Classical and Quantum Spin Systems

Jung Hoon Han

November 4, 2015
2
Contents

1 Ginzburg-Landau Theory of Critical Phenomena 7

1.1 Introductory remark . . . . . . . . . . . . . . . . . . 7
1.2 What is an order parameter? . . . . . . . . . . . . . . 8
1.3 Order parameter expansion of free energy . . . . . . . 9
1.4 Its uses and limits . . . . . . . . . . . . . . . . . . . . 11
1.5 Multiple order parameters . . . . . . . . . . . . . . . 13

2 Mean-field Theory 15
2.1 Mean-field theory of Ising model . . . . . . . . . . . . 15
2.2 Ginzburg-Landau free energy from mean-field theory 18
2.3 Ising model coupled to magnetic field . . . . . . . . . 19
2.4 Mean-field theory of XY spins . . . . . . . . . . . . . 20

3 Fluctuations and Excitations 23

3.1 Domain walls for discrete spin . . . . . . . . . . . . . 23
3.2 Spin waves for continuous spin . . . . . . . . . . . . . 25
3.3 Topological excitations . . . . . . . . . . . . . . . . . 30

4 XY Model 33
4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Spin-wave Theory . . . . . . . . . . . . . . . . . . . . 34
4.3 High-temperature expansion . . . . . . . . . . . . . . 36
4.4 Vortex Excitation and Kosterlitz-Thouless transition 37
4.5 Helicity Modulus . . . . . . . . . . . . . . . . . . . . 39
4.6 Binder cumulent . . . . . . . . . . . . . . . . . . . . 40
4.7 Villain transformation . . . . . . . . . . . . . . . . . 40
4.8 Frustrated XY model . . . . . . . . . . . . . . . . . . 44

5 Heisenberg Model 47
5.1 Saddle-point Method . . . . . . . . . . . . . . . . . . 47
5.2 Renormalization Group Analysis . . . . . . . . . . . . 48
5.3 Skyrmion Excitations . . . . . . . . . . . . . . . . . . 49

3
4 CONTENTS

6 Monte Carlo Methods and Critical Phenomena 51

6.1 Introductory Remark . . . . . . . . . . . . . . . . . . 51
6.2 Monte Carlo (MC) for Ising spins . . . . . . . . . . . 53
6.3 Calculation of physical quantities . . . . . . . . . . . 56

7 Miscellaneous 59
7.1 Two-state Problem . . . . . . . . . . . . . . . . . . . 59
7.2 Three-state problem . . . . . . . . . . . . . . . . . . 59
7.3 d-orbitalogy . . . . . . . . . . . . . . . . . . . . . . . 59
7.4 p-orbitalogy . . . . . . . . . . . . . . . . . . . . . . . 60

8 Path Integral for Spins 61

8.1 Single spin . . . . . . . . . . . . . . . . . . . . . . . . 61
8.2 Antiferromagnetic spin chain . . . . . . . . . . . . . . 61

9 1D Quantum Spin Chain 63

9.1 The model . . . . . . . . . . . . . . . . . . . . . . . . 63
9.2 Ferromagnetic case . . . . . . . . . . . . . . . . . . . 63
9.3 Antiferromagnetic case . . . . . . . . . . . . . . . . . 63

10 Hubbard Model 69
10.1 The Model . . . . . . . . . . . . . . . . . . . . . . . . 69
10.2 Projection to truncated Hilbert space . . . . . . . . . 71
10.3 Mean-field Theory and Antiferromagnetism . . . . . . 74
10.3.1 Square lattice . . . . . . . . . . . . . . . . . . 74
10.3.2 Triangular lattice . . . . . . . . . . . . . . . . 77
10.4 Spin dependent hopping and DM interaction . . . . . 78
10.5 Orbitally degenerate Hubbard model . . . . . . . . . 80
10.6 Four-boson Theory . . . . . . . . . . . . . . . . . . . 81
10.6.1 Non-uniform Case . . . . . . . . . . . . . . . . 81
10.6.2 Uniform Case . . . . . . . . . . . . . . . . . . 82

11 Spin Interactions 87
11.1 Higher-order exchange processes . . . . . . . . . . . . 87

12 Spin Representation and Spin Excitation 91

12.1 Semiclassical Equation of Motion . . . . . . . . . . . 91
12.1.1 Ferromagnet . . . . . . . . . . . . . . . . . . . 91
12.1.2 Antiferromagnet . . . . . . . . . . . . . . . . 93
12.1.3 Spiral ferromagnet . . . . . . . . . . . . . . . 94
12.1.4 Multiple spiral ferromagnet . . . . . . . . . . 95
12.2 Holstein-Primakoff theory . . . . . . . . . . . . . . . 95
12.2.1 Ferromagnet . . . . . . . . . . . . . . . . . . . 96
CONTENTS 5

12.2.2 Antiferromagnet . . . . . . . . . . . . . . . . 97
12.2.3 Spiral ferromagnet . . . . . . . . . . . . . . . 98
12.2.4 Multiple spiral ferromagnet . . . . . . . . . . 98
12.3 Schwinger bosons . . . . . . . . . . . . . . . . . . . . 99
12.3.1 Ferromagnet . . . . . . . . . . . . . . . . . . . 100
12.3.2 Antiferromagnet . . . . . . . . . . . . . . . . 100
12.3.3 Spiral ferromagnet . . . . . . . . . . . . . . . 102
12.3.4 Multiple spiral ferromagnet . . . . . . . . . . 102
12.4 Slave boson . . . . . . . . . . . . . . . . . . . . . . . 102
12.5 Slave-fermion Schwinger boson . . . . . . . . . . . . . 103

13 t-J Model 105

14 Exact Spin Hamiltonians 113

14.1 AKLT States . . . . . . . . . . . . . . . . . . . . . . 113
14.2 Majumdar-Ghosh states . . . . . . . . . . . . . . . . 114
6 CONTENTS
Chapter 1

Ginzburg-Landau Theory of
Critical Phenomena

1.1 Introductory remark

In my own view of what the field of condensed matter physics is all about, I
have the following statement: It is a field in which one deals with the question,
“What are the possible phases, its excitations, and the phase transitions of a
state of matter?”1
Let me explain this using the magnesium diboride (MgB2 ) as an illustration.
It is a material which superconducts at about 40K. Hence, the answer to the
first question (what is the possible phase?) in this case is that it is a supercon-
ductor. Superconductor is a type of ground state, that is the lowest possible
energy state of the system as dictated by the laws of quantum mechanics. If the
temperature is raised slightly above absolutely zero, this material is no longer
a simple superconductor. Rather, it is a superconductor with a small density
of the so-called elementary excitations floating above the condensate. The ele-
mentary excitations which are in this case the broken Cooper pairs determine
the transport, and much of the thermodynamic properties of the material. Un-
derstanding the nature of elementary excitations is therefore synonymous with
the understanding of the material properties themselves. As you increase the
temperature further, there are more and more elementary excitations until the
ground state property itself loses its meaning as it gets swamped by the sheer
number of elementary excitations. At a critical temperature it is no longer a
superconductor, but rather behaves like an ordinary metal. We have a phase
transition of the classical sort.
Properties of a piece of solid is characterized in several ways: by their elec-
trical resistivity, susceptibility which is a measure of the response to external
magnetic field, and specific heat, to name a few. Because there are so many
excitations and so little signature of the original ground state properties near
the transition temperature one might think that understanding the physical
phenomena near Tc may be very hard. Instead, it turns out that all the physi-
cal properties I listed above behave rather similarly, regardless of the materials
1 There are also the dominant issues of materials and their growth in laboratories. Without

an interesting new material to look into, none of these issues will be relevant, of course.

7
8CHAPTER 1. GINZBURG-LANDAU THEORY OF CRITICAL PHENOMENA

chosen, if one is doing the measurement sufficiently near Tc . The properties

of the materials become universal near the transition, and the detailed under-
standing of the universal behavior is one of the most active research interests in
the condensed matter community.
This lecture is about outlining the basic concepts and methods in theoreti-
cally understanding the above three questions of condensed matter. Although
the traditional avenue of teaching universality phenomena is in terms of scaling,
and the subsequent development of renormalization group, I have chosen to en-
tirely avoid them in this lecture. Instead, I present a more practical method of
“seeing” the universality emerge in a variety of model systems by Monte Carlo
techniques.

1.2 What is an order parameter?

Phase transition is a change of the state of matter from a disordered to a more
ordered form as the temperature is lowered. Mathematical description for such
change is often given by defining an order parameter, X. Using this quantity,
one says the phase transition is from a state in which order parameter has a
zero expectation value hXi = 0 at high temperature, to a state with a finite
average hXi =6 0 at low temperature. The expectation value is defined in the
usual sense of statistical mechanics:

Xe−βEC
P
hXi = PC −βEC . (1.1)
Ce

The sum is over all allowed configurations C with a probability weight according
to the Boltzmann factor and β = 1/kB T .
We are dealing with spin systems in this lecture, so let’s try to identify
the right order parameter for them. At high temperature the system is de-
magnetized, meaning that magnetic moments can point in Pany arbitrary direc-
N
tion so that its average over the whole sample is zero: i=1 hSi i = 0. With
some N = 1023 atoms present in a typical sample we are looking at, the av-
erage
PN cannot of course be exactly zero. A more precise statement will be that
i=1 hSi i = O(1). That is, only a tiny, unobservable fraction of all the moments
survives the averaging.
It so happens that, by simply cooling the sample, the spins begin to talk to
each other and order. Mathematically, one expresses this fact by
N 
X 
Si  = O(N ), (1.2)
 

 
i=1

where O(N ) means the sum is proportional to the number of spins, because all
the spins point more or less in
Pthe same direction. Appropriate order parameter
for the spin system is Stot = i Si . Its average is a macroscopic number if spins
are in an ordered phase, zero if disordered2 .
2 Not all transitions have a simple order parameter description. For example, a gas-to-liquid

transition (change of state of matter from disordered to less disordered form) is lacking an
obvious order parameter. Still, there is a well-defined transition temperature at which this
change of phase occurs.
1.3. ORDER PARAMETER EXPANSION OF FREE ENERGY 9

1.3 Order parameter expansion of free energy

A brief lesson in statistical mechanics is in order here. Say you are given a
box containing a large number of atoms or molecules and the box is isolated
from its environment in the sense that no particles can escape from or enter
into it. However, the box is a thermal conductor, and you can be certain that
the temperature inside the box is the same as that of its surroundings. Now,
we shake the box violently for a while and let it sit still for a long, long time,
while maintaining the environment at the same constant temperatue T . After
a sufficiently long wait you open the box, and what do you expect to see? The
answer, according to the rules of statistical mechanics, is “Whatever minimizes
F !” F is called the Helmholtz free energy and is defined by

F = E − T S. (1.3)

As you know very well, E is the total energy (sum of kinetic, potential, and
whatever else is relevant) of the system, and S is the total entropy. In other
words, what you find in a box is a maximal harmony between the efforts to
minimize the energy (which lowers F ), and simultaneously try to maximize the
entropy (which also lowers F ). Loss of energy is often accompanied by loss of
order, so the two quantities, E and T S, are typically in competition with each
other. For a very high temperature T , entropically favorable state will be found
because T S  E, and for a very low temperature we can largely ignore the
entropy effect and expect to find a minimum energy state being realized.
Here comes the great revelation of Landau and his famous disciple, Ginzburg3 .
They assumed that it is possible to write down the free energy F as a function
of its own order parameter which, for spin systems, is Stot . For convenience
we will introduce an average magnetization S ≡ Stot /N . Free energy is an ex-
tensive quantity so it will be proportional to N , but otherwise the free energy
density f ≡ F/N is assumed to depend solely on S. Further assume that f (S)
is an analytic function of S. All that means is that one can Taylor-expand f (S)
in powers of S for small S:
1
f (S) = f (0) + (∂α f )Sα + (∂α ∂β f )Sα Sβ + · · · . (1.4)
2
As it stands, however, it doesn’t look very useful. But, wait, let’s think a little
harder.
Say you are holding a piece of magnet with its magnetization pointing in
the +ẑ direction. You flip it, and the magnetization is pointing in −ẑ direction.
But it is the same piece of magnet after all! No matter how you rotate the
magnet on the palm of your hand it will be the same magnet, having the same
free energy f . In the mean time, the order parameter S has clearly changed its
orientation from +ẑ to −ẑ to whichever direction you rotated the magnet into.
This means that free energy cannot depend on the orientation of S, but only on
its magnitude. Then, f must be a function of S · S. Given this, we can re-write
the Taylor expansion as

t λ
f (S2 ) = f (0) + S2 + (S2 )2 + · · · . (1.5)
2 4
3 Ginzburg shared the 2003 Nobel prize for his development of Ginzburg-Landau theory.
10CHAPTER 1. GINZBURG-LANDAU THEORY OF CRITICAL PHENOMENA

We have obtained a free energy expansion for a ferromagnet in powers of S2 .

Since S is assumed small, it is probably OK to keep the few lowest-order terms
such as given above, and add higher order only if necessary. This type of free
energy expansion in terms of the putative order parameter is known as Ginzburg-
Landau (GL) free energy4 . The assumption of small-S is consistent only near
Tc , where the order parameter magnitude is small.
It turns out there are other types of magnetic ordering in solids. Most
notable is the antiferromagnetic (AFM) ordering in which magnetic moment at
each site is opposite to those of its surrounding. One can express AFM as

Si = ẑS(−1)ix +iy , (1.6)

the symbol (−1)ix +iy alternating between ±1 as the site i = (ix , iy ) changes
by one lattice constant. Two dimensional coordinates are used for convenience.
The magnetic moments are random as usual at high temperature, and begin to
take on the form, Eq. (1.6), below TN , which is called the Néel temperature
after Louis Néel who first discovered the existence of AFM.
What is the order parameter for P a Néel-ordered state? A naive application
of the prior definition S = (1/N ) i Si will give zero whether above or below
the transition. Rather, we should define a new order parameter
1 X
SN = (−1)i Si , (1.7)
N i

where the prefactor (−1)i = (−1)ix +iy has the effect to cancel out the alternat-
ing directions of spin alignment in Eq. (1.6). Expansion of the free energy in
terms of the Néel order parameter SN gives Eq. (1.5) with S → SN . From the
GL perspective, AFM and FM are the same entities5 .
Now imagine a peculiar kind of magnet, in which the magnetic moment
varies in the manner
     
2π 2π
Si = S cos ix x̂ + sin ix ŷ (1.8)
3 3
at low temperature. That is, spin direction rotates with a period of 3 lattice
constants when traversing the lattice along the x̂-direction. Once again we are
in 2 spatial dimensions. Can we write down an order parameter for such a
magnet and a free energy expansion in terms of it?
In defining an order parameter we should try to multiply each spin Si with a
suitable pre-factor so that, when ordered, each rotated spin S0i is pointing in the
same direction regardless of the site. Well, it’s pretty easy to guess the correct
answer in this case: you simply rotate each spin by the angle −2πix /3!
   
0 2πix 2πix
Six = cos Six + sin Siy
3 3
   
0 2πix 2πix
Siy = cos Siy − sin Six . (1.9)
3 3
The new spin S0i is always pointing in the x̂ direction in the ordered phase.
4 Some people call it Landau-Ginzburg (LG) free energy.
5 Laterwhen we study mean-field theory this equivalence between AFM and FM becomes
more exact.
1.4. ITS USES AND LIMITS 11

Having covered a progressly more complicated sort of magnets, I hope it is

clear to you that we need a systematic way to deal with a general case of mod-
ulation of spin at the ordering wave vector k. An arbitrary spin configuration
can be Fourier-expanded as
X
Si = Sq eiq·ri (1.10)
q

according to the rules of Fourier analysis. Ordering of spins with a particu-

lar
P wave vector k means that average of its inverse Fourier transform Sq =
−iq·ri
S
i i e is nonzero if and only if q equals k. Hence, the right order pa-
rameter to use is Sk . Since −k is equivalent to k after reflection, we should
expect S−k to appear as well. From the above equation it is also readily seen
that S−k = S∗k , in order to make Si a real number. A pair of vectors Sk and its
complex conjugate is sufficient to describe spin ordering at wave vector k.
In writing down the free energy we must first consider what kind of invariants
can be formed out of S and S∗ (We shall omit the redundant k subscript).
First a uniform rotation of spins changes Si to RSi , where R is an appropriate
rotation matrix, and consequently S to RS, and S∗ → RS∗ , without changing
the free energy. Invariants under a uniform rotation can be formed by taking
inner products such as S · S, S∗ · S∗ , S∗ · S, and higher-order terms such as
(S × S∗ ) · (S × S∗ ).
We have so far overlooked another important symmetry of the system, and
that is the symmetry under a lattice translation. For a change of reference
coordinates by any of the lattice vector R, we get

Sk → e−ik·R Sk , S∗k → eik·R S∗k . (1.11)

Out of the previously defined invariants, S·S becomes e−2ik·R S·S, and similarly
for its complex conjugates. This is in general not an invariant, unless 2k happens
to coincide with a reciprocal lattice vector, G. And only when this happens are
such terms allowed. Otherwise, we should not keep them. Such terms are called
“Umklapp” terms in analogy with the Umklapp processes of the solid state
physics.
Barring Umklapp terms we can write down the general free energy form as

1 λ µ
f= t|S|2 + |S|4 + (S × S∗ )2 + · · · . (1.12)
2 4 2

Allowing Umklapp processes gives us a further κ∗ S2 + κS∗2 . FM and AFM

are special cases which are examples of collinear spins. For collinear spins,
S × S∗ = 0. Furthermore, if k = (0, 0) (FM) or (π, π) (AFM), we have S = S∗ ,
and Eq. (1.12) reduces to Eq. (1.5).

1.4 Its uses and limits

In the previous section we have derived the free energy form assuming that it
depends solely on the order parameter in question. Ultimately, this statement
awaits verification from a more microscopic theory. In the next chapter we will
do this for the simplest of the magnetic models: the Ising model. In the mean
12CHAPTER 1. GINZBURG-LANDAU THEORY OF CRITICAL PHENOMENA

time we will proceed to work out the consequences of the GL free energy we
have just derived.
In accordance with the principles of statistical mechanics, a system that’s
realized at a given temperature T is the one that minimizes f . For Eq. (1.5)
this condition gives
S[t + λS2 ] = 0. (1.13)
The solution of this is either S = 0 or, if t < 0, S2 = −t/λ. S = 0 means
there is no macroscopic average of spins, and implies a paramagnetic p phase.
If S2 = −t/λ, there is a net magnetic moment of magnitude −t/λ. This
is a sensible description of the ordered phase provided that −t were in fact
proportional to Tc −T . Then t changes sign precisely at the critical temperature
Tc and the magnetic moment develops6 . Whether t behaves as T − Tc is once
again up to the microscopic theory to verify. In the case of Ising model this can
be derived rather straightforwardly.
In summary, the assumption of the free energy expanded in terms of an
order parameter, plus the dependence of the coefficient t as T − Tc is sufficient
to guarantee that there will be √ a phase transition. Moreover, the expected
temperature dependence is |S| ∼ −t, which is subject to a ready experimental
test. Such square-root dependence of the order parameter on the temperature
is found in a wide variety of systems.
For the general case of Eq. (1.12) the analysis gets more complicated, but
not its spirit. One differentiates Eq. (1.12) with respect to S and S∗ , which are
independent quantities, and solve for the two coupled equations. If there are
several solutions, the one which gives the lowest free energy is the true solution.
The interesting exercise of working out the phase diagram for Eq. (1.12) is left
as an exercise.
So, as we have seen, the beauty of GL theory is that it is capable of producing
a qualitative phase diagram of the system being probed without knowing much
of the microscopic details about it. All that’s required is that you identify the
right order parameter.
A drawback of GL theory, on the other hand, is that it entirely ignores the
effects of fluctuations. When you measure, say, the magnetization, it does not
always give the exact same value from measurement to measurement. You might
think that’s always the case, but no, that’s not always the case. If you measure
the charge of an electron using some probe, you do so expecting to find always
the same number regardless of how, and when you measure it. If two persons
got different numbers of the charge, that’s because the measurements were not
perfect. With the magnetization, on the other hand, I’m saying that you should
expect to get a different number each time you measure even if your apparatus
were operating in perfect order!
Let me explain this using the total magnetization Sz as an example. Its
squared average, hSz i2 , is in general different from the average of the square,
hSz2 i, because there is nothing in the mathematics to guarantee that they ought
2 2
to be the same. This in turns implies that p h(Sz − hSz i) i = h(∆Sz ) i 6= 0.
2
The square root of this quantity, ∆Sz ≡ h(∆Sz ) represents the degree of
deviation in the measured value of Sz from its average value hSz i. If ∆Sz is
much greater than hSz i, the very concept of average loses its meaning.
6 For t < 0 one must compare the free energies of the two solution, S = 0 and S2 = −t/λ.

The latter is shown to have a lower free energy by −t2 /4λ.

1.5. MULTIPLE ORDER PARAMETERS 13

We saw that hSz i vanishes continuously to zero near Tc . Unless ∆Sz vanishes
faster than hSz i near the critical temperature, we have no reason to trust the
predictions of the GL theory. In reality, the fluctuation ∆Sz grows larger as Tc
is approached. So, at some temperature TG (after Ginzburg) below Tc we have
∆Sz /hSz i ∼ 1 and the GL theory breaks down.
Ginzburg himself noted this fact and was able to estimate the temperature
at which his theory begins to break down. His method of estimation is known
as the Ginzburg criterion. GL theory is only trustworthy for temperatures
below TG , not below Tc ! In the following chapter we will discuss how to derive
Ginzburg criterion for the Ising model. In the mean time let me assure you that
in certain systems Tc − TG is a very small fraction of Tc , so that one can in
practice never distinguish the two temperatures. In that case, it’s OK to ignore
the fluctuations.

1.5 Multiple order parameters

Well, you have seen the uses and limits of the GL theory. Sufficiently removed
from the phase boundary where fluctuations are small, yet close enough to it
that the order parameter is still considered small, GL theory and its predictions
are expected to be good. Given that, the right attitude is to try to make the
most out of the GL theory before venturing upon anything more sophisticated.
A nice illustration of what GL theory can provide us is the following example,
taken from Kivelson’s paper on stripes7 .
Some high-Tc compounds show ordering of spins (in the manner we have
discussed in the previous sections) at some wave vector ks . The same compound
also shows a modulation of the electronic charge density with its own ordering
wave vector kc . Order parameters for spins has beenP defined previously, and
we shall call it S. For charge ordering, we define ρ = i δρi e−ikc ·ri where δρi
is the local modulation of the charge density. Now the task is to write the free
energy using these two order parameters. The easy part is to write the free
energy containing only one order parameter:
1 λs µs
fs = ts |S|2 + |S|4 + (S × S∗ )2 ,
2 4 2
1 2 λc 4
fc = tc ρ + ρ . (1.14)
2 4
Left to themselves, the two order parameters need not condense at the same
temperature, ts ∼ (T − Ts ), tc ∼ (T − Tc ), and Ts 6= Tc . The free energy arising
from interaction of the two order parameters may be written down by collecting
all terms consistent with the symmetry:

fint = κ1 [S · Sρ∗ + S∗ · S∗ ρ] + κ2 |S|2 ρ2 . (1.15)

The total free energy is a sum of three parts: f = fs + fc + fint .

Why are the terms we wrote down in fint above the only low-order terms
allowed in the free energy? First S is a vector, while ρ is a scalar. To form a
quantity which is invariant under rotation, S must take an inner product with
itself or its conjugate, S∗ . So, the lowest order terms allowed must have at least
7 Oron Zachar, S. A. Kivelson, and V. J. Emery, Phys. Rev. B 57, 1422 (1998)
14CHAPTER 1. GINZBURG-LANDAU THEORY OF CRITICAL PHENOMENA

two S’s and one ρ. This can be S · Sρ∗ , S∗ · S∗ ρ or |S|2 ρ. The last term is
forbidden by translational symmetry.
Under a lattice translation by R, we find S → e−iks ·R S, and ρ → e−ikc ·R ρ.
The combination |S|2 is invariant under this transformation, but |S|2 ρ becomes
e−ikc ·R |S|2 ρ. Unless kc equals one of the reciprocal lattice vectors, we cannot
use this term in the GL expansion. We are not considering Umklapp terms here,
and therefore kc must be zero. But that means there is no charge modulation!!
On the other hand, S · Sρ∗ transforms as ei(kc −2ks )·R S · Sρ∗ . Such a term is
OK provided the two modulation vectors are correlated: 2ks ≡ kc . Spins are
modulated along the same direction as the charge with a period which is twice
that of the charge.
In LSCO (La2−x Srx CuO4 ) family of cuprates, spin and charge modulations
are observed with the modulation periods in accordance with the predictions
of the GL theory. We were able to show from simple GL analysis that, should
there be a simultaneous ordering of spin and charge, and should there be any
interaction between the two, then it must happen that the modulation vectors
are related by the condition 2ks = kc .
The argument has a bit of a flavor of Dirac’s argument for the quantization
of electric charge. GL theory does not say there must exist a coupling between
spin and charge degrees of freedom (Dirac never said magnetic monopoles should
exist). But should there be an interaction between the two (If there is a magnetic
monopole somewhere), then they must obey the rule 2ks = kc (electric charge
must be quantized). The rule follows from translational symmetry (Dirac’s
quantization follows from gauge symmetry).
These days, more and more people are devoting their attention on com-
pounds that have complex ground states with ordering of spins, orbitals, and
even electric dipole moments. Microscopic description of this is undoubtedly
very hard. Even writing down the right Hamiltonian for such systems appears
to be very difficult, not to mention solving one. I believe GL theory for multiple
order parameters can do nicely here, and show some interesting results before
a heavier machinery can be brought to bear. And remember, now you know as
much about how to write down GL free energy as anybody else on earth!!
Chapter 2

Mean-field Theory

2.1 Mean-field theory of Ising model

For a given compound, some of the constituent atoms are magnetic, while others
are not. The magnetic atoms interact with each other for various reasons1 , and
the total energy of interacting spins usually takes the form
X
E= Jij Si · Sj . (2.1)
ij

Classical spins are vectors, of fixed magnitude |Si | = S, but having an arbitrary
orientation on the 3-sphere. Often the magnitude S is absorbed in the definition
of Jij and the rotational degrees of freedom is expressed as a uni-modular vector,
Si · Si = 1. Furthermore, we treat the case of nearest-neighbor interaction only,
so that Jij is non-zero if and only if i and j are the two adjacent sites of
the lattice. The sign of J can be either negative, in which case the model is
ferromagnetic, or positive (antiferromagnetic).
The discussion of phase transition in the previous chapter was based on
a bunch of assumptions: First you needed to define an order parameter, then
assume that free energy can be expanded in terms of it. Secondly, the coefficient
in front of the square term in the expansion had to change sign at some finite
temperature in order to ensure that phase transition does occur. How do we
know that all these things happen in a realistic calculation of the free energy?
To dispel doubts and place GL theory on a firmer ground we must turn to a
microscopic derivation of the free energy, and be able to prove the existence of a
phase transition from the model such as the one given in Eq. (2.1). But before
we do that, let us simplify the model even further.
What do the other, non-magnetic atoms do in a compound? Well, the wave
functions of the surrounding non-magnetic ions can affect the spin dynamics of
the magnetic atoms in various ways. As a result, sometimes the environment
around each magnetic ion favors the spin direction to lie along, say, ẑ direction
much more than along the x̂ or ŷ. In that case, Eq. (2.1) is not quite correct,
because it says that two spins lying in the x̂ direction is just the same as the
spins lying along the ẑ direction as far as their interaction energy is concerned
1 Details of the mechanism for the spin-spin interaction is outside the scope of this lecture.

