Computational Studies

Computational Studies of Quantum Spin Systems
Anders W. Sandvik
Citation: AIP Conf. Proc. 1297, 135 (2010); doi: 10.1063/1.3518900

View online: http://dx.doi.org/10.1063/1.3518900
View Table of Contents: http://proceedings.aip.org/dbt/dbt.jsp?KEY=APCPCS&Volume=1297&Issue=1
Published by the American Institute of Physics.
Related Articles
Phononic and magnetic excitations in the quasi-one-dimensional Heisenberg antiferromagnet KCuF3
Low Temp. Phys. 38, 419 (2012)
Coexistence of two cycloid-phases in multiferroic RMnO3: Consequence of the next-nearest-neighbor spin
interaction modulations
J. Appl. Phys. 111, 083918 (2012)
Counterexamples to ferromagnetic ordering of energy levels
J. Math. Phys. 53, 043302 (2012)
Macroscopic thermal entanglement in a spin chain caused by the magnetic field: Inhomogeneity effect
Low Temp. Phys. 38, 210 (2012)
Anomalous low-temperature magnetic and magnetotransport properties in Ru-deficient SrRuO3
J. Appl. Phys. 111, 07E121 (2012)
Additional information on AIP Conf. Proc.

Journal Homepage: http://proceedings.aip.org/
Journal Information: http://proceedings.aip.org/about/about_the_proceedings
Top downloads: http://proceedings.aip.org/dbt/most_downloaded.jsp?KEY=APCPCS
Information for Authors: http://proceedings.aip.org/authors/information_for_authors
Downloaded 04 Jul 2012 to 139.184.30.132. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
Computational Studies
of Quantum Spin Systems
Anders W. Sandvik
Department of Physics, Boston University,
590 Commonwealth Avenue, Boston, Massachusetts 02215, USA
Abstract. These lecture notes introduce quantum spin systems and several computational methods
for studying their ground-state and finite-temperature properties. Symmetry-breaking and critical
phenomena are first discussed in the simpler setting of Monte Carlo studies of classical spin sys-
tems, to illustrate finite-size scaling at continuous and first-order phase transitions. Exact diago-
nalization and quantum Monte Carlo (stochastic series expansion) algorithms and their computer
implementations are then discussed in detail. Applications of the methods are illustrated by results
for some of the most essential models in quantum magnetism, such as the S = 1/2 Heisenberg an-
tiferromagnet in one and two dimensions, as well as extended models useful for studying quantum
phase transitions between antiferromagnetic and magnetically disordered states.
Keywords: Quantum spin system, antiferromagnet, valence-bond solid, quantum phase transition,
finite-size scaling, quantum Monte Carlo, exact diagonalization, Lanczos method
PACS: 75.10.Jm, 75.40.Mg, 75.40.Cx, 02.70.Ss
1. INTRODUCTION
One of the primary goals of theoretical physics is to provide the simplest models cap-
turing various complex physical phenomena. In condensed matter physics, as well as in
statistical mechanics more broadly, spin systems often serve in this role. As the Ising
model and other idealized classical spin models have been invaluable in forming our un-
derstanding of thermal phase transitions and critical phenomena, so are various quantum
spin models now instrumental in developing a theoretical framework for exotic quantum
many-body states and quantum phase transitions, i.e., phase transitions driven by quan-
tum fluctuations (controlled by some tunable interaction parameter) at temperature T = 0
[1, 2]. With their many possible ordered and disordered ground states and different types
of excitations arising from them, quantum spin systems also provide rich opportunities
to study other manifestations of collective quantum behavior [3, 4].
Although often used as simplified prototypical model systems for various phenom-
ena, and not always intended for describing fully all the details of specific real materials,
quantum spin systems have also been very successful in explaining quantitatively the an-
tiferromagnetic properties of a variety of Mott insulators with localized electronic spins.
A prime example of this is provided by the undoped parent compounds of the high-
temperature superconductors and other related quasi-two-dimensional and quasi-one-
dimensional copper oxides. Their experimentally measured magnetic response func-
tions can be remarkably well reproduced by the S = 1/2 Heisenberg model on two-
dimensional (2D) planes [5, 6], isolated chains [7, 8], and “ladders” [9] consisting of
two or more
CREDIT coupled
LINE chains.
(BELOW) TO In
BEagreement
INSERTED with
ONmodel calculations,
THE FIRST PAGE2D OFlayered
EACH systems
PAPER
CP1297, Lectures on the Physics of Strongly Correlated Systems XIV: Fourteenth Training Course
edited by A. Avella and F. Mancini
© 2010 American Institute of Physics 978-0-7354-0851-7/10/$30.00
135
exhibit an exponentially divergent correlation length as T is lowered (until ordering sets
in below some critical temperature due to 3D couplings or anisotropies), while chain
and ladder compounds exhibit only short-range (power-law or exponentially decaying)
correlations. In addition to cuprates, many other inorganic and organic antiferromagnets
also show similarly good agreement between theory and experiments [10].
A prominent research theme in contemporary condensed matter physics is to model
and explain magnetically disordered ground states of 2D or quasi-2D materials with
non-uniform or frustrated (competing) antiferromagnetic interactions [11, 12, 13, 14].
Quantum phase transitions in 2D spin systems challenge the classical Ginzburg-Landau
framework [15] for understanding and classifying phase transitions based on order
parameters, as exemplified by the recent theory of “deconfined” quantum critical points
[2, 16], which separate antiferromagnetic (Néel) and non-magnetic valence-bond solid
(VBS) ground states [17]. In a field-theory proposed to describe this quantum phase
transition (the non-compact CP1 model), deconfined spinons (collective S = 1/2 degrees
of freedom) are the “elementary particles”, out of which the two order parameters can
be formed due to condensation (in the Néel state) or confinement (in a VBS state) [16].
Apart from the interest in such unusual phase transitions in condensed matter physics,
there are also intriguing connections to deconfinement in gauge theories in particle
physics [18]. Interacting quantum spins have also recently become interesting in the
context of ultra-cold atoms in optical lattices [19, 20], as well as in quantum information
theory [21]. Fundamental many-body concepts such as entanglement entropy [22] are
currently explored in various ground states of quantum spin systems [23].
Exact solutions of quantum spin systems are very rare beyond one dimension, where
there are several important cases (enough to fill a whole encyclopedia [24], in fact). In
two dimensions there are also some examples [25, 26], but normally analytical calcula-
tions rely on approximations or assumptions that cannot be rigorously justified. Purely
computational studies of model hamiltonians are therefore also essential. Unbiased nu-
merical results are important for testing theories and analytical calculations (in particu-
lar, continuum field theories for the low-energy physics). Moreover, numerical “simula-
tions” can also in their own right serve as laboratories for exploration and discovery, and
may thus stimulate further theoretical and experimental developments.
In classical statistical physics, almost any model can be studied in detail using Monte
Carlo or molecular dynamics simulations (although there are also challenging classical
systems, e.g., ones with very slow, “glassy” dynamics [27]). The situation is different
in quantum mechanics. There are still enormous hurdles limiting computational studies
of generic quantum spin hamiltonians, especially ones with frustrated interactions and,
going beyond pure spin models, strongly correlated fermion systems. Devising efficient
practically useful numerical algorithms for these types of systems is one of the greatest
challenges in theoretical physics. Thanks to a series of significant developments over the
past couple of decades, large-scale computational studies have already become possible
for some important classes of quantum lattice models. Very large 1D systems can be
studied using the density matrix renormalization group (DMRG) method [28, 29] or
related methods formulated using matrix-product states [21, 30]. Quantum Monte Carlo
(QMC) methods with loop-cluster updates [31, 32, 33] can be used to study a wide
range of spin and boson models in any number of dimensions, typically on lattices with
up 104 sites or more in the ground state, and much larger still at elevated temperatures.
136
In addition to breakthroughs in efficient algorithms, the impressive improvements in
computer performance have of course played an important role in recent progress, too.
The most dramatic gains have, however, been achieved as a result of better algorithms,
and there is reason to believe that this will continue to be the case in the future as well.
Topics. Two classes of computational methods will be discussed in these lecture

notes: Numerical (exact) diagonalization and quantum Monte Carlo simulation. Exact
diagonalization methods will be developed primarily for 1D systems, followed by some
discussion of extensions to 2D square-lattice systems. The use of symmetries for block-
diagonalization will be developed and used in both complete diagonalization (T > 0
calculations) and with the Lanczos method (for obtaining the ground state and low-
energy excitations in a given symmetry sector). QMC simulations based on the series
expansion of the partition function (stochastic series expansion; SSE) will be developed
for T > 0 calculations (and applied also in the limit T → 0). Computer programs written
in close correspondence with the pseudocodes are available on-line [34].
Beyond describing the technical aspects of the numerical methods, an integral goal of
these lecture notes is to introduce the most essential quantum spin models (hamiltonians)
and to present some of their physical properties from a computational perspective. While
the discussion is largely self-contained as far as the algorithms and implementations are
concerned, the physics of the systems is for the most part discussed in a rather “light”
fashion, in the form of elementary calculations (e.g., spin-wave theory) and qualitative
descriptions illustrated by numerical results. Connections to complementary analytical
approaches (e.g., predictions based on field theories) are also pointed out, with key
references for further study. The topics range in maturity from well established basics to
very recent and ongoing research on exotic quantum phase transitions.
Outline. The quantum spin models to be discussed in the subsequent sections are
first introduced in Sec. 2, along with a brief summary of various types of ground states
and quantum phase transitions. Classical phase transitions, Monte Carlo simulations, and
finite-size scaling techniques are reviewed Sec. 3 in order to set the stage for quantum-
mechanical finite-lattice calculations and data analysis. Exact diagonalization techniques
and their applications to 1D spin systems are discussed in Sec. 4. Basic properties of the
Heisenberg chain and its extension with frustrated interactions are illustrated with nu-
merical results (including the frustration-driven quantum phase transition into a dimer-
ized VBS state). Extensions of the methods to 2D systems are also summarized, and
used to study the low-energy states (quantum-rotor states) of small antiferromagnetic
systems. Sec. 5 begins with a general discussion of path integrals, followed by the al-
ternative series-expansion formulation of quantum statistical mechanics, on which the
SSE QMC method is based. The SSE method is then developed in detail for the S = 1/2
Heisenberg model. Illustrative results for chains, ladders, and 2D planes are presented,
including a study of quantum-criticality in dimerized 2D systems. Applications of the
SSE method to “J-Q” models with four- and six-spin interactions are also discussed, and
the Néel–VBS transitions occurring in these systems as a function of the strength of the
multi-spin iteractions are studied. Sec. 6 concludes with a brief survey of other recent
works related to the topics of the lecture notes.
137
2. QUANTUM SPIN MODELS, THEIR GROUND STATES
AND QUANTUM PHASE TRANSITIONS
In solid-state physics, quantum spin hamiltonians describe the effective magnetic inter-
actions between localized electronic spins. As such, they can be derived starting from the
full problem of interacting electrons [35, 36, 37]. Here we will not discuss their relation-
ships with real materials in detail, but just introduce some well established models with
Heisenberg couplings and other, similar interactions. To motivate and prepare for the
quantitative numerical calculations in the later sections, in this introductory part we will
first survey some of the possible types of ground states and quantum phase transitions.
We will discuss the nature of the quantum fluctuations in spin systems from different
perspectives, including spin wave theory and singlet pairing (valence bonds).
The Heisenberg exchange is the most important spin-spin interaction and forms the
starting point for understanding many materials and phenomena in quantum magnetism.
Two spins are coupled according to the hamiltonian
y y
Hi j = Ji j Si · S j = Ji j (Six Sxj + Si S j + Siz Szj ). (1)
Often this pair interaction is summed over only nearest-neighbor sites (i, j), but longer-
range interactions can also be included. The type of ground state, the nature of the ex-
citations, and the finite-temperature properties of a system with Heisenberg interactions
depend strongly on the underlying lattice. The dimensionality plays a crucial role. Ac-
cording to the Mermin-Wagner-Hohenberg theorem [38, 39], a continuous symmetry
of a quantum system with short-range interactions, here the global SU(2) spin rotation
symmetry, can be broken neither at T ≥ 0 in one dimension nor at T > 0 in two dimen-
sions. This normally rules out magnetic order in 1D Heisenberg models (unless there
are long-range interactions, in which case the theorem does not apply—we will discuss
an example of this in Sec. 4.3.3), but in two dimensions the ground state can be mag-
netic, i.e., hSi i =
6 0 (with all these vector expectation values parallel in a ferromagnet
and alternating between two opposite directions in a bipartite antiferromagnet). Beyond
dimensionality, the microscopic details of the lattice and the coupling strengths Ji j (e.g.,
uniform, modulated in some periodic pattern, or in some random, disordered way) are
also decisive, and many different types of ground states and quantum phase transitions
can be realized. Some of these states and transitions are still not very well understood
and subjects of ongoing research.
Apart from a brief review of spin wave theory for general S, we will in these lecture
notes focus on the simplest case of S = 1/2 spins, corresponding to individual uncom-
pensated electronic spins. This is often the most interesting case, as S → ∞ is the classi-
cal limit, and S = 1/2, thus, maximizes the effects of quantum fluctuations (although in
some cases, S = 1 or higher can actually lead to even larger quantum effects, e.g., in the
case of the “Haldane state” of the S = 1 chain [26]). We will only consider antiferromag-
netic interactions, Ji j > 0 in (1), which from a theoretical perspective are more interest-
ing than ferromagnetic couplings. Antiferromagnetic interactions in strongly-correlated
systems are also more prevalent in nature. A much broader area of quantum many-body
physics can be entered by also allowing anisotropies in spin space, i.e., different x, y, and
z coupling strengths in (1). In addition to the relevance of such anisotropies in many real
138
magnetic materials, the mapping between S = 1/2 spins and “hard-core” bosons makes
such models interesting also for other reasons. The spin-isotropic Heisenberg interac-
tion can be considered the essence of quantum magnetism, however, and we will focus
almost exclusively on this case here.
Section Outline. After discussing the antiferromagnetic (Néel) state and the nature
of its quantum fluctuations based on spin wave theory in Sec. 2.1, two important classes
of non-magnetic states—spin liquids and valence-bond solids—will be discussed in
Sec. 2.2. In Sec. 2.3 the special properties of 1D systems are briefly reviewed. Sec. 2.4
introduces the extended quantum Heisenberg models and the types of quantum phase
transitions that we will study in much more detail in the later sections, in applications of
the various computational methods.
2.1. The Néel state and its quantum fluctuations

On the 2D square lattice (and other bipartite 2D and 3D lattices with uniform inter-
actions) the ground state of the Heisenberg model with only nearest-neighbor interac-
tions is antiferromagnetically (Néel) ordered, with neighboring spins being oriented, on
average, in an antiparallel (staggered) fashion. Note that a maximally ordered antiferro-
magnetic state, e.g., | ↑↓↑↓, . . .i on a chain (or a checkerboard pattern of ↑ and ↓ spins
on the 2D square lattice), is not an eigenstate of the Heisenberg hamiltonian, whereas a
fully polarized ferromagnetic state, e.g., | ↑↑↑↑ . . .i, is an eigenstate (and, in the case of
ferromagnetic interactions, has the minimum energy). This can be easily seen with the
pair interaction (1) written as
y y
Hi j = Ji j (Six Sxj + Si S j + Siz Szj ) = Ji j [SizSzj + 12 (Si+ S−j + Si− S+j )]. (2)
When acting on the perfect Néel state, the raising and lowering operators flip pairs
of spins, causing local defects, whereas they destroy the perfect ferromagnetic state.
The antiferromagnetic order must therefore always be reduced by quantum fluctuations,
whereas the fully polarized ferromagnetic state is the ground state also in the presence
of the off-diagonal interactions. The amount of magnetic order (if any) remaining in the
true ground state of a system with antiferromagnetic interactions depends sensitively on
details of the lattice and the interactions included in the hamiltonian.
Note again that the magnetic order parameter of a system with Heisenberg interactions
(e.g, the magnetization of a ferromagnet or the sublattice magnetization of an antiferro-
magnet) is a vector in spin space. The sublattice magnetization operator is
1 N
ms = ∑ φiSi,
N i=1
(3)
where φi = ±1 is the staggered phase factor, e.g., on the 2D square lattice φi = (−1)xi +yi ,
where xi and yi are the lattice (integer) coordinates of site i. In a Néel state the expectation
value hms i = mxs x̂ + mys ŷ + mzs ẑ is non-zero in the thermodynamic limit. The order can
form—the spin-rotational symmetry can be broken—in any direction in spin space. For
139
convenience one normally associates the z spin components with this direction, so that
the staggered magnetization hms i = |hSizi|. In a finite system a non-zero magnitude ms of
the Néel order can form, even though the direction of the vector remains fluctuating over
all angles and, thus, hms i = 0. In a calculation for a finite system one should therefore
detect the presence of order using a quantity which is independent of the direction, e.g.,
hm2s i or h|ms|i. We will see many examples of finite-lattice calculations later. First, let us
discuss some of the basic properties of the symmetry-broken Néel state in more detail.
2.1.1. spin wave theory

The Néel state and its excitations in 2D and 3D systems can be understood within a
simple linear spin wave theory (discussed in more detail in many standard works, e.g.,
the review article by Manousakis [6] and the book by Mattis [36]). Such a calculation
starts from a maximally ordered state of staggered spins, which is the exact ground state
in the classical limit where the spin magnitude S → ∞. This is regarded as a vacuum
state, on top of which quantum effects are included systematically by adding bosons,
representing the deviations of the spins from |Siz| = S, in such a way as to obtain a good
approximation to the ground state of the system for finite S.
The relationship between spins and bosons for S = 1/2 is illustrated in Fig. 1. The
reason for using this mapping is that it is easier to carry out calculations with the bosons,
due to their simpler commutation relations. Linear spin wave theory corresponds to non-
interacting bosons and is, by construction, exact for S → ∞. Results for finite S can be
systematically improved by including interactions in the form of an 1/S expansion. Here
we just outline the lowest-order calculation.
If we neglect the constraint that the boson number ni for each site should be within the
range 0, . . ., 2S (the physical subspace), we can use the following simple (lowest order
in 1/S) mapping between the spin operators in the original hamiltonian (2) and boson
creation and destruction operators a+ +
i and ai (and the number operators ni = ai ai );
1
√ √
i ∈ ↑ sublattice : Siz = S − ni , Si+ = √2Sai , Si− = √2Sa+ i (4)
j ∈ ↓ sublattice : Szj = n j − S, S+j = 2Sa+j , S−j = 2Sa j .
It is useful to look at Fig. 1 to understand this mapping—apart from the obvious way in
which the off-diagonal operators can affect the states, one just has to be careful with the
different factors associated with boson creation/destruction and spin raising/lowering
(discussed in standard quantum-mechanics texts). The factors in (4) are correct for
ni ≪ S, but note that they are also exact for S = 1/2 in the physical subspace. In principle,
one can write down more complicated expressions that are formally correct for any S and
ni in the physical subspace (and also automatically decouple the physical and unphysical
subspaces), but these are only relevant when going beyond linear spin wave theory.
1 Often two species of bosons, ai and b j , are used, corresponding to the two sublattices. In momentum
space, the use of a single boson species implies that we are here working in the full Brillouin zone of the
original lattice with N momenta, instead of one with N/2 momenta corresponding to a two-site unit cell.
140
spins
bosons
FIGURE 1. Correspondence between S = 1/2 spins and bosons, here illustrated for a 1D chain. The
fully staggered reference state (left) is the vacuum for bosons (with – indicating zero boson occupation).
For each spin flipped with respect to the reference state there is a boson (•) at the corresponding site
(right). For S = 1/2 the boson occupation numbers are 0 or 1 on each site, while for arbitrary spins they
are 0, . . . , 2S. Linear spin wave theory corresponds to non-interacting bosons, and the occupation number
constraints are not enforced. Higher-order calculations in 1/S gradually restore the constraints.
We now transform the terms in the hamiltonian (2) using the mapping (4). Because
we are considering a bipartite system, where the two sites i and j are always on different
sublattices, we obtain the off-diagonal term
1 + − − + + +
2 (Si S j + Si S j ) → S(ai a j + ai a j ), (5)
and the diagonal interaction is
SizSzj → −S2 + S(ni + n j ) − ni n j . (6)
spin wave theory should formally be regarded as a large-S calculation. In the lowest-
order calculation we should therefore discard the interaction term ni n j in (6), because
it is a factor 1/S smaller than the non-interacting contributions from (5) and (6). Let us
for definiteness consider the two-dimensional square lattice with N = L2 sites. We then
have the effective hamiltonian
N
H = −2NS2 J + 4SJ ∑ ni + SJ ∑ (ai a j + a+ +
i a j ). (7)
i=1 hi ji
Here it should be noted that we consider a finite system but assume that the symmetry
is broken (the direction of the staggered magnetization has been locked) which can be
strictly true only in the limit N → ∞ (unless some symmetry-breaking field is added to
H). This is fine, however, because we will anyway take the limit N → ∞ at the end.
The boson hamiltonian (7) can be easily diagonalized (i.e., written in terms of number
operators). To construct this solution, we first Fourier transform,
ak = N −1/2 ∑ eik·r ar , ar = N −1/2 ∑ e−ik·r ak , (8)
r k
where the real-space operators have been labeled by their lattice coordinate vectors r
instead of just the site index i, and the momentum k is in the first Brillouin zone of the
square lattice (i.e., the reciprocal square lattice of N sites). The hamiltonian is then
H = −2NS2 J + 4SJ ∑ nk + 2SJ ∑ γk (ak a−k + a+ +
k a−k ), (9)
k hi ji
where, with the lattice constant set to 1 (i.e., kx , ky = n2π /L, n ∈ {0, . . ., L − 1}),
γk = 21 [cos(kx ) + cos(ky )]. (10)
141
The next step is to carry out a Bogolubov transformation to boson operators mixing the
original +k and −k operators;
αk = cosh(Θk )ak + sinh(Θk )a+
−k , (11)
which has the inverse
ak = cosh(Θk )αk − sinh(Θk )α−k
+
. (12)
It is easy to verify that the operators αk obey the standard bosonic commutation relations
for any Θk . The trick is to choose these “angles” for each k such that all operators of the
form αk α−k and αk+ α−k +
cancel out in the hamiltonian (9). This is the case if
2 cosh(Θk ) sinh(Θk )
= γk . (13)
cosh2 (Θk ) + sinh2 (Θk )
The Bogolubov-transformed hamiltonian (the spin wave hamiltonian) is then diagonal
in the occupation numbers;
H = E0 + ∑ ωk αk+ αk , (14)
k
where, after some algebra making use of Eq. (10), the constant can be written as
γ2
E0 = −2SJ ∑ qk − 2NS2 J, (15)
k 1 + 1 − γk
2
and the dispersion relation in (14) for the spin wave states αk+ |0i is given by
q
ωk = 4SJ 1 − γk2 . (16)
For momenta close to (0, 0) and (π , π ), this dispersion is linear, ωk√= ck and ωk =
c|(π , π ) − k|, respectively, with velocity (the spin wave velocity) c = 2 2SJ.
The ground state |0i of (14) is the vacuum for spin waves, where the energy is just E0
given by (15). The sum can be evaluated numerically, most easily by a straight-forward
summation over the momenta on large lattices, and extrapolating E0 /N to N → ∞ (or
converting the sum divided by N to an integral, the numerical evaluation of which gives
the N = ∞ value directly). The result is E0 /JN = −0.65795 for S = 1/2.
Note that while the ground state does not contain any Bogolubov α -bosons (spin
waves), it does contain some amount of the original a-bosons. The sublattice magne-
tization is directly related to the density of these bosons, which is uniform and can be
computed at any site or averaged over the sites;
1 N
hms i = S − h0|a+
i ai |0i = S − ∑ h0|a+i ai|0i.
N i=1
(17)
Using the Bogolubov transformation, this becomes

1
N∑
hms i = S − sinh2 (Θk ). (18)
k
142
In the most interesting case of S = 1/2, this evaluates to hmsi = 0.3034, or ≈ 61% of the
“classical” value 1/2. Thus, the quantum effects (zero-point fluctuations, represented by
the presence of of bosons) reduce, but do not destroy, the long-range order.
In principle it is not clear whether spin wave theory should be reliable for small S.
There has been much discussion of this issue, but the method does in fact give a good
description of the 2D Heisenberg model on the square lattice. As we will see later,
unbiased QMC calculations give results for E0 and hms i differing from those quoted
above by only 1 − 2%. This can be traced, a posteriori, to the true value of hms i being
quite large (i.e., spin wave theory can be expected to be accurate when the density of
bosons is low). In cases where the true sublattice magnetization is small or vanishes,
spin wave theory normally does not work that well, even when going to higher orders in
1/S (which increases the complexity of the calculation very significantly [40, 41, 42]).
2.1.2. Destruction of the Néel order

When going beyond bipartite lattices with uniform interactions, or by supplementing
the Heisenberg model with additional interactions (e.g., including more than two spins)
the quantum fluctuations can become so significant that the ground state loses its long-
range Néel order (retaining only short-range antiferromagnetic correlations). There are
several other possible types of ordered and disordered ground states, some of which have
no classical counterparts. Much of the current interest in quantum spin models is related
to the existence of non-magnetic states and quantum phase transitions between them
and the Néel state [2]. This is also the main theme of the numerical calculations to be
discussed in these lecture notes. One long-standing motivation for studying such transi-
tions stems from the cuprate high-Tc superconductors, the undoped parent compounds
of which are antiferromagnets corresponding closely to weakly coupled Heisenberg lay-
ers (many properties of which can be understood based on a single layer) [6]. In these
systems the magnetic order is destroyed upon doping by mobile charge carriers. This is
a very challenging electronic many-body problem, where computational studies are also
playing an important role [43, 44, 45]. While the full solution of the high-Tc problem
will of course require more complicated models (perhaps some variety of the t-J or Hub-
bard model), some generic aspects of the physics close to a quantum phase transition
out of the Néel state can, however, be understood based on spin-only models [46].
Apart from the cuprates and related antiferromagnetic systems, there are also many
materials with non-uniform or frustrated spin interactions [11, 12, 13], which can lead
to non-magnetic low-temperature states. Many of these states, and the possible quan-
tum phase transitions between them, are still not well understood. It is therefore useful
to search for and study prototypical quantum spin models that realize various types of
ground states and quantum phase transitions. Studies of quantum phase transitions also
have a broader context of understanding “exotic” manifestations of quantum mechanics
at the collective many-body level [2]. There are even interesting connections with par-
ticle physics—close analogies exist between phase transitions in supersymmetric gauge
theories in 2+1 dimensions and “deconfined” quantum-critical points that may separate
the 2D Néel and VBS ground states [18].
143
2.2. Spin liquids and valence-bond solids
Within linear spin wave theory, a non-magnetic state corresponds to boson density
hni i = S. This kind of state is, however, not a good representation of actual non-magnetic
ground states of Heisenberg and related quantum spin models, because it does not
contain any correlation effects. To discuss more interesting non-magnetic states, it is
useful to first look at the quantum fluctuations from a different perspective.
For two isolated spins i, j (a dimer, N = 2), the ground state of the S = 1/2 antiferro-
magnetic Heisenberg hamiltonian (1) is the singlet;
|↑i ↓ j i − |↓i ↑ j i
|φisj i = √ . (19)
2
Although the two spins in such a singlet are always perfectly anti-correlated (entangled),
the individual spins are strongly (maximally) fluctuating, and there is no static spin
order; hSi i = hS j i = 0. In contrast, the perfect Néel states for N = 2, |↑i ↓ j i and |↓i ↑ j i, are
product states (i.e., of the form |φi i ⊗ |φ j i) with no fluctuations (and no entanglement—
loosely speaking, the degree of entanglement corresponds to the deviations from a
product state). Note that for N = 2 the (eigen) energy of the singlet is −3Ji j /4, whereas
the expectation value of the energy in the Néel states is much higher, −Ji j /4 (and the
states are not eigenstates). The tendency of interacting spins to entangle by forming pair-
wise singlets to minimize the energy remains in multi-spin systems, but when N > 2 a
spin cannot simultaneously form pure singlet pairs with all its neighbors. The system
can instead be thought of as a superposition of different singlet pairings. No pair is
then in a pure singlet, and the energy contribution from each interaction Hi j is therefore
always higher than the singlet energy −3Ji j /4. This can be regarded as a reduction of
quantum fluctuations (leading to less than maximal two-spin entanglement) for N > 2,
bringing the state (or, more correctly, the density matrix) of each interacting pair closer to
a product state. A state with antiferromagnetic long-range order, breaking the rotational
invariance of the hamiltonian, can form in the thermodynamic limit if there are enough
fluctuations in the singlet pairings (and note that, from the perspective of singlets, the
Néel state has larger fluctuations than, e.g., a state of fixed singlet pairings—to be
precise, fluctuations should always be specified with respect to some reference state). If
the system is one-dimensional, or if the interactions are strongly frustrated (competing),
or if the lattice geometry and couplings favor the formation of singlets in a specific
pattern (e.g., in a system of weakly coupled dimers), antiferromagnetic order may not
be present in the ground state when N → ∞.
Valence-bond states. The above intuitive picture of a state of fluctuating singlets

can in fact be made rigorous. Any singlet state can be expanded in basis states that are
products of singlet pairs, or valence bonds. Denoting by (i, j) a singlet of spins i and j,
as in Eq. (19), a normalized valence-bond basis state for N (even) spins is of the form
|φ i = N −1/4 |(i1, j1 )(i2 , j2) · · · (iN/2 , jN/2 )i, (20)

where each site index appears exactly once (i.e., each spin belongs to one singlet).
This basis is over-complete in the singlet subspace, and, thus, any singlet state can be
144
(a) (b)
FIGURE 2. Valence-bond states in two dimensions. The thick lines represent singlets. The arrows show
a spin configuration compatible with the valence bonds (i.e., the spins on each bond are antiparallel—the
valence-bond state is a superposition of all such spin configurations). A valence-bond state with only short
bonds (or any superposition of such states), as shown here in the extreme case of all bonds of length one,
has no magnetic order in two dimensions. There can, however, be order in the valence bonds. (a) and (b)
represent typical configurations of a disordered RVB spin liquid and a columnar VBS, respectively.
expanded in these states, but the expansion coefficients are not unique. The valence-bond
basis and computational methods using it are discussed in Refs. [47, 48, 49, 50]. Here
we just note that some types of states are more naturally expressed in the valence-bond
basis than in the standard basis of individual ↑ and ↓ spins.
For the valence-bond basis to be (over-)complete, arbitrary lengths of the bonds must
be allowed. Restricting the lengths renders the basis incomplete (although one can
actually restrict the bonds to only connect sites on two different sublattices). Some types
of states are still completely dominated by short bonds (i.e., the probability of a bond
of length r decreases rapidly with r). Such short-bond states in two dimensions have
no magnetic long-range order (while in three dimensions they can be Néel ordered)
and are often called spin liquids or resonating valence-bond (RVB) states. Various
types of crystalline order can also form in the bond configurations, leading to periodic
modulations in observables such as hSi · S j i (where i, j are nearest-neighbor sites). Such
ordered states (which break lattice symmetries) are called valence-bond solids (VBSs),
or valence-bond crystals. Representative configurations of valence bonds in RVB and
VBS states are illustrated in Fig. 2.
2.3. One-dimensional systems

1D systems are rather special and deserve their own introduction. Studies of quantum
spin chains date back to Bethe’s exact solution of the S = 1/2 Heisenberg model, which
was published in 1931 [51] and worked out in greater detail some time thereafter [52].
The solution is very complicated (often requiring complex numerical calculations [53,
54]), however, and many properties of the heisenberg chain were only obtained much
later and with complementary methods (in particular, renormalization-group treatments
of effective low-energy field-theories [55, 56, 57, 58, 59]).
As we have already noted, there can be no magnetic order in a 1D Heisenberg sys-
tem (but VBS order is allowed, since it breaks a discrete symmetry). The spin correlation
function C(r) = hSi · Si+r i of the Heisenberg chain decays with the distance r as (−1)r /r
(with a multiplicative logarithmic correction, which we will discuss later). Thus, the
ground state is critical (or quasi long-range ordered, on the verge of ordering). Includ-
ing a frustrated second-nearest-neighbor interaction causes, when sufficiently strong, a
quantum phase transition into a VBS state (some times also called the spin-Pierls state),
145
order parameter
order parameter
(a) (b)
Tc [gc] T [g] Tc [gc] T [g]
FIGURE 3. Temperature (T ) or coupling (g) dependence of the order parameter (e.g., the magnetization
of a ferromagnet) at a continuous (a) and a first-order (b) phase transition. A classical, thermal transition
occurs at some temperature T = Tc , whereas a quantum phase transition occurs at some g = gc at T = 0.
where the spin correlations decay exponentially with distance [60].

