Thermal and Quantum Phase Transitions

Lectures given at the

Les Houches Doctoral Training School in Statistical Physics 2015

Matthias Vojta, TU Dresden

October 9, 2015
Contents

Contents 2

1 Introduction 4
1.1 Recommended literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 What is a phase? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 What is a phase transition? . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 What is a quantum phase transition? . . . . . . . . . . . . . . . . . . . . 5

2 Classical phase transitions and universality 7

2.1 Generic phase diagrams of fluids and magnets . . . . . . . . . . . . . . . 9
2.2 Landau theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Landau functional . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Ginzburg-Landau theory and spatial correlations . . . . . . . . . 14
2.2.3 Fluctuations and Ginzburg criterion . . . . . . . . . . . . . . . . . 17
2.3 Critical exponents and universality . . . . . . . . . . . . . . . . . . . . . 18
2.4 Scaling hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Renormalization group 21
3.1 Concept of RG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Scaling transformation and scaling dimension . . . . . . . . . . . . . . . 22
3.3 Momentum-shell RG for the O(N ) φ4 model . . . . . . . . . . . . . . . . 23
3.3.1 Fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.2 Field renormalization and anomalous exponent . . . . . . . . . . . 32
3.4 Phase transitions and critical dimensions . . . . . . . . . . . . . . . . . . 33

4 Theoretical models for quantum phase transitions 35

4.1 Quantum Ising model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Quantum rotor model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.3 Coupled-dimer model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5 Quantum phase transitions: Primer 40

5.1 Classical and quantum fluctuations . . . . . . . . . . . . . . . . . . . . . 40
5.2 Phenomenology: Phase diagrams and crossovers . . . . . . . . . . . . . . 41

2
CONTENTS 3

5.3 Quantum φ4 theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

5.4 Quantum scaling hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.5 Quantum-to-classical mapping . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.1 Classical Ising chain . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.5.2 Scaling limit and universality . . . . . . . . . . . . . . . . . . . . 47
5.5.3 Mapping to a quantum spin . . . . . . . . . . . . . . . . . . . . . 49
5.5.4 Mapping: XY-chain ↔ O(2) quantum rotor . . . . . . . . . . . . 50
5.5.5 Mapping: Heisenberg chain ↔ O(3) quantum rotor . . . . . . . . 50
5.5.6 Quantum-to-classical correspondence: Rules and exceptions . . . 50

6 Magnetic quantum phase transitions 52

6.1 Order parameters and response functions . . . . . . . . . . . . . . . . . . 52
6.2 Properties of the quantum φ4 model . . . . . . . . . . . . . . . . . . . . . 54
6.3 Quantum Ising chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.3.1 Strong-coupling limit g  1 . . . . . . . . . . . . . . . . . . . . . 56
6.3.2 Weak-coupling limit g  1 . . . . . . . . . . . . . . . . . . . . . . 57
6.3.3 Exact spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
6.3.4 Results near criticality . . . . . . . . . . . . . . . . . . . . . . . . 60

7 Quantum phase transitions of bosons and fermions 63

7.1 Bose-Hubbard model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.2 Bose-Hubbard model: Mean-field theory . . . . . . . . . . . . . . . . . . 64
7.3 Superfluid-insulator transition: Field theory . . . . . . . . . . . . . . . . 66
7.3.1 Quadratic vs. linear time derivative . . . . . . . . . . . . . . . . . 67
7.4 Dilute Bose gas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
7.5 Dilute spinless Fermi gas . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
7.6 Fermi-Hubbard Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.6.1 Antiferromagnetism in the Mott insulator . . . . . . . . . . . . . 73
7.6.2 Phases of the Hubbard model . . . . . . . . . . . . . . . . . . . . 74

Bibliography 75
Chapter 1

Introduction

1.1 Recommended literature

• N. Goldenfeld, Lectures on Phase Transitions and the Renormalization
Group

• S. Sachdev, Quantum Phase Transitions (2nd ed)

• S. K. Ma, Modern Theory of Critical Phenomena

• H. E. Stanley, Introduction to Phase Transitions and Critical Phenomena

• P. M. Chaikin, T. C. Lubensky, Principles of Condensed Matter Physics

• J. W. Negele, H. Orland, Quantum Many-Particle Systems

• M. E. Peskin, D. V. Schroeder, An Introduction to Quantum Field Theory

• A. Auerbach, Interacting Electrons and Quantum Magnetism

• K. Yosida, Theory of Magnetism

1.2 What is a phase?

Phase (a definition attempt): Equilibrium state of matter whose qualitative character-
istics do not change upon small changes of external parameters (alternatively: which
is stable under infinitesimal parameter changes). Within a phase, the thermodynamic
potential varies in an analytic fashion.
Phases can be characterized by their symmetries (more precisely: the symmetries of
their density operator). Different phases are separated by phase transitions.

4
Chapter 1. Introduction 5

1.3 What is a phase transition?

Phase transition: Point in parameter space where the equilibrium properties of a sys-
tem change qualitatively. The system at a phase transition is unstable w.r.t. infinitesimal
changes of external parameters; this implies non-analytic behavior of the thermodynamic
potential as function of system parameters.
Thermal phase transitions occur upon varying the temperature. Often, the transi-
tion is from an ordered low-temperature phase to a disordered high-temperature phase.
Thus, order is destroyed by thermal fluctuations. Phase transitions can be continuous
or discontinuous, see Chapter 2.

1.4 What is a quantum phase transition?

Quantum phase transitions (QPT) are phase transitions at temperature T = 0 which
occur upon varying a non-thermal control parameter (such as pressure, magnetic field, or
chemical composition). A QPT implies non-analytic behavior of the ground-state energy
as function of the control parameter.
As there are no thermal fluctuations at zero temperature, a QPT is apparently driven
by “quantum fluctuations”. Note that the term “fluctuations” has to be used with care:
At zero temperature, a quantum system is described by a single phase-coherent (many-
body) wavefunction. The term “fluctuations” is most appropriately used to characterize

T quantum critical region

(with experimentally accessible
signatures of the QPT)

classical PT

ordered disordered
0
QPT r
QCP

Figure 1.1: Generic phase diagram in the vicinity of a continuous quantum phase tran-
sition. The horizontal axis represents the control parameter r used to tune the system
through the QPT, the vertical axis is the temperature T . The dashed lines indicate
the boundaries of the quantum critical region, located above the quantum critical point
(QCP). The solid line marks the finite-temperature boundary between the ordered and
disordered phases. Close to this line, the critical behavior is classical.
6 Chapter 1. Introduction

deviations from a reference state (e.g. an ordered magnet).

Experimentally, zero temperature cannot be accessed. Why are QPT interesting
(and more than just an academic curiosity)? To understand this, we have to consider
the generic “quantum critical” phase diagram as function of both the non-thermal control
parameter and temperature, Fig. 1.1. For a continuous QPT, the zero-temperature axis
is divided into two stable phases separated by a quantum critical point (QCP). The
finite-T part contains two low-temperature regimes derived from the stable phases and
a so-called “quantum critical” regime at elevated temperatures above the QCP.
The stable phases typically display conventional quasiparticle excitations, i.e., objects
with conventional quantum numbers (and statistics) which can be identified through
sharp peaks in appropriate spectral functions; examples are particle-hole excitations in a
Fermi liquid or spin waves in an ordered magnet. In such a situation, the low-temperature
thermodynamics can be described as thermal occupation of these quasiparticle states,
e.g., leading to a specific heat of C(T ) ∝ T in a Fermi liquid. In contrast, the elementary
excitations of the quantum critical ground state can usually not be characterized as
quasiparticles, but instead there is a critical continuum of excitations. (Examples will
be presented in the course of this lecture.) Now, upon increasing temperature starting
from the QCP, this critical continuum will be excited, resulting in power-law behavior
of thermodynamic observables as function of temperature with non-trivial exponents.
These power laws are the experimentally accessible signatures of quantum criticality;
they signal the ground state at the QCP being a “novel state of matter”.
Chapter 2

Classical phase transitions and

universality

To set the stage, we start by summarizing a few terms and definitions which are useful
to characterize phase transitions. In cases where we need to be specific we will use the
terminology appropriate for an order-disorder transition of a ferromagnet; the concepts
are much more general.

Order parameter (OP): The OP is an observable ϕ, which is


= 0 in the disordered
hϕi phase. (2.1)
6= 0 in the ordered

Here, the brackets h·i refer to a thermodynamic average in a suitable Gibbs ensem-
ble and, for quantum systems, to a quantum-mechanical expectation value.
In most cases, the OP is chosen to be a local observable where “local” means that
ϕ can be (in principle) defined at every point in space and time. Starting from a
microscopic description, such a definition may require appropriate coarse-graining,
i.e., a suitable average over a larger (i.e. mesoscopic) number of microscopic con-
stituents. (Counter-example: The volume enclosed by the Fermi surface of a metal
is not a local observable.)
For the ferromagnet the natural order parameter is the local magnetization. In gen-
eral, the choice of an order parameter is not unique, and in some cases a suitable or-
der parameter is not even known (e.g. for the interaction-driven metal-to-insulator
transition).

First-order transition: The OP changes discontinuously at the transition.

Continuous transition: The OP varies continuously across the transition.

Critical point: The transition point of a continuous transition.

7
8 Chapter 2. Classical phase transitions and universality

Figure 2.1: Discontinuous and continuous behavior of the order parameter as function
of a control parameter r for a first-order and continuous phase transitions, respectively.

Correlations: For an order parameter depending on space and time, ϕ = ϕ(~r, t), cor-
relation functions can be defined by hϕ(~r, t)ϕ(~r 0 , t0 )i.

Correlation length ξ: In stable phases the order-parameter correlation function typ-

|~
r −~r 0|
ically follows an exponential law hϕ(~r)ϕ(~r 0 )i − hϕ(~r)ihϕ(~r 0 )i ∝ e ξ , with a
correlation length ξ.

Upon approaching a critical point this correlation length ξ diverges, such that
ξ = ∞ at criticality where the exponential decay with distance is replaced by a
power-law decay (recall that power laws are scale-invariant). Near criticality, the
correlation length is the only length scale characterizing the low-energy physics.

Spontaneous symmetry breaking: Consider a system, described by the Hamiltonian

Ĥ, and a symmetry operation Ŝ, with Ĥ being symmetric under Ŝ: [Ĥ, Ŝ]− = 0.
If the system’s density operator ρ̂ is not symmetric under Ŝ, [ρ̂, Ŝ]− 6= 0, then the
symmetry Ŝ is spontaneously broken.

To allow for symmetry breaking in Gibbs ensembles (where [Ĥ, Ŝ]− = 0 implies
[e−β Ĥ , Ŝ]− = 0) it is necessary to include a (small) field conjugate to the order
parameter: Ĥ → Ĥ − hϕ. Then, spontaneous symmetry breaking is equivalent to
the limit h → 0 being singular: limh→0 ρ̂(h) 6= ρ̂(h = 0).

Symmetry breaking requires the thermodynamic limit, N → ∞, as a finite-size

system cannot show a finite-temperature phase transition (e−β Ĥ being a product
of a finite number of exponentials cannot be singular). Consequently, sponta-
neous symmetry breaking is equivalent to the non-commutativity of two limits:
limh→0 limN →∞ . . . 6= limN →∞ limh→0 . . .
Chapter 2. Classical phase transitions and universality 9

2.1 Generic phase diagrams of fluids and magnets

Fig. 2.2 illustrates the phase diagrams of systems near critical points, showing similarities
and differences between (a) a fluid which in particular displays a liquid–gas transition
and (b) a magnet which displays ferromagnet–paramagnet transition. One key difference
is that both the line of first-order transitions and the critical endpoint have enhanced
symmetry in the magnetic case, but not in the fluid case. Many aspects of these phase
diagrams will be discussed in more detail below.

Figure 2.2: Different representations of the generic phase diagrams of (a) fluids and (b)
magnets.
10 Chapter 2. Classical phase transitions and universality

2.2 Landau theory

We continue with a brief summary of Landau’s theory of phase transitions. Landau the-
ory is based on the concept of a local order parameter ϕ, which is assumed to be small
– the latter assumption is justified near critical points. The goal is to derive thermody-
namic properties near phase transitions. A major obstacle in simple calculations is that
the thermodynamic potential and its derivatives are non-analytic (as function of external
parameters) at the critical point, i.e., no Taylor expansion is possible. In Landau theory,
this problem is circumvented by considering a generalized potential which includes an
explicit dependence on the order parameter itself.

2.2.1 Landau functional

Landau’s approach starts by introducing a generalized thermodynamic potential f , which
depends not only on the external variables (of the considered ensemble), but also on
the order parameter. Consider e.g. a ferromagnet (FM): In equilibrium, specifying
the external magnetic field H and the temperature T fixes the magnetization ϕ ≡ M .
Landau’s approach implies to generalize the equilibrium potential f (T, H) to f (T, H, ϕ).
Subsequently, it is assumed that f (T, H, ϕ) is non-singular as function of the order
parameter ϕ. Further, it is postulated that the equilibrium state of the system is given
by ∂f (T,H,ϕ)
∂ϕ
= 0 which defines ϕeq . The (conventional) thermodynamic potential follows
as f (T, H) = f (T, H, ϕeq (T, H)).
In general, the form of f (T, H, ϕ) is not known. However, near criticality where ϕ is
small, we can expand – based on the assumption of analyticity – f (T, H, ϕ) in a Taylor
series in ϕ:
 
a 2 b 4 c 6
f = fn + f0 ϕ + ϕ + ϕ + · · · − ϕh (2.2)
2 4 6

where h = fH0 is the external field that couples to (or “is conjugate to”) the order
parameter, and fn is a weakly T -dependent term which can usually be ignored.
The above expansion keeps all symmetry-allowed terms: In the absence of a field, the
potential should be independent of the sign of ϕ – reflecting an underlying Z2 (or Ising)
symmetry ϕ ↔ (−ϕ) – this forbids odd powers of ϕ. For a vector order parameter ϕ ~
with underlying O(N ) symmetry, all terms should be invariant under O(N ) rotations,
such that ϕ ~2 = ϕ
~ ·ϕ ϕ2 )2 , (~
~ , (~ ϕ2 )3 etc. are allowed.
As Eq. (2.2) represents an effective model for the order parameter, its coefficients a,
b, . . . can be assumed to be smooth functions of those external parameters which preserve
the symmetries of the problem (in particular, of the temperature).

Discussion for h = 0 for the expansion up to 4th order: The minima of the
∂f
potential f (see Fig. 2.3) as a function of ϕ is determined by ∂ϕ = 0 and they are given
Chapter 2. Classical phase transitions and universality 11

by
(
0 a > 0 disordered state
ϕ= p (2.3)
± −a/b a < 0 ordered state

where b > 0 is required by stability (otherwise f would be unbounded from below, and
∂f
|ϕ| would tend to infinity). Note that ϕ = 0 is also a solution of ∂ϕ = 0 for a < 0, but
here it corresponds to a maximum of f , to be interpreted as an unstable state.

Figure 2.3: Functional dependence of the Landau potential f (2.2) on the order parameter
ϕ for h = 0.