15
16 CHAPTER 2. MEAN-FIELD THEORY

(rotational invariance). Instead, we should adopt the model

X X X
E = Jx Six Sjx + Jy Siy Sjy + Jz Siz Sjz (2.2)
ij ij ij

with |Jz |  |Jx |, |Jy |. Well, since Jx and Jy are tiny, we shall drop it altogether
in the first approximation and treat Jz only:
X
E = Jz Siz Sjz . (2.3)
ij

This is the celebrated Ising model, probably the most famous model in all of
many-body physics. From here and forward we shall drop the z altogether,
and treat the ferromagnetic case Jz = −J < 0 first. Due to the uni-modular
condition, Siz can only point either up (Siz = +1), or down (Siz = −1). It is
claimed that a simple model like the Ising model above is sufficient to capture
the qualitative features of the phase transition behavior.
Ising model is a many-body model in the sense that spin at site i interacts
with its neighboring spins at j each of which interacts with its own neighbors
at k, ad infinitum. In essence, two spins which are very very far apart manage
to affect, and be affected by, each other. In this regard, solving for the Ising
model is a lot like solving coupled, multi-variable equations. For the Ising
model, the number of variables equals the number of spins, which is a huge
number of Avogadroian scale. To get the answers you must know the values
of all the variables at once. In most cases, of course, exact, analytic answers
won’t be available, and one must rely on approximations. However, the nature
of approximation must be such that you can later make systematic improvement
should you wish to get a more accurate answer. And often times, a reasonable
approximation is the one based on sound physical picture. A general guiding
principle for reducing a hard, interacting problem to a simple, non-interacting
problem has been found a long time ago, and goes under the title of “mean-field
theory”2 .
Let’s look at the energy Eq. (2.3) from the perspective of a spin at one
particular site, i. Then the piece of the total energy that it cares about is given
by X
Ei = −Si (J Sj ). (2.4)
j∈i

Here we introduced j ∈ iPto indicate a sum over all nearest neighbors of i. For
a moment write hi ≡ J j∈i Sj , then Ei becomes −hi Si , which looks exactly
like the energy of a single, isolated spin subject to an external magnetic field
hi . Of course hi is not really an external field; it is a variable, that depends
on the spin values of sites j. If Sj changes, then so does hi . But rewritten in
this suggestive form, it shows that what the neighboring spins do is basically to
provide an effective magnetic field on the i-th spin.
Here comes the key logical loop:

2 Sometimes you hear words like “molecular field theory” or “Weiss theory” which mean the

same thing as the mean-field theory. I prefer to collect all the techniques, both classical and
quantum, which rely on reducing the many-body problem to a tractable one-body problem
under one roof, and call it mean-field theory.
2.1. MEAN-FIELD THEORY OF ISING MODEL 17

If spins do order, then Sj are more or less a fixed number, and so is hi . So

in effect, Si is subject to a more or less constant magnetic field which polarizes
the orientation of Si and gives rise to a finite average of Si .

The nature of the argument, as you can see, is self-consistent, and the self-
consistency is one of the defining characteristics of any mean-field theory.
So, let’s do make the replacement
X X
hi = J Sj → J mj , (2.5)
j∈i j∈i

where mj ≡ hSj i is the thermodynamic average, and define the mean-field

energy as X X
Emf = Ei = − hi Si . (2.6)
i i

In this form, the spins are no longer interacting with each other. Partition
function can be readily derived,
X
Zi = e−βEi = 2 cosh[βhi ]
Si =±1
Y Y
Zmf = Zi = (2 cosh[βhi ]). (2.7)
i i

Magnetization of the spin at each site follows immediately:

mi = hSi i = tanh(βhi ). (2.8)

We have thus managed to reduce the initial, difficult many-body problem to

something that involves only one site index i at a time. But wait.... We have
reduced the problem to something that looks simpler, but have we really solved
it yet? P
Remember that hi = J j∈i mj , and that mi = tanh(βhi ). From these we
derive
X
mi = tanh(βJ mj ), (2.9)
j∈i

which is a set of N coupled equations in N variables {mi }!! As I mentioned

earlier, solving Ising model is really lot like solving coupled, multi-variable,
and non-linear, equations! The many-body nature of the original problem is
embedded in the above, non-linear equations.
The only way to solve equations like these is to make an ansatz. For a
ferromagnet, fortunately, we already know that all spins are equal, mi = m,
and the above, coupled equations reduce to a single equation

m = tanh(βJzm) (2.10)

where z is the coordination number indicating how many neighbors there are for
a given site. For a square lattice z equals 4, and for cubic, z = 6. For generality,
we keep z in the expression. The solution of this equation can be found by
graphical means, or by two short lines of programming on Mathematica. You
18 CHAPTER 2. MEAN-FIELD THEORY

will find a familiar curve of m

√ as a function of T which saturates to 1 at low
temperature and vanishes as Tc − T near Tc ≡ Jz.
As advertised, mean-field theory successfully predicts ordering of spins at
low temperature, but has nothing to say about T > Tc , where hi = 0, and
Emf = 0. The same situation arises in GL theory where free energy simply
equals zero if the order parameter is zero. The analogy is not accidental. In the
following section we will derive GL theory starting from the mean-field theory
of the Ising model.

2.2 Ginzburg-Landau free energy from mean-

field theory
Once we are given a mean-field energy like Eq. (2.6) it is quite easy to write
down the corresponding free energy. But before doing that, let me re-write Eq.
(2.6) as
X 1X
Emf = − hi Si + hi mi . (2.11)
i
2 i
We are
P effectively re-adjusting the zero-point of the energy by introducing
(1/2) i hi mi . The reason for this change is that
X 1X X
hEmf i = − hi mi + hi mi = −J mi mj , (2.12)
i
2 i ij

while Eq. (2.6) without the adjustment will give twice this value as the energy.
So, Eq. (2.11) is right, and Eq. (2.6) is wrong. After correcting for this mistake,
free energy is derived from
X P Y
Zmf ≡ e−βFmf = e−βEmf = e−βJ ij mi mj (2 cosh[βhi ]), (2.13)
C i

and,
X X
Fmf = J mi mj − T ln[2 cosh(βhi )]
ij i
X X X
= J mi mj − T ln[2 cosh(βJ mj )]. (2.14)
ij i j∈i

This is the free energy derived in the mean-field theory. Taking advantage of
our knowledge about the ferromagnetic state we can set mi = m everywhere
and get (f ≡ F/N )

f = (Jz/2)m2 − T ln[2 cosh(βJzm)]. (2.15)

Free energy minimum is the one that minimize f . By plotting f as a function of

m (using Mathematica or other favorite math package of yours) one finds that
the minimum free energy occurs at m = 0 if T > Tc , with Tc = Jz, but at a
finite m if T < Tc .
We can make this observation more mathematically founded. Near the criti-
cal temperature the magnetic moments are small, allowing the Taylor expansion
2.3. ISING MODEL COUPLED TO MAGNETIC FIELD 19

of f in powers of m. The whole function is even in m, so only even powers of

m should appear. The result is (Tc ≡ Jz)
   3
Tc Tc Tc Tc
f ≈ 1− m2 + m4 + · · ·
2 T 12 T
1 Tc
≈ (T − Tc )m2 + m4 + · · · . (2.16)
2 12
which is precisely the form expected from the GL theory, with t = T − Tc , and
λ = Tc /3. Refer to Eq. (1.5) in the previous chapter for comparison.
Having another look at Eq. (2.16) one realizes that −(Tc /2)m2 is nothing
but the energy per site for the condensed spins. Then, from F = E − T S, it
follows that everything else in the expansion is due to entropy,
1 1
S ≈ − m2 − m4 . (2.17)
2 12
Clearly ordering is not good for the entropy. The (T −Tc )m2 term nicely express
this competition between energy and entropy effects. For T > Tc entropy domi-
nates energy, and vice versa for T < Tc . However, even below Tc entropy effects
are not completely gone, and it manifests itself as m4 term in the expansion.
The exact value of the magnetization below Tc is obtained as a compromise of
the two competing tendencies.
Once you have obtained the free energy, you can derive other thermody-
namic quantities such as the specific heat, magnetization, and the magnetic
susceptibility according to the formulas of thermodynamics3 .

2.3 Ising model coupled to magnetic field

We will generalize the mean-field treatment of the Ising model to include the
influence of site-dependent external magnetic field.

X 1X X
Emf = − hi Si + hi mi − Hi Si ,
i
2 i i
1X X
Fmf = hi mi − T ln (2 cosh[β(hi + Hi )]) ,
2 i i
X
mi = hSi i, hi = J mj . (2.18)
j∈i

The self-consistency equation governing mi is similar to the previous one,

mi = tanh[β(hi + Hi )], (2.19)

where now both the internal (hi ) and external (H) magnetic fields influence the
local moment mi .
The magnetization behavior for the uniform case mi = m, Hi = H, is
examined through
3 Unfortunately there is not enough time in this lecture to cover this more practical aspect

of the calculation. All textbooks on statistical mechanics discuss this point.

20 CHAPTER 2. MEAN-FIELD THEORY

 
m+h
m = tanh[βTc (m + H/Tc )] = tanh . (2.20)
t
where Tc = Jz stands for the mean-field critical temperature when H = 0, and
t = T /Tc the reduced temperature. For h = 0, the magnetization is nonzero
and grows with lowering temperature for T < Tc . For h > 0, the magnetization
remains positive and grows with lowering temperature. For h < 0, the magne-
tization remains negative and grows in magnitude with lowering temperature.
If one fixes the temperature at T > Tc and sweeps h from positive to negative,
the magnetization will change continuously from a positive, through zero, to
a negative value. If the temperature fixed is less than Tc , the magnetization
upon the field sweep remains positive for h > 0, then suddenly jumps over to a
negative value as h turns negative. The amount of magnetization jump is given
by twice the value determined by m = tanh[m/t] for a fixed t.
The magnetic susceptibility dm/dh is obtained from differentiating Eq. (2.20),

dm 1 1 − m2
= 2 = 2 . (2.21)
dh T cosh [(m + h)/t] − 1 m +t−1
For t > 1 the susceptibility never√diverges. For t < 1 the susceptibility diverges
at the point m2 = 1 − t, m = ± 1 − t.

2.4 Mean-field theory of XY spins

We go back to Eq. (2.2) to consider other limits of the model. Ising model
was derived as a special limit in which anisotropy of the crystal forced the spin
to preferentially lie along the ẑ-axis. What if the environment is such that
the preferred spin direction is away from the ẑ-axis? Then it is appropriate to
consider the planar model given by
X
E=J (Six Sjx + Siy Sjy ). (2.22)
ij

This model is also known as the XY model for an obvious reason. A spin is free to
rotate within the XY plane. What are the possible phases and thermodynamic
properties of this model?
As a first step, we again use mean-field theory to reduce the energy functional
Eq. (2.22) to a non-interacting form. Writing Si = (Six , Siy ), we get
X X 1X
E=J Si · Sj → Emf = hi · Si − hi · mi (2.23)
ij i
2 i
P
where mi = hSi i, and hi = J j∈i mj . OnceP again we have to make adjustment
for the zero-point energy ∆Emf = −(1/2) i hi · mi . I hope these steps are
familiar to you now. Once we have obtained the non-interacting model this
way, it is trivial to get the partition function (I0 =modified Bessel function)
Z 2π
Zi = dθi e−βhi ·Si = 2πI0 (β|hi |). (2.24)
0
2.4. MEAN-FIELD THEORY OF XY SPINS 21

Magnetization is obtained by differentiation:

∂ ln Zi I1 (β|hi |)
mi = − =− ĥi ≡ −R(β|hi |)ĥi . (2.25)
∂(βhi ) I0 (β|hi |)

We get the magnetization pointing opposite to the local field ĥi ≡ hi /|hi | be-
cause of the antiferromagnetic nature of the coupling.
How are these results different from the Ising spins? Well, not very much. If
you plot Zi obtained above, it looks pretty similar to cosh[x] which was the parti-
tion function for the mean-field Ising solution. The function R(x) ≡ I1 (x)/I0 (x)
looks formidable, but if you plot it, you won’t see a dramatic difference from the
tanh(x) curve we found for the Ising model. Finally, self-consistency condition
expressing hi in terms of magnetization becomes
X
hi + J R(β|hj |)ĥj = 0. (2.26)
j∈i

We have all the ingredients to construct the mean-field free energy for the
XY spins:
X 1X X
F = ∆E − T hi · mi − T
ln Zi = − ln[I0 (β|hi |)]
i
2 i i
J XX X X
= − mi · mj − T ln[I0 (βJ| mi |)]
2 i j∈i i j∈i
 
Jz 2
→ N× m − T ln[I0 (βJzm)] . (2.27)
2

The last line is derived assuming a uniform (staggered) magnetization, |mi | = m.

If you like, you can expand this in powers of small |mi | and derive the GL free
energy. Else, if you are a more practical person, you can ask Mathematica to
plot Eq. (2.27) vs temperature. You will find that F develops a minimum for
non-zero m when T is lower than Tc = Jz/2.
Either solving Eq. (2.25) or (2.26) gives the low-temperature ordered phase
of the XY model. In solving the mean-field equations, we have the following
possibilities.

• Ferromagnetic coupling: When J < 0, mean-field equation Eq. (2.25)

becomes
mi = R(β|hi |)ĥi = R(β|hi |)m̂i . (2.28)
Taking all the magnetic moments to be the same magnitude, we get

|m| = R(βJz|m|). (2.29)

The solution of this should be qualitative similar to the Ising model. We

have the critical temperature Tc = Jz/2.
• Antiferromagnetic coupling: On the square (cubic) lattice4 the spin direc-
tions alternate from site to site. Except for this, the magnetic moment
has the same temperature dependence as in the ferromagnetic coupling.
4 More generally, on any bipartite lattice.
22 CHAPTER 2. MEAN-FIELD THEORY

Now, we come to the important case of antiferromagnetic coupling on the

frustrated lattice. The new word on the table is frustration. It means the follow-
ing. The ideal situation for an antiferromagnet is if every spin is surrounded by
spins of exactly opposite sign, because such configuration lowers the exchange
energy JSi · Sj the most. There are cases where this condition cannot be sat-
isfied for all bonds of the lattice. A prime example is the triangular lattice,
where a quick inspection shows that not all three corners of a given triangle can
have all the spins anti-parallel. In this case, how are the spins going to arrange
themselves to minimize the total free energy?
Minimum free energy configuration of spins is the one that solves the coupled
equations, Eq. (2.25). On the triangular lattice, one finds that the spins make
a 120◦ angle with its neighbors. This is not quite as good as the full 180◦ angle,
but rather offers the most “democratic” solution for all the spins concerned.
The critical temperature for the continuous spin is half that of the Ising spin
for the same coordination number z and the exchange energy J: Tc = Jz/2
(XY), Tc = Jz (Ising). Because of the continuous nature of the spin, the
interaction strength between spins is effectively weaker for the XY spins, and
has a correspondingly lower energy. Other than these quantitative differences,
the predictions of the mean-field theory is very similar whether the spins are
discrete, or continuous. I haven’t done the mean-field calculation for the full
Heisenberg spin, Jx = Jy = Jz = J yet, but I bet the conclusions are again
similar, with a Tc that is presumably lower than in the XY model.
As you will discover in the next chapter, this does not mean that the behav-
ior of the spin systems are similar in reality. In fact, they are vastly different as
far as their critical properties are concerned. The critical exponents one derives
for Ising, XY , and Heisenberg models are different, putting the three models
into different “universality classes” as they are often called in the critical phe-
nomena literature. The reason why these models fall into different universality
classes despite their similarities at the mean-field level can be glimpsed at from
the nature of the excitations at finite temperature. For the Ising model, the
important excitations are the domain walls. For XY , it is the spin waves, and
the vortices. For Heisenberg spins, it is the spin waves, vortices, and hedgehogs.
Excitations of all kinds tend to destroy long-range order. If the spin waves
destroy long-range order below the mean-field transition temperature, then in
real systems the transition is driven by spin waves, and the critical properties
at Tc will not be those of the mean-field type. Similarly, if vortices drive the
phase transition well below the mean-field transition temperature, the critical
properties must be understood in terms of the proliferation of vortices.
Chapter 3

Fluctuations and
Excitations

Neither Ginzburg-Landau theory nor the mean-field theory describes in detail

how the long-range order gets destroyed at high temperatures. They are in
reality destroyed by the proliferation of elementary excitations, which are the
systematic deviations away from the ground-state configuration. Such excita-
tions are a consequence of the nature of the minimum free energy state at finite
temperature. Because F = E − T S, a certain amount of disorder implied by
a large S is favored at finite temperature T . The elementary excitations pro-
vide such a disorder, hence they are favored at higher temperature. The precise
nature of elementary excitations depends in turn on the nature of the spins.
Discrete, Ising spins have a completely different kind of elementary excitations
than those possessed by continuous spins. Even for continuous spins, the nature
of excitations changes depending on whether the spins are 2- or 3-dimensional.

3.1 Domain walls for discrete spin

For an Ising ferromagnet the ground state is obtained when all the spins are
pointing in the same direction, which we here take as +ẑ, Si = S ẑ. As we
raise the temperature other configurations begin to contribute to the partition
function. These “other configurations” are obtained by flipping some of the
spins to lie along the −ẑ direction, which raises the energy above that of the
ground state of perfect spin alignments. A border separating the up spins from
the flipped spins is known as the domain wall (Fig. 3.1). Presence of domain wall
leads to lowering of average magnetization, defined as (N↑ − N↓ )/N , where N ,
N↑ , N↓ are the total number of spins, up-spins, and down-spins, respectively.
At a sufficiently high temperature, proliferation of domain walls will lead to
N↓ ≈ N↑ , and the vanishing of magnetization, signaling the loss of long-range
order. It is in this sense that we attribute the proliferation of domain walls as
the mechanism for the loss of order in Ising ferromagnet.
Rudolf Peierls, in 1936, produced an elegant argument as to why in two
dimensional Ising ferromagnet the long range order must exist, by estimating
hN↓ i arising from the domain wall excitation. Let me sketch the argument

23
24 CHAPTER 3. FLUCTUATIONS AND EXCITATIONS

below1 .
A single domain wall of the perimeter s contains approximately (s/4)2 num-
ber of flipped spins. The total number of down spins is then
X  s 2
N↓ ∼ Ns P s , (3.1)
s
4

where Ns is the number of different ways in which domain walls of a given length
s can be formed, and Ps is the Boltzmann probability for the occurrence of such
a domain wall.

Figure 3.1: A domain wall configuration of spins in Ising model. Regions shaded
in blue have their spins reversed from the background spins.

A single domain wall costs an energy +2Js to excite, because there are a
total of s “bad bonds”, each of which costs an energy 2J. The probability of
thermally exciting such a configuration is given by Ps = e−β∆E = e−2βJs . This
is an exponentially small number at a sufficiently low temperature, so to a first
approximation we may assume that there is at most one such domain wall in
the entire system, leading to the estimate in Eq. (3.1).
The number of configurations Ns of a given perimeter s can be estimated
by noting that in constructing such a domain wall, one can proceed by laying
down a series of “sticks” connecting the nearest neighbors until all s sticks are
used up. For each step there are at most 3 different ways in which the stick can
be laid, and assuming that each move is independent of all previous moves, we
have
Ns ∼ N · 3s . (3.2)
The prefactor N arises from N different points from which a domain wall of a
given shape can originate. Thus,
∞  2  2 X∞
X s N d
N↓ ∼ N 3s e−2βJs = x xs
s=0
4 16 dx s=0
1 If you like to read a more rigorous argument, refer to Robert B. Griffiths, Phys. Rev.

136, A437 (1966).

3.2. SPIN WAVES FOR CONTINUOUS SPIN 25

 2
N d 1 N x(1 + x)
= x = . (3.3)
16 dx 1−x 16 (1 − x)3

We have introduced x ≡ 3e−2βJ . Therefore the average magnetization m is

2N↓ x(1 + x)
m=1− ∼1− . (3.4)
N 8(1 − x)3
At a sufficiently low temperature x is a small number, and we may approximate
m ≈ 1 − x/8 ≈ 1 − (3/8)e−2βJ . The low-temperature magnetization deviates
only by an exponentially small amount from its mean-field ground-state value,
proving the existence of long-range order in two-dimensional Ising model.
Historically this was a very important argument firmly establishing the ex-
istence of long range order for the first time in any many-body model. Onsager
later (1944) derived the partition function of the two-dimensional Ising model
exactly, and definitely laid the issue to rest.

3.2 Spin waves for continuous spin

Visualize throwing a small piece of rock into a pond in the pristine hours of the
morning and watching the ripples emanate from the point of the rock’s entry.
Pretty much the same thing happens in the system composed of many spins.
The ground state we have worked out in the previous chapter corresponds to
an utterly still surface of the pond. A disturbance, such as throwing a rock,
can cause ripples on the pond’s surface, which in the case of spin systems we
call “spin waves”. A spin wave is just a collective motion of the spins in the
same way that a ripple is a collective dance of all the water droplets composing
the surface. Various things, or “perturbations” can excite the spin waves. For
instance, a magnetic field applied locally over a small portion of the magnet will
tilt the spins in that region to lie along the direction of applied field. If you
suddenly turn it off, the spins will try to move back to its ground state position.
In so doing, they will necessarily cause the spins outside the region to dance
together with them because spins tend to interact with its neighboring spins.
How each spin will dance in coordination with the others is determined by
the Hamiltonian X
H = −J Si · Sj , (3.5)
ij

and the equation of motion for the individual spin Si that follows from it. We
first remind ourselves of the fundamental commutation relations of spin

[Siα , Sjβ ] = iδij αβγ Siγ (~ ≡ 1). (3.6)

From this one can calculate, for instance,

dSix X y X
x
= −i[Six , H] = −J Sj Siz − Sjz Siy = −J


Sj × Si . (3.7)
dt j∈i j∈i

Combined with the equation of motion for the y- and z-components, we get
dSi X
= −J Sj × Si . (3.8)
dt j∈i
26 CHAPTER 3. FLUCTUATIONS AND EXCITATIONS

Figure 3.2: A spin wave configuration

This equation is of a typical, nonlinear type because the quantities we must

solve for, Si , appears as a product on the r.h.s. In other words, one can never
solve the above equation exactly. However we must remind ourselves that it
is the small fluctuation away from the ground state that concerns us, and as
such we can decompose the spin operator as Si = hSi i + δSi = mi + δSi ,
where mi represents the ground state spin average. For a ferromagnetic ground
state, mi = S ẑ. A small fluctuation means |hSi i|  |δSi |. In making such a
comparison of magnitudes we must be careful to remember that mi is a number,
but δSi is not. So, it’s not altogether clear what exactly we mean by the
magnitude of δSi . Nevertheless, we can at least make the substitution Si =
mi + δSi as a matter of formal definition, and proceed to plug it into Eq.
(12.11).

d X


δSi = −J mj + δSj × mi + δSi
dt j∈i
X  X  X 
≈ −J mj × mi − J mj × δSi − J δSj × mi (3.9)
j∈i j∈i j∈i

2
ignoring terms of order δS . Once again, the justification is that δS is a small
quantity, and (δS)2 , even smaller. We take the quantization axis ẑ and write
mi = mẑ everywhere. Then2
X 
d
δSi = −Jzmẑ × δSi − Jm δSj × ẑ. (3.10)
dt j

d z
The ẑ-component of this equation reads dt δSi = 0, hence the fluctuation in the

2 I apologize for the multiple use of z here: z is the number of neighbors, or the coordination

number, whereas ẑ is the unit vector along the z-axis.

3.2. SPIN WAVES FOR CONTINUOUS SPIN 27


ẑ-component of spin is zero, Siz t = m. For the x̂-and ŷ-components we get
d x X
S = JzmSiy − Jm Sjy ,
dt i j∈i
d y X
S = −JzmSix + Jm Sjx . (3.11)
dt i j∈i

We have dropped the δ symbol in front. Introducing the complex notation

Zi ≡ Six + iSiy simplifies the equation a lot:
 
dZi X
= iJm  Zj − zZi  . (3.12)
dt j∈i

We will solve this equation using the ansatz

Zi = Z0 eik·ri −iωk t , (3.13)

which yields the self-consistency condition

 
X
ωk = Jm z − eik·(rj −ri )  . (3.14)
j∈i

For a 3-dimensional cubic lattice,P j∈i eik·(rj −ri ) = 2(cos kx + cos ky + cos kz ),
P
and with z = 6, ωk = 2Jm(3 − α cos kα ). For 2D square lattice we will get
ωk = 2Jm(2 − cos kx − cos ky ). This is the desired spin wave energy for a spin
disturbance associated with a wavevector k.

Holstein-Primakoff theory 3 : There is a neater way of deriving the same

spin wave spectrum we have just derived above, using the method of Holstein-
Primakoff (HP) representation of spins. First, I will introduce the representation
(ni = b+
i bi ),

Si+ = Six + iSiy = (2S −ni )1/2 bi ,

Si− = Six − iSiy = b+
i (2S −ni )
1/2
,
Siz = S − ni , (3.15)

then verify that the funny new expressions on the r.h.s. of the above equations
satisfy all the commutation algebras of ordinary spin operators: [Siα , Sjβ ] =
iδij αβγ Siγ , provided the new operator bi is a canonical boson operator, with
the familiar bosonic commutation relations, [bi , b+
j ] = δij , etc.
How this comes about does not concern us so much here; rather we will
proceed to apply the brand new technique and see if something good comes out.
First rewrite
1
Si · Sj = Siz Sjz + (Si+ Sj− + Si− Sj+ )
2
1 1 1
= (S − ni )(S − nj ) + (2S −ni ) 2 (2S −nj ) 2 (b+ +
i bj + bj bi ). (3.16)
2
3 This section is largely optional on first reading.
28 CHAPTER 3. FLUCTUATIONS AND EXCITATIONS

OK, so it doesn’t appear particular illuminating after the substitution. But

let’s try to appeal to our physical senses to see if some simplifications can occur.
We already know the ground state of the Hamiltonian is given by a ferromagnetic
arrangement of spins, with all the spins pointing in the ẑ-direction. If we take
the average of the operator Siz in the ground state, at zero temperature, we
should get hSiz i = S. The same can be said of the r.h.s. of the last of Eq.
(12.28) because it is just a mathematically equivalent way of writing down the
same quantum-mechanical operator, so we must get hS − ni i = S, or hni i = 0!
This means that the quantum-mechanical ground state |0i of the ferromagnetic
Heisenberg spin model is defined by the condition that h0|ni |0i = 0 for all sites i.
An excitation is just a small wiggle of the spins away from the absolutely frozen
spin configuration that is the ground state, hence we can suppose that hSiz i
deviates only slightly from its ground state value S, or in the language of bosons,
that ni is some small, but non-zero number. In this limiting√ circumstance,
√ we
can simplify the square roots in Eq. (12.28) by taking 2S − ni ≈ 2S, and
writing
1
Si · Sj ≈ (S − ni )(S − nj ) + · 2S(b+ +
i bj + bj bi ).
2
≈ S 2 − S(ni + nj ) + S(b+ +
i bj + bj bi ). (3.17)

The ni nj is a product of two small numbers, so we delete them. The fully

HP-transformed Hamiltonian, after the approximations we have just made, will
be X
HHP ≈ −JS (b+ + + +
i bj + bj bi − bi bi − bj bj ). (3.18)
ij

Well, now the new Hamiltonian is just a quadratic Hamiltonian involving only
a product of two boson operators. Such operators can be diagonalized easily, in
this case by going to the Fourier space:
X
bi = eik·ri bk . (3.19)
k

Substitution of the Fourier expression for bi into Eq. (12.31) immediately gives
X
H= ωk b+
k bk (3.20)
k

where

ωk = JSz(1 − γk )
X
γk = z −1 eik·(rj −ri ) . (3.21)
j∈i

The eigenenergy ωk agrees with the spin-wave spectrum derived earlier, in Eq.
(12.10), using the equation-of-motion theory.
We have derived the same spin wave excitation energy using two completely
disjoint methods. The real power of the HP boson theory lies in the simplicity it
offers in dealing with excitation spectra of complicated magnetic systems, such
as those of an antiferromagnet. The strategy is essentially the same, but the
technique is more advanced. The following segment is for those of you who like
3.2. SPIN WAVES FOR CONTINUOUS SPIN 29

the technical challenge, dealing with the application of HP theory to diagonalize

the spin wave spectrum of an antiferromagnet on a square lattice.
First of all, we define a new quantum-mechanical operator Siy , Siz related to
y
Si , Siz by the relation

Siy + iSiz = eiθi (Siy + iSiz ) (3.22)

where θi is the angle of the classical ground state spins, θi = π(ix + iy ). This
represents a rotation of spins, with respect to the x-axis, by the angle that just
equals the orientation of the classical ground state spins. The x-component
remains unchanged, hence Six = Six . Furthermore, one can easily verify that
the new operators Siα satisfy all the commutation rules of the original spins, Siz ,
hence they are as good a representation of the local spin operator as the previous
one we used. In terms of Si , we have the antiferromagnetic Hamiltonian written
as
X JX + + X
Si Sj + Si− Sj− − J Siz Sjz .