While quasi-1D antiferromagnets were actively studied experimentally already in the
1960s and 70s, these efforts were further stimulated by theoretical developments in the
1980s. Haldane conjectured [55], based on a field-theory approach, that the Heisenberg
chain has completely different physical properties for integer spin (S = 1, 2, . . .) and
“half-odd integer” spin (S = 1/2, 3/2, . . .). It was known from Bethe’s solution that the
S = 1/2 chain has a gapless excitation spectrum (related to the power-law decaying
spin correlations). Haldane suggested the possibility of the S = 1 chain instead having a
ground state with exponentially decaying correlations and a gap to all excitations; a kind
of spin liquid state [26]. This was counter to the expectation (based on, e.g., spin wave
theory) that increasing S should increase the tendency to ordering. Haldane’s conjecture
stimulated intense research activities, theoretical as well as experimental, on the S = 1
Heisenberg chain and 1D systems more broadly. There is now completely conclusive
evidence from numerical studies that Haldane was right [61, 62, 63]. Experimentally,
there are also a number of quasi-one-dimensional S = 1/2 [64] and S = 1 [65] (and
also larger S [66]) compounds which show the predicted differences in the excitation
spectrum. A rather complete and compelling theory of spin-S Heisenberg chains has
emerged (and includes also the VBS transitions for half-odd integer S), but even to this
date various aspects of their unusual properties are still being worked out [67]. There are
also many other variants of spin chains, which are also attracting a lot of theoretical and
experimental attention (e.g., systems including various anisotropies, external fields [68],
higher-order interactions [69], couplings to phonons [70, 71], long-range interactions
[72, 73], etc.). In Sec. 4 we will use exact diagonalization methods to study the S = 1/2
Heisenberg chain, as well as the extended variant with frustrated interactions (and also
including long-range interactions). In Sec. 5 we will investigate longer chains using the
SSE QMC method. We will also study ladder-systems consisting of several coupled
chains [9], which, for an even number of chains, have properties similar to the Haldane
state (i.e., exponentially decaying spin correlations and gapped excitations).
2.4. Models with quantum phase transitions in two dimensions

The existence of different types of ground states implies that phase transitions can
occur in a system at T = 0 as some parameter in the hamiltonian is varied (which
146
experimentally can be achieved in quantum magnets, e.g., as a function of pressure
or external magnetic field [74]—it should be noted, however, that the possibilities are
more limited than in models). As in classical, temperature driven phase transitions,
such quantum phase transitions can be continuous or first-order, as illustrated in Fig. 3.
Normally a phase transition is associated with an order parameter, which is zero in
the disordered phase and non-zero in the ordered phase. e.g., the magnetization m of a
ferromagnet or the sublattice magnetization ms of an antiferromagnet. In Secs. 4 and 5
we will also discuss examples of order-order (Néel-VBS) transitions. Continuous phase
transitions are associated with scaling and universality, which we will discuss in more
detail in Sec. 3. In these lecture notes we will not discuss topological phase transitions,
which are not associated with any local order parameter. The different states across such
transitions are distinguishable only through some global topological quantity [76].
In this section we introduce some spin models exhibiting quantum phase transitions
in two dimensions. As we have already seen, the 2D S = 1/2 Heisenberg model with
nearest-neighbor interactions has Néel order at T = 0. The question is then how to
destroy this order and create some different kind of ground state. Specific studies based
on QMC calculations will be presented in Sec. 5. The discussion here serves as an
introduction to the kind of models and physical quantities that we will have in mind when
discussing computational techniques. Some QMC results will be shown for illustration
purposes already in this section, but at this point we do not need to worry about how they
were obtained—it suffices to say that the data are numerically exact (to within statistical
errors that are in most cases too small to be discerned in the figures).
There are also 1D analogues of most of the transitions discussed here and some of
them will be discussed in Sec. 4.3. Systems in three dimensions can exhibit transitions
similar to those in two dimensions, but most computational research is currently focused
on 1D and 2D systems, partially because 3D systems are much more challenging. The
quantum fuctations are normally also more prominent in 1D and 2D systems, and some
of the most interesting open questions related to real materials are associated with the
quasi-2D nature of the systems (although there are also interesting 3D systems [74]).
2.4.1. Dimerized systems

The perhaps simplest way to obtain a non-magnetic ground state of a 2D Heisenberg
model is to dimerize the system [75], i.e., to introduce weak and strong antiferromagnetic
couplings (bonds), J1 > 0 and J2 > J1 > 0, respectively, in a pattern such that each spin
belongs exactly to one strong bond. There are several ways to do this, three of which
are illustrated in Fig. 4. When J1 = 0, these systems consist of N/2 independent pairs
of spins (dimers) with intra-dimer coupling J2 , and, as we have already discussed, the
ground state of each such a pair is a singlet; thus the ground state of the whole system
is a singlet-product (valence-bond) state, which clearly has no magnetic order. On the
other hand, for J2 = J1 the ground state has Néel order. The question is then how the
ground state evolves as a function of the coupling ratio g = J2 /J1 . One might perhaps
think that some amount of Néel order should appear once the dimers are coupled, i.e.,
for any g < ∞. It turns out, however, that there is actually a phase transition at some
147
(a) (b) (c)
J2 J1 J1
J1 J1 J2 J1 J2
FIGURE 4. Dimerized systems with two different coupling strengths between nearest neighbors; a
bilayer (a) with the dimers across the layers and single layers with columnar (b) and staggered (c) dimers.
critical value gc (which depends on the dimer arrangement). While this transition can be
analyzed using several analytical and numerical approaches, finite-size scaling of QMC
data is currently the only way to obtain unbiased quantitative results (see, e.g., [77] for
a discussion of how spin wave theory breaks down close to the phase transition).
Fig. 5 shows some QMC results for the columnar dimer model of Fig. 4(b). The order
parameter is the staggered magnetization, with the corresponding operator defined in
Eq. (3). This is a vector operator, and, for a finite lattice, its expectation value vanishes
due to the spin-rotational symmetry of the hamiltonian. Its square, hm2s i, was computed
in the simulations. Fig. 5(a) shows results for L × L lattices versus the coupling ratio
g, and Fig. 5(b) shows data for several values of g graphed versus 1/L (which is
often the most convenient way of graphing data when examining the convergence to
a non-zero value for L → ∞). Here it is clear that the behavior changes at g ≈ 1.9;
below this coupling the sublattice magnetization extrapolates to a non-zero value in the
thermodynamic limit, whereas for larger g it decays to zero.
In the Néel state, the leading finite-size corrections to hm2s i are ∝ 1/L (a result which
can be obtained using spin wave theory [78, 79]). This can be seen quite clearly in
Fig. 5(b) for g = 1 and 1.5. The extrapolation for g = 1 (see Ref. [50], from which the
results for this coupling are taken) gives hms i = 0.3074. This is only about 1% higher
than the linear spin wave result, hmsi = 0.3034. As discussed above in Sec. 2.1.1, the
success of spin wave theory for this model can be explained a posteriori by the fact that
the quantum fluctuations are rather weak, reducing the sublattice magnetization only
by about 40% from the classical value. For larger g, the agreement between spin wave
theory and QMC results quickly becomes much worse. In this kind of dimerized model
linear spin wave theory typically over-estimates the critical g [80, 81, 77], while in other
cases, such as a system with uniform couplings with different strengths in the x and
y lattice directions (coupled chains) linear spin wave theory predicts a critical point at
non-zero coupling when, in fact, there is none [82]. Various improvements can be made
of the spin wave theory (going to higher order in 1/S, incorporating self-consistency,
etc. [6, 40, 41, 80, 77, 83]), but a system at or close to a quantum phase transition cannot
be captured correctly by these approximative methods.
An important question regarding the quantum phase transition in dimerized systems
is its universality class (a concept to be discussed further in Sec. 3). A quantum many-
body system in d dimensions can formally be mapped, using path integrals, onto an
effective classical system in 2+1 dimensions (as we will discuss in Sec. 5), and effective
continuum field-theories can be constructed for the low-energy behavior, including the
quantum phase transition. Many properties have been predicted in this way based on
148
(a) L=8 g = 1.0 (b)
0.25
L = 16 g = 1.5
0.15
L = 32 g = 1.9
0.20 g = 3.0
<ms2>
<ms2>
0.10 0.15
0.10
0.05
0.05
0.00 0.00
1.0 1.5 2.0 2.5 3.0 0 0.05 0.1 0.15 0.2 0.25
g 1/L
FIGURE 5. QMC results for the squared sublattice magnetization in the two-dimensional Heisenberg
model with columnar dimerization. (a) shows results versus the coupling ratio g for different lattice sizes
and (b) shows the size dependence for several values of g. A quantum phase transition where the Néel
order vanishes occurs at g ≈ 1.9.
renormalization-group treatments of one such field theory—the nonlinear σ -model in

2+1 dimensions [5, 84]. Based on symmetry arguments alone, one would then expect
the transition to be in the universality class of the 3D classical Heisenberg model. There
are, however, subtle issues in the quantum-classical mapping, and QMC simulations are
therefore needed to test various predictions. We will see examples of such comparisons
between results of simulations and field theories in Sec. 5. While results for the transition
in the bilayer (a) [85] and columnar dimer (b) [86] systems in Fig. 4 (and several
other cases [87, 88]) are in good agreement with the expectations, recent studies of
the staggered dimers (c) show unexpected deviations [89] that are still not understood.
2.4.2. Frustrated systems

The prototypical example of frustration is a system with antiferromagnetic inter-
actions on a triangular lattice. Looking at this problem first within the Ising model,
the spins on a single triangle cannot simultaneously be anti-parallel to both their
neighbors—there are six configurations with minimum energy, and these all have one
“frustrated” bond (two parallel neighbors), as shown in Fig. 6. Being a consequence of
the lattice, this is often referred to as geometric frustration. Upon increasing the system
size, the ground-state degeneracy grows with the system size, and in the ensemble in-
cluding all these configurations there is no order of any kind [90, 91]. In the case of the
classical XY (planar vector) or Heisenberg (vectors in three dimensions) model, there is,
however, order at T = 0 (but not at T > 0, according to the Mermin-Wagner theorem).
The energy is minimized by arranging the spins in a plane at 120◦ angle with respect
to their neighbors on the same triangle, as shown for a single triangle in Fig. 6. This is
referred to as a three-sublattice Néel state. There have been many studies of the S = 1/2
variant of this model. This was, in fact, the system for which the RVB spin-liquid state
was initially proposed [92]. There is now, however, compelling numerical evidence for
the three-sublattice Néel order actually surviving the quantum fluctuations [93, 94].
149
FIGURE 6. Ising (left) and planar (right) spin configurations with minimum energy on a triangle with
antiferromagnetic nearest-neighbor interactions.
(a) (b) (c)
FIGURE 7. The square-lattice S = 1/2 Heisenberg model with only nearest-neighbor interactions
J1 > 0 has an antiferromagnetic ground state, in (a) illustrated by open and solid circles for hSiz i > 0
and hSzj i < 0. In (b), next-nearest-neighbor interactions J2 > 0 are shown as dashed lines. When 0.4 <
J2 /J1 < −0.6 (approximately) the ground state may be a columnar VBS, with hSiz i = 0 on all sites but
modulations in the bond correlations hSi ·S j i (where i, j are nearest-neighbors) as shown with thicker lines
for the more strongly correlated bonds forming a columnar pattern. For J2 /J1 > 0.6, the ground state has
collinear (striped) magnetic order, as shown in (c).
On the square lattice, frustration effects can be investigated by adding interactions

beyond nearest neighbors. A well studied case is the J1 -J2 model, where J1 and J2 refer,
respectively, to the strengths of the nearest- and next-nearest-neighbor interactions. The
system is frustrated if both J1 > 0, J2 > 0 or if J1 < 0, J2 > 0. Even with Ising spins,
this is a highly non-trivial system, with unresolved questions still attracting interest
[95, 96, 97]. We will discuss the frustrated Ising system further in Sec. 3.4.
In the case of quantum spins, the S = 1/2 J1 -J2 Heisenberg model with all antiferro-
magnetic couplings is one of the prototypical models with a quantum phase transition
out of the standard Néel state, in this case as a function of the coupling ratio g = J2 /J1 .
While many different calculations show rather consistently that the Néel order vanishes
at a critical coupling ratio gc ≈ 0.4 [98, 99, 100, 101, 102, 103, 104], the order of the
phase transition and the nature of the non-magnetic ground state are still controversial
issues. Most studies indicate some type of VBS state, with a columnar one being the
prime candidate, but an RVB spin liquid state has also been proposed [105].
For larger g, above g ≈ 0.6, there is again magnetic order. This can be understood in
the limit g → ∞, where the system decouples into two separate square-lattice Heisenberg
antiferromagnets. At g = ∞ (J1 = 0) the relative direction in spin space of the antiferro-
magnetic order within these subsystems is arbitrary, but for any J1 > 0 the subsystems
lock to each other and form collinear spin order, with vertical or horizontal stripes of
parallel spins (a state breaking the 90◦ lattice rotation symmetry, in addition to the global
spin rotation symmetry). The transition between the non-magnetic and collinear states
is most likely first-order.
The three different ground states of the J1 -J2 Heisenberg model with S = 1/2 spins are
illustrated in Fig. 7. Note that a classical version of this model (Ising, XY, or Heisenberg)
150
has a direct (first-order) Néel-collinear T = 0 transition exactly at g = 1/2 (which can be
easily verified just by computing the energies of those states for g < 1/2 and g > 1/2).
The magnetically disordered state is thus induced by quantum fluctuations, and it has no
direct classical analogue. It is not clear whether this kind of state persists for S = 1 or
higher spins—for some large S, perhaps already S = 1, it must give way to a first-order
Néel-Collinear transition, as in the classical system.
The reason why it has been so difficult to reach a firm conclusion on the nature of the
non-magnetic state and the quantum phase transition between it and the Néel state is that
large-scale unbiased computational studies of the S = 1/2 J1 -J2 model are currently not
possible, because of “sign problems” affecting QMC calculations of frustrated systems.
We will discuss this issue further in Sec. 5. There are no other unbiased method that can
reach sufficiently large lattices, e.g., exact diagonalization can reach only N ≈ 40 spins.2
2.4.3. The J-Q class of models

VBS states of quantum spin systems in two dimensions were predicted theoretically
more than two decades ago [107]. The VBS formation is associated with a spontaneously
broken translational lattice symmetry. The VBS state and quantum phase transitions
into it are therefore quite different from the non-magnetic states and transitions in the
“manually” dimerized systems discussed above in Sec. 2.4.1. Although there is a pattern
of “strong” and “weak” spin correlations in both cases, the quantum fluctuations in
systems with manually and spontaneously broken translational symmetry are different
(being much more interesting in VBS states). Until recently, large-scale computational
studies of VBS states and Néel-VBS quantum phase transitions had not been carried out
starting from microscopic hamiltonians, because of the QMC sign problems affecting
frustrated Heisenberg models (which at first sight are the most natural systems in which
to explore the physics of non-magnetic states).
VBS states and the Néel-VBS transition have come into renewed focus with a pro-
posal by Senthil et al. [16] that this transition is generically continuous and, thus, violates
the “Landau rule”, according to which an order-order transition (between states breaking
unrelated symmetries) should be first-order (except at fine-tuned multi-critical points).
In the “deconfined” quantum-criticality scenario, the VBS and Néel order parameters
are manifestations of spinon confinement and condensation, respectively. Spinons can
be thought of as S = 1/2 degrees of freedom, but not just corresponding to the bare
individual spins on the lattice sites, but more complex collective objects “dressed” by
interactions. A non-compact CP1 field theory was proposed to describe such spinons
coupled to an emergent gauge field (which corresponds to the properties of the valence-
bond background in which the spinons exist) [16]. A central question is then whether
this low-energy physics of a continuum field theory really can arise starting from a rea-
2 Recently developed variational methods based on tensor-product states in principle become unbiased
in the limit of large tensors [21]. They have been applied to various frustrated spin models [106]. The
computational complexity of such calculations is, however, currently too demanding to reach tensors
sufficiently large to produce unbiased results in practice.
151
j j l j l n
Sij SijSkl SijSklSmn
i i k i k m
FIGURE 8. Graphical representation of possible arrangements of products of singlet-projector oper-

ators Si j in the J-Q model and its generalizations. (a) is the Heisenberg exchange, (b) the four-spin
interaction of the original J-Q model, and (c) a six-spin interaction which leads to more robust VBS order.
These operators, and their 90◦-rotated analogues, are summed over all positions on the square lattice.
sonable microscopic hamiltonian. Answering this question requires large-scale compu-

tational studies of models exhibiting Néel-VBS transitions.
Since the QMC sign problem prohibits large-scale studies of the J1 -J2 Heisenberg
model and other similar frustrated systems, we have to try something else. In the “J-
Q” class of models [17, 108, 109], the Néel order is destroyed by an interaction (Q)
which is not frustrated, in the standard sense, but still competes with the Heisenberg (J)
interaction. To understand these J-Q models, note first that the Heisenberg interaction is,
up to a constant, equal to a singlet projector operator: Hi j = −Si j + 41 , where
Si j = 41 − Si · S j . (21)
The pair-singlet, Eq. (19), is an eigenstate of this operator with eigenvalue 1, whereas a
triplet state is destroyed by it;
Si j |φisj i = |φisj i, Si j |φit,m
j i = 0, (m = 0, ±1). (22)
Thus, when Si j acts on a singlet-triplet superposition, only the singlet component sur-
vives (is “projected out”—note that the property Si2j = Si j required of a projection op-
erator is satisfied). The standard Heisenberg interaction thus favors the formation of
singlets on pairs of nearest-neighbor sites, but, as we discussed in Sec. 2.2, the fluctua-
tions of these singlets among many different pairings of the spins leads to Néel order in
the ground state. The idea behind the J-Q models is to project singlets on two or more
bonds in a correlated fashion, using products of several Si j operators on a suitable set of
different bonds. This favors a higher density of short valence bonds, thereby reducing or
completely destroying the antiferromagnetic order.
The original J-Q hamiltonian [17] on the square lattice can be written as
H = −J ∑ Si j − Q ∑ Si j Skl , (23)
hi ji hi jkli
where both the J and Q terms are illustrated in Fig. 8. The Q interaction involves four
spins on a 2 × 2 plaquette. An interaction with three singlet projectors in a columnar
arrangement is also shown, and operators with even more projectors, or with the projec-
tors arranged on the lattice in different (non-columnar) patterns, can also be considered
[109]. With J > 0 and Q > 0 [and the minus signs in front of the interactions in Eq. (23)],
correlated singlets are favored on the lattice units formed by the product of singlet pro-
jectors. It is still not clear just from the hamiltonian whether a VBS state is realized for
152
(a) (b) L = 16
0.015 0.010 L = 32
L = 16 L = 64
D*xx(L/2,L/2)
L = 32
C(L/2,L/2)
0.010 L = 64
0.005
0.005
0.000 0.000
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1
J/Q J/Q
FIGURE 9. QMC results for the long-distance spin (a) and staggered dimer (b) correlation functions
(corresponding the Néel and VBS order, respectively) in the ground state of the J-Q model (23) versus the
coupling ratio J/Q. A careful analysis of these correlation functions, as well as other quantities, indicates
a single critical point (J/Q)c ≈ 0.045 where both the Néel order and VBS order vanish continuously.
large Q/J—since the hamiltonian does not break any symmetries (the interactions in
Fig. 8 are summed over all distinct lattice translations and rotations), a VBS state only
forms if the hamiltonian also contains in it, implicitly, some effective interactions that
favor singlets in some ordered pattern. The J-Q model (23) does exhibit a Néel-VBS
transition, at Qc /J ≈ 22 [17, 108, 110], and at lower Qc if more than two singlet projec-
tors are used [108, 109]. We will discuss this in more detail in Sec. 5. Here we just look
briefly at examples of correlation functions useful for characterizing the Néel and VBS
states and the quantum phase transition between them.
Order parameters. As an alternative to the square of the full spatially averaged

sublattice magnetization (3), we can detect the presence or absence of of Néel order
using the spin correlation function,
C(ri j ) = hSi · S j i, (24)
at long-distances. Fig. 9(a) shows J-Q model results for the largest separation of the
spins, rmax = (L/2, L/2), on periodic L × L lattices. The results have to be analyzed
carefully to determine the transition point, but already this raw data suggest that the
Néel order vanishes, i.e., C(rmax ) → 0 when L → ∞, for J/Q < 0.04.
VBS order can be detected in the dimer (or bond) correlation function, defined as
Dxx (ri j ) = hBx (ri )Bx (r j )i, (25)
where the bond operator is given by
Bx (ri ) = S(ri ) · S(ri + x̂). (26)
Here, instead of using a subscript i on the spin operator for some arbitrary site labeling,
it is more convenient to use the corresponding lattice position vector ri . Then ri + x̂
corresponds to the site immediately next to ri in the positive x direction. The subscripts
153
xx in (25) indicate that the two bond operators Bx are both oriented in the x direction, and
this correlation function is, by symmetry, equal to Dyy on an L × L lattice. One can also
consider bond correlations Dxy , where the two bond operators are oriented differently. In
a VBS state such as the one illustrated in Fig. 7, one expects Dxx (r) to exhibit a columnar
pattern of smaller and larger values. A VBS order parameter can then be defined as a
suitable difference between these modulated correlations, e.g.,
D∗xx (r) = Dxx (r) − 12 [Dxx (r − x̂) + Dxx (r + x̂)]. (27)
This correlation function is shown in Fig. 9(b) at the largest lattice separation for several
different system sizes. Here it is clear that VBS order exists for J = 0, up to J/Q ≈ 0.04,
roughly where the Néel order sets in. As we will discuss in Sec. 5, all calculations
so far point to a single critical point, without any intervening third phase or region of
coexistence of both Néel and VBS order.
3. CLASSICAL PHASE TRANSITIONS, MONTE CARLO

SIMULATIONS, AND FINITE-SIZE SCALING
Many aspects of quantum phase transitions, and how to analyze them based on numerical
finite-lattice data, are very similar to classical phase transitions. Here we discuss this
common formalism in the simpler context of classical phase transitions, before turning
to calculations and data analysis for quantum systems.
In classical statistical mechanics the prototypical example of a system with a continu-
ous phase transition is the 2D Ising ferromagnet, the exact solution [111, 112] of which
shows rigorously that such a phase transition exists. This model is defined by
N
Eσ = −J ∑ σi σ j − h ∑ σi , (28)
hi ji i=1
which is just the (potential) energy as a function of the spins σi = ±1. We use σ to
collectively denote an entire spin configuration; σ = (σ1 , . . ., σN ). With the interaction
bonds hi ji restricted to nearest-neighbors on the simple 2D √ square lattice, the exact
critical temperature for an infinite system is Tc /J = 2/ ln(1 + 2) ≈ 2.269. The order
parameter is the magnetization hmi = hσi i. It has the asymptotic T → Tc (T < Tc )
form m ∝ |t|β , where t is the reduced temperature, t = (T − Tc )/Tc, and the exponent
β = 1/8. Such critical exponents and other aspects of scaling behavior at continuous
phase transitions will be discussed in this section.
The exact solution of the Ising model is very special, and normally one studies phase
transitions in other ways. Critical exponents appear already in simple mean-field theo-
ries, but their values are typically not correct. Mean-field theory is nevertheless essential
as a starting point, which we here outline for the Ising model. The most important theo-
retical framework for phase transitions is the renormalization group, which explains how
universal (depending only on symmetries and dimensionality, not details of the interac-
tions) non-trivial exponents (i.e., different from the generic mean-field values) can arise
(see, e.g., the book by Cardy [15]). To compute critical exponents and other properties in
154
an unbiased way, one normally uses Monte Carlo simulations, where spin configurations
are stochastically sampled according to the Boltzmann distribution.3 In this section the
basics of Monte Carlo simulations are outlined (for more details, see, e.g., the books by
Landau and Binder [115] and Newman and Barkema [116]). To motivate and illustrate
finite-lattice calculations in general, we will also discuss how properties in the thermo-
dynamic limit can be extracted from Monte Carlo data.
Section outline. The standard mean-field treatment of the Ising model is discussed
in Sec. 3.1. In Sec. 3.2 the Monte Carlo method is introduced and used to illustrate
how phase transitions and symmetry-breaking can occur in practice for large systems
(formally in the limit of infinite system size). In Sec. 3.3 we review the key aspects
of critical behavior (defining exponents, etc.) in the thermodynamic limit, then discuss
the finite-size scaling hypothesis and demonstrate its usage by analyzing Monte Carlo
results for the 2D Ising model. First-order transitions are discussed in Sec. 3.4, using
a frustrated 2D Ising model as an example of finite-size scaling methods for detecting
discontinuities. In Sec. 3.5 the important concept of spin stiffness (which corresponds
to an elastic modulus of a solid) is introduced in the context of XY (planar vector)
spin models. The scaling properties of the spin stiffness are illustrated with Monte Carlo
results for the 3D and 2D XY models, the latter of which does not exhibit a normal phase
transition into an ordered state, but exhibits a different Kosterlitz-Thouless transition
into a critical low-temperature phase. In Sec. 3.6 we briefly discuss how the classical
criticality concepts are generalized to quantum phase transitions.
3.1. Mean-field theory of the Ising model

In mean-field theories the environment of a subsystem of an infinite lattice is replaced
by an external field representing the average interactions between the subsystem and the
environment. The subsystem can be a single spin or a cluster of several spins. Here we
will just consider the simplest case of a single-spin calculation for the Ising model, i.e.,
the infinite environment of a single spin is replaced by an effective field.
It is convenient to write the hamiltonian (28) in an extended form with general
interactions Ji j between all the spins (on an arbitrary lattice), and also include an external
magnetic field;
1 N N N
Eσ = − ∑ ∑ Ji j σi σ j − h ∑ σi . (29)
2 i=1 j=1 i=1
Note the factor 1/2, which compensates for each interacting pair being included twice
in the sum (and Jii = 0). We do not impose any restrictions of the signs and magnitudes
of Ji j , but for simplicity we assume that if there is a phase transition, the ordered state is
ferromagnetic. To construct and motivate the single-spin approximation, we first group
3 Other important methods include high-temperature series expansions [113] and ε -expansions (where
ε = du − d) around systems at their upper critical dimension du [114], where the exponents take their
mean-field values. Monte Carlo simulation is normally the most reliable method, however.
155
together all the interactions in (29) involving an arbitrary spin i;
!
Ei = −σi ∑ Ji j σ j + h . (30)
j
Note that there is no factor 12 in front of the sum here, and we have used Ji j = J ji .
We now assume that h > 0, so that m = hσi i > 0. We will investigate spontaneous
ordering in the absence of the field, by eventually letting h → 0. Adding and subtracting
a constant, we can write the terms within the parentheses in (30) as
∑ Ji j σ j + h = m ∑ Ji j + h + ∑ Ji j (σ j − m). (31)
j j j
In its most basic formulation, mean-field theory amounts to neglecting the second sum
in (31)—the fluctuation term—after which we are left with the easily solvable problem
of a single spin in an effective magnetic field of strength Js m + h;
Eσ = −(Js m + h)σ , (32)
where Js is the sum of the original couplings
Js = ∑ Ji j , (33)
j
and we assume a translationally invariant system, so that this sum is independent of i.

The magnetization m in (32) is at this stage unknown and will be determined through a
self-consistency condition; hσ i = m.
Mean-field theory in the present formulation can be justified if the neglected fluctua-
tions are much smaller, on average, than the other terms in (31), i.e., if
∑ j Ji j (σ j − m)

δm = ≪ 1. (34)
Js m + h
Here hi denotes the expectation value under the actual probability distribution (the
Boltzmann distribution) of the spins, which we cannot compute exactly. In principle the
fluctuations can also be calculated within mean-field theory, as an internal consistency
check. Even without doing any calculations, we can roughly deduce the conditions
under which δm will be small. Clearly, δm is small if there is substantial order, i.e.,
when (1 − m) = (1 − hσ j i) ≪ 1, because then most σ j = 1 ≈ m. This is the case if
h is large. It is also true for h = 0 if the symmetry is spontaneously broken and m is
close to 1, i.e., for T ≪ Tc if there is a phase transition. Moreover, δm can be small
even if m is not very large, if the sum over j involves many non-zero (and relatively
large) coupling constants Ji j —because of cancellations of fluctuations, the sum ∑ j Ji j σ j
will then typically (in most of the statistically important spin configurations) be close
to m ∑ j Ji j . In the extreme case of infinite-range uniform interactions, Ji j = J/N for
all i, j, and N → ∞ (where we regard J as a finite constant, e.g., J = 1, in order to
have a finite energy) all the fluctuations cancel out exactly and δm = 0 (and the mean-
field theory is then exact). This holds true exactly also for short-range interactions on
156
an infinite-dimensional lattice. Thus, in general, even if m is not large, the fluctuation
measure δm can be expected to be small for systems in high dimensions and/or for long-
range interactions. These are then the conditions under which mean-field theory can
be expected to be quantitatively accurate. Even in cases where it is not quantitatively
accurate, mean-field theory can still provide valuable insights qualitatively.
Let us now actually solve the mean-field problem (32), i.e., the single-spin problem
(32) under the self-consistency condition hσ i = m. The magnetization is
∑σ σ eσ (Js m+h)/T
hσ i = = tanh[(Js m + h)/T ], (35)
∑σ eσ (Js m+h)/T
and thus the self-consistency condition reads
m = tanh[(Jsm + h)/T ]. (36)
This equation in general has to be solved numerically, which can be done easily using
successive bracketing of the solution. For small m and h, we can proceed analytically
by expanding to leading order in m, h. First, when the external field h = 0, m = 0 is a
solution for all T . Looking for other possible solutions, expanding (36) to third order in
x = Js m + h, tanh(x) = x − x3 /3, we have
T 2 Js − T
m2 = 3 , (37)
Js2 Js
from which we can identify the critical temperature Tc = Js below which the magnetiza-
tion can be non-zero. The asymptotic T → Tc behavior is
1/2
Tc − T
m= 3 , (h = 0, T < Tc ). (38)
Tc
We can also obtain the h → 0 field dependence of the magnetization at Tc from (36).
Keeping the leading-order terms in h of the third-order expansion at T = Js we get
1/3
h
m= 3 , (T = Tc ). (39)
Js
It is instructive to also look at the full numerical solution for m. Fig. 10(a) shows some
examples of the field dependence of m at different temperatures. For T > Tc the behavior
is analytic across h = 0, with a singularity described by (39) developing as T → Tc .
Below Tc the behavior is discontinuous, which corresponds to a first-order transition
versus h between the m < 0 and m > 0 states. The discontinuity corresponds to a
spontaneous magnetization at zero field, the full numerical solution of which is graphed
in Fig. 10(b) and the asymptotic T → Tc− behavior of which is given by (38). Here the
role of h → 0+ or h → 0− in the calculation is to break the degeneracy between the ±m
solutions. The degeneracy for h = 0 also corresponds to co-existence of the two ordered
states exactly at the first-order transition, as we will discuss more generally in Sec. 3.4.
157
1.0
T/Tc=0.8 1.0
h=0
T/Tc=1.0
0.5 0.8
T/Tc=2.0
0.6
m
0.0
0.4
-0.5
(a) 0.2 (b)
-1.0
0.0
-0.4 -0.2 0.0 0.2 0.4 0.0 0.2 0.4 0.6 0.8 1.0
h/Js T/Js
FIGURE 10. Mean-field solution of the Ising model. (a) Magnetization versus external field for temper-
atures above, at, and below Tc . The discontinuity for h → 0+ and h → 0− corresponds to the spontaneous
magnetization at h = 0. (b) Temperature dependence of the spontaneous magnetization.
Translating the result Tc = Js with Js given in Eq. (33) gives Tc = 4J for the 2D
Ising model with nearest-neighbor interactions of strength J. This is well above the
correct value, Tc /J ≈ 2.269. An over-estimation of Tc can be expected on account of the
neglected fluctuations (which naturally lower Tc ).
The exponents β = 1/2 and δ = 1/3 in (38) and (39) are generically the values of
these critical exponents within mean-field theories. Other exponents can also be com-
puted [15]. While power laws are indeed correct generic features of continuous phase
transitions, the mean-field values are not correct in general. As already mentioned, for
the 2D Ising model, β = 1/8 from Onsager’s exact solution. In three dimensions its value
in the Ising universality class is β ≈ 0.33, as determined using Monte Carlo simulations
and series expansion methods (and also, less precisely, using field-theoretical methods
and other analytical approaches). The mean-field critical exponents are exact in four and
higher dimensions (four being the upper critical dimension for the Ising model—the
dimensionality above which the mean-field critical exponents become exact). We will
discuss critical behavior in greater detail further below, after developing the technical
aspects of Monte Carlo simulations to study criticality in practice.
3.2. Monte Carlo simulations of the Ising model

In a Monte Carlo simulation, the goal is to generate a sequence of spin configurations,
σ (1), σ (2), . . ., σ (K), representing a statistically unbiased sample from the Boltzmann
distribution, i.e., the probability P(σ ) of an arbitrary configuration σ to be among the
sampled ones should be proportional to the Boltzmann weight at temperature T ;
Wσ = e−Eσ /T , (40)
where we work in units such that kB = 1. The actual (properly normalized) Boltzmann
probability is Pσ = Wσ /Z, where Z is the partition function
Z = ∑ e−Eσ /T . (41)
σ
158
In Monte Carlo simulations we do not need the full partition function, only the un-
normalized weights (40). It is also important to note that the sequence σ (1), . . ., σ (K)
can be correlated (as we will discuss further below), but we only need the probability
P(σ ) ∝ Wσ and for now do not need to worry abot joint probabilities such as P(σ1 , σ2 ).
The Metropolis algorithm. The simplest way to generate a valid sequence of con-
figurations is with the Metropolis algorithm [117]. This kind of simulation starts from
an arbitrary spin configuration σ (1) (e.g., randomly generated). Thereafter, each suc-
cessive σ (k + 1) is obtained stochastically from its predecessor σ (k) according to a few
simple steps based on flipping randomly selected spins with a probability related to the
desired distribution. We only need to store the current configuration and from now on
suppress the “time” index k. We denote by σ −i the configuration obtained when the ith
spin of σ has been flipped; σ −i = (σ1 , . . . , −σi , . . ., σN ). Normally one defines a Monte
Carlo sweep as N such flip attempts, so that ∝ N spins are flipped, on average, during
this size-normalized unit of simulation time. A Monte Carlo sweep can be carried out
according to the following simple algorithm:4
do j=1,N
i = random[1, . . ., N] {1}
if (random[0 − 1]) < Wσ −i /Wσ ) σi = −σi
enddo
This Metropolis algorithm is based on the detailed balance principle, which is a general
theorem for a stochastic process (a Markov chain in some arbitrary configuration space)
that should generate a probability distribution W . By this we mean, roughly speaking,
that the set of sampled configurations should approach the distribution W as the number
of configurations is increased, independently of the initial condition. Denoting by P(A →
B) the transition probability of “moving” to configuration B if the current one is A,
the detailed balance principle states that the desired distribution is generated if all the
transition probabilities satisfy the condition
P(A → B) W (B)
= , (42)
P(B → A) W (A)
for all pairs of configurations A, B for which P(A → B) > 0. In addition, the sampling
should be ergodic, i.e., any configuration C with non-zero weight W (C) must be reach-
able, in principle, with non-zero probability through a series of moves starting from an
arbitrary configuration.
4 This pesudocode segment resembles a computer language such as Fortran, but with mathematical nota-
tion making it easier to read. Such pseudocodes will be used throughout these notes to illustrate imple-
mentations of algorithms (showing the essential steps, but without trivial details that may in practice make
real working code somewhat longer). Numbers in curly brackets, here {1}, will be used to label codes.
In most cases the syntax will be rather self-explanatory and does not require any detailed discussions of
definitions. In the segment above, and in many other ones to follow, random denotes a random-number
generator, with the square brackets [1, . . . , N] and [0 − 1] indicating the range of uniformly distributed
integers and floating-point numbers, respectively.
159
In the Metropolis algorithm, as implemented in code {1}, the transition probability
consists of two factors;
P(σ → σ −i ) = Pselect (i)Paccept (σ −i ), (43)
where Pselect (i) is the probability of randomly selecting spin i, which here always equals
1/N. The full transition probability in (42) can therefore be replaced by the probability
Paccept of actually carrying out (accepting) the spin flip. This probability is not unique.
In the Metropolis algorithm it is taken as

