A. M. N. Niklasson, S. Mirbt, H. L. Skriver, and B. Johansson. Quantum-

well-induced ferromagnetism in thin films. Physical Review B, 56:3276,
1997.

300