H=J Si · Sj → (3.23)
ij
2 ij ij

Note that we have two raising (lowering) operators instead of one raising and one
lowering, as was the case before. The ground state of the new Hamiltonian, Eq.
(12.36), is given by the rotation of the antiferromagnetic spins, hSiz i = Seiθi .
But this implies that in the new representation, hSiz i = S everywhere! Since
hSiy i = 0 initially, we have hSiy i = 0, too. So the new Hamiltonian has a
ferromagnetic ground state with all the spins pointing in the +ẑ direction, and
we can make the assumption hSiz i ≈ S, or hni i  S after the HP substitution,
for the small-fluctuation analysis.
Under these assumptions the HP-transformed Hamiltonian becomes
X
H ≈ JS (b+ +
i bj + bi bj + ni + nj ). (3.24)
ij

Unlike its ferromagnetic counterpart, Eq. (12.31), this Hamiltonian has a prod-
uct of two annihilation (creation) operators. As before, we take the Fourier
transform and obtain
Xh γk + + i
H = JSz b+
k bk + (bk b−k + bk b−k ) . (3.25)
2
k

Whereas doing the Fourier transform was sufficient to yield excitation en-
ergies in the previous example of ferromagnetic spins, now the Hamiltonian
written in momentum space fails to give eigenenergies. The final trick lies in
working out the equation of motion of the boson operators, which gives

[H, bk ] = −JSz bk + γk b+


−k . (3.26)

Here it’s clear why we could not get the energy from Fourier transform alone.
The dynamics of the operator bk is coupled to that of b+ −k , and the dynamics of
b+
−k is coupled to that of bk by the equation

[H, b+ +


−k ] = JSz b−k + γk bk . (3.27)

Noting that [H, X] = −idX/dt in quantum mechanics, and that for the eigen-
modes we can re-write the time derivative as −iE, where E is the eigenenergy,
30 CHAPTER 3. FLUCTUATIONS AND EXCITATIONS

the coupled oscillator equation we must solve becomes

    
bk 1 γk bk
Ek = JSz . (3.28)
b+
−k −γk −1 b+
−k
p
Eigenvalues are readily found as Ek = ±JSz 1 − γk2 . Here I state, without
proof, that the negative energies are just artifacts of the theory, and that only
the positive branch matters. Hence, the spin p waves in the antiferromagnetic
background carries the energy given by JSz 1 − γk2 for a wave vector k.

3.3 Topological excitations

In the previous two sections we have reviewed several kinds of excitations that
can arise in spin systems. They are states which are higher in energy than
the ground state, but are nonetheless important at finite temperatures because
thermal fluctuations produce them with a probability e−β∆E . For a discrete spin
system excited states can be achieved by flipping some spins, creating domain
walls. For continuous spins, excitations can be a small rotation of the spin angle
from its ground state direction - the spin waves.
There are another class of important fluctuations that arise in continuous
spin systems, in particular for XY , or planar spins. The XY model, as it is
called, is achieved as a limit Jz  Jx = Jy = J in the ferromagnetic Heisenberg
model:
X
Six Sjx + Siy Sjy .


HXY = −J (3.29)
ij

Each spin is like the arms of a clock, free to rotate within the plane. Indeed, a
spin can be characterized by an angle as (Six , Siy ) = (cos θi , sin θi ) if we normalize
the magnitude to be one. In terms of the angle variable the XY model becomes
X
HXY = −J cos(θi − θj ). (3.30)
ij

There are of course spin wave excitations in this model. Their properties can
be understood using the techniques already explained in the previous section.
Another important class of excitations can be found in the planar model which
are known as the “vortices”. A vortex is a configuration of spins given by
 
yi − y0
θi = ± arctan . (3.31)
xi − x0

It is a rotation of the spin directions in the (counter)clockwise direction as

depicted in Fig. 3.3, about the center located at (x0 , y0 ). On following a closed
loop encircling the origin, the angle of the spins has rotated by ±2π. A non-zero
circulation, as it is called, is the hallmark of a vortex.
This is a dramatic departure from the spin wave pattern displayed in Fig.
3.2 which involves slight twist of the spin direction from its ground state con-
figuration. Although both refer to the deviation of the spin direction, a class
3.3. TOPOLOGICAL EXCITATIONS 31

Figure 3.3: A vortex configuration of spins in XY model

of excitations involving a non-zero circulation are commonly known as vortex

excitations to distinguish it from the smooth spinwave-like excitations.
Creating a single vortex costs an energy of order J log(L/a) where a refers
to the lattice spacing, and L is the linear dimension of the entire system. This
can be obtained by inserting Eq. (4.22) into Eq. (3.30) and subtracting off the
ground state energy from it. It is a big cost in energy, diverging to infinity as we
try to create it for a larger and larger system size L. So, why do we even bother
to consider such an expensive excitation? After all, high energy excitations are
exponentially suppressed due to Boltzmann weight.
Again, the main consideration for whether an excitation is important or not
is not in terms of its energy, but its contribution to the free energy F = E − T S.
We just found out that E = J log(L/a) for a single vortex. Now what about S?
A vortex is characterized by its center coordinate (x0 , y0 ), and no matter
where it is placed in the system it costs the same amount of energy. In other
words, the degeneracy associated with the creation of a single vortex roughly
grows as (L/a)2 , which are the number of different ways in which the center of
a vortex can be located. The logarithm of this is just the entropy, so we have
S ∼ log(L/a)! In total, the vortex contribution to the free energy is4

Fv = (J − T ) log(L/a). (3.32)

As long as the temperature is lower than Tc ≡ J, we are better off without a

vortex, because the energetic cost to create one well surpasses the entropy gain.
For T > Tc , however, the situation is reversed, and we gain free energy through
the proliferation of vortices!
Think of a situation where we have a proliferation of vortices in the lattice.
Clearly, the spins will get disordered because depending on where the vortex
is located, the nearby spins can point in any direction. Averaging over the
4 I’m being sloppy about constants of order unity multiplying both J and T here.
32 CHAPTER 3. FLUCTUATIONS AND EXCITATIONS

vortex coordinate will lead to null magnetization5 . The simple thermodynamic

argument just presented shows that vortices can be important excitations -
a driving mechanism for a phase transition - for the planar XY spins in two
dimensions. The theory for the nature of vortex-driven phase transition was
worked out by Kosterlitz and Thouless in 1973 and bears the name of KT
transition6 .
In the full 3-dimensional Heisenberg model there are even more elaborate
kinds of spin excitations. One example of it is the “skyrmion” excitation illus-
trated in Fig. 5.1. In 3 dimensions however the skyrmions do not drive the
phase transition to a non-magnetic phase, and not as important as the vortices
in two dimensions.

Figure 3.4: A skyrmion configuration of spins in Heisenberg model

5 There is a subtlety involved, because long-range order of spins is already lost even before

we consider effects of vortices. Spin wave excitations turn out to be sufficiently violent to
destroy all long range order. So a more precise statement is that proliferation of vortices
above Tc makes the spins from being disordered to even more disordered. This is therefore
quite different situation from the usual kind of transition with the order parameter vanishing
to zero.
6 Some people call it BKT transition, named after another physicist, V. L. Berezinskii, who

worked on the same model a few years earlier than Kosterlitz and Thouless. KT knew about
Berezinkii’s work, but that did not stop them from working out their own, and more complete
theory of the phase transition in the planar model. Thouless himself calls it a “topological
phase transition” in his book.
Chapter 4

XY Model

4.1 Definition
The XY model, as it is commonly known, is given by the classical energy function
X
EXY = −J cos(θi − θj ), (4.1)
ij

where J can be of either sign, and ij spans all the nearest neighbors of the lattice.
The model can be defined in arbitrary dimension, with the two-dimensional case
being the most interesting in view of its critical properties.
The PXY model can be obtained as the planar limit of the spin model,
H = J ij Si · Sj where each spin is represented by the planar component
Si =(cos θi , sin θi ). The signs of J then corresponds to ferromagnetic (J > 0) or
antiferromagnetic (J < 0) spin exchanges, respectively.
The energy function, Eq. (4.1) is invariant under the global rotation of
all spins, θi → θi + α. In addition, it is periodic in each variable θi with
the periodicity of 2π. This latter symmetry reflects the angular nature of the
variable θi . In dealing with the statistical mechanics of the XY model, one often
defines the classical action and the partition function as

X
S = K (1 − cos[θi − θj ])
ij
!
Z 2π Y
Z = dθi e−S . (4.2)
0 i

The constant K plays the role of the inverse temperature, K = J/T .

The low-temperature phase of this model is prohibited from possessing long-
range order, due to the Mermin-Wagner theorem which states that the abun-
dance of soft fluctuations (Goldstone modes) will destroy the long-range ordered
state of continuous order parameter in two or lower dimensions. The spin-wave
theory, to be presented in the next section, explicitly demonstrates that the
spin-spin correlation at sufficiently low temperature is quasi-long-range ordered
(QLRO), in the manner that

33
34 CHAPTER 4. XY MODEL

Figure 4.1: A phase diagram of the XY model in two dimensions. A critical

point at T = TKT separates the low-temperature critical phase from the high-
temperature disordered phase.

1
hSi · Sj i ∼ hei(θi −θj ) i ∼ η(T )
. (4.3)
rij
Spontaneous symmetry breaking, which would imply hSi ·Sj i → m2 as rij → ∞,
clearly does not occur.
At a sufficiently high temperature, T  J, or K  1, one must arrive at
a disordered phase characterized by the exponential decay of the correlation
function

e−rij /ξ
hSi · Sj i ∼ . (4.4)
rij
The behavior of the two correlation functions, Eqs. (4.3) and (4.4), are qualita-
tively different. A phase transition separating the two regimes is expected at a
critical temperature, TKT . The phase diagram of the XY model is qualitatively
described in Fig. 4.1.
In Sec. 4.2, we present the spin-wave theory which describes adequately
the low-temperature behavior of the model. Then in Sec. 4.4 we introduce the
concept of vortex excitation and the vortex-drive phase transition known as the
Kosterlitz-Thouless (KT) transition.

4.2 Spin-wave Theory

Strictly at zero temperature, the ground state of the XY model is the ferromag-
netic state where all the spins lie along a fixed axis, θi = θ0 . For finite but very
low temperature, T  J, one would expect the fluctuations or small deviations
of the spin angle from the ground state value θ0 . Such small fluctuations in the
spin direction are known as the spin wave. In the spin wave approximation or
the Gaussian approximation, Eq. (4.1) becomes
JX KX
E≈ [θi − θj ]2 , S ≈ [θi − θj ]2 . (4.5)
2 ij 2 ij

Although the global symmetry under θi → θi + α is still preserved, the lo-

cal invariance of the action under θi → θi + 2π is lost in the above Gaussian
approximation.
angle θi , which now extends over −∞ to ∞, can be Fourier-transformed
The P
as θi = k θk eik·ri , and

XX 0
h ih 0
i
[θi − θj ]2 = ei(k−k )·ri eik·(ri −rj ) − 1 e−ik ·(ri −rj ) − 1 θk θ−k0 , (4.6)
k k0
4.2. SPIN-WAVE THEORY 35

0
ei(k−k )ri = δkk0 , Eq. (4.5) can be rewritten as
P
Because i

K XX K XX
S= [θi − θj ]2 = (1 − cos[k · rij ]) θk θ−k . (4.7)
4 i j∈i 2 j∈i k

An abbreviation j ∈ i is introduced to indicate all nearest neighbor site j with

respect to i. Since θi is a real variable, we also have θ−k = θk∗ . We will introduce
D independent unit vectors eα to denote the nearest neighbor displacement in
D dimensional lattice and write
D
XX
S=K (1 − cos[k · eα ]) |θk |2 . (4.8)
k α=1

In evaluating the spin-spin correlation function, it is convenient to introduce

the complex quantity

Si = Six + iSiy = eiθi , (4.9)

with which one has
1 2
hSi Sj∗ i = hei[θi −θj ] i = e− 2 h(θi −θj ) i . (4.10)
The last equality follows due to the quadratic nature of the classical action S,
Eq. (4.5). The thermodynamic average of h(θi − θj )2 i is evaluated as

XX 0
h ih 0
i
h(θi − θj )2 i = ei(k−k )·ri eik·(ri −rj ) − 1 e−ik ·(ri −rj ) − 1 hθk θ−k0 i.
k k0
(4.11)

The average hθk θ−k0 i is non-zero only if k = k0 , for which we get

1 1
hθk θ−k i = ,
2K D
P
α=1 (1 − cos[k · eα ])
1 X 1 − cos[k · (ri − rj )]
h(θi − θj )2 i = PD . (4.12)
K α=1 (1 − cos k · eα )
k
P
Contribution to the sum k is dominated by the small-k region, where one can
PD
approximate α=1 (1 − cos k · eα ) ≈ k2 /2,

d2 k 1 − cos[k · (ri − rj )]
Z
2
h(θi − θj )2 i = . (4.13)
K 4π 2 k2
Integrating over all angular orientations of ri − rj with the same magnitude rij
gives
Z
1 dk
h(θi − θj )2 i = [1 − J0 (krij )] . (4.14)
πK k
One can separate the domain of integration into two: 0 < k < 1/rij , and 1/rij <
k < 1/a. The first integration gives a numerical factor which is independent
36 CHAPTER 4. XY MODEL

of rij and can be ignored. For the second integration one can ignore the fast-
oscillating J0 (krij ) and obtain

Z 1/a
1 dk 1 r 
ij
h(θi − θj )2 i ≈ ≈ ln ,
πK 1/rij k πK a
1
  2πK   2πJ T

i[θi −θj ] a a
he i≈ = . (4.15)
rij rij

This calculation indeed shows that the thermally activated small spin fluctuation
is capable of destroying the long-range order of spins at arbitrary non-zero
temperature. Note, however, the critical exponent η(T ) ≡ T /2πJ controlling
the rate of decay becomes smaller and eventually becomes zero as T → 0.

4.3 High-temperature expansion

As we saw in the previous section the spin-wave analysis gives the spin-spin
correlation that falls off with separation in the power-law manner. At a very
high temperature, T  J, another type of scheme for calculating the spin-spin
correlation function becomes accurate.
P
The Boltzmann weight, exp(K ij cos θij ), can be Taylor expanded assum-
ing K ≡ J/T  1,
 s  s
∞ ∞  s X
X X Ks X X 1 K
eiθij + e−iθij  .
 
exp(K cos θij ) =  cos θij  = 
ij s=0
s! ij s=0
s! 2 ij
(4.16)
We can think of each factor eiθij as a directed bond going from site i to site
j. Eventually, one will integrate over all the angles θi in the evaluation of
the partition function, and the only non-vanishing contribution to the partition
function arises from products of the form
Y
eiθij (4.17)
ij

where the sequence of ij bonds forms a closed loop. Evaluation of the numerator
involves the integration of the function
   s
∞  s
X X 1 K X
ei(θI −θJ ) exp K eiθIJ  eiθij + e−iθij  .
 
cos θij  =
ij s=0
s! 2 ij
(4.18)
By the same reasoning as above, the only non-vanishing contribution arises from
the path that joins the two points I and J. If the separation rIJ is about m,
then the nonzero integral arises from s ≈ m in Eq. (4.18) and there are roughly
m! different ways to connect the two end points, leading to the estimate of the
numerator  m
K
NUMERATOR ≈ . (4.19)
2
4.4. VORTEX EXCITATION AND KOSTERLITZ-THOULESS TRANSITION37

The denominator, at each order s of the expansion, has roughly s! different ways
to produce the non-zero combination of bond angles, and gives
∞  s
X K 1
DENOMINATOR ≈ ≈ ≈ 1. (4.20)
s=0
2 1 − K/2

Thus the spin-spin correlation function at high temperature goes as

 rij
K
∼ e−rij ln( J ) = e−rij /ξ .
2T
iθij
he i ∼ (4.21)
2

Here ξ = 1/ln(2T /J) defines the correlation length. Higher the temperature,
shorter the correlation length as expected.

4.4 Vortex Excitation and Kosterlitz-Thouless

transition
Another important class of excitations can be found in the planar model which
are known as the “vortices”. A vortex is a configuration of spins given by the
angular pattern  
yi − y0
θi = ± arctan . (4.22)
xi − x0
It gives a rotation of the spin directions in the (counter)clockwise direction as
depicted in Fig. ??, about the center located at (x0 , y0 ). If one follows a closed
loop encircling the origin, then one will discover that the angle of the spins has
rotated by ±2π. More directly, one can see this by taking the gradient

 
−(yi − y0 ) xi − x0
∇θi = ± , , (4.23)
(xi − x0 )2 + (yi − y0 )2 (xi − x0 )2 + (yi − y0 )2
R
and then integrating it ∇θi · dr along a closed path encircling the center
(x0 , y0 ). The quantized value ±2π is independent of the path chosen, as long as
the different paths enclose the same origin. A finite circulation, as it is called,
is the definition of a vortex, and defines the “topological” (path-independent)
property of this peculiar excitation.
Sufficiently far from the core of the vortex where the change of the spin angle
is too swift between nearby spins to allow a continuum description, one can still
apply Eq. (4.5) to evaluate the energy cost of creating a single vortex.

Z Z L  
J 2 2 J rdr L
E≈ [∇θ(r)] d r ≈ × 2π = πJ × log . (4.24)
2 2 a r2 a

The lower cutoff a represents the typical size of the core where the continuum
approximation would fail, and L is the linear dimension of the lattice we are
considering. It is a big cost in energy, in principle diverging to infinity as we
try to create it for a larger system size L. Making such an excitation would be
suppressed due to the tiny Boltzmann weight associated with it.
38 CHAPTER 4. XY MODEL

However, considering whether an excitation is important or not is made

in terms of its contribution to the free energy F = E − T S. The “entropy”
S associated with a single vortex can be evaluated as follows. A vortex is
characterized by its center coordinate (x0 , y0 ), and no matter where it is placed
in the system it costs the same amount of energy. In other words, the degeneracy
associated with the creation of a single vortex roughly grows as (L/a)2 , which
are the number of different ways in which the center of a vortex can be located.
The logarithm of this is just the entropy, so we have S = 2 log(L/a). In total,
the vortex contribution to the free energy is

Fv = (πJ − 2T ) log(L/a). (4.25)

As long as the temperature is lower than Tc ≡ TKT πJ/2, we are better off
without a vortex, because the energetic cost to create one well surpasses the
entropy gain. For T > Tc , however, the situation is reversed, and we gain free
energy by proliferating the system with vortices.
A second type of consideration leads to the same estimate of the critical
temperature. Imagine creating a pair of vortices each having the opposite cir-
culation, or vorticity. For a size much larger than the separation of the vortex
cores all the spins point in the same direction without an energy cost. Then one
can estimate the energy cost of creating a vortex-antivortex pair as 2πJ ln(r/a),
where r now represents the separation of the two cores. Then the mean separa-
tion of the vortices is calculated to be

R∞ 3
2 r dre−2πβJ ln(r/a) πβJ − 1 TKT − T /2
hr i = Ra∞ 3 −2πβJ
= = . (4.26)
a
r dre ln(r/a) πβJ − 2 TKT − T

The mean separation grows and eventually diverges at T = TKT , so we may

think of TKT as the temperature at which the vortex and antivortex unbind
and become “free.” Below TKT , one has a collection of vortex-antivortex pairs,
or vortex dipoles, due to the thermal excitation. These pairs will affect the
spin-spin correlation in the low temperature phase in addition to those already
given by the spin wave calculation p as in Eq. (4.3), on a length scale that is less
than the typical size of the dipole, hr2 i. The ultimate long-distance behavior
is still controlled by the small angle fluctuation of spins.
However, as T > TKT , the vortices and antivortices which became free can
destroy the spin-spin correlation much more effectively than the spin waves.
Thus the correlation becomes exponentially correlated, in the manner of Eq.
(4.4). Although this type of short-range correlation is common for paramagnetic
states of other magnetic systems, the uniqueness of the 2D XY model is that
the short-range correlation is caused by the proliferation of vortices.
There is a subtlety involved, because long-range order of spins is already lost
even before we consider effects of vortices. Spin wave excitations turn out to be
sufficiently violent to destroy all long-range order. So a more precise statement
is that proliferation of vortices above Tc makes the spin from being disordered to
even more disordered. This is, therefore, not the same as the usual kind of tran-
sition with the order parameter vanishing to zero. The simple thermodynamic
argument just presented shows that vortices are indeed important excitations
- a driving mechanism for a phase transition - for the planar XY spins. An
4.5. HELICITY MODULUS 39

elaborate theory for the nature of vortex-driven phase transition is worked out
by Kosterlitz and Thouless in 1973 and bears the name of KT transition.
A more thorough analysis of the KT transition involves the use of the real-
space renormalization group analysis, which lies outside the scope of this thesis.

4.5 Helicity Modulus

Due to the lack of order parameter in the conventional sense, determining the
critical temperature of the KT transition is a bit tricky. For ordinary order
parameters, Binder cumulant method is probably the most powerful method to
fix the critical temperature in the thermodynamic limit. In determining the
KT transition temperature TKT , the best and the most popular method is by
working out the helicity modulus ρs (T ). The definition of the helicity modulus
dates back to a paper by Fisher, Barber, and Jasnow1 or even earlier, and
one should refer to their paper for a physical foundation of the quantity. It is
alternatively known as the superfluid density (in the context of the theory of
two-dimensional superfluids) or phase stiffness. Although it is most commonly
used to study the transition in XY spin model, the definition itself is general
enough to accommodate Heisenberg spin models and some studies of phase
stifness for Heisenberg spins exist in the literature2
We phrase the discussion of the helicity modulus in the broader context of
the XXZ spin Hamiltonian on a square lattice

X JX + −
H = −Jz Siz Sjz − (Si Sj + Si− Sj+ ), (4.27)
2
hiji hiji

with Si± = Six ± iSiy . Jz = 0 corresponds to the XY spin model. Now a phase
twist θ is imposed across the x-direction,

Si+ → Si+ eiθxi /L . (4.28)

The difference in H is evaluated to second order in θ:

J X iθ/L −
H[θ]−H[0] = − (e − 1)Si+ Si+x + c.c.
2 i
J iθ X + − Jθ2 X + −
≈− (Si Si+x − Si− Si+x
+
)+ (Si Si+x + Si− Si+x
+
). (4.29)
2L i 4L2 i

At this point it is already clear that the difference between XY model and XXZ
model disappears and one can use the same definition of the helicity modulus
for both spin systems. Difference in free energy evaluated to second order in θ
gives
1 Michael E. Fisher, Michael N. Barber, and David Jasnow, Phys. Rev. A 8, 1111 (1973).
2 For instance, see N. Schultka, Phys. Rev. B 55, 41 (1997).
40 CHAPTER 4. XY MODEL

Dθe−βH[θ]
R
Z[θ]
=R = e−β(F [θ]−F [0])
Z[0] Dθe−βH[0]
!2
βJθ2 X + − − + β 2 J 2 θ2 X + − − +
=1− h (Si Si+x + Si Si+x )i − h (Si Si+x − Si Si+x ) i,
4L2 i 8L2 i

!2
Jθ2 X − βJ 2 θ2 X + −
F [θ]−F [0] ≈ 2 h (Si+ Si+x + Si− Si+x
+
)i + 2
h (Si Si+x − Si− Si+x
+
) i.
4L i
8L i
(4.30)

Helicity modulus follows from the definition F [θ] − F [0] = (ρs /2)θ2 :

!2
J X J2 X
ρs = 2 h (Six Si+x
x
+Siy Si+x
y
)i− 2
h y
(Six Si+x −Siy Si+x
x
) i. (4.31)
L i
T L i

4.6 Binder cumulent

Binder cumulent is the method of finding the transition point in 2nd order phase
transition. The formula is written by

hM 4 i
UL = 1 − , (4.32)
3hM 2 i2

where M is a order parameter. By using this formula, we can find the transition
temperature for chirality and Ising orders. If you calculate the Binder cumulent
for different lattice size, the line is cross at some point. This crossing point is
the phase transition temperature and this is the same effect to get the critical
temperature at thermodynamic limit.

4.7 Villain transformation

The treatment of the vortices and their interactions becomes simpler through
the transformation of the original partition functioni of the XY model first
invented by J. Villain. It also makes explicit the important concept of “duality
mapping,” by which the vortex degrees of freedom becomes the primary variable
in the transformed partition function. The duality transformation for the 2D
XY model proceeds as follows. The classical 2D ferromagnetic (FM) XY model
assumed the following partition function
Z 2π Y ! X
Z= dθi exp[−βJ (1 − cos[φi − φj ]). (4.33)
0 i hiji

Define βJ ≡ K, and φi − φj ≡ φij , and the probability weight for ij bond

eK(cos[ϕij ]−1) is replaced by the Villain form
4.7. VILLAIN TRANSFORMATION 41

∞
X 1 2
P [φij ] ∝ e K(cos[ϕij ]−1)
→ e− 2 K[ϕij +2πnij ] (4.34)
nij =−∞

to ensure periodicity of the function under φi → φi +2π. A Poisson sum formula

states
X X∞ Z ∞
f (n) = dθ h(θ)e2πilθ . (4.35)
n l=−∞ −∞

For each link ij, the formula can be used to re-write the probability weight
∞ ∞ Z ∞
− 12 K[ϕij +2πnij ]2 1 2
X X
e = dθij e− 2 K[ϕij +2πθij ] +2πilij θij
. (4.36)
nij =−∞ lij =−∞ −∞

Since θij is a continuous variable the Gaussian integration can be performed

easily to yield

∞
X 1 2
P [ϕij ] ∝ e− 2K lij −ilij ϕij (4.37)
lij =−∞

up to an irrelevant overall constant. Compare this with an earlier formula, Eq.

(4.34), the main difference lies in the linear dependence of the weight function
on φij . The periodicity of P [ϕij ] under ϕi → ϕi + 2π is ensured since lij are
integer-valued. The net probability weight is given by the product of all pairwise
P [ϕij ]’s:
Y
P [{ϕi }] = P [ϕij ]. (4.38)
ij

Partition function Z is the integral of P [{ϕi }] over all ϕi ’s

!
Y Z 2π
Z= dϕi P [{ϕi }]. (4.39)
i 0

Each bond ij carries with it an integer-valued field lij , or a link variable. On

a square lattice each
Psite i carries
P two independent link variables li,i+x , li,i+y .
The coupling term ij lij ϕij = ij lij (ϕi − ϕj ) can be reorganized in terms of
ϕi , noting that each ϕi couples to lij in the manner

li,i+x (ϕi − ϕi+x ) + li,i+y (ϕi − ϕi+y ) + li−x,i (ϕi−x − ϕi ) + li−y,i (ϕi−y − ϕi )
 
X
→ ϕi [li,i+x + li,i+y − li−x,i − li−y,i ] = ϕi [li,i+x + li,i+y + li,i−x + li,i−y ] = ϕi  lij  .
j∈i

(4.40)

The last manipulation follows from the antisymmetry lij = −lji of the link
field. The coupling term has a geometric interpretation. Think of lij as a
current flowing from i to j. It is also clear from such an interpretation that
lji = −lij . Then the expression that couples to ϕi is the total current flowing
42 CHAPTER 4. XY MODEL

Figure 4.2: Integer-valued link (current) variable lij emanating from site i on a
square lattice. The sum of the current flowing away from i must be zero.