Wσ −i
Paccept (σ ) = min
−i
,1 . (44)
Wσ
One can easily confirm that this satisfies the detailed balance condition (42). A proba-
bility formally has to be ≤ 1, and this is taken care of above with the “minimum of”
function. In code {1}, comparing the weight ratio with a random number in the range
[0, 1) automatically achieves the same result. Note that the ratio of the Boltzmann fac-
tors depends only on the spins interacting with the flip-candidate σi (in the simplest case
just its nearest neighbors) and can be rapidly evaluated. In more physical terms, a spin
flip leading to a lower energy is always accepted, whereas a configuration with higher
energy is accepted only with probability Paccept (σ −i ) = exp[(Eσ − Eσ −i )/T ].
The Metropolis algorithm leads to the correct distribution after some transient time,
which depends on the initial configuration, the temperature, and the system size. In a
simulation one should therefore discard some number of configurations before “measur-
ing” physical observables. After this equilibration, measurements are normally carried
out after every or every few sweeps.
In the case of the ferromagnetic Ising model, the most important quantity to calculate
is the magnetization. It should be computed using all the spins to take advantage of
self-averaging to improve the statistics. Thus, we will normally use
1 N
m= ∑ σi.
N i=1
(45)
We will here consider only simulations with the external field h = 0. We have already
discussed that fact that relevant symmetries of the hamiltonian (here the discrete spin-
inversion symmetry) are not broken in simulations of finite systems. For the Ising model,
one should therefore compute spin-inversion invariant expectation values such as hm2 i
or h|m|i in order to detect the phase transition into the ferromagnetic state.
Symmetry-breaking and finite systems. To obtain a qualitative understanding of how

symmetry-breaking in the thermodynamic limit is manifested in practice for large lat-
tices, it is useful to first look at an actual Monte Carlo time series for m. Examples for
two small systems at a temperature below Tc are shown in Fig. 11. Here one can see that
the magnetization fluctuates between positive and negative values, and that the typical
time taken to reverse the sign of m is longer for the larger system. Plotting the series over
a longer time makes this clearer, but note that the time the system spends close to m = 0
160
1.0
0.5
L=8
m
0.0
-0.5
-1.0
1.0
0.5
L=16
m
0.0
-0.5
-1.0
0 5000 10000
Monte Carlo sweep #
FIGURE 11. Time series for the magnetization generated in Monte Carlo simulations of the Ising model
on L × L lattices with L = 8 (top) and L = 16 (bottom) at temperature T /J = 2.2 (< Tc ≈ 2.269). In both
cases the starting configuration was fully polarized, m = 1, and the subsequent points are separated by
N = L2 Metropolis spin flip attempts (constituting one Monte Carlo sweep).
is visibly much smaller for L = 16 than L = 8, and this is of course directly related to a
typically longer reversal time. It is clear from results such as these that the distribution
of m values is peaked at non-zero positive and negative values for T < Tc . For T > Tc the
distribution is instead a single peak centered at m = 0. Examples of such distributions
are shown in Fig. 12. In thermodynamics language, these two qualitatively different dis-
tributions can be understood as a consequence of the free energy F(m) = E(m) − T S(m)
at low T being dominated by the internal energy E (which is low in large-|m| configura-
tions), and at high T by the entropy S (which is large for small |m|).
A double-peaked magnetization distribution at low temperature suggests that even a
finite system can be considered as ordered, although the spin-inversion symmetry is not
broken in the simulations. It appears very plausible that the typical reversal time should
diverge with L (and it is not difficult to check this with simulation data for several L
values and a suitably definition of a reversal), and then no reversals would take place for
large L, even during very long simulations. The reason for this divergent time scale is
that, in order for the magnetization to reverse, a series of local spin flips must necessarily
take the system through many configurations with m ≈ 0, which have increasingly
high energy for increasing system size (and, thus, lower Boltzmann probability). For
a large system the distribution for T < Tc is therefore only sampled among the subset
of configurations with fixed sign of m—the stochastic process in practice becomes non-
ergodic. Broken symmetry and non-ergodic sampling are manifested strictly only for
N = ∞, but in practice also for large but finite systems on time scales less than the
typical magnetization reversal time. This time scale of course depends on the details of
how the spin configurations are thermally sampled—in Monte Carlo simulations and in
a real magnets—but diverges as N → ∞ for any local sampling scheme.
Normally, in Monte Carlo simulations one does not investigate the time series and
the full distribution of physical quantities (although some times this is useful). By
computing hm2 i or h|m|i, we do not have to worry about the time scale of reversals.
We normally want to extrapolate finite-N results to the thermodynamic limit, where
161
2.5 5.0
L=16 L=64
2.0 T/J=2.20 4.0
T/J=2.60
P(m)
P(m)
1.5 3.0
1.0 2.0
0.5 1.0
0.0 0.0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
m m
FIGURE 12. Magnetization distributions for L × L Ising models with L = 16 (left) and 64 (right) at two
temperatures below (T /J = 2.2) and above (T /J = 2.6) the critical temperature Tc /J ≈ 2.27.
|hmi| = h|m|i = hm2 i1/2 (but note that hm2 i1/2 6= h|m|i for finite N, because of the finite
width of the peaks in the m-distribution).
Autocorrelations and statistical errors. Before calculating expectation values, we

have to discuss how to analyze Monte Carlo data statistically. Consecutive configura-
tions generated with the Metropolis algorithm are not statistically independent—only
configurations separated by a number of sweeps much larger than the autocorrelation
time are statistically independent. The autocorrelation function for a quantity Q is de-
fined as
hQ(i + t)Q(i)i − hQi2
AQ (t) = , (46)
hQ2 i − hQi2
where t and i denote simulation time, normally in units of the Monte Carlo sweeps
defined above, and averages are over the reference time i. The normalization is such that
AQ (0) = 1 and AQ (t → ∞) = 0. The asymptotic decay is exponential, AQ (t) ∼ e−t/τQ ,
which can be used to the define the autocorrelation time τQ . Normally one instead uses
the integrated autocorrelation time, which also contains contributions from the often
dominant non-asymptotic behavior;
1 ∞
τQint = + ∑ AQ (t). (47)
2 t=1
Here we will not discuss autocorrelations at length, but only summarize their underlying
role in determining the statistical precision (“error bars”) of computed quantities.
The autocorrelation time for m of the Ising model roughly corresponds to the typical
time between magnetization reversals (as in Fig. 11). Other quantities, such as m2 and
|m|, that are not sensitive to the sign of m, have shorter autocorrelation times (however,
often much longer than a single Monte Carlo sweep). A long autocorrelation time does
not bias a computed average (i.e., it is not wrong to measure Q after every Monte Carlo
sweep even if τQ ≫ 1), provided that the total simulation time is much longer than τQ .
The autocorrelation time does, however, come into play (explicitly or implicitly) when
computing statistical errors.
To calculate the statistical errors, one can subdivide a simulation into a number B
of bins, each containing some number M of Monte Carlo sweeps. For some quantity
162
Q, averages Q̄b , b = 1, . . ., B are computed over each bin, and the final average Q̄ and
error bar σQ (one standard deviation of the average of the bin averages) are calculated
according to
1 B 1 B
Q̄ = ∑ Q̄b , σQ2 = ∑ (Q̄b − Q̄)2 . (48)
B b=1 B(B − 1) b=1
The final estimate of the true expectation value hQi should then be quoted as Q̄ ± σQ .
The reason for binning the data is that, according to the central limit theorem, the
distribution of bin averages is Gaussian for large M (unlike the distribution of individual
measurements, as seen clearly for the magnetization in Fig. 12), and the computed error
bar then has a well defined unique meaning (e.g., we know that the probability of hQi
being within one error bar of Q̄ is about 2/3). This is only true if the bin length M is also
much longer than the autocorrelation time, so that the bin averages can be regarded as
statistically independent. If that
√ is the case, the error bar should depend only√on the total
number of sweeps; σQ ∝ 1/ MB, where the factor of proportionality is ∝ τQ (and of
course also depends on the detailed form of the distribution of the individual measured
Q values). It is not necessary to calculate τQ explicitly. If there is any doubt about the
bins being sufficiently long, one can check this by using a rather large number of bins
(e.g., in the range 100-1000) and saving all the bin averages on disk. The data can then
be re-binned into longer bins post-simulation, and the convergence of σQ as a function
of the bin length can be tested. Saving the bin averages on disk is always advisable, not
only for the purpose of analyzing the error bars, but also in order to make it easy to add
more data at some later time (to improve the results, if needed).
An important point that has to be mentioned is that the autocorrelation time of any
local updating scheme, e.g., the Metropolis algorithm, diverges for T → Tc and N → ∞.
We previously saw an example of this in the magnetization reversal time of the Ising
model, but divergent autocorrelation times also affect any quantity (in particular m2 and
|m|) that is sensitive to the fluctuations of the magnitude of m (which in practice is the
case for most quantities of interest). It is therefore difficult to obtain good results for large
systems close to a critical point. In many cases, including the Ising model, this problem
can be solved (or almost solved) by using cluster algorithms [118, 119], where clusters
of spins (constructed so as to satisfy detailed balance) are flipped collectively (instead of
flipping individual spins one-by-one). While we will consider an analogous loop-cluster
algorithm for QMC simulations in Sec. 5, we do not have to discuss classical cluster
algorithms here; they have been described extensively in the literature (e.g., in [115]
and [116]). It should be noted, however, that the Ising results to be discussed below, for
systems with up to 10242 spins, were in fact obtained using a cluster algorithm. It would
be very difficult to generate (within reasonable time) data of the same quality using the
Metropolis scheme for such large systems in the critical region.
3.3. Finite-size scaling and critical exponents

Fig. 13(a) shows the temperature dependence of the squared magnetization hm2 i of
the 2D Ising model for several lattice sizes. An increasingly sharp feature develops with
increasing L in the neighborhood of the known Tc , and it appears very plausible that
163
1.0 0
L=8 10
L = 16, 32, 64
0.8 L = 128
(a) 10
-1 (b)
0.6
<m >
<m >
2
2
0.4 T/J = 2.25
-2
10 T = Tc
0.2 T/J = 2.30
-3
0.0 10
0 1 2 3 4 5 10 100 1000
T/J L
FIGURE 13. (a) The squared magnetization as a function of temperature for several L×L Ising systems.
(b) The finite-size dependence at, below, and above Tc . At Tc the behavior is a power law, m2 ∼ L−1/4 ,
which on this log-log scale corresponds to a line with slope −1/4. For T > Tc the decay to zero is of the
form L−2 , and this is also the rate of convergence for T < Tc .
hm2 i vanishes for N → ∞ for T > Tc and remains non-zero for T < Tc . In this plot, error
bars are not visible because they are much smaller than the line widths. The data were
generated on a very fine T -grid to produce continuous-looking curves.
The size dependence of hm2 i at three different temperatures is shown in Fig. 13(b).
For T < Tc it converges with L to a non-zero value, while for T > Tc it decays to 0 as L−2 .
Exactly at Tc the decay follows a non-trivial power-law; hm2 i ∝ L−2β /ν , where β = 1/8
is the same exponent as in the N = ∞ magnetization for T < Tc , and ν (the correlation-
length exponent, to be discussed below) equals 1 for this model. This behavior is an
example of finite-size scaling. A perfect power-law behavior at criticality holds strictly
only when L → ∞, while for small systems there can be significant corrections to scaling
(which are unusually small in the case of the Ising model discussed here). Normally we
do not know the exact value of Tc (which may of course be one reason why a simulation
is carried out), and procedures for locating the critical point then have to be devised.
There are many ways to do this, all building in some way or another on the fact that
the order parameter and related quantities should behave as non-trivial power laws (with
known or unknown exponents) for large lattices at a critical point.
Note that the squared magnetization is related to the spin correlation function,
C(ri j ) = hσi σ j i, (49)
where ri j is the distance between the spins. If there is long-range order, then C(r) → hm2 i
when r → ∞ in an infinite lattice. For finite systems, the same is true for C(rmax√ ) in the
limit L → ∞, where rmax is the longest distance on a periodic lattice, e.g., rmax = 2L for
an L × L lattice. If there is no order, then C(r) → 0, according to a power-law at Tc and
exponentially for T > Tc . Below Tc , in the infinite syztem the “connected” correlation
function,
C∗ (r) = C(r) − hmi2, (50)
decays to zero exponentially as r → ∞. The exponential forms of both C(r) and C∗ (r)
are characterized by a correlation length ξ , which diverges as T → Tc from either side.
164
Note also that hm2 i can be written exactly as a sum over the spin correlations;
1
N∑
hm2 i = C(r). (51)
r
The L−2 convergence of hm2 i (both below and above Tc ) is simply related to the finite
correlation length.
Here it should be pointed out that the nature of the spin correlations below Tc depends
on the symmetry of the order parameter we are dealing with. For a vector order param-
eter, e.g., in the Heisenberg model (with classical or quantum spins), the correlations of
the spin components parallel (longitudinal) to the order parameter of a symmetry-broken
system decay exponentially, as in the Ising model. The transverse correlations decay as
a power law, however. This is related to the continuous symmetry of the order parameter.
In the Ising model the symmetry is discrete—the ordered state breaks the spin-inversion
symmetry and the free-energy cost of a local magnetization fluctuation is large (propor-
tional to the boundary of a flipped domain). This leads to the exponentially decaying
(connected) correlation function in the ordered phase. In the Heisenberg model, on the
other hand, there are gapless spin-wave excitations (which are excitations of the spins in
the directions transverse to the ordering direction and are more generally called Gold-
stone modes). Fluctuations due to these lead to a power-law form of the transverse spin
correlations. The transverse correlation length is then formally infinite, and in a finite-
lattice calculation in which the rotational symmetry is not explicitly broken (so that the
computed hm2s i contains contributions from both longitudinal and transverse correla-
tions), the size corrections to hm2s i are ∝ 1/L. We already discussed this behavior in the
ground state of the 2D S = 1/2 Heisenberg antiferromagnet in Sec. 2.4, and Fig. 5(b)
shows the 1/L corrections very clearly.
3.3.1. Scaling and critical exponents

To discuss finite-size scaling close to a critical point in more detail, we first have to
review some basic aspects of critical phenomena in the thermodynamic limit. We will
only list some of the key results and definitions here; see any standard text on critical
phenomena (e.g., the book by Cardy [15]) for more details.
The correlation length is one of the most important concepts underlying the theory of
phase transitions and critical phenomena. In an infinite system, as the critical tempera-
ture is approached, the correlation length diverges according to a power law;
ξ ∼ |t|−ν . (52)
The exponent ν is the same upon approaching Tc from above or below, but the pref-
actor in (52) is in general different for t → 0+ and t → 0− (however, ratios of these
prefactors—called amplitude ratios—are often universal). Within mean-field theory,
ν = 1/2 . Exactly at Tc , although the correlation length is formally infinite, the sys-
tem is not yet ordered. Instead, for an infinite system the correlation function exactly at
Tc has the power-law form,
C(r) ∼ r−(d−2+η ) , (53)
165
where d is the dimensionality of the system and η is another critical exponent (also
called the anomalous dimension, because it can be related to the fractal dimensionality
of ordered domains at the critical point), the mean-field value of which is η = 0. Long-
range order sets in only infinitesimally below Tc , where the asymptotic long-distance
correlation approaches a constant; C(r → ∞) = hm2 i.
We have already discussed the behavior of the order parameter in the symmetry-
broken state as t → 0− ;
hmi ∼ |t|β . (54)
We will also be interested in the corresponding susceptibility, defined as
dhmi N
χ= hm2 i − h|m|i2 ,

= (55)
dh h→0 T

where h is the strength of a field coupling to the order parameter, e.g., in the case of
the Ising ferromagnet a term −h ∑ σi in the hamiltonian. The last expression in (55)
shows explicitly how χ is directly related to the fluctuations of the order parameter. The
susceptibility diverges at Tc ;
χ ∼ |t|−γ . (56)
Thus, upon approaching a critical point, the system becomes infinitely sensitive to a field
h coupling to the order parameter, and exactly at Tc the linear response form hmi = χ h
ceases to be valid. Instead, at Tc , the order parameter depends on the (weak) field as h1/δ .
According to the result in Eq. (39), the mean-field value of the exponent here is δ = 3.
The specific heat is also singular at Tc ,
C ∼ |t|−α , (57)
where α can be positive or negative. In mean-field theory α = 0. When α < 0, there is

no divergence, only a cusp singularity at Tc . In some cases, e.g., for the 2D Ising model,
α = 0 but there is still a weak, logarithmic divergence of the specific heat.
The critical exponents ν , β , η , etc., that we have encountered above are not all inde-
pendent of each other. Relationships between the exponents are explained by the renor-
malization group theory, which, in the type of order-disorder transitions we are discussed
here, shows that there are two more fundamental exponents, in terms of which the physi-
cally observable exponents can be written [15]. Exponent relations had been found using
other arguments even before the advent of this ultimate theory of phase transitions (in
the most complete form by Widom in the mid 1960s), e.g.,
γ = ν (2 − η ),
νd = 2 − α, (58)
α + 2β + γ = 2.
Such relations are very useful for checking the consistency of numerical calculations
of the exponents. Exponent relations involving the dimensionality d are called hyper-
scaling relations and are less generic than the other relations. They are not applicable,
e.g., within mean-field theory, and, therefore, for any system at or above the upper
critical dimension.
166
3.3.2. The finite-size scaling hypothesis
The basic assumption underlying finite-size scaling theory [120] is that deviations
from the infinite-size critical behavior should occur when the correlation length ξ (of
the infinite system) becomes comparable with the finite system length L. If L ≫ ξ , the
fact that the system is finite should be irrelevant and the infinite-size behavior applies.
On the other hand, if L ≪ ξ , then L, not ξ , should be the most relevant length-scale. In
order to see how the two length scales come into play, it is useful to express quantities
of interest in terms of the correlation length. Consider a quantity Q which exhibits a
power-law divergent behavior at Tc (reduced temperature t → 0),
Q ∼ |t|−κ , (59)
e.g., the susceptibility (in which case κ = γ ). We can use Eq. (52) to express |t| as a
function of the correlation length;
|t| ∼ ξ −1/ν , (60)

and using this we can write Q as
Q ∼ ξ κ /ν . (61)
This form should apply for ξ ≪ L, but when ξ ∼ L the divergence can no longer continue
on the finite lattice. The maximum value Qmax attainable by Q on the finite lattice should
then be obtained by replacing ξ → L in Eq. (61), giving
Qmax (L) ∼ Lκ /ν . (62)

In the same way, from Eq. (60) we can also deduce the scaling of the reduced tempera-
ture at which ξ reaches L, which should also be the temperature at which the maximum
value of the quantity Q is reached;
|tmax (L)| ∼ L−1/ν . (63)
It should be noted that this is just a proportionality, and the number in front of L−1/ν
depends on the quantity. The shift also applies to non-divergent quantities—any feature
which develops singular behavior as T → Tc should shift at the rate L−1/ν . If there is a
well-defined maximum or other distinguishable feature in some quantity at T = T ∗ (L),
then this temperature can be used as a size-dependent critical temperature (and, again,
this temperature is not unique but depends on the quantity considered).
The finite-size scaling laws (62) and (63) follow from a more general finite-size
scaling hypothesis [120], which, like the scaling theory for infinite systems, was initially
proposed based on phenomenological considerations, but was later derived using the
renormalization group theory. The hypothesis is that an observable which is singular
at Tc in the thermodynamic limit scales with the system size close to Tc as a power of
L multiplied by a non-singular function of the ratio ξ /L. Any singular quantity (not
necessarily divergent) should thus be of the form
Q(t, L) = Lσ f (ξ /L), (64)
167
0.05
2 L = 128
10 L = 16, 32, 64 0.04
L=8
0.03
-γ/ν
1
10
χ
χL
0.02
0
10 0.01
0
2.0 2.5 3.0 3.5 -2 -1 0 1 2
T/J 1/ν
tL
FIGURE 14. Monte Carlo results for the susceptibility (55) of the Ising model on several different L × L
lattices. (a) shows the temperature dependence, with the vertical line indicating Tc . Note the vertical log
scale. In (b) the data has been scaled using the exact values of the Ising exponents, γ = 7/4 and ν = 1,
and the exact value of Tc in t = (T − Tc )/Tc .
which, using ξ ∼ |t|−1/ν , we can also write as
Q(t, L) = Lσ g(tL1/ν ). (65)

This scaling law should hold both above (t > 0) and below (t < 0) the critical point.
Exactly at Tc , we recover the size-scaling Q(0, L) ∼ Lσ . To relate σ to the standard
critical exponents, we can use the fact that, for fixed t close to 0, as the system grows the
behavior for any t 6= 0 eventually has to be given by Eq. (59); Q(t, L → ∞) ∼ |t|−κ (where
κ is negative for a singular non-divergent quantity, e.g., the for the order parameter we
have κ = −β ). To obtain this form, the scaling function g(x) in (65) must asymptotically
behave as g(x) ∼ x−κ for x → ∞. In order for the size-dependence in (65) to cancel out,
we therefore conclude that σ = κ /ν , i.e.,
Q(t, L) = Lκ /ν g(tL1/ν ). (66)

To extract the scaling function g(x) using numerical data, one can define
yL = Q(t, L)L−κ /ν , xL = tL1/ν , (67)

and plot yL versus xL for different system sizes. If the scaling hypothesis is correct,
data for different (large) system sizes should fall onto the same curve, which then is
the scaling function (this is referred to as curves collapsing onto each other); g(x) =
yL→∞ (x). Fig. 14 illustrates this using Monte Carlo data for the magnetic susceptibility
of the 2D Ising model. The peak location in panel (a) clearly moves toward the known
Tc with increasing L. After scaling the data according to the above procedures, as shown
in panel (b), the curves indeed collapse almost onto each other close to t = 0, but further
away from the critical point deviations are seen for the smaller systems. These are due to
corrections to scaling, which in principle can be described with subleading exponents.
We can apply the scaling form (66) to the correlation length itself, for which κ = ν and
the L-scaling is independent of model-specific exponents. In cases where the universality
168
class is not known a priori, this is useful for extracting the exponent ν by curve-
collapsing ξ /L data without having to simultaneously adjust another exponent κ in (66).
Practical definitions of the correlation length. The correlation length can be defined
in various ways, not necessarily just based on the asymptotic decay of the correlation
function (which is often difficult to extract reliably). One practical and commonly used
correlation length definition is based on the Fourier transform of the correlation function,
often called the (static) structure factor,
S(q) = hσ−q σq i = ∑ e−iq·rC(r) = ∑ cos(q · r)C(r), (68)

r r
where σq is the Fourier transform of an individual spin configuration,
1
σq = √ ∑ σ j e−iq·r j . (69)
N j
We denote by Q the wave-vector of the dominant correlations—for a ferromagnet Q = 0,

for a 2D antiferromagnet Q = (π , π ), etc. To simplify the notation, q will be used for the
deviation from Q. Then q1 = 2π /L corresponds to one of the wave-vectors closest to Q,
e.g., Q + (2π /L)x̂, where x̂ is the reciprocal-space unit vector in the x-direction.
A correlation length ξa can be defined using the structure factors at q = 0 and q1 ;
s
1 S(0)
ξa = − 1. (70)
q1 S(q1)
One can show that, for a d-dimensional lattice, if the correlation function is given by the
Ornstein-Zernike form (obtained in mean-field treatments [15]),
1
COZ (r) ∼ r− 2 (d−2) exp(−r/ξ ), (71)
then ξa is related to the original correlation length ξ appearing in the exponential decay
of this correlation function according to
r
(1 + d)(3 + d)
ξa = ξ . (72)
8d
Thus, for d = 1 and 3, ξa = ξ , whereas the 2D case is special, with ξa = ξ (15/16)1/2

(or, one may say that d = 1, 3 are special cases, since the factor differs from one also for
all d > 3). The Ornstein-Zernike form is normally valid in disordered phases for r ≫ ξ
[15]. Deviations from this form at short distances imply that (72) is not true exactly, but
regardless of the short-distance behavior this relation holds exactly when ξ → ∞.
In the case of a long-range ordered classical system, ξa normally diverges as T →
0, for any L, because there are no fluctuations in the ground state (and, thus, the
structure factor vanishes for q 6= 0. In order to remove the contributions from the non-
decaying part of the correlation function in an ordered system, we can use the connected
169
2
10
L=8 0.3
1 L = 16,32,64
10
L = 128
ξa/L
ξb/L
0 0.2
10
-1 0.1
10
-2
10 0.0
2.0 2.5 3.0 2.0 2.5 3.0
FIGURE 15. Temperature dependence of the two correlation length definitions, Eqs. (70) and (73),
normalized by the size L for L × L Ising models. The vertical lines indicate Tc .
correlation function (50). Although hmi2 is not uniquely defined for a finite system, one
can, e.g., subtract C(rmax ). As an alternative, we can use a different correlation length
definition, based on the structure factor at q1 and q2 = 2q1 = 4π /L;
s
1 S(q1 )/S(q2) − 1
ξb = . (73)
q1 4 − S(q1 )/S(q2)
One can show that (72) also holds for this definition if C(r) is given by the Ornstein-
Zernike correlation function, also when a constant corresponding to the long-range order
is added for T < Tc [since this affects only S(0)].
Fig. 15 shows results for ξa /L and ξb /L for the 2D Ising model. Note that a logarith-
mic scale is used in the case of the T → 0 divergent ξa /L, while ξb /L is convergent and
graphed on a linear scale. Both quantities exhibit size independence (curves crossing
each other) at Tc , but the values there are clearly different. This is because the Ornstein-
Zernike form of the correlation function applies asymptotically only for T > Tc , and
there is no reason why the two definitions ξa and ξb should agree exactly at Tc (although
their values should be related). Their values are very similar for large systems close to
Tc in the disordered phase. The crossing points (or peak location of ξb ) can be used to
extract Tc in systems where it is not known. The temperature axis can also be scaled in
the same way as in Fig. 14 to extract the correlation-length exponent.
Note that for the small number of q-points needed, S(q) can be efficiently evaluated
using the Fourier transform (69) of the spin configurations generated in Monte Carlo
simulations. Since the structure factor is real-valued we have
S(q) = hRe{σq }2 i + hIm{σq }2 i. (74)
Computing the full correlation function C(r) and Fourier-transforming it post-simulation

is much more time-consuming.
Binder ratio and cumulant. Besides ξ /L, there are also other dimensionless quanti-
ties that are size-independent at the critical point and useful for extracting Tc indepen-
dently of the values of the critical exponent. The perhaps most frequently used one is
170
1 1.00
L = 128
0.8 L = 16, 32, 64
L=8 0.95
0.6
U2
U2
0.4 0.90
L = 1024
L = 128, 256, 512
0.2
0.85 L = 64
0
1 2 3 4 5 2.24 2.25 2.26 2.27 2.28
T/J T/J
FIGURE 16. The Binder cumulant (76) for L × L Ising models. In (a) on can see the approach to the
limiting values U2 → 0 (for T > Tc ) and U2 → 1 (for T < Tc ) for increasing L. In (b) the data close to Tc
(vertical line) are graphed on a more detailed scale and for larger L to show the crossings of the curves.
the Binder ratio [121, 122]:

hm4 i
R2 = . (75)
hm2 i2
At Tc , the power laws cancel out, and the ratio is L-independent (and also universal), up
to subleading finite-size corrections. Normally, graphing R versus T for different system
sizes produces curves that intersect each other close to Tc . Locating the points where R2
for pairs of system sizes (e.g., L and 2L) cross each other, one obtains a size dependent
critical point which typically converges faster than the L−1/ν shift in (63). One can think
of this as being a results of the leading corrections canceling in a quantity involving two
system sizes, and one is then left with something which approaches Tc according to a
faster, higher-order scaling correction. One can also define ratios similar to (75) based
on other powers of m, e.g., R1 = hm2 i/h|m|i. The curve-crossing method for locating Tc
can also be applied in the same way with ξ /L.
The Binder ratio also has other interesting properties. In the case of a scalar order
parameter (e.g., for the Ising model), the Binder cumulant is defined as

3 1
U2 = 1 − R2 . (76)
2 3
In an ordered state, U2 → 1 for N → ∞, because the magnetization distribution P(m)

then approaches two delta-functions at ±h|m|i, and, therefore, R2 → 1. In contrast, in a
disordered phase, the fluctuations of m are Gaussian around m = 0 (following from the
central limit theorem, since fluctuations in regions separated by distance ≫ ξ in a large
system are uncorrelated). Based on Gaussian integrals, R2 → 3 and U2 → 0.
Generalizing the Binder cumulant to an n-component order parameter (where n = 1
for the Ising magnetization, n = 2 for the XY model, etc.), one should keep in mind that
m2 = m · m and m4 in (75) are insensitive to angular fluctuations of the order parameter.
Integrating a Gaussian distribution of |m| over the n-dimensional space to compute the
averages in (75) does, however, introduce n-dependent factors. To reproduce the above
171
-4
2×10
0.04 -4
1×10
χL -g(x)
−γ 0
χL
−γ
0.03 L = 64
L = 128
-4
L = 256 -1×10
0.02 (a) L = 512 (b)
-4
-2×10
-0.4 -0.2 0.0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
1/ν 1/ν
x=tL x=tL
FIGURE 17. (a) Best-fit scaling collapse of the susceptibility of the 2D Ising model; the same data
as in Fig. 14, but including larger lattices and adjusting Tc in t = (T − Tc )/Tc as well as the exponents
ν , γ to minimize χ 2 with respect to a scaling function g(x) in the form of a polynomial (here of fourth
order, shown as the solid curve). The optimal values for this data set are: Tc /J = 2.26921 ± 0.00002,
ν = 0.9985 ± 0.0011, and γ = 1.750 ± 0.002, where the error bars (one standard deviation) were computed
by repeating the fit several times with Gaussian noise (of magnitude equal to the Monte Carlo error bars)
added to the data. (b) shows the data with the fitted scaling function g(x) subtracted, so that the error bars
become visible. The fit is statistically sound, with χ 2 ≈ 0.9 (per degree of freedom).
properties of U2 one therefore has to define it for a general order parameter as

n+2 n
U2 = 1− R2 . (77)
2 n+2
Fig. 16 shows Monte Carlo results for U2 as a function of T for several 2D lattices. The
evolution into a step-function at Tc with increasing system size can be seen clearly. In
this case all the curves cross each other very close to the known Tc , reflecting very small
subleading corrections in this model. It is common in other systems that the crossings
exhibit some drift, and a careful extrapolation of crossing points have to be carried out
(e.g., based on data sets for sizes L and 2L, or some other aspect ratio).
Finite-size scaling in practice. We briefly discuss how to carry out finite-size data-
collapse in practice. The number of parameters involved (i.e., Tc as well as one or two
exponents, and possibly also exponents of subleading corrections to be discussed below)
is rather small, and normally one has some rough knowledge of their values just from
looking at raw data and doing some initial experimentation, e.g., by just locating a non-
trivial power-law behavior as in Fig. 13 [and this may often be enough to determine the
exponent ratio κ /ν in Eqs. (62) and (66)]. An analysis of the Binder cumulant or ξ /L
may already have given Tc to adequate precision, but it may still be useful to check the
sensitivity of other fits to its value. Thanks to the power of modern computers, as an
alternative to using some complicated multi-dimensional optimization procedure, one
can write a simple brute-force computer program to search for the best set of parameters
on a suitable finite mesh. The goodness of the data collapse produced by a set of
parameters can be quantified as the χ 2 -value obtained by fitting a single high-order
polynomial through all the scaled data points (xL , yL ), defined in (67), simultaneously
for all L for which data are available. Normally there is a large number of data points,
172
for different couplings and system sizes, and, since the scaling function should be well
behaved, a polynomial of reasonable order (3th-8th, as a rough guideline) should work
well within the window where the data can be collapsed. The size of this window also
has to be adjusted until the data collapse well. As an example, for the data in Fig. 14,
x in the range (−0.5, 0.5) should be appropriate, although the window also depends on
the system sizes included in the analysis and the the error bars (which determine the
sensitivity to neglected subleading scaling corrections).
It is not always easy to determine reliable error bars for parameters obtained in this
kind of fitting. Beyond the purely statistical errors (which can be determined, e.g., by
repeating the χ 2 minimization several times with Gaussian noise, of magnitude equal
to the error bars, added to the data), there are also systematical errors due to scaling
corrections, which can be difficult to estimate. If χ 2 is statistically reasonable (i.e., close
to 1 per degree of freedom), one would normally assume that the neglected corrections
have not influenced the parameters beyond the statistical uncertainties.
Fig. 17 shows an example of data collapse. The critical exponents and Tc of the 2D
Ising model were determined using data for L ∈ {64 − 512} in the window |t|L1/ν < 0.5
(where t contains the variable Tc adjusted in the procedure). The results, listed in the
figure caption, are in excellent agreement with the known Tc and the 2D Ising exponents,
and the fit is also statistically very good. Using the same fitting window, a statistically
acceptable fit cannot be obtained if much smaller lattices are included, e.g., including
L = 32 gives χ 2 ≈ 2 (per degree of freedom), which is marginally too large for the
number of degrees of freedom of the fit. In this case Tc = 2.26924 ± 0.00002, about 3
error bars above the true value, while the exponent are correct to within two error bars.
Corrections to scaling . As a check, it is often a good idea to include some correc-

tions to the leading scaling forms, and some times it is even necessary to do so in order
to obtain good fits. The most commonly used method is to do the data collapse with
yL = Q(t, L)L−κ /ν (1 + aL−ω )−1 , xL = tL1/ν , (78)
where ω is a subleading exponent and a is a constant. Here there is no correction to

the xL scaling (the finite-size shift of the critical point), but such a correction can in
principle be included as well [85]. Normally, good fits with corrections can be obtained
also when lattice sizes are included that have to be left out if no corrections are used.
A larger range of system sizes can partially compensate for the fact that the statistical
uncertainties of all parameters increase when more parameters are included. If consistent
leading critical exponents are obtained in fits both with and without corrections, then one
can be reasonably certain that the results are correct.
3.4. First-order transitions

The scaling properties we discussed in the preceding section apply at continuous
phase transitions, where the correlation length diverges. At first-order (discontinuous)
transitions, the correlation length remains finite at the transition point and the order
parameter, as well as other quantities, exhibit discontinuous jumps. The discontinuities
173
develop in the limit of infinite system size, normally according to power-laws which
can also be studied using finite-size scaling techniques. The exponents associated with
these powers-laws are typically trivially related to the dimensionality of the system
[123, 124, 125]. For instance, the specific heat diverges with the system size as Ld at
a first-order transition, instead of Lα /ν , with a typically very small (or even negative) α ,
at a continuous transition. The shift of the critical point with the system size scales as
L−d , instead of the L−1/ν shift of a continuous critical point.
Although finite-size scaling with exponents equal to d in principle makes it easy to
recognize a first-order transition, studies of weak first-order transitions are difficult, be-
cause they exhibit large corrections to the leading scaling forms. It may then be difficult,
with system sizes accessible in practice, to clearly distinguish slowly developing discon-
tinuities from weaker singular behavior at a continuous transition.
Strongly first-order transitions are also difficult to study, for a completely different
reason. A Monte Carlo simulation may get stuck in a meta-stable state, in which case
computed quantities do not correspond to correct thermal averages (which, on the other
hand, is completely analogous to real systems, for which metastability and hysteresis ef-
fects are hall-marks of first-order transitions). It may then even be difficult to accurately
locate the transition point. To alleviate such problems, various multi-canonical Monte
Carlo methods have been developed in which the configurations are sampled in an ex-
tended ensemble where the temperature of the system is also fluctuating (tempering or
parallel tempering [126, 127] methods), or with a distribution different from the Boltz-
mann probability [128, 129, 130] (to which the measurements are re-weighted), which
makes it easier for the system to explore the configuration space.
3.4.1. Phase coexistence

One of the most important characteristics of a first-order transition is phase coexis-
tence. In Monte Carlo simulations at the transition point (in practice in a small window
which shrinks to a single temperature with increasing system size) this is manifested in
the generation of two types of configurations, corresponding to the two distinct phases
existing just above and below the transition temperature. This is provided that the full
configuration space can be ergodically sampled, which, as discussed above, is not always
possible in practice. In small systems, there will also be configurations that cannot be
clearly associated with one of the phases—they correspond to the fluctuations (domain
walls) required for the system to transition between the two phases. Fig. 18 illustrates
this schematically for an Ising order parameter. The transition window in which a three-
peak distribution can be observed narrows rapidly with increasing system size, and the
peaks develop into delta-functions. When the weight between the peaks becomes very
small, it may in practice not be possible to ergodically sample the configurations, as the
system gets trapped within the sub-space corresponding to just one of the peaks (in a
way exactly analogous to the symmetry-breaking discussed in Sec. 3.2).
Note that the type of order-parameter distribution in Fig. 18 does not apply to the field-
driven first-order transition of the Ising model, with magnetization curves illustrated in
Fig. 10. In that case there would be just two peaks, with weight transferring between
174
P(m) T<Tc P(m) T=Tc P(m) T>Tc
m m m
-1 0 1 -1 0 1 -1 0 1
FIGURE 18. Evolution (schematic) of the magnetization distribution of a finite Ising ferromagnet close
to a first-order transition. There is a rapid transfer of weight between the central peak (corresponding to
the disordered phase) and the two peaks at non-zero magnetization (the two ferromagnetic states) as the
temperature is tuned through the transition region. The peaks become delta-functions in the limit of infinite
system size. The transition temperature Tc can be defined as a point with some specific (but essentially
arbitrary) feature of the three peaks, e.g., equal weight or equal height of the peaks.
them as the field is tuned through h = 0. The phase coexistence at h = 0 (and T < Tc )
corresponds to the two symmetric peaks in the distributions in Fig. 12. The similarity
between this case and a paramagnetic-magnetic transition (with a general n-component
order parameter m) can be made clearer by considering the one-dimensional distribution
of |m|. This distribution has two peaks, one at |m| = 0 and one at |m| > 0, when an
ordered and disordered phase coexist. Note again that in continuous transitions (Fig. 12)
the central peak continuously splits into two peaks below Tc , in contrast to the two
ordering peaks emerging at a non-zero value of |m| in the first-order case.
The Binder ratio, Eq. (77) in the case of a generic n-dimensional order parameter
depends only on the distribution of |m|. It has a very interesting property at a first-order
transition. One can easily check, using, e.g., an idealized co-existence distribution of the
type P(|m|) = (1 − p)G(|m|) + pδ (|m| − m∗), where G is a normalized “half-Gaussian”
(defined for |m| ≥ 0) and 0 < m∗ ≤ 1, that U2 exhibits a negative divergence when the
ordered-phase weight p is tuned from 0 to 1, and if the width of the Gaussian vanishes
(corresponding to infinite system size). While the full m distribution of course contains
more information, locating a window of negative Binder cumulant and checking for a
divergence of the peak value can be a practical way to analyze a first-order transition.
3.4.2. Frustrated Ising model