Apparently, a = 0 corresponds to the phase transition point, T = Tc . As a should

be a smooth function of temperature, we can expand to lowest non-vanishing order
a(T ) = α(T − Tc )/Tc ≡ αt where we have introduced the reduced temperature

T − Tc
t= . (2.4)
Tc

The order parameter is then given by


 0 T > Tc
ϕ= (2.5)
r
± α
(−t) T < Tc
b

This square-root behaviour is characteristic of Landau theory. The two possible signs
correspond to spontaneous breaking of the p assumed Z2 symmetry; for a vector order
parameter its magnitude would be fixed by −a/b with its direction being arbitrary.
Thermodynamic observables follow from inserting ϕ into the expression for the Lan-
12 Chapter 2. Classical phase transitions and universality

dau functional and taking appropriate derivatives:

(
fn (T ) for T > Tc
f (T ) = 2
  2
fn (T ) − f0 α4b TcT−T
c
for T < Tc
(
∂f s0 (T ) for T > Tc
S/V = − ∂T = 2
s0 (T ) − f0 α2b TcT−T
2 for T < Tc
c


∂(S/V ) c0 for T > Tc
c=T ∂T
= α2 1
c0 + f0 2b Tc for T < Tc

The results for entropy and specific heat are schematically shown in Fig. 2.4. In par-
ticular, within Landau theory the specific heat displays a jump at Tc (not a power-law
singularity).

Figure 2.4: Landau-theory results for entropy and specific heat near the transition tem-
perature Tc .

Discussion for h 6= 0: Minimizing f with respect to ϕ yields the equation

αtϕ + bϕ3 = h . (2.6)

For t > 0 there is one solution, whereas t < 0 yields three solutions for small h (one again
corresponding to a maximum of f ) and a single solution for large h. The interpretation
is obvious, see Fig. 2.5: There is a global minimum of f for one sign of ϕ (depending
on the sign of h), and possibly a local (i.e. metastable) minimum for the other sign of
ϕ. For the ferromagnet, this reflects the fact that the field selects the state with spins
parallel to the field. 
It is useful to define ϕ(h) = ϕspont + χ(T ) · h, with χ := ∂ϕ being the order-

∂h 
h→0
parameter susceptibility and ϕspont being the zero-field value of the order parameter
Chapter 2. Classical phase transitions and universality 13

Figure 2.5: Landau potential f (2.2) for h 6= 0, displaying minima which correspond to
either stable or metastable states.

Figure 2.6: Singular behavior of the order-parameter susceptibility χ near TC . For

ferromagnets, the behavior for T > TC , χ = C/(T − TC ) with C the Curie constant, is
also known as Curie-Weiss law.

from Eq. (2.5). The calculation yields the susceptibility χ to be

 1
 T > Tc
t

χ(T ) = 1 . (2.7)

 T < Tc
2|t|
Hence, χ grows upon approaching the transition, and is infinite at Tc , see Fig. 2.6.
At Tc the ϕ2 term is absent in f , yielding the field dependence of the order parameter
along the critical isotherm,
ϕ ∝ h1/3 , (2.8)
see Fig. 2.7. (The general definition is h ∝ |ϕ|δ , here δ = 3.) In the language of a
ferromagnet, this singular behavior reflects the fact that, at criticality, it is very easy to
polarize the spins.
14 Chapter 2. Classical phase transitions and universality

Figure 2.7: Location of the critical isotherm and corresponding order-parameter behav-
ior.

Analysing the phase diagram

• Fig. 2.8 shows that there is a line of first-order phase transitions for T < TC at
h = 0 (i.e. from h < 0 to h > 0). This is a special case of a first-order PT
since the two phases are related by symmetry, so there is no latent heat in the
transition. (In general, the two phases on both sides of a first-order transition are
not symmetry-related, like for the liquid–gas transition, resulting in latent heat
given by the entropy difference Q = ∆S.)

• For T < TC and small h there is a coexistence region where we have two minima of
f , the stable and the metastable minimum. The metastable phase can be accessed
experimentally in a non-equilibrium setting: if h is varied with a finite rate through
h = 0, the system switches from the metastable to the stable phase in a delayed
fashion, leading to hysteresis.

• For larger h the metastable solution disappears, this marks the end of the coexis-
tence region.

• The continuous temperature-driven PT occurs at Tc at h = 0 in Fig. 2.8. In

contrast, for h 6= 0 there is no PT upon varying the temperature, because the
symmetry is already broken by nonzero h.

2.2.2 Ginzburg-Landau theory and spatial correlations

Up to now the order parameter has been assumed to be spatially homogeneous. This
assumption can be relaxed, by introducing a spatially varying order-parameter field:

ϕ → ϕ(~r) . (2.9)
Chapter 2. Classical phase transitions and universality 15

Figure 2.8: Continuous and first-order transitions in the temperature–field phase diagram
of a magnet as described by Landau theory.

Figure 2.9: Varying the temperature in the presence of a finite field results in a smooth
variation of the order parameter.
16 Chapter 2. Classical phase transitions and universality

Here, ~r is a continuous space coordinate. Such a continuum-limit description requires

appropriate coarse-graining of microscopic degrees of freedom, as noted above. This also
implies that, conceptually, ϕ(~r) should be a smooth function of ~r which captures spatial
variations on length scales much larger than microscopic lengths. Such variations can be
expected to come with an energy cost, such that the Landau functional generalizes to
 
a 2 b 4 2 ~ 2
f (T, H, ϕ(~r)) = fn + f0 ϕ(~r) + ϕ(~r) + ξ0 (∇~r ϕ(~r)) − ϕ(~r)h (2.10)
2 4
corresponding to Ginzburg-Landau theory. The gradient term can be understood as the
lowest additional term arising from an expansion in powers and gradients of ϕ which
is allowed by symmetry: Spatial inversion symmetry forbids odd powers of gradients.
Importantly, f (T, H, ϕ(~r)) is a local energy density, and the Ginzburg-Landau (free)
energy is Z
F = dd rf (~r) . (2.11)

Ginzburg-Landau theory has been particularly used in the field of superconductivity,

where a complex scalar field ϕ describes the superconducting condensate which can
become inhomogeneous in the presence of an external magnetic field (leading to vortex
states).
Using Ginzburg-Landau theory one can also calculate spatial correlations of the OP
ϕ. These are measured by the correlation function
G(~r, ~r 0 ) = hϕ(~r)ϕ(~r 0 )i − hϕ(~r)ihϕ(~r 0 )i = hϕ(~r)ϕ(~r 0 )i − ϕ20 (2.12)
where ϕ0 refers to the spatially homogeneous equilibrium value of ϕ. Note that in the
ordered phase, where ϕ0 6= 0, we have
lim hϕ(~r)ϕ(~r 0 )i = ϕ20 (2.13)
|~ r 0 |→∞
r−~

reflecting long-range order, and G(~r, ~r 0 ) measures correlations of fluctuations around ϕ0 .

We now sketch the calculation of G(r), for details see e.g. the book of Goldenfeld [2].
A deviation δϕ of the order parameter from its equilibrium value ϕ0 increases the free
energy by Z
f0  
~ 2 δϕ(~r)
δF = dd r δϕ(~r) αt + 3bϕ20 − ξ02 ∇ ~
r (2.14)
2
After a Fourier decomposition of δϕ the correlation function can be calculated as expec-
tation value weighted by exp(−βδF ). The result is
kT
G(k) = hδϕ(~k)δϕ(−~k)i = (2.15)
f0 (A + ξ02 k 2 )
where A = αt + 3bϕ20 . Fourier-transforming back to real space yields G(r). For large
r = |~r − ~r 0 | the correlation function has the limit
e−r/ξ(T ) (3−d)/2
G(r) ∝ ξ (2.16)
r(d−1)/2
Chapter 2. Classical phase transitions and universality 17

which is dominated by the exponential at large distances. The so defined correlation

length depends in Ginzburg-Landau theory on the reduced temperature t as
( √
ξ0 / αt T > TC
ξ(T ) = p (2.17)
ξ0 / 2α|t| T < TC

The divergence of ξ is then ξ ∝ |t|−ν , ν = 12 . A diverging correlation length implies that,

as we approach the critical point, there are locally ordered islands of increasing size. The
fact that, near criticality, the order parameter varies slowly in space is an a-posteriori
justification of the gradient expansion in Ginzburg-Landau theory.
As will be discussed later, the diverging correlation length is accompanied by a di-
verging correlation time, i.e., the fluctuations of these island become slower and slower
upon approaching criticality – a phenomenon known as critical slowing down.
Using the Ginzburg-Landau result (2.15), we find that the correlation function at the
critical point decays in a power-law fashion:

G(r) ∝ r2−d (ξ = ∞) (2.18)

Note that this result can obviously only apply to d > 2.

2.2.3 Fluctuations and Ginzburg criterion

So far, we have described Landau theory as a simple phenomenological theory describing
a phase transition, which can be applied to systems with one (or more) local order
parameters. As will become clear later, Landau theory can be interpreted as the saddle-
point solution to a more general theory (formulated in the path-integral language) where
the order-parameter field ϕ is allowed to fluctuate in space and time. As the saddle-
point solution fixes ϕ, its fluctuations are neglected in Landau theory. This is a serious
(mean-field-type) approximation.
We can assess the validity of the neglect of fluctuations using a simple estimate based
on Landau theory itself:

hδϕ(ξ) δϕ(0)i ∝ ξ 2−d ξ ∝ |t|−1/2

2−d

h δϕ( ~r = ξ(T ) ) δϕ( ~r = 0 ) i ↓ ξ(T ) ↓ d−4 0 d>4
∝ ∝ |t| 2 −→
ϕ20 ↑ |t| |t|→0 ∞ d<4
ϕ0 ∝ (−t)1/2
(2.19)
This estimate, known as Ginzburg criterion, shows that fluctuations become in-
evitably important upon approaching criticality in dimensions d < 4, whereas they are
unimportant for d > 4. This implies that Landau theory is asymptotically exact for
d > 4, in particular the critical exponents predicted by Landau theory (often dubbed
mean-field exponents) are exact for d > 4. In contrast, for d < 4 fluctuations cannot
be neglected close to criticality, and hence Landau theory fails. d+
c = 4 is called upper
critical dimension.
18 Chapter 2. Classical phase transitions and universality

Two remarks are in order: (i) It is interesting to see that Landau theory can predict
its own failure (many other approximate theories don’t). (ii) There are systems in d = 3
where mean-field exponents can be observed experimentally: This happens if the nu-
merical prefactor in front of relevant fluctuation terms is small. The Ginzburg criterion
remains valid, but applies only to the immediate vicinity of the critical point which may
be experimentally inaccessible. One important example is the superconducting transition
of conventional superconductors in d = 3: Here fluctuations become important only if
t < 10−10 or so, due to the small ratio of transition temperature and Fermi temperature.

2.3 Critical exponents and universality

As we have seen above, the critical point is characterized by a diverging correlation
length: ξ → ∞. This implies that the order-parameter fluctuations do not display a
characteristic length scale at criticality, hence fluctuations exist on all length scales. The
system is said to be scale-invariant, i.e., looks “similar” on all length scales.
A consequence of scale invariance is that observables depend on parameters (like t h
etc.) in the form of power laws, because power laws A ∼ xα are the only scale-invariant
dependencies. These power laws define critical exponents and are an important part of
critical phenomena. Commonly used critical exponents are given in Table 2.1, together
with their mean-field values (as obtained from Ginzburg-Landau theory).

Table 2.1: Table of critical exponents

Exp. Definition Conditions MF value (GL)

Specific heat α C ∝ |t|−α t → 0, h = 0 α=0

Order parameter β ϕ ∝ (−t)β t → 0− , h = 0 β= 1
2

Susceptibility γ χ ∝ |t|−γ t → 0, h = 0 γ=1

Critical isotherm δ h ∝ |ϕ|δ sgn(ϕ) t = 0, h → 0 δ=3
Correlation length ν ξ ∝ |t|−ν t → 0, h = 0 ν= 1
2

Correlation function η G(r) ∝ |r|−d+2−η t = 0, h = 0 η=0

dynamic z τc ∝ ξ z ∝ |t|−νz t → 0, h = 0 N/A

We have also quoted the dynamical exponent z which describes the relation between
the diverging correlation time, τc , and the diverging correlation length, ξ, according to
τc ∝ ξ z . As we will discuss in Chapter 5, this exponent plays a special role in classical
critical phenomena, as statics and dynamics decouple in a well-defined sense.
Critical phenomena display a high degree of universality. This means that critical
exponents are identical for classes of phase transitions, the so-called universality classes:
The exponents depend only on the dimensionality of the system and the symmetry of
Chapter 2. Classical phase transitions and universality 19

the order parameter.1 Consequently, e.g. all systems with Ising order parameter in three
dimensions share the same critical exponents.
Universality is rooted in the divergence of the correlation length: Given that critical
phenomena are determined by the physics at large length scales, microscopic details
become unimportant.

2.4 Scaling hypothesis

The scaling hypothesis is based on the assumption that, near a critical point, the order-
parameter correlation length ξ is the only relevant length scale.2 The concept of scaling
is exposed by considering the behavior of the system under the scaling transformations.

x → bx lengths
t → by t t reduced temperature
h → by h h external field (2.20)

Here, the changes of the parameters t, h compensate for the change in length scale: The
physics should be invariant under such a transformation (for appropriate yt , yh ), because
the change of ξ upon a change of parameters t, h corresponds to a change of the reference
length scale.

Scaling hypothesis for the free energy density (more precisely for the singular
part of it, fs ) is
fs (t, h) = b−d fs (byt t, byh h) (2.21)
for any value of b, which is the mathematical statement of a generalized homogeneity
law.
To gain insight, we first consider h = 0. The transformation |t| → byt |t| causes the
correlation length to change as ξ → b−νyt ξ (because ξ ∼ |t|−ν ). Requiring invariance
of physics under the combined transformation (2.20), this change must be undone by
the changes of length scales x → bx, which implies νyt = 1 or yt = ν1 . Thus, use
b = |t|−1/yt = |t|−ν and get the power law behaviour of the free energy density

fs (t, h = 0) = |t|dν fs (1, 0) (2.22)

where fs (1, 0) = const.

From fs one can derive all critical exponents and relate them to yt , yh . There exist
only two independent critical exponents for static critical phenomena. Therefore there
1
In addition to dimensionality and order-parameter symmetry, the character of the interactions can
also be relevant for determining the universality class of a system: systems with long-ranged interactions
behave differently as compared to those with short-ranged interactions.
2
See Goldenfeld’s book [2] for an exposition on why microscopic length scales are required for the
existence of anomalous exponents.
20 Chapter 2. Classical phase transitions and universality

exist several scaling relations, like e.g.

2 − α = dν ,
2 − α = 2β + γ ,
2 − α = β(δ + 1) .