Figure 4.3: Representation of the original and dual lattice sites labeled by lower
case and upper case letters, respectively.

out from i, which is a conserved quantity in the case of conserved charge. This
is in fact the discretized version of divergence ∇ · l.
the P
−i i ϕi ( j∈i lij )
P
Each ϕi appears linearly, e P , in the partition function. Inte-
grating over ϕi yields zero unless j∈i lij = 0, which acts as a constraint on the
possible values of lij . The constraint is solved by writing lij as the difference

lij = hI − hJ (4.41)

of another integer fields hI defined on the dual lattice. In the picture shown in
Fig. 4.3, we will have

li,i+x = hI − hI−y
li,i+y = hI−x − hI
li,i−x = hI−x−y − hI−x
li,i−y = hI−y − hI−x−y . (4.42)

Indeed li,i+x + li,i+y + li,i−x + li,i−y = 0. Initially there were 2N link variables
lij for a lattice of N sites. The constraint being imposed at each site reduces
the number of independent fields to N , which is precisely the number of dual
variables hI . Now the probability weight P [{ϕi }] has lost its meaning because
ϕi ’s have been integrated out. What we have is an equivalent representation of
the partition function Z using the new dual variables,
X − 1 P (h −h )2
Z= e 2K hIJi I J . (4.43)
{hI }
4.7. VILLAIN TRANSFORMATION 43

The new model defined on the dual lattice is called the “height model” because
the greater difference in the heights of the nearby sites results in a larger en-
ergy difference. The high-temperature phase of the XY model corresponds to
the low-temperature phase of the height model and vice versa, because of the
reversed location of K in the Boltzmann factor. The 2D height model is in the
same universality class as the 2D XY model.
We have not yet come to the full vortex representation of the partition
function. To achieve this, invoke once more the Poisson sum formula to get
XZ ∞ 1
P 2 P
Z= dθI e− 2K hIJi (θI −θJ ) +2πi I mI θI . (4.44)
mI −∞

Here mI is another integer field defined at the dual sites I. We will integrate
out θI to obtain the partition functionPsolely in terms of the “vortex variable”
mI . Using the Fourier expansion θI = k θk eik·rI , the quadratic part becomes

1 X 1 X |θk |2
(θI − θJ )2 = . (4.45)
2K 2K Gk
hIJi k

where the Green’s function Gk = (4 − 2 cos kx − 2 cos ky )−1 is used. The linear
part becomes
X X
−2πi mI θI = −πi (θ−k mk + θk m−k ). (4.46)
I k

Upon completing the square, the action becomes

1 X 1
(θ−k − 2πKim−k Gk ) (θk − 2πKiGk mk ) + 2π 2 Km−k Gk mk . (4.47)
2K Gk
k

Integrating out θI gives a factor ZSW to the partition function:

!
X X
Z = ZSW exp −2π 2 K mI GIJ mJ . (4.48)
{mI } IJ

The real-space Green’s function GIJ is given by

d2 k eik·rIJ
Z
GIJ = a2 2 . (4.49)
(2π) 4 − 2 cos(kx a) − 2 cos(ky a)
It is convenient to use the regularized Green’s function G0IJ

G0IJ = GIJ − G0 , (4.50)

 !2  !
X X X
Z= exp −2π 2 KG0 mI  × exp −2π 2 K mI G0IJ mJ .
{mI } I IJ
(4.51)
Since G0 ∼ (1/2π) ln(L/a), the allowed configurations are those with total vor-
ticity zero. With the help of the result at large distances
44 CHAPTER 4. XY MODEL

1 rIJ 1
G0IJ ∼ −
ln − , (4.52)
2π a 4
the partition function may be rewritten as

 
0 2
X π K X X r 
 IJ 
Z= exp − m2I + πK mI mJ ln   . (4.53)
2 a
{mI } I I6=J

We have used the relation ( I mI )2 = 0 = I m2I + I6=J mI mJ in deriving the

P P P

final expression above. Often the fugacity term is defined by y = exp(−π 2 K/2).
This term controls the number of (total) vortices in the system. Equation
(4.53) is the desired action expressed solely in terms of the vortex variables mI .
Equation (4.53) is the start of renormalization group analysis of Kosterlitz.
With K large, the only kind of vortex states that can make a meaningful
contribution to the partition function are those with
P m I = 0, ±1. Going back to
2 2
Eq. (4.44), inserting the fugacity term e−(π K/2) I mI by hand, and restricting
the sum of mI to 0, ±1,

Z ∞ Yh i
1
P 2 2
Z ≈ dθI e− 2K hIJi (θI −θJ ) 1 + e−π K/2
(e2πiθI + e−2πiθI ) .
−∞ I
Z ∞ Yh i
1
P 2 2
− 2K hIJi (θI −θJ )
≈ dθI e 1 + 2e−π K/2
cos(2πθI )
−∞ I
Z ∞
1
P 2
− 2K hIJi (θI −θJ ) +2y cos(2πθI )
≈ dθI e
−∞
Z
1
d2 r[∇θ(r)]2 +2y d2 r cos[2πθ(r)]
R R
≈ Dθ(r) e− 2K . (4.54)

We have arrived at the sine-Gordon model. One can analyze this model using
the Wilson’s renormalization group idea. The existence of the confinement-
deconfinement transition in the sine-Gordon model is consistent with the vortex
binding-unbinding transition in the original XY model or the Villain model. All
analysis point to the existence of the defect(vortex)-mediated phase transition
at a temperature set by Tc = (π/2)J. A simpler, variational analysis of the
confinement-definement transition of the sine-Gordon model can be found in
Nagaosa’s book.

4.8 Frustrated XY model

An important variation of the XY model is the frustrated version given by
X
E = −J cos[θi − θj − αij ]. (4.55)
ij

If I define a boson hopping model on the same lattice in the presence of an

external field, the Hamiltonian would read
X
−iαij
b+ + b+ iαij


H = −t j bi e j bi e . (4.56)
ij
4.8. FRUSTRATED XY MODEL 45

√
In writing the boson operator bi = ρi eiθi , ρi = ρ, one obtains Eq. (4.55)
through 2tρ ≡ J. Thus one finds that αij in Eq. (4.55) reflects the influence of
externally imposed flux. The flux around an elementary plaquette is given by
the directed sum
1 X
ϕ= αij . (4.57)
2π
ij∈plaquette

Now the optional choice of angles θi would be to satisfy θi − θj = αij for all the
bonds. For certain geometries and depending on the value of the flux, this may
not be possible to achieve - hence the “frustration”.
The frustration effect can be introduced purely due to the geometric reason.
To see how this happens, we start with an antiferromagnetic XY model on the
triangular lattice

X X X
E=J cos(θi − θj ) = J cos(θi − θj ). (4.58)
ij i j=i+eα
√ √
The three unit vectors are given by e1 = (1, 0), e2 = (− 21 , 23 ), e3 = (− 12 , − 23 ).
We already know that the ground state of the model is obtained with the spin
configuration  
0 4π
θi = k · ri , k = ,0 . (4.59)
3
With this knowledge we can replace θi in Eq. (4.58) by θi + θi0 where θi0 is given
by the above equation. On substitution one finds that Eq. (4.58) becomes

XX XX π
E=J cos[θi −θi+eα −k · eα ] = −J cos[θi −θi+eα − ]. (4.60)
i eα i eα
3

It turns out the antiferromagnetic model without flux is equivalent to a ferro-

magnetic model in the presence of π flux per elementary plaquette.
The frustrated XY model has been the subject of lively research particularly
in the 80’s with the work of Teitel. Here we step aside from the historical
development concerning the model and describe the main intriguing physics.
It is strongly related between the chirality and the frustration according to
Villain. The model considered by Villain is
X
E = −J εij cos(ϕi − ϕj ), (4.61)
ij

where ε = ±1 which +1 and −1 refer ferromagnetic and antiferromagnetic

(AF) bond, respectively. It has the different formula of the ground state for
the number of AF bond. If the number of AF bond is an even, the formula of
ground state is
π
ϕi − ϕj = (1 − εij ). (4.62)
2

Figure 4.4(a) shows the system which is not frustated. If there is an odd
number of AF bond, the formula of ground state is written by
π π
ϕi − ϕj = (1 − εij ) + τij , (4.63)
2 4
46 CHAPTER 4. XY MODEL

Figure 4.4: (a) Even number of AF bond. (b) and (c) Odd number of AF bond.

Figure 4.5: Only one frustrated lattice induces half-integer vortex.

where τij has the value of ±1 and is the plaquette variable. The rotation of
spin, which is clockwise or counter-clockwise, is determined by the value of τij .
The chirality of plaquette is defined by this τij . The triangular lattice always
has the chirality because it is fully frustrated. However, the square lattice has
no chirality because it is not frustrated.
Another interesting phenomenon by frustration occurs in vortex. If there is
only one frustrated plaquette in all lattice like Fig. 4.5, it has a half-integer
vortex of ground states. If there is other frustration nearby the original one,
two possible cases for the rotation of vortex exist. First, the rotation of vortex
is equal to original half-integer vortex. In this case, it looks like a large vortex
and hence it costs much energy. Second, the rotation of vortex is opposite to
original half-integer vortex. In this case, they cancel out each other. Therefore,
it has a lower energy so that the system prefer this state.
Chapter 5

Heisenberg Model

As we learned from the previous chapter, the elementary excitation of a pla-

nar ferromagnetic model are spin waves and vortices. One may ask the same
question regard the ferromagnetic Heisenberg model in two dimensions,

X
H = −J Si · Sj , (5.1)
ij

where now the uni-modular spin Si can point in any direction in space. The
ground state of the model is the fully polarized state, Si = S0 . Taking this
arbitrary spin orientation to be the z-direction, a small fluctuation of the spin
vector away from equilibrium is denoted as Si = ni . Then the Heisenberg model
can be re-written as

Z
JX J h i
H= (ni − nj )2 ≈ d2 r (∂x nr )2 + (∂y nr )2 . (5.2)
2 ij 2

The final expression obtained under the smoothly varying configuration of ni is

called the non-linear σ-model (NLσM).

5.1 Saddle-point Method

Saddle-point method is a way to treat the NLσM in a mean-field manner. One
writes the partition function together with the constraint n2 = 1 as

Z Z !
1 X
Z = δ(n2 − 1) exp − dd r (∂µ nr )2
g µ
Z Z Z !
1 d
X
2 d 2
= exp − d r (∂µ nr ) − d rλr (nr − 1)
g µ
Z  Z   Z 
d 1 2 d
= exp − d r nr − ∂µ + λr nr + d r λr (5.3)
g

47
48 CHAPTER 5. HEISENBERG MODEL

Different components of n can be integrated out without constraints. Taking

G−1 = λ − (1/g)∂µ2 ,

Z 
3
Z = exp dd r λr − Tr ln G . (5.4)
2
The factor 3 reflects the three components of the spin vector. The saddle-point
equation follows from taking λr = λ and minimizing with respect to it,

dd k
Z
3 3 1
1 − TrG−1 = 1 − = 0. (5.5)
2 2 k<Λ (2π) λ + k2 /g
d

In two dimensions the integral relation yields

Λ
Λ2 Λ2
Z  
3 g 1 3g
1= kdk = ln 1 + → λg = . (5.6)
2 2π 0 λg + k 2 8π λg exp(8π/3g) − 1
For any coupling strength g (or temperature T ), there exists a finite solution λ.
In turn, λ gives the mass gap, or the inverse correlation length.
Extending the saddle-point analysis to D = d + 1 dimensions is easy with
the substitution

dd k
Z
1
→
(2π)d k 2 +λg
 p 
β 2 +λg
Z
d kd X 1
Z d
d k coth 2 k 2
= = . (5.7)
(2π)d iω ωn2 +k 2 +λg (2π)d
p
2
2 k +λg 3g
n

Size of the mass gap p

can be estimated forptwo dimensions in the high-temperature
limit β → 0, coth(β k 2 +λg/2) ≈ 2/β k 2 +λg,

Λ2 Λ2
Z  
2 kdk 1 T
≈T 2
= ln 1 + → λg = . (5.8)
3g 2π k +λg 4π λg exp(8π/3gT ) − 1

5.2 Renormalization Group Analysis

In doing the RG analysis
√ of the NLσM one first decomposes the spin vector
n = σn0 + π, σ = 1 − π 2 being the amplitude of longitudinal component
along the direction of magnetic order n0 and π representing the Goldstone
modes which satisfy n0 · π = 0. In this decomposition the partition function
becomes

Z Z !
2 2 1 d
X
2 2
Z = dπdσδ(σ + π − 1) exp − d r [(∂µ π) + (∂µ σ) ] (, 5.9)
g µ

assuming ∂µ n0 = 0. Without the sign of σ changing to the negative side, we

can integrate out σ first and obtain
5.3. SKYRMION EXCITATIONS 49

Z Z !
dπ 1 X 2
p
2
Z = √ exp − [(∂µ π) + (∂µ 1 − π 2 ) ]
2 1 − π2 g µ
Z Z Z !
1 X 2
p
2 2 ρ 2
= dπ exp − [(∂µ π) + (∂µ 1 − π ) ] − ln(1 − π ) .
g µ
2
(5.10)

The density ρ = 1/ad (a=lattice spacing) is introduced in the second line.

Assuming small amplitudes of fluctuation |π|  1 we expand the nonlinear
terms to obtain

Z Z X Z !
1  2 2
 ρ 2
Z = dπ exp − (∂µ π) + (π · ∂µ π) + π . (5.11)
g µ
2

The action can be separated as two pieces

Z X
1
S0 = (∂µ π)2
g µ
Z X Z
1 2 ρ
S1 = (π · ∂µ π) − π2
g µ
2
Z
Z = dπ exp (−S0 − S1 ) . (5.12)

5.3 Skyrmion Excitations

An important class of topological excitations obtained in the nonlinear σ-model
is the “skyrmion” excitation illustrated in Fig. 5.1. A single Skyrmionic spin
configuration with the center at the origin has the spins pointing according to

r2 − 4R2
nz = cos θ = ,
r2 + 4R2
4R(x + iy)
nx + iny = eiφ sin θ = 2 . (5.13)
r + 4R2
Here R serves as the “radius” of the Skyrmion. A Skyrmion carries with it a
topological number, obtained as the integral

Z
1
S= dr n · (∂x n × ∂y n). (5.14)
4π

For a singly-quantized Skyrmion given in Eq. (5.13) this number equals +1.
There is a convenient formalism for dealing with rotational excitations of
spins by introducing the complex fields z1 and z2 , and re-write the orientation
vector as
50 CHAPTER 5. HEISENBERG MODEL

n = zα∗ σαβ zβ . (5.15)

Since n · n = α |zα |2 , the unit norm is preserved by requiring α |zα |2 = 1.

P P
Indeed, one can parameterize zα as z1 = e−iφ/2 cos(θ/2) and z2 = eiφ/2 sin(θ/2)
one recovers n = (sin θ cos φ, sin θ sin φ, cos θ). Using these new variables one
finds

1
n · (∂x n × ∂y n) = ∂x Ay − ∂y Ax (5.16)
2
where the “vector potential” Ai (i = x, y) is given by

X
Ai = zα∗ ∂i zα . (5.17)
α

To complete the mapping, we also find

X X X 
(∂i n)2 = 4 (∂i zα∗ )(∂i zα ) + 4 zα∗ ∂i zα zβ∗ ∂i zβ . (5.18)
α α β

Figure 5.1: A Skyrmion configuration of spins in 2D Heisenberg model.

Chapter 6

Monte Carlo Methods and

Critical Phenomena

6.1 Introductory Remark

We have learned a great deal about the collective behavior of spins both in their
classical ground states and for low temperatures where fluctuations in terms of
a small number of elementary excitations offered a fairly adequate picture of
what was going on.
As we approach the critical temperature Tc where the order parameter van-
ishes, the picture of elementary excitations begins to fail, as there are increas-
ingly larger number of them, which will inevitably interact. For instance, do-
main walls which were the elementary excitations of an Ising ferromagnet will
begin to have their boundaries overlap. Then we have to understand how to
describe the dynamics of two domain walls colliding to form a new domain wall,
or perhaps a single domain wall splitting off into several smaller pieces. Like-
wise, a spin wave of momentum k will collide with another wave of momentum
k 0 and scatter. The elaborate tools of manybody physics are designed to com-
pute the effects of collision of such elementary excitations as far as we can -
perturbatively.
The drawback of the perturbative approach is that while a single collision can
be described quite accurately, the effects of multiple collisions and the possible
new qualitative behavior that arise as a result are exceedingly difficult to predict,
as the task corresponds to calculating ever higher orders of Feynman diagrams.
Unfortunately, the critical temperature corresponds to such a regime since many,
many elementary excitations populate the system so as to completely wash out
the remnants of the order present in the ground state.
Luckily for us, there is a new “organization principle”1 that emerges in the
critical region which are broadly known as the “scaling” and “universality”.
Much of the critical phenomena is understood in terms of these two key con-
cepts.

1 This word is a favorite of professor R. B. Laughlin, the president of KAIST and a Nobel

laureate for his work on fractional quantum Hall effects.

51
52CHAPTER 6. MONTE CARLO METHODS AND CRITICAL PHENOMENA

• Scaling: Measurable, physical quantities exhibit a behavior that depends

on the reduced temperature t ≡ (T − Tc )/Tc , as some power of t. The
power law is a defining characteristic of the critical region. For instance,
magnetization, specific heat, and the susceptibility depends on t in the
manner

m(t) ∼ (−t)β
C(t) ∼ |t|−α
χ(t) ∼ |t|−γ . (6.1)

The three numbers (α, β, γ) (and there are a few others) are the criti-
cal exponents which define the critical properties of a given system. For
instance, the mean-field theory discussed in chapter 2 gives (α, β, γ) =
(0, 1/2, 1). This is an example of the scaling behavior, although we did
not make too big a deal about it when we discussed mean-field theory.
Understanding critical phenomena is reduced to the understanding of (i)
why the scaling behavior arises near Tc , and (ii) how to calculate the crit-
ical exponents.

• Universality: According to the principle of scaling a given material in its

critical state is characterized by a number of critical exponents. This it-
self is a remarkable observation, but there is something more. Take two
materials made of different compositions which both undergo a magnetic
phase transition at TcA for material A, at TcB for material B. The transi-
tion temperatures are different, of course, because A and B have different
exchange energies J. However, when you plot the magnetizations mA and
mB as a function of the reduced temperature t using their respective crit-
ical temperatures, then suddenly both materials seem to exhibit exactly
the same behavior for m(t)! Not only is this true for magnetization, but
also for other physical quantities such as specific heat, susceptibility, etc.
As far as the critical properties are concerned, you can’t really distinguish
material A from material B. In other words, there is a class of materials
whose critical properties are defined by the same set of exponents (α, β, γ),
another class of materials characterized by a different set of critical ex-
ponents. We say the materials having the same exponents belong to the
same “universality class”.
Why this happens, of course, is because all materials belonging to the same
universality class have its critical properties described by the same model,
and the same free energy (Remember thermodynamic quantities are all
just derivatives of the free energy). On the other hand, two materials
which are different in chemical composition cannot possibly have exactly
identical microscopic Hamiltonians, but possess

HA = H0 + VA
HB = H0 + VB . (6.2)

H0 is a piece of the Hamiltonian that’s common to both materials, whereas

the remaining pieces VA and VB are unique to a given material. From these
6.2. MONTE CARLO (MC) FOR ISING SPINS 53

you write down the partition functions

X X
ZA = e−βEA , ZB = e−βEB . (6.3)
C C

Assuming that you can compute these quantities, the free energy is written
as the logarithm, FA = −T log ZA , FB = −T log ZB , and from the free
energies one can derive all the physical quantities through differentiation.
Hence, the fact that physical observables display identical behavior (when
plotted vs. t) is a statement ultimately about the identity of the partition
functions ZA ∼ ZB . But this is remarkable, given the fact that generally
EA 6= EB !

If we want to check theoretically if universality is really a property of the

critical phenomena, we will have to write down two models which differ some-
what, calculate the respective free energies and plot the magnetization and other
physical quantities. As mentioned earlier, the physics near the critical point is
extremely difficult to track down analytically because of the proliferation of
elementary excitations. Various approximation schemes can be devised, and
have been used over the decades, which at best gives circumstantial proof of the
above two important statements about the critical phenomena.
To be really sure, one will have to resort to a brute-force, yet exact, cal-
culation of the partition function itself. The number of configurations allowed
in a system increases exponentially with the number of particles, and to fully
count all the contributions of all the configuration in the partition function is
hopeless. An alternative method, which actually offers an (essentially) exact
answer for the partition function, exists. In the following section we will discuss
the method, and demonstrate the uses of the method in computing physical
quantities.

6.2 Monte Carlo (MC) for Ising spins

Let’s review the Boltzmannian principle of statistical mechanics. For concrete-
ness, we take the Ising model on L × L square lattice. A configuration refers
to a set of spin values assigned for each site of the square lattice. For example,
S1 = +1, S2 = −1, S3 = −1, · · ·, where 1, 2, 3, · · · refer to lattice sites, are a
configuration. A given configuration is collectively denoted {Si }, or C from now.
Once a configuration is given, it is trivial to calculate all the physical quantities
of interest for that configuration. Magnetization is a sum over all spins in that
configuration divided by N = L2 . Spin-spin correlation hSi SjP i is just a product
of two spin values located at i and j. Total energy E = −J ij Si Sj is a sum
over all nearest-neighbor products Si Sj . Everything is so simple for just one
configuration.
Imagine now that wePwant to calculate the thermodynamic average of the
magnetization hStot i = h i Si i for the Ising model. Boltzmann formula for the
magnetization reads

e−βEC
P
S
P tot,C
X
C
hStot i = −βEC
= Stot,C × pC . (6.4)
Ce C
54CHAPTER 6. MONTE CARLO METHODS AND CRITICAL PHENOMENA

P
In the above, Stot,C = i Si is the total spin for a given configuration C. There
are a total of 2N configurations accessible in the Ising model for L × L lattice.
Each one of 2N allowed configuration contribute to the thermodynamic average,
with a probability weight given by pC = e−βEC /Z, Z = C e−βEC .
P
One way to calculate the average magnetization will be like this. (Of course
no one can do this awful sum exactly, so we will have to put the question to a
computer.)

• Generate each one of 2N configurations by computer.

• For each configuration thus generated, calculate Stot,C by summing the

spin values of all sites.

• Calculate also the energy EC for that configuration and multiply e−βEC
with Stot,C previously calculated.

• Add the product over all configurations.

• Divide it by the partition function, which can be calculated in the similar

manner.

This process is OK, in principle, except for the astronomical cost in time it
takes to actually do it! Take L = 100. Total number of allowed configurations
are 2100×100 = (210 )1000 ≈ (103 )1000 = 103000 . This is a truly awesome number.
No one, not even the best computers in the world all collaborating on this single
project, can perform this many operations in a reasonable time frame. Clearly,
we need an alternative.
Metropolis and co-workers devised a clever way2 to overcome this difficulty.
Let me first explain the scheme, known as the Metropolis algorithm. It may
be helpful to read the following instruction in comparison with the previous
instructions for evaluating Boltzmann average.

• Take some random, initial configuration, {Si0 }. Pick one spin, at site k,
and change it to some other value, Sk0 → Sk1 . The change in energy for
Ising model as a result is
X X
∆E = −J Sj0 (Sk1 − Sk0 ) = 2J Sj0 Sk0 . (6.5)
j∈k j∈k

• If ∆E is less than zero, you have found a way to lower the energy, therefore
you accept it.

• If ∆E > 0, you are moving to a higher energy configuration. In deciding

whether to accept or reject such a move, you throw a coin. In practice, you
run a random number generator in the interval [0, 1], and if the random
number thus generated is less than e−β∆E you accept the move. If not,
reject! Having decided in this manner, we have either Sk1 = Sk0 (reject) or
Sk1 = −Sk0 (accept).
2 N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.H. Teller, E. Teller, “Equation of

State Calculations for Fast Computing Machines”, Journal of Chemical Physics 21, 1087-1092
(1953).
6.2. MONTE CARLO (MC) FOR ISING SPINS 55

• Move to the next site and play the same game over. Having scanned the
whole lattice this way, we will have arrived at a new configuration given
by {Si1 }.
• Go back to step 1. Repeat the Metropolis run using {Si1 } as an initial
configuration.

You may wonder why we graciously accept a move to a higher energy in

the above. For zero-temperature physics of course the lowest-energy state -
the ground state - is the only allowed configuration. Remember, however, that
finite-temperature physics is not just about energy. The entropy S also plays a
role, and the quantity T S competes with the energy E for dominance. In the
end, the free energy F = E − T S is the right quantity to minimize, not the
energy alone. So, instead of unilaterally rejecting a move to a higher energy
state, we sometimes accept it, but with the acceptance rate given by e−β∆E .
Either at very high temperature or for a slight increase in the energy for which
β∆E  1, e−β∆E ≈ 1, and the acceptance rate will be very high. In general the
rate of acceptance will depend on the temperature T at which we perform the
Metropolis algorithm. At a high temperature where a large number of excited
configurations are allowed it is difficult to expect a large magnetization hStot,C i
and its value will be low. Ultimately when a sufficiently large number of excited
states occur in the MC run, the long-range order will be lost altogether.
Following the above instructions, there is no guarantee that all of 2N al-
lowed configurations of the Ising model will be realized in the course of the
MC run. Higher energy configurations are less likely to emerge from Metropolis
run because such states are discouraged by an exponential factor e−β∆E from
happening. What Metropolis et al. showed, remarkably, was that each configu-
ration that does occur in the Metropolis run in fact does so with the probability
that is exactly its own Boltzmann weight pC . The deceptive little algorithm
described above ensures this to happen!! So, every configuration which comes
out at the end of one Metropolis run has already embodied its own probability.
Therefore, the thermal average of Eq. (6.4) gets replaced by
PMmax
M =1Stot,MC
hStot i = , (6.6)
Mmax
where Mmax is the total number of Metropolis runs. Stot,MC is evaluated for
each configuration generated by Metropolis algorithm, denoted MC to distin-
guish from the Boltzmannian configuration C. Now that the statistical weight
is automatically taken care of, all one has to do is a simple average over each
configuration. The same principle applies for evaluating other quantities such
as correlation function, etc, as will be discussed in the subsequent section.
Equivalence of Eq. (6.6) to Eq. (6.4) is guaranteed only in the limit Mmax →
∞ so, to be really precise, we must run Metropolis algorithm infinitely many
times. But, hey, life is short! We are usually content with a very large run,
Mmax  1, and experience has shown that a few million Metropolis steps are
good enough for most of the simple models we encounter in physics. This is a
vast reduction over 103000 samplings expected from a naive Boltzmann sum. The
problem can be put on just about any computers available nowadays. Professor
Boltzmann should say this from his grave: “Thank you, Mr. Metropolis, for
making my dream a reality!”
56CHAPTER 6. MONTE CARLO METHODS AND CRITICAL PHENOMENA

6.3 Calculation of physical quantities

The previous section described how to run a Metropolis algorithm and to cal-
culate the average magnetization, Eq. (6.6). Other quantities can be calculated
similarly. The susceptibility χ(T ) and the specific heat C(T ) at a given tem-
perature T are defined by (kB ≡ 1)
2
hStot i − hStot i2
χ(T ) =
T
hE 2 i − hEi2
C(T ) = (6.7)
T2
The first term in χ is calculated according to
−βEC
P 2
C Stot,C e
X
2 2
hStot i= P −βE = Stot,C × pC . (6.8)
C e C
C

If you run Metropolis algorithm, the probability pC is automatically taken care

of, and you instead end up with
PMmax 2
2 M =1 Stot,MC
hStot i = . (6.9)
Mmax

For a given Metropolis run, you calculate Stot,MC and square it, then you sum
over all the squared total spins obtained over the Metropolis runs, and that
2
will be the hStot i. We already discussed how to obtain hStot,MC i in Eq. (6.6)
and combining the two numbers you can easily access χ(T ). The same princi-
ple applies for the specific heat calculation, where hEi and hE 2 i are equal to
PM PMmax 2
( Mmax=1 EMC )/Mmax and ( M =1 EMC )/Mmax , respectively.
Actual MC calculation runs on a finite lattice of size L×L. Figure 6.1 is a plot
of the average magnetization as a function of temperature T /J, for progressively
larger system sizes L = 4 through L = 60. The decrease of magnetization with
temperature is more abrupt the larger the system size, and it appears that it
will drop to zero at some sharply defined temperature Tc as L → ∞. But the
thing is that we can never run a simulation for the truly infinite!!
The new rule of the game is finite size scaling. The rough idea is that by
looking at the simulation data for a number of different sizes L, one tries to guess
what will be the correct answer in the truly thermodynamic limit. In fact, one
can do better than just guess. A well-established rule of critical phenomena says
that for a finite-sized system near its critical temperature the physical quantities
must behave as

m(L, t) = L−β/ν ms (L1/ν t)

C(L, t) = Lα/ν Cs (L1/ν t)
χ(L, t) = Lγ/ν χs (L1/ν t). (6.10)

(α, β, γ) are the critical exponents already introduced in Eq. (6.1) and ν is
another critical exponent defining the correlation length ξ as

ξ(t) ∼ |t|−ν . (6.11)

6.3. CALCULATION OF PHYSICAL QUANTITIES 57

The quantities on the r.h.s. of Eq. (6.10) with a subscript s are the scaling
functions which depend on a particular combination L1/ν t only. According to
Eq. (6.10) magnetization obtained in the MC calculation at the same tempera-
ture t but for a different size L will give different results, but with the difference
behaving in the predictable manner. So if we plot, instead of m(L, t), the quan-
tity Lβ/ν m(L, t) vs. L1/ν t rather than t itself, all data should collapse onto a
single curve!!
Figure 6.2 is a plot of the same data as in Fig. 6.1 using the scaled equation.
Indeed, all data points collapse onto more or less a single curve, verifying the
validity of the finite-size scaling hypothesis. Similar plot of the susceptibility for
the 2D Ising model on a square lattice in Fig. 6.3 shows that finite-size scaling
works well for that quantity, too. By fitting the numerical data to a scaling
form one can do a rather nice job of extract the critical exponents3 .