Here we look at a particular example of a first-order transition, in the frustrated 2D
square-lattice Ising model with hamiltonian
Eσ = −J1 ∑ σiσ j + J2 ∑ σiσ j , (79)
hi ji1 hi ji2
where both couplings J1 , J2 > 0 (but note the different signs in front of the parameters).
The first term (where hi ji1 denotes nearest-neighbor spins) is then the standard Ising fer-
romagnet, whereas the second term (where hi ji2 refers next-nearest-neighbor spins, i.e.,
across the diagonals on 2 × 2 plaquettes, as in Fig. 7) is antiferromagnetic and causes
175
T=0.771 T=0.772 T=0.773
FIGURE 19. Order parameter distributions in the (mx , my ) plane (in the full space |mx | ≤ 1, |my | ≤ 1)
of the 2D frustrated Ising model of size L = 128 at coupling ratio g = 0.55. Brighter features correspond
to higher probability density. The temperatures (indicated in the panels in units of J1 ) are in the first-order
transition region, with the phase coexistence (a central peak as well as four peaks corresponding to x- and
y-oriented stripes) seen most clearly at g = 0.772.
frustration. For coupling ratios g = J2 /J1 < 1/2 the ground state of the system is fer-
romagnetic (fully polarized). In the limit g → ∞ the system reduces to two decoupled
antiferromagnets, with four striped (or collinear) ground states, such as the one illus-
trated in Fig. 7(c). These remain the ground states for all g > 1/2. At the point g = 1/2
the ground state is massively degenerate. Note that the case of J1 < 0 is equivalent to
J1 > 0, through the invariance of Eσ under the transformation σi → −σi on one of the
checker-board sublattices; thus in this case the ferromagnetic phase is replaced by an
antiferromagnetic one, whereas the striped phase is not changed.
The frustrated Ising model has been studied for a long time, but some features of its
phase diagram are still debates or were resolved only recently [95, 96, 97]. For g < 1/2,
the transition is of the standard Ising type, but is difficult to study close to g = 1/2,
due to large scaling corrections and long Monte Carlo autocorrelation times. Here we
will consider only g > 1/2, for which there is a first-order transition between a high-
temperature paramagnet and a low-temperature striped phase, up to a coupling g∗ ≈ 0.8
[97] after which the transition becomes continuous. We analyze results obtained with
the standard single-spin Metropolis algorithm (as cluster Monte Carlo methods do not
work in the presence of frustration). Temperatures will be quoted in units of J1 .
Order-parameter distribution. The ordered phase can have its stripes oriented either
along the x or y-axis, with the corresponding order parameters
1 N 1 N
mx = ∑ σi(−1)xi ,
N i=1
my = ∑ σi(−1)yi ,
N i=1
(80)
where xi and yi are the (integer) lattice coordinates of site i. Let us first look at the
order-parameter distribution P(mx , my). If the four possible ordered states are sampled
equally below Tc , the distribution should be four-fold symmetric, with peaks located
on the negative and positive x- and y-axis. In the paramagnetic phase there should be a
single central peak. Coexistence at a first-order transition should hence be reflected in the
presence of five peaks in this case (instead of the 3-peak distribution for the scalar order
176
1.0 1.0
g=0.55 g=0.70
0.5
0.8
0.0
0.6
-0.5
U2 0.4
-1.0 L=4
L = 8, 16, 32 0.2
-1.5 L = 64
L = 128 0.0
-2.0
-0.2
0.6 0.8 1 1.2 1.2 1.4 1.6 1.8
T/J1 T/J1
FIGURE 20. Binder cumulant across the transition to the striped phase of the frustrated Ising model
for coupling ratios g = 0.55 (left) and g = 0.70 (right). The divergent negative peak developing with
increasing L is an unambiguous signal of a first-order transition.
parameter in Fig. 18). Fig. 19 shows results for an L = 128 system at g = 0.55, for three
temperatures in the transition window. A very distinct four-fold symmetry can be seen in
all cases. Four symmetric peaks clearly indicate an ordered phase at T = 0.771, but there
are strong fluctuations, reflected in weight extending to the center of the distribution. At
T = 0.773, the distribution is peaked at the center, but weight also extends far into the
ordered regions. Between these temperatures, at T = 0.772, one can discern both the four
ordering peaks and a central peak. All these plots show the hall-marks of coexistence;
even when there are not five peaks present, there is still significant probability for both
ordered and paramagnetic configurations. Away from the narrow transition window, the
distribution rapidly turns into one with either distinctly ordered or disordered features.
Binder cumulant. Let us analyze the Binder cumulant, using m2 = m2x + m2y in
Eq. (77). The question now is what order-parameter dimensionality n to use in this
definition. At high temperatures, fluctuations of mx and my in different parts of a large
system are uncorrelated. This implies a rotationally-symmetric (circular) distribution
P(mx , my). The n-dependent factors are intended to make U2 → 0 for T → ∞, and in this
case we should therefore use n = 2. This is correct also at low temperatures (to guarantee
U2 → 1), because in an ordered phase the Binder ratio, Eq. (76), R2 → 1 when the
system size diverges, regardless of the order parameter structure. Fig. 20 shows results
as a function of temperature at two coupling ratios; g = 0.55 and 0.70. For g = 0.55,
a negative Binder cumulant can be seen for L ≥ 8, with the negative peaks becoming
very narrow and apparently diverging as the system size is increased. At g = 0.70,
the transition is still first-order, but with weaker discontinuities that start to manifest
themselves as co-existence and a negative Binder cumulant only around size L = 32.
In both cases, the Monte Carlo simulations were still ergodic, but a large number of
updating sweeps (≈ 108 for each case) had to be carried out to obtain smooth curves.
Discontinuities and finite-size scaling. In first-order transitions that are not very
strong, as in the systems above, it may not be easy to accurately extract the size of
discontinuities in physical observables. An example is shown in Fig. 21, where the
177
-0.7
L = 256
g = 0.55
L = 8,...,128
-0.8 L=4 g = 0.51
10
Cmax
E -0.9
-1.0 1
-1.1
0.4 0.6 0.8 1 10 100
T/J1 L
FIGURE 21. Left: Internal energy versus temperature for the frustrated Ising model at g = 0.55. The
discontinuity developing with increasing lattice size corresponds to the latent heat. Right: Finite-size
scaling of the peak value of the specific heat at g = 0.51 and 0.55. The line shows the expected asymptotic
L2 scaling at a first-order transition.
temperature dependence of the internal energy is graphed for several lattice sizes. While
a discontinuity (latent heat) clearly develops with increasing L, it is not easy to determine
exactly between which two energies the jump will eventually take place. This requires a
careful analysis using larger lattices.
Fig. 21 also shows the maximum value of the specific heat versus the system size for
two different coupling ratios. At g = 0.51, the transition is rather strongly first-order,
and a scaling consistent with the expected Cmax ∼ L2 can be observed for the largest
lattices. In contrast, at g = 0.55 the behavior appears to follow a different power law
(with an exponent close to 1.2) up to L ≈ 64. For larger lattices the results start to show
a somewhat more rapid divergence, however, and eventually, for very large lattices, one
would expect the exponent to be equal to d = 2 also in this case.
3.5. Spin stiffness and the Kosterliz-Thouless transition

An important aspect of a system with long-range magnetic order is that it exhibits a
non-zero spin stiffness. For a system with continuous vector spins (XY or Heisenberg
models), the spin stiffness is the analogue of an elastic modulus of a solid. It is also often
refereed to as the helicity modulus [131]. To study this concept, we here consider the XY
model, with the hamiltonian written as
H = −J ∑ cos(Θi − Θ j ), (81)
hi, ji
where Θi ∈ [0, 2π ] is the angle characterizing spin i. For simplicity we assume only
nearest-neighbor interactions, but generalizations to arbitrary interactions are straight-
forward. When deriving an expression for the spin stiffness, we will consider the 2D
square lattice (and some times use a 1D chain for simplicity), but later will study
also a 3D cubic system. In 2D XY systems, the spin stiffness is an important quantity
178
characterizing the unconventional (topological) Kosterlitz-Thouless transition exhibited
by this system, which we also discuss briefly in this section.
3.5.1. Definition of the spin stiffness

Loosely speaking, the spin stiffness ρs characterizes the tendency of ordered spins to
adapt in response to perturbations imposing modulations of the direction of the order
parameter (in contrast to the susceptibility, which measures the tendency of the order
parameter to change in response to a field applied in a fixed direction). It is analogous to
the shear modulus in continuum mechanics, which characterizes the tendency to shape
deformation of an elastic object (while the compressibility corresonds to the tendency to
volume change with maintained shape). The definition of the spin stiffness is easiest to
understand at T = 0, which we consider first, before generalizing to T > 0.
The stiffness at T = 0. We will first consider a system with open boundaries in the
x direction, while periodic boundaries in the y direction can be assumed (and later we
will generalize to periodic boundaries also in the x direction). Fig. 22 illustrates how
ferromagnetic XY spins at T = 0 adapt in order to minimize the energy when an over-all
change in the spin angle Φ between the left and right boundaries is imposed (e.g., due to
strong magnetic fields applied at the boundary columns). To minimize the energy, each
column is twisted with respect to the following column by an angle φ = Φ/(Lx − 1),
where Lx is the length of the system in the x direction (the number of columns). At
T = 0, we are interested in the energy in the presence of this twist, which for the 2D XY
model is simply given by
E(φ ) = −J(Lx − 1)Ly cos(φ ) = E(0) + J(Lx − 1)Ly [1 − cos(φ )]. (82)
For small φ we get E(φ ) − E(0) = (J/2)(Lx − 1)Ly φ 2 to leading-order. Motivated by
this result, the T = 0 spin stiffness is defined as
1 ∂ 2 E(φ )
ρs = , (83)
N ∂φ2
and we have ρs = J for large N [or for any N if we normalize by the number Ly (Lx − 1)
of interacting x bonds, over which the energy cost due to the twist is distributed].
Normally it is more convenient to consider a periodic system. To derive an expression
for the spin stiffness in this case, a phase twist is first imposed at the boundary. For
simplicity we work this out for a 1D chain, with spin angles Θx , x = 0, . . ., N − 1, but the
calculation can be trivially generalized to higher dimensionality. The interaction energy
for each bond is Ex = −J cos(Θx+1 − Θx ), except at the boundary, where a twist Φ
corresponds to EN = −J cos(Θ0 −ΘN−1 +Φ). The configurations minimizing the energy
have Θx+1 − Θx = δ , where δ is independent of x, which gives the total energy
E(Φ) = −(N − 1) cos(δ ) − cos(Φ − [N − 1]δ ), (84)
which is minimized by δ = φ ≡ Φ/N, with E(φ ) = E(0) + JN[1 − cos(φ )], and the spin
stiffness defined according to (83) is again ρs = J.
179
Φ
Ly
Lx
FIGURE 22. A 2D classical XY model with a phase twist in the x-direction imposed by fixing the spins
in the boundary column at a relative angle Φ. To minimize the energy the total twist Φ is distributed
evenly, so that the spins in neighboring columns are twisted by φ = Φ/Lx relative to each other.
It is useful to consider also another way of twisting the spins in a periodic system,
by introducing a twist field Φx = xφ in the hamiltonian. For a 1D system the energy for
each bond in the presence of this field is −J cos(Θx+1 − Θx + Φx+1 − Φx ). To treat the
boundary correctly, Φx should be considered as a function of a continuous variable x,
which jumps discontinuously from N φ to 0 at x = N. The phase difference appearing in
the XY interaction should then be interpreted as
Z x+1 Z x+1
Φx+1 − Φx = Φx dx = φ dx = φ , (85)
x x
which holds also at the boundary (x + 1 = N). Apart from factors, φ is analogous to a
flux threading the ring. The energy in the presence of this twist is minimized for Θx = Θ
independently of x, giving E(φ ) = E(0) + JN[1 − cos(φ )] as before in the case of the
twisted boundary condition. One can consider the twist field as a way of transforming
away the twisted boundary condition (rotating to the reference frame of the spins in the
presence of the boundary twist), which makes many calculations easier in practice. We
will work with the twist field from now on.
The stiffness at T > 0. Now we consider non-zero temperatures, in which case the
spin stiffness is defined as
1 ∂ 2 F(φ )
ρs = , (86)
N ∂φ2
where F(φ ) is the free energy in the presence of a twist field (or, equivalently, a twisted
boundary condition), which in turn is related to the partition function according to
1
F(φ ) = − ln[Z(φ )]. (87)
β
For T → 0, Eq. (86) clearly reduces to the ground-state energy definition (83).
Applying a twist field which only depends on x, Φ(x, y) = xφ , in the 2D XY model,
we have the hamiltonian
H(φ ) = −J ∑ cos(φ + Θ j − Θi) − J ∑ cos(Θ j − Θi), (88)

hi, jix hi, jiy
180
where we can assume periodic boundary conditions in both directions (although in
principle the y boundary could be open). We can simplify the dependence on φ by using
a standard trigonometric equality,
cos(φ + Θ j − Θi ) = cos(Θ j − Θi ) cos(φ ) − sin(Θ j − Θi ) sin(φ ), (89)
which we expand to second order in φ :
cos(φ + Θ j − Θi ) → cos(Θ j − Θi )(1 − φ 2 ) − sin(Θ j − Θi )φ + O(φ 3 ). (90)
The hamiltonian in the presence of a small (φ → 0) twist can then be written as
H(φ ) → H(0) + 12 φ 2 Hx − φ Ix , (91)
where Hx is the x-bond part of the hamiltonian at φ = 0 and Ix is the total spin “current”
in the x lattice direction:
Ix = J ∑ sin(Θ j − Θi ). (92)
hi, jix
The partition function in the presence of a small twist can then be written as
Z
Z(φ ) = d[Θ]e−β H(φ ) (93)
Z
→ d[Θ]e−β H(0) [1 − 12 β φ 2 Hx + . . .][1 + β φ Ix + 12 (β φ Ix )2 + . . .],
where the exponentials involving Hx and Ix have been Taylor expanded to the orders
needed. We can now write this in a form with expectation values over the distribution
for φ = 0. To second order in φ :
Z(φ ) → Z(0)[1 − 12 β φ 2 hHx i + β φ hIx i + 12 β 2 φ 2 hIx2i]. (94)
By symmetry hIx i = 0 and the free energy (87) is given by
F(φ ) = F(0) + 12 φ 2 hHxi − β hIx2 i ,

(95)
and from this we can extract a simple expression for the spin stiffness (86):
1
ρs =
(hHxi − β hIx2i), (96)
N
which can be evaluated using Monte Carlo simulations. This result is for a twist field in
the x lattice direction. For a d-dimensional isotropic system, we can average over all the
equivalent directions and write the spin stiffness as
!
d
1
ρs = hHi − β ∑ hIa i ,
2
(97)
Nd a=1
where the index a corresponds to the current in lattice direction a. For an anisotropic
d-dimensional system, the stiffness in general is different for all lattice directions, i.e.,
there are d different stiffness constants, each of them given by a form like (96).
The twist-field definition of the stiffness can be easily generalized to systems with
longer-range interactions. The forms (96) and (97) remain valid, with the current Ix
containing contributions from all interactions exactly as in the hamiltonian.
181
1.0 1.8 L=4
(a) 1.6 L = 6, 8, 10, 12
0.8
L = 16, 20, 24
1.4
0.6 L = 32
Lρs/J
ρs/J
1.2
0.4 L=4 1.0
L = 6, 8
0.2 0.8 (b)
L = 10
0.6
0.0
0 1 2 3 2.16 2.18 2.2 2.22 2.24 2.26
T/J T/J
FIGURE 23. (a) Temperature dependence of the spin stiffness of the 3D XY model for different lattices
of size N = L3 . (b) Finite-size scaling according to Eq. (99). Curves of Lρs for different L cross each other
(asymptotically for large L) at the critical temperature. These results were obtained using a Monte Carlo
cluster algorithm [119], and in all cases the error bars are too small to be visible.
Relation to superfluidity and superconductivity. A very interesting aspect of the spin

stiffness of the XY model is that it can be directly related to the superfluid density of a
superconductor (or a superfluid such as 4 He) [131, 132]. Like the magnetization of the
XY model, the order parameter of a superconductor or a superfluid is U (1) symmetric
(corresponding to the global phase of the wave function), and the twist field we discussed
above is directly analogous to a magnetic flux (in the case of a superconductor). A non-
zero stiffness corresponds to the Meissner effect exhibited by a superconductor. Monte
Carlo simulations of the 2D XY model directly aimed at properties of thin superfluid
films are discussed in Refs. [133, 134].
Spin stiffness scaling. The critical scaling properties of the spin (or superfluid)
stiffness were first worked out in the context of superfluids [131, 132]. In the infinite
d-dimensional system with d > 2, it was shown that
ρs ∼ (Tc − T )(d−2)ν , (98)
for T → Tc from below. Here ν is the standard correlation length exponent. According to
the general finite-size scaling relation (66), the size dependence of the stiffness exactly
at Tc is then given by
ρs ∼ L2−d . (99)
The stiffness is therefore, like the correlation length and the Binder ratio, a useful
quantity for locating the critical point without having to adjust any unknown (or not
precisely known) exponents. This is illustrated with Monte Carlo results for the 3D XY
model in Fig. 23, where Lρs is size independent at the critical point.
3.5.2. Kosterliz-Thouless transition in two dimensions

In the 2D XY model there can be no transition into a phase with long-range magnetic
order at T > 0, according to the Mermin-Wagner theorem [38] (as discussed in Sec. 2).
182
Remarkably, this system exhibits a different kind of phase transition, where no long-
range order develops but the spin correlations change from exponentially decaying to
a power-law form [135, 136]. This Kosterlitz-Thouless (KT) transition is topological in
nature, being a consequence of proliferation of unbound vortices (which are topological
defects) in the spin configurations at temperatures T > TKT . For T ≤ TKT , vortices
also exist (at a density which vanishes at T → 0) but they are all bound in vortex-
antivortex pairs, which have no net vorticity (and therefore vanish upon course graining
of the spins). The power-law form of the spin correlations, C(r) ∼ r−η , applies for all
0 < T ≤ TKT . The exponent η is temperature dependent, with the value η = 1/4 exactly
at TKT and η → 0 as T → 0, so that true long-range order exists at T = 0.
Although there is no long rage order for 0 < T ≤ TKT , the spin stiffness is actually
non-zero in the KT phase—power-law correlations with a sufficiently small exponent are
enough to support an energy cost of a boundary twist. There is no power-law onset of the
stiffness, but instead an even more prominent signal of the transition; a discontinuous
jump at TKT from ρs = 0 to a non-zero value. Renormalization-group calculations for
the continuum field theory corresponding to the 2D XY model have given very detailed
information about the KT transition. One of the most important results is a rigorous
relationship (due to Nelson and Kosterliz [139]) between the transition temperature TKT
and the spin stiffness exactly at this temperature (i.e., the size of the discontinuity);
2TKT
ρs (TKT ) = . (100)
π
For finite lattices, the stiffness at TKT approaches the infinite-size value with a logarith-
mic size correction [138]:

1
ρs (TKT , L) = ρs (TKT , ∞) 1 + , (101)
2 ln(L) + c
where c is a system dependent parameter. Thus, one can say that the general finite-size
scaling law (99) for ρs at a critical point holds, but with a logarithmic correction to the
leading-order size-independent form obtaining when d = 2.
Fig. 24(a) shows Monte Carlo results for the stiffness of the 2D XY model versus
temperature. The jump expected in the thermodynamic limit is approached very slowly
as a function of size. This can be related to the fact that the correlation length for T > TKT
does not diverge as a power-law, but according to the exponential form
1/2
ξ ∼ ea/(T −TKT ) , (102)
where a depends on details of the lattice and the interactions. Therefore, using the same
arguments as we in the case of a standard critical point in Sec. 3.3.2, the finite-size shift
of TKT [defined using, e.g., the temperature T ∗ (L) at which ρs drops the most rapidly] is
given by the form,
1
T ∗ (L) − TKT ∼ 2 , (103)
ln (L)
which is much slower than the conventional power-law shift T ∗ (L) − Tc ∼ L1/ν .
183
1.0
0.5
[1+1/(2ln(L)+c)] ρs/J
0.8
L=8
-1
(a) 0.4
L = 16
0.6
ρs/J
0.3 L = 32
L=4 L = 64
0.4
L=8 0.2 L = 128
L = 16
0.2 L = 32 0.1
L = 64 (b)
0.0 0.0
0 0.5 1 1.5 -20 -15 -10 -5 0
T/J 1/2
ln(L)-a/(T-TKT)
FIGURE 24. (a) Monte Carlo results (obtained with a cluster algorithm [119]) for the temperature
dependence of the spin stiffness of the 2D XY model for several lattices of size L = 2n . A discontinuity
develops at TKT as L → ∞, as indicated with the vertical line at the known transition temperature (TKT ≈
0.8933 [137]). According to the Nelson-Kosterlitz relation (100), the stiffness exactly at TKT , for any
system exhibiting a KT transition, must fall on the line shown; ρs = 2T /π . (b) Finite-size data collapse
according to the combined (T, L) scaling hypothesis (105), using the known TKT = 0.8933 and with the
two parameters, a = 1.5 and c = 0.7, chosen to obtain good data collapse. The vertical line shows the
asymptotic T → TKT , L → ∞ value expected according to the Nelson-Kosterlitz relation.
Using the standard finite-size scaling hypothesis (64) and replacing Lσ (the size
dependence of the singular quantity exactly at the critical point) by the logarithmic size
correction in (101), we can write a hypothesis for the size and temperature dependence
of the spin stiffness at the KT transition as

1 1/2
ρs (T, L) = 1 + f (ea/(T −TKT ) /L), (104)
2 ln(L) + c
which, after taking the logarithm of the argument of the scaling function f (x), can be
written in terms of another function g[ln(x)] as
−1
1
ρs(T, L) 1 + = g[ln(L) − a/(T − TKT )1/2 ]. (105)
2 ln(L) + c
Here T → TKT corresponds to the argument z = ln(L) − a/(T − TKT )1/2 → −∞.
The KT transition temperature of the 2D XY model has been extracted in different
ways in many studies, e.g., in [133, 137]. Fig. 24(b) shows a test of the scaling form
(105), using TKT = 0.8933, as obtained in [137], and with the constants a and c adjusted
to obtain the (approximately) best data collapse onto a common scaling function for sys-
tem sizes L = 8, 16, . . ., 128. The data are indeed well described by this form. This kind
of plot confirms three different aspects of the KT transition (which were theoretically
deduced at different stages of the history of the KT transition) at the same time; the
exponentially divergent correlation length (102) [135, 136], the logarithmic correction
(101) [138], and the Nelson-Kosterlitz relation (100) [139].
184
3.6. Quantum Phase Transitions
In the following sections of these notes, we will study several examples of quantum
phase transitions, which take place in the ground state of a system as a function of some
model parameter [1]. The scaling properties we have discussed above for classical con-
tinuous and first-order transitions still apply, with some important extensions and modi-
fications. The nature of a classical thermal phase transition can be traced to singularities
in the free energy. At a quantum phase transition, it is instead the ground state energy
which exhibits singular behavior, which is manifested also in other quantities. This can
be understood as arising from the T > 0 free energy of the quantum system, which when
T → 0 becomes the ground state energy. That is of course also true classically, but the
ground states of classical models normally do not evolve continuously as a function of
the parameters (as we saw in the example of the frustrated Ising model in Sec. 3.4.2)
and are therefore strongly first-order. In contrast, as already discussed in Sec. 2.4, quan-
tum systems have non-trivial ground states, with zero-point fluctuations that evolve as
the parameters are varied. Continuous phase transitions driven by divergent quantum
fluctuations are common.
As we will see in connection with quantum Monte Carlo methods in Sec. 5, a d-
dimensional quantum system can formally be mapped, using path integrals, onto an
equivalent classical statistical-mechanics problem in d + 1 dimensions (albeit some
times with a non-positive-definite distribution function). The size of the system in the
new “imaginary time” dimension is the inverse temperature; Lτ = c/T , where c is a ve-
locity. At T > 0, this dimension is finite, while the spatial dimensions can be infinite. The
system is then d-dimensional with a “thickness” Lτ . The strict (d + 1)-dimensionality
applies when also T → 0. In that case, a tunable parameter in the hamiltonian can play
a role very similar to the temperature in a classical system. Interestingly, changing the
temperature in this case is analogous to finite-size scaling in Lτ [5], which can be used
to deduce finite-temperature scaling properties close to quantum-critical points [84].
In some cases, the d-dimensional quantum system has the low-energy properties of the
same kind of classical system in d +1 dimensions. This is the case, e.g., for the dimerized
Heisenberg models discussed in Sec. 2.4. The low-energy physics of these models can
be mapped onto a 3D classical Heisenberg model, which is normally done via continuum
field theories, such as the nonlinear σ -model [5, 84] (and, it, should be noted, the
dimerization is only a means of tuning the strength of the quantum fluctuations, which in
a course-grained uniform effective model is just represented by the coupling constant of
a field theory, or the temperature in an effective 3D uniform classical model [140]). One
can then presume that the quantum phase transition driven by tuning the dimerization
strength should be in the universality class of the temperature-driven transition of the
3D classical Heisenberg model. In other cases, the low-energy mapping may give an
effective system that does not correspond to any familiar classical model, but one can
still say that the system corresponds to some (d + 1)-dimensional effective model.
The dynamic critical exponent. In many cases, such as the dimerized Heisenberg
models mentioned above, the time dimension arising in the mapping to d +1 dimensions
is equivalent (in an asymptotic sense) to the spatial dimensions. The scaling properties
of such a system at a quantum phase transition are then obtained by just replacing d
185
by d + 1 in the critical correlation function (53), in hyperscaling relations such as the
second Eq. (59), and in the scalin forms (98) and (99) of the spin stiffness. In other
cases, the correlations in the new dimension may be fundamentally different from those
in the spatial dimensions. The dynamic exponent z relates the power-laws associated
with spatial and temporal correlations, e.g., if the spatial correlation length ξ diverges
when some parameter g is tuned to its critical value as |g − gc|−1/ν , then the temporal
correlation length diverges as |g − gc|z/ν , and in classical scaling relations one should
replace d by d + z for the quantum critical point.
The dynamic exponent derives its name from the fact that it also governs the dis-
persion ω (q) ∼ qz of excitations of wave-number q. An important aspect of this is
that the finite-size excitation gap is obtained by setting q ∝ 1/L, giving the gap scal-
ing ∆L ∼ 1/Lz . This result can be used directly in numerical calculations, and also has
indirect consequences for the scaling properties of quantities that depend on the ex-
citation spectrum (e.g., various susceptibilities). Extracting the dynamic exponent is an
important aspect of computational studies of quantum phase transitions. Apart from this,
quantum phase transitions, both continuous and first-order ones, can be analyzed with
the same finite-size scaling methods as the classical transitions discussed above.
It should be noted that there is a dynamic exponent also in classical systems, but
this does not come into play in equilibrium statistical mechanics (because if a system
has a kinetic-energy part of the hamiltonian, the associated phase-space probability
distribution cancels out in expectation values). The classical dynamic exponent depends
on how dynamics is introduced into the system [141]. The Metropolis Monte Carlo
algorithm for the Ising model is an example of classical dynamics (called Glauber
dynamics), and the long-time decay of the autocorrelations at criticality is governed
by an associated dynamic exponent. This dynamics does not, however, determine the
equilibrium properties. In quantum mechanics, the dynamics cannot be separated out,
but is an integral part of the equilibrium properties, because the full hamiltonian always
enters and contains in it the static as well as dynamic properties [142].
4. EXACT DIAGONALIZATION METHODS

By exactly diagonalizing its hamiltonian, complete knowledge of a quantum spin system
can be obtained—with the eigenstates available, any static or dynamic quantity can be
computed. In principle, all eigenstates can be computed exactly for a finite quantum
system, by constructing the hamiltonian matrix and diagonalizing it numerically. In
practice, however, such exact diagonalization studies are limited to rather small lattices,
a few tens of spins, because of the exponential increase of the basis size with the number
of spins (2N states in the case of S = 1/2). Great care therefore has to be taken in drawing
conclusions about the thermodynamic limit, which may not even be possible (if the
available lattices are too small to accommodate the infinite-size physics). Insights gained
from exact diagonalization studies are nevertheless very useful, in their own right and as
a complement to other calculations. Exact results for small lattices are also indispensable
for testing the correctness of, e.g., quantum Monte Carlo programs. In addition, exact
diagonalization methods provide a concrete path for learning many important aspects of
quantum mechanics, in particular the symmetry properties of many-body states.
186
FIGURE 25. Schematic illustration of block diagonalization. In the original basis, the hamiltonian has
no apparent structure (left). By constructing states labeled by a conserved quantum number, the matrix
breaks up into blocks (with all matrix elements zero outside the shaded squares) that can be diagonalized
independently of each other (middle). Applying another symmetry (conservation law), the blocks can be
further broken up into smaller blocks (right) labeled by two different quantum numbers, etc.
Block diagonalization. Given a hamiltonian H, the first step of an exact diagonal-

ization calculation is to choose a basis in which it and other operators of interest will
be expressed. The working basis for a spin-1/2 system normally consists of the single-
spin states ↑i and ↓i , i = 1, . . . , N (with the quantization direction normally taken as z).
However, because of the 2N growth of the Hilbert space with the number of spins in the
system, symmetries should be used whenever possible to first reduce the hamiltonian
to a block-diagonal form, as illustrated in Fig. 25. In such a scheme, the spin states are
combined and ordered with the aid of applicable symmetry operations. The blocks corre-
spond to states with different conserved quantum numbers related the symmetries, e.g.,
crystal momentum conservation following from lattice translational symmetry or the
conserved z-component of the total spin (the magnetization). The blocks can be diago-
nalized independently of each other, at a much reduced computational cost. In addition
to the reduced computational effort, immediate access to the quantum numbers is very
useful for classifying excitations.
Some symmetries are relatively easy to take advantage of (e.g., the conservation of the
magnetization) whereas others require some more work and lead to more complicated
basis states (e.g., momentum states). Some symmetries that could be used in principle
are normally not implemented, because the practical complications of the calculation
may outweigh the benefits. Total spin conservation is an example of this.
The use of symmetries in exact diagonalization can be discussed using the language of
group theory [143]. This formalism is not necessary (and often confusing), however, and
here a more practical approach is taken, with no reference to group theory terminology.
Group theory is actually very useful when dealing with complex lattices, but the power
of its formalism can perhaps be better appreciated after a thorough understanding of
symmetry operations and block-diagonalization has been gained through less formal
methods for simple lattices. Here we consider 1D chains and 2D simple square lattices.
Outline of this Section. Our discussion will first be framed around the S = 1/2
Heisenberg chain as a concrete example in Sec. 4.1. We first discuss its symmetries and
introduce a computer representation of the basis states using bits of integers in 4.1.1. In
4.1.2 we use this representation in pseudocodes to construct the full hamiltonian without
any symmetries, and then block-diagonalize using magnetization conservation. Momen-
tum states are discussed in 4.1.3, the use of parity (reflection symmetry) in 4.1.4, and
187
spin-inversion symmetry in 4.1.5. Complete diagonalization to obtain finite-temperature
thermodynamic properties is illustrated with some results in 4.1.6. To compute a small
number of eigenstates for larger chains, the Lanczos method is developed in 4.2. The
utility of this technique for several 1D systems is illustrated in Sec 4.3. The ground state
and low-energy excitations of the Heisenberg chain are discussed in 4.3.1. The dimer-
ization transition taking place in the presence of a frustrating second-nearest-neighbor
interaction is investigated in 4.3.2, and in 4.3.3 an extended model with long-range in-
teractions is considered. Diagonalization of 2D systems is briefly described in 4.4. Mo-
mentum states and other square-lattice symmetries are considered. Results pertaining to
the Néel ground state of this system are discussed.
4.1. Diagonalization of the Heisenberg chain

We will study the S = 1/2 antiferromagnetic Heisenberg chain with hamiltonian
N−1 N−1
H=J ∑ Si · Si+1 = J ∑ [SizSi+1
z
+ 12 (Si+ Si+1
−
+ Si− Si+1
+
)], (106)
i=0 i=0
where, for reasons that will become apparent below, it will here be practical to label the
spins i = 0, . . . , N − 1. Periodic boundaries, SN = S0 , will be assumed when we consider
momentum states, but before that the boundary condition is arbitrary.
4.1.1. Representations of states and symmetries

Lattice transformations. We use the standard notation |S0z , . . . , SN−1 z
i for the basis
z
states, with Si from left to right always corresponding to the spin states on lattice sites
numbered 0, 1, 2, ..., irrespective of the ordering of the subscripts i. Thus, if we write a
state |S1z , S0z i, this is different from |S0z , S1z i unless S0z = S1z . The former could refer to the
state obtained from the latter when the two spins are switched by a permutation operator
P, which can be operationally defined so as to affect the site indices; P|S0z , S1z i = |S1z , S0z i,
e.g., P| ↑↓i = | ↓↑i. Generalizing this to N-spin lattice transformations, reflections, trans-
lations, and rotations (in two and three dimensions) are defined in terms of permutations
of the spin indices. As an example, for a periodic chain we define the translation operator
as moving the spins one step cyclically to the “right”;
T |S0z , S1z , . . ., SN−1

z z
i = |SN−1 , S0z , . . . , SN−2
z
i. (107)
This corresponds to decreasing the spin index by one (modulus the system size N) at
each location i in the ket: Siz → Si−1
z
. When writing specific states with up and down
spins denoted by ↑ and ↓, e.g., | ↑↓↑↓ . . .i, indices are normally not needed (and, for, the
sake of compactness of the notation, the arrows are also not be separated by commas).
The Heisenberg hamiltonian (106) with periodic boundary conditions is invariant with
respect to translations, i.e., it commutes with T ; [H, T ] = 0. We can therefore construct
188
momentum states |Ψ(k)i, which by definition are eigenstates of the translation operator,
T |Ψ(k)i = eik |Ψ(k)i. (108)
Here the allowed momenta are k = 2nπ /N, with n = 0, . . . , N −1, following from the fact
that T N = 1. States with different k form their own individually diagonalizable blocks
of the hamiltonian. How the momentum states are constructed in practice and used in a
computer program will be discussed in Sec. 4.1.3.
The Heisenberg hamiltonian also commutes with the reflection (parity) operator,
which we will define in a way generalizing the two-spin permutations already consid-
ered, in terms of the spin index transformation i → N − 1 − i;
P|S0z , S1z , . . . , SN−1

z z
i = |SN−1 , . . ., S1z , S0z i. (109)
For an eigenstate of P, T |Ψ(p)i = p|Ψ(p)i, where p = ±1 since P2 = 1. We will use