Other forms of the scaling hypothesis: Eliminate yt by fixing b = |t|−ν and get,
e.g.
h
fs (t, h) = |t|dν fs (1, νy ) (2.23)
|t| h
The scaling hypothesis for the correlation function, e.g. reads:
 
1 r h
G(r; t, h) = d−2+η fG ; (2.24)
r ξ |t|νyh
1
where at the critical point ξ → ∞, h → 0 and G(r) = rd−2+η fG (0, 0). Now we can relate
η to other critical exponents by looking at the definitions of the exponents
Z Z  
h
d ν
χ ∼ d rG(r) ∝ drfG rt , νy r1−η
t h
Z  
h 1
= dxfG x, νy x1−η (2−η)ν
t h t
−γ
∼t (at h = 0)

using the substitution x = r|t|ν , and after doing the last integral, one gets a relation
γ = ν(2 − η).
At the level of the present discussion, the scaling hypothesis is a hypothesis. It can be
justified using RG tools. As we will see later, this simple form of the scaling hypothesis
(often dubbed “naive scaling”) only holds below the upper critical dimension, d < d+ c
(recall d+ +
c = 4 for magnets). For d > dc naive scaling is violated due to the presence of
dangerously irrelevant variables. In this case, some of the scaling laws do not hold – this
applies to so-called hyperscaling laws which involve the spatial dimension d (and scaling
laws derived from them).
Chapter 3

Renormalization group

3.1 Concept of RG
In a typical condensed-matter setting, we are often interested in the physics at large
length or time scales and at low energies, because the long-distance behavior of cor-
relation functions contains the information about spontaneous symmetry breaking and
hence about the phase. This is obvious for a magnet: a short-distance correlator, e.g.,
between neighboring spins, is finite at any temperature, whereas the correlation function
in the long-distance limit is zero above TC and finite below TC .
To access the long-distance, low-energy behavior efficiently, renormalization-group
approaches successively eliminate short-distance/high-energy degrees of freedom such
that the long-distance behavior remains unchanged. Generically, a decimation step mod-
ifies both the model itself and its coupling constants. Often, however, one can combine
the decimation step with a scaling transformation such that the model remains invariant,
and only its coupling constants change. This “renormalization-group flow” of coupling
constants under successive decimation of high-energy degrees of freedom (leading to
“running coupling constants”) is a central object in RG calculations.
Following the RG flow to longer and longer length scales (or smaller and smaller ener-
gies), the coupling constants will usually flow to asymptotic values (so-called fixed point
values). Fixed point values are either trivial – 0 or ∞, corresponding to stable phases –
or they are non-trivial, i.e., finite, corresponding to continuous phase transitions.1

Example: For a lattice spin model of a ferromagnet one can think of a simple real-
space decimation scheme, dubbed block-spin RG, see figure 3.1: Here, in each step the
lattice is divided into supercells, and the spins in each cell are joined into a superspin.
This maps the coupling constant J onto a new coupling constant J 0 .
Successive decimation steps define a discrete RG flow of J. The fixed points of this
flow are easily interpreted: J → 0 implies decoupled spins and hence corresponds to the
1
First-order phase transitions do not have a straightforward signature in RG. Often, the RG is
characterized by run-away flow, but other situations are possible. We will not discuss this further.

21
22 Chapter 3. Renormalization group

Figure 3.1: Schematic representation of block-spin RG.

disordered (i.e. high-temperature) phase. In contrast, J → ∞ implies spins being locked

together and corresponds to the symmetry-broken ordered (i.e. low-temperature) phase.
A typical flow diagram is shown for the block-spin RG in figure 3.2.

Figure 3.2: Generic RG flow diagram for a lattice spin system, as can be obtained from
block-spin RG.

3.2 Scaling transformation and scaling dimension

Each RG step involves – in addition to the decimation of degrees of freedom – a rescaling
of lengths and momenta to account for the change of the “observation scale”. Explicitly,
lengths and momenta change according to

k 0 = sk , x0 = x/s (3.1)

where s > 1 is a rescaling factor. In many cases, RG transformations are performed in

infinitesimal steps, such that a parametrization s = edl with infinitesimal dl is useful.
Then, the RG flow is described by differential equations.
Chapter 3. Renormalization group 23

Model parameters (fields, coupling constants) will change under scaling transforma-
tions as well; usually their transformation behavior is fixed within a certain RG scheme
by imposing suitable invariance conditions onto the model, see below for examples.2
For infinitesimal transformations, it is useful to introduce the concept of a scaling
dimension of a model parameter Q: Under scaling transformation x0 = x/s it changes
as
Q0 = Qsdim[Q] (3.2)

which defines dim[Q]. The concept of scaling dimensions is particularly useful to classify
small perturbations to a particular RG fixed point: The sign of dim[Q] decides whether
Q grows or shrinks along the RG flow. Accordingly, one distinguishes the cases

dim[Q] > 0 relevant parameter

dim[Q] < 0 irrelevant parameter
dim[Q] = 0 marginal parameter

In a differential RG equation, the scaling dimension appears as prefactor of the linear

term:
dQ
= dim[Q]Q + . . . (3.3)
dl
with the dots representing higher-order terms: The left-hand side is a Taylor expansion
in powers of Q, with the constant term being absent as Q = 0 is assumed to be a fixed
point. For dim[Q] = 0 the higher-order terms determine the fate of Q under RG flow,
and one distinguishes marginally relevant, marginally irrelevant, or exactly marginal
perturbations (in the latter case, Q does not flow to all orders).
As a relevant perturbation to a fixed point grows under RG, it implies that the state
represented by the fixed point is not stable at low energies. Conversely, if a particular
fixed point represents a stable phase, then all symmetry-preserving perturbations should
be irrelevant. For example, chemical-potential disorder is a relevant perturbation to a
Fermi gas in one or two space dimensions, leading to Anderson localization, whereas it
is an irrelevant perturbation in d = 3.

3.3 Momentum-shell RG for the O(N ) φ4 model

We now demonstrate the working principles of RG using a concrete example. Consider
a classical N -component φ4 model in d space dimensions, described by the effective
2
In principle, the choice of an RG scheme for a given model is not unique. Different RG schemes may
lead to different intermediate-step results, but their answers for physical observables agree. In many
cases, a direct interpretation of the RG flow is useful and justified, but needs to be done with care as
the running couplings are in general not observable.
24 Chapter 3. Renormalization group

Hamiltonian
Z  
d 1 2 1 2 u 4
H= d x (∂x ϕ) + rϕ (x) + ϕ (x) (3.4)
2 2 4!
Z Λ d Z d d d
d k 1 2 2 2

u d k1 d k2 d k3
= d
k ϕ(k) + rϕ + ϕ(k1 )ϕ(k2 )ϕ(k3 )ϕ(−k1 − k2 − k3 )
0 (2π) 2 4! (2π)3d
(3.5)
where r is the tuning parameter of the phase transition (often dubbed “mass”), u the
order-parameter self-interaction, and Λ the physical ultraviolet momentum-space cutoff
related to the lattice constant a as Λ ∼ πa . The form of this theory is identical to
that of Ginzburg-Landau theory (2.10), and the fields have been scaled such that the
prefactor of the gradient term is unity. As we will see below, it pays to consider the φ4
model (3.4) for arbitrary values of the space dimension d, including non-integer ones. The
critical exponents will then be continuous functions of d, and the perturbative calculation
described below will evaluate exponents in d = 4 −  dimensions.
In particular in view of quantum generalizations to be described later, it is useful to
re-interpret H (3.4) – up to a factor
R β = 1/(kB T ) –3 as an action S which is related to
the partition function Z by Z = D[φ(r)] exp(−S).
The recipe for the so-called momentum-shell RG is to eliminate degree of freedom in
momentum space in a differential fashion. It can be summarized as follows.
1. Separate degrees of freedom
Λ

 ϕ<
 k< with (s > 1)
ϕ(k) = s . (3.6)
 ϕ> Λ
<k<Λ

s
2. Integrate out ϕ> (while preserving Z), i.e. renormalize the couplings r, u.
3. Rescale momenta, lengths, fields, couplings, etc. to restore the action with cutoff
Λ.

RG for Gaussian model: We first demonstrate these steps for the Gaussian model,
with quartic coupling u = 0. We combine the rescaling (3.1) with a rescaling of the fields
to
k 0 = sk ,
x0 = x/s ,
ϕ0 (k 0 ) = ϕ< (k)sa .
The RG steps are:
3
We note that the kinetic-energy piece is missing from H, as thermodynamic properties of classical
critical phenomena are determined by the potential energy only. Hence, there is no distinction between
Lagrangian and Hamiltonian at this point.
Chapter 3. Renormalization group 25

1. Separate degrees of freedom:

!
Z Z Z Λ/s Z Λ
Z= D[ϕ< ] D[ϕ> ] exp − dd k(k 2 + r)ϕ2< − dd k(k 2 + r)ϕ2> (3.7)
Λ/s

where prefactors of 2 and π have been omitted.

2. Integrate over fields with large momenta. The resulting renormalization is trivial:
For non-interacting fields, eliminating ϕ> does not change the action for ϕ< .

3. Rescale momenta, lengths, fields, couplings, etc. to restore the action – this fixes
the parameter a of the scaling transformation and results in a transformation rule
for r.
Z Λ/s Z Λ d 0
d k 0 0 2 −2a
d 2 2
(k 0 /s)2 + r


S< = d kϕ< (k) (k + r) = d
ϕ (k ) s
s
Z Λ
= dd k 0 ϕ0 (k 0 )2 (k 02 + r0 )

where a = − d+2
2
and therefore r0 = s2 r.4

Now we consider a differential RG step using s = edl , hence dr = 2dlr, and obtain
the RG equation for the coupling r:
dr
= 2r . (3.8)
dl
Fig. 3.3 shows two stable fixed points at r = +∞ and r = −∞ which correspond to the
disordered and ordered phases of Ginzburg-Landau theory, respectively. At r = 0 there
is an unstable fixed point, where the model is scale-invariant. This fixed point, dubbed
Gaussian fixed point, is a critical fixed point controlling the phase transition.

Figure 3.3: RG flow for the Gaussian model derived from Eq. (3.8), showing two stable
fixed points at r = ±∞ and one unstable fixed point at r = 0.

4
Our rescaling fixes the prefactor of the gradient term to unity; other rescaling schemes are possible
as well and lead to the same physical results.
26 Chapter 3. Renormalization group

RG for φ4 model: We now include the quartic coupling u of the ϕ4 model, treating
u perturbatively. As we will see below, this is justified close to and above four space
dimensions. We start with the contributions to first order in u:
Z
Z = D[ϕ> ]D[ϕ< ] exp −S0< − S0> − “uϕ4 ”

Z Z
= D[ϕ< ] exp (−S0 ) D[ϕ> ] exp (−S0> ) × (1 − “uϕ4 ”)
<

where “uϕ4 ” is a term linear in u given by ϕ> ϕ> ϕ< ϕ< . Upon performing the integral
over ϕ> , the contribution from the u term can be represented in the form of Feynman
diagrams, shown in figure 3.4. Keeping only the zero-momentum piece of the external

Figure 3.4: Top: Graphical representation of the quartic interaction. Bottom: Feynman
diagrams to linear order in u – both induce a renormalization of the coupling constant
r.

fields ϕ< , both diagrams contribute to the renormalization of r. Together with the
rescaling step, we arrive at

N + 2 Λ dd k
 Z 
0 2 1
r =s r+u d 2
. (3.9)
6 Λ/s (2π) k + r

At this point, a few remarks on the construction of Feynman diagrams for the per-
turbative corrections to the action are in order. (i) All internal lines must represent ϕ>
propagators, because these are the fields which are integrated over, whereas all external
lines must corresponds to ϕ< . (ii) All contributions with odd powers of ϕ> drop out, as
they yield Gaussian integrals over odd functions.
It turns out that a consistent determination of a non-trivial RG flow requires to go
to second order in u. The resulting diagrams are shown in Figs. 3.5, 3.6, and 3.7. The
diagram in Fig. 3.5 is of two-loop order; as we will see below, this contributes only to
second order in an expansion in  = 4 − d and will be neglected here. The diagram in
Fig. 3.6 generates a ϕ6 -term which can be shown to be strongly irrelevant near d = 4; it
Chapter 3. Renormalization group 27

Figure 3.5: Second-order Feynman diagram leading to a renormalization of r to order u2

(neglect this).

Figure 3.6: Second-order Feynman diagram inducing a ϕ6 term (neglect this).

can also be neglected. Finally, the diagram in figure 3.7 induces a renormalization of u,
being important for the RG flow. Together with the rescaling step, the transformation
for u reads  Z Λ d 
0 4−d 2N + 8 d k 1
u =s u−u d 2 2
. (3.10)
6 Λ/s (2π) (k + r)

It is useful to trade r and u for dimensionless variables t and g, defined by:

r = tΛ2
Λ4−d (2π)d
u=g
Sd
where Sd is the surface of a d-dimensional sphere.

RG equations: Evaluating the diagrams then yields the RG equations for the variables
t and g:

dt N +2 g
= 2t + + O(g 2 ) , (3.11)
dl 6 1+t
dg N + 8 g2
= (4 − d)g − 2
+ O(g 3 ). (3.12)
dl 6 (1 + t)

Figure 3.7: Second-order Feynman diagram leading to a renormalization of u (keep this).

28 Chapter 3. Renormalization group

where the last terms represent the influence of higher-order diagrams – recall that these
RG equations are based on an expansion in the quartic coupling and as such require g
to be small.

3.3.1 Fixed points

We now discuss the fixed points of the RG equations (3.11), concentrating on the vicinity
of the phase transition for small |t|. (As before, t → ±∞ correspond to the disordered
and ordered phases, respectively.) First, we observe that the Gaussian fixed point

t? = g ? = 0 (3.13)

remains a fixed point of the RG flow. It is useful to state the scaling dimensions of the
couplings with respect to this fixed point which can be read off from the linear term of
the corresponding RG equation:

dim[r] ≡ dim[t] = 2 , (3.14)

dim[u] ≡ dim[g] = 4 − d , (3.15)

i.e., the quartic coupling is irrelevant for d > 4, as will be discussed in more detail below.
We also note that higher-order terms are typically more irrelevant, e.g., the coefficient
of a φ6 term is characterized by

dim[u6 ] = 6 − 2d (3.16)

and hence irrelevant for d > 3.

In addition to the Gaussian fixed point there is another fixed point at finite t and g,
with g ? ∼ (4 − d)(1 + t)2 . Requiring that g should be small – only then the perturbative
treatment is valid – implies that (4−d) needs to be small (assuming |t|  1). Introducing

=4−d (3.17)

we find the second fixed point as

6
g? =  + O(2 ) , (3.18)
N +8
1N +2
t? = −  + O(2 ) . (3.19)
2N +8
This is the so-called Wilson-Fisher fixed point. Note that the (1 + t) terms in the
denominators in Eq. (3.11) have been approximated by unity, as they contribute to higher
orders in  only. It turns out that the expansion terms for the fixed-point values (and for
physical observables) can be systematically organized in a loop expansion: Contributions
at order n arise from Feynman diagrams with n closed loops.
For d > 4 the Gaussian fixed point controls the transition, as seen in figure 3.8, while
the Wilson-Fisher fixed point controls the transition for d < 4 (figure 3.9). The flow
Chapter 3. Renormalization group 29

Figure 3.8: RG flow of the φ4 model for d > 4.

Figure 3.9: RG flow of the φ4 model for d < 4.