Figure 6.1: Magnetization obtained for 2D square Ising model for varying system
sizes L. (excerpt from Landau’s book)

3 Due to lack of time and also of my own depth in this subject, I’m leaving out a lot of

important details of the critical phenomena in general, and of Monte Carlo in specific. The
in-depth discussion will have to wait another summer school and another expert.
58CHAPTER 6. MONTE CARLO METHODS AND CRITICAL PHENOMENA

Figure 6.2: Same magnetization as in the previous figure plotted in the scaling
form. (excerpt from Landau’s book)

Figure 6.3: Susceptibility for 2D square lattice Ising model, in scaling form
(excerpt from Newman’s book)
Chapter 7

Miscellaneous

7.1 Two-state Problem

All two-state quantum-mechanical problems can be mapped onto the Hamilto-
nian
sin θe−iφ
 
cos θ
H = ĥ · σ = , (7.1)
sin θeiφ − cos θ

with ĥ = (sin θ cos φ, sin θ sin φ, cos θ) denoting the spin orientation. This is
diagonalized by the unitary rotation U + HU with
 
cos(θ/2) − sin(θ/2)
U= (7.2)
sin(θ/2)eiφ cos(θ/2)eiφ

implying that the two eigenstates are

 
cos(θ/2)
|1i = , E1 = +1
sin(θ/2)eiφ
 
− sin(θ/2)
|2i = , E1 = −1. (7.3)
cos(θ/2)eiφ

7.2 Three-state problem

7.3 d-orbitalogy
The five d-orbitals are classified using the spherical harmonics

r
5
Y22 (θ, φ) = 3 sin2 θe2iφ = |2i
96π
r
5
Y21 (θ, φ) = −3 sin θ cos θeiφ = |1i
24π
r  
5 3 2 1
Y20 (θ, φ) = cos θ − = |0i
4π 2 2

59
60 CHAPTER 7. MISCELLANEOUS
r
5
Y21 (θ, φ) = 3 sin θ cos θe−iφ = |1i
24π
r
5
Y22 (θ, φ) = 3 sin2 θe−2iφ = |2i. (7.4)
96π
Three t2g orbitals are

i
|xyi = √ (|2i − |2i) ∼ sin2 θ sin φ cos φ ∼ xy,
2
i
|yzi = √ (|1i + |1i) ∼ sin θ cos θ sin φ ∼ yz,
2
1
|zxi = √ (|1i − |1i) ∼ sin θ cos θ cos φ ∼ zx. (7.5)
2
The two eg orbitals are

1
|x2 − y 2 i = √ (|2i + |2i) ∼ sin2 θ(cos2 φ − sin2 φ) ∼ x2 − y 2 ,
2
|3z 2 − r2 i = |0i ∼ 3z 2 − r2 . (7.6)

7.4 p-orbitalogy
The three p-orbitals are classified using the spherical harmonics

r
3
Y11 (θ, φ) = − sin θeiφ = |1i
8π
r
3
Y10 (θ, φ) = cos θ = |0i
4π
r
3
Y11 (θ, φ) = sin θe−iφ = |1i. (7.7)
8π
x, y, z-orbitals are given by

1
|xi = √ (|1i − |1i)
2
i
|yi = √ (|1i + |1i)
2
|zi = |0i. (7.8)
Chapter 8

Path Integral for Spins

8.1 Single spin

In the coherent state representation of spins, the spin state is expressed by a
spinor

 
z1 θ θ
|ni = z = , z1 = e−iφ/2 cos , z2 = eiφ/2 sin . (8.1)
z2 2 2

The angles are those of the classical spin vector n = (sin θ cos φ, sin θ sin φ, cos θ).
The path integral for spin contains the term

i
hṅ|ni = ż† z = φ̇ cos θ. (8.2)
2
For an arbitrary spin length I there is a Berry phase contribution to the action
e−iSB with

Z
SB = I d2 rdt (1 − cos θ) φ̇. (8.3)

R
The term dt(1 − cos θ)φ̇ gives the solid angle subtended by the unit vector
over time t. Equation of motion for spin follows from varying the Berry action

δSB
= I ṅ × n. (8.4)
δn

8.2 Antiferromagnetic spin chain

For a spin chain the Berry phase term is a collection of spin-spin Berry phase
action

X
AF
S1D = SB [ni ] (8.5)
i

61
62 CHAPTER 8. PATH INTEGRAL FOR SPINS

Owing to the staggered nature of spins we can write ni = (−1)i Li + Mi , as-

suming |Li |  |Mi |. We can expand the Berry action as

X X X δSB 
SB [(−1)i Li + Mi ] ≈ SB [(−1)i Li ] + Mi · . (8.6)
δni ni =(−1)i Li

i i i

When the spin direction is reversed, the sign of the Berry phase is also reversed,
SB [(−1)i Li ] = (−1)i SB [Li ]. We get

X X
AF
S1D ≈ (−1)i SB [Li ] + I Mi · L̇i × Li . (8.7)
i i

In the case of the one-dimensional chain one can manipulate the first term in
the Berry phase action a little more. For a chain of length N (N =even) the
first term can be re-arranged as

N/2 N/2
X X δSB IX
(SB [L2i ] − SB [L2i−1 ]) ≈ · (L2i − L2i−1 ) ≈ L̇i × Li · ∂x Li .(8.8)
i=1 i=1
δL 2 i

In the final expression all sites are included in the sum. The continuum limit of
the Berry phase action for 1D AF chain reads

Z Z
AF I
S1D ≈ L · ∂x L × ∂t L + I M · L̇ × L. (8.9)
2
What about the Heisenberg part? By using the same strategy one can re-
write the Heisenberg spin interaction as

X X
(−1)i Li + Mi −(−1)i Li+1 + Mi+1



H = J ni · ni+1 = J
i i
JX X
≈ (Li+1 − Li )2 + J M2i
2 i
Z X Z i
J
→ (∂µ L)2 + J M2 . (8.10)
2 µ
Chapter 9

1D Quantum Spin Chain

9.1 The model

The one-dimensional quantum S = 1/2 spin model on a periodic lattice of length
N is given by
X X
H = J⊥ Sn · Sn+1 + Jz Snz Sn+1
z
. (9.1)
n n

9.2 Ferromagnetic case

Ground state: For Jz < 0 and |Jz |/J⊥ sufficiently large, the ground state is the
ferromagnetic one |F i with all the spins up. Its energy H|F i = (Jz /4)|F i is
considered as zero.

9.3 Antiferromagnetic case

The general solution of Eq. (9.1) with J⊥ = Jz = J which is an eigenstate of
S z = j Sjz is
P

X
|ψi = φ(n1 , n2 , · · · , nM )|n1 , n2 , · · · nM i,
1≤n1 <n2 <···<nM ≤N

|n1 , n2 , · · · nM i = Sn−1 · Sn−M |F i. (9.2)

The ansatz solution that diagonalizes the isotropic spin Hamiltonian is

 
X M
X
φ(n1 , · · · , ·nM ) = AP exp i kP (j) nj  . (9.3)
P j=1

Here kP (j) spans all permutations of the set {k1 , k2 , · · · kM }, P ({k1 , k2 , · · · kM }) =

{kP (1) , kP (2) , · · · kP (M ) }, and AP is a coefficient that depends on each permuta-
tion. The kj ’s are the pseudo-momenta for each of the M down spins, and the
total energy and momenta in terms of the pseudo-momenta is

63
64 CHAPTER 9. 1D QUANTUM SPIN CHAIN

M
X M
X
E=J (cos kj − 1), P = kj . (9.4)
j=1 j=1

The coefficients AP and the set of allowed pseudo-momenta {kj } can be

derived by substitution of Eq. (9.3) into Eq. (9.1). One can separate the
Hamiltonian as two pieces H = H⊥ + Hz where

N
1 X − + −
H⊥ = J (S S + Sn+ Sn+1 )
2 n=1 n n+1
X
Hz = J Snz Sn+1
z
. (9.5)
n

The action of H⊥ on |ψi is

1 X X
J φ(n1 , · · · , nM ) |{n0j }i. (9.6)
2 0
{nj } {nj }

Here |{n0j }i refers to all set of configurations related to the original configuration
|{nj }i by the reversal of the nearest pair of up and down spins. For example,
if a given basis configuration is |3, 6i (spin down position at 3 and 6), the set
of |{n0j }i is |2, 6i, |4, 6i, |3, 5i, |3, 1i. The last configuration arises in the periodic
boundary condition geometry. The total number of configurations {n0j } arising
from {nj } equals the the total number of spin up-down pairs, or the number of
kinks. The action of Hz gives a numerical factor proportional to the number of
kinks,
1
Hz |{nj }i = − J × (number of kinks)|{nj }i. (9.7)
2
The Schrodinger equation becomes (J = 2)
X
Eφ({n}) = (φ({n0 }) − φ({n})) . (9.8)
{n0 }

One-spin down state: The one-spin down state, M = 1, has Eq. (9.8) reduced
to

φ(n + 1) + φ(n − 1) − 2φ(n) = Eφ(n). (9.9)

Ansatz c(n) = Aeikn solves the equation with the energy given by

Ek = 2(cos k − 1). (9.10)

Imposing the periodic boundary condition φ(n + N ) = φ(n) quantizes the k
values
2π
kj =
Ij (9.11)
N
with the allowed quantum number Ij = 1, · · · N .
9.3. ANTIFERROMAGNETIC CASE 65

Two-spin down states: For M = 2, the case of n2 = n1 + 1 needs to be treated

separately from the others. If n2 = n1 + 1, EoM gives

Eφ(n1 , n1 + 1) = −2φ(n1 , n1 + 1) + φ(n1 − 1, n1 + 1) + φ(n1 , n1 + 2). (9.12)

Otherwise EoM gives

Eφ(n1 , n2 ) = −4φ(n1 , n2 )
+φ(n1 − 1, n2 ) + φ(n1 + 1, n2 ) + φ(n1 , n2 − 1) + φ(n1 , n2 + 1).
(9.13)

Suppose Eq. (9.13) can be analytically continued to n2 = n1 + 1. Then sub-

tracting out Eq. (9.12) from Eq. (9.13) for n1 = n, n2 = n + 1 would give

2φ(n, n + 1) = φ(n, n) + φ(n + 1, n + 1). (9.14)

One can try the solution

φ(n1 , n2 ) = A1 eik1 n1 +ik2 n2 + A2 eik2 n1 +ik1 n2 . (9.15)

If Eq. (9.13) were valid everywhere, we would immediately conclude that Eq.
(9.15) is the solution with the energy
2
X
E=2 (cos kj − 1). (9.16)
j=1

Substituting the ansatz, Eq. (9.15), into the constraint, Eq. (9.14), gives
k1 −k2
A1 eik1 +ik2 + 1 − 2eik1 ei 2 − cos k1 +k
2
2

− = ik1 +ik2 ik
= k2 −k1 . (9.17)
A2 e + 1 − 2e 2
ei 2 − cos k1 +k2
2

The ratio |A1 /A2 | is one, hence one can write A1 = eiθ12 /2 and A2 = eiθ21 /2 ,
θ21 = −θ12 , and show that θ12 in terms of k1 , k2 is obtained from

θ12 sin k1 −k
2
2

cot = . (9.18)
2 cos k1 +k
2
2
− cos k1 −k
2
2

The two particle solution becomes

φ(n1 , n2 ) = eiθ12 /2+i(k1 n1 +k2 n2 ) + e−iθ12 /2+i(k2 n1 +k1 n2 ) . (9.19)

Imposing the periodic boundary condition becomes tricky for two particles
and beyond. We choose the condition that φ(n1 , n2 ) = φ(n2 , n1 + N ), which
renders

eik1 N = eiθ12 , eik2 N = eiθ21 . (9.20)

The quantum numbers are quantized accordingly,

k1 N = 2πI1 + θ12 , k2 = 2πI2 + θ21 , Ii = 1, · · · , N. (9.21)

66 CHAPTER 9. 1D QUANTUM SPIN CHAIN

M-spin down state: We return to the general, M -spin down state with the
ansatz solution given by
 
X M
X
φ(n1 , · · · , ·nM ) = AP exp i kP (j) nj  . (9.22)
P j=1

The general form of the Schrodinger equation becomes more explicitly

M 
X
φ(n1 , · · · , nj − 1, · · · , nM ) + φ(n1 , · · · , nj + 1, · · · , nM )
j=1

− 2φ(n1 , · · · , nM ) = Eφ(n1 , · · · , nM ). (9.23)

The terms which contains two adjacent spins of the same orientation are handled
by the constraints

φ(· · · , nj , nj , · · ·) + φ(· · · , nj + 1, nj + 1, · · ·) = 2φ(· · · , nj , nj + 1, · · ·). (9.24)

Inserting the M -particle wave function into this constraint yields the relation

X X X
AP + AP eikP (j) +ikP (j+1) = 2 AP eikP (j+1) ,
P P P
 
X
AP = AP exp i nl kP (l) + nj kP (j) + nj kP (j+1)  . (9.25)
l6=j,j6=j+1

It can be seen that only a pair of coefficients AP and AP (j,j+1) need to be

involved in the equality. I define P (j, j + 1) as the original permutation P plus
the interchange of kP (j) and kP (j+1) . Hence, Eq. (??) reduces to the pair-wise
equaiton

AP + AP (j,j+1) + (AP + AP (j,j+1) )eikP (j) +ikP (j+1) = 2(AP eikP (j+1) + AP (j,j+1) eikP (j) )

AP (j,j+1) eikP (j) +ikP (j+1) + 1 − 2eikP (j+1)

− ≡ −e−iθP (j),P (j+1) = ik +ik . (9.26)
AP e P (j) P (j+1) + 1 − 2eikP (j)
The relation is satisfied by the ansatz

P eikP (i) +ikP (j) + 1 − 2eikP (j)

AP = e i i<j θP (i),P (j) /2
, − e−iθP (i),P (j) = . (9.27)
eikP (i) +ikP (j) + 1 − 2eikP (i)
The fact that AP must be of the form given above is easily illustrated with an
example M = 4. In that case, the phase factor for AP reads (1/2) times

θP (1),P (2) + θP (1),P (3) + θP (1),P (4) + θP (2),P (3) + θP (2),P (4) + θP (3),P (4) . (9.28)
9.3. ANTIFERROMAGNETIC CASE 67

The phase factor for AP (1,2) , for instance, is given by 1 and 2 in the above
equation replaced by 2 and 1. An inspection reveals that the net difference
only occurs in replacing θP (1),P (2) by θP (2),P (1) , hence the ratio AP (1),P (2) /AP
is indeed equal to e−iθP (1),P (2) . The same result obtains for interchange of any
other consecutive indices j and j + 1.
Following the earlier practice we impose the periodic boundary condition

φ(n1 , · · · , nM ) = φ(n2 , · · · , nM , n1 + N ) →

X P X
AP e i j kP (j) nj
= AP eikP (1) n2 +···+ikP (M −1) nM +ikP (M ) (n1 +N ) .
P P
(9.29)
One can define a permutation P0 which takes

P0 (1) = M, P0 (j + 1) = j (j > 1). (9.30)

Then the r.h.s. of Eq. (9.29) becomes

X
AP eikP P0 (2) n2 +···+ikP P0 (M ) nM +ikP P0 (1) (n1 +N ) =
P
X P X P
AP P −1 ei j kP (j) nj +ikP (1) N
= AP e i j kP (j) nj
. (9.31)
0
P P

From the last equality it follows that

AP = AP P −1 eikP (1) N . (9.32)

0

The amplitudes are written out explicitly

 
i X
AP = exp  θP (i),P (j) 
2 i<j
 
i X
AP P −1 = exp  θ −1 −1 . (9.33)
0 2 i<j P P0 (i),P P0 (j)

As long as j < M , θP P −1 (i),P P −1 (j) gives another element belonging to the

0 0
family θP (i),P (j) with i < j. When j = M , one gets

θP P −1 (i),P P −1 (M ) = θP (i+1),P (1) = −θP (1),P (i+1) . (9.34)

0 0

It follows that Eq. (9.32) becomes

PM −1
eikP (1) N = ei i=1 θP (1),P (i+1)
, (9.35)
or
P
eikj N = ei l6=j θjl
, (9.36)
68 CHAPTER 9. 1D QUANTUM SPIN CHAIN
Chapter 10

Hubbard Model

10.1 The Model

In the previous chapter we learned how to analyze electron motion in solids by
employing a simple model - the tight-binding model. An easy generalization of
the model also allowed us to study the disordered system. While disorder is an
essential aspect of all real materials, there is an equally or in some instances even
more important aspect of the many-particle system. And that is the interaction
between the particles. For electrons, the interaction is due to the Coulomb
force. Hubbard model, invented by John Hubbard in the 1950’s, is the most
representative of the models dealing with electron-electron interaction. It is
written as the Hamiltonian

H = T +V
X X
tij c+ +
c+


T = jσ ciσ + ciσ cjσ − µ iσ ciσ
hijiσ iσ
X
V = U ni↑ ni↓ (10.1)
i

T represents hopping of electron between a pair of sites hiji with hopping am-
plitude tij = tji . For simplicity we take tij = −t for all the bonds. Unlike
in the previous chapter we include electron spin σ explicitly because, after all,
real electrons
P do carry spin. Diagonalizing T in the absence of the interaction
V leads to kσ k c+ kσ ckσ , where k is the energy of the eigenstate |ki. Thus T
alone describes delocalized electrons that form a band inside the solid. This has
been the essence of the discussion of previous chapter.
Let’s examine the properties of the V -term. First it is completely local, i.e.
it only acts at a given site i: niσ = c+iσ ciσ (no sum on σ) counts the number of
electrons of spin σ at site i. This being an electron model, niσ can only be 0 or
1. When a given site i is occupied by zero electron (ni↑ = ni↓ = 0) or by only
one electron of spin σ (niσ = 1, niσ = 0, ni↑ ni↓ = 0), V has an expectation
value of zero. When the site is doubly occupied, i.e. ni↑ = ni↓ = 1, the site
has an energy increase of U . We may therefore understand U as the Coulomb
repulsion energy of the two electrons. The challenge we must face now is to
understand the properties of, and to solve, the model Hamiltonian (10.1) in the
simultaneous presence of T and V .

69
70 CHAPTER 10. HUBBARD MODEL

state ni E
|0i 0 0
c†i↑ |0i, c†i↓ |0i 1 −µ
c†i↓ c†i↑ |0i 2 U − 2µ

Table 10.1: Possible atomic states and their energies

P In the atomic limit tij → 0, the Hamiltonian reduces to the atomic form H =
i Hi , Hi = −µni + U ni↑ ni↓ , that lacks communication between neighboring
sites. Diagonalizing Hi is trivial since each fixed occupation of electron gives an
eigenstate. Eigenstates and their corresponding energies are
Suppose the chemical potential is chosen to satisfy 0 < µ < U2 . Then among
the three possible eigenenergies, only the −µ, corresponding to single occupa-

−1
tion, lies below zero energy. According to the Fermi factor F (E) = eβE + 1
that determines the occupation probability of a given eigenstate, only the ni = 1
state will occur at zero temperature. That means each site of the lattice will
be occupied by one, and only one electron at zero temperature. Since both spin
states give rise to the same energy, each electron occupying the site can be either
spin up or spin down. For N electrons occupying N different lattice sites, this
leads to enormous ground state degeneracy 2 × 2 × · · · × 2 = 2N . In other words,
any state of the spins given by |σ1 σ2 · · · σN i, σi = ±1, is an eigenstate having
the same energy −µN .
Next we raise the question: when we add the hopping part T to the Hamil-
tonian will the eigenstates of V alone remain eigenstates, and will their energies
differ due to the hopping effect?

Before we answer these equations we re-visit the foundations of quantum me-

chanics. The basic postulates of quantum mechanics is that there P are a set of
states, labelled |mi, that together form a complete, orthonormal set, m |mihm| =
1, hm|ni = δmn . Each |mi is not necessarily an eigenstate of H, but due to
the completeness
P any eigenstate of H can be written as a linear combination
|ψi = m cm |mi, ! !
X X
H cm |mi = E cm |mi . (10.2)
m m
By taking projection on both sides with hn|, we get
X
cm hn|H|mi = Ecn . (10.3)
m

We will call matrix element hn|H|mi = Vnm , then we have a linear matrix
equation X
Vnm cm = Ecn . (10.4)
m
By diagonalizing the matrix Vnm , the eigenvectors {cmP } are obtained. Then
the eigenstate of the Hamiltonian is expressed as |ψi = m cm |mi, having the
energy E.
We apply precisely this strategy to solve the Hubbard model. And before
doing so, we must first identify the basis states to be used in the diagonalization
10.2. PROJECTION TO TRUNCATED HILBERT SPACE 71

process. For a given site i there are four possible states, listed in Table 10.1,
corresponding to zero, one (of spin up and down), and two electron occupation,
which we will label |0i i, |σi i, and |di i in obvious notation. When there are N
sites, the basis states are given by

|α1 i|α2 i · · · |αN i. (10.5)

where each αi is one of the four possible states. Working out Vnm corresponds
to calculating the overlap of H between two arbitrary basis states (10.5). For
N sites, there are 4N such basis states, and the size of the matrix Vnm is
4N × 4N . Even for N = 10, the dimension of the matrix is 410 × 410 ' 106 × 106 !!
Diagonalizing such a huge matrix is impossible, even with the world’s best
computer available right now. In this chapter we will try to diagonalize H up
to N = 4.
To get a flavor of the steps needed to diagonalize H, we start with the
two-site Hubbard model
X X
c+ +


H = U n1↑ n1↓ + U n2↑ n2↓ − µ (n1σ + n2σ ) − t 2σ c1σ + c1σ c2σ . (10.6)
σ σ

First there are sixteen basis states given as the direct product

{|01 i, | ↑1 i, | ↓1 i, |d1 i} ⊗ {|02 i, | ↑2 i, | ↓2 i, |d2 i}. (10.7)

We know how to work out the matrix element of T already. Since T conserves
the total spin, nonzero element only occurs between states such as

|σ1 , 02 i ↔ |01 , σ2 i, |d1 , 02 i ↔ |σ1 , σ 2 i, or |d1 , σ2 i ↔ |σ 1 , d2 i. (10.8)

You should carefully work out the matrix element for each of these.
Next, the interaction term V has a nonzero element if and only if a site is
doubly occupied. Thus

hd1 , d2 |V |d1 , d2 i = U + U = 2U
hd1 , α2 |V |d1 , α2 i = U (α2 6= d2 )
hα1 , d2 |V |α1 , d2 i = U (α1 6= d1 )
hα1 , α2 |V |α1 , α2 i = 0 (α1 6= d1 , α2 6= d2 ) (10.9)

In this way we can work out all the matrix elements among 16 basis states. The
16 × 16 matrix can be diagonalized easily, and will yield 16 energy levels. The
lowest energy and the corresponding eigenvector defines the ground state of the
Hubbard Hamiltonian.

10.2 Projection to truncated Hilbert space

When the on-site interaction energy U is very large, double occupation of any
given quantum state is realistically forbidden. It is as if the Hilbert space for
each atomic site is restricted to three, consisting of empty, or singly-occupied
with either spin-up or spin-down. Within such a restricted Hilbert space the
Hubbard term no longer exists because ni↑ ni↓ acting on such a space will iden-
tically give zero.
72 CHAPTER 10. HUBBARD MODEL

It is often the case that the effective Hamiltonian acting within a restricted
Hilbert space looks quite different from the original Hamiltonian defined in the
larger space. Being able to systematically carry out the transformation to the
restricted space is of practical value and the techniques are presented in this
section using the Hubbard model as the starting point.

Tji c†jσ ciσ + U

X X
H= ni↑ ni↓ . (10.10)
ijσ i

We have used Tji above which reduces to Tji = −t in the case of the nearest-
neighbor hopping only. The following derivation is largely taken from Yoshioka,
Girvin, and MacDonald.
Using the fact that the electron occupation plus the hole occupation at any
site for spin σ adds to one, niσ + hiσ = 1, one can split the kinetic term into
three pieces, T = T0 + T1 + T−1 where

Tji njσ c†jσ ciσ hiσ

X
T1 =
ijσ

Tji hjσ c†jσ ciσ niσ

X
T−1 =
ij

Tji hjσ c†jσ ciσ hiσ + Tji njσ c†jσ ciσ niσ .
X X
T0 = (10.11)
ijσ ijσ

T1 increases the number of doubly occupied site by one because it has a non-zero
matrix element if and only if the initial state has the configuration | ↑i ↓j i, or
| ↓i ↑j i. Then T1 acting on the initial state produces |0i dj i or |di 0j i, d implying
the doubly occupied site. Similarly, T−1 decrease the number of doubly occupied
sites by one by acting on the initial state |0i dj i or |di 0j i. T0 does not change
the number of double occupancy because it acts only on the initial state |0i σj i,
|σi 0j i, or |di σj i, |σi dj i, σi =↑, ↓.
Now consider a canonical transformation
0
H = eiS He−iS (10.12)
using iS = (T1 − T−1 )/U . T±1 P satisfies the commutation [V, T1 ] = U T1 ,
[V, T−1 ] = −U T−1 where V = U i ni↑ ni↓ is the Hubbard term. By a straight-
0
forward expansion of H up to second order in S and keeping the leading terms
in O(1/U ) we get
0 1
H = T0 + V + ([T1 , T−1 ] + [T1 , T0 ] + [T0 , T−1 ]) . (10.13)
U
0
When we let the Hamiltonian H act between states with no doubly occupied
sites it is further reduced to
0 1
H = T0 − T−1 T1 (10.14)
U
because the other two terms would change the number of doubly occupied sites.
Write out T−1 T1 as

Tlk Tji hlλ c†lλ ckλ nkλ njσ c†jσ ciσ hiσ .
X
(10.15)
ijklσλ
10.2. PROJECTION TO TRUNCATED HILBERT SPACE 73

Excluding the three-site interaction we get k = j, l = i,

Tij Tji hiλ c†iλ cjλ njλ njσ c†jσ ciσ hiσ .
X
T−1 T1 = (10.16)
ijσλ

A short consideration gives that the operators have the following action on
the possible spin states on sites i, j:
X
T−1 T1 = t2ij (| ↑i ↓j i − | ↓i ↑j i) (h↑i ↓j | − h↓i ↑j |) = 2t2ij |sij ihsij |. (10.17)
ij

The symbol sij refers to the singlet state in the ij-bond. For less than half-filled
case it is possible to have an empty site either at i or j, for which T−1 T1 would
give zero. Taking this into account, |sij ihsij | may be re-written

1
|sij ihsij | = ni nj − Si · Sj (10.18)
4
employing the spin operator Si now. In all, we get
 
0 2 X 2 1
H = T0 + tij Si · Sj − ni nj ,
U ij 4

Tji hjσ c†jσ ciσ hiσ .