T and P for block-diagonalization, although, as we will discuss in detail in Sec. 4.1.4,
they cannot always be used simultaneously because [T, P] = 0 only in a sub-space of the
Hilbert space. For a system with open boundaries, T is not defined, but P can be used.
Spin quantum numbers. Since the hamiltonian is spin-rotationally invariant, its

eigenstates also have to be eigenstates of the square S2 of the total spin, where
N−1
S= ∑ Si. (110)
i=0
For an eigenstate we have S2 |ψ (S)i = S · S|ψ (S)i = S(S + 1)|ψ (S)i. Here there is
potential for confusion, as the same symbol S is used for both the spin magnitude of
the individual spins (i.e., Si = S) and the total spin of a many-body state. The context
should always make the meaning clear (and we anyway only consider Si = 1/2 here).
With the total spin conserved, we know that the states can form blocks labeled by the
quantum numbers (S, mz), where mz ∈ {−S, −S + 1, . . . , S} is the total magnetization in
the direction of the quantization axis,
N−1
mz = ∑ Siz. (111)
i=0
If we use momentum states, each k-block of course also splits into smaller blocks,
(k, S, mz), because [T, S] = 0. It is easy to block diagonalize using mz , but implementing
the conservation of total S is normally cumbersome (except for a very small number of
spins) and therefore rarely used (although S = 0 states in the valence-bond basis are some
times used [144]). We will use mz conservation in combination with lattice symmetries.
Note that mz conservation is more general than conservation of the total S—even if we
introduce some anisotropy in the hamiltonian by giving a different prefactor to, e.g., the
Ising (diagonal) term in the hamiltonian (106), mz is still conserved although total S is
not. All the techniques discussed in this section can therefore be applied directly also to
such anisotropic models.
189
For the special (and most important) case mz = 0 (for even N), we can block-
diagonalize using a discrete subset of all the possible rotations in spin-space; the spin-
inversion symmetry, i.e., invariance with respect to flipping all the spins. This is defined
formally by an operator we call Z;
Z|S0z , S1z , . . . , SN−1

z
i = | − S0z , −S1z , . . ., −SN−1
z
i. (112)
For this operator we again have Z 2 = 1 and the eigenvalues z = ±1. Since Z commutes
with both P and T , it can be used together with these operators to further block-
diagonalize H, which we will do in Sec. 4.1.5.
The total-spin operator S2 can be written in a form resembling the Heisenberg model
with equal interactions among all the spins;
N−1 N−1
3
S2 = ∑ ∑ Si · S j = 2 ∑ Si · S j + 4 N. (113)
i=0 j=0 i< j
Constructing the matrix form of this operator, which we need for computing the quantum
number S, is therefore very similar to constructing the hamiltonian matrix.
Bit representation of spin states. S = 1/2 models are special because ↓ and ↑ spins
can be represented directly in the computer by the bit values 0 and 1 of an integer. We
will take advantage of this here. The bits are conventionally labeled starting from 0, and
that is why we also number the spins in that way. To refer to the bits i = 0, . . ., 31 of
an integer s (or i = 0, . . . , 63 for a “long” integer), we will use the notation s[i]. A basis
state |S0z , . . . , SN−1
z
i for N spins is thus represented in the computer by an integer s with
z
s[i] = Si + 1/2 for i = 0, . . ., N − 1 and s[i] = 0 for i > N − 1.
Most computer languages have functions for examining and manipulating bits. The
off-diagonal terms of the Heisenberg hamiltonian (106) flip two spins. In pseudocodes
we will accomplish this operation using a bit function flip(s, i, j), which flips (0 ↔ 1)
bits i and j of the integer s representing the state. How this is implemented in practice
depends on the language used. One possibility is to use a bitwise exclusive-or operation
with a mask, as illustrated in Fig. 26. In Fortran 90, this operation can be implemented
as ieor(s, 2i+ j ). Later, we will also need functions that accomplish the various symmetry
transformations of the states; translation, reflection, and spin-inversion.
With standard 4-byte integers, the bit representation with a single integer works up to
N = 32. Using long integers, one can extend the scheme up to N = 64. The latter is well
beyond the maximum size for which exact diagonalization techniques can be used in
practice, except for magnetization sectors with 2mz = n↑ − n↓ large enough for the block
size N!/(n↑ !n↓ !) to be manageable. Magnetization mz = 0 and other low-mz sectors are
typically of primary interest, however.
Discussing algorithms for constructing the basis set and the hamiltonian matrix,
we will start from the simplest method, in which no symmetries at all are employed,
and then implement the magnetization conservation. In Secs. 4.1.3-4.1.5 we include
more symmetries. The actual diagonalization of the hamiltonian matrix and the use its
eigenstates to calculate physical observables will be deferred to Sec. 4.1.6.
190
i j j i
0 1 0 1 0 0 1 1 s = 83
i j
0 0 0 1 1 0 0 0 f = 2i+j = 24
0 1 0 0 1 0 1 1 XOR(s,f)
01234567 7 6 5 4 3 2 1 0
FIGURE 26. The top line shows an N = 8 spin state |si and its representation as the eight first bits of
an integer s. Note that we label spins i = 0, . . . , N − 1 from left to right, while the bits are conventionally
labeled from right to left, as a binary number. To flip two spins i, j, a bitwise exclusive-or (XOR) operation
with a mask f (middle line) can be used. Bits i and j of f are set to 1, and all other bits are 0 (i.e., f = 2i+ j ).
The bottom line shows the outcome of the XOR operation [ieor(s, f ) in Fortran 90].
4.1.2. Computer generation of the hamiltonian

If we do not make use of any conservation laws, the hamiltonian consists of a single
2N × 2N matrix. We then simply use the numbers a = 0, 1, . . ., 2N − 1 to label the basis
states. The bit-values of these integers correspond directly to the spin states. Determining
the diagonal contributions ha|SizS1+1
z
|ai = ±1/4 to the hamiltonian matrix just involves
examining the corresponding bit pairs a[i], a[i + 1] (same or different), whereas an
off-diagonal operator (Si+Si+1−
+ Si− Si+1
+
)/2 acting on a state |ai with a[i] 6= a[i + 1]
generates the state |bi where the two bits have been flipped. The matrix element is then
hb|H|ai = 1/2. For a state with a[i] = a[i + 1] there is of course no off-diagonal matrix
element. The following piece of pseudocode generates the full hamiltonian matrix H:
do a = 0, 2N − 1
do i = 0, N − 1
j = mod(i + 1, N) {2}
if (a[i] = a[ j]) then
H(a, a) = H(a, a) + 14
else
H(a, a) = H(a, a) − 41
1
b = flip(a, i, j); H(a, b) = 2
endif
enddo
enddo
Here j = mod(i +1, N) is the “right” nearest neighbor of i, with the mod function taking
care of the periodic boundary. In Fortran 90, the test for a[i] = a[ j] can be implemented
with the boolean function btest(a, i) to examine bit i of a. Each bond operator in the
hamiltonian corresponds to a single off-diagonal matrix element, while the diagonal
elements have N different contributions. Matrices corresponding to other operators of
interest can of course be constructed in an analogous way, by examining and flipping
bits according to whatever combinations of Siz and Si± operators that are involved.
Using fixed-magnetization blocks. Moving up in sophistication, we next implement

magnetization conservation. We want to construct the block hamiltonian acting on all
191
states with given mz = (n↑ − n↓ )/2. There are M = N!/(n↑ !n↓ !) such states, and we
need a list of them. The order within this list will be used as a label of the states
of the block, i.e., we write |ai for the a:th state in the list. We then also need a list
of integers {sa }, the bits of which represent the spin configuration of the a:th state.
Later, we will have to search the list sa to find the position label a of a particular given
state-integer s, and it is therefore practical to make the list ordered; sa < sa+1 . We will
some times use the notation |sai instead of |ai when referring explicitly to the spins,
|sai = |sa [0] − 1/2, . . ., sa [N − 1] − 1/2i. The context will make it clear if a label inside
|i refers to the position in the list of M states or to the integer containing the spins.
To construct the state list, we loop over the integers s = 0, . . ., 2N − 1 and check
whether the number of set bits (the number n↑ of spins ↑ in the state) corresponds to
the target sector; n↑ = mz + N/2. After initializing a state counter a = 0, we can use the
following pseudocode to generate all the states with given magnetization:
do s = 0, 2N − 1
if (∑i s[i] = n↑ ) then a = a + 1; sa = s endif {3}
enddo; M = a
We now have a basis of size M stored as the integers sa , a = 1, . . . , M.
To construct the hamiltonian, we loop over the labels a = 1, ..., M and use bit op-
erations as before to act on the corresponding state-integers sa . When an off-diagonal
operation on |sa i leads to another state |s∗ i = |sb i, and have to find the position b of the
integer s∗ in the state list. Since this list is ordered, we can do this by a bisectional search
in ∼ log2 (M) steps on average. Such a search proceeds through a series of bracketings,
where in each step we can halve the possible range of b by examining the state at the
mid-point of a range [bmin , bmax ], with the brackets bmin = 1 and bmax = M initially. The
following subroutine finds the position b of a state-integer s∗ ;
subroutine findstate(s∗ , b)
bmin = 1; bmax = M {4}
do
b = bmin + (bmax − bmin )/2
if (s∗ < sb ) then
bmax = b − 1
elseif (s∗ > sb ) then
bmin = b + 1
else
exit
endif
enddo
Division of an integer by 2 should here be regarded in the standard way, i.e., i/2 for odd
i equals (i − 1)/2. The exit from the loop occurs when the basis state sb equals the target
state s∗ . One can also use more efficient search procedures, using so-called hash-tables
[145]. The bisection is much simpler, however, and sufficiently fast in most cases.
Using subroutine {4}, the part of the hamiltonian originating from operation on a spin
pair (i, j) in the state |ai can be constructed with the following modification of code {2};
192
if (sa [i] = sa [ j]) then
H(a, a) = H(a, a) + 41 {5}
else
H(a, a) = H(a, a) − 41
s∗ = flip(sa , i, j); call findstate(s∗ , b); H(a, b) = 1
2
endif
If one is just interested in obtaining quick results for some very small lattice, or if the
system is not periodic (an open chain or a system with random couplings, in which case
the momentum is not conserved), it may be sufficient to construct the hamiltonian in
this form and proceed to diagonalize it (successively for all the mz -blocks desired). For
serious work on translationally invariant (periodic) systems, it is worth implementing
additional symmetries to further block-diagonalize the fixed mz blocks.
4.1.3. Momentum states

We now construct eigenstates of the translation operator T defined Eq. (107). Trans-
lating N steps brings the spins back to their original state. Thus, T N = 1, which implies
eigenvalues eik , where the set of N non-equivalent momenta can be chosen as,
2π
k=m , m = −N/2 + 1, . . . , N/2, (114)
N
with the lattice constant equal to 1. A momentum state can be constructed using a
reference state |ai (a single state in the z-component basis) and all its translations;
1 N−1 −ikr r
|a(k)i = √ ∑ e T |ai.
Na r=0
(115)
It can easily be verified (by a shift of the summation index allowed due to the peri-
odic boundaries) that operating with the translation operator (107) on this state gives
T |a(k)i = eik |a(k)i, which is the definition of a momentum state.
To construct the momentum basis for given k (and normally also given mz ) we have
to find a set of representatives resulting in a complete set of normalizable orthogonal
states. Clearly, for two states |a(k)i and |b(k)i to be orthogonal, the corresponding
representatives must obey T r |ai = 6 |bi for all r. Therefore, among all the states of the
set of translated states |a(r)i = T r |ai, r = 0, . . . , N − 1, only one should be used as a
representative. With the labels referring to the bit representation, it will be practical to
always choose the representative as the one for which the integer a(r) is the smallest (as
determined by carrying out all translations).
Normalization and excluded representatives. If all the translated states T r |ai are
distinct, the normalization constant in (115) is just Na = N. Some reference states have
periodicities less than N, however, and this affects the normalization. The periodicity of
193
a state is defined as the smallest integer Ra for which
T Ra |ai = |ai, Ra ∈ {1, . . ., N}. (116)
If R < N, then there are multiple copies of the same state in the sum in (115), and the
normalization constant must be modified accordingly. We could then also restrict the
summation in (115) to r = 0, . . . , Ta − 1, but it is more practical in formal manipulations
of the states to keep all N terms regardless of Ra .
An important aspect of the momentum basis is that the periodicity of the representa-
tive has to be compatible with the momentum in order for (115) to be a viable state. The
compatibility is related to normalizability. The sum of phase factors associated with the
representative state |ai in the sum in (115) is
N/Ra −1
N/Ra , if kRa is a multiple of 2π ,

F(k, Ra ) = ∑ e −iknRa
=
0, otherwise.
(117)
n=0
The normalization constant is then
Na = ha(k)|a(k)i = Ra |F(k, Ra )|2, (118)
and, therefore, if F(k, Ra ) = 0, no state with momentum k can be defined using the
reference state |ai. Thus, for given |ai the allowed momenta are those for which kRa is
a multiple of 2π , or
2π
k= m, m = 0, 1, . . ., Ra − 1. (119)
Ra
For the allowed momenta, the normalization constant in (115) is
N2
Na = . (120)
Ra
When the reference state is not equal to any non-trivial translation of itself, then Ra = N
and Na = N. Note again that a given reference state can only appear in a single basis
state. It is then clear that the momentum states are orthonormal; hb(k′ )|a(k)i = δab δkk′ .
Also note again that reference states that are not compatible with a given momentum will
not appear in that block of states. An important aspect of constructing the momentum
basis is to check the compatibility of a potential representative state with the momentum.
The hamiltonian matrix. Next, we construct the hamiltonian matrix in the momen-
tum basis. The periodic Heisenberg hamiltonian under consideration here is translation-
ally invariant and consists of N bond operators Si · Si+1 . It is convenient to lump all the
diagonal terms together and consider the off-diagonal terms separately. To simplify the
formalism to follow, we define operators accordingly;
N
H0 = ∑ Szj Szj+1, (121)
j=1
1 + − + −
Hj = 2 (S j−1 S j + S j−1 S j ), j = 1, . . . , N, (122)
194
so that H = J ∑Nj=0 H j . We now set J = 1. We need to find the state resulting when H
acts on the momentum state (115). Since [H, T ] = 0 we can write
1 N−1 −ikr r 1 N N−1

H|a(k)i = √ ∑
Na r=0
e T H|ai = √
Na
∑ ∑ e−ikr T r H j |ai, (123)
j=0 r=0
and we need to operate with the hamiltonian operators H j only on the reference state.
For each operation we get a different state, or, in the diagonal ( j = 0) case, the same
state. In either case we can write H j |ai = h j (a)|b′j i, where h j (a) is the matrix element
coming from (121) or (122), and we do not, for simplicity of the notation, include any
explicit indicator that |b′j i also depends on |ai. The prime in |b′j i is there to indicate
that this new state is not necessarily one of the reference states used to define the basis
and, therefore, a momentum state should not be written directly based on it. Provided
that |b′j i is compatible with the momentum, there must be a reference state |b j i which is
related to it by some number of translations;
|b j i = T l j |b′j i, (124)
and using this relation we have
H j |ai = h j (a)T −l j |b j i, l j ∈ {0, 1, . . ., N − 1}. (125)
Here the notation is again simplified by not making explicit that l j depends on the actual
state |b j i (and therefore on |ai). We can now write (123) as
N h j (a) N−1 −ikr (r−l j )
H|a(k)i = ∑ √ ∑ e T |b j i,
Na r=0
(126)
j=0
and by shifting indices in the summation, and also noting that |b j i may have a different
normalization factor (periodicity) than |ai, we obtain
s
N Nb j
H|a(k)i = ∑ h j (a)e −ikl j
|b j (k)i. (127)
j=0 Na
We can now simply extract the matrix elements of the hamiltonian operators H j ;
s
Nb j
hb j (k)|H j |a(k)i = h j (a)e−ikl j . (128)
Na
Strictly speaking, for the off-diagonal bond operators ( j > 0), this is not the only matrix
element, because individually these are not translationally invariant operators. They
therefore have matrix elements also between states with different momenta. Here we
have in mind summing over all j, after which only the elements diagonal in k survive.
There can be several terms in H that contribute to the same matrix element (128),
because it is possible (in fact very likely) that H j |ai ∝ T −li |bi i and H j |ai ∝ T −l j |b j i
195
with |bi i = |b j i (but li 6= l j ). Note also that the matrix element h j (a) in (125) is zero
for an off-diagonal operator acting on two parallel spins. It should also be kept in mind
that |b j (k)i may not exist even if h j (a) > 0 in in (125), if |b j i is incompatible with
the momentum. The momentum-state matrix element (128) then does not exist. With
these caveats, which we will deal with in the implementations below, (128) specifies
all the non-zero matrix elements of the hamiltonian. For the specific Heisenberg model
considered here, we can substitute h j (a) by the actual values of the diagonal and off-
diagonal matrix elements and obtain
N
ha(k)|H0|a(k)i = ∑ Szj Szj , (129)
j=1
s
−ikl j 1 Ra
hb j (k)|H j>0|a(k)i = e , |b j i ∝ T −l j H j |ai. (130)
2 Rb j
Matrix elements of other translationally invariant operators (which is what should

be used in the momentum basis) can be obtained in exactly the same way if the mo-
mentum transfered by the operator is 0 (i.e., a sum of identical local operators Oi ,
i = 0, . . . , N − 1). For an operator Aq which transfers momentum q 6= 0 (such as Sqz ,
the Fourier transform of Siz ) the procedures differ only in that the basis states obtained
when operating, Aq |a(k)i, have momentum k + q (and in that case the basis sets for both
momenta involved have to be stored).
Constructing the momentum-state basis in a program. We will now again use a label
a to refer to the position of the state in the basis, and store the corresponding spins in
the form of the bits of integers sa , a = 1, . . ., M. A momentum state is defined in terms
of its representative |sa i, from which the full state is generated using the translation
operator T according to (115). In this case we do not know the basis size M a priori,
because the number of compatible representatives depends on the momentum. If all the
translations of all states were unique, and there were no states incompatible with the
momentum, then the number of states would equal [N!/(n↑ !n↓ !)]/N. However, since
many states have periodicities less than N and some are incompatible with k (unless
k = 0), this is only an approximate basis size. The k = 0 momentum block is always
the largest, because there are no states incompatible with this momentum. For large
systems the fractions of disallowed states and periodicities < N are small, and the above
approximation to the basis size is then quite good.
To implement the momentum basis in a computer program, we first need to decide
how the representatives are chosen—in principle the representative of a state |si could
be any one of the members of the group of states related to it through translations;
|s(r)i = T r |si, r = 0, . . . , N −1. Fig. 27 shows an example of all the unique translations of
an 8-spin state in the bit representation. We have to construct a list of the representatives
and should be able to easily identify the representative corresponding to an arbitrary spin
state. As already mentioned, when using the bit representation it is natural to pick as the
representative the translated state |s(r)i for which the integer s(r) is the smallest.
To generate the basis, we loop over the integers s = 0, . . . , 2N − 1, and, as in code {3},
process only those corresponding to a chosen magnetization. The operations needed to
196
r s(r) Tr s
0 153 1 0 0 1 1 0 0 1
1 51 0 0 1 1 0 0 1 1
2 102 0 1 1 0 0 1 1 0
3 204 1 1 0 0 1 1 0 0
4 153 1 0 0 1 1 0 0 1
FIGURE 27. Bit representation of the 8-spin state |si = | ↑↓↓↑↑↓↓↑i (top row) along with its transla-
tions, T r |si, with r = 1, . . . , R, where R = 4 is the periodicity. The integers s(r) correspond to r successive
cyclic permutations (to the left) of the bits. The representative is s(1) = 51; the lowest integer in the set.
determine whether a state is a valid new representative are carried out by bit operations.
When translating |si, if some s(r > 0) < s = s(0), then the representative of |si is
already in the basis and should not be used again [as in Fig. 27, where the original
integer s = s(0) = 153 but the translation s(1) = 51 < s(0)]. If all s(r) ≥ s(0), then the
representative of |si is not yet in the list. However, |si may still not be allowed, due to
its periodicity potentially being incompatible with the momentum, according to (119).
In the pseudocode segments below, the momentum will be represented just by the
integer ∈ {−N/2 + 1, . . . , N/2} multiplying 2π /N in (114), which we here call k. We
will define a subroutine which checks whether or not a state-integer s is a new valid
representative and, for a valid representative, delivers its periodicity R. To carry out the
translations of the state |si, we use cyclic permutations of its bits. We assume that this
is accomplished with a function cyclebits(i, n), which performs a cyclic permutations
to the “left” (as exemplified in Fig. 27) of the n first bits of the integer i. In Fortran
90, exactly this operation is available with the intrinsic function ishftc(s, d, n), which
cyclically right-shifts the first n bits of s by d steps. The definition of T according to
(107) corresponds to d = −1. Our state checking subroutine can be implemented as:
subroutine checkstate(s, R)
R = −1; t = s {6}
do i = 1, N
t = cyclebits(t, N)
if (t < s) then
return
elseif (t = s) then
if (mod(k, N/i) 6= 0) return
R = i; return
endif
enddo
Here the integer R is first initialized to −1, and this value will be returned (and later
used as an indicator of a disallowed representative) unless the input state-integer s is
the smallest among all the translated integers and the periodicity is compatible with
the momentum (and then R will equal the periodicity upon return). We can now easily
construct the list of M basis states;
197
do s = 0, 2N − 1
call checkstate(s, R) {7}
if R ≥ 0 then a = a + 1; sa = s; Ra = R endif
enddo
M=a
The list Ra of periodicities will be needed when constructing the hamiltonian.
Constructing the hamiltonian matrix. To generate the hamiltonian, we loop over the
representatives |sa i, a = 1, ..., M. For each of them, we check the bits corresponding to
all nearest-neighbor pairs (i, j). The diagonal matrix element can be handled exactly as
in code {5}. For the off-diagonal part, we need a few minor modifications. After two
spins have been flipped, the resulting state |s′ i is typically not a representative, and we
need to find it using Eq. (124). The matrix element (130) also requires the number l j of
translations used to bring |s′ i to its representative |sb i. We implement these tasks as a
subroutine, the contents of which are rather similar to the subroutine checkstate, code
{6}, that we used when constructing the basis. Given the state-integer s, we translate
its bits in all possible ways and store the corresponding smallest integer r (the potential
representative) found so far, along with the corresponding number of translations, l;
subroutine representative(s′ , r, l)
r = s′ ; t = s′ ; l = 0 {8}
do i = 1, N − 1
t = cyclebits(t, N)
if (t < r) then r = t; l = i endif
enddo
Having found the representative r, we need to locate its position b in the list {sa }. This
is done in the same way as before, with the subroutine findstate, in code {4}. Note,
however, that because of the periodicity constraint imposed by the momentum, it is now
possible that the potential representative r is actually not present in the list. Therefore
findstate is slightly modified, so that b = −1 is returned if there is no element r in the
list, using the following piece of code after the if...endif statements in code {4};
if (bmin > bmax ) then
b = −1; return {9}
endif
For each spin flip, we can now add the contribution to the hamiltonian with the following
code replacing the next-to-last statement in code {5};
s′ = flip(sa , i, j)
call representative(s′ , r, l) {10}
call findstate(r, b)
if (b ≥ 0) H(a, b) = H(a, b) + 21 (Ra /Rb )1/2 exp(i2π kl/N)
The hamiltonian is complex, except when k = 0 or N/2 (actual momentum 0 or π ).
198
4.1.4. Reflection symmetry and semi-momentum states
We will next consider in addition to the translated states T r |ai those that are generated
by the reflection (parity) operator P defined in (109). The operators T and P do not
commute, and so it would at first sight appear that we cannot construct states that are
eigenstates of both operators simultaneously. These operators do, however, commute in
the k = 0, π momentum blocks, as we will show explicitly below. In addition, we will
construct semi-momentum states that are also parity eigenstates for any k. An advantage
of such states is that they (and the hamiltonian) are real-valued, in contrast to the standard
complex momentum states.
States with parity. Consider the following extension of the momentum state (115);
1 N−1 −ikr r
|a(k, p)i = √ ∑ e T (1 + pP)|ai,
Na r=0
(131)
where p = ±1. Clearly, this is a state with momentum k [i.e., it satisfies Eq. (108)], but
is it also an eigenstate of P with parity p? We can check this by explicit operation with
P, using P2 = 1, p2 = 1, and the relationship PT = T −1 P:
1 N−1 −ikr −r
P|a(k, p)i = √ ∑ e T (P + p)|ai
Na r=0
1 N−1 ikr r
= p√ ∑ e T (1 + pP)|ai.
Na r=0
(132)
This is not exactly of the form (131), unless k = 0 or π , for which eikr = e−ikr (i.e.,
the momentum k is equivalent to −k, and there is no directionality associated with the
state). Thus, in these two special cases, parity and translational invariance can be used
simultaneously for block-diagonalization and |a(k, p)i is indeed a momentum state with
parity p (or, in other words, [T, P] = 0 in the sub-spaces with momenta k = 0 and π ).
Semi-momentum states and parity. Except for the special cases k = 0, π , parity
cannot be used to further block diagonalize a momentum block of H. We can, however,
use parity in combination with the momentum in a different way, by mixing momentum
states with ±k. We consider the sum and difference of these states;
1 N−1 σ
|aσ (k)i = √ ∑ C (r)T r |ai,
Na r=0 k
(133)
where σ = ±1 and we have, for convenience, introduced a function Ckσ (r);
cos(kr), σ = +1

σ
Ck (r) = (134)
sin(kr), σ = −1.
We we will here refer to k in (133) as the semi-momentum. Note that σ is not a
conserved quantum number, just an indicator for how the momentum states have been
199
combined into semi-momentum states. Strictly speaking, the special values k = 0, π are
still conventional (crystal) momenta (and the σ = −1 states do not exist for these k) and
only k in the range 0 < k < π should be referred to as a semi-momenta. We will here
consider all 0 ≤ k ≤ π on the same footing (where it should be noted that only half of
the first Brillouin zone is used, as the other half corresponds to the same states). The
normalization constant is different for the special cases;
2 Ra
N 2 gk

N
Na =
Ra ∑ [Ckσ (r)]2 = 2Ra
, (135)
r=1
where we have introduced the factor

1, 0 < k < π ,

gk = (136)
2, k = 0, π .
In practice, gk will not matter here, because we are only considering matrix elements
of H, which are diagonal in k. The g-factors therefore cancel out in the ratios of
normalization constants appearing in the matrix elements.
Checking overlaps, states with the same k but different σ are orthogonal,
1 Ra
ha−σ (k)|aσ (k)i = ∑ sin(kr) cos(kr) = 0,
Na r=1
(137)
and other requirements for orthonormality, haτ (k′ )|aσ (k)i = δσ τ δkk′ , of the semi-
momentum basis can also easily be verified. The advantage of semi-momentum states
is that they are real-valued for all k, in contrast to the complex momentum states.
The following equalities—standard trigonometric identities—are useful when manip-
ulating semi-momentum states;
Ck± (−r) = ±Ck± (r), (138)

Ck± (r + d) = Ck± (r)Ck+(d) ∓Ck∓(r)Ck− (d), (139)
where ± stands for σ = ±1. Using (139) it is easy to see that the Hamiltonian acting on
a semi-momentum state mixes σ = ±1 states. With H j |ai = h j (a)T −l j |b j i we get
s
N
Ra +
H|a (k)i = ∑ h j (a)

±
Ck (l j )|b±j (k)i ∓Ck−(l j )|b∓j (k)i , (140)
j=0 Rb j
from which the hamiltonian matrix elements can be extracted and written as
s
τ σ (σ −τ )/2
Nb j σ τ
hb (k)|H j |a (k)i = h j (a)τ C (l j ). (141)
Na k
Incorporating parity. Since the hamiltonian (141) is not diagonal in σ , τ , the number
of states in a semi-momentum block is twice that in a conventional momentum block,
200
and thus it would appear that there is not much to be gained over complex momentum
states by making hamiltonian real in this way (by making the states real). However, we
can also incorporate parity in a semi-momentum state, by defining
N−1
1
|aσ (k, p)i = p σ
Na
∑ Ckσ (r)(1 + pP)T r |ai. (142)
r=0
This is also a semi-momentum state, because the reflected component P|aσ (k)i =
σ |Paσ (k)i is the semi-momentum state obtained by using the reflected representative
state P|ai instead of |ai. It can also easily be verified that (142) is an eigenstate of the
parity operator for any semi-momentum; P|aσ (k, p)i = p|aσ (k, p)i with p = ±1. This is
simply due to the fact that in (142) the operator (1 + pP) appears before the translation
operators T r , in contrast to the momentum state (131) where (1 + pP) is written after
T r . The hamiltonian is thus diagonal in p, and the number of states in each (k, p) block
is roughly half of that in the original semi-momentum k blocks. We are then back to the
same block size as with the conventional momentum states, but with states with purely
real coefficients.
Orthogonality and normalization of semi-momentum states. Examining the orthog-

onality of the states (142), we have an apparent problem: When T m P|ai = |ai for some
m, the states |a+(k, p)i and |a− (k, p)i are not orthogonal. We can then use Eqs. (138) and
(139) to write the parity-conserving semi-momentum state (142) as a linear combination
of non-parity semi-momentum states (133);
r
± Na +
± − ∓

|a (k, p)i = 1 ± pC (m) |a (k)i − pC (m)|a (k)i . (143)
Na± k k
It is then clear that ha−σ (k, p)|aσ (k, p)i can be non-zero. To look at this more closely,
we first determine the normalization constant Naσ in (142). Note that we have attached
σ as a superscript to indicate that the normalization constant of the parity-conserving
semi-momentum states can, unlike the normalization (135) of the plain semi-momentum
states, depend on σ (in addition to the implicit dependence on k, p). In the case T m P|ai =
6
|ai for all m, the calculation of Naσ is trivial. When T m P|ai = |ai for some m we can use
(143) and the orthonormality of the pure semi-momentum states |aσ (k)i, resulting in
N 2 gk T m P|ai =

1, 6 |ai ∀m,
Naσ = × (144)
Ra 1 + σ p cos(km), T m P|ai = |ai.
In the same way, we can calculate the overlap between the σ = ±1 states when T m P|ai =
|ai for some m and find
ha∓(k, p)|a±(k, p)i = −p, (T m P|ai = |ai for some m). (145)
This is of course under the assumption that both the σ = ±1 states exist (i.e., Naσ 6= 0).
Thus, the σ = ±1 states in the case T m P|ai = |ai differ at most by a sign, and we
should include only one of them in the block of states. For definiteness, we can choose
|a+(k, p)i if Na+ 6= 0 and |a− (k, p)i else. We still have to pay attention to the non-zero
overlap (145) in some formal manipulations with the semi-momentum states, as we shall
see shortly.
201
The semi-momentum hamiltonian. To calculate the hamiltonian matrix elements,
we first note that, instead of (124) for a plain momentum state, now when we act
with a hamiltonian operator H j on |ai we get a state |b′j i which is related to another
representative state |b j i by a number of translations and possibly a reflection:
|b j i = T l j Pq j |b′j i, l j ∈ {0, . . ., N − 1}, q j ∈ {0, 1}. (146)

The result of H acting on a semi-momentum state (or a k = 0, π momentum state with
parity) can therefore be written in in the form
N h j (a)(σ p)q j N−1 σ
H|aσ (k, p)i = ∑ p
N σ ∑ Ck (r + l j )(1 + pP)T r |b j i. (147)
j=0 a r=0
where the representative |b j i is related to H j |ai
H j |ai = h j (a)Pq j T −l j |b j i, l j ∈ {0, . . ., N − 1}, q j ∈ {0, 1}. (148)

Using the relation (139) we can write (147) as
Nbσj
s
N
H|aσ (k, p)i = ∑ h j (a)(σ p)q j Naσ
× (149)
j=0
v
u −σ
u Nb j
cos(kl j )|b j (k, p)i − σ t σ sin(kl j )|b−σ (k, p)i .
σ

Nb j
The ratio of the σ = ±1 normalization constants is 1 if T m P|b j i =

6 |b j i for all m, and
otherwise it can be written as
v
u −σ s
u Nb j 1 − σ p cos(km) | sin(km)|
= = . (150)
Nbσj
t
1 + σ p cos(km) 1 + σ p cos(km)
Using the overlap (145) we can extract the matrix elements. The ones diagonal in σ are,
Nbτ j
s
σ σ
hb j (k, p)|H j |a (k, p)i = h j (a)(σ p) qj
×
Naσ
(
cos(kl j ), 6 T m |b j i,
P|b j i =
cos(kl j )+σ p cos(k[l j −m]) (151)
1+σ p cos(km) , P|b j i = T m |b j i,
whereas the off-diagonal ones are
Nbτ j
s
hb−j σ (k, p)|H j |aσ (k, p)i = h j (a)(σ p)q j ×
Naσ
−σ sin(kl j ),
(
6 T m |b j i,
P|b j i =
−σ sin(kl j )+p sin(k[l j −m])| (152)
1−σ p cos(km) , P|b j i = T m |b j i.
202
r Tr T rP
0 27 0 0 0 1 1 0 1 1 216 1 1 0 1 1 0 0 0
1 54 0 0 1 1 0 1 1 0 177 1 0 1 1 0 0 0 1
2 108 0 1 1 0 1 1 0 0 99 0 1 1 0 0 0 1 1
3 216 1 1 0 1 1 0 0 0 198 1 1 0 0 0 1 1 0
4 177 1 0 1 1 0 0 0 1 141 1 0 0 0 1 1 0 1
5 99 0 1 1 0 0 0 1 1 27 0 0 0 1 1 0 1 1
6 198 1 1 0 0 0 1 1 0 54 0 0 1 1 0 1 1 0
7 141 1 0 0 0 1 1 0 1 108 0 1 1 0 1 1 0 0
FIGURE 28. A state |si = | ↑↑↓↑↑↓↓↓i (upper left) along with all its transformations by the translation,
T , and reflection, P, operator in the bit representation. The corresponding base-10 integer s and its
transformations are also shown–the smallest integer, s = 27, corresponds to the representative state. For
this particular state, T 5 P|si = |si, i.e., m = 5 in the normalization constant (144).
Degeneracies. In the conventional momentum basis, states with ±k are degenerate,

but are orthogonal states (except when k 6= 0, π ). In the semi-momentum basis, ±k states
are really the same state, differing at most (in the case σ = −1) by a sign (and that is
why the semi-momentum Hilbert space only includes 0 ≤ k ≤ π ). The degeneracy has
here been moved to the parity sector, with the p ± 1 states being degenerate (but still
orthogonal) for k 6= 0, π . This can be seen by letting (p, σ , τ ) → −(p, σ , τ ) in the matrix
elements (152), which leaves unchanged the set of matrix elements for τ , σ = ±1. Thus,
in calculations we need to consider only, e.g., p = +1 for k 6= 0, π (while for k = 0, π we
only have σ = τ = 1 and the p = ±1 sectors are not degenerate).
Constructing the basis of semi-momentum states. Generalizing the convention we

used for momentum states, we now choose as the representative of a state |si the state
|s(r, q)i = T r Pq |si for which the corresponding integer s(r, q) is the smallest (where
r ∈ {0, . . ., N − 1}, q ∈ {0, 1}). States can now be disallowed in a given block not only
due to incompatibility with the momentum, but also because of restrictions imposed by
the reflection quantum number p and the non-conserved state label σ . We thus have
to determine which one, or both, of the σ = ±1 semi-momentum states (142) should
be included, according to the conditions discussed above. Fig. 28 illustrates all the
combination of translations and reflections in the bit representation of a state which
is related to its own reflection according to |si = T m P|si, in which case only the σ = −1
or σ = +1 variant of the state should be included in the basis (and the state is compatible
with all momenta, because its periodicity R = N).
When both σ = ±1 states are required, we will store two consecutive copies of the
same representative, so that we can continue to use the location in the list sa as the
label of each state within the block. We then also need to store the σ labels for each
representative, as well as the number of translations m by which the reflection of a state
is brought back onto itself, which is needed in the normalization and the hamiltonian.
As before, we will use a subroutine checkstate to determine whether or not a state
is a new representative to be added to the list. In addition to the periodicity R of a new
representative, this subroutine now also delivers the reflection-translation number m,
when applicable. We can use m = −1 as a flag indicating that there is no m for which
203
T m P|si = |si (and otherwise m ≥ 0). The operations on the reflected state are carried out
after the instructions in {6} have been performed, by first reflecting the spin bits in the
state integer s and storing it as t, t[i] = s[N − 1 − i] for i = 0, . . ., N − 1. This reflection is
accomplished by a function reflectbits(s, N). Then the resulting integer is translated as
before. We thus modify code {6} according to:
subroutine checkstate(s, R, m)
... {11}
t = reflectbits(s, N); m = −1
do i = 0, R − 1
if (t < s) then
R = −1; return
elseif (t = s) then
m = i; return
endif
t = cyclebits(t, N)
enddo
Here ... represents all operations in code {6}. Note that the loop in the code above starts
at i = 0, and the translations are carried out after comparing t with s, since a state can
be its own reflection (in which case m = 0). Also, since we have already determined
the periodicity R of the state, we need to go only up to i = R − 1 in this loop. The
function reflectbits has to be implemented by hand; a corresponding internal function is
normally not available. In Fortran 90 one can use the functions btest(i, b), setbit(i, b),
and clrbit(i, b), to examine, set, and clear individual bits b of an integer i.
A state which has passed the above checks is still not necessarily a valid representative
for both σ = ±1, and the state must therefore be further examined to determine which
(one or both) of the σ = ±1 copies of the representative should be added to the basis.
The following code segment stores the state information for σ = ±1 states that are to be
included in the block based on the criteria discussed above;
call checkstate(s, R, m)
do σ = ±1 (do only σ = +1 if k = 0 or k = N/2) {12}
if (m 6= −1) then
if (1 + σ p cos(ikm2π /N) = 0) R = −1
if (σ = −1 and 1 − σ p cos(ikm2π /N) 6= 0) R = −1
endif
if R > 0 then a = a + 1; sa = s; Ra = σ R; ma = m endif
enddo
Recall that here m = −1 means that there is no m such that T m P|si = |si, whence both
σ = ± states should be included. On the other hand, if there is such an m, then only
one of the states is included—we pick σ = 1 if the normalization constant (144) for
it is non-zero. This is accomplished above by the two consecutive if statements (where
R = −1 is set for an invalid representative). We have to store σ , and since the periodicity
R is positive and non-zero, we can store it and σ jointly as σ R in a list Ra . A more dense
packing of R, σ , m into a single integer is of course also possible.
204
Constructing the semi-momentum hamiltonian. For general semi-momentum, the
same representative can appear once or twice in the list sa , whereas for the special cases
k = 0, π there is always only a single (σ = 1) copy of each representative. We here
construct code that can treat both cases, but simpler code can be written for the two
special momenta. When building the hamiltonian by looping over the state indices a,
with the corresponding representatives sa , we first check the previous and next items in
the list. If the previous item sa−1 = sa , then we skip over this representative, because it
will already have been taken care of in the previous pass through the loop. If the next
representative, sa+1 equals sa , then we store as n = 2 the number of same representatives,
and else set n = 1. This part of the hamiltonian building is accomplished by;
do a = 1, M
if (a > 1 and sa = sa−1 ) then {13}
skip to next a
elseif (a < M and sa = sa+1 ) then
n=2
else
n=1
endif
...
enddo
Here the skip command skips to the next iteration of the loop and ... stands for the bulk
of the loop, the main features of which we discuss next.
Considering first the diagonal matrix elements H(a, a), for a given a we now first sum
up all the diagonal contributions from the bit configurations in sa and store the result as
Ez . Taking into account the possibility of one or two entries of the representative (n = 1
or 2), we can use the following loop to assign the diagonal matrix elements accordingly;
do i = a, a + n − 1
H(a, a) = H(a, a) + Ez {14}
enddo
Turning next to the off-diagonal operations, looping over site pairs (i, j), if the bits
sa [i] 6= sa [ j], then the spins are flipped, resulting in a state-integer named s′ as in code
{10}. We again have to identify the corresponding representative state |ri = T l Pq |s′ i
by carrying out all the symmetry operations to determine the transformation indices l
and q appearing in the hamiltonian matrix elements (151) and (152). The subroutine
accomplishing this is a simple modification of code {8};
subroutine representative(s′ , r, l, q)
... {15}
t = reflectbits(s′ , N); q = 0
do i = 1, N − 1
t = cyclebits(t, N)
if (t < r) then r = t; l = i; q = 1 endif
enddo
205
Here ... stands for the code after the header in {8}. After having called this subroutine,
we can search for the representative r in the list {sa }. This is done in the same way as
we did before, with the subroutine findstate(r, b).
Since the hamiltonian is more complicated than in the standard momentum basis, we
also define a function helement(a, b, l, q), which returns the matrix element according to
(151) or (152). The function takes as input the labels a and b (locations in the matrix) of
the two representatives and the transformation indices l, q delivered by representative.
The matrix element also of course depends on the momentum k and the parity quantum
number p = ±1, which we do not indicate explicitly here. The function is straight-
forward to implement based on (151) and (152), and we do not list any code here.
As with the state-integer sa in code segment {13}, we also here need to take into
account that the representative sb can appear once or twice (indicated below with m = 1
or 2) in the list of representatives. In case there are two copies, we do not know whether
the subroutine findstate has returned its first or second location. Thus, we again need
to examine positions in the list of representatives adjacent to the one delivered. The
following code segment does all that and then assigns the matrix elements:
s′ = flip(sa , i, j)
call representative(s′ , r, l, q) {16}
call findstate(r, b)
if (b ≥ 0) then
if (b > 1 and sb = sb−1 ) then
m = 2; b = b − 1
elseif (b < M and sb = sb+1 ) then
m=2
else
m=1
endif
do j = b, b + m − 1
do i = a, a + n − 1
H(i, j) = H(i, j) + helement(i, j, l, q)
enddo
enddo
endif
This piece of codes replaces the much simpler four-line code segment {10} for the pure
momentum basis. The advantage is that the matrix H is now real-valued.
4.1.5. Spin-inversion symmetry

Spin-inversion symmetry, with the operator Z defined in (112), can be used to reduce
the hamiltonian block size in the magnetization sector mz = 0. The fact that the magne-
tization is conserved implies, however, that there is nothing to be gained by using this
symmetry in the |mz | > 0 sectors (since such a basis would consist of mixed ±|mz | states,
and the blocks of symmetric and anti-symmetric combinations would be of the same size
206
as the original blocks). In models that do not conserve mz , e.g., the transverse-field Ising
model, spin inversion can be exploited for all states. We denote by z the eigenvalue of Z.
Since Z 2 = 1, we again have z = ±1.
Unlike the parity operation considered in the previous section, the spin-inversion
operator Z in (112) commutes with the translation operator, [T, Z] = 0. The associated
quantum number z is therefore conserved together with k in all momentum sectors. In
the magnetization sector mz = 0, we can therefore always split a momentum block into
two smaller ones by using states of the form
1 N−1 −ikr r
|a(k, z)i = √ ∑ e T 1 + zZ)|ai,
Na r=0
(153)
where the normalization constant is easily obtained as
2N 2 T m Z|ai =

1, 6 |ai ∀ m,
Na = × m Z|ai = |ai. (154)
Ra 1 + z cos(km), T
For a hamiltonian operation on |α i resulting in
H j |ai = h j (a)Z g j T −l j |b j i, l j ∈ {0, . . ., N − 1}, g j ∈ {0, 1}, (155)
we obtain the matrix element

s
g j −ikl j
Nb j
hb j (k, z)|H j |a(k, z)i = h j (a)z e , (156)
Na
which is valid for any k. As always, for j = 0 this reduces to just h0 (a).
Spin-inversion symmetry with semi-momentum states. We can also consider semi-

momentum states incorporating spin-inversion symmetry;
N−1
1
|aσ (k, p, z)i = p σ
Na
∑ Ckσ (r)(1 + pP)(1 + zZ)T r |ai, (157)
r=0
These states are eigenstates of Z as well as P.