30 Chapter 3. Renormalization group

diagrams feature a separatrix which separates the flow to the disordered fixed point from
the one to the ordered fixed point; the critical fixed point is located on this separatrix.5
It is useful to make contact between the RG flow diagram and experimental systems.
In an experiment, one typically varies a single parameter (most often temperature) to
access a phase transition. Each set of experimental parameters corresponds to one ini-
tial set of couplings of the φ4 model; hence, the experimental conditions define a line
(parameterized by temperature) in the RG flow space. This line usually does not hit the
critical fixed point, but crosses the separatrix as illustrated in figure 3.10. This crossing
corresponds to the physical phase transition point, T = Tc , where the RG flow is towards
the critical fixed point. At any other temperature, the flow is either towards t → +∞,
i.e., the disordered phase, or t → −∞, i.e., the ordered phase.

Figure 3.10: RG flow diagram with trajectory of typical experimental initial conditions
together with the resulting RG flow.

Perturbations to the Wilson-Fisher fixed point: It is useful to analyze the flow

in the vicinity of the critical fixed point in more detail. To this end, we re-write the RG
equations using the parametrization

t = t? + δt , (3.20)
r = r? + δr , (3.21)

5
In a n-dimensional space of coupling constants, the separatrix is a (n − 1) dimensional surface.
Chapter 3. Renormalization group 31

with the fixed-point values g ? = N6+8  + O(2 ) and t? = − 2(N

N +2
+8)
 + O(2 ). Keeping only
first-order terms in the deviations δt, δr we arrive at the linearized RG equations
   
dδt N +2 N +2 N +2
= δt 2 −  + δg 1+ (3.22)
dl N +8 6 N +8
dδg
= −δg . (3.23)
dl

These equations can be solved by writing them in a matrix form: diagonalizing the
matrix yields eigenvalues, representing the scaling dimensions of the perturbations about
the fixed point, and the corresponding eigenvectors, reflecting the flow directions. For
the Wilson-Fisher fixed point in spatial dimensions d < 4 there is one irrelevant direction
and one relevant direction.
In general, the scaling dimension of most relevant perturbation X to a critical fixed
point is identical to the inverse correlation length exponent,

1
dim[X] = , (3.24)
ν

because a generic deviation from the critical fixed point has a component along X, hence
X represents a tuning parameter of the phase transition with ξ ∼ X −ν . The behavior
under scaling transformations X 0 = Xsdim[X] and ξ 0 = ξ/s (as dim[ξ] = −1) then dictates
Eq. (3.24). This reasoning also implies that ν1 is a measure of how fast (in terms of RG
flow) a system is driven away from criticality.
For the Wilson-Fisher fixed point we have X ≡ δt and hence

1 (N + 2)
ν= + (3.25)
2 4(N + 8)

1
whereas the Gaussian fixed point simply has ν = 2
as in Ginzburg-Landau theory.
A few further remarks:

• A critical fixed point has a single relevant perturbation, i.e., one flow eigenvector
with positive scaling dimension.

• Fixed points with more than one relevant perturbation correspond to multi-critical
points: To reach such a point experimentally requires fine tuning of more than one
parameter.

• The RG demonstrated here for the classical d-dimensional φ4 model applies un-
changed to the quantum ϕ4 model (see Section 5.3) at T = 0 in (d − 1) space
dimensions.
32 Chapter 3. Renormalization group

3.3.2 Field renormalization and anomalous exponent

We now supplement the RG analysis with a determination of the anomalous exponent
η of the order-parameter correlations. This analysis also illustrates how to actually
calculate observables at criticality in renormalized perturbation theory: One performs
bare low-order perturbation theory in u, then replaces u by its fixed-point value, and
expands the result for small  to yield the desired form (e.g. a power law). This strategy
of a double expansion (in both u and ) is key in applications of perturbative RG.
The exponent η is defined by the behavior of the correlator at criticality:
1
hφ(k)φ(−k)i ∼ . (3.26)
k 2−η
Treating the φ4 perturbatively yields the form
1
hφ(k)φ(−k)i = (3.27)
k2 + r − Σ(k)

where r is the bare mass, and Σ(k) contains all contributions arising from the quartic
coupling u. The critical point is defined by r = Σ(k = 0), i.e., a vanishing renormalized
mass. Hence, a non-trivial momentum dependence at criticality arises from Σ(k)−Σ(k =
0). The lowest-order contribution to this arises from the diagram in Fig. 3.5 which,
evaluated for d = 4, yields

Σ(k) − Σ(k = 0) = c1 u2 k 2 ln Λ/k (3.28)

with c1 a constant. Expressing this using the coupling t and inserting its fixed-point
value gives
1
hφ(k)φ(−k)i = 2 . (3.29)
k − c2 g k 2 ln Λ/k
?2

This is not yet the desired power law, the reason being that higher-order contributions
will produce additional log-divergent terms, and only the combination of all of them yields
a power law. A small- approximation to the anomalous exponent can be obtained using
the relation k x ≈ 1 + x ln k, valid for small x:
 c2 g?2
1 1 1 Λ
hφ(k)φ(−k)i = 2 ≈ . (3.30)
k 1 − c2 g ?2 ln Λ/k k2 k

The last step, which is an expansion for small g ? , effectively re-sums a series of higher-
order contributions. Comparing to Eq. (3.26) we read off η = −c2 g ?2 . Re-instating the
constants we finally have
N +2 2
η= . (3.31)
2(N + 8)2
The first contribution to η being of two-loop order implies, combined with the small
prefactor, that η is numerically small for phase transitions in the φ4 model.
Chapter 3. Renormalization group 33

3.4 Phase transitions and critical dimensions

We now turn to a more general discussion of the order–disorder phase transition of φ4
models in arbitrary dimension d and for arbitrary number of order-parameter components
N . First, we establish the notion of upper and lower critical dimension.

Critical dimensions:

• The upper critical dimension d+ +

c is defined by the fact that for d ≥ dc mean-field
theory is asymptotically exact.

• The lower critical dimension d− −

c is defined by the fact that for d ≤ dc fluctuations
destroy the ordered phase at any temperature.

• Critical exponents typically take non-trivial d-dependent values for d− +

c < d < dc
whereas they take d-independent mean-field values for d > d+c .
6

• Classical magnets (described by φ4 models) with short-range interactions have

− −
d+
c = 4 and dc = 2 for N > 2 or dc = 1 for N = 1 (see below for d = 2, N = 2).

Let us discuss the physics of the upper critical dimension in more detail, based on
the RG flow for the φ4 model derived above.

• For d < d+ c = 4, the quartic coupling u is relevant, such that the phase transition
is controlled by the Wilson-Fisher fixed point. In calculations of observables, the
coupling u can be replaced by its fixed-point value u? and the cutoff Λ can be
sent to ∞. As a result, suitably defined observables (like dimensionless crossover
functions) will be fully universal, i.e., independent of the initial values of u and Λ.
Exponents will depend on the fixed-point value of u and hence vary with d. Naive
scaling is typically valid, and the critical exponents fulfill hyperscaling relations.

• For d > 4, u is irrelevant and the exponents will take mean-field values. Further,
integrals for observables are typically ultra-violet divergent, such that the cutoff Λ
has to be kept finite and crossover functions acquire non-universal contributions.
Naive scaling does necessarily not apply, and hyperscaling relations are violated.

• For d = 4, u is marginally irrelevant, resulting in logarithmic corrections to mean-

field behavior.

• Attention: The quartic coupling u is dangerously irrelevant for d ≥ 4. Although

u? = 0 at the Gaussian fixed point, a finite u is required in the calculation of certain
observables (like the magnetization in the ordered phase). Near-critical observables
will then depend on the initial value of u, although the fixed-point value is zero.
6
For models with long-range interactions, mean-field exponents may depend on d.
34 Chapter 3. Renormalization group

Figure 3.11: Overview of transitions of the O(N ) model in a plane spanned by the number
of spatial dimensions d and the number of order-parameter components N , showing the
applicability ranges of mean-field theory, (4 − ) expansion, and 1/N expansion.

The momentum-shell RG for the φ4 model, combined with the expansion in the
quartic coupling u, is controlled by the small parameter  = 4 − d. Hence, it gives
reliable results close to and above the upper-critical dimension. Further, it turns out
that it gives qualitatively correct results even for  = 1, i.e., for d = 3.7
There are various alternative ways to analytically calculate critical exponents, which
employ different limits. One option is a 1/N -expansion where N is the number of order-
parameter components: It turns out that the case N = ∞ can be solved exactly using
a certain mean-field theory, and a controlled expansion in 1/N is feasible. Another
option is an expansion about the lower critical dimension: For d = d− c +  the ordering
temperature Tc ∼ O(), such that Tc itself becomes a small parameter of the theory. For
magnets, this “2+” expansion, which is performed as an expansion around the ordered
phase, can be used to access critical exponents as well.
A summary of the utility of expansions for magnetic system is given in figure 3.11.
A special case is d = 2, N = 2 which is not correctly described by any of the expansions:
While long-range order is absent at any finite T , there exists quasi-long-range order
with power-law correlations below a transition temperature Tc . The transition between
this low-temperature and the disordered high-temperature phase is a Kosterlitz-Thouless
transition driven by the unbinding of vortex–antivortex pairs, i.e., topological defects.

7
Higher-order  expansions, together with suitable extrapolation schemes, can often be used to obtain
accurate quantitative estimates of exponents for  = 1.
Chapter 4

Theoretical models for quantum

phase transitions

With this chapter, we turn to phase transitions taking place at zero temperature, so-
called quantum phase transitions (QPT). We start by introducing a few basic lattice
models which such QPT. In general, the models are characterized by the competition
between two Hamiltonian terms, one favoring long-range order and one promoting local
quantum fluctuations which tend to destroy order.

4.1 Quantum Ising model

A quantum Ising model – often also called transverse-field Ising model – consists of
localized spin-1/2 objects placed on a regular lattice, with the Hamiltonian
X X
HI = −J σiz σjz − Jg σix (4.1)
hiji i

where σ α with α = x, y, z are Pauli matrices representing spins 1/2. The first term is
the conventional Ising interaction, and the second encodes an external magnetic field
in a direction perpendicular to the Ising axis, such that the model features a Z2 sym-
metry corresponding to σiz ↔ −σiz . Here J represent the overall energy scale, and the
dimensionless parameter g controls the behavior of the system at T = 0.
The limiting cases are easily discussed:
Q
Limit g →Q0: The ground state will be the perfectly aligned state |↑i = i |↑ii or
|↓i = i |↓ii .This state has a spontaneously broken Z2 symmetry. For small finite
g, flipped spins will admix into the ground state, but the Z2 symmetry remains
broken.
This ordered state is characterized by the order parameter h0| σiz |0i = ±N0 , rep-
resenting a uniform magnetization (per site), i.e., a ferromagnet.

35
36 Chapter 4. Theoretical models for quantum phase transitions
Q
Limit g → ∞: The ground state is |0i = i |→ii , i.e., the spins are polarized in field
direction. For large finite g, pairs of antialigned spins admix into the ground state.
This disordered state does not break any symmetries of the Hamiltonian.
Hence, a T = 0 phase transition must occur upon variation of g: lowering g from large
values leads to the onset of order at g = gc , see Fig. 4.1. As we will discuss later,
this quantum phase transition in d space dimensions is in the universality class of the
d + 1-dimensional classical Ising model.
For space dimensions d ≥ 2 the Ising order also exists at finite temperature, resulting
in a phase diagram as shown in Fig. 4.2.

Figure 4.1: Order parameter of the quantum Ising model as function of the tuning
parameter g.

Figure 4.2: Finite-temperature phase diagram of the quantum Ising model.

The quantum Ising model, with an experimentally accessible quantum phase transi-
tion, can be realized in magnetic insulators with ferromagnetic interactions and a strong
Ising anisotropy. Two examples are LiHoF4 [5], a three-dimensional Ising system (which,
however, has long-range dipolar instead of nearest-neighbor interactions) and CoNb2 O6
[6], a quasi-one-dimensional system.
Chapter 4. Theoretical models for quantum phase transitions 37

4.2 Quantum rotor model

A quantum rotor can be visualized as a particle moving on a N -dimensional unit sphere
(N ≥ 2), with orientation ~n and the constraint ~n2 = 1. The quantum dynamics of a
rotor is captured by a conjugate momentum p~, defined by

[nα , pβ ] = iδαβ (4.2)

in units where ~ = 1. It is convenient to define the rotor angular momentum by

Lαβ = nα pβ − nβ pα (4.3)

which has N (N − 1)/2 components. For the important case N = 3 we may introduce
the familiar vectorial notation Lα = (1/2)αβγ Lβγ with αβγ the antisymmetric tensor.
Then the commutation relations read:

[Lα , Lβ ] = iαβγ Lγ
[Lα , nβ ] = iαβγ nγ

For any N ≥ 2 the kinetic part of the Hamiltonian of a single rotor is

Jg ~ 2
HK = L (4.4)
2
with 1/(Jg) representing the rotor moment of inertia. For N = 2 this Hamiltonian has
2
the eigenvalues Jg l2 with l = 0, 1, 2, . . . , while for N = 3 its eigenvalues are Jg l(l+1)
2
with
l = 0, 1, 2, . . . .
For a lattice quantum rotor model we combine the kinetic energy with a (ferromag-
netic) interaction promoting long-range order, with the result
X Jg X ~ 2
HR = −J ~ni · ~nj + L (4.5)
2 i i
hiji

As before, the dimensionless parameter g controls the T = 0 behavior of the system, and
the limits are easily understood:

Limit g  1: The dominant interaction leads to an ordered state with O(N ) symmetry
broken. Long-range order is described by | h0| ~ni |0i | = N0 and lim|~ri −~rj |→∞ h0| ~ni ·
~nj |0i = N02 .

Limit g  1: The dominant kinetic energy leads to a fluctuating, i.e., disordered, phase,
with h0| ~ni · ~nj |0i ∼ e−|~ri −~rj |/ξ .

As a result, a quantum phase transition occurs at an intermediate value of g; this is in

the universality class of the (d + 1)-dimensional classical O(N ) model. We note that
38 Chapter 4. Theoretical models for quantum phase transitions

the T = 0 ordered phase exists only in spatial dimension d ≥ 2, as angular fluctuations

destroy order otherwise (Mermin-Wagner theorem).1
While elementary quantum rotors do not exist in nature, quantum rotors often occur
as effective models for low-energy degrees of freedom. For instance, N = 2 quantum
rotors may capture phase degrees of freedom of superconducting islands, and the lattice
model then represents an array of Josephson-coupled islands. N = 3 quantum rotors
appear in the description of spin models where they can represent degrees of freedom of
spin pairs. An explicit example is given by the coupled-dimer model in the next section.

4.3 Coupled-dimer model

The final example is a class of Heisenberg models of spins 21 , with the special property
that each crystallographic unit cell contains two spins (a “dimer”). We require antifer-
romagnetic interactions, such that the Hamiltonian reads
X
H= ~i · S
Jij S ~j , Jij > 0 (4.6)
hiji

The interactions obey (

J intradimer bonds
Jij = . (4.7)
λJ interdimer bonds
As will become clear below, the dimensionless parameter λ controls the quantum phase
transition in the system. A sample lattice is shown figure 4.3, where spins live on a
square lattice, but the bond strengths are modulated according to Eq. (4.7).