X
T0 = (10.19)
ij

Equation (10.19) can easily be understood within the ordinary second-order

perturbation calculation, the same language used in the original derivation of
superexchange by Anderson. Let us assume the case of the nearest hopping only
and half-filling for simplicity. Action of T on the low-energy Hilbert space takes
one out of the truncated Hilbert space since it always gives rise to the doubly
occupied site. The second-order perturbation theory on the other hand gives
non-trivial transition amplitudes within the low-energy space

X hg|T |exihex|T |g 0 i
hg|Heff |g 0 i = , (10.20)
ex
Eg − Eex

where |gi = | ↑i ↓j i or | ↓i ↑j i is one of the two low-energy configurations for the

nearest-neighbor pair ij, and |exi is one of the excited states or “virtual states”
containing doubly occupied sites. Due to Pauli principle only the initial states
for a given pair of nearest sites whose spins are opposite will acquire non-zero
transition amplitude:

ground state T virtual state T ground state

| ↑i ↑j i −→ 0
(10.21)
|di 0j i | ↑i ↓j i
| ↑i ↓j i −→ −→ .
|0i dj i | ↓i ↑j i

By working out the nonzero elements of Eq. (10.20) explicitly as shown in Table
10.2, hg|Heff |g 0 i is found to have the same 2 × 2 matrix elements as those of the
effective Hamiltonian
74 CHAPTER 10. HUBBARD MODEL

| ↑i ↑j i | ↑ i ↓j i | ↓ i ↑j i | ↓i ↓j i
h↑i ↑j | 0 0 0 0
h↑i ↓j | 0 2 −2 0
h↓i ↑j | 0 −2 2 0
h↓i ↓j | 0 0 0 0

Table 10.2: Transition amplitudes in units of t2 /U within the low-energy states

of two neighboring sites.

4t2 X
 
1
Heff = Si · Sj − , (10.22)
U 4
hiji

where the summation runs over all nearest-neighbor pairs.

10.3 Mean-field Theory and Antiferromagnetism

10.3.1 Square lattice
The previous section described a strategy as to how one would exactly diagonal-
ize the Hubbard Hamiltonian for a fixed size and particle number. Undoubtedly
the process will have to stop and, with the present capacity of computing power,
stop at a rather miserably level of the system size. The thermodynamic behavior
which often emerges only as the limit of the infinite system size is taken is hard
to obtain by direct numerical diagonalization. If one has a good guess what the
ground state of the model will be, on the other hand, one can often define a
mean-field theory that easily produces such a ground state in a self-consistent
manner. Below we show how one such phase, namely the antiferromagnetic
ground state, is “predicted” from the Hubbard model.
First one rewrites the interaction term of the Hubbard model
X X X
H = −t c+
jσ cjσ + U ni↑ ni↓ − µ ni (10.23)
i,j,σ i i

using the simple identity

1 1 1 1
n↑ n↓ = (n↑ + n↓ )2 − (n↑ − n↓ )2 ≡ n2 − m2 , (10.24)
4 4 4 4

U 2
U ni↑ ni↓ = (n − m2i ). (10.25)
4 i
We will carry out the meanfield decoupling at half-filling, hni i = 1.
U 2 U U U U
ni − m2i → hni ini − hmi imi = ni − si mi .


(10.26)
4 2 2 2 2
The average of the magnetic moment is denoted si . The meanfield Hubbard
Hamiltonian is
 X
X U UX
HM F = −t c+ c
jσ iσ + − µ ni − si mi . (10.27)
ijσ
2 i
2 i
10.3. MEAN-FIELD THEORY AND ANTIFERROMAGNETISM 75

To describe antiferromagnetism we assume the staggered field si = (−1)i s where

s is the size of the magnetic moment at each site. At half-filling the effective
chemical potential µef f = µ − U2 must equal zero. Hence

X X X X
HM F = −t c+
jσ ciσ − η (−1)i mi = k c+
kσ ckσ − η σc+
k+Qσ ckσ ,
ijσ i kσ kσ
(10.28)
where η = U s/2. The last line follows from Fourier transform. Q implies
the AF wavevector (π, π)(2D) or (π, π, π)(3D). We divide the original Brillouin
zone [−π, π] ⊗ [−π, π] into two: one covers the region bounded by the four lines
kx + ky = π, kx + ky = −π, kx − ky = π, kx − ky = −π, and the other, the
remaining part of the BZ. The first region, shaded in Fig. ??, is the reduced
Brillouin zone (RBZ).
Now one can rewrite the Hamiltonian
X 0   X0
k c+ +
σ c+ +


HM F = kσ ckσ + k+Q ck+Qσ ck+Qσ −η k+Qσ ckσ +ckσ ck+Qσ
kσ kσ
0   
X k −ησ ckσ
c+ c+


= kσ k+Qσ . (10.29)
−ησ k+Q ck+Qσ
kσ

σ = ±1 corresponds to ↑, ↓ spins. The 2 × 2 matrix given above can be diago-

nalized exactly with the canonical transformation
    
ckσ cos θk sin θk γ1kσ
= , (10.30)
ck+Qσ − sin θk cos θk γ2kσ
provided we take

r
ση k − k+Q 1
sin 2θk = , cos 2θk = , Ek = η 2 + (k −k+Q )2 . (10.31)
Ek 2Ek 4
In the standard cubic(square)
p lattice k+Q = −k , and thus sin 2θk = ση/Ek ,
cos 2θk = k /Ek , Ek = 2k + η 2 . Using this transformation the Hamiltonian is
brought to a diagonalized form:
0
X
+ +
H= Ek [γ1σ γ1kσ − γ2kσ γ2kσ ], (10.32)
kσ
assume k = −k+Q . We have managed to reduce the Hamiltonian to the diag-
onalized form, and obtain the meanfield energies ±Ek . Furthermore, we have
explicitly derived a relation between the quasiparticle operators (γ) and the orig-
inal, electron operators. The eigenvectors of the matrix act as a “connection”
between the two operators. Since Ek > 0, all the k-states in the RBZ for the
lower energy branch are occupied, and completely empty for the upper branch
at zero temperature, in agreement with the original prescription of half-filling.
The ground state is expressed as
0
Y
+ +
|GSi = γ2k↑ γ2k↓ |0i. (10.33)
k
76 CHAPTER 10. HUBBARD MODEL

The inverse transformation

    
γ1kσ cos θk − sin θk ckσ
= , (10.34)
γ2kσ sin θk cos θk ck+Qσ
with
s   s  
1 εk 1 εk
cos θk = 1+ , sin θk = σ 1− , (10.35)
2 Ek 2 Ek
gives

+
γ2kσ = uk c+ +
kσ + σvk ck+Qσ ,
s   s  
1 εk 1 εk
uk = 1− , vk = 1+ . (10.36)
2 Ek 2 Ek

The ground state is expressed as

Y  
|GSi = uk c+ +
k↑ + vk ck+Q↑ uk c+
k↓ − v +
k k+Q↓ |0i.
c (10.37)
k

The real-space wave function is a product of Slater determinants, D[φ↑ (k, r)]D[φ↓ (k, r)]:

φ↑ (ki , rj ) = eiki ·rj uk + ei(ki +Q)·rj vk

φ↓ (ki , rj ) = eiki ·rj uk − ei(ki +Q)·rj vk . (10.38)

The mean-field condition on η reads

0  
X 1 Ek
η = Uη tanh . (10.39)
Ek 2T
k

This equation has an exact, T = 0 solution in the case of the constant density
of states D() = 1/(2D),
2D/U
s= , (10.40)
sinh[2D/U ]
The transition temperature TN is given when s = 0, Ek = |k |.
In this section I have given a rather detailed derivation of the gap equation
for the antiferromagnetic ordering. The BCS problem is entirely similar to this
in its mathematical structure. Technically, if you know how to solve the AF
problem, then you are (almost) equally well prepared for the BCS or any other
kind of mean-field problems of condensed matter physics.
The antiferromagnetic ordering at (π, π) introduced a spatial modulation in
an initially translationally invariant Hamiltonian that led to (i) doubling of the
unit cell and halving of the Brillouin zone and (ii) folding of the band into two
subbands and a gap between the bands. Full occupation of the lower band and a
complete absence in the upper band leads to the opening of the gap at half-filling.
10.3. MEAN-FIELD THEORY AND ANTIFERROMAGNETISM 77

10.3.2 Triangular lattice

For the triangular lattice the ordered spins no longer lie along a common axis.
The general belief is that the ordered moments form a collinear state, with a 120◦
angle between adjacent pair of spins. To capture this in the mean-field theory
one must first generalize the mean-field decoupling of the Hubbard interaction
to the spin-rotation-invariant form

U U U U
ni↑ ni↓ → hni ini − hsi i · si = ni − si · mi . (10.41)
2 2 2 2
The average moment mi are taken to lie within the plane, and it is useful to
use the characterization in the complex notation

mix + imiy = mi = m0 eiQ·ri (10.42)

Taking Q = (4π/3)x̂ produces the right spin orientation of√ the triangular lattice
where the lattice sites are given by ri = px̂ + q(x̂/2 + 3ŷ/2) for a pair of
integer (p, q). The initial configuration m0 is an arbitrary complex number of
unit magnitude or less. The Zeeman term becomes

UX ∗ − Um − Um X +
− (mi si + mi s+
i )=− (sQ + s+ Q) = − (ck+Q↑ ck↓ + c+
k↓ ck+Q↑ ).
2 i 2 2
k
(10.43)
The Brillouin zone of the triangular lattice can be broken up into three pieces
of equal area which are connected to each other by the translation of +Q or
−Q in momentum space. The Hamiltonian within the reduced Brillouin zone
becomes

0 
X 
k c+ + +
kσ ckσ + k+Q ck+Q,σ ck+Q,σ + k−Q ck−Q,σ ck−Q,σ
kσ
0 
X 
−η c+ + +
k+Q↑ ck↓ + ck−Q↑ ck+Q↓ + ck↑ ck−Q↓ + h.c.
k
T 
c+
  
k↑ k 0 0 0 0 −η ck↑

 c+
k↓
 
  0 k −η 0 0 0  ck↓ 
0  +  
 ck+Q↑ 0 −η k+Q 0 0 0   ck+Q↑
X   
=
 
 +  
 ck+Q↓ 0 0 0 k+Q −η 0   ck+Q↓
  
 
k  +  
0 0 −η 0 k−Q 0   ck−Q↑
 
 ck−Q↑  
c+ −η 0 0 0 0 k−Q ck−Q↓
k−Q↓
 +
T
ck↑
  
k −η 0 0 0 0 ck↑
 c+
−η k−Q 0 0 0 0   ck−Q↓

 k−Q↓   
0  +   
 ck+Q↑  0 0 k+Q −η 0 0   ck+Q↑
X   
= +
  
 c 0 0 −η k 0 0   ck↓
   
k  +k↓
   
0 0 0 0 k−Q −η   ck−Q↑

 ck−Q↑   
c+ 0 0 0 0 −η k+Q ck+Q↓
k+Q↓
(10.44)
78 CHAPTER 10. HUBBARD MODEL

where η = mU/2. Although the Hamiltonian looks six-dimensional at first, there

is a pairwise scattering of one type of fermion into another, and no scattering
among the fermions belonging to different pairs. It can be diagonalized by using

c+
    
k↑ cos θ1 − sin θ1 0 0 0 0 a1↑
 c+
  sin θ1 cos θ1 0 0 0 0 a1↓
 
 k−Q↓  
 c+  
  0 0 cos θ2 − sin θ2 0 0 a2↑
   
 k+Q↑
=
 c+
 
  0 0 sin θ2 cos θ2 0 0 a2↓
  
 +k↓  
0 0 0 0 cos θ3 − sin θ3 a3↑
 
 ck−Q↑   
c+
k+Q↓
0 0 0 0 sin θ3 cos θ3 a3↓
(10.45)
with

2η 2η 2η
tan 2θ1 = , tan 2θ2 = , tan 2θ3 = .
k−Q − k k − k+Q k+Q − k−Q
(10.46)
The six eigenenergies are given by

1
q
±
E1k = [(k + k−Q ) ± (k − k−Q )2 + 4η 2 ]
2
1
q
±
E2k = [(k+Q + k ) ± (k+Q − k )2 + 4η 2 ]
2
1
q
±
E3k = [(k−Q + k+Q ) ± (k−Q − k+Q )2 + 4η 2 ].
2
√
√ bare dispersion is given by k = −2t[cos kx +cos(kx /2+ 3ky /2)+cos(kx /2−
The
3ky /2)] − µ.
+
Once we fix the chemical potential at zero, µ = 0, one can see that Eαk >0
−
but Eαk < 0 for all α = 1, 2, 3. (See Fig. ??.) That means as long as we
take µ = 0 the three lowest bands are completely filled, and the three highest
completely empty, with the net electron density of one per site, or half-filling.
+ −
The gap magnitude is given by min(Eαk − Eαk ) = 2η.
In the triangular lattice, the antiferromagnetic ordering at Q = (4π/3, 0)
introduces a spatial modulation that led to (i) tripling of the unit cell and 1/3-
ing of the Brillouin zone and (ii) folding of the band into six subbands and a gap
separating the lower three bands from the upper three. Full occupation of the
lower bands and a complete absence in the upper bands leads to the opening of
the gap at half-filling. The reduced Brillouin zone has the shape of a hexagon
with the edge-to-edge separation exactly equalling 4π/3.

10.4 Spin dependent hopping and DM interac-

tion
Under time reversal operation the spin-up and spin-down operators transform
as

ci↑ → ci↓ , ci↓ → −ci↑ . (10.47)

10.4. SPIN DEPENDENT HOPPING AND DM INTERACTION 79

The most general fermion hopping between a pair of sites hiji obeying the
 time-
ci↑
reversal symmetry can be written down using the spinor notation ψi = ,
ci↓
ˆ
Hij = tψi† [cos θ + i(dˆ · σ) sin θ]ψj = tψi† [eiθd·σ ]ψj . (10.48)
Here dˆ is an arbitrary
 3-dimensional
 real unit vector. The time-reversal opera-
0 1
tion renders ψi → ψi = iσy ψi and one can easily see that the above
−1 0
expression for Hij is invariant under such a change of ψi plus the conjugation
of idˆ · σ to −idˆ · σ ∗ . Note that σ ∗ = (σx , −σy , σz ). Most generally t, θ and dˆ
will depend on the bond, so we can write
ˆ
Hij = tij ψi† [eiθij dij ·σ ]ψj . (10.49)
To ensure that the Hamiltonian remains Hermitian we will require dˆji = −dˆij .
Let’s consider a specialized case where dˆij is uniform, so that we can take
dˆij = +dˆ for j = i + x̂ and j = i + ŷ, and θij = θ, tij = t everywhere. (I’m
assuming a square lattice.) Then the Hamiltonian can be written in the form

X  † ˆ † ˆ † ˆ † ˆ

H= t ψi+x eiθd·σ + ψi−x e−iθd·σ + ψi+y eiθd·σ + ψi−y e−iθd·σ ψi . (10.50)
i

The superexchange calculation in the case of general hopping matrix can be

worked out from the two-site model
ˆ
tij ψi† [eiθij dij ·σ ]ψj + U ni (ni − 1) + U nj (nj − 1). (10.51)
1
Following Aharonny et al. we can eliminate the unitary matrix by the trans-
formation
ˆ ˆ
ψi → [e−i(θij /2)dij ·σ ]ψi , ψj → [ei(θij /2)dij ·σ ]ψj . (10.52)
P
The Hubbard interaction U i ni (ni − 1) is invariant under the proposed trans-
formation. Then in the rotated basis we obtain the superexchange Hamiltonian

t2ij
Jij Si · Sj , Jij =
. (10.53)
U
Now we must rotate the spins back to the original basis.

1 † −i(θij /2)dˆij ·σ i(θij /2)dˆij ·σ

Si → ψ e σe ψi = cos θij Si + (1 − cos θij )dˆij (dˆij · Si ) − sin θij (dˆij × Si )
2 i
1 † −i(θij /2)dˆij ·σ i(θij /2)dˆij ·σ
Sj → ψ e σe ψj = cos θij Sj + (1 − cos θij )dˆij (dˆij · Sj ) + sin θij (dˆij × Sj ).
2 j
(10.54)

We have invoked the identities

(dˆ · σ)σ(dˆ · σ) = 2d(σ

ˆ · d)
ˆ − σ, σ(dˆ · σ) − (dˆ · σ)σ = 2i(dˆ × σ), (10.55)
1 L. Shekhtman, O. Entin-Wohlman, and Ammon Aharony, Phys. Rev. Lett. 69, 836

(1992).
80 CHAPTER 10. HUBBARD MODEL

to show that

ˆ ˆ ˆ dˆ · σ) − sin θ(dˆ × σ). (10.56)

e−i(θ/2)d·σ σei(θ/2)d·σ = (cos θ)σ + (1 − cos θ)d(

Now the inner product Si · Sj becomes

Si · Sj → (cos 2θij )Si · Sj − (sin 2θij )dˆij · (Si × Sj ) + (1 − cos 2θij )(Si · dˆij )(Sj · dˆij ).
(10.57)
The terms generated are superexchange, DM, and Kitaev interactions, respec-
tively.

10.5 Orbitally degenerate Hubbard model

Hubbard model treats a localized site subject to a strong Coulomb repulsion.
In real materials, the atomic site treated within the Hubbard framework has
not one, but several orbitals associated with the site. For example, transition
metals have partially occupied d-shells that play a dominant role in the mag-
netic properties. The d-orbitals are fairly localized, so that the local Coulomb
repulsion within an orbital and between different orbitals centered at the same
site are considered important. For such a system, the concept of Hubbard inter-
action needs to be generalized. In addition, multiorbital systems have the Hund
interaction which tends to align the spins for different orbitals. One may write
down a general interaction term between local charge densities and the spins,
U1 X 2 U2 X X X
V = nα + nα nβ − J1 sα · sβ + J2 d+
α dβ . (10.58)
2 α 2
α6=β α6=β α6=β

We have the intra(inter)-orbital interaction energy U1 (U2 ), the Hund coupling

energy J1 , and J2 which represents the process of moving two electrons simul-
taneously from α-orbital to β-orbital. All four terms shown above are invariant
under the global spin rotation.
In the absence of crystal field splitting, all orbital states must be degenerate.
Real-space rotation mixes different orbital states α, but the interaction must
be written in a manner independent of the choice of real-space axis. The re-
quirement of rotational invariance dictates the relations between the coefficients,
U1 , U2 , J1 , and J2 .
From Dworin and Narath [Phys. Rev. Lett. 25, 1287 (1970)] we find that
the choice
U1 = U + J, U2 = U − J/2, J1 = J, J2 = 0 (10.59)
for the l = 2, five-fold degenerate d-orbitals. The sum over α and β goes from
−2 to +2. This choice of parameters was adopted in later works such as M. J.
Rozenberg [Phys. Rev. B 55, 4855 (1997)] and J. E. Han et al. [ Phys. Rev.
B 58, 4199 (1998)].
On the other hand, Fujimori [Phys. Rev. B 51, 12880 (1995)] and Koga et
al. [Phys. Rev. Lett. 92, 216402 (2004)] adopted a different convention

U1 = U + 2J, U2 = U − J/2, J1 = J = J2 . (10.60)

It is possible that the choice of rotation-invariant interaction is not unique.

10.6. FOUR-BOSON THEORY 81

10.6 Four-boson Theory

We are interested in implementing a numerical scheme for finding the self-
consistent solution of the n-fold degenerate Hubbard model using the four-boson
theory (and its n-fold generalization) introduced by Kotliar and Ruckenstein
(KR). Below we work out the self-consistent equations which arise in the n = 1
case, for (I) the non-uniform situation (assuming no symmetries) and for (II) a
bi-partite case with magnetization.

10.6.1 Non-uniform Case

For the non-uniform situation the effective Hamiltonian to be solved is

X tij X
H = − (d+ + + + +
j pj−σ + pjσ ej )(pi−σ di + ei piσ )fjσ fiσ +
+
(ρiσ − µ)fiσ fiσ
mi mj iσ
hijiσ
X X X X
+ U d+
i di − λi (e+
i ei + p+ +
iσ piσ + di di − 1) − ρiσ (p+ +
iσ piσ + di di )
i i σ iσ
(10.61)
+
p
where we have introduced the abbreviation mi = niσ (1 − niσ ), niσ = hfiσ fiσ i.
In the KR theory this factor is expressed in terms of bosons, e.g. niσ →
1 − e+ +
i ei − pi−σ pi−σ , and becomes equal to the present definition only if the
constraints are satisfied exactly. Our scheme has the advantage that it will
simplify the minimization equation
PN considerably. Fermionic part of the Hamil-
tonian is diagonalized as fiσ = m=1 umiσ ψmσ for each species of spin σ. The
number of lattice sites is N .
Minimizing the free energy with respect to d+ +
i , piσ lead to the condition
2

 
1 X tij
di = P − pi−σ (p+ + +
j−σ dj + ej pjσ )hfiσ fjσ i , (10.62)

σ ρiσ + λi − U jσ
m i mj

 
1 X tij
piσ = − [ei (p+ + + + + +
j−σ dj + ej pjσ )hfiσ fjσ i + di (dj pjσ + pj−σ ej )hfj−σ fi−σ i] .

λi + ρiσ j
mi mj
(10.63)
Minimization with respect to e+i leads to
X tij
λi ei = − piσ (d+ + +
j pj−σ + pjσ ej )hfjσ fiσ i. (10.64)
jσ
m i mj

We canPcalculate di and piσ from Eqs. (10.62)-(10.63) and, using e+ i ei = 1 −

di di − σ p+
+
iσ p iσ
3
, obtain e i . Then Eq. (10.64) is used for λi .
2 Formally this can be achieved by differentiating H with respect to all the variables,

and setting them equal to zero. Expressions involving fermions are replaced by their finite-
temperature averages. Although in the actual calculation complex fields such as d+ i will be
treated as a real number, and thus d+ i = di , it is convenient during the derivation to keep the
complex nature of the boson fields.
3 In invoking this relation we are already carrying out the minimization with respect to λ .
i
82 CHAPTER 10. HUBBARD MODEL

+
ρiσ imposes the constraint hfiσ fiσ i = p+ +
iσ piσ + di di . They can be inde-
pendently obtained by diagonalization the fermion Hamiltonian and from Eqs.
+
(10.62)-(10.63). If hfiσ fiσ i turns out to be greater than p+ +
iσ piσ + di di one raises
(by hand) the value of ρiσ in the next diagonalization step. This will lower the
+
expectation value hfiσ fiσ i in the next iteration. Iteration will continue until the
identity is achieved.
Chemical potential is adjusted (also by hand) to satisfy the global constraint
1 X +
hf fiσ i = ν (10.65)
N iσ iσ

for the filling factor ν. Fermionic averages are given by

X
+
hfjσ fiσ i = u∗mjσ umiσ F (Emσ ). (10.66)
m

F (Emσ ) is the Fermi distribution function corresponding to energy Emσ .

10.6.2 Uniform Case

We specialize to the case of the above Hamiltonian where operators can take on
two different values on sublattice A and B. One can write
X
fiσ = eik·ri fAkσ
k
X
fjσ = eik·rj fBkσ (10.67)
k

for i and j belonging to sublattice A and B, respectively. k is restricted to

half the Brillouin zone. An example is the diamond-shaped space covered by
connecting (π, 0) − (0, π) − (−π, 0) − (0, −π) in the BZ.
Introducing similar notations for other variables, the effective Hamiltonian
becomes

X
+ + + +
H = k (zBσ zAσ fBkσ fAkσ + zAσ zBσ fAkσ fBkσ )
kσ
X X
+ +
+ (ρAσ − µ)fAkσ fAkσ + (ρBσ − µ)fBkσ fBkσ
kσ kσ
N X
+ × {U d+ A dA − λ A (e +
A e A + p+ +
Aσ pAσ + dA dA − 1)
2 σ
X
+ +
− ρAσ (pAσ pAσ + dA dA ) + (A → B)}. (10.68)
σ

Expressions for z are

1
zAσ = p (p+ +
A−σ dA + eA pAσ )
nAσ (1 − nAσ )
1
zBσ = p (p+ +
B−σ dB + eB pBσ ). (10.69)
nBσ (1 − nBσ )
10.6. FOUR-BOSON THEORY 83

The eigenenergies for the fermion part are given by4

r
1 1
Ekσ± = (ρAσ + ρBσ ) − µ ± (ρAσ − ρBσ )2 + 2k |zAσ |2 |zBσ |2 . (10.70)
2 4
Free energy of the system becomes

X  N
F = Ekσ± − T ln(1 + eβEkσ± ) + ×
2
kσ±
( )
X X
U d+
A dA − λA (e+
A eA + p+
Aσ pAσ + d+
A dA − 1) − ρAσ (p+
Aσ pAσ + d+
A dA ) + (A → B) .
σ σ
(10.71)

Minimization of F takes place with respect to d+ + + + + +

A , dB , eA , eB , pAσ , pBσ , ρAσ , ρBσ , λA , λB ,
+
and the chemical potential µ. ∂F/∂dA,B = 0 gives

1 2 X ∂Ekσ±
dA = × × F (Ekσ± )
ρA↑ + ρA↓ + λA − U N ∂d+
A
kσ±
1 2 X ∂Ekσ±
dB = × × F (Ekσ± ) . (10.72)
ρB↑ + ρB↓ + λB − U N ∂d+
B
kσ±

∂F/∂p+
A,Bσ = 0 gives

1 2 X ∂Ekσ0 ±
pAσ = × × F (Ekσ0 ± )
ρAσ + λA N 0 kσ ±
∂p+
Aσ

1 2 X ∂Ekσ0 ±
pBσ = × × F (Ekσ0 ± ) . (10.73)
ρBσ + λB N 0 kσ ±
∂p+
Bσ

∂F/∂e+
A,B = 0 gives

2 X ∂Ekσ±
λ A eA = × F (Ekσ± )
N
kσ±
∂e+
A

2 X ∂Ekσ±
λB dB = × F (Ekσ± ) . (10.74)
N
kσ±
∂e+
B

∂F/∂ρA,Bσ = 0 gives

2 X ∂Ekσ±
p+ +
Aσ pAσ + dA dA = F (Ekσ± )
N ∂ρAσ
k±
2 X ∂Ekσ±
p+ +
Bσ pBσ + dB dB = F (Ekσ± ) . (10.75)
N ∂ρBσ
k±

4 Since for each k and σ there is mixing between A and B, one ends up diagonalizing a 2 × 2

matrix, whose engenvalues are distinguished as Ekσ± .