When calculating the normalization and constructing the hamiltonian we now have
to consider five different types of reference states, depending on which combinations of
symmetry operations transform the reference state into itself;
1) T m P|ai =
6 |ai, T m Z|ai =
6 |ai T m PZ|ai =
6 |ai,
2) T m P|ai = |ai, T m Z|ai =6 |ai T m PZ|ai =6 |ai,
3) T m P|ai =6 |ai, T m Z|ai = |ai T m PZ|ai = 6 |ai, (158)
4) T m P|ai = 6 |ai, T m Z|ai = 6 |ai T m PZ|ai = |ai,
5) T m P|ai = |ai, T n Z|ai = |ai ⇒ T m+n PZ|ai = |ai.
Here the inequalities should hold for all m and the equalities for some m, n. The calcu-
lations follow the procedures of the previous section, and we just list the results. The
207
normalization constants are:

1, 1)
1 + σ p cos(km),

2)

2

2N 
Naσ = × 1 + z cos(km), 3) (159)
Ra g k 
 1 + σ pz cos(km), 4)


[1 + σ p cos(km)][1 + z cos(kn)]. 5)
For cases 2),4), and 5), only one state out of a pair with σ = ±1 should be included in
the basis as the two states within such a pair differ merely by a factor. To be definite, we
can again pick σ = 1 if that makes Naσ > 0 and σ = −1 else. Acting with a hamiltonian
operator on a representative now leads to a new representative |b j i according to
H j |ai = h j (a)Pq j Z g j T −l j |b j i. (160)
The σ -diagonal matrix elements of H are
Nbτ j
s
hbσj (k, p)|H j |aσ (k, p)i = h j (a)(σ p)q j zg j ×
Naσ
cos(kl j ),

1), 3)
σ p cos(k[l j −m])

 cos(kl j )+
1+σ p cos(km) , 2), 5), (161)
 cos(kl j )+σ pz cos(k[l j −m]) , 4),

1+σ pz cos(km)
whereas the ones off-diagonal in σ are,
Nbτ j
s
hb−j σ (k, p)|H j |aσ (k, p)i = h j (a)(σ p)q j zg j ×
Naσ
−σ sin(kl j ),

1), 3),
 −σ sin(kl j )+p

sin(k[l j −m])
1−σ p cos(km) , 2), 5), (162)
 −σ sin(kl j )+pz sin(k[l j −m]) , 4),

1−σ pz cos(km)
These expressions may seem rather complicated, but it should be noted that they are
completely general for 1D systems, not just for the Heisenberg chain considered here.
Once they have been implemented and tested for some model, the code can easily be
reused for other systems.
Program implementation; combining all symmetries. Incorporating spin-inversion

symmetry in the semi-momentum basis with mz = 0, with states |aσ (k, p, z)i defined in
(157), we have to augment the subroutine checkstate in code segment {11}. It should
return the translation numbers corresponding not only to reflection P of a representative,
T m p P|si = |si, but also translation numbers in symmetry relationships involving Z and
PZ; T mz Z|si = |si and T m pz PZ|si = |si. To accomplish the spin inversion in the bit
representation, we define a function invertbits(s, N) which flips all the bits s[i] →
1 − s[i], i = 0, N − 1. This can be easily accomplished without individual bit-level
208
TABLE 1. Size of the k = 0 state blocks for magne-
tization mz = 0 and different parity and spin-inversion
quantum numbers (p, z).
N (+1, +1) (+1, −1) (−1, +1) (−1, −1)
8 7 1 0 2
12 35 15 9 21
16 257 183 158 212
20 2518 2234 2136 2364
24 28968 27854 27482 28416
28 361270 356876 355458 359256
32 4707969 4690551 4685150 4700500
operations as s = 2N−1 − s (for N less than the number of bits in the integers used;
otherwise a different formula has to be used). Codes {6} and {11} can be easily modified
to also check the inverted and reflected-inverted states to determine mz and mzp .
If the modified checkstate returns a potential representative, R 6= −1, then again the
state has to be examined further to determine whether it satisfies the further criteria for a
representative for one or two states σ = ±1. An allowed representative now falls into one
of five different classes depending on its symmetry properties as summarized in (158).
The class, c ∈ {1, . . ., 5} can be determined from the translation integers m p , mz , mzp
delivered by checkstate. After c has been determined, the three translation integers can
be reduced to two; the m and n in the normalization constants (159). The class c can be
packed along with m, n into a single integer, e.g., as m + n(N + 1) + c(N + 1)2 , which is
stored along with the periodicity and σ index packed as σ R as before.
For the construction of the hamiltonian, the subroutine representative, the main parts
of which are described in code segments {8} and {15}, has to be amended further in
order to also return an index g corresponding to the number, 0 or 1, of spin-inversions
needed, along with q reflections and l translations, to transform the state |sigenerated by
a spin flip into the corresponding representative; |ri = T l j Pq j Z g j |si The construction
of the hamiltonian then proceeds exactly as before, with a modification only of the
function helement to include also g as an argument in code segment (5.15). This is
straight-forward, and there is no need to list any code here.
Examples of state block sizes. This is as far as we will go with applying symme-
tries for 1D systems. For the Heisenberg model, the total spin is also conserved, but
incorporating that symmetry in the basis is much more complicated. In principle it can
be done for total spin singlets, using the valence-bond basis [144], but it is rarely done
in practice because the resulting hamiltonian is very dense, unlike the sparse matrices
obtaining with the symmetries implemented here (the number of non-zero elements is
proportional to M × N for an M × M matrix, and N ≪ N). The sparseness will be very
useful in the Lanczos calculations discussed in Sec. 4.2.
Table 1 shows the size of the Hilbert space blocks for several different chain sizes
N when all the symmetries are used. Here the momentum k = 0 and the number of
states is shown for all combinations of the parity and spin-inversion quantum numbers
(p = ±1, z = ±1). The largest block is always the fully symmetric one. As the system
209
size grows, the relative variations in the block size diminish, and for large N all sizes are
rather well approximated by N!/[4( N2 !)2 N]. For k = π , the blocks are approximately of
the same size as in the table for k = 0, and for other momenta, where both σ = 1 and −1
states are allowed, the blocks are roughly twice as large.
4.1.6. Expectation values and thermodynamics

We are now ready to diagonalize the hamiltonian and calculate physical observables.
Here we will first consider complete diagonalization, meaning that we compute all the
eigenvalues and eigenstates of H. In principle, the eigenvalues λn , n = 1, . . ., M, of a
non-singular M × M matrix H can be obtained by solving the secular equation,
det[H − λn I] = 0, (163)
where det[] denotes the determinant and I is the unit matrix. The eigenvectors vn can
subsequently be obtained by solving the linear system of equations
Hvn = En vn . (164)
However, since the secular equation is very complicated for a large matrix, this method
is not used in practice. Most numerical matrix diagonalization methods are based on an
iterative search for a unitary transformation U such that
U −1 HU = E, (165)
where E is the diagonal eigenvalue matrix with Enn = En and, for a complex hermitian
matrix H, the inverse U −1 of the unitary matrix U is the transpose of its complex
conjugate matrix; U −1 = U ∗T . If the matrix is real and symmetric, we have U −1 = U T .
The columns of the diagonalizing matrix U contain the eigenvectors of H. This can be
seen by multiplying (165) with U from the left, giving HU = U H. Since E is diagonal,
the nth column of U on the right side of this equality is multiplied by the nth eigenvalue
(i.e., the matrix element Enn = λn ). In the multiplication by H on the left side, the nth
column of U gives the nth column (HU )n of HU , i.e., HUn = EnnUn , and thus the
eigenvectors vn of H are identified as vn = Un .
The expectation value of some observable (operator) A in an eigenstate of H is given
by the diagonal element of the corresponding matrix transformed to the energy basis;
hn|A|ni = [U −1AU ]nn . (166)
One would typically be interested in the ground state, n = 1 (assuming the eigenvalues
to be ordered from the lowest to the highest), of the block in question (especially the
actual ground state, in the block with the lowest energy of all). For a thermal average,
all eigenvalues and/or eigenstates of all symmetry blocks are needed;
j M Mj
1
hAi = ∑ ∑
Z j n=1
e−β E j,n [U j−1A jU j ]nn , Z = ∑ ∑ e−β E j,n . (167)
j n=1
210
TABLE 2. The eight lowest energies for a 16-site chain with momentum k = 0
in the blocks with parity and spin-inversion quantum numbers (p = ±1, z = ±1).
n E(+1, +1) S E(+1, −1) S E(−1, +1) S E(−1, −1) S
1 -7.1422964 0 -4.9014133 1 -4.1926153 2 -5.7475957 1
2 -6.1223153 2 -4.1067293 1 -3.6537528 0 -4.9014133 1
3 -5.5912905 0 -3.9398439 1 -3.6085498 2 -4.6986358 1
4 -5.0981578 2 -3.7756347 1 -3.2192241 2 -4.1067293 1
5 -4.8142442 0 -3.6808576 1 -3.2129324 0 -4.0007340 1
6 -4.5657878 0 -3.5785191 3 -3.1695648 2 -3.9398439 1
7 -4.3243602 2 -3.3678831 1 -3.1652647 0 -3.7756347 1
8 -4.1926153 2 -3.3605397 3 -3.1169772 2 -3.6808576 1
Here β = T −1 is the inverse temperature (in units where kB = 1), and the index j
collectively denotes the different quantum numbers, mz , k, p, z, of the blocks of size M j .
Matrix diagonalization is a standard linear algebra operation, and sophisticated sub-
routines are available in many software libraries. We will therefore not discuss the inner
workings of matrix diagonalization procedures here. Almost always, a diagonalization
routine delivers the eigenvalues in ascending order in a vector, along with the matrix U
with the corresponding eigenvectors.
The number of operations needed to diagonalize an M × M matrix generally scales
as M 3 , and the memory required for storage is ∼ M 2 (even for a sparse matrix, as
intermediate steps normally do not maintain sparsity). This severely limits the size of the
matrices that can be fully diagonalized in practice. Currently M ≈ 104 can be handled
without too much effort on a workstation, and a few times larger on a supercomputer.
Looking at Table 1, it is then clear that one cannot realistically carry out complete
diagonalizations for Heisenberg chains larger than N ≈ 20. Calculations aiming at just
the ground state, and possibly some number of excited states, can be carried out in other
ways for larger systems, using, e.g., the Lanczos method discussed in Sec. 4.2.
Total spin. It is useful to know the total spin S of the energy eigenstates. Since we do
not use the conservation of S when block-diagonalizing the hamiltonian, we have to cal-
culate S2 using the states. As we have already noted, this operator is formally equivalent
to a Heisenberg hamiltonian with long-range interactions, as written explicitly in (113).
We can use a slightly modified version of the procedures we developed for constructing
the hamiltonian to obtain the matrix for S2 . We then transform it with U , after which we
have S(S + 1) in the form of the diagonal matrix elements as in (166).
Let us look at some results obtained for a 16-site chain in the k = 0 momentum sector.
The eight lowest energies are listed for each of the symmetry sectors (p, z) in table 2,
along with the calculated spin of the states. The lowest-energy state is a singlet with
(p = 1, z = 1). The first excited state of the system is actually not in the k = 0 sector; it
is a k = π triplet with (p = −1, z = −1), at energy E = −6.87210668.
One interesting feature to note in the table is that, in every symmetry sector, all the
states have either even or odd spin. This is due to the spin-inversion symmetry. An mz = 0
state with N sites and total spin S can be written√ as a linear combination of states √ made
up of N/2 − S singlets, (| ↑i ↓ j i − | ↓i ↑ j i)/ 2, and S triplets, (| ↑i ↓ j i + | ↓i ↑ j i)/ 2 [48].
211
Since the singlet is odd under spin inversion but the triplet is even, it follows that a spin
S state of an N-site system has spin-inversion quantum number z = +1 if N/2 and S are
both even or both odd, and has z = −1 else. Thus, even if we do not calculate S we have
some limited knowledge of it also from z. The low-energy states typically also have low
spin, S = 0 or 1, and these can then be distinguished by z alone. Note, however, that the
lowest state in a given (p, z) sector does not necessarily have the lowest (even or odd)
spin, as exemplified in Table 2 by the lowest (p = −1, z = 1) level, which has S = 2, not
S = 0. However, this is not a low-energy state, as there are 11 states with lower energy in
the k = 0 sector (and numerous additional ones in other momentum sectors). The k = 0
state with second-lowest energy also has S = 2, but there are four S = 1 states below it
in other momentum sectors.
Magnetic susceptibility and specific heat. As an example of thermodynamics, let us

calculate two important properties of the Heisenberg chain; the specific heat C and the
magnetic susceptibility χ z . These can be evaluated according to the formulas,
dhHi 1
= 2 hH 2 i − hHi2 ,

C = (168)
dt T
dhmz i 1
χ = hm2z i − hmzi2 ,

= (169)
dh T
which can both be easily derived using (167). In the definition of χ , h is an external
magnetic field in the z-direction, which couples through a term −hmz added to the
hamiltonian. We will here consider the zero-field case only, and thus hmz i = 0.
Both C and χ are special quantities in the sense that they do not depend on the states,
just the energy spectrum. For the Heisenberg model one can calculate these quantities
by just considering the mz = 0 sector, because of the conservation of the total spin S. A
spin S state is (2S + 1)-fold degenerate with magnetization mz = −S, . . .S. In the case
of C, we therefore just have to sum 168 over the mz = 0 levels and weight them by
this degeneracy factor to obtain the full average over all mz . In the case of χz , we can
proceed in the same way, using also the fact that, in a spin-S multiplet, the average
hm2z i = hS2 i/3 = S(S + 1)/3 in (169).
The situation is slightly complicated by the fact that there can be “accidental” de-
generacies not related to any apparent symmetry of the hamiltonian. If such degenerate
states have different S, then the diagonalization procedure will give some random (de-
pending on exactly how the diagonalization is done) mixed spins for the states. One
should then in principle diagonalize S2 in the degenerate subspace to obtain the spin
eigenvalues. However, in practice these accidental degeneracies are very rare. For small
Heisenberg chains they occur only in the k = π , p = 1, z = 1 sector. For N = 10, there
are two such states with E = 0.5, for N = 12 and 14 there are no accidental mixed-S
degeneracies, and for N = 16 there are four mixed levels at energy 2J. With such a small
number of degeneracies, and at such high energies, we can just ignore them.
Fig. 29 illustrates how the results for C and χ converge to the thermodynamic limit
with increasing system size. Going up to N = 16, the results are well converged down
to T /J ≈ 0.25. It is natural that the convergence is more rapid at high temperatures,
because in the limit T → ∞ the spins are statistically decoupled and the system behaves
212
0.5
N=4 0.15
0.4 N=8
N=12
0.3 N=16 0.10
high-T
χ
0.2
0.05
0.1
0.0 0.00
0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
T T
FIGURE 29. Temperature dependence of the specific heat and the magnetic susceptibility for chains of
length N = 4, 8, 12, and 16. The thin curves show the leading high-temperature forms.
as a set of independent spins. Thus the susceptibility per spin for T → ∞ has the single-
spin Curie form; χ → 1/4T . For C, an analytic high-temperature form can be obtained
by calculating the energy exactly for the 2-site system (a single spin has constant energy,
thus giving C = 0), taking the derivative, and keeping only the leading term of an
expansion in T −1 . This gives C → 3/13T 2 when T → ∞. The high-T limits are also
shown in Fig. 29. In the case of C, this form describes the behavior well down to
T /J ≈ 1.5, whereas in the case of χ the agreement is good only at higher temperatures.
When T → 0, both C and χ approach zero exponentially for a finite system, as a
consequence of the finite gap between the ground state and excited states (due to which
only the ground state contributes to thermal averages as T → 0) and, in the case of χ , the
fact that the ground state of the Heisenberg chain for even N is always a singlet (while
for odd N the ground state is an S = 1/2 doublet and χ thus diverges as T −1 ). In the
infinite system, C → 0 but χ approaches a constant non-zero value.
Note that here, as well as in many cases we will encounter later on, when we talk
about T = 0 properties it is important to consider the order in which the limits T → 0
and N → ∞ are taken. The susceptibility of a chain with finite (even) N always vanishes
as T → 0. The temperature at which the exponential drop to zero commences depends
on the finite-size spin gap, which in the case of the Heisenberg model is ∝ N −1 (which
we will also discuss further and illustrate with data in Sec. 4.3.1). Thus, for given N, one
can expect essentially thermodynamic-limit results down to some T /J ∝ N −1 (which
can also be roughly seen in Fig. 29), and if one takes the limit N → ∞ before T → 0,
then the susceptibility remains non-zero at T = 0.
It is not possible to use exact diagonalization results to extrapolate reliably the prop-
erties of the Heisenberg chain to the thermodynamic limit at low temperatures. For the
susceptibility, an extrapolation of exact diagonalization results [146], using a particular
functional form [147], was already attempted decades ago—the result is known as the
Bonner-Fisher susceptibility. This extrapolation used also the known exact value of χ at
T → 0 (N = ∞), which was available from the exact Bethe ansatz solution of the ground
state. The Bonner-Fisher curve has been commonly used in analyzing experimental re-
sults for quasi-one-dimensional antiferromagnets (weakly coupled chains). The most
important feature here is the location and shape of the maximum at T /J ≈ 0.6, which
213
is already well converged for small systems, as seen in Fig. 29. It has for some time
now been possible to use other computational methods to study various properties of
the Heisenberg model on very long chains (as we will discuss in Sec. 5.3.1), and thus
the low-temperature limit can be accessed more reliably. There are also essentially exact
results from field-theoretical studies combined with the Bethe ansatz [7]. These cal-
culations show an anomalous feature; the approach to the T → 0 limit is logarithmic,
and therefore very difficult to reproduce with finite-N calculations. The low-T form
of the Bonner-Fisher curve is therefore incorrect. Experimentally the most important
feature is the broad maximum. Often effects beyond the Heisenberg chain appears at
lower temperatures (e.g., due to inter-chain couplings, spin anisotropies, or disorder).
The spin susceptibilities of some quasi-one-dimensional materials actually follows χ of
the Heisenberg chain down to temperatures as low as ≈ J/50, where the logarithmic
behavior is prominent [8].
4.2. The Lanczos method

As we have seen in the preceding section, a complete diagonalization of the hamilto-
nian (or individual blocks) becomes prohibitively time consuming for S = 1/2 systems
with more than ≈ 20 spins. For higher S, the situation is of course even worse. If we re-
strict ourselves just to the ground state, and possibly a number of low-lying excitations,
we can reach systems roughly twice as large, by using a Krylov-space technique, such
as the Lanczos method.
4.2.1. The Krylov space

The Krylov space is a sub-space of the full Hilbert space, constructed in such a way
that the low-lying states of a hamiltonian H of interest should be well approximated
within it. Consider an arbitrary state |Ψi, e.g., a one with randomly generated vector
elements Ψ(i), i = 1, . . . , M, in the M-dimensional Hilbert space in which H is defined,
and the expansion of this state in terms the hamiltonian eigenstates |ψn i (in order of
increasing eigenvalues, which we here label n = 0, 1, . . ., M − 1). We operate with a
power of the hamiltonian on this chosen state;
En Λ
" #
M−1
cn
H |Ψi = ∑ cn En |ψn i = cmax Emax |ψmax i + ∑
Λ Λ Λ
|ψn i . (170)
n=0 n6=max cmax Emax
If the power Λ is large, the state corresponding to the eigenvalue |Emax | with largest
magnitude (i.e., max = 0 or max = M − 1) will dominate the sum, provided that the
expansion coefficient cmax 6= 0. Hence, acting many times with the hamiltonian on the
state will project out the eigenvector with the eigenvalue Emax . If we want to make sure
that the ground state |ψ0 i is obtained (exactly for Λ → ∞ or approximatively for finite
Λ), we can instead of H Λ use (H − c)Λ , where c is a positive number large enough to
ensure that |E0 − c| > |EM−1 − c|. Here we will in the following assume that such a
constant, if required, has already been absorbed into H.
214
While Eq. (170) is guaranteed to produce the ground state when Λ is sufficiently large,
a more efficient way to construct a state which approaches the ground state as Λ → ∞
is to consider not only H Λ |Ψi, but the whole subspace of the Hilbert space spanned
by the set of states H m |Ψi, m = 0, . . ., Λ. These states can be constructed one-by-one
by successive operations with H on the initial state |Ψi. In this subspace, an optimal
linear combination of vectors approximating the ground state (minimizing the energy
expectation value) exists, and the way to find it is to diagonalize H in the generated sub-
space of Λ + 1 vectors. In addition to projecting out the ground state for relatively small
Λ (often many orders of magnitude smaller than the size M of the full Hilbert space),
this approach can also accurately reproduce a number of low-lying excited states.
The subspace of the Hilbert space obtained by acting multiple times with H on an
initial state is called the Krylov space. We are here discussing hamiltonians in quantum
mechanics, but Krylov space methods of course apply to eigenvalue problems more
generally as well, and are very widely used in many areas of science and engineering.
4.2.2. The Lanczos basis and hamiltonian

In the Lanczos method [148], an orthogonal basis is constructed using linear combina-
tions of the Krylov space states such that the hamiltonian written in this basis is tridiago-
nal. In a standard approach, which will be described below, a basis {| fm i}, m = 0, . . . , Λ,
is first constructed that is orthogonal but not normalized. These states will subsequently
be normalized to yield an orthonormal set {|φm i}, in which the hamiltonian takes a par-
ticularly simple tridiagonal form. We will also discuss a slightly different approach of
directly generating the normalized basis {|φm i}.
Generation of the Lanczos basis. The construction of the basis {| fm i} starts from
an arbitrary normalized state | f0 i, of which it is required only that it is not orthogonal
to the ground state of H (if the goal is to find the ground state). This should be the case
for a randomly generated | f0 i, but one can also start from some vector which is known
to have a substantial overlap with the ground state. The next state is given by
| f1 i = H| f0 i − a0 | f0 i, (171)
where the constant a0 should be determined such that | f1 i is orthogonal to | f0 i. To this
end, we examine the overlap between the two states;
h f1 | f0 i = h f0 |H| f0 i − a0 h f0 | f0 i = H00 − a0 N0 , (172)
where we have introduced notation for the normalization constants and diagonal matrix
elements of the hamiltonian that will also be used for subsequent states;
Nm = h fm | fm i, Hmm = h fm |H| fm i. (173)
For the overlap (172) to vanish we must choose a0 = H00 /N0 . The following, m = 2 state
is written in terms of the two preceding ones as
| f2 i = H| f1 i − a1 | f1 i − b0 | f0 i, (174)
215
where a1 and b0 can be chosen such that | f2 i is orthogonal to both | f0 i and | f1 i. The
overlaps are, using H| f0 i and H| f1 i obtained from Eqs. (171) and (174);
h f2 | f1 i = H11 − a1 N1 , h f2 | f0 i = N1 − b0 N0 . (175)
The appropriate coefficients are thus a1 = H11 /N1 and b0 = N1 /N0 . For all subsequent
iterations, Eq. (174) generalizes to
| fm+1 i = H| fm i − am | fm i − bm−1 | fm−1 i, (176)
and the coefficients rendering this state orthogonal to | fm i and | fm−1 i are generalizations
of the expressions we already found for a0 , a1 , and b0 ;
am = Hmm /Nm , bm−1 = Nm /Nm−1 . (177)
This can easily be checked by direct computation of the overlaps. It remains to be shown
that with these coefficients the state | fm+1 i is orthogonal also to all previous states | fk i
with k < m − 1. In an inductive proof, we can use the fact that all previously generated
states are orthogonal to each other, and obtain,
h fm+1 | fm−k i = h fm |H| fm−k i − am h fm | fm−k i − bm−1 h fm−1 | fm−k i (178)
= h fm | fm−k+1 i + am−k h fm | fm−k i + bm−k−1 h fm | fm−k−1 i = 0,
where we also used H| fm−k i in the form given by Eq. (176).
This iterative procedure (176) is continued until m = Λ, where Λ can be determined
automatically, on the fly, according to some convergence criterion for computed quanti-
ties, as we will discuss below. First we need the hamiltonian matrix elements.
The hamiltonian in the Lanczos basis. Having constructed a set of Λ + 1 Lanczos

vectors | fm i, m = 0, . . . , Λ, the hamiltonian in this basis can be constructed. We can use
the expression for H acting on one of the basis states, obtained from Eq. (176);
H| fm i = | fm+1 i + am | fm i + bm−1 | fm−1 i, (179)
and thus the non-zero matrix elements are
h fm−1 |H| fm i = bm−1 Nm−1 = Nm ,
h fm |H| fm i = am Nm , (180)
h fm+1 |H| fm i = Nm+1 .
The normalized basis states are
1
|φm i = √ | fm i, (181)
Nm
and bm = Nm+1 /Nm according to (177). The non-zero matrix elements are therefore
hφm−1 |H|φm i =
p
bm−1 ,
hφm |H|φm i = am , (182)
hφm+1 |H|φm i =
p
bm .
216
This is a tridiagonal matrix, which can be diagonalized using special methods that are
faster than generic diagonalization algorithms (available in many linear algebra subrou-
tine libraries). The main advantage is not, however, that the matrix can be diagonalized
more easily—it is anyway typically not very large and even if a generic diagonalization
routine is used the time spent on that part of the calculation is negligible. The advan-
tage is that this matrix can be constructed relatively quickly, especially when the matrix
H is sparse and only its non-zero elements have to be considered (stored or generated
on the fly when acting with H on a state). Some times, when pushing the method to
very large matrices, it is also useful that only three of the large state vectors | fm i have
to be stored in the process. The tridiagonality is also very convenient when calculating
spectral functions (dynamic correlations), as discussed, e.g., in Refs. [43, 143].
Note that with the basis states labeled m = 0, . . . , Λ, the size of the tridiagonal matrix
(the actual basis size) is Λ, not Λ + 1, because for m + 1 = Λ in (179) the last diagonal
matrix element generated for use in (183) is aΛ−1 . The state | fΛ i does not have to be
constructed; only the term H| fΛ−1 i is needed to calculate the last coefficient aΛ−1 .
Since the cost of diagonalizing the Lanczos hamiltonian matrix is very low, one can
do that after each new basis state has been generated, and follow the evolution of the
eigenvalues. One can stop he procedure based on some suitable convergence criterion,
e.g., the desired eigenvalues changing by less than some tolerance ε between iterations.
The basis size required for convergence of course depends on the system studied, but
typically, for quantum spin systems, Λ in the range tens to hundreds should suffice.
Alternative formulation with normalized vectors. The generation of the Lanczos

basis is normally discussed in terms of the un-normalized states | fm i, as we have done
above. However, a direct computer implementation of this procedure occasionally leads
to numerical problems, because the normalization constants Nm can become exceedingly
large (if the eigenvalues of H are large). It may then be better to work directly with the
normalized states |φm i (as an alternative to multiplying H by a suitable factor). In fact, in
practice this formulation is even simpler, and the additional computational cost is merely
the normalizations at each step (which is normally small compared to other costs).
We start with a normalized state |φ0 i and generate the second state according to
1
| φ1 i = H|φ0 i − a0 |φ0 i .