Figure 4.3: Columnar coupled-dimer

model on a square lattice: Each unit
cell contains two spins connected by
a coupling of strength J, while the
remaining square-lattice bonds have
strength λJ.

The limit λ = 0 corresponds to disconnected spin pairs, each of them having a singlet
S = 0 ground state and a triplet S = 1 excited state, separated by an excitation energy
J. The full lattice model has two distinct phases, which can be easily discussed:
1
A special case is d = 1, N = 2, where the ordered phase is replaced by a phase with quasi-long-range
order (i.e., a power-law behavior of the correlation function) which is bounded by a Kosterlitz-Thouless
transition.
Chapter 4. Theoretical models for quantum phase transitions 39

Limit λ  1: This implies weakly coupled dimers, leading to a disordered (or quantum
paramagnetic) phase with no broken symmetries and exponentially decaying spin
correlations.

Limit λ ∼ 1: Here the dimers are strongly coupled, and long-range antiferromagnetic
order with broken SU(2) symmetry emerges. (For the lattice shown in Fig. 4.3
λ = 1 represents a square lattice which is known to display long-range order.)2

Again, a quantum phase transition must occur at an intermediate value of λ. As the order
parameter is of O(3) type, the QPT is in the universality class of the (d + 1)-dimensional
classical O(3) (or Heisenberg) model. The phase diagram is the square-lattice coupled-
dimer model of Fig. 4.3 is depicted in Fig. 4.4.

Figure 4.4: Phase diagram for columnar coupled-dimer model of Fig. 4.3.

Coupled-dimer magnets are frequently realized in Mott-insulating materials. Exam-

ples are TlCuCl3 [7], a three-dimensional system which can be driven through the QPT
by applying pressure, and BaCuSi2 O6 [8] which realizes weakly coupled layers of dimers
(i.e. a quasi-two-dimensional system).

2
The system in Fig. 4.3 becomes disordered again for λ  1, as this limit corresponds to decoupled
spin ladders.
Chapter 5

Quantum phase transitions: Primer

Having illustrated the possibility of phase transitions in the ground state of many-body
system, we now turn to a phenomenological description of systems in the vicinity of such
phase transitions, in particular at finite temperature. This discussion will also highlight
the interplay of quantum and classical fluctuations near QPT.

5.1 Classical and quantum fluctuations

The partition function in a (canonical) Gibbs ensemble,
h i
Z = Tr e−β Ĥ (5.1)

factorizes for classical systems into kinetic and potential pieces:

H = T + V,
Z = Tr e−βH = Tr e−βT e−βV .
   

P 2
The kinetic energy, T = i pi /(2m), generically contributes an analytic factor to Z
because e−βT is a product of Gaussian integrals. Hence any non-analyticity in Z, corre-
sponding to a phase transition, can only arise from the potential term e−βV . For classical
systems we conclude:

• Statics and dynamics decouple.

• The exponents listed in Sec. 2.3 (with the exception of z) are determined by the
static piece of Z (i.e. the potential V ) and constitute a “static” universality class.
(Recall: Ginzburg-Landau theory did not involve dynamics.)

• The dynamical exponent z is independent of the other exponents. A given “static”

universality class can come with different dynamics.

40
Chapter 5. Quantum phase transitions: Primer 41

In contrast, for quantum systems the situation is different: In general, [T̂ , V̂ ]− 6= 0.

As a result, kinetic and potential part in the partition function do not commute:

e−βH 6= e−βT e−βV

Therefore statics and dynamics are always coupled, and the dynamical exponent z is
integral part of the set of exponents of a given universality class.

Quantum vs. classical transitions: Near any continuous phase transition the cor-
relation length and the correlation time diverge

ξ→∞ τc → ∞ (5.2)

The divergence of τc implies the existence of an energy scale ~ωc , representing the typical
energy for order-parameter fluctuations, which goes to zero at criticality.
This allows for a sharp distinction between classical and quantum phase transitions:
A transition is classical if the order-parameter fluctuations follow classical statistical
mechanics, i.e., if
kB Tc  ~ωc . (5.3)
We conclude:
• All phase transitions with Tc > 0 are (asymptotically) classical, in the sense that
their long-distance fluctuations are in the classical limit. Note that this is not a
statement about the physics underlying the ordered state: For instance, the critical
behavior of a finite-T transition into a superconducting state is classical, although
the physics leading to superconductivity is “quantum”.

• for Tc = 0, quantum statistics is inevitably needed to describe the order-parameter

fluctuations, hence all zero-temperature transitions are quantum in nature.
In a phase diagram spanned by temperature and a non-thermal control parameter, a
quantum phase transition is typically the end point of a line of finite-temperature transi-
tions, see figure 5.1(b) below. The criterion (5.3) is fulfilled in a region near the classical
transition whose width goes to zero as Tc → 0.

5.2 Phenomenology: Phase diagrams and crossovers

Quantum phase transitions are driven by varying a non-thermal control parameter r at
T = 0. Experimentally relevant is, of course, the finite-temperature behavior, such that
a minimal phase diagram involves the parameters r and T . Schematic phase diagrams of
this type, with a continuous quantum phase transition at r = rc , are shown in Fig. 5.1,
and we will quickly discuss the main aspects.
The zero-temperature axis features two stable ground states, separated by a quantum
critical point at rc . The low-temperature behavior can be understood by considering the
42 Chapter 5. Quantum phase transitions: Primer

Figure 5.1: Schematic phase diagrams in the vicinity of a quantum critical point (QCP).
The horizontal axis represents the control parameter r used to tune the system through
the quantum phase transition, the vertical axis is the temperature T . a) Order is only
present at zero temperature. The dashed lines indicate the boundaries of the quantum
critical region where the leading critical singularities can be observed; these crossover
lines are given by kB T ∼ |r − rc |νz . b) Order can also exist at finite temperature. The
solid line marks the finite-temperature boundary between the ordered and disordered
phases. Close to this line, the critical behavior is classical.

elementary excitations on top of the ground state. In the disordered (i.e. non-symmetry-
broken) phase, these excitations are typically well-defined quasiparticles with a finite
energy gap ∆. The corresponding low-temperature regime will keep the characteristics of
the ground state, with an exponentially small density of thermally excited quasiparticles
– this is the regime dubbed quantum disordered. A similar discussion applies to the
ordered side of the QPT, with two complications: (i) Thermal fluctuations may destroy
order already at infinitesimal temperature: This applies to Ising systems in d = 1 and
Heisenberg systems in d = 2 and leads to a thermally disordered regime at any
finite as shown in Fig. 5.1(a). (ii) Even if order survives at finite T , Fig. 5.1(b), there
will be no excitation gap in cases with spontaneous breaking of a continuous symmetry.
Nevertheless, the density of thermally excited quasiparticles will be small.
This is to be contrasted with the finite-temperature regime, labelled quantum crit-
ical in Fig. 5.1 and located above the quantum critical point. This regime is bounded by
crossover lines kB T ∼ |r − rc |νz (equivalent to kB T ∼ ∆ above a ground state with gap
∆), indicating that pictures of dilute quasiparticle excitations do not apply. Instead, one
might think about thermal excitations of the quantum critical ground state. As we will
discuss in detail later, the quantum critical state typically does not feature conventional
quasiparticles, but a quantum critical continuum of excitations. Thermally exciting this
critical continuum leads to unconventional thermodynamic and transport properties in
the quantum critical regime, often characterized by fractional temperature exponents.
Technically, the quantum critical regime is the universal high-temperature regime of the
Chapter 5. Quantum phase transitions: Primer 43

field theory underlying the quantum critical point.

In Fig. 5.1(b), where order exists at finite T , there is a finite-temperature transition.
According to the discussion above, the critical behavior is classical if kB Tc  ~ωc , which
applies to a narrow regime close to the finite-T transition line.
Finally, the behavior far away from criticality, in particular at high temperatures,
becomes non-universal. This is the case once the order-parameter correlation length ξ
is no longer large – this applies for temperatures kB T & J where J is a representative
microscopic energy scale. Numerical simulations have shown that quantum critical be-
havior in insulating magnets can be observed up to temperatures of J/3 where J is the
exchange constant [9] – in some systems this may even include room temperature.

5.3 Quantum φ4 theory

As we have seen in the previous chapters, the φ4 (or Ginzburg-Landau) theory is concep-
tually important in the discussion of critical phenomena. We now construct a quantum
version of the classical φ4 theory, the latter given by
Z  
d 1 2 r 2 u 4
H= d x (∂i ϕ) + ϕ + ϕ (5.4)
2 2 4!

where i runs over x, y, z. In the quantum case, the expression for the partition function
needs to be formulated as an imaginary-time path integral,
Z
Z = D[ϕ(r, τ )]e−S . (5.5)

The key new ingredient is a derivative along the (imaginary) time (τ ) direction, such
that the quantum action is given by
Z Z β  2 
d c 2 1 2 r 2 u 4
S= d x dτ (∂i ϕ) + (∂τ ϕ) + ϕ + ϕ (5.6)
0 2 2 2 4!

and c plays the role of a velocity. Hence, going from classical to quantum essentially
amounts to the replacement
Rβ
e−βH → e− 0 dτ H
= e−S . (5.7)

The form of the time derivative in (5.6) follows – as usual in Ginzburg-Landau theory –
from symmetry considerations: In the presence of time-reversal invariance, the quadratic
term is the lowest one allowed by symmetry (and higher ones are ignored).
We see that, at T = 0 where the imaginary-time axis is infinitely long, the quantum
problem in d dimensions is equivalent to the classical problem in d + z dimensions, with
the dynamic exponent z = 1: Imaginary time enters as z additional space dimensions
(compare τc ∼ ξ z ). In general, z can take any value (including fractional ones), but in
44 Chapter 5. Quantum phase transitions: Primer

the simple models we have introduced so far (like quantum Ising, quantum rotor etc.) z
is locked to unity. This equivalence between quantum problems and classical problems in
higher dimensions is known as quantum-to-classical correspondence and will be discussed
in more detail later. While this correspondence holds for a number of important models
of quantum phase transitions, we note that it breaks down in numerous models of current
interest.
The form of the quantum action (5.6) can also be used to discuss the crossover from
quantum to classical behavior upon approaching a finite-temperature phase transition.
The diverging correlation time τc implies that fluctuations along the imaginary-time axis
get slow. If β < ∞, then these fluctuations are effectively frozen once τc & β – this is
equivalent to the condition (5.3). Then, all time dependencies in (5.6) drop out, and the
integral over imaginary time can be replaced by a factor β, corresponding to a classical-
limit evaluation. For later reference, we note that this reasoning does not apply to the
quantum critical regime: In this finite-temperature regime, τc is of order β, such that
thermal and quantum fluctuations are equally important.

5.4 Quantum scaling hypothesis

Recall the classical hypothesis
fs (t, h) = b−d fs (tbyt , hbyh ) (5.8)
which expresses a homogeneity law for the free-energy density or, alternatively, states
how the parameters t and h need to be changed upon a rescaling of lengths in order to
restore the physics. It is also useful to remember that this simple form of scaling only
applies below the upper critical dimension.
The key difference of the quantum case is that there are now two “directions” which
can be used to tune the system away from the zero-temperature critical point, namely
the non-thermal control parameter and temperature itself. Hence, fs acquires T as a
new parameter, and we will denote by t the dimensionless distance to criticality on the
control-parameter axis, e.g., t = r − rc or t = (p − pc )/pc if the transition is tuned by
pressure p. The quantum scaling hypothesis then takes the form
fs (t, h, T ) = b−(d+z) fs (tbyt , hbyh , T byT ) (5.9)
where the exponent (d + z) accounts for the fact that the rescaling factor b acts on both
length and time directions, and we work in units where both temperature and field are
dimensionless (i.e. measured in suitable microscopic units). As in the classical case, we
conclude yt = ν1 based on ξ ∝ |t|−ν . The definition of the dynamical exponent implies
that T scales as
1
∼ (length)z −→ yT = z . (5.10)
T
Choosing b = t−ν yields the scaling relation
 
ν(d+z) h T
fs (t, h, T ) = t fs 1, νy , νz (5.11)
t h t
Chapter 5. Quantum phase transitions: Primer 45

Figure 5.2: Crossover lines at

T ∼ |r − rc |νz

The finite-temperature behavior is controlled by the ratio T /|t|νz . We distinguish the

regimes (
T  1 quantum critical regime
. (5.12)
|t|νz  1 stable-phase regimes
Accordingly, the relation T ∼ |t|νz defines crossover lines which divide the universal part
of the phase diagram in three regimes, Fig. 5.2.
Specializing to the case without external field, h = 0, we have
 
ν(d+z) T
fs (t, T ) = t f1 νz
t
 
T
= T (d+z)/z f2 νz
t

with f1 (x) = f2 (x) · xν/z . In the quantum critical regime, t → 0, we obtain

fs (t = 0, T ) = T (d+z)/z · const. (5.13)

This allows to deduce the temperature dependence of (the critical part of) the specific
heat as
∂ 2 fs
Ccrit = V T ∝ T d/z (5.14)
∂T 2
exclusively by scaling arguments.1
Interestingly, the power-law behavior realized by physical observables upon approach-
ing the QCP either by tuning t or by tuning T can often be related. Consider the entropy
S which vanishes at the QCP, but its derivatives are singular. The specific heat C will
show power-law behavior, as does the observable B = ∂S/∂t. At a pressure-tuned phase
transition, t = (p − pc )/pc , B measures the thermal expansion,
 
1 ∂V  1 ∂S 
α= =− . (5.15)
V ∂T p V ∂p T

B/C defines the Grüneisen parameter Γ,

α 1 (∂S/∂p)T
Γ= =− (5.16)
Cp Vm T (∂S/∂T )p
1
C ∝ T d/z requires hyperscaling and does not necessarily apply above the upper critical dimension.
46 Chapter 5. Quantum phase transitions: Primer

where Vm = V /N the molar volume. Taking the ratio of the singular parts of B and C
one observes that the scaling dimensions of T and S cancel, and therefore B/C scales as
the inverse of the tuning parameter t. Thus, one obtains a universal divergence2 in the
low-T limit [4]

Bcr
Γcr (T = 0, t) = = Gt |t|−1 , (5.17)
Ccr
Γcr (T, t = 0) = GT T −1/(νz) . (5.18)

Given that Γ does not diverge at a finite-T phase transition, a divergence of Γ is a unique
signature of a continuous QPT.

5.5 Quantum-to-classical mapping

This section gives deeper insights into the correspondence between quantum phase tran-
sitions in d dimensions and classical phase transitions in (d + 1) dimensions. We start
by establishing the equivalence of the classical Ising spin chain (d = 1) and the quantum
problem of a single spin in a field (d = 0), valid in the scaling limit of large correlation
length. Afterwards, we list more mappings between classical and quantum problems and
finally discuss general aspects of the quantum-to-classical correspondence.