84 CHAPTER 10. HUBBARD MODEL

Writing out the derivatives explicitly gives

+ +
1 1 X 2 |zBσ |2 pA−σ (pA−σ dA + eA pAσ )
dA = × × ±F (Ekσ± ) √k
ρA↑ + ρA↓ + λA − U N D nAσ (1 − nAσ )
kσ±
+ +
1 1 X 2 |zAσ |2 pB−σ (pB−σ dB + eB pBσ )
dB = × × ±F (Ekσ± ) √k .
ρB↑ + ρB↓ + λB − U N D nBσ (1 − nBσ )
kσ±
(10.76)

1 1 X 2
pAσ = × × ±F (Ekσ± ) √k ×
ρAσ + λA N D
k±
" #
2
|zB−σ | dA (pAσ d+
+ eA p+
A A−σ ) |zBσ |2 eA (pAσ e+ +
A + dA pA−σ )
+
nA−σ (1 − nA−σ ) nAσ (1 − nAσ )
1 1 X 2
pBσ = × × ±F (Ekσ± ) √k ×
ρBσ + λB N D
k±
" #
2
|zA−σ | dB (pBσ d+
+ eB p +
B B−σ ) |zAσ |2 eB (pBσ e+ +
B + dB pB−σ )
+ .
nB−σ (1 − nB−σ ) nBσ (1 − nBσ )
(10.77)

1 X 2 |zBσ |2 pAσ (pA−σ d+ +

A + eA pAσ )
λ A eA = × ±F (Ekσ± ) √k
N D nAσ (1 − nAσ )
kσ±

1 X 2 |zAσ |2 pBσ (pB−σ d+ +

B + eB pBσ )
λ B eB = × ±F (Ekσ± ) √k .
N D nBσ (1 − nBσ )
kσ±
(10.78)

 
1 X ρAσ − ρBσ
p+ +
Aσ pAσ + dA dA = F (Ekσ± ) 1 ± √
N 2 D
k±
 
1 X ρAσ − ρBσ
p+ p
Bσ Bσ + d+
d
B B = F (Ekσ± ) 1 ∓ √ (10.79)
N 2 D
k±

where D = (1/4)(ρAσ − ρBσ )2 + 2k |zAσ |2 |zBσ |2 . Condition on the chemical

potential reads
1 X
F (Ekσ± ) = ν. (10.80)
N
kσ±

Strategy for updating the variables are as follows. From Eqs. (10.76) and
+
(10.77) one obtains
P + updated values for dA , dB , pAσ , and pBσ . Using eA eA =
+
1 − dA dA − σ pAσ pAσ and the corresponding relation for B sublattice one
obtains eA , eB . From Eq. (10.78), using known values of eA , eB from above
10.6. FOUR-BOSON THEORY 85

one obtains λA , λB . Finally, Eq. (10.79) determines ρAσ − ρBσ . The even
combination ρAσ + ρBσ leads to the uniform shift of the chemical potential (see
Eq. (10.70)), or it can be absorbed into µ. Taking the difference of the two
equations in Eq. (10.79) gives

p+ + + + + +
Aσ pAσ + dA dA − pBσ pBσ + dB dB = hfAσ fAσ i − hfBσ fBσ i
1 X F (Ekσ± )
= (ρAσ − ρBσ ) × ± √ . (10.81)
N D
k±

The occupation of electrons with spin σ will be different on the two sublattices
provided ρAσ 6= ρBσ , indicating antiferromagnetic component of spin. In prac-
tice I think one should set ρAσ = −ρBσ = ρσ from the outset. This simplifies
the equation to

2 X F (Ekσ± )
p+ + + +
Aσ pAσ + dA dA − pBσ pBσ + dB dB = ρσ × ± √ (10.82)
N D
k±
√
for D = ρ2σ + k |zAσ |2 |zBσ |2 and Ekσ± = −µ ± D.
The ground state can be (i) uniform and paramagnetic, (ii) uniform and
spin-polarized (for nonzero ρσ ) or (iii) possess staggered magnetic moment.
86 CHAPTER 10. HUBBARD MODEL
Chapter 11

Spin Interactions

11.1 Higher-order exchange processes

The Heisenberg spin exchange model describes spin interactions of the type
Si · Sj for two neighboring spins. In real materials the interactions between
spin degrees of freedom are undoubtedly more complicated, but it is a highly
nontrivial matter to figure out what sort of terms, or processes, contribute
significantly to the Hamiltonian.
It is a curious fact of quantum mechanics that, for S = 1/2 spins, this
exchange term has a close connection with the pair exchange operator Pij which
simply swaps the quantum states located at i and j; Pij |αi βj i = |βi αj i. Here
α and β may represent any of the available quantum states. For S = 1/2, one
can readily verify
 
1
Pij = 2 Si · Sj + . (11.1)
4
To prove that the two operator expressions are identical requires that the action
of each operator on any given state yields exactly the same new state whether
we use Pij or 2(Si · Sj + 1/4). So, let’s prove it. With S = 1/2, there are only
four available states, | ↑↑i, | ↑↓i, | ↓↑i, | ↓↓i, so our task is reduced to checking
the validity of Eq. (11.1) for each of the four states. For | ↑↑i one gets

Pij | ↑↑i = | ↑↑i.

Now, let’s see what happens when we act with the r.h.s. of Eq. (11.1)
     
1 1 11 1
2 Si · Sj + | ↑↑i = 2 Siz Sjz + | ↑↑i = 2 + | ↑↑i = | ↑↑i.
4 2 22 4

OK, so both operators produce the same state, | ↑↑i, when acting on | ↑↑i. For
| ↑↓i, one has Pij | ↑↓i = | ↑↓i, because the up- and down-spins simply swap
their positions. Before acting with the r.h.s. of Eq. (11.1) again, first remember
that

1
Si · Sj = Siz Sjz + (Si+ Sj− + Si− Sj+ ). (11.2)
2

87
88 CHAPTER 11. SPIN INTERACTIONS

After collecting the result of action with each of the three operators of Eq. (11.2)
on | ↑↓i we obtain
 
1 1 1
2 Siz Sjz + Si+ Sj− + Si− Sj+ + | ↑↓i
2 2 4
 
1 1 1
= 2 − | ↑↓i + 0 + | ↑↓i + | ↑↓i
4 2 4
= | ↑↓i.

Again, we obtain agreement. The case for | ↑↓i and | ↓↓i states proceeds
similarly and need not be reproduced here. Hence we learn that the cherished
Heisenberg exchange model may equally well be written
JX
H= Pij .
2
hiji

In reality, there is no reason to just stop with exchanging two particles at a time.
Why not exchange the spins of three, or even four sites at once. For example,
the 3-site exchange process is depicted in Fig. ??.
The exchange process of this type is denoted Pijk , in obvious generaliza-
tion of the two particle exchange. In turn, the three particle exchange can be
decomposed as a series of two-particle exchanges:

Pijk = Pik Pij .

Proof : Label the states of i, j, k sites by A, B, C and denote the state

|Ai Bj Ck i. The cyclic exchange Pijk produces a state |Ci Aj Bk i. Acting on
the original state |Ai Bj Ck i with Pij gives |Bi Aj Ck i. Acting on this state with
Pik gives |Ci Aj Bk i, same as Pijk |Ai Bj Ck i. Acting with Pji Pjk or Pkj Pki leads
−1
to the same state. There is also the clockwise permutation of the states Pijk
which renders
−1
Pijk |Ai Bj Ck i = |Bi Cj Ak i
The same result is achieved if we act with Pij Pik , hence

−1
Pijk = Pij Pik = Pjk Pji = Pki Pkj .

In any sensible physical system the two processes - clockwise and anti-
clockwise - occur with same amplitude, and thus the most general 3-site ex-
change process is described by a term in the Hamiltonian
X
−1 1 X
(Pijk + Pijk )= (Pik Pij + Pji Pjk + Pkj Pki + Pij Pik + Pjk Pji + Pki Pkj ).
3
hijki hijki

Insert Pij = 2(Si · Sj + 1/4) into this, one gets

X
∼ (Si · Sj )(Si · Sk ) + (Sj · Sk )(Sj · Si ) + (Sk · Si )(Sk · Sj ) (11.3)
hijki

plus 2-site spin interactions which are readily absorbed into the Heisenberg term.
So, allowing a 3-particle exchange generates a new spin interaction term in the
11.1. HIGHER-ORDER EXCHANGE PROCESSES 89

Hamiltonian of the form shown in Eq. (11.3). The general spin Hamiltonian
involving 2- and 3-particle exchanges can be written down
X
H = J1 Si · Sj
ij
X
+J2 (Si · Sj )(Si · Sk ) + (Sj · Sk )(Sj · Si ) + (Sk · Si )(Sk · Sj ).(11.4)
hijki

Generalization to 4-particle exchange is also obvious, and is left to the reader.

90 CHAPTER 11. SPIN INTERACTIONS
Chapter 12

Spin Representation and

Spin Excitation

There are several kinds of representations for spin operators. Some of them are
exclusively used for S = 1/2, and others are applicable for general S. They also
lead to different ways of obtaining the spin excitations of a given spin interaction
model.

12.1 Semiclassical Equation of Motion

12.1.1 Ferromagnet
In the ferromagnetic spin model, how each spin will dance in coordination with
the others is determined by the Hamiltonian
X
H = −J Si · Sj , (12.1)
ij

and the equation of motion for the individual spin Si that follows from it. We
first remind ourselves of the fundamental commutation relations of spin

[Siα , Sjβ ] = iδij αβγ Siγ (~ ≡ 1). (12.2)

From this one can calculate, for instance,

dSix X y X
x
= −i[Six , H] = −J Sj Siz − Sjz Siy = −J


Sj × Si . (12.3)
dt j∈i j∈i

Combined with the equation of motion for the y- and z-components, we get

dSi X
= −J Sj × Si . (12.4)
dt j∈i

This equation is of a typical, nonlinear type because the quantities we must

solve for, Si , appears as a product on the r.h.s. In other words, one can never

91
92 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

solve the above equation exactly. However we must remind ourselves that it
is the small fluctuation away from the ground state that concerns us, and as
such we can decompose the spin operator as Si = hSi i + δSi = mi + δSi ,
where mi represents the ground state spin average. For a ferromagnetic ground
state, mi = S ẑ. A small fluctuation means |hSi i|  |δSi |. In making such
a comparison of magnitudes we must be careful to remember that mi is a
number, but δSi is not. So, it’s not altogether clear what exactly we mean
by the magnitude of δSi . Nevertheless, we can at least make the substitution
Si = mi + δSi as a matter of formal definition, and proceed to plug it into Eq.
(12.11).

d X


δSi = −J mj + δSj × mi + δSi ≈
dt j∈i
X  X  X 
−J mj × mi − J mj × δSi − J δSj × mi (12.5)
j∈i j∈i j∈i

2 P
ignoring terms of order δS . The sum j∈i comprises all the nearest-neighbor
sites j for i. Once again, the justification is that δS is a small quantity, and
(δS)2 , even smaller. We take the quantization axis ẑ and write mi = mẑ
everywhere. Then1
X 
d
δSi = −Jzmẑ × δSi − Jm δSj × ẑ. (12.6)
dt j

d
The ẑ-component of this equation
reads dt δSzi = 0, hence the fluctuation in the
z
ẑ-component of spin is zero, Si t = m. For the x̂-and ŷ-components we get
d x X
S = JzmSiy − Jm Sjy ,
dt i j∈i
d y X
S = −JzmSix + Jm Sjx . (12.7)
dt i j∈i

We have dropped the δ symbol in front. Introducing the complex notation

Zi ≡ Six + iSiy simplifies the equation a lot:
 
dZi X
= iJm  Zj − zZi  . (12.8)
dt j∈i

We will solve this equation using the ansatz

Zi = Z0 eik·ri −iωk t , (12.9)
which yields the self-consistency condition
 
X
ωk = Jm z − eik·(rj −ri )  . (12.10)
j∈i

1 I apologize for the multiple use of z here: z is the number of neighbors, or the coordination

number, whereas ẑ is the unit vector along the z-axis.

12.1. SEMICLASSICAL EQUATION OF MOTION 93

For a 3-dimensional cubic lattice,P j∈i eik·(rj −ri ) = 2(cos kx + cos ky + cos kz ),
P
and with z = 6, ωk = 2Jm(3 − α cos kα ). For 2D square lattice we will get
ωk = 2Jm(2 − cos kx − cos ky ). This is the desired spin wave energy for a spin
disturbance associated with a wavevector k.

12.1.2 Antiferromagnet
For an antiferromagnet, the equation of motion reads
dSi X
=J Sj × Si . (12.11)
dt j∈i

The linearized equation of motion is

X  X 
d
δSi = J mj × δSi + J δSj × mi . (12.12)
dt j∈i j∈i

We take the quantization axis ẑ and write mi = (−1)i mẑ. Then, the whole
square lattice can be divided into two sublattices: i ∈ A if (−1)i = 1, and i ∈ B
if (−1)i = −1. A lattice for which such a division into two sublattices can be
carried out is known as a bipartite lattice.
We write δSi as Si and note that mj for the neighbors j surrounding the
site i is always opposite to mi : mj = −mi .

X 
d
Si = zJSi × mi + J Sj × mi . (12.13)
dt j∈i

Finally, we note that the staggered nature of the average magnetization mi

implies that the spin fluctuation Si must contain both uniform and staggered
components. With this knowledge, one can write Si as

Si = S0i + (−1)i S1i . (12.14)

Inserting this expression into the equation of motion and writing down the
equations separately for the uniform and the staggered components, one obtains

X 
d 0
S = zJS1i × msi −J S1j × msi
dt i j∈i
X 
d 1
S = zJS0i × msi +J S0j × msi (12.15)
dt i j∈i

The staggered magnetization msi = (−1)i mi = mẑ is uniform. Since all the
quantities appearing in the above equation is slowly varying, one can write
S0i = S0 eik·ri −iωt , S1i = S1 eik·ri −iωt , and obtain

 
−iωS0 = mzJ S1 × ẑ − εk S1 × ẑ ,
 
−iωS1 = mzJ S0 × ẑ + εk S0 × ẑ . (12.16)
94 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

where we write zεk = j∈i eik·(rj −ri ) . Using the complex notation S 0 = Sx0 +
P

iSy0 , S 1 = Sx1 + iSy1 , we can re-write the equation as

ωS 0 = mzJ(1 − εk )S 1 ,
1
= mzJ(1 + εk )S 0 .
ωS (12.17)
p
It readily follows that ωk = mzJ 1 − ε2k .

12.1.3 Spiral ferromagnet

We work out the elementary excitations for the spiral ferromagnet given by the
Hamiltonian

X X 
HHDM = −J Si · (Si+x̂ +Si+ŷ ) − K Si ×Si+x̂ · x̂+Si ×Si+ŷ · ŷ
i i
X  
=− Si · JSj + KSj ×êji . (12.18)
ij

where we have introduced the unit vector êji extending from site i to site j.
Using the previously derived equation of motion for a ferromagnet, one can
write down the equation of motion for the spiral ferromagnet easily
dSi X 
= Si × JSj + KSj × êji . (12.19)
dt j∈i

As before, we use Si = mi + δSi . The zeroth order term vanishes by default.

In fact, the relation
X 
mi × Jmj + Kmj × êji = 0 (12.20)
j∈i

defines the classical ground state. The spiral spin state that concerns us is given
by mi = meik·ri + c.c. where

1
m = √ (ê1 − iê2 ), ê1 × ê2 = k. (12.21)
2
√
Here k̂ = (1, 1)/ 2 is the propagation vector of the spiral spins. One can show
that

X 
Jmj + Kmj × êji = αmi (12.22)
j∈i

with the constant α given by

√
α = 4J cos Q + 2 2K sin Q. (12.23)

From
√ an independent calculation we know that in two dimensions tan Q =
K/ 2J, so that
12.2. HOLSTEIN-PRIMAKOFF THEORY 95

√ √ p
α = 4J cos Q + 2 2K sin Q = 2 2 K 2 + 2J 2 . (12.24)
Then the linearized equation of motion for the spiral ferromagnet becomes
d X 
Si = mi × JSj + KSj × êji + αSi × mi . (12.25)
dt j∈i

We introduce a set of orthogonal vectors explicitly,

1
k = √ (1, 1, 0)
2
 1 1 
mi = − √ cos[Q · ri ], √ cos[Q · ri ], sin[Q · ri ]
2 2
 1 1 
ni = − √ sin[Q · ri ], √ sin[Q · ri ], − cos[Q · ri ] .
2 2
(12.26)
One can easily verify that the three vectors (ni , mi , k) form a right-handed
orthogonal basis at each site i. And using these orthogonal vectors, the equation
(12.25) can be decomposed into three components, Sik , Sim , Sin ,where
Sik = Si · k, Sim = Si · mi , Sin = Si · ni .
Inserting the vectors (12.26) into (12.25) and using some vector identities and
d mi
trigonometric formulas, one immediately obtains dt Si = 0, hence S mi = 0.
The other two equations are

d k  
S = −αSin − J cos Q Si−x̂n n
+ Si+x̂ n
+ Si−ŷ n
+ Si+ŷ
dt i
Kh 
k k k k

− √ cos[Q · ri ] Si+x̂ − Si−x̂ + Si−ŷ − Si+ŷ
2
 i
n n n n
+ sin Q Si−x̂ + Si+x̂ + Si−ŷ + Si+ŷ
d n  
S = αSik + J Si−x̂
k k
+ Si+x̂ k
+ Si−ŷ k
+ Si+ŷ
dt i
K n n
− √ cos[Q · (ri + x̂)]Si+x̂ − cos[Q · (ri − x̂)]Si−x̂
2

n n
+ cos[Q · (ri − ŷ)]Si−ŷ − cos[Q · (ri + ŷ)]Si+ŷ . (12.27)

where i in sub and superscript means ri , so i ± x̂ = ri ± x̂ and i ± ŷ = ri ± ŷ.

12.1.4 Multiple spiral ferromagnet

12.2 Holstein-Primakoff theory

There is a neater way of deriving the same spin wave spectrum we have just
derived above, using the method of Holstein-Primakoff (HP) representation of
spins. The technique applies somewhat differently for an antiferromagnet as it
does for a ferromagnet, and we start off with the simpler case of a ferromagnet.
96 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

12.2.1 Ferromagnet
First, I will introduce the representation (ni = b+
i bi ),

Si+ = Six + iSiy = (2S −ni )1/2 bi ,

Si− = Six − iSiy = b+
i (2S −ni )
1/2
,
Siz = S − ni , (12.28)

then verify that the new expressions on the r.h.s. of the above equations satisfy
all the commutation algebras of ordinary spin operators: [Siα , Sjβ ] = iδij αβγ Siγ ,
provided the new operator bi is a canonical boson operator, with the familiar
bosonic commutation relations, [bi , b+
j ] = δij , etc.
How this comes about does not concern us so much here; rather we will
proceed to apply the brand new technique and see if something good comes out.
First rewrite

1
Si · Sj = Siz Sjz + (Si+ Sj− + Si− Sj+ )
2
1 1 1
= (S − ni )(S − nj ) + (2S −ni ) 2 (2S −nj ) 2 (b+ +
i bj + bj bi ). (12.29)
2
OK, so it doesn’t appear particular illuminating after the substitution. But
let’s try to appeal to our physical senses to see if some simplifications can occur.
We already know the ground state of the Hamiltonian is given by a ferromagnetic
arrangement of spins, with all the spins pointing in the ẑ-direction. If we take
the average of the operator Siz in the ground state, at zero temperature, we
should get hSiz i = S. The same can be said of the r.h.s. of the last of Eq.
(12.28) because it is just a mathematically equivalent way of writing down the
same quantum-mechanical operator, so we must get hS − ni i = S, or hni i = 0!
This means that the quantum-mechanical ground state |0i of the ferromagnetic
Heisenberg spin model is defined by the condition that h0|ni |0i = 0 for all sites i.
An excitation is just a small wiggle of the spins away from the absolutely frozen
spin configuration that is the ground state, hence we can suppose that hSiz i
deviates only slightly from its ground state value S, or in the language of bosons,
that ni is some small, but non-zero number. In this limiting√ circumstance,
√ we
can simplify the square roots in Eq. (12.28) by taking 2S − ni ≈ 2S, and
writing

1
Si · Sj ≈ (S − ni )(S − nj ) + · 2S(b+ +
i bj + bj bi ).
2
≈ S 2 − S(ni + nj ) + S(b+ +
i bj + bj bi ). (12.30)

The ni nj is a product of two small numbers, so we delete them. The fully

HP-transformed Hamiltonian, after the approximations we have just made, will
be
X
HHP ≈ −JS (b+ + + +
i bj + bj bi − bi bi − bj bj ). (12.31)
ij
12.2. HOLSTEIN-PRIMAKOFF THEORY 97

Well, now the new Hamiltonian is just a quadratic Hamiltonian involving only
a product of two boson operators. Such operators can be diagonalized easily, in
this case by going to the Fourier space:
X
bi = eik·ri bk . (12.32)
k

Substitution of the Fourier expression for bi into Eq. (12.31) immediately gives
X
H= ω k b+
k bk (12.33)
k

where

ωk = JSz(1 − γk )
X
γk = z −1 eik·(rj −ri ) . (12.34)
j∈i

The eigenenergy ωk agrees with the spin-wave spectrum derived earlier, in Eq.
(12.10), using the equation-of-motion theory.

12.2.2 Antiferromagnet
The strategy for dealing with elementary spin excitations of an antiferromagnet
is essentially the same, but the technique is more advanced.
First of all, we define a new quantum-mechanical operator Siy , Siz related to
y
Si , Siz by the relation

Siy + iSiz = eiθi (Siy + iSiz ) (12.35)

where θi is the angle of the classical ground state spins, θi = π(ix + iy ). This
represents a rotation of spins, with respect to the x-axis, by the angle that just
equals the orientation of the classical ground state spins. The x-component
remains unchanged, hence Six = Six . Furthermore, one can easily verify that
the new operators Siα satisfy all the commutation rules of the original spins, Siz ,
hence they are as good a representation of the local spin operator as the previous
one we used. In terms of Si , we have the antiferromagnetic Hamiltonian written
as
X JX + + X
Si Sj + Si− Sj− − J Siz Sjz .


H=J Si · Sj → (12.36)
ij
2 ij ij

Note that we have two raising (lowering) operators instead of one raising and one
lowering, as was the case before. The ground state of the new Hamiltonian, Eq.
(12.36), is given by the rotation of the antiferromagnetic spins, hSiz i = Seiθi .
But this implies that in the new representation, hSiz i = S everywhere! Since
hSiy i = 0 initially, we have hSiy i = 0, too. So the new Hamiltonian has a
ferromagnetic ground state with all the spins pointing in the +ẑ direction, and
we can make the assumption hSiz i ≈ S, or hni i  S after the HP substitution,
for the small-fluctuation analysis.
Under these assumptions the HP-transformed Hamiltonian becomes
X
H ≈ JS (b+ +
i bj + bi bj + ni + nj ). (12.37)
ij
98 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

Unlike its ferromagnetic counterpart, Eq. (12.31), this Hamiltonian has a prod-
uct of two annihilation (creation) operators. As before, we take the Fourier
transform and obtain
Xh γk + + i
H = JSz b+ b k + (b b + bk b −k ) . (12.38)
k
2 k −k
k

Whereas doing the Fourier transform was sufficient to yield excitation en-
ergies in the previous example of ferromagnetic spins, now the Hamiltonian
written in momentum space fails to give eigenenergies. The final trick lies in
working out the equation of motion of the boson operators, which gives

[H, bk ] = −JSz bk + γk b+


−k . (12.39)

Here it’s clear why we could not get the energy from Fourier transform alone.
The dynamics of the operator bk is coupled to that of b+
−k , and the dynamics of
+
b−k is coupled to that of bk by the equation

[H, b+ +


−k ] = JSz b−k + γk bk . (12.40)

Noting that [H, X] = −idX/dt in quantum mechanics, and that for the eigen-
modes we can re-write the time derivative as −iE, where E is the eigenenergy,
the coupled oscillator equation we must solve becomes
    
bk 1 γk bk
Ek = JSz . (12.41)
b+
−k −γk −1 b+
−k
p
Eigenvalues are readily found as Ek = ±JSz 1 − γk2 . Here I state, without
proof, that the negative energies are just artifacts of the theory, and that only
the positive branch matters. Hence, the spin p waves in the antiferromagnetic
background carries the energy given by JSz 1 − γk2 for a wave vector k.

12.2.3 Spiral ferromagnet

We work out the HP theory of elementary excitations for the spiral ferromagnet
given by the Hamiltonian

X X 
HHDM = −J Sr · (Sr+x̂ +Sr+ŷ ) − K Sr ×Sr+x̂ · x̂+Sr ×Sr+ŷ · ŷ . (12.42)
r r

12.2.4 Multiple spiral ferromagnet

X X 
HHDM = −J Sr · (Sr+x̂ +Sr+ŷ ) − K Sr ×Sr+x̂ · x̂+Sr ×Sr+ŷ · ŷ
r r
X  X y

+A1 (Srx )4 +(Sry )4 +(Srz )4 − A2 Srx Sr+x̂
x
+Sry Sr+ŷ
r r
X
−H · Sr . (12.43)
r
12.3. SCHWINGER BOSONS 99

12.3 Schwinger bosons

Schwinger found a way to represent the spin operator using a pair of canonical
boson operators (b1 , b2 ) as

S+ = S x + iS y = b+
1 b2
S− = S x − iS y = b+2 b1
1 +
Sz = (b b1 − b+2 b2 ). (12.44)
2 1
More concisely one can write S = 12 b+
α σαβ bβ . The spin operators thus defined
obey the canonical spin operator relation

[S α , S β ] = iαβγ S γ . (12.45)
The original spin operator has the additional property S · S = S(S + 1) which
translates into (b+ + + +
1 b1 + b2 b2 )(b1 b1 + b2 b2 + 2)/4 upon the substitution of Eq.
(12.44). To match the original relation, the sum of the two boson occupation
numbers must equal 2S:

b+ +
1 b1 + b2 b2 = 2S. (12.46)
This puts constraint on the the allowed boson Hilbert space of a and b operators.
The spin state with total spin S and Sz = m is represented

(b+ )S+m (b+ )S−m

|Smi = p 1 p2 |0i. (12.47)
(S + m)! (S − m)!

For spin half the representation becomes | ↑i = b+ +

1 |0i, and | ↓i = b2 |0i.
We can also consider the parametrization
r r
S θ S θ
b1 = cos , b2 = sin e−iφ (12.48)
2 2 2 2
from which we deduce S = S(sin θ cos φ, sin θ sin φ, cos θ). For a pair of sites i
and j we have the relation2
 1/2
iωij /2  1 + Ωi · Ωj 
X
b+
 
b
jσ iσ = e  . (12.49)
σ
 2

ωij is the solid angle subtended by ẑ, Ωi and Ωj . Substituting Eq. (12.48) in
the action b+ i ∂τ bi also gives rise to the Berry phase action first discussed by
Haldane.
Now let’s see how the Heisenberg Hamiltonian looks like in the Schwinger
boson representation. Unlike the Holstein-Primakoff theory, we do not need to
know a priori what the classical ground state looks like. For this reason the
Schwinger boson theory is best applied to cases where the ground state does
not possess broken symmetry.
It is useful to decompose the spin-spin interaction term Si · Sj as
2 Patrick Lee, PRL 63, 680 (1989)
100 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

1
S i · S j = S 2 − A+ Aij (12.50)
2 ij
where Aij = b1i b2j − b2i b1j if it is an antiferromagnet, and as
1 + 1 +
Si · Sj = Bij Bij − S(S + 1) = : Bij Bij : −S 2 (12.51)
2 2
where Bij = b+ +
i1 bj1 + bi2 bj2 if it is a ferromagnet. With this substitution the
Heisenberg Hamiltonian, which originally expressed how the nearby spins inter-
act, becomes one for two bosons interacting with each other through the A+ ij Aij
term. Not surprisingly, such interaction Hamiltonian is not exactly solvable.
Instead one relies on an approximation scheme which renders the Hamiltonian
more analytically tractable. We call it the Schwinger boson mean-field theory
(SBMFT) which we describe below.