(183)
N1
Here N1 is a normalization constant, which is determined by direct computation of the
scalar product of the constructed state H|φ0 i − a0 |φ0 i with itself. Orthogonality with
|φ1 i again requires a0 = H00 . For each following state one can easily show that
1 |γm+1 i
|φm+1 i = √ H|φm i − am |φm i − Nm |φm−1 i = (184)
Nm+1 Nm+1
is orthogonal to all previous states. Here the definitions of am and Nm differ from the
previous ones in Eq. (177);
am = hφm |H|φm i,
Nm = hγm |γm i, (185)
217
where |γm i is the generated state before normalization as in Eq. (184) above. The
hamiltonian matrix elements are
√
hφm−1 |H|φm i = Nm ,
hφm |H|φm i = am , (186)
hφm+1 |H|φm i =
p
Nm+1 .
In this formulation all the stored numbers are well behaved.
Degenerate states. It should be noted that the Lanczos method cannot produce more
than one member of a multiplet; out of a degenerate set of states, only a particular linear
combination of them will be obtained (which depends on the initial state | f0 i). To see
the reason for this, we again look at the expansion (170) of a state H Λ |Ψi, in which we
assume that there are two degenerate states |ψi i and |ψ j i, Ei = E j . In the expansion we
can isolate these states from the rest of the terms;
H Λ |Ψi = Eim (ci |ψi i + c j |ψ j i) + ∑ cm EmΛ |ψm i. (187)

m6=i, j
For any Λ, the expansion contains the same linear combination of the states |ψ j i and
|ψ j i. Hence, in the subspace spanned by the set of states H m |Ψi, m = 0, . . ., Λ − 1, there
is no freedom for obtaining different linear combinations of the two degenerate states.
This of course generalizes also to degenerate multiplets with more than two states.
Loss and restoration of orthogonality. When the basis size Λ becomes large, the
Lanczos procedure typically suffers from numerical instabilities. Round-off errors ac-
cumulated in the course of constructing the basis set will eventually introduce some
non-orthogonality among the states. Such numerical errors can escalate and lead to suc-
cessive sudden appearances (within some narrow ranges of Λ) of several identical eigen-
values (recall that the Lanczos scheme should never produce degenerate states). We will
see an example of this further below.
Loss of orthogonality and the appearance of multiple copies of the same states is nor-
mally not a problem when the aim is to obtain only the ground state. It can complicate
calculations of excited states, however. To remedy this, one can add to the basis con-
struction a step where each new Lanczos vector constructed is explicitly orthogonalized
with respect to all previous basis vectors. In the simplest implementation of such a stabi-
lization procedure, all the Lanczos vectors are stored (in primary memory or secondary
storage). This can become problematic when dealing with very large basis sets, but the
scheme is very simple. Working with the normalized states, each |φm+1 i constructed
according to (184) is orthogonalized with respect to all previous states, according to
|φm+1 i − q|φi i
|φm+1 i → , q = hφi |φm+1 i, (188)
1 − q2
successively for i = 0, . . . , m. This makes it possible to study a much larger number of
excited states (in principle only limited by computer memory).
218
Instead of converging several excited states in the same run, one can also target excited
states one-by-one, starting each time from a vector which is orthogonal to all previous
ones. Re-orthogonalization should then also be done with respect to those.
Eigenstates and expectation values. Diagonalizing the tridiagonal Lanczos hamil-

tonian results in eigenvalues En and eigenvectors vn , n = 0, . . ., Λ − 1. We want these
eigenvectors expressed in the original basis {|ai}, in which we are able to evaluate the
matrix elements hb|O|ai of operators O of interest. First, the Lanczos basis states are
M
| φm i = ∑ φm(a)|ai, m = 0, . . . , Λ − 1, (189)
a=1
and we denote the desired eigenvectors of the hamiltonian as

M
|ψn i = ∑ ψn (a)|ai, n = 0, . . . , Λ − 1. (190)
a=1
The first few eigenvectors vn of the tridiagonal matrix accurately represent eigenstates
of the hamiltonian (essentially exactly for sufficiently large Λ) in the Lanczos basis;
Λ−1 Λ−1 M
|ψn i = ∑ vn(m)|φmi = ∑ ∑ vn (m)φm(a)|ai, (191)
m=0 m=0 a=1
and thus the wave function coefficients we want to construct are given by
Λ−1
ψn (a) = ∑ vn(m)φm(a), a = 1, . . ., M. (192)
m=0
If all the Lanczos vectors have been stored during the basis construction, this transfor-
mation can be carried out in a straight-forward manner. If there is enough computer
memory available, one should store all Lanczos vectors, but often calculations are push-
ing the limits of computer capacity, and then only the bare minimum of information can
be stored, at the cost of longer calculation times. If we did not store the full set of Lanc-
zos vectors during the basis construction, we need to generate them again, in the same
iterative fashion as before, except that we already have the coefficients am and Nm avail-
able (and bm if the scheme with un-normalized Lanczcos vectors | fm i is used) and do not
need to recompute them. We can transform the states according to (192) on the run with
the eigenvectors vm , building up one or several of the eigenstates of the hamiltonian.
Having generated one or several eigenstates, now assumed to be stored in the form of
the coefficients ψm (a) in (190), an expectation value of some operator O can be obtained
by first acting on the state, giving an un-normalized state that we call |ψnO i,
M
O|ψn i = |ψnO i = ∑ ψn (a)O|ai
a=1
M M
= ∑ ∑ ψn (a)|bihb|O|ai. (193)
a=1 b=1
219
M M
= ∑ ψnO (a)|ai, ψnO (a) = ∑ ψn (b)hb|O|ai.
a=1 b=1
We then evaluate the scalar product of this state with the original state |ψn i;
M
hψn |O|ψn i = hψn |ψnO i = ∑ ψn (a)ψnO(a). (194)
a=1
This somewhat cumbersome way of writing the expectation value corresponds to a typ-
ical computational procedure of first computing the state vector ψnO and then computing
its scalar product with ψn .
It is interesting to note that, in Lanczos calculations, we are often dealing with four
different bases: First, we have the original “computational” basis of single states of ↑
and ↓ spins. From these we generate a basis incorporating symmetries, e.g., momentum
states (which normally would be the states denoted as |ai above). We then construct
the Lanczos vectors, which are particular linear combinations of those states. Finally,
diagonalizing the tridiagonal matrix (which is an effective hamiltonian in the low-energy
sector), we obtain the desired energy eigenstates. To do calculations with those states,
we effectively do the basis transformations in reverse. The matrix elements in (193) and
(194) are in the end carried out in the computational basis of ↑ and ↓ spins (i.e., with the
representative states used, e.g., to build the momentum states). We next discuss all these
procedures in practical program implementations.
4.2.3. Programming the Lanczos method

While the Lanczos method can be applied to any symmetric (or hermitian) matrix,
in the case of a spin hamiltonian there is an added advantage in that the hamiltonian is
a sparse matrix. Although the size of the hamiltonian (an individual symmetry-block)
can be very large, the number of non-zero matrix elements is much smaller. For a model
with short-range interactions on a lattice of N sites, a hamiltonian block of M states has
on the order of NM non-zero elements, which for large M is much smaller than the total
number of elements M 2 . Since the most time consuming part of the construction of the
Lanczos basis is the repeated operations with the hamiltonian matrix on a state vector,
to generate the next basis state according to Eq. (176) or (184), the sparseness allows for
enormous time savings. There are similarly significant memory savings advantages as
well. The non-zero elements of the hamiltonian should then be stored in a compact form,
or generated on the fly as needed (which, when many symmetries are used, typically
takes longer than even reading the elements from disk storage). We here implement the
Lanczos method with the hamiltonian stored in a compact form in primary memory.
We will generate the Lanczos basis of states {|φm i} that are normalized at each step,
using Eqs. (184) and (183). In a computer program, these states are stored in the form
of their vector components φm (a), in terms of which the Lanczos states are given by
Eq. (189). Here a = 1, . . . , M labels the states of our working basis, which in the case
of maximal use of symmetries would be momentum or semi-momentum states, e.g.,
|ai = |k, p, z, mz = 0, ai, but for simplicity we do not write out all the quantum numbers.
220
The nature of the basis states only come into play when acting with the hamiltonian (or
some operator to be measured) on the basis states, and for this we need exactly the matrix
elements that have already been discussed for the various symmetry implementations in
Secs. 4.1.3, 4.1.4, and 4.1.5. We will write pseudocodes assuming real-valued wave-
function coefficients; the changes needed for complex states are self-evident.
Lanczos basis construction and eigenstates. To discuss the general structure of

a Lanczos program, we begin by assuming that we have implemented a subroutine
hoperation(φ , γ ) which acts with the hamiltonian on a state vector; |γ i = H|φ i. Later,
we will describe the implementation of this subroutine, which is the only part of the
Lanczos procedure that requires a specification of the model and the symmetries imple-
mented. We also use a subroutine normalize(φ , n), which first computes n = hφ |φ i and
then rescales φ so that hφ |φ i = 1 upon return.
To start the basis generation, we first load the initial state |φ0 i with a randomly
generated normzalized state. We denote the elements φ0 (i), i = 1, . . . , M and hφ0 |φ0 i =
∑i φ0 (i)2 = 1. The next Lanczos state is generated according to (183);
call hoperation(φ0 , φ1 )
a0 = hφ0 |φ1 i; φ1 = φ0 − a0 |φ1 i {17}
call normalize(φ1 , n1 )
If we are not storing all the Lanczos vectors, we can cycle between three vectors,
φ0 , φ1 , φ2 , that contain the information we need at each step when implementing the
iterative basis generation according to (184);
do m = 1, Λ − 1
call hoperation(φ1 , φ2 ) {18}
am = hφ1 |φ2 i; φ2 = φ2 − am φ1 − nm φ0
call normalize(φ2 , nm+1 )
φ0 = φ1 ; φ1 = φ2
enddo
If possible, we should store all the states, to avoid having to regenerate them at the
later stage when computing expectation values (which doubles the computation time).
We have to store all the states if we want to carry out additional re-orthogonalization,
to ensure that numerical truncation errors do not eventually degrade the Lanczos basis.
In that case we insert an additional loop to orthogonalize with respect to all previously
generated states, according to Eq. (188);
do m = 1, Λ − 1
call hoperation(φm , φm+1 ) {19}
am = hφm |φm+1 i; φm+1 = φm+1 − am φm − nm φm−1
call normalize(φm+1 , nm+1 )
do i = 0, m
q = hφm+1 |φi i; φm+1 = (φm+1 − qφi )/(1 − q2)
enddo
enddo
221
After completing either code segment {18} or {19}, we have the contents of the tridi-
agonal matrix in Eq. (187) and can proceed to feed this information into a diagonal-
ization routine which delivers the eigenvalues En and eigenvectors with elements vn (i),
n = 0, . . ., Λ − 1, i = 1, . . ., M. It should be noted again that, with the conventions we
have been using, the Lanczos basis size (the size of the matrix) is Λ; the last state gener-
ated in {18} and {19}, with m + 1 = Λ, is actually not needed (only the coefficient aΛ−1
generated in the last iteration is needed).
The computational effort of the diagonalization is very small compared to the time
spent on the basis construction. We may therefore as well diagonalize after each new
Lanczos vector has been generated. We can then monitor how the energies evolve with
the basis size. One can then stop when some convergence criterion is satisfied. One can,
e.g., demand that the change in the ground state energy (or the highest excitation of
interest) changes between steps m − 1 and m by less than some small number ε . The
Lanczos method is normally capable of converging energies to the numerical precision
of the computer (and double precision should always be used). As we will see below,
other properties converge slower than the energies.
Expectation values. First, let us consider calculations of operator expectation values

(besides the energy) in a program where we have not stored all the Lanczos basis vectors.
Then, in order to be able to transform the ground state from the Lanczos basis to the
original basis, where we can carry out “measurements” on the state, we have to repeat
the Lanczos basis construction starting from the same initial state as in the first run. If
we had initialized with a random φ0 , as in code {17}, and did not save this state, we
can re-generate it by initializing the random number generator with the same seeds as in
the first construction. In any case, with φ0 at hand, we can proceed as in code segment
{18}, with the minor simplification that we now do not have to calculate the coefficients
am , nm , because we already have them. We can use nm for normalizing, instead of calling
the subroutine normalize in {18}. During the basis re-construction, we transform the
states with the ground state vector v0 (or vn for an excited state i) that resulted from the
diagonalization of the tridiagonal matrix, building up the ground state in a vector ψ ;
√
call hoperation(φ0 , φ1 ); φ1 = (φ1 − a0 φ0 )/ n1
ψ = v0 (0)φ0 + v0 (1)φ1 {20}
do m = 2, Λ − 1
call hoperation(φ1 , φ2 ) √
φ2 = (φ2 − am−1 φ1 − nm−1 φ0 )/ nm
ψ = ψ + v0 (m)φ2
φ0 = φ1 ; φ1 = φ2
enddo
The elements of the vector ψ now contain the ground-state wave-function coefficients in
whatever basis is used and implemented through hoperation. If we have stored all the
Lanczos vectors, we of course do not need this step and can directly transform the states
generated in code {19}.
It is particularly easy to calculate expectation values of operators that depend only
on the z-components of the spins and is invariant under all the symmetry operations
222
used in the basis. We then just have to weight the quantity calculated using the rep-
resentative states with the corresponding wave-function coefficient squared (the state’s
probability). As an example, the spin-spin correlation function hSiz Si+r
z
i for all distances
r = 0, . . . , N/2 can be obtained as:
do a = 1, M
do r = 0, N/2 {21}
do i = 0, N − 1
j = mod(i + r, N)
if (sa [i] = sa [ j]) then
C(r) = C(r) + ψ 2 (a)
else
C(r) = C(r) − ψ 2 (a)
endif
enddo
enddo
enddo
C = C/4N
Here ψ 2 (a) is the probability of the state |ai, which when symmetries are incorporated is
represented by a number sa . As before, bit tests can be used to determine the relative spin
orientation, now of spins separated by r lattice spacings. Note that although the final spin
correlation hSiz Szj i only depends on r = |i − j|, we still have to average over all i above,
because we are only using the representative states (which do not by themselves, without
acting with the symmetry operators, obey any lattice symmetries). In principle we should
also average over both reflections and spin-inversions of the representative, but after
the translational averaging has been done the spin correlation function is also invariant
with respect to the other two symmetries (which is not the case for all operators). Note
that since the Heisenberg hamiltonian is spin-rotation invariant, the correlation function
hSizSi+r
z
i calculated here equals hSi · Si+r i/3.
Calculating expectation values of z off-diagonal operators that cannot be simply re-
lated to z diagonal ones require explicit operations on the states, and thereafter eval-
uation of a scalar product, according to Eqs. (193) and (193). This is in principle
easy—essentially proceeding as in the construction of the hamiltonian in the preced-
ing sections—but more time consuming than diagonal operators.
Compact storage of the hamiltonian. We now discuss the inner workings of the
subroutine hoperation(φ , γ ) that we employed in the Lanczos procedures. It implements
the operation
M M
H|φ i = |γ i = ∑ ∑ φ (a)hb|H|ai|bi. (195)
a=1 b=1
This is of course where all the details of the symmetries employed will enter. The main
difference with respect to the construction of the hamiltonian in complete diagonaliza-
tion is that we do not want to store H as a full M × M matrix, because we have in mind
calculations were the number of basis states M can be up to many millions (e.g., for
223
N = 32 in Table 1). We therefore have to devise a convenient way of only storing its
non-zero elements, of which there are of the order NM in (195).
We define the following data structures for the compact hamiltonian storage: For each
a = 1, . . ., M in (195) we store the number ea of non-zero matrix elements hb|H|ai.
We store the locations b of these non-zero elements as consecutive integers in a vector
with elements B(i). In a program, it is convenient to use the elements i = 1, . . ., e1
for a = 1, followed by i = e1 + 1, . . . , e1 + e2 for a = 2, etc. The required size of the
vector B (the number of non-zero elements) is initially not known, and, depending on
the programming language used, may have to be allocated using an estimated size. We
can store the corresponding non-zero matrix elements in a vector with floating-point
values H(i). We can also take advantage of the fact that the hamiltonian is a symmetric
matrix and only store one of its “triangles”.
With the above notation, carrying out the operation (195) is now as simple as:
subroutine hoperation(φ , γ )
γ = 0; i = 0 {22}
do a = 1, M
do j = 1, ea
i = i+1
γ (B(i)) = γ (B(i)) + H(i)φ (a)
γ (a) = γ (a) + H(i)φ (B(i))
enddo
enddo
The counter i keeps track of the position of the elements in the data structures H and
B. Diagonal matrix elements, B(i) = a, are double counted in this procedure since the
contributions from the upper and lower triangle of the hamiltonian matrix are added. It
is therefore assumed that their stored values have been divided by 2.
The matrix elements H(i) and their locations B(i) are generated in a way similar to
what we did when constructing the complete hamiltonian, e.g., as in code segments
{13}–{16} in the semi-momentum basis with parity. We also discussed the simple ex-
tensions involving spin-inversion symmetry. Now we discuss the modifications needed
when we wish to load only the non-zero elements into our compact storage.
In the case of the semi-momentum basis, which we will consider here, the procedures
are again somewhat complicated by the fact that the same representative can appear once
or twice (σ = ±1) in the state list, and we want to take care of these at the same time
in order to avoid repeating tasks unnecessarily. We therefore carry out the loop over the
basis states a and determine the number of same representatives n as in code {13}. We
cannot put the matrix elements directly into the storage vector H, because our scheme
in code {22} requires consecutive storage of all matrix elements for each column. With
the way the individual bond operators in the hamiltonian are treated one-by-one in code
{16}, when n = 2 the two columns would be mixed up if we store each matrix element
as it is generated. We therefore use temporary storage for the one or two columns being
currently processed, and then later copy their contents into the appropriate positions of
the final storage. Referring to the first (a) and potential second (a + 1) column as c = 1
and c = 2, we store the data temporarily as Bc (k) and Hc(k), k = 1, . . ., nc , where nc is
224
the number of elements for the column in question. The diagonal elements (divided by
2, as discussed above) are entered first into these temporary storage lists;
n1 = 0; n2 = 0
do c = 1, n {23}
Bc (1) = a + c − 1; Hc (1) = Ez /2; nc = 1
enddo
For the off-diagonal matrix elements we proceed as in code segment {16}. Note, how-
ever, that several bond operations may lead to the same state |bi when acting on a given
basis state |ai, but in the compact storage we do not want to store individual contribu-
tions to the same matrix element separately. Everything is taken care of by replacing the
loops over i and j in code {16} by this extended version;
do i = a, a + n
do j = max(b, i), b + m {24}
c = i−a+1
E = helement(i, j, l, q)
do k = 1, nc
if Bc (k) = j exit
enddo
if (k > nc ) then nc = k; Bc (k) = j endif
if ( j = i) then Hc (k) = Hc (k) + E/2 else Hc(k) = Hc (k) + E endif
enddo
enddo
Here max(i, b) on the second line ensures that only the elements in matrix triangle j ≥ i
are constructed. The innermost loop over k checks weather there is already a stored
contribution to the matrix element h j|H|ii. It has been assumed that when the loop over
k has been completed (without exiting before k = nc ), then k takes the value nc +1 (which
is the case in many computer languages). Then k = nc + 1 if there is no prior location
j in the list, which means that it should be added to the list (and the size of the list is
then k). We have also assumed that all Hc (k) are initially set to zero, so that each new
contribution E can be added in the appropriate location of Hc . Note that although we are
carrying out off-diagonal operations here, in the basis we are using, such operations can
also lead to diagonal matrix elements, in which case we have to divide E by 2.
After having generated all the matrix elements originating from the current represen-
tative (i.e., completed the loop over all nearest-neighbor spin pairs), we copy the contents
of the temporary storage vectors into the permanent full storage;
do c = 1, 2
do i = 1, nc {25}
nH = nH + 1; B(nH ) = Bc (i); H(nH ) = Hc (i)
enddo
enddo
Here nH is a counter for the total number of non-zero matrix elements added so far.
After this, the loop over state labels a is closed, as in code {13}. This completes the
225
-2
4
n=0
-4 n=1
n=2 3
-6 n=3
En
S
2
-8
1
-10
0
0 10 20 30 40 0 10 20 30 40
Λ Λ
FIGURE 30. Example of the convergence as a function of the Lanczos basis size of the energy (left)
and the total spin (right) of the four lowest levels of the 24-site Heisenberg chain in the symmetry sector
k = 0, p = 1, z = 1. The spin S is extracted using the assumption that hS2 i = S(S + 1).
construction of the hamiltonian.

For other operators, it is not worth storing the matrix elements, because normally we
do not reuse operators for observables many times. The hamiltonian is used repeatedly,
however, up to a few hundred times, and so using code segment {21} instead of having
to carry out all the operations associated with extracting each individual hamiltonian
matrix element from scratch every times can amount to a very significant speed-up.
It is possible to further compactify the hamiltonian by not storing the matrix elements
as double-precision numbers, but instead use a mapping to the actual numbers based on
a table of integers. The number of unique values is often very small (tens or hundreds of
values, so that one can even use “short” integers as pointers to the actual values), and this
can save some memory, at the cost of a somewhat more time consuming construction of
the hamiltonian. The locations B(i) have to be stored as four-byte integers, however.
When pushing the limits of the largest treatable system sizes, one may have to store the
hamiltonian on disk (and reading successive portions of it when executing code {22}),
or generate it on the fly without storing it. The latter essentially amounts to executing
code based on {24} every time when acting with the hamiltonian.
4.2.4. Convergence of Lanczos calculations

The Lanczos method is essentially exact if a sufficiently large number of Lanczos
vectors Λ is used, and typically this number does not even have to be very large; on
the order of a few tens to hundreds. The convergence should be checked by carrying
out calculations for several Λ, until no changes can be detected in the energies and
expectation values of interest. The ground state converges the fastest, and energies
converge faster than expectation values. As an example, Fig. 30 shows results for an
N = 24 chain. The energy and the total spin of the four lowest levels in the symmetry
sector of the ground state are shown versus Λ. The spin quantum number is calculated
by acting with the squared total spin operator S2 on the states, according to Eqs. (193)
and (194), and extracting S assuming S2 = S(S + 1) (which is valid only when the
226
-3 -3
(a) (b)
-4 -4
E -5 -5
-6 -6
-7 -7
0 20 40 60 80 0 20 40 60 80
Λ Λ
FIGURE 31. (a) The four lowest energies as a function of the Lanczos basis size for a 16-site Heisen-
berg chain with quantum numbers (k = 0, p = 1, z = 1). Multiple copies of the same state appear suc-
cessively due to loss of orthogonality. (b) The five lowest states of the same system obtained with re-
orthogonalization of the basis set.
states have converged to eigenstates of the operator). The energies are seen to converge
monotonically, whereas this is not necessarily the case for other quantities, as seen
clearly for the S = 2 states in this case. The details of the convergence of course depend
on the initial state from which the Lanczos basis is constructed (which in this case was
a random state). In this case all the four levels shown (E as well as S) were converged to
better than 10 decimal places at Λ = 60, with the ground state having converged at that
level already at Λ = 30. Going to larger system sizes, the convergence becomes a little
slower, but for this particular model there are no difficulties in converging several levels
up to the largest system sizes that can feasibly be studied.
One can accelerate the convergence of a Lanczos calculation by starting from a
state which is already close to the ground state. Such states may be constructed in a
number of ways, e.g., based on some approximate analytical method. But if there are no
convergence problems this may not be worth the additional effort. However, if a series
of calculations are carried out as a function of some parameter in the hamiltonian, then
subsequent calculations can be started from the ground state of the preceding parameter
value, which is likely to have a significant overlap with the next ground state. However,
it should be noted that if the initial state is a good approximation to the ground state,
it will have very small overlaps with the first few excited states, and hence only the
ground state is likely to converge rapidly in such a calculation. If excited states are also
needed, this problem can be circumvented by starting the next calculations using a linear
combination of eigenstates from prior calculations.
Loss and recovery of orthogonality. The Lanczos basis vectors should all be com-
pletely orthogonal to each other, but numerical truncation errors build up and eventually
lead to escalating loss of orthogonality for some Λ. This manifests itself as artificial de-
generacies, with excited states “falling down” onto lower states. An example of this is
shown in Fig. 31(a), where the four lowest energies of a 16-site chain in the ground-state
symmetry sector are graphed versus Λ. The higher energies are seen to successively col-
lapse onto the immediately lower energies, with only a few iterations taken for the levels
227
to become completely degenerate after signs first appear of the imminent collapse. The
basis then includes multiple copies of the same states.
The loss of orthogonality may at first sight seem like a serious problem, but in practice
this is no necessarily the case. Calculations for the ground state are not much affected
by there being more than one copy of it, as long as one makes sure that it is properly
normalized before using it to calculate expectation values, and the first excited states are
normally (but not always) well converged before they fall down onto lower states.
Loss of orthogonality often occurs sooner for small systems than large ones. This is
related to the fact that there are less low-energy states for small systems, whence the
lowest states in the Lanczos basis can become “overcomplete” in this subspace (if the
Lanczos basis is larger than the total number of states, it clearly is truly overcomplete
and the calculation will not work).
One can easily supplement the Lanczos method by an explicit re-orthogonalization
step, as in Eq. (188) and implemented in code {19}. The drawback is that this requires
storage of all the Lanczos basis vectors, which may not be possible for large systems.
The procedure may also become time consuming if the number of states is large. As
shown in Fig. 31(b), re-orthogonalization enables convergence of many more eigenstates
(limited only by memory and time constraints).
4.3. 1D states and quantum phase transitions

In this section we discuss several related 1D calculations using the Lanczos method
and finite-size scaling methods for up to N = 32 spins. First we investigate the critical
ground state of the standard Heisenberg chain, and also analyze some properties of
excited states. Then we introduce frustration, studying the dimerization transition in the
J1 -J2 chain. Finally, we add long-range interactions in addition to frustration, in which
case the continuous dimerization transition evolves into a first-order transition between
a Néel state (which is possible even in a 1D system if the interactions are sufficiently
long-ranged) and a dimerized VBS state.
4.3.1. Ground state and excitations of the Heisenberg chain

Although the Heisenberg chain has an exact Bethe ansatz solution [51], the wave
function is very complicated and in many cases numerical calculations for finite-size
systems have to be used to extract information on physical properties from it [53, 54].
The energy of the ground state [52] and the low-lying excitations [149] can be calcu-
lated exactly both for finite chains and in the thermodynamic limit, however. Some other
quantities can also be extracted for very large chains [150]. In addition to using the exact
solution, many properties of this class of system (i.e., with a wider range of interactions
maintaining the symmetries of the system) are known based on an asymptotically exact
low-energy field-theory description; the Weiss-Zumino-Witten non-linear σ model with
topological coupling (or central charge) c = 1 [56]. Equivalently, the Heisenberg chain
also represents a special case (because of its spin-rotational invariance) of the Luttinger-
228
S=2
1.5 S=1
S=0
ωS(k)
1.0
0.5
0.0
0 0.2 0.4 0.6 0.8 1
k/π
FIGURE 32. Dispersion of the lowest excitations, relative to the ground state, in the sectors with spin-
inversion quantum number z = 1 (giving low-energy states with spin S = 0 or 2) and z = −1 (which always
gives S = 1 for the lowest state) for an N = 32 chain. The solid curve below the data points is the exact
infinite-size dispersion derived by des Cloizeaux and Person [149] using the Bethe ansatz.
liquid state, which describes a broad range of interacting 1D spin, fermion, and boson
systems [151, 152]. Numerical studies based on exact diagonalization and other unbi-
ased finite-lattice techniques (such as QMC and DMRG) have played an important role
in guiding and confirming these theories (see, e.g., [153] and [7]). Computational studies
are also required for extracting non-universal (short-distance, higher-energy) properties
that are not captured by universal continuum field theories (including quantities directly
accessible to experiments, as discussed in, e.g, Refs. [154, 155]). Numerical results for
the Heisenberg chain are also very important for enabling rigorous bench-mark tests of
extrapolation techniques when studying other spin models, for which less is known from
analytical calculations.
Ground state and low-energy excitations. For a Heisenberg chain of size N = 4n,
the ground state has momentum k = 0, parity p = 1, and spin-inversion number z = 1; it
is a fully symmetric singlet state. For N = 4n + 2 the lowest-energy state is in instead in
the completely antisymmetric sector; k = π , p = −1, z = −1, S = 0.
Let us look at the most important excitations of the Heisenberg chain. For k not equal
to the ground-state momentum, the lowest-energy state of a finite system is a triplet;
hence z = −1 and +1 for chains of size N = 4n and 4n + 2, respectively. Fig. 32 shows
the momentum dependence of the excitation energies ω (k) = ES (k) − E0 for both the
lowest z = +1 and z = −1 states of a 32-site chain. The lowest z = +1 state always
has S = 1, while the lowest z = −1 state has either S = 0 or S = 2. The triplet energies
are quite close to the exact N = ∞ triplet dispersion obtained from the Bethe ansatz
[149], ωS (k) = π | sin(k)|/2, especially close to k = 0. Close to k = π the deviations are
larger. The S = 0, 2 excitations are mostly slightly higher in energy, except that the S = 2
state close to k/π = 0.85 is actually marginally lower than the triplet (likely a small-N
anomaly), and at k/π = 2/N it is much higher.
For large N → ∞, one would expect the singlet and triplet excitations to become de-
generate with dispersion ω0,1 (k) = π | sin(k)|/2. This can be understood as originating
229
8
2.0 7
∆N
6
1.5 5
4
∆
0 0.1 0.2
1.0 1/N
0.5 singlet
triplet
0.0
0 0.05 0.1 0.15 0.2 0.25
1/N
FIGURE 33. Singlet and triplet gaps of the Heisenberg chain versus the inverse system size. The inset
shows the gaps multiplied by N, illustrating the presence of multiplicative logarithmic corrections to the
dominant ∼ 1/N scaling (dynamic exponentz = 1).
from excitations of pairs of spin-1/2 soliton-like degrees of freedom called spinons.

These spinons are very weakly interacting and in the infinite chain behave as indepen-
dent particles, hence forming four degenerate levels (from which S = 0 and 1 states can
be formed). The dispersion graphed in Fig. 32 is only the lowest edge of a continuum of
spinon excitations, which can be calculated in detail using the Bethe ansatz [53] and has
also been observed experimentally in quasi-1D antiferromagnets [156].
The very lowest excitation for a N = 4n chain is a triplet at k = π , with p = −1, z = −1,
whereas for 4n + 2 the lowest triplet has k = 0, p = 1, z = 1 (i.e., the difference in
momentum with respect to the ground state is always π , and p, z are minus their ground-
state values). Fig. 33 shows the finite-size scaling of these lowest triplet and singlet
excitation energies (the singlet and triplet finite-size gaps) versus the inverse system size.
They are both seen to scale to zero as 1/N, corresponding to a dynamic exponent z = 1.
There is a weak correction to this form originating from logarithmic corrections, which
can be seen clearly when plotting the gaps multiplied by N (in the inset of Fig. 33).
This gap scaling has been predicted in detail based on the the continuum field theory
approaches [57, 153]. The prefactor of the gap scaling is directly related to the velocity
of the spinon excitations, but it is not easy to extract it reliably based on the finite-size
data because of the log corrections. The velocity can also be extracted from the linear
parts of the dispersion relation close to k = 0 and π , ωS (k)/k or ωS (k)/(π − k), with the
k = 0 behavior being the easier to analyze, as seen clearly in Fig. 32.
Spin correlation function. The spin correlation function is one of the key character-
istics of of any quantum spin system. The correlations in the ground state are of primary
interest. For a spin-isotropic and translationally invariant chain we can write it as
C(r) = hSi · Si+r i = 3hSzi Sz i+r i, (196)
for any reference site i. As we discussed in Sec. 2, the Mermin-Wagner theorem rules
out antiferromagnetic long-range order in the Heisenberg chain. Instead, in the ground
230
0.25
(-1) rC(r)
0.20
r
r=N/2
0.15 N=32
N=24
N=16
0.10
0 2 4 6 8 10 12 14 16
r
FIGURE 34. Absolute value of the spin correlation function, multiplied by the separation r, in the
ground state of the Heisenberg chain. Results are shown as a function of distance for different system
sizes, and also at r = N/2 versus the chain length N.
state C(r) decays as (−1)r /r for large r, up to a multiplicative logarithmic correction

(originating from a marginally operator in the field-theoretical description [57, 58, 59]).
Fig. 34 shows ground-state results for C(r) as a function of r for system sizes N =
16, 24, and 32. The results have been multiplied by r, so that an r-independent behavior
should obtain for large r if the asymptotic form is ∼ 1/r. It is also useful to multiply by
(−1)r , in order to cancel out the ± oscillations of the staggered phases. Note that there
are remaining even-odd oscillations in C(r)(−1)r . Such oscillations are quite common
for various correlation functions of 1D systems. They diminish with increasing system
size but are still strong for the small system studied here. The form of the oscillations is
also predicted by the Luttinger liquid theory; they decay as 1/r2 [152].
In Fig. 34, clear deviations from the 1/r form of C(r) can be seen which are
mainly due to the periodic boundary conditions, which enhance the correlations close
to r = N/2. Open boundaries cause even more severe finite-size effects (in addition to
not allowing the use of translational symmetry for block-diagonalization). The logarith-
mic corrections should play some role as well. As the chain length is increased, the
correlation function at fixed r converges to the infinite-N limiting form. However, as can
be seen in the figure, the maximum r for which the results are approximately converged
is a rather small fraction of N. For N = 32, one could safely say that C(r) is converged
for N only up to r ≈ 4, which is not enough to say much about the long-distance behav-
ior. An alternative, also shown in the figure, is to investigate the correlation function at
fixed r/N as a function of N, with r/N = 1/2 the most natural choice for checking the
long-distance behavior. Even though periodic boundaries enhance the correlations sig-
nificantly at this point, the functional form of C(r = N/2) is proportional to the infinite-
size C(r) [although the overall prefactor of function C(r = N/2) versus N will clearly be
different from that in the actual N → ∞ converged C(r)]. The leading 1/r form appears
quite plausible based on the data in Fig. 34, and the remaining enhancement is consistent
with the predicted logarithmic corrections. In Sec. 5.3.1 we will further investigate the
long-distance correlations based on quantum Monte Carlo calculations for much longer
chains.
231
4.3.2. Frustration-driven quantum phase transition
A very interesting aspect of the S = 1/2 Heisenberg chain is that it exhibits several
ground-state phase transitions (quantum phase transitions) when a next-nearest-neighbor
interaction is added to the hamiltonian;
N N
H = J1 ∑ Si · Si+1 + J2 ∑ Si · Si+2 . (197)
i=1 i=1
This model goes under the name of the J1 -J2 chain, or the Majumdar-Ghosh model (after
the authors of the first comprehensive study of the system [60]). Here we will discuss
the most important (and most well understood) transition, which occurs when both J1
and J2 are antiferromagnetic (positive). For convenience we define the ratio g = J2 /J1 .
For g < gc , gc ≈ 0.2411 [157], the system is in the same phase as the Heisenberg
chain with J2 = 0 discussed in the previous section. The ground state is critical with
antiferromagnetic correlations decaying as 1/r and the finite-size gaps scale as 1/N (up
to log corrections in both cases). Only the prefactors (e.g., the velocity of the excitations)
and the strength of the log corrections change as a function of g.
For g > gc the system is in a completely different kind of state, with exponentially de-
caying spin correlations and a triplet excitation gap which remains finite in the thermo-
dynamic limit. The ground state has long-range dimer order (a simpler, 1D version of the
2D VBS states discussed in Sec. 2.2). The nearest-neighbor bond strengths (the J1 energy
contribution), hBi i = hSi · Si+1 i, are of the alternating (period two) form Bi = B + δ (−1)i
in the symmetry-broken infinite-size state (of which there are two degenerate ones, with
oscillations out-of-phase relative to each other). Unlike magnetic order, this kind of or-
der is allowed at zero temperature in one dimension because the broken symmetry (the
lattice translational symmetry) is discrete. The modulation δ becomes non-zero at g = gc
and thereafter increases with g. The spin correlations are initially staggered (peaked at
momentum k = π in reciprocal space), but at g ≈ 0.52 the change to k = π /2 [158], and
for g > 1 the peak-value may change continuously with g (a spiral state) [159].
The model with ferromagnetic (negative) J1 is also interesting. There are several
transitions between states with different periodicities [160]). Here we will consider
exclusively the antiferromagnetic case, in the regime g < 1. We will use Lanczos results
to investigate the phase transition into the VBS state, and also discuss the properties of
this ordered state.
The Majumdar-Ghosh point. Before we discuss the Lanczos results, it is wort noting
that the existence of VBS order can be shown exactly at the special point g = 1/2 (the
Majumdar-Ghosh point), where the ground state is very simple [60]. On a ring with even
N, the ground state is a two-fold degenerate singlet product, as illustrated in Fig. 35. One
can demonstrate this rather easily by just acting on the states with the hamiltonian, to
show that they are eigenstates (while the proof of them being the lowest states is more
involved [161, 162, 163]). Out of these degenerate states, |ΨA i and |ΨB i, which break
the translational symmetry, one can form symmetric and anti-symmetric states, which
have momentum k = 0 and π , respectively:
√ √
|Ψ(0)i = (|ΨA i + |ΨB i)/ 2, |Ψ(π )i = (|ΨA i − |ΨB i)/ 2. (198)
232
A B
FIGURE 35. Illustration of the degenerate ground states at the Majumdar-Ghosh point. The thick arcs
illustrate singlets; the product of these singlets on alternating bonds is an exact ground state at g = 1/2.
These are the states obtaining in Lanczos calculations with conserved momentum. Thus,
the order parameter hBi i = E1 is featureless when calculated on finite periodic systems.
As always in exact finite-lattice calculations, the symmetry-breaking has to be observed
by calculating correlation functions (unless we add some perturbation that breaks the
symmetry between the two possible ordering patterns, which has its own complications
as the limit of vanishing perturbation has to be taken). The ground state away from the
point g = 1/2 is more complicated than than the simple singlet-products in Fig. 35,
with fluctuations in the singlet pairings, but with remaining alternating higher and lower
density of singlets on the nearest-neighbor bonds. The translationally invariant k = 0
and π states on finite rings correspond approximately to symmetric and anti-symmetric
combinations as in Eq. (198). It is, however, only at g = 1/2 that these two states are
exactly degenerate for finite N. In other cases, the states become degenerate only in the
limit N → ∞.
Dimer order parameter. We already discussed VBS ground states in Sec. 2.2 and
an example of how to characterize this kind of order by dimer (or bond) correlations
in Sec. 2.4.3. We now investigate the dimer correlation function introduced in Eq. (25),
written explicitly for a 1D system as
D(r) = hBi Bi+r i = h(Si · Si+1 )(Si+r · Si+1+r )i. (199)
This function should alternative between “weak” and “strong” values for large r if there
is VBS order (even if the symmetry is not broken). The calculation of this correlation
function with the Lanczos method can be simplified by taking advantage of the rotational
z z
symmetry and compute h(Si · Si+1 )Si+r Si+1+r i, which is 1/3 of D(r). In principle, we
could make it even simpler by defining the order parameter solely in terms of the z
components, hSizSi+1
z z
Si+r z
Si+1+r i, which is not just a constant times (199) but still a valid
order parameter for a dimerized state. Here we consider the full D(r).
Fig. 36(a) shows results for a 32-site system at three different couplings. At g = 0, the
standard Heisenberg model, the correlations decay rapidly with r and are very small at
r = N/2. At the transition point, gc ≈ 0.2411 (which we will determine more precisely
below), the correlations are clearly stronger. The Luttinger-liquid theory applied to spin
chains predicts dimer correlations decaying as 1/r for g ≤ gc (with stronger logarithmic
corrections when g < gc ) [152]. Going to larger coupling ratios, Fig. 36(a) shows
correlations clearly indicative of long-range order at g = 0.4. To confirm the presence
of long-range order, a finite-size scaling analysis has to be carried out. As we did with
the spin correlation functions of the Heisenberg chain in the preceding section, it is
233
0.4 (a) g=0 0.15 (b)
g=gc
D(N/2)-D(N/2-1)
0.3 g=0.4
0.10
D(r) 0.2
0.05 N=16
0.1 N=24
N=32
0.0 0.00
0 5 10 15 0 0.2 0.4 0.6 0.8
r g
FIGURE 36. (a) Bond correlations for N = 32 chains at three different coupling ratios g. (b) Bond cor-
relations versus g for three different system sizes. The solid and dashed vertical lines indicate, respectively,
the critical coupling gc and the exactly solvable point g = 1/2.
then typically best to look at the correlations at the longest distance, r = N/2, versus the
system size. In the case of the dimer correlations, where the “bare” correlation function is
non-zero even in the disordered phase, one has to subtract off the constant corresponding
to this background. One possibility is to subtract hBi i2 , which is the value to which D(r)
should converge for large r. Another option, which we will choose here, is to instead
use the difference D(N/2) − D(N/2 − 1). This quantity will be non-zero for N → ∞
only if D(r) oscillates, i.e., if there is long-range dimer order. Results for three different
system sizes are shown versus the coupling ratio in Fig. 36(b). Here the order parameter
appears to be well converged in the range 0.4 < g < 0.6 confirming that there is indeed
long-range order.
Finite-size extrapolations of spin and dimer correlations. To extract the infinite-