5.5.1 Classical Ising chain

We consider the Hamiltonian function of a finite Ising chain,
M
X X
H = −K σiz σi+1
z
−h σiz , (5.19)
i=1 i

and the parameters K and h are measured in units of the temperature T (formally
T = 1).

e−H is
P
Transfer matrix method: The partition function Z = {σi }

M
XY
Z= z
T1 (σiz , σi+1 )T2 (σiz ) (5.20)
{σi } i=1

with

T1 (σ1z , σ2z ) = exp (Kσ1z σ2z )

T2 (σ z ) = exp (hσ z )
2
For T → 0 even the prefactor Gt is universal and given by a combination of critical exponents.
Chapter 5. Quantum phase transitions: Primer 47

which is
eK e−K
   
eh 0
T1 = , T2 = (5.21)
e−K eK 0 e−h
where in zero field h = 0, T2 = 1. Assuming periodic boundary conditions we get

Z = Tr [T1 T2 T1 T2 . . . ]
= Tr (T1 T2 )M
 
h 1 1 i
= Tr (T22 T1 T22 )M
= M M
1 + 2

with the eigenvalues

12
12 = eK cosh(h) ± e2K sinh2 (h) + e−2K . (5.22)

Correlation function: Restricting ourselves to zero field we have

1 X −H z z
hσiz σjz i = e σi σj
Z
{σ}
1  i z j−1 z M −j 
= T1 σ T1 σ T1
Z
..
.
−j+i j−i −j+i j−i
M
1 2 + M2 1
= M M
1 + 2

In the large system limit M → ∞ the correlation function becomes

hσiz σjz i = (tanh K)j−i (5.23)

To formulate the scaling limit, it is useful to introducing coordinates τ = ja where a is

the lattice spacing, and to identify σjz → σ z (τ ). The correlation function takes the form

hσ z (τ )σ z (0)i = e−|τ |/ξ (5.24)

1
where ξ is the correlation length, with 1/ξ = a
ln coth K. At low temperatures K  1
we have ξ/a ∼ 21 e2K  1.

5.5.2 Scaling limit and universality

Universality is made transparent if physical observables are expressed in terms of param-
eters characterizing the low-energy/long-distance behavior of the system. The present
system features the following length scales:
48 Chapter 5. Quantum phase transitions: Primer

• Large: correlation length ξ, observation scale τ , system size: Lτ = M a.

• Small: lattice constant a.

The scaling limit is the limit where the ratio of “large” to “small” scales is sent to infinity
which is suitable to eliminate microscopic details. Here, we take a → 0 while keeping
ξ, τ, Lτ finite. The free-energy density is

ln Z
F =− . (5.25)
Ma
We need the transfer-matrix eigenvalues in the scaling limit
  12  
2ξ a 2 2 21
1,2 ∼ 1 ± (1 + 4h̃ ξ ) (5.26)
a 2ξ

where K is expressed using the correlation length ξ, and h̃ = h/a. Hence, the free-energy
density takes the form
s !!
1 1
F = E0 − ln 2 cosh Lτ + h̃2 (5.27)
Lτ (4ξ)2

where the ground state energy at zero temperature is E0 = − Ka . For h̃ = 0 the correlation
function is
e−|τ |/ξ + e−(Lτ −|τ |)/ξ
hσ z (τ )σ z (0)i = (5.28)
1 + e−Lτ /ξ
and we recover
hσ z (τ )σ z (0)i = e−|τ |/ξ (5.29)
for Lτ  ξ.

Universality: Expressions for the observables (expressed in terms ξ, Lτ , h̃) are identi-
cal for all models with the same dimensionality and the same symmetry as HI , provided
that we take the scaling limit a → 0. This can be cast in universal scaling functions. For
instance, the for the free-energy density can be written as
 
1 Lτ
F = E0 + φF , h̃Lτ (5.30)
Lτ ξ
p
with the scaling function φF (x, y) = − ln (2 cosh ( (x/2)2 + y 2 )). Similarly the correla-
tion function can be written as
 
z z τ Lτ
hσ (τ )σ (0)i = φσ , , h̃Lτ . (5.31)
Lτ ξ
Chapter 5. Quantum phase transitions: Primer 49

5.5.3 Mapping to a quantum spin

We now rewrite the partition in a manner suitable to derive a classical-to-quantum
correspondence. First, we analyze the transfer matrices T1 and T2 in the scaling limit.
T1 can be written as

T1 = eK (1 + e−2K σ x )
a
≈ eK (1 + σ x )
2ξ
1 x)
≈ ea(−E0 + 2ξ σ

at low temperatures where e−2K ≈ a/2ξ. Also,

z
T2 = eah̃σ . (5.32)

The product T1 T2 can be simplified in the limit a → 0 where

eaÔ1 eaÔ2 = ea(Ô1 +Ô2 ) (1 + O(a2 )) (5.33)

such that
T1 T2 = e−aĤQ (5.34)
with
∆ x
ĤQ = E0 − σ − h̃σ z (5.35)
2
and ∆ = 1ξ . Then the partition function is

Z = Tr (T1 T2 )M = Tr e−HQ /T
   
(5.36)

with T = L1τ . This Z apparently represents the partition function of a quantum spin 12
in two perpendicular fields h̃, ∆2 at a quantum temperature T given by T = 1/Lτ . The
free energy r !
1 ∆2
F = E0 − T ln 2 cosh + h̃2 (5.37)
T 4
q
2
is that of a quantum spin in an external field of size ∆4 + h̃2 .
Hence, a quantum system at temperature T is equivalent to a classical system of
finite length Lτ = 1/T . To make this intuitive, let us consider a small ∆: this implies
a small flipping rate for the quantum spin (i.e., a large autocorrelation time) which
corresponds to a large correlation length in the classical system. The correspondence
between quantum (imaginary) time and classical length is summarized in table 5.1. It is
important to realize that in classical systems temperature can always be absorbed by a
rescaling of Hamiltonian parameters. In contrast, in a quantum system the temperature
is an independent parameter.
50 Chapter 5. Quantum phase transitions: Primer

Table 5.1: Correspondence between classical quantum

classical spin chain and quantum system size Lτ inverse temperature 1/T
spin. correlation length ξ inverse excitation gap ∆

5.5.4 Mapping: XY-chain ↔ O(2) quantum rotor

The classical XY-chain can be written by the Hamiltonian
M
X M
X
Hcl = −K ~ni · ~ni+1 − ~h · ~ni (5.38)
i=1 i=1

with ~n2i = 1 and ~ni = (nix , niy ). Using methodology similar to the one above, one can
show that this model is equivalent to an O(2) quantum rotor, with the Hamiltonian
∂2
HQ = −∆ − h̃ cos θ . (5.39)
∂θ2

5.5.5 Mapping: Heisenberg chain ↔ O(3) quantum rotor

The classical Heisenberg-chain Hamiltonian is similar to the classical XY-chain Hamil-
tonian
XM M
X
Hcl = −K ~ni · ~ni+1 − ~h · ~ni (5.40)
i=1 i=1

with ~n2i
= 1, but now with three-component spins ~ni = (nix , niy , niz ). Again, one can
shown that this maps onto an O(3) quantum rotor, described by
∆ ˆ~2
HQ = − L − h̃ · ~n . (5.41)
2
It is important to note that the mapping is to a rotor instead of a spin model: Instead, a
quantum spin is characterized by Berry-phase dynamics which has no classical analogue.

5.5.6 Quantum-to-classical correspondence: Rules and excep-

tions
The previous examples illustrate that certain classes of quantum phase transition in d
dimensions correspond to classical transition in D = d + z dimensions. As emphasized
before, this correspondence applies to T = 0 in the quantum system. It is useful to
deduce both thermodynamic properties and correlation functions, and we will employ
this in the next chapter. However, it has to be kept in mind that the correspondence is at
work on the imaginary time axis of the quantum system, and approximate information
on the imaginary axis cannot be easily transformed into approximate information on the
real axis, because analytic continuation is an ill-conditioned problem.
More importantly, there are many classes of quantum phase transitions which do not
obey any quantum-to-classical correspondence. This applies in particular to
Chapter 5. Quantum phase transitions: Primer 51

• systems with quenched disorder, as disorder is frozen in (imaginary) time;

• systems with quantum or Berry-phase dynamics – such terms are purely imaginary
and have no classical analogue. A simple example is the linear-time-derivative
describing the dynamics of canonical quantum particles: A single quantum particle
has the action
S = φ∂τ φ − φα φ (5.42)
where φ∂τ φ cannot be interpreted classically, see Sec. 7.3.1 for a discussion. An-
other example are Berry-phase terms describing the dynamics of quantum spins
(see the books of Sachdev [1] or Fradkin [3] for spin coherent states).

• systems where a symmetry-breaking order parameter is coupled to low-energy

particle-hole pairs of a metal, i.e., quantum phase transitions in metallic or semimetal-
lic systems [10];

• various types of topological phase transitions.

Chapter 6

Magnetic quantum phase transitions

In this chapter we describe in more detail the physics of quantum phase transitions in
magnets. We concentrate on insulating magnets, i.e., Mott insulators of local moments;
quantum phase transitions in metallic magnets are more complicated due to the interplay
between order-parameter fluctuations and low-energy particle-hole pairs [10].

6.1 Order parameters and response functions

We start with a few remarks on order parameters. The definition of a local order param-
~ i , t) whose thermodynamic average is parameterized
eter ϕ involves local operator Ô(R
as
~ R
~ i )i = Re(eiQ· ~i ~ i ))
hÔ(R ϕ(R (6.1)
where ϕ(R ~ i ) is a slowly varying function of R ~ i , which can be a scalar, a vector, tensor,
~ R
iQ· ~i
etc. The factor e captures possible fast oscillations in the local observable Ô ,
for instance he up–down oscillation of the magnetization in an antiferromagnet (i.e.
~ is called the ordering wavevector, and eiQ·~ R
~i
staggered magnetization). Q must be chosen
~
such that ϕ(Ri ) varies slowly on the lattice scale.
Examples for ordered states are

• Charge density wave (CDW):

~ ~
~ = ρ0 + Re(eiQ·Ri ϕC (R))
~
hρ̂(R)i (6.2)

where ϕ has N = 1 components.

• Spin density wave (SDW):

~ R
~ = Re(eiQ· ~i ~
hŜα (R)i ϕSα (R)) (6.3)

where ϕ has N = 3 components.

52
Chapter 6. Magnetic quantum phase transitions 53

~ = 0 then the order parameter is real, otherwise ϕ can be complex. Exceptions

If Q
~ R
~ = (π, π, π), where eiQ· ~i ~ i . In general,
are, e.g., on the cubic lattice with Q is real for all R
the phase of a complex ϕ takes discrete values if Q ~ is commensurate with the lattice, i.e.
the ratio of Q~ and the lattice constant is a rational number.
For spin density waves, different modulation types are possible due to the vector
character of the order. One class, with all hSi ~ pointing along the same axis, is dubbed
collinear and shown in figure 6.1. A proper parametrization of the corresponding order

Figure 6.1: Collinear arrangement of spins

parameter is ϕSα = eiθ nα with nα real.1 A second class are spiral states as shown in
figure 6.2. Here the order parameter is ϕSα = n1α + in2α with n1,2α real and ~n1 · ~n2 = 0.

Figure 6.2: Spiral (non-collinear) arrangement of spins

Experiments typically observe correlation functions of the order parameter; for mag-
nets this can be done for instance using magnetic neutron scattering. We now define the
most important correlation and response functions to be discussed below.
We start with a real-time correlation function
C(~r, t; ~r0 , t0 ) = hϕ(~r, t)ϕ(~r 0 , t0 )i (6.4)
 −βH 
where h. . .i = (1/Z)Tr e . . . is the thermodynamic average. Its Fourier transform
yields the dynamic structure factor
Z Z ∞
~ ~
S(k, ω) = d r d
dtC(~r, t; 0, 0)e−ik·~r+iωt (6.5)
−∞

where translation invariance in space and time has been assumed. In addition, it is useful
to consider an imaginary-time correlation function
C(~r, τ ; ~r0 , τ 0 ) = hT̂τ ϕ(~r, τ )ϕ(~r 0 , τ 0 )i (6.6)
where T̂τ is the time-ordering operator on the imaginary-time axis. This can be used to
define a dynamic susceptibility
Z Z 1/T
~ ~
d
χ(k, iωn ) = d r dτ C(~r, τ ; 0, 0)e−ik·~r+iωn τ (6.7)
0
1
θ denotes sliding degree of freedom for a density wave.
54 Chapter 6. Magnetic quantum phase transitions

where ωn = 2πnT are called (bosonic) Matsubara frequencies, and we recall that the
imaginary-time correlator is periodic in time with period 1/T . The fluctuation-dissipation
theorem, which can be derived via the spectral representation of the correlation func-
tions, states that
2
S(~k, ω) = Imχ(~k, ω) (6.8)
1 − e−ω/T
where Imχ(~k, ω) has been obtained from the imaginary-frequency dynamic susceptibility
by analytic continuation.2 Note that Imχ(~k, ω) is also called the spectral density.

6.2 Properties of the quantum φ4 model

We now discuss a few observable properties of systems described by the quantum φ4
model: Z Z β  2 
d c 2 1 2 r 2 u 4
S= d x dτ (∂i ϕ) + (∂τ ϕ) + ϕ + ϕ (6.9)
0 2 2 2 4!
Zero-temperature properties can be deduced from the classical results, discussed
in Chapters 2 and 3, via the quantum-to-classical (here: classical-to-quantum) corre-
spondence. The present model has a dynamical exponent of z = 1, hence the map-
ping is from (d + 1) space dimensions to d space and one (imaginary) time dimension:
|~k| ↔ (c2~k 2 + (iω)2 )1/2 where a velocity factor has been added. This implies that at
criticality the dynamical two-point correlation function is given by
1
χ(k, ω) ∝ . (6.10)
(c2 k 2 − (ω + iδ)2 )1−η/2
This is the advertised critical continuum of excitations, Fig. 6.3: Above a threshold given
by ω = ck this function has branch cuts instead of poles, hence quasiparticles are absent
for η 6= 0.3

Figure 6.3: Dynamic structure fac-

tor S(k, ω) ∝ Imχ(k, ω) at the crit-
ical point, displaying a critical con-
tinuum of excitations.

Technically, χ(~k, ω) is a retarded Green’s function which requires an analytic continuation iωn ↔
2

ω + iδ|δ→0 .
3
The quantum critical response (6.10) displays quasiparticle poles for η = 0. In cases where η is
non-zero but numerically small the critical continuum is difficult to distinguish experimentally from
conventional quasiparticle response.
Chapter 6. Magnetic quantum phase transitions 55

Figure 6.4: Dynamic structure

factor in the disordered phase.

Figure 6.5: Evolution of the

quasiparticle weight and the
energy gap in the disordered
phase as well as the order pa-
rameter in the ordered phase
as function of the control
paramter r of a quantum phase
transition.