12.3.1 Ferromagnet
12.3.2 Antiferromagnet
Assume that one of the operators in A+ ij Aij can be replaced by its average,
Qij = hb1i b2j − b2i b1j i. Then we obtain the mean-field Hamiltonian
1X ∗
Qij Aij + Qij A+


HM F = − ij . (12.52)
2 ij
The new Hamiltonian contains only a pair of boson operators, and like all quan-
tum theories involving only a product of two operators, it is solvable.
Although the theory of diagonalization of the mean-field Hamiltonian can be
readily derived, we first introduce a method that will prove useful for treating
a variety of lattice situations. In spirit, this is similar to the rotation of the
classical spin angles to make all the spins aligned ferromagnetically along a
specific direction.
Let’s say the putative classical spin average is given by hSi i = SΩi and
Ωi = (sin θi cos φi , sin θi sin φi , cos θi ). This will in turn impose the condition on
the Schwinger boson average
hb+
i1 bi2 i = Se
iφi
sin θi
−iφi
hb+
i2 bi1 i = Se sin θi
1 +
hb bi1 − b+
i2 bi2 i = S cos θi . (12.53)
2 i1
Introduce the unitary matrix Ui
− sin(θi /2)e−iφi
 
cos(θi /2)
Ui = (12.54)
sin(θi /2)eiφi cos(θi /2)
which has the relation
Ui+ (Ωi · σ)Ui = σ z . (12.55)
It is now claimed that the rotation of the Schwinger boson spinor ψi = (bi1 bi2 )
by ψi → Ui ψi yields for the new Schwinger boson operators the average
1 +
hb+ +
i1 bi2 i = hbi2 bi1 i = 0, hb bi1 − b+
i2 bi2 i = S. (12.56)
2 i1
12.3. SCHWINGER BOSONS 101

It also implies that in the rotated basis the classical spin average becomes fer-
romagnetic, hSi i = S ẑ.
To make the case more concrete, consider a square-lattice antiferromagnetic
for which the classical spin orientation is hSi i = (−1)i S ẑ. The unitary matrix
(12.54) becomes a unit matrix for A sublattice sites and σx for all the B sub-
lattice sites. Therefore, for B lattice sites, (b1j , b2j ) becomes (−b2j , b1j ). The
pair amplitude Qij , i ∈ A, j ∈ B becomes b1i b1j + b2i b2j . Of course it doesn’t
really matter
P2 what we call the A sublattice, so for every pair of sites ij we have
Aij = m=1 bim bjm .
To carry out the mean-field analysis of the transformed Hamiltonian one
writes, similar to Eq. (12.57),
1X
Qij Aij + Qij A+


HM F = − ij . (12.57)
2 ij

with Aij = bi1 bj1 + bi2 bj2 , Qij = hAij i. We treat the mean-field amplitude as
real and uniform, Qij = Q, and obtain

2
QXX
HM F = − [bim bjm ] + h.c.
2 ij m=1
 
Q X ik·(rj −ri )  X
= − e bkm b−km + h.c.
4 j∈i
km
  
X
+ 0 −Qk b1k
= ( b1k b2k ) (12.58)
−Qk 0 b+
2k
k

where Qk = (Q/4) j∈i eik·(rj −ri ) . Inclusion of the Lagrange multiplier i λi (a†i ai +
P P

b+
i bi − 2S) renders the mean-field Hamiltonian
  
X λ −Qk b1k
HM F = ( b+ b ) . (12.59)
1k 2k −Qk λ b+
2k
k

One can bring the Hamiltonian to its diagonalized form by the rotation
    
b1k cosh θk sinh θk γ1k
= (12.60)
b+
2k
sinh θk cosh θk +
γ2k

where the angle θk is fixed by

λ Qk
cosh 2θk = , sinh 2θk = , (12.61)
Ek Ek
and the eigenenergy Ek is given by
q
Ek = λ2 − Q2k . (12.62)

The minus sign in the energy dispersion (12.62) raises concern for what
happens when λ < |Qk |. In fact, this never happens since the moment when
λ equals min[Qk ] is when the bosons begin to condense, and due to the Bose
condensation λ is always fixed to maintain λ = min[Qk ].
102 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION

When boson condensation happens, λ equals Q, and we have

q
Ek = Q 1 − (1/4)(cos kx + cos ky )2 . (12.63)

In fact, this is identical to the spin wave spectrum obtained using the Holstein-
Primakoff theory, provided we identify Q with JSz, J =exchange energy, S=spin,
z=coordination number. And this is no coincidence. One will find that the same
Schwinger boson calculation on the triangular lattice, at the onset of Bose con-
densation, also has an energy spectrum of the spin waves on the triangular
lattice. Schwinger boson theory has the power to capture the physics of both
the ordered and the disordered phases of the continuous magnet.
To complete the analysis, the self-consistency relation for Q is derived as

1X
Q= (12.64)
4 i

12.3.3 Spiral ferromagnet

12.3.4 Multiple spiral ferromagnet

12.4 Slave boson

In a strong correlation environment there often arises a situation for which each
site has a low probability of being occupied by two electrons at once, because of
the high energy cost of on-site Coulomb repulsion. Because of this it is useful to
formulate the theory using the Hilbert space in which the doubly occupied state
is truncated out from the outset. In such a limited Hilbert space consisting of
empty (|0i), or singly occupied (|σi) site, the electron commutation is modified
in an odd way. See for example,

{cσ , c+
σ } = |0ihσ|σih0| + |σih0|0ihσ| = |0ih0| + |σihσ|. (12.65)
Normally this should have been one, but due to the limitations of the Hilbert
space, the electron creation operator can only connect an empty state to a
singly occupied state, but not a singly occupied state to a doubly occupied
state. Besides, the result (12.65) equals 1 − |σihσ| because we have
X
|0ih0| + |σihσ| = 1, (12.66)
σ

also due to the limited Hilbert space imposed by the no double occupancy
constraint. Using a similar idea, we can derive

{cσ , c+
σ} = 1 − cσ+ cσ
{cσ , c+
σ} = c+
σ cσ . (12.67)
P
On the other hand, commutation relation with the density operator n = σ |σihσ|
is the same.

[n, c+ +
σ ] = cσ , [n, cσ ] = −cσ . (12.68)
12.5. SLAVE-FERMION SCHWINGER BOSON 103

The slave-boson approach re-write the electron operator in this restricted

Hilbert space as a product of a boson (bi ) and a fermion (fiσ ) operator

cσ = b+ fσ . (12.69)
Quite remarkably, this prescription recovers all the commutations of the re-
stricted Hilbert space, (12.67)-(12.68), provided we impose the constraint
X
b+ b + fσ+ fσ = 1. (12.70)
σ

In this representation, the electron number operator becomes

X
c+ + + + +
σ cσ = bb fσ fσ = (1 + b b)fσ fσ = (2 − fσ+ fσ )fσ+ fσ = fσ+ fσ . (12.71)
σ

Similarly one can show that the spin operator has the expression
1 + 1
S= c σαβ cβ = fα+ σαβ fβ . (12.72)
2 α 2

12.5 Slave-fermion Schwinger boson

The Schwinger boson formalism introduced in the previous section gave a good
way to handle the spin dynamics, especially in the incoherent regime before the
magnetic long-range order has taken place.
A system where all the dynamics arise from that of the spins is by definition
an insulator. On the other hand, a metallic state allows the motion of individual
electrons from one site to the next, and this requires the introduction of an
electron operator ciσ in addition to the spin operator Si in the Hamiltonian.
The Schwinger boson theory by itself does not tell us how one can write down
the electron operator in terms of the Schwinger bosons. To do that, one must
introduce yet another kind of operator, called the slave-fermion operator denoted
fi . Using this operator, an electron operator gets written as

cσ = f + bσ . (12.73)
Compared to the slave-boson substitution, here the role of boson and fermion is
simply reversed. The advantage is that now one obtains an easy generalization
of the Schwinger boson theory to treat the conducting electrons. Similar to the
slave-boson case, we have

c+ +
σ cσ = bσ bσ
c+ +
α σαβ cβ = bα σαβ bβ (12.74)

The latter equation is precisely the Schwinger boson prescription.

104 CHAPTER 12. SPIN REPRESENTATION AND SPIN EXCITATION
Chapter 13

t-J Model

The “standard model” for the high-Tc cuprates is the t-J model. A popular
method used in the solution of the t-J model is first to introduce the slave
boson coordinates to rewrite the Hamiltonian, and apply mean-field theory.
Other, more sophisticated techniques exist, e.g. DMRG, exact diagonalization,
QMC, etc, which I do not talk about here. While the mean-field theory is the
least convincing of all the techniques, it is the most versatile and can be applied
to all sorts of problems involving t-J Hamiltonian.
The t-J model is defined by the Hamiltonian

1 Vc X 1
(c†jα ciα + h.c.) + J
X X
H = −t (Si · Sj − ni nj ) + (ni − n̄)(nj − n̄),
4 2 rij
hiji hiji i6=j

(13.1)

with typically the Coulomb term Vc set to zero. Each site i is roughly like an
atomic orbital, whose wavefunction hybridizes with its neighbors’ to produce
itinerancy. Because of the large Coulomb energy of putting two electrons in a
given atomic orbital (a few eV), it is usually believed that double occupation of
orbitals never happens.
A naive diagonalization of the above Hamiltonian would naturally involve
states which have two electrons at a given site, | ↑↓ii , which should cost very
high energy. In the t-J model this energy is treated as infinity, and the diago-
nalization takes place within the restricted Hilbert space consisting of an empty
site, and single occupied sites of either spin orientation. This sort of prob-
lem was confronted sometime in early 80’s by using the so-called slave-boson,
or auxiliary-boson technique. In a nutshell, one writes an electron operator
as a composite of a boson operator bi and another fermion operator fiσ as
ciσ = b†i fiσ . In this new choice of variables, the no-double-occupancy constraint
becomes

b†i bi +
X †
fiσ fiσ = 1 (13.2)
σ

for all i. In loose terms it means that the occupation of a given orbital by a
vacancy (b†i bi ) or by a single spin (fiσ
†
fiσ ) should always add up to one. Double
P †
occupancy is then by definition excluded because σ fiσ fiσ = 1 − b†i bi ≤ 1.

105
106 CHAPTER 13. T-J MODEL

In the path-integral picture the constraint is implemented through the La-

grange multiplier technique, which amounts to adding to the original Lagrangian
the term
†
X
LtJ → LtJ + i (b+
i bi + fi fi − 1). (13.3)
i

The Hamiltonian can be accordingly rewritten as

1
(bj b†i fjα
†
X X
H = −t fiα + h.c.) + J (Si · Sj − ni nj )
4
hiji hiji
Vc X 1 †
λi (b†i bi +fiσ
X X †
+ (ni − n̄)(nj − n̄)+ fiσ − 1)−µ( fiσ fiσ −Ne ).
2 rij i iσ
i6=j
(13.4)

The Lagrange multipliers appear as λi and the chemical potential µ is adjusted

so as to obtain the desired total electron number Ne . The spin-spin interaction
term can be written using the fermion coordinates,
1 † αβ
Si = f σ fiβ . (13.5)
2 iα i
In other words,

† †
2Six = fi↑ fi↓ + fi↓ fi↑
† †
2Siy = −i(fi↑ fi↓ − fi↓ fi↑ )
† †
2Siz = fi↑ fi↑ − fi↓ fi↓ . (13.6)

This fermion representation of spin satisfies all the commutator algebra of the
original spin. By substituting Eq. (6) back into Eq. (5) one completes the
expression of the t-J model in terms of slave-boson coordinates bi and fiσ .
It is horribly difficult to solve a Hamiltonian which contains a quartic inter-
action term. The J-term is quartic in the fermions, and the t-term is quartic
because it contains two bi ’s and two fiσ ’s. One can reduce the problem to the
fermion-only problem by treating the boson fields bi as condensed, and satisfying
the constraint

hb†i bi i = 1 − hfiσ
†
fiσ i. (13.7)
In reality one replaces the boson fields in the Hamiltonian by a complex number
†
whose magnitude is determined from |bi |2 = 1 − hfiσ fiσ i. Because the Hamilto-
nian is invariant under a local phase change bi → eiθi bi , fiσ → eıθi fiσ , one can
always choose bi to be real and positive without loss of generality.
The density-density interaction term can be rewritten as
Vc X 1 2
(b − x)(b2j − x), (13.8)
2 rij i
i6=j

with b2i given as one minus the local electron density.

 107

Having removed the boson degrees of freedom, one can now face the quartic
interaction term by mean-field decoupling technique. One groups the quartic
term as a product of two bilinear operators, and replace one of them by its
expectation values. There are three distinct ways to pair up the fields, hence
three different order parameters in the meanfield theory. Keeping track of the
Fock and pairing terms in the mean-field decoupling scheme gives

3 s 1 t
Si · Sj → − (∆ij Pijs + χsij Hij
s
) + (∆ij Pijt + χtij Hij
t
)
8 8
1 1 s s 1 t
ni nj → (∆ P − χsij Hij s
) + (∆ij Pijt − χtij Hij
t
)
4 8 ij ij 8
(13.9)

plus their hermitian conjugates. The singlet/triplet operators are defined by

s/t
Pij = fi1 fj2 ∓ fi2 fj1 ,
s/t † †
Hij = fj1 fi1 ± fj2 fi2 , (13.10)

and their expectation values by

s/t s/t s/t s/t
∆ij = hPij i, χij = hHij i. (13.11)
In actual treatments, the triplet components are assumed to be zero. When I
didn’t know better, I decoupled the Si · Sj fully in all three channels, but did
the Hartree decoupling only with respect to the − 41 ni nj in the J-term1 . As a
result what I had for the mean-field Hamiltonian is

† 3J X
(∆ij Pij† + χij Hij
†
X
HM F = −t (bj bi fjα fiα + h.c.) − + h.c.)
8
hiji hiji
JX JX
+ Mi σiz − (1 − b2i )(1 − b2j )
4 i 4
hiji
Vc X 1 2 X †
X †
+ (bi − x)(b2j − x) + λi (b2i + fiσ fiσ − 1) − µ fiσ fiσ ,
2 rij i i
i6=j
(13.12)

where Pij = Pijs , and Hij = Hij

s
and ∆ij = ∆sij , Kij = Kij
s
, respectively. σiz is
the Pauli matrix for the local magnetic moment whose axis is chosen in the z-
direction,
P and Mi is the local magnetic field generated by the neighbouring spins,
Mi = j hσjz i (j=four nearest neighbours of i). Summation of the spin index
σ is implicitly assumed. One can also find treatments where coefficients are J/2
for the pairing term, and J/4 for the hopping term, because they decoupled
ni nj according to Eq. (13.9). We will distinguish them as scheme I (sI) and
scheme II (sII). The maximal gap value will be equal to 4 × (3J/8)|∆ij | (sI), or
to 4 × (4J/8)|∆ij | (sII).
1 Also other people decoupled it this way in order to preserve the SU(2) symmetry which

came from the Heisenberg model.

108 CHAPTER 13. T-J MODEL

The fermion part of the Hamiltonian is expressed in the Nambu form by

introducing the two spinors
   
fi↑ fi↓
ψi↑ = + , ψ i↓ = + , (13.13)
fi↓ −fi↑
and the matrices

−χ∗ij − (8t/3J)bi bj
 
∆ij
Tij = ,
∆∗ij χij + (8t/3J)bi bj
 
λi + σmi 0
Miσ = , (13.14)
0 −λi + σmi
as

3 X + 1X +
HM F = J ψjσ Tij ψiσ + ψ Miσ ψiσ
8 2 iσ iσ
hijiσ
3 XX + 1X +
= J ψjσ Tij ψiσ + ψ Miσ ψiσ . (13.15)
16 iσ j∈i 2 iσ iσ

For actual calculation it is easier if we re-arrange the ψjσ to appear on the right:

3 XX + 1X +
HM F = J ψiσ Tji ψjσ + ψ Miσ ψiσ ,
16 iσ j∈i 2 iσ iσ
 
−χij − (8t/3J)bi bj ∆ij
Tji = .
∆∗ij χ∗ij + (8t/3J)bi bj
(13.16)

From now on we will write χ0ij = χij + (8t/3J)bi bj .

The above fermion Hamiltonian can be diagonalized by an appropriate Bo-
goliubov rotation of the operators. We will first work out the case without
magnetism, so that Miσ = Mi = λi σz . The spinors are rotated according to the
Bogoliubov matrix (1=spin up, 2=spin down)

X  uni −v ∗ 
ψiσ = ni Γnσ ,
vni u∗ni
n
   
γi1 γi2
Γn1 = + , Γ i2 = + . (13.17)
γi2 −γi1

Provided the eigenfunctions (uni , vni ) obey the equation (taking J ≡ 1)

3 3
− χ0ij unj + ∆ij vnj + λi uni = En uni
8 8
3¯ 3
∆ij unj + χ̄0ij vnj − λi vni = En vni , (13.18)
8 8
we will have
 109

∗ ∗
   
X unj −vnj uni vni
Tji = En
vnj u∗nj vni −u∗ni
j∈i
X X  u∗ ∗ ∗
  
ni vni unj −vnj
Tji = E n σz ,
−vni uni vnj u∗nj
i j∈i
(13.19)

and the mean field Hamiltonian in diagonal form

3 X 3X
HM F = En Γ+
nσ σz Γnσ =
+
En γnσ γnσ . (13.20)
16 n 8 nσ

Self-consistent parameters are calculated according to the formulae (F (x) ≡

1/(exp(x/T ) + 1)):

X X 
χij = h +
fjσ fiσ i = uni u∗nj F (En ) + vni
∗
vnj F (−En )
σ n
X 
∗ ∗
∆ij = hαβ fiα fjβ i = uni vnj F (−En ) − vni unj F (En )
n
X † X 
ni = h fiσ fiσ i = |uni |2 F (En ) + |vni |2 F (−En ) .
σ n
(13.21)
P
The sum n runs over both positive and negative energy sets. The usual
∗
identity that (uni , vni ) of energy −E is equal to (−vni , u∗ni ) at energy +E still
holds and allows the self-consistency equations to simplify to

X
χij = 2 uni u∗nj F (En )
n
X
∗
∆ij = −2 vni unj F (En )
n
X
ni = 2 |uni |2 F (En ). (13.22)
n

Now we put the magnetism back in and see how to diagonalize the mean field
Hamiltonian. We try a slightly different Bogoliubov rotation than the previous
case,

X X
∗ + ∗ +
fi1 = (uni γn1 − yni γn2 ), fi2 = (xni γn2 + vni γn1 ),
n n
X  uni ∗ ∗
 X  xni 
−yni −vni
ψi1 = ∗ Γn1 , ψi2 = ∗ Γn2 .(13.23)
vni xni yni uni
n n

This time the correct equations to be satisfied by (uni , vni ) are

110 CHAPTER 13. T-J MODEL

3 3
− χ0ij unj + ∆ij vnj + (mi + λi )uni = En uni
8 8
3¯ 3
∆ij unj + χ̄0ij vnj + (mi − λi )vni = En vni
8 8
(13.24)

with χ0ij ≡ χij + (8t/3)bi bj (sI), or

2 4
− χ0ij unj + ∆ij vnj + (mi + λi )uni = En uni
8 8
4¯ 2
∆ij unj + χ̄0ij vnj + (mi − λi )vni = En vni
8 8
(13.25)

with χ0ij ≡ χij + (8t/2)bi bj (sII). Eigenfunctions (xni , yni ) are determined from
the same equation with mi → −mi . Alternatively one can obtain (xni , yni ) of
energy E as (v̄ni , −ūni ) where (uni , vni ) is an eigenstate of energy −E. This
statement holds for arbitrary magnetic states: ferromagnetic, antiferromagnetic,
and what not.
Self-consistency equation after considering the relation between (xni , yni )
and (uni , vni ) becomes Eq. (13.21) without it being reducible to Eq. (13.22).
Additional relation for local magnetic moment is obtained

X † X 
Mi = h σfiσ fiσ i = |uni |2 F (En ) − |vni |2 F (−En ) . (13.26)
σ n

Specializing to uniform case: We think of a uniform situation where each

eigenstate
P ik·ri n is associated
P ik·rwith a momentum vector k and one can write uni =
k e u k , v ni = k e i
v k . For the square lattice one has

χi,i+x̂ = χ∗i+x̂,i = χx , χi,i+ŷ = χ∗i+ŷ,i = χy ,

∆i,i+x̂ = ∆i+x̂,i = ∆x , ∆i,i+ŷ = ∆i+ŷ,i = ∆y
(13.27)

with four complex parameters (χx , χy , ∆x , ∆y ). If the local magnetic moment

mi is staggered, a unit cell would have to be doubled. When we consider a
uniform magnetization due to, say, a strong external magnetic field, one can take
mi = m and the BdG equation in momentum space reduces to (χ0x = χx +8tδ/3,
etc.)

3 3
− [χ0x eikx +χ0x e−ikx + χ0y eiky +χ0y e−iky ]uk + [∆x cos kx +∆y cos ky ]vk +(m+λ)uk = Ek uk
8 4
3 3 0 −ikx
[∆x cos kx +∆y cos ky ]uk + [χx e +χx e +χ0y e−iky +χ0y eiky ]vk +(m − λ)vk = Ek vk .
0 ikx
4 8
(13.28)
 111

With the definitions

3 0 ikx 3
Ak = [χ e + χ0x e−ikx + χ0y eiky + χ0y e−iky ], Bk = [∆x cos kx + ∆y cos ky ],
8 x 4
(13.29)

I can reduce it to a simple-looking form

    
−Ak + λ Bk uk uk
= (Ek − m) . (13.30)
Bk∗ A−k − λ vk vk
0
Presumably it is OK to take χij to be real in this case, so that Ak =
p(3/4)[χx cos kx +
0
χy cos ky ] = A−k . The eigenenergies follow as Ek = m±εk , εk = (λ − Ak )2 + |Bk |2 .
The average quantities are

X 
χx = |uk |2 e−ikx F (Ek ) + |vk |2 eikx F (−Ek )
Ek
X 
χy = |uk |2 e−iky F (Ek ) + |vk |2 eiky F (−Ek )
Ek
X  
∆x = uk vk∗ e−ikx F (−Ek ) − eikx F (Ek )
Ek
X  
∆y = uk vk∗ e−iky F (−Ek ) − eiky F (Ek )
Ek
X 
n= |uk |2 F (Ek ) + |vk |2 F (−Ek ) .
Ek
(13.31)

For the positive energy branch, Ek = m + εk , we obtain

   
1 λ−Ak 1 λ−Ak 1 Bk
|uk |2 = 1+ , |vk |2 = 1− , uk vk∗ = . (13.32)
2 εk 2 εk 2 εk

For the negative energy branch, Ek = m − εk , we obtain

   
1 λ−Ak 1 λ−Ak 1 Bk
2
|uk | = 1− 2
, |vk | = 1+ , uk vk∗ = − . (13.33)
2 εk 2 εk 2 εk
112 CHAPTER 13. T-J MODEL
Chapter 14

Exact Spin Hamiltonians

14.1 AKLT States

The AKLT (Affleck, Kennedy, Lieb, Tasaki) Hamiltonian is given by

H AKLT =
X
P2S (ij), (14.1)
hiji

where P2S projects the spin operators Jij = Si + Sj for nearest-neighbor pair of
adjacent spins hiji onto a subspace of magnitude 2S. There are two constraints
on the spin operators in writing down the AKLT states,

1. Si2 = S(S + 1) for any spin Si .

2. The value of S must satisfy z = 2S, where z is the coordination number.

The projection operator is given by

2 2 2
Jij Jij −1·2 Jij − (2S − 1) · 2S
P2S (ij) = × × ··· × (14.2)
.
2S(2S + 1) 2S(2S + 1) − 1 · 2 2S(2S + 1) − 2S(2S − 1)

Any pair of spins with a net spin less than 2S is projected out, i.e. P2S (ij)|J <
2Si = 0. For J = 2S, which is the maximal value attainable for the sum of two
spin-S operators, we get P2S (ij)|2Si = |2Si. For S = 1,
2 2
Jij (Jij − 2) 1 2 2
P2 (ij) = = J (J − 2) (14.3)
2·3·4 24 ij ij
2
and since Jij = (Si + Sj )2 = 2S(S + 1) + 2Si · Sj ,

1 2
P2 (ij) = Si · Sj + (Si · Sj )2 + . (14.4)
3 3
Hence the AKLT Hamiltonian for a one-dimensional spin-1 chain is written as
Xh 1 i
H AKLT = Si · Sj + (Si · Sj )2 . (14.5)
3
hiji

113
114 CHAPTER 14. EXACT SPIN HAMILTONIANS

A similar reasoning leads to the AKLT Hamiltonian in two dimensional square

and triangular lattices. For a two-dimensional square lattice with spin-2 the
Hamiltonian is
AKLT =
X
H (Si · Sj )(Si · Sj + 3)(Si · Sj + 5)(Si · Sj + 6) (14.6)
hiji

The Hamiltonian of a two-dimensional triangular lattice with spin-3 is

AKLT =
X
H4 (Si · Sj + 12)(Si · Sj + 11)(Si · Sj + 9) ×
hiji

(Si · Sj + 6)(Si · Sj + 2)(Si · Sj − 3) (14.7)

By construction, the AKLT Hamiltonian is positive semi-definite and therefore

the ground state can be at most the zero energy eigenstate. Such a state can
be constructed in the Schwinger boson representation as
Y † †
|AKLTi = (ai bj − b†i a†j )|0i . (14.8)
hiji

Each hiji bond is covered by a dimer, a spin singlet, made up of two spin-1/2
constituents. The AKLT state has, for each hiji bond, a maximum Jij,z value of
2S−1. Then we must have J of that bond less than 2S, and P2S (ij)|AKLTi = 0.
To see that Jij,z does not exceed 2S−1, we take the square lattice as an example.
From Eq. (14.8), it is clear that a†i b†j − b†i a†j for the hiji bond gives Jij,z = 0.
According to Fig. ??, bonds hi2i, hi3i, and hi4i contribute a†i b†2 , a†i b†3 , and a†i b†4 ,
respectively and bonds hj5i, hj6i, and hj7i contribute a†j b†5 , a†j b†6 , and a†j b†7 ,
respectively. We have up to six up S = 1/2 spins or (a†i )3 (a†j )3 for the hiji bonds.
max
So Jij,z = 3 = 2(z − 1) · 21 = z − 1 = 2S − 1. Since the wavefunction does not
2
possess any J = 2S component for hiji bonds, one must have J2S (ij)|AKLTi =
0.

14.2 Majumdar-Ghosh states

The Majumdar-Ghosh (MG) model offers an exactly solvable model of spin-1/2
chains.
N
4J X 1 NJ
HMG = (Si · Si+1 + Si · Si+2 ) + , SN +1 = S1 . (14.9)
3 i=1 2 2

HMG possesses two degenerate ground states one of which is given by

N/2
Y 1  † †

|MGi = √ (a2i−1 b+
2i − b+
2i−1 2i |0i,
a (14.10)
i=1
2

and the other is a translation of this state by one lattice unit. The proof is as
follows: construct the total spin operator for a triad of adjacent spins,

Ji = Si−1 + Si + Si+1
1 3
Ji2 = J(J + 1), J = , . (14.11)
2 2
14.2. MAJUMDAR-GHOSH STATES 115

Since one is adding up three S = 1/2 operators, the Hilbert space for the triad
of spins must be decomposed into either J = 3/2 or J = 1/2 sectors. The
appropriate projection operator is constructed as

 
1 3
Ji2 − ≡ P3/2 (i − 1, i, i + 1) . (14.12)
3 4

If {Si−1 , Si , Si+1 } form a J = 1/2 state, then P3/2 (i−1, i, i+1) |J = 1/2i = 0,
but if J = 3/2 then P3/2 P(i − 1, i, i + 1) |J = 3/2i = |J = 3/2i. HMG is a sum
of projectors, HMG = J i P3/2 (i − 1, i, i + 1). Due to the preceding argument
about the allowed values of J, each average hP3/2 (i − 1, i, i + 1)i must be non-
negative, and we can anticipate that the zero-energy state, if it exists, is the
ground state.
Expanding
 
1 3
P3/2 (i − 1, i, i + 1) = (Si−1 + Si + Si+1 )2 −
3 4
 
1 3
= + 2(Si−1 · Si + Si · Si+1 + Si−1 · Si+1 )
3 2
1 2
= + (Si−1 · Si + Si · Si+1 + Si−1 · Si+1 ) . (14.13)
2 3
Hence,

N
!
X NJ 4 X 1X
HMG =J P3/2 (i − 1, i, i + 1) = + J Si · Si+1 + Si · Si+2 .
i=1
2 3 i
2 i
(14.14)

Two of the three sites in the |MGi states are already bound into a singlet. The
portion of the wave function featuring the three sites i − 1, i, i + 1 are linear
combinations of

(a†i−1 b+ + † † † + + † +
i − bi−1 ai )ai+1 & (ai−1 bi − bi−1 ai )bi+1 . (14.15)

The singlet combination (a†i−1 b+ + †

i − bi−1 ai ) remains invariant under an arbitrary
†
SU(2) rotation of spins, and ai+1 and b+ i+1 rotates into linear combinations of
each other. So, no matter what the basis to write down the Schwinger bosons,
one always finds Jz = 1/2 for the triad, and thus J = 1/2. If no triad possesses
J = 3/2 component, P3/2 (i − 1, i, i + 1) |di± = 0 for all i, so HMG |di± = 0.
Since HMG is a sum of positive semi-definite operators, the zero-energy eigen
state is the ground state. The spin-spin correlation for |MGi is

 3/4, i = j
hSi · Sj i = −3/4, |i − j| = 1 (14.16)
0, |i − j| > 1