Away from criticality in the disordered phase we have

1
χ(k, ω) = (6.11)
c2 k 2 + r − (ω + iδ)2 ) − Σ(k, ω)

where Σ(k, ω) contains the contributions from the quartic coupling. Away from critical-
ity, Σ is non-singular, and χ(k, ω) will display quasiparticle poles:

A
Imχ(k, ω) = (δ(ω − k ) − δ(ω + k )) (6.12)
2k

where 2k = c2 k 2 + r − Σ refers to the quasiparticle dispersion. The role of Σ is to modify

both dispersion and the pole weight A (but not to remove the pole). The energy position
of the pole at k = 0 defines the energy gap, ∆2 = r −Σ(k = 0, ω = ∆), and the dispersion
near k = 0 can be parameterized as k = ∆ + c2 k 2 /(2∆). At higher energies, scattering
processes induce n-particle continua in Imχ(k, ω) with n = 3, 5, 7, . . ., see Fig. 6.4.
56 Chapter 6. Magnetic quantum phase transitions

Near criticality, one may cast the T = 0 susceptibility into the scaling form
 
1 ck ω
χ(k, ω) = 2−η fχ , (6.13)
∆ ∆ ∆
which shows that the weight of the single-particle pole scales as A ∝ ∆η .
On the magnetically ordered side, the structure factor displays a Bragg peak
S(k, ω) = N02 (2π)d+1 δ(ω)δ(k) + . . . (6.14)
where the dots represent finite-frequency contributions. For N > 1 the transverse suscep-
tibility displays poles from gapless Goldstone modes, while the longitudinal susceptibility
is generically gapped – it shows a damped by well-defined mode near criticality which is
often dubbed amplitude or Higgs mode.
Fig. 6.5 illustrates some of the results collected above, for a discussion of finite-
temperature results we refer the reader to the literature [1].

6.3 Quantum Ising chain

After having discussed general properties of the continuum φ4 model, we now supplement
this by similar considerations for the quantum Ising chain, i.e., the model described in
Sec. 4.1 in d = 1. This is a microscopic model which can be essentially solved exactly, and
allows to obtain concrete results which are consistent with the general phenomenology
discussed above. We consider the Hamiltonian
X
HI = −J (gσ x + σiz σi+1
z
) (6.15)
i

which we will first examine at both strong coupling, g  1, and weak coupling, g  1.

6.3.1 Strong-coupling limit g  1

When
Q
g = ∞: the ground state is |Ψ0 i = i |→ii , where |→i denotes an eigenstate of σx .
g < ∞:
Y 1 X
|Ψ0 i = |→ii − |· · · →←i ←i+1 →→ . . .i − . . . (6.16)
i
2g i
which means that C(x, 0) ∼ e−x/ξ (a disordered phase).

One-particle excitations: In the cases

Q
g = ∞: the first excited state is |ii = |←ii j6=i |→ij with  = 2gJ.

g < ∞: then |ki = √1N j eikxj |ji with k = gJ(2 − g2 cos ka + O( g12 )). For this disper-
P
sion relation see figure 6.6.
Chapter 6. Magnetic quantum phase transitions 57

Figure 6.6: Dispersion of

single-spin-flip excitations of
the quantum Ising chain in
the strong-coupling limit.

Two-particle excitations
Q
g = ∞: then |iji = |←ii |←ij l6=i,j |→il

g < ∞: when far apart, particles are independent, when close together, they scatter.
For details see Sachdev’s book.

Dynamic structure factor at T = 0 is

X
S(k, ω) = 2π | hΨα | σ z |Ψ0 i |2 δ(ω − Eα ) (6.17)
|Ψα i

see figure 6.7.

Figure 6.7: Dynamic structure factor of the quantum Ising chain in the strong-coupling
limit.

6.3.2 Weak-coupling limit g  1

The system is in the ordered phase at T = 0, which is implied by

lim C(x, 0) = N02 6= 0 (6.18)

|x|→∞

The ground state is

58 Chapter 6. Magnetic quantum phase transitions

for g = 0 Y

 |↑ii

i
|Ψ0 i = Y (6.19)


 |↓ii
i

and for g > 0 Y gX

|Ψ0 i = |↑ii − |. . . ↑↓i ↑↑ . . .i − . . . (6.20)
i
2 i

One-particle excitation, if
g = 0, is |ii = |. . . ↑↓i ↓↓ . . .i, with  = 2J, where at i the particle is the domain wall.
g > 0, is |ki = √1N i eikxi |ii with k = J(2 − 2g cos ka) + O(g 2 ), see figure 6.8
P

Figure 6.8:
Dispersion of
domain-wall ex-
citations of the
quantum Ising
chain in the weak-
coupling limit.

Dynamic structure factor. There are two special things to note:

• We have an ordered phase, which gives a Bragg peak at k = 0 and ω = 0.
• We are in 1D, thus we can only create 2 domain walls at a time, thus the one
particle peak is absent in 1D.4
See figure 6.9.

6.3.3 Exact spectrum

Representing spins by fermions by the Jordan-Wigner transformation
σiz = 1 − 2c†i ci
Y
σi+ = (1 − 2c†j cj )ci
j<i
Y
σi− = (1 − 2c†j cj )c†i
j<i

4
In higher dimensions, the elementary excitations of the ordered phase would be spin flips instead of
domain walls, resulting in a single-particle peak in the dynamic structure factor.
Chapter 6. Magnetic quantum phase transitions 59

Figure 6.9: Dynamic structure factor of the quantum Ising chain in the weak-coupling
limit.

where the factor j<i (1 − 2c†j cj ) counting the number of fermions to the left ensures
Q
anticommutation rules of the fermions on different sites, while spins on different sites
have commutation rules. The inverse transformation is
Y
ci = ( σjz )σi+
j<i
Y
c†i = ( σjz )σi−
j<i

The transformation is not local! Applying this to quantum Ising chain rotated by 90◦
gives
σix = 1 − 2c†i ci
Y
σiz = − (1 − 2c†j cj )(ci + c†i )
j<i

and plugging into the Hamiltonian

X †
HI = −J (ci ci+1 + c†i+1 ci + c†i c†i+1 + ci+1 ci − 2gc†i ci + g) (6.21)
i

after Fourier transform we have

X 
HI = J 2(g − cos ka)c†k ck − i sin ka(c†−k c†k + c−k ck − g) (6.22)
k

A Bogoliubov transformation
γk = uk ck − ivk c†−k
†
ck = uk γk + ivk γ−k
with u2k + vk2 = 1, u−k = −uk and v−k = −vk correctly chosen, gets rid of the anomalous
terms. The correct choice (with parameter θ defined by uk = cos θk /2, vk = sin θk /2) is
sin ka
tan θk = . (6.23)
cos ka − g
60 Chapter 6. Magnetic quantum phase transitions

The diagonalized Hamiltonian takes the form:

X 1
HI = k (γk† γk − ) , (6.24)
k
2
1
k = 2J(1 + g 2 − 2g cos ka) 2 ≥ 0 . (6.25)
Given that k = 0 only for g = 1 and k = 0, we conclude that we have a quantum critical
point at g = 1. The dispersion minimum near the critical point obeys
∆ ≡ k=0 = 2J|1 − g| (6.26)
from which we conclude νz = 1.
The exact solution can now be used to calculate correlation functions, structure
factor, etc. see Sachdev’s book [1]. We note a remarkable aspect of the solution: The
form (6.24) shows no obvious signature of spontaneous symmetry breaking, and the
qualitative distinction between the two phases appears lost. In fact, the physics is hidden
in the Bogoliubov transformation together with the non-local character of the Jordan-
Wigner transformation. We further note that the two phases of the quantum Ising chain
can be mapped onto each other by a suitable duality transformation which exchanges
g ↔ 1/g – this is consistent with Eq. (6.26).

6.3.4 Results near criticality

The results for the excitation gap, the ferromagnetic moment (the order parameter) and
the flipped-spin quasiparticle weight are shown in figures 6.10 and 6.11

Figure 6.10: Excitation gap and ferromagnetic moment (order parameter) of the quantum
Ising chain near criticality.

The exact solution might suggest that there are well-defined quasiparticles (spin
flips or domain walls) even at the critical point g = 1, but this is incorrect:5 The
5
The non-local character of the Jordan-Wigner transformation is responsible for the non-trivial be-
havior of physical observables.
Chapter 6. Magnetic quantum phase transitions 61

Figure 6.11: Flipped-spin

quasiparticle weight in the
disordered phase of the
quantum Ising model.

Figure 6.12: Dynamic structure factor of

the quantum Ising model at criticality.

structure factor, Fig. 6.12, displays a dissipative critical continuum instead, consistent
with the general arguments made previously. To understand how the structure evolves
from its strong-coupling form, Fig. 6.7, to the quantum critical form, we note (i) both
the weight of the quasiparticle peak and its energy location go to zero upon approaching
the critical point, Z → 0, ∆ → 0. Further, ∆ → 0 also implies that all multi-particle
continua move towards zero energy, and their superposition results in the critical power
law (ω − c|k|)−7/8 . This is illustrated in Fig. 6.13 which shows the structure factor for g
slightly above gc = 1.

Figure 6.13: Dynamic structure factor

in the disordered phase very close to
criticality.

The structure of the ground states and elementary excitations derived so far allows
to deduce the finite-temperature phase diagram, see Fig. 6.14. Finally we note that non-
universal physics takes over at large T where ξ is no longer large but becomes comparable
to the lattice constant a.
62 Chapter 6. Magnetic quantum phase transitions

Figure 6.14: Finite-temperature phase diagram of the quantum Ising chain.

Chapter 7

Quantum phase transitions of

bosons and fermions

In this last chapter we turn to quantum phase transitions where the dynamics of quantum
particles plays a role (as opposed to those with collective density fluctuations only).

7.1 Bose-Hubbard model

We start by considering a microscopic model of bosons on a lattice which interact via an
on-site density-density interaction:
X  X 1 X
H = −w b†i bj + h.c. − µ ni + U ni (ni − 1) (7.1)
i
2 i
hiji

with w being a hopping matrix element and U the strength of the interaction. This
model is known as the Bose-Hubbard model. It applies to systems of charge-neutral
bosons with contact interactions, as realized for helium atoms adsorbed on surfaces (like
graphite) or for ultra-cold bosonic atoms on optical lattices.
Let us discuss the behavior in the low-temperature limit: First, non-interacting
bosons (U/w = 0) form a Bose-Einstein condensate, with all particles condensed at
T = 0. Weak interactions (U/w  1) turn this condensate into a proper superfluid
with phase stiffness, but also reduce the condensate fraction. Strongly interacting lattice
bosons can, in general, host a variety of non-trivial phases. For the simple model in
Eq. (7.1), however, the only other T = 0 phase is a Mott insulator – this phase exists
at strong interactions, U/w  1, and integer filling hni. It is characterized by bosons
being localized on lattice sites, with number fluctuations being suppressed.1
The model (7.1) can be conveniently realized for ultracold bosonic atoms in the
presence of a periodic potential arising from an optical lattice. Figures 7.1 and 7.2
1
Note that the Mott insulator is a genuine lattice phase whereas the superfluid can exist in a contin-
uum system as well.

63
64 Chapter 7. Quantum phase transitions of bosons and fermions

illustrate the lattice potential, with modulation amplitude v. Small v will lead to large
w,2 and the system will be a superfluid. Large v leads to small w: Here the system is a
Mott insulator for integer densities, whereas for non-integer (incommensurate) densities,
hni ∼ 1 + δn, extra particles are mobile such that a superfluid with a condensate density
∼ δn emerges.

Figure 7.1: Bosons in a shallow potential (implying U/w  1): The ground state is
generically superfluid.

Figure 7.2 shows a deep potential.

Figure 7.2: Bosons in a deep potential (i.e. U/w  1): The ground state is superfluid
for non-integer densities.

7.2 Bose-Hubbard model: Mean-field theory

We now describe a simple mean-field theory which captures both the superfluid and the
Mott insulator phase. Given that the Mott insulator is featureless whereas the superfluid
breaks a U(1) symmetry, it is useful to introduce an order parameter for the condensate:
1
ΨB = zwhbi i (bi = ΨB + δbi ) (7.2)
zw
where z is the number of nearest neighbor on the lattice (i.e. coordination number). This
order parameter is used to decouple the kinetic-energy term (!). Then the mean-field
Hamiltonian is purely local:
X 1 †

?
HMF = −µni + U ni (ni − 1) − ΨB bi − ΨB bi + const. (7.3)
i
2

The possible mean-field phases are:

2
For a shallow potential, higher Bloch bands not described by Eq. (7.1) might play a role.
Chapter 7. Quantum phase transitions of bosons and fermions 65

Figure 7.3: Mean-field phase diagram of the Bose-Hubbard model as function of µ/U
and zw/U , showing Mott and superfluid phases.

• ΨB 6= 0: superfluid, which breaks the U(1) phase symmetry ΨB = |ΨB |eiφ , with
the condensate phase φ.

• ΨB = 0: Mott insulator, with no kinetic energy at the mean-field level.

Then the mean-field ground state energy EMF is calculated, by minimizing EMF w.r.t.
Ψb . In general this has to be done numerically, but the case w → 0 can be analyzed
easily: This always yields a Mott insulator having exactly n0 particles per site, with µ
and U determining n0 : 

 0 for µ/U < 0

 1 for 0 < µ/U < 1

n0 = 2 for 1 < µ/U < 2 (7.4)


 ..


.
Figure 7.3 shows the mean-field phase diagram, µ/U as a function of zw/U . The super-
fluid phase is compressible with ∂hni/∂µ 6= 0. The Mott insulator phase is incompressible
with ∂hni/∂µ = 0. This mean-field phase diagram has been checked and confirmed by
Monte Carlo simulations, and mean-field theory has been shown to be qualitatively (and
even semi-quantitatively) correct.
66 Chapter 7. Quantum phase transitions of bosons and fermions

7.3 Superfluid-insulator transition: Field theory

We now turn to discuss the quantum phase transition between Mott insulator and su-
perfluid in more detail. It turns out that the behavior of the boson density play a key
role. This density is fixed in each Mott insulator phase, but varies continuously in the
superfluid phase. This is illustrated in Figure 7.4 which shows lines of constant density.

Figure 7.4: Lines of constant boson density hni in the phase diagram of the Bose-Hubbard
model.

As we will derive below, there exist two distinct universality classes for the Mott–
superfluid quantum phase transiton:

• If the density is constant across the transition (this applies to the tip of the Mott
lobe, as derived below), the critical fluctuations are phase fluctuations of the con-
densate, and consequently the transition is that of an O(2) rotor model.

• If density varies across the transition (away from the Mott lobe), both condensate
phase and density are critical. The corresponding critical theory will be derived
now.

We start from a path-integral representation of the Bose-Hubbard model:

Z Z 1/T
ZB = D[bi (τ )]D[b?i (τ )] exp (− LB ) (7.5)
0
 
b?i ∂bi − µb?i bi + 1 U b?i bi (b?i bi − 1) − w
X X
LB = (b?i bj + c.c.) (7.6)
i
∂τ 2
hiji

Introduce an auxiliary field ψBi (τ ) via Hubbard-Stratonovich transformation of the hop-

ping term, which eliminates the non-local term. Then
Z Z 1/T
ZB = D[bi (τ )]D[b?i (τ )]D[ψB ]D[ψB? ] exp (− L0B ) (7.7)
0
Chapter 7. Quantum phase transitions of bosons and fermions 67

with the new Lagrangian

X  ∂bi 1 ?
 X
0 ? ? ? ? ? −1
LB = bi − µbi bi + U bi bi (bi bi − 1) − ψBi bi − ψBi bi + ψB? i wij ψBj (7.8)
i
∂τ 2
hiji

where (
w if i, j are neighbors
wij = (7.9)
0 otherwise
Now we integrate out the b field – this is possible since the terms are all local – and
arrive at an action for the ψB alone,
Z  Z 
? 00
ZB = D[ψB ]D[ψB ] exp − dτ LB − F0 /T , (7.10)

with F0 /T being the constant arising from the transformation. Instead of explicitly
deriving the action for ψB , we guess the result by writing down all the terms allowed by
symmetry. This yields:
Z
00
ZB = DΨB DΨ?B e−S (7.11)
Z Z  2 !
∂Ψ B ∂Ψ B 1
 + r |ΨB |2 + K3 |∇ΨB |2 + U |ΨB |4 + · · ·
S 00 = dd x dτ K1 Ψ?B
 
+ K2 
∂τ ∂τ  2
(7.12)

where r is the control parameter of the theory. Here we have included both a linear
and a quadratic time derivative, which we analyze now. To this end, we note that L0B is
invariant under the U(1) gauge transformation:

bi → bi eiφ(τ )
ψBi → ψBi eiφ(τ )
∂φ
µ→µ+i
∂τ
Demanding that L00B is also invariant under this transformation gives
∂r
K1 = − (7.13)
∂µ
i.e., the linear-time-derivative term is present only if ∂r/∂µ is non-vanishing.

7.3.1 Quadratic vs. linear time derivative

The theory (7.11) can have two qualitatively distinct low-energy behaviors, depending
on whether or not K1 is finite. (i) For K1 = 0 there is a quadratic time derivative only,
68 Chapter 7. Quantum phase transitions of bosons and fermions

as in the quantum φ4 theory. From Eq. (7.13) we see that K1 = 0 applies to the tip of
the Mott lobe where the transition occurs with at fixed density hni. Consequently, this
transition is that of an O(2) φ4 theory, or, equivalently, of an O(2) rotor model. The
critical variable is the phase of the condensate (and the density is fixed). (ii) For K1 6= 0,
away from the tips of the Mott lobes, a linear time derivative is present which dominates
at low energies over the quadratic one. Hence, the critical theory is different: As we will
see below, the density itself becomes a critical degree of freedom, and the transition does
not have a straightforward quantum-to-classical correspondence. A detailed discussion
of the transition follows in Section 7.4.
Having realized that the order of time derivative is crucial in determining quantum
critical properties, let us recall where these time derivatives arise from: In the quantum
φ4 theory the quadratic time derivative is dictated from the gradient expansion, combined
with demanding time-reversal invariance, for the order-parameter field theory. The linear
time derivative instead can be traced back to the dynamics of canonical quantum particles
which genuinely has a time direction. Phrased differently, for a quantum particles forward
and backward propagation need to be distinguished (recall that Feynman diagrams have
arrows). However, for hermitian objects (like order-parameter fields) propagators have
no time arrow – in a sense, their propagators are sums of forward and backward processes.
This can be nicely seen by thinking about harmonic oscillators, phonons, and the like:
The Hamiltonian is written in terms of canonical bosons, and their propagators are
directed. In contrast, both position and momentum operators are hermitian, and their
propagators are undirected. A critical theory involving these variables only would be of
φ4 type, with quadratic time derivative.

7.4 Dilute Bose gas

We continue by studying a model of weakly interacting bosons with a chemical potential
near zero, describing a dilute gas of bosons:
X k2 † X X
HB = bk bk − µnk + U ni (ni − 1) (7.14)
k
2m k i

we get a continuum Lagrangian

Z  
d ? ∂ψB 1 2 2 1 4
LB = d x ψB + |∇ψB | − µ|ψB | + U |ψB | (7.15)
∂τ 2m 2
At T = 0 this model has a quantum critical point at µ = 0. For µ < 0 the boson density
vanishes, whereas a superfluid with finite boson density emerges for µ > 0. Clearly,
this transition is identical to any of the Mott–superfluid transition of the Bose-Hubbard
model in Fig. 7.3 away from the tip of the Mott lobe (if an integer background density
of bosons is added).
The field theory in Eq. (7.15) has dynamical exponent z = 2, and hence the upper
critical dimension is d+c = 2. For d < 2 boson self interactions are relevant at criticality,
Chapter 7. Quantum phase transitions of bosons and fermions 69

Figure 7.5: Phase diagram of the dilute Bose gas (in dimensions d > 2) as function of
chemical potential µ and temperature T . The superfluid phase is bounded by a finite-
temperature phase transition which ends at the QCP at µ = 0, T = 0.

and the theory needs to be analyzed by RG techniques, see Sachdev’s book [1] for details.
Remarkably, for d = 1 the critical fixed point is identical to that of the dilute Fermi gas
to be discussed below.
For d > 2 boson self interactions are irrelevant. Hence, mean-field theory applies but
70 Chapter 7. Quantum phase transitions of bosons and fermions

the quartic interaction u is dangerously irrelevant and needs to be kept. The density
obeys (
0 µ<0
hψB? ψB i = (7.16)
µ/u + . . . µ > 0
The finite-temperature crossovers are somewhat non-trivial: There is a dilute classical
gas regime for T < −µ where the particle density is exponentially small. The high-
temperature quantum critical regime is bounded by crossover lines T ∼ (|µ|/u)2/3 in
d = 3, and another crossover regime exists for µ < 0. The phase diagram for d = 3 is
sketched in figure 7.5.
We note a special property of the dilute Bose-gas model: (i) For µ < 0, T = 0
the system is “empty”, i.e., there are no particles and no fluctuations. This implies that
there is no diverging correlation upon approaching the QCP at T = 0 from the negative-µ
side. Nevertheless, the finite-temperature crossovers follow the standard quantum-critical
phenomenology.

7.5 Dilute spinless Fermi gas

We finally turn to systems of fermions. The simplest situation is given by spinless non-
interacting fermions – this is, of course, a textbook example of an exactly solvable model
system. Focussing on the situation with the chemical located near the band bottom, the
following discussion will cast well-known results into the scaling perspective of quantum
phase transitions, and hence will be useful for a broader picture.
We consider a model of free fermions (defined either on a lattice or in the continuum)
X k2 † X
HF = ck ck − µnk (7.17)
k
2m k

which is equivalent to the continuum Lagrangian

Z  
d ? ∂ψF 1 2 2
LB = d x ψF + |∇ψF | − µ|ψF | . (7.18)
∂τ 2m
It can be shown that interactions (which must be non-local in space due to the Pauli
principle) are irrelevant, hence are not written down – this is fundamentlly different from
the dilute Bose gas discussed above.
At T = 0 the Fermi-gas model has a quantum critical point at µ = 0 where the
fermion density develops singular behavior:
(
0 µ<0
hψF? ψF i = d/2
(7.19)
(Sd /d)(2mµ) µ>0

with Sd = 2/[Γ(d/2)4π d/2 ]. In other words, a quantum phase transition occurs upon
tuning the chemical potential from zero to finite density of particles. In contrast to
Chapter 7. Quantum phase transitions of bosons and fermions 71

Figure 7.6: Phase diagram of the dilute Fermi gas as function of chemical potential µ and
temperature T . The QCP determining the structure of the phase diagram is at µ = 0,
T = 0; it is the endpoint of a critical line located at µ > 0, T = 0 which controls the
Fermi-liquid phase. The dashed lines mark the crossover T ∗ ∝ |µ|νz with νz = 1; the
solid lines denote lines of constant density. Figure taken from [1].

the case of bosons discussed in the previous section, the present QPT does not involve
spontaneous symmetry breaking.3
Using standard thermodynamic relations, one can derive the free energy density which
reads
dd k
Z  
(µ−k2 /(2m))/T
FF = −T ln 1 + e (7.20)
(2π)d
which can be cast into the form
µ
FF = T d/2+1 ΦFF (7.21)
T
where ΦFF is a universal scaling function. Consequently, all thermodynamic observables
show scaling behavior as function of µ/T . The quantum-critical regime must therefore
be bounded by lines ±µ ∼ T , i.e., νz = 1. The quadratic dispersion of the particles at
µ = 0 suggests z = 2, hence ν = 1/2. Of course, these exponents also follow from a
thorough scaling analysis [1].
The crossover phase diagram is shown in Fig. 7.6. The low-temperature regime for
µ < 0 is characterized by an exponentially small density of particles, hence their behavior
is classical even in the (fixed-µ) low-temperature limit (recall that the thermal de-Broglie
wavelength only diverges in a power-law fashion as T → 0.) In contrast, in the high-
temperature quantum critical regime both the inter-particle spacing and the thermal
3
Sometimes the transition of the dilute Fermi gas is referred to as Lifshitz transition, because the
topology of the Fermi surface changes (from non-existent to pocket-like).
72 Chapter 7. Quantum phase transitions of bosons and fermions
√
de-Broglie wavelength scale as 1/ T , such that quantum and thermal effects are equally
important. Finally, the low-temperature regime for µ > 0 is a standard finite-density
Fermi gas (which turns into a Fermi liquid upon including interactions, except for d = 1
where it should be properly labelled Luttinger liquid).
As noted for the Bose gas, the µ < 0, T = 0 state is fluctuationless. Further,
the Fermi liquid itself can be understood as a critical system – it displays power-law
correlations due to the presence of a Fermi surface - such that the T = 0, µ > 0 line is a
line of quantum critical points. Thanks to the linear dispersion near the Fermi surface,
this critical line is characterized by a dynamical exponent z = 1 (as opposed to z = 2 at
the µ = 0 QCP).

7.6 Fermi-Hubbard Model

In this final section, we discuss the fermionic (=original) version of the Hubbard model,
which reads X † X X
H = −t (ciσ cjσ + h.c.) − µ niσ + U ni↑ ni↓ (7.22)
hijiσ iσ i

for a single band of spin-1/2 fermions where σ =↑, ↓ is a spin label. This model was
proposed by Hubbard, Gutzwiller, and Kanamori in the 1960s, mainly as a model to
describe the occurrence of ferromagnetism in materials with significant electron–electron
interactions.
As for bosons, this model features an incompressible Mott-insulating phase at filling
n = 1 (dubbed half filling), while n = 0 and n = 2 are commonly called band insulators.
The compressible phase realized at small interactions or non-integer filling is a Fermi-
liquid metal (which may be unstable to superconductivity, see below). A minimal phase
diagram, similar to its bosonic counterpart, is sketched in figure 7.7.

Figure 7.7: Simplified generic phase dia-

gram of the Hubbard model. Symmetry-
broken states (magnetic, superconduct-
ing, etc.) are not shown.

As compared to the Bose-Hubbard model discussed above, the main difference is

(apart from the restricted local Hilbert space of fermions) the fact that fermions carry
spin. As a result, the Mott insulator is not a trivial featureless phase, because spin
Chapter 7. Quantum phase transitions of bosons and fermions 73

degrees of freedom (i.e. local moments) remain once the charges are frozen. As a re-
sult, magnetism plays a central role, which cannot only occur within the Mott-insulating
phase, but also in the metal. Local moments and magnetic fluctuations can lead to a
plethora of ordering phenomena, including the occurrence of high-temperature super-
conductivity. In fact, the one-band Hubbard model (7.22) is one of the simplest models
which is believed to describe the physics of cuprate superconductors. However, many
properties of this model are not understood to a sufficient degree to date, and continue
to constitute an active area of current research.

7.6.1 Antiferromagnetism in the Mott insulator

To illustrate the emergence of magnetism, we consider the Mott-insulating phase in
the limit of U/t  1. In this regime n = 1 (with density fluctuations suppressed as
U/t → ∞), i.e., we have localized spins 21 per lattice site.

Figure 7.8: Spin degrees of

freedom at half filling.

For U/t = ∞ (or t = 0) particles are strictly immobile, hence all 2N spin states are
degenerate. Departing from this limit, virtual hopping processes like

↑↓ ⇒ · l ⇒ ↓↑ (7.23)

become possible for antiparallel spins on neighboring sites. These lead to an energy gain
of order t2 /U . One can derive an effective model for the spin degrees of freedom valid
for t  U and hni = 1, which turns out to be an antiferromagnetic Heisenberg model:4
X 1

H=J ~ ~
Si · Sj − (7.24)
4
hiji

2
with an antiferromagnetic exchange coupling J = 4 tU .
The Heisenberg model (7.24) displays an antiferromagnetic ground state on bipartite
lattices, but can feature other types of phases (including spin liquids, valence-bond solids
etc) on frustrated lattices.
4
The Heisenberg form of the effective Hamiltonian is dictated by SU(2) symmetry.
74 Chapter 7. Quantum phase transitions of bosons and fermions

7.6.2 Phases of the Hubbard model

The physics of the Hubbard model, both at and away from half-filling, is extremely rich
and is not fully understood even for the square lattice.[11]
At half-filling where a metal is realized for small U/t (except for cases with perfect
nesting of the Fermi surface) and an antiferromagnetic Mott insulator appears at large
U/t, it is not a priori clear whether the phase boundary for antiferromagnetism and the
one for the metal-insulator transition coincide on the U/t axis. While their coincidence
may result in a simple first-order transition, a more interesting scenario seems to be
realized on the triangular lattice, where a spin-liquid paramagnetic Mott insulator exists
between the paramagnetic metal and the antiferromagnetic Mott insulator.
Away from half-filling, where the system is conducting, a plethora of symmetry-
breaking phases have been discussed, and most research has been focussed on the square
lattice believed to be relevant for the description of cuprate high-temperature supercon-
ductors. Most important is d-wave superconductivity which appears over a sizable range
of dopings and interaction strengths U/t. Other possible phases include metallic charge
and spin density waves as well as nematic states with broken lattice rotation but intact
translation symmetries.
Bibliography

[1] S. Sachdev, Quantum Phase Transitions (2nd ed.), Cambridge University Press
(2011).

[2] N. Goldenfeld, Lectures on Phase Transitions and the Renormalization Group, West-
view Press (1992).

[3] E. Fradkin, Field Theories of Condensed Matter Physics (2nd ed.), Cambridge Uni-
versity Press (2013).

[4] L. Zhu, M. Garst, A. Rosch, and Q. Si, Phys. Rev. Lett 91, 066404 (2003).

[5] D. Bitko, T. F. Rosenbaum, and G. Aeppli, Phys. Rev. Lett. 77, 940 (1996).

[6] R. Coldea et al., Science 327, 177 (2010).

[7] Ch. Rüegg et al., Nature 423, 62 (2003).

[8] M. Jaime et al., Phys. Rev. Lett. 93, 087203 (2004).

[9] A. Kopp and S. Chakravarty, Nature Phys. 1, 53 (2005).

[10] H. von Löhneysen, A. Rosch, M. Vojta, and P. Wölfle, Rev. Mod. Phys. 79, 1015
(2007).

[11] P. A. Lee, Rep. Prog. Phys. 71, 012501 (2008).

75