size dimer order parameter quantitatively, it is in most cases necessary to perform an
extrapolation, unless the system size is sufficiently large for there to be no remaining
size-dependence of significance [which based on Fig. 36(b) is the case close to g = 1/2].
Fig. 37 shows results for both dimer and spin correlations versus the inverse system size
for representative g values both inside and outside the VBS phase.
The dimer correlations, shown in the right panel of Fig. 37, are of the 1/r form for
g < gc . There are logarithmic corrections, which are small exactly at gc . For g > gc ,
inside the VBS phase, the correlations extrapolate to a non-zero value. The asymptotic
N → ∞ convergence should be exponential in this case (as can be demonstrated explicitly
at g = 1/2) but close to the transition it is in practice not possible to reach system sizes
sufficiently large to observe this behavior. Instead, closer to gc the behavior appears to be
essentially linear in 1/N, as seen in the figure at g = 0.4. However, at g = 0.45 one can
see that the behavior is actually non-monotonic, with the large-N behavior consistent
with an approach to the infinite-N value from below, as at g = 1/2. As seen in the right
panel of Fig. 36, for g > 0.6 the dimer order parameter decreases with g. Larger systems
are required for proper extrapolations in this case, and other methods have to be used. It
is believed that the system remains dimerized for all g > gc [158, 159].
The spin correlations should change from 1/r to exponentially decaying as the VBS
234
0.8 0.15
D(N/2)-D(N/2-1)
0.6
C(N/2)(N/2)
0.10
0.4
0.05
0.2
0.0 0.00
0 0.05 0.1 0 0.05 0.1
1/N 1/N
FIGURE 37. Spin (left panel) and bond (right panel) correlations at the longest distance in periodic
chains versus the inverse system size. The couplings ratios for the five curves are g = 0, gc , 0.40, 0.45, 0.50,
in the order top to bottom (the left panel) and bottom to top (right panel), i.e., the three sets shown with
black circles correspond to the VBS phase.
phase is entered. To detect this change, the correlation function C(r = N/2) is multiplied
by N/2 in the left panel of Fig. 37. As we saw before, for g = 0 there is a log-correction
which enhances the long-distance spin correlations. At g = gc these corrections are much
smaller—it is known that the leading log corrections vanish at the dimerization transition
[164]—although the behavior is still not purely 1/r (which can be expected only in the
limit N → ∞). For g inside the VBS phase, rC(r) decays to zero, as expected. Close to
gc one of course has to go to large system sizes—larger than the spin correlation length
(which diverges at gc )—in order to observe a pure exponential fall-off.
For the g-values used in Fig. 37, the spin correlations are staggered, i.e.,peaked at
k = π in reciprocal space. For g ≈ 0.52 the peak position changes rapidly to π /2 [158],
and for g > 1 the spin structure most likely evolves into a spiral with continuously
varying pitch [159] (as is the case for the classical version of the model). The correlations
always decay exponentially, however (unlike the classical long-range spiral).
Determining the dimerization transition point. The dimerization transition is known

to be similar to the Kosterliz-Thouless transition of the classical 2D XY model. Unlike
a conventional phase transition, the order parameter does not follow a power law at
gc , but is exponentially small close to gc . It is therefore not possible to extract the
infinite-N order parameter close to gc based on the small systems accessible to Lanczos
calculations—due to the cross-over behavior it is even difficult relatively deep inside
the VBS phase. It would then appear to be very difficult to determine the location of
the dimerization transition. There is, however, a very elegant way to extract the critical
coupling in an indirect way, based on excited-state energies [157].
As we saw in Sec. 4.3.1, the lowest excited state of the pure Heisenberg chain (g = 0)
is a triplet. On the other hand, we also know that in the VBS state the ground-state should
be two-fold degenerate, and both these ground states must be singlets. At the exactly
solvable point g = 1/2, the degeneracy is exact for any N, but away from this special
point the two states become degenerate only in the infinite-N limits. The approach to
degeneracy should be exponential in N (as will be illustrated with data below), whereas
235
gcross
-5.5
S=0, k=0
S=0, k=π
S=1, k=π
-6.0 0.243
E
-6.5
0.242
-7.0
(a) (b)
0.241
0.0 0.2 0.4 0.6 0 0.002 0.004 0.006
g 2
1/N
FIGURE 38. (a) The three lowest energies for a 16-site chain as a function of the coupling ratio g. The
crossing point gcross of the k = π singlet and triplet levels can be used as a size-dependent definition for the
critical (dimerization) coupling gc . The vertical line indicates the exactly solvable Majumdar-Ghosh point,
at which the singlet states are degenerate. (b) The crossing point versus 1/N 2 . The curve is a polynomial
fit in 1/N (without a ∝ 1/N term, as the leading correction is ∝ 1/N 2 ).
in the case of the triplet excitation of the Heisenberg chain the gap closes as 1/N, and the
lowest singlet also approaches the ground state as 1/N (as shown in Fig. 33). This being
the case, for fixed N there must be some coupling ratio gcross (N) at which the singlet
and triplet excitations cross each other. This point can be taken as a size-dependent
definition of the transition point, and as N → ∞ it should approach the actual critical
coupling; gcross (N → ∞) = gc .
Fig. 38(a) shows the three energy levels of interest for a 16-site chain, with the
corresponding quantum numbers indicated as well (these quantum numbers apply to
all system sizes N = 4n). Note how the two singlets become degenerate as g → 1/2. As
expected, the singlet and triplet excited levels cross; for this system size at g ≈ 0.242.
The crossing points converge very rapidly as a function of N, as shown in Fig. 38(b)—the
leading corrections are known to be proportional to 1/N 2 [164]. Based on these results
it is possible to determine gc very precisely. Fitting a high-order polynomial to crossing
points for 8 ≤ N ≤ 32 gives gc = 0.2411674(2), where (2) indicates the uncertainty in
the last digit based on fluctuations in the extrapolated value when different ranges of
system sizes are included in the fits and the order of the polynomial is varied.
Finite-size gaps in the VBS phase. As already mentioned, to accommodate symme-

try breaking in the thermodynamic limit, one would expect the singlet-singlet finite-size
gap to close exponentially fast with increasing system size inside the VBS phase (with
the exception of the Majumdar-Ghosh point, where the degeneracy is exact for any size).
The exponential gap scaling is demonstrated in the left panel of Fig. 39 for some repre-
sentative values of g. At gc the behavior is instead ∝ 1/N, which is not clearly seen in
this lin-log plot, but more clearly in the log-log plot in the right panel. In the right panel
triplet gaps are also shown versus 1/N. Here extrapolations to non-zero values inside the
VBS phase are apparent. The finite gap essentially corresponds to the energy required to
promote a singlet bond into a triplet. Note that only the ground state is exactly solvable
236
0.5
-1
10 0.4
0.3
∆s -2
∆t
10 g=0.50
g=gcross 0.2 g=0.48
-3 g=0.40 g=0.45
10
g=0.45 0.1 g=0.40
g=0.48
g=gcross
-4
10 0.0
10 15 20 25 30 0 0.05 0.1
N 1/N
FIGURE 39. Singlet (left panel) and triplet (right panel) gaps as a function of system size for different
coupling ratios. Here gcross is the size-dependent coupling at which the lowest singlet and triplet excita-
tions are degenerate (and thus ∆s = ∆t ). The gap at gcross decays as 1/N (as seen more clearly in the right
panel). In the left panel a log scale for ∆s versus N is used to show the exponentially rapid closing of the
gap between the two singlet states that become degenerate for N → ∞ in the ordered VBS phase. To the
right, the triplet gap is plotted versus 1/N to show its convergence to a non-zero value in the VBS phase.
at the Majumdar-Ghosh point, and the triplet gap is size dependent also here. At gc the
gaps scale as 1/N—in the figure the behavior at the size-dependent crossing points gcross
is shown and also scales as 1/N.
4.3.3. Chains with long-range interactions

While long-range spin ordering is not possible in Heisenberg chains with finite-range
interactions, long-range interactions make magnetic order possible at T = 0. A transition
between a Néel state and a quasi-long-range-ordered (QLRO) state (the Heisenberg
critical state with spin correlations decaying as 1/r, discussed in the preceding sections)
takes place in a system with distance-dependent couplings of the form Jr ∝ (−1)r−1 /rα
[72]. The signs here favor antiferromagnetic order, and there is no frustration. When
the exponent α < αc ≈ 2, the ground state has true long-range Néel order, while for
α > αc the system is in the QLRO phase. The critical value of the long-range interaction
parameter αc depends on details of the couplings, e.g., on the nearest-neighbor coupling
J1 when all other Jr are fixed, and the critical exponents at the transition are continuously
varying (in contrast to the constant exponents throughout the QLRO phase).
Another interesting example of a system with long-range interactions is the Haldane-
Shastry chain [165, 166], with frustrated interactions Jr = 1/r2 . It has a critical ground
state similar to that of the standard Heisenberg chain, but, in field-theory language, the
marginal operator causing the leading log-corrections vanishes [167]. The system is,
thus, right at a dimerization transition such as the one discussed above for the J1 -J2
chain. It can be noted that J2 /J1 = 1/4 in the Haldane-Shastry model, which is quite
close to the critical ratio gc = 0.2411 of the J1 -J2 chain. Thus, the interactions beyond
distance r = 2 in the Haldane-Shastry chain only play a minor role (but, importantly,
actually make the system exactly solvable [165]).
237
α− 1
Neel VBS + QLRO(π/2)
1/2
QLRO(π) VBS
q =π q = π/2
0 0.2411 1/2 g
FIGURE 40. Ground state phases of the chain with frustration parameter g and exponent α in the long-
range interaction, reproduced from Ref. [73]. The dashed curves indicate continuous phase transitions,
whereas the thick solid curve represents a first-order transition. The thin solid curve for g < 0 corresponds
to the system with unfrustrated interaction studied in [72]. At α −1 = 0 (the J1 -J2 chain) the dominant spin
correlations in the VBS state change from q = π to π /2 at g ≈ 0.52 [158]. The curve where this change
occurs for α −1 > 0 connects to the multi-critical point where all the phase boundaries come together.
In the presence of long-range interactions one can also realize a direct 1D quantum
phase transition between a Néel state and a VBS. Here we discuss the system introduced
in Ref. [73], which combines unfrustrated long-range interactions and short-range frus-
tration (at separation r = 2) according to the hamiltonian
N/2 N
H= ∑ Jr ∑ Si · Si+r , (200)
r=1 i=1
where the distance dependence of the couplings is given by
N/2
1 1 (−1)r−1 1
J1 = , J2 = g, Jr>2 = , JΣ = 1 + ∑ . (201)
JΣ JΣ rα r=3 r
α
Here, with the normalization by JΣ of all couplings but J2 , the sum of all non-frustrated
interactions |Jr | equals 1, and the limit α → ∞ therefore corresponds exactly to the J1 -
J2 chain with g = J2 /J1 . This will be useful when comparing the two models and also
guarantees a finite energy per spin for N → ∞, even for α < 1. Instead of summing Jr
up to r = N/2, one could also include N/2 < r < N. This should not affect the phase
boundaries and critical exponents for α > 1, however.
We will study the evolution if the dimerization transition occurring as a function of
the frustration strength g as the inverse α −1 of the long-distance interaction exponent
is increased from 0. As shown in the semi-quantitative phase diagram in Fig. 40 (con-
structed on the basis of Lanczos results, as will be discussed below) this continuous
transition persists until α ≈ 2, while for smaller α it evolves into a first-order transition
between the Néel state and a state with coexisting VBS order and critical (or possibly
long-ranged) spin correlations at wave-number q = π /2. This state is denoted in the
phase diagram as VBS+QLRO(π /2).
Evolution of the dimerization transition. We have discussed how to study the dimer-
ization transition based on crossings of excited-state energies for the J1 -J2 chain, with
238
-10 α=4 α=1
-6
E -11 -7
E
-5.2
S=0, k=0 (2) -5.4
S=1, k=π
S=0, k=π -8 -5.6
-12
S=0, k=0 (1) -5.8
0.35 0.4
-9
0 0.2 0.4 0.6 0 0.1 0.2 0.3 0.4 0.5
g g
FIGURE 41. Low-energy levels of a 16-spin system at two values of the long-range exponent; α = 4
(left) and α = 1 (right). The spin S and the momentum k of the states are indicated in the left panel (and
note that two levels are shown with S = 0, k = 0). The inset in the right panel shows the avoided level
crossing of the two k = 0 singlets in greater detail. There is a very small gap between these singlets where
they come together, but this is much smaller than what can be resolved in this figure.
an illustration in Fig. 38(a). The same physics applies in the presence of the long-range
interaction as well, if α is sufficiently large. This is shown for a 16-spin chain at α = 4
in the left panel of Fig. 41. The lowest k = 0 and k = π singlets should become degen-
erate in the VBS phase for N → ∞ (so that a symmetry-broken dimerized states can be
formed). A region of very near degeneracy for g > 1/2 can be seen in the figure. The
region of approximate degeneracy, which is not easy to demarcate precisely, expands
very slowly toward smaller g with increasing N. As shown in Fig. 42, the singlet-triplet
crossing point converges with the system size [although slower than the 1/N 2 conver-
gence shown for J1 -J2 the chain in Fig. 38(b)] and can reliably give the QLRO(π )–VBS
phase boundary gc (α ) for α > 2.
Another interesting feature of the energy levels is that, upon decreasing α below ≈ 2,
the broad maximum in the ground state energy versus g becomes increasingly sharp. As
seen in the right panel of Fig. 41, at α = 1 it has developed into a sharp tip due to an
avoided level crossing with the second singlet at k = 0. The real singlet-triplet crossing
has moved to the same region. An avoided level crossing between two states with the
same quantum numbers, leading to a discontinuity in the derivative of the ground state
energy with respect to g for N → ∞, is the hall-mark of a first-order transition. It should
be noted that it is not just the two lowest singlets that exhibit this kind of avoided level
crossing. Other low-energy states as well come together in the neighborhood of the
transition, in a complicated cascade of level crossings—all of these should converge
to a single point at the first-order transition when N → ∞. The nature of the phases at
this transition will be discussed below. First, let us investigate in more detail how the
dimerization transition evolves from continuous to first-order.
Fig. 42 shows the size dependence of the level crossing gcross and the location gpeak
of the maximum in the ground state energy. In the standard Majumdar-Ghosh frustrated
chain the size correction to the crossing point is ∝ 1/N 2 [as seen in Fig. 38(b)], which
also can be seen for large α . For smaller α , the corrections instead seem to be ∝ 1/N,
but a cross-over to 1/N 2 for large N seems likely as long as the transition remains
239
0.50
0.45
gcross, gpeak
0.40
α = 3.0 (cross)
0.35 α = 2.0 (cross)
α = 1.7 (cross)
α = 3.0 (peak)
0.30 α = 2.0 (peak)
α = 1.7 (peak)
0 0.02 0.04 0.06 0.08

1/N
FIGURE 42. Dependence on the inverse chain length of the singlet-triplet crossing point gcross and the
location gpeak of the ground state energy maximum for different long-range interaction exponents α . The
two lines show extrapolations of the g = 2.0 data to N = ∞.
continuous. The peak location moves in the opposite direction. For some α and N → ∞,
gcross and gpeak should coincide. The results indicate that both gcross and gpeak have
dominant 1/N corrections at this point. Line fits are shown in Fig. 42 at α = 2, were
there is still a small gap between the two extrapolated values (and the extrapolations
may not be completely accurate, if the linear scaling holds asymptotically only at the
special α -value for which gcross = gpeak ). At α = 1.7, where the transition is first-order,
they should coincide (and then the asymptotic size correction should be exponential).
First-order Néel–VBS transition. To confirm an avoided level crossing with a dis-

continuous energy derivative for α > 1.8 (approximately), the second derivative of the
ground state energy at its maximum is graphed on a lin-log scale in Fig. 43. These results
were obtained based on calculations of the energy on a very dense grid of points close
to the peak value, to which a polynomial could be reliably fitted. The second derivative
extracted from this polynomial grows exponentially with N for α = 1.5, showing that
the slope of the energy curve indeed changes discontinuously for an infinite chain. In
contrast, at α = 3 the second derivative decreases for large N. For α = 2 convergence to
a finite value also seems plausible, whereas α = 1.7 and 1.8 appear to be close to a sepa-
ratrix (where the form of the divergence is consistent with a power law) between the two
different behaviors. This analysis suggests that the continuous dimerization transition
changes smoothly into a first-order transition at (gm ≈ 0.41, αm ≈ 1.8). The singlet-
triplet crossing moves toward the ground-state energy maximum and coincides with it at
the multi-critical point (gm , αm ), beyond which it develops into a first-order singularity.
To analyze the states involved in the first-order transition, we next investigate the
standard spin correlation function C(r) and the dimer correlation function D(r) defined
in (199). We will also study the Fourier transforms of these correlation functions; the
static spin (s) and dimer (d) structure factors:
N/2−1
Ss (q) = C(0) +C(N/2) + 2 ∑ cos(qr)C(r), (202)
r=1
240
α = 1.5
3 α = 1.7
10 α = 1.8
α = 2.0
α = 3.0
2
d E/dg
2
2
10
1
10 10 15 20 25 30
N
FIGURE 43. Size dependence of the second derivative of the ground state energy with respect to the
frustration parameter g at the point gpeak where the ground state energy takes its maximum value.
N/2−1
Sd (q) = D(0) + D(N/2) + 2 ∑ cos(qr)D(r). (203)
r=1
If there is long-range order at some wave-vector q = Q, then the corresponding structure

factor is proportional to N (for large N), and order parameters can therefore be defined
as Ss,d (Q)/N. Here Q = π for both the Néel state and the dimerized VBS state. Fig. 44
shows the g-dependence of Sd (π )/N and Ss (π )/N at α = 1.5. Discontinuities are seen to
develop at g ≈ 0.42, in agreement with the first-order scenario (and at g = 1, not shown
here, the discontinuities are much sharper). Both the Néel and VBS order parameters are
well converged to non-zero values in their respective phases, and small in the other phase
(where they should vanish when N → ∞, but it is not possible to observe this clearly
because of the small system sizes). The spin structure factor at q = π /2 is also shown in
the figure. Interestingly, it also becomes large in the VBS phase, while in the Néel phase
it should decay to zero with increasing N (which is plausible based don this data, but,
again, not possible to see clearly for these small systems). Based on these results one
might conclude that the VBS state obtaining at the first-order transition has co-existing
long-range magnetic order at k = π /2 (i.e., a spiral with period four). Considering the
small system sizes, this cannot be guaranteed, however. The behavior could also reflect
a slow power-law decay of the spin correlations.
Fig. 45 shows the real-space spin and dimer correlations at α = 1 for two g values, at
either side of the first-order transition. At g < gc , in the Néel state, the spin correlations
are staggered and clearly long-ranged (with almost no decay seen as a function of r
for r > 1). At this coupling there is no structure in D(r), i.e., there is no VBS order. For
g > gc there is clearly VBS order, with D(r) showing the characteristic staggered pattern.
Strong period-four spin correlations are also observed, but it is not clear whether these
are long-ranged or decay slowly to zero with increasing r. The behavior is consistent
with an 1/r decay, but much longer chains would be needed to extract the behavior
reliably. An interesting point to note here is that the dimer correlations oscillate around
zero, even though no constant has been subtracted. This is in contrast to the oscillations
241
0.15 0.10 N = 24
0.20 N = 28
order parameter
N = 32
Sd(π)/N Ss(π)/N
0.10
Ss(π/2)/N
0.10 0.05
0.05
0.00 0.00 0.00

0.2 0.3 0.4 0.5 0.6 0.2 0.3 0.4 0.5 0.6 0.2 0.3 0.4 0.5 0.6
g g g
FIGURE 44. VBS and magnetic order parameters defined in terms of the size-normalized dimer
structure factor Sd (π ) (left) and spin structure factors Ss (π ) (middle) and Ss (π /2) (right) versus the
frustration strength g in systems with long-range interaction exponent α = 1.5. Discontinuities develop
with increasing system size at the first-order transition.
around a positive value in the J1 -J2 chain, as seen in Fig. 36. Looking at the definition
of the dimer correlation function in Eq. (199), this behavior is another manifestation
of period-four spin structure, i.e., at odd separation of the bond operators Si + Si+1 , a
correlation between a ferromagnetic and an antiferromagnetic bond is measured. While
the nature of the magnetic structure in this VBS is not completely clear, the correlation
functions seem to indicate a tendency to bond-ordering of the form |ststst · · ·i, where s
and t represent nearest-neighbor bonds with high singlet and triplet density, respectively.
In Fig. 40 the state has been denoted as magnetically quasi-long-range ordered, a VBS–
QLRO(π /2) coexistence state, but it could in fact also be a state with coexisting VBS
and long-range magnetic order. It should also be noted that this phase could in principle
be very sensitive to boundary conditions on small lattices, and it cannot be excluded that
the actual magnetic structure is spiral-like, with continuously varying pitch (as in the
J1 -J2 chain with g > 1 [159]).
The VBS+QLRO(π /2) state should have gapless spin excitations, regardless of the
spin correlations as long as they are not exponentially decaying. The lowest triplet is at
k = π /2. It is, however, difficult to demonstrate the gaplessness based on data for small
systems, because the size dependence of the gaps (and other quantities) for system sizes
N = 4n exhibit large even-odd n oscillations (as well as other irregular size effects). In
the conventional VBS phase (in the low-right part of the phase diagram in in Fig. 40)
the lowest triplet is at k = π , even when the spin correlations (which are exponentially
decaying in this phase) are peaked at k = π /2. The level crossing between the lowest
k = π and k = π /2 triplets can therefore in principle be used to extract the boundary
between the VBS and VBS+QLRO(π /2) phases. The size dependence of the crossing
point is not smooth, however, and cannot be extrapolated very reliably. The boundary
between dominant k = π and k = π /2 spin correlations has also not been extracted
accurately. This change in the spin correlations may be associated with a transition to a
state with periodicity four [158], although there are no sign of the VBS order changing.
QLRO–Néel transition. Let us return to Fig. 41 for another interesting feature of the
level spectrum: The lowest singlet excitation for small g has momentum k = π for α = 4
but k = 0 for α = 1. The switching of the order of these levels as a function of α for
242
0.6 g = 0.25 g = 0.25
g = 0.45 0.3 g = 0.45
0.4
0.2
0.2
D(r)
C(r) 0.1
0.0
0.0
-0.2
-0.4 -0.1
0 5 10 15 0 5 10 15
r r
FIGURE 45. Spin (left) and dimer (right) correlations in a 32-spin chain at α = 1 for two values of
the frustration parameter. At g = 0.25 and 0.45 the system is in the Néel and VBS–QLRO(π /2) phase,
respectively. The first-order transition in this case occurs at g ≈ 0.39.
g < gc is associated with the Néel–QLRO(π ) transition. The level crossings can be used
to extract this phase boundary very accurately up to g ≈ 0.25 (while for higher g the
N → ∞ extrapolations become difficult). While it is not completely clear why the k = 0
and k = π excited singlets cross each other at this transition, it can again be noted that
the QLRO(π ) phase the low-energy excitations arise from two deconfined spinons. The
lowest (for N = 4n) is a triplet at k = π . The lowest singlet is also at k = π , with a small
finite-size gap to the triplet due to weak spinon-spinon interactions (and Fig. 33 shows
and example of how these levels become degenerate as N → ∞). In the Néel phase the
spinons are no longer deconfined, and the structure of the low-energy spectrum changes.
As we will discuss below in Sec. 4.4.2, the 2D Néel state on finite lattices has
low-energy excitations analogous to quantum-rotor states, which have S = 1, 2, 3, . . .
and become degenerate with the ground state as the system size increases. They have
momenta k = (0, 0) and k = (π , π ) for even and odd S, respectively. One might expect
this kind of quantum rotor “tower” in the 1D Néel state as well. Apparently, the lowest
k = π singlet, which would not be part of such a rotor tower, is then pushed up much
higher in energy, and the k = 0 singlet takes over as the lowest singlet excitation.
It is useful to compare the level crossing approach with the QMC calculation
in Ref. [72] for the unfrustrated model with J1 = 1 and Jr>1 = λ (−1)r−1/rα . A
reparametrization of this model for λ = 1, to the convention used in (201), gives
the curve for g < 0 shown in Fig. 40. Finite-size scaling of QMC data for the Néel
order parameter gave αc = 2.225 ± 0.025, for N up to 4096 [72] . Extrapolating
the k = 0, π singlet crossing points (which here have size corrections ∝ 1/N β , with
β ≈ 1.50) for N ≤ 32 gives a marginally higher (and probably more reliable) value;
αc = 2.262 ± 0.001. Analyzing the singlet and triplet gaps at the crossings, assuming
∆ ∼ N −z , gives the dynamic exponent z = 0.764 ± 0.005, in very good agreement with
Ref. [72]. For the frustrated model (201) on the QLRO(π )–Néel boundary, the gap scal-
ing in 1/N give z ≈ 0.75 for g up ≈ 0.25, while for larger g it is not possible to reliably
extract αc and z this way, because of large scaling corrections and the absence of level
crossings for increasingly large systems as the multi-critical point (gm ≈ 0.41, αm ≈ 1.8)
is approached (for reference αc = 2.220 ± 0.005 for g = 0 and 2.170 ± 0.01 for g = 0.2).
243
At the multi-critical point z can be extracted using gaps at the singlet-triplet crossing
point and the ground state energy maximum, which gives z ≈ 0.8. It is, thus, possible
that the dynamic exponent is constant, z ≈ 3/4, on the whole QLRO(π )–Néel boundary,
including the multi-critical point where all phase boundaries come together.
Technical notes. The model we have studied here is rather challenging for a 1D
system, from a technical standpoint, because of the long-range interactions. The combi-
nation of long-range interactions and frustration makes it difficult to apply other com-
putational techniques. While efficient QMC techniques can be applied to systems with
unfrustrated long-range interactions [168, 72], this is no longer possible in the presence
of the frustrating J2 term, due to the sign problem (discussed in Sec. 5). The DMRG
method [28, 29], on the other hand, can handle frustration but not easily long-range
interactions. It would still be interesting to try to apply DMRG, or related techniques
based on matrix-product states [21], to this system, in particular to study further the spin
structure in the putative VBS-QLRO(π /2) state.
The Lanczos calculations presented above exploited all the symmetries discussed in
Sec. 4.1. The hamiltonian was stored in compact form on disk, with a separate file for
each interaction distance r in (201). These files were read and used one-by-one in each
operation with the hamiltonian according to code {22}. For N = 32, the number of
Lanczos iterations needed was typically 60 ∼ 80 (and less for smaller systems). All basis
states were saved in primary memory, and re-orthogonalization to all previous states was
carried out after each iteration (as in code {19}).
4.4. Two-dimensional systems

We now discuss exact diagonalization in two dimensions, using the important case of
the Heisenberg model on the simple periodic square lattice as an example. As in one
dimension, we can use momentum conservation to block diagonalize the hamiltonian
within the sectors of fixed magnetization. It is a little more complicated to take maximum
advantage of lattice symmetries to further split some of the blocks, because of the
larger number of symmetry operations commuting with the hamiltonian. On the positive
side, the reward for implementing all the lattice symmetries is that, for high-symmetry
momenta, where lattice reflections and/or rotations can be used, the blocks are smaller
for given N than in one dimension. In practice, this may be of little help, since the
maximum linear size L of an L × L lattice that can be diagonalized is still pitifully small
(L = 6). Exact diagonalization studies are still of great value.
In one dimension, we used the reflection (parity) operation P and constructed momen-
tum eigenstates based on representatives |ai and their reflections, (1 + pP)|ai (and later
we added also spin-inversion). Although P does not commute with the translation opera-
tor, we found that for the two special momenta k = 0, π , both k and p are still good quan-
tum numbers. For general k, p is not a valid quantum number, but with semi-momentum
states, mixing ±k states, we could use P to accomplish a real-valued representation. In
two dimensions one can in principle also construct a real-valued hamiltonian for any
momentum, but this is much more complicated in practice than in one dimension, and
244
29 30 31 32
13 14 15 16
12 13 14 15
25 26 27 28
9 10 11 12
8 9 10 11
21 22 23 24
5 6 7 8
4 5 6 7
17 18 19 20
1 2 3 4
0 1 2 3
FIGURE 46. Labeling of sites and bonds on a 4 × 4 periodic square lattice. With the bit representation,
it is practical to start the site labels from 0, here according to i = x + yL with x, y ∈ {0, . . . , L − 1} with
L = 4. A bond is here labeled according to the “left-down” reference site i to which it is connected; the
horizontal and vertical bonds are bx = i and by = L2 + i, respectively.
in the end it is not worth the effort (since it does not increase the system sizes that can
be studied). 5 Here we will therefore work with standard complex momentum states.
Defining the lattice. As in one dimension, we use the state notation |S0z , . . . , SN−1 z
i,
which is practical with the bit representation of the spins. For a 1D system the spin
indices correspond directly to the chain geometry, but for a 2D lattice we have to
establish a labeling convention for the sites. We will consider rectangular lattices with
N = Lx × Ly sites, at coordinates (x, y) with x = 0, . . ., Lx − 1 and y = 0, . . ., Ly − 1. A
natural choice is to number the sites i = 0, . . ., N − 1 such that i = xi + yi Lx , as illustrated
for a 4 × 4 system in Fig. 46. In a notation not explicitly dependent on the lattice, we can
write the Heisenberg hamiltonian as
Nb
H=J ∑ Si(b) · S j(b), (204)
b=1
where [i(b), j(b)] are the two nearest-neighbor sites connected by bond b. Specifying
the lattice then just amounts to creating this list of site pairs.
4.4.1. Momentum states in two dimensions

We again define translations of the spin indices, now in both the x and y directions, as
illustrated in Fig. 47, with corresponding operators Tx and Ty defined by
Tx |S0z , . . . , SN−1
z
i = |STz x (0) , . . . , STz x (N−1) i,
Ty |S0z , . . ., SN−1
z
i = |STz y (0) , . . ., STz y (N−1) i, (205)
5 A real maximally blocked basis exists because the hamiltonian is real and symmetric, which guarantees
real eigenvalues and eigenvectors. If the hamiltonian is diagonalized numerically as a single block, the
eigenvectors in degenerate multiplets will be mixed, and the translation operators will not be diagonal
(because their eigenvalues are complex). The degenerate complex momentum states have been mixed in
such a way as to make the linear combinations real. Such a superposition can in principle be applied to
block-diagonalize the original hamiltonian matrix into semi-momentum like real-valued blocks.
245
Ty Px
Pe
Tx Py
Pd
FIGURE 47. Symmetries used for the two-dimensional square lattice. Tx and Ty translate the spins by
one lattice spacing in the positive x and y directions, respectively. The lattice is periodic; the open circles
represent sites of the opposite edges. Px and Py are reflections with respect to the x- and y-axis, and Pd
and Pe reflect with respect to the two diagonal axes. For even L (which normally should be used for an
antiferromagnet), the Px and Py axes pass between lattice sites. The reflections Pd and Pe are about the
lines connecting the far corners, which go through the sites on the diagonals (and hence leave the spins on
those sites unchanged upon reflection).
where the translated spin indices are

Tx (i) = [xi − 1]Lx + yi Lx ,
Ty (i) = xi + ([yi − 1]Ly )Lx , (206)
with [γi − 1]Lγ denoting the modulus of γi − 1 with respect to Lγ , i.e., [−1]Lγ = Lγ − 1.
Using these translations, a momentum state based on a representative |ai is defined as
L −1
1 Lx −1 y
|a(k)i = |a(kx , ky )i = √ ∑ ∑ e−i(kxx+ky y)Tyy Txx|ai,
Na x=0 y=0
(207)
where the possible momenta are

2π
kγ = mγ , mγ = 0, 1, . . ., Lγ − 1, γ ∈ {x, y}. (208)
Lγ
The normalization constant Na depends on the translational properties of the representa-
tive, i.e., the number of different states Da obtained among the group of Lx × Ly transla-
tions of the representative |ai (e.g., for a state with the spins in a checker-board pattern
Da = 2). A representative is incompatible with the momentum if the sum of phases Fa in
(207) over the translations bringing |ai onto itself vanishes. For a compatible state, the
normalization constant Na = Da |Fa |2 . The easiest way to compute this in a program is
simply to carry out all the translations and sum up Da and Fa in the process, instead of
using explicit formulas as we did for 1D systems (where one can of course also use the
more brute-force approach).
The construction of the hamiltonian matrix proceeds as in the 1D case. We again split
the hamiltonian into a diagonal piece H0 and off-diagonal bond operators H j as in (121)
and (122), where now j = 1, . . . Nb . Acting with these operators on the representative
|ai, we again may have to translate the resulting state in order to obtain the new
representative |b j i corresponding to H j |ai, i.e.,
y
−l j −l xj
H j |ai = h j (a)Ty Tx |b j i. (209)
246
Acting on a momentum state, we thus get
L −1
h j (a) Lx −1 y y
(y−l ) (x−l x )
H j |a(kx , ky )i = √ ∑
Na x=0 y=0
∑ e−i(kx x+ky y) Ty j Tx j |b j i (210)
y Lx −1
Ly −1
h j (a)
= √ e−i(kx l j +ky l j ) ∑ ∑
x
e−i(kx x+ky y) Tyy Txx |b j i,
Na x=0 y=0
and taking the overlap with hb j (kx , ky )| gives the matrix element
s
x y Nb j
hb(kx, ky )|H j |a(kx , ky )i = h j (a)e−i(kx l j +ky l j ) . (211)
Na
The normalization constants (or some integer mapping to their possible numerical val-
ues) should be stored along with the representatives.
Incorporating other lattice symmetries. We here consider a quadratic, L×L, system.

There are eight independent rotations and reflections of such a lattice, which can be taken
as, e.g., the four 90◦ rotations and a reflection about a horizontal or diagonal axis. It can
be easily verified that additional reflections can be written as combinations of rotations
and the first reflection. The easiest way to see this is to just consider all rotations and
reflections of a 2 × 2 array;

4 3 3 2 2 1 1 4
1 2 4 1 3 4 2 3

3 4 2 3 1 2 4 1
. (212)
2 1 1 4 4 3 3 2
Here the first row contains all the rotations of the first array, and in the second row those
arrays have been reflected by exchanging the two columns. Any other reflection will just
produce an array which is already in the above set of eight. Other permutations of the
elements do not correspond to a combination of rotations and reflections (and hence do
not correspond to symmetries of hamiltonians we are normally interested in). In practice,
we can use any convenient set of reflections, or rotations and reflections, with which the
eight unique transformations can be generated.
Using translations as well as a set of other symmetry operators, which we for now
collectively add together into a single operator Q, with corresponding quantum numbers
{q}, a momentum state can be defined as
1 L−1 L−1 −i(kx x+ky y) ry rx

|aσ (k, {q})i = √ ∑ ∑e
Na rx =0 rx =0
Ty Tx Q|ai. (213)
As we have seen above, there are more symmetry operations than unique transforma-
tions. We can select the ones that are most convenient for given momentum. Instead of
using the rotations of the square lattice, we will here use the reflections Px and Py about
247
the x and y-axis, respectively. These correspond to symmetries of the hamiltonian for
any rectangular (Lx × Ly ) lattice (and note that the 180◦ rotation is equivalent to Px Py).
For generic momenta, the reflections do not commute with the translations, and there
can be no further blocking using Px and Py, but there are high-symmetry momenta for
which Px or Py , or both, can be used. For kx = ±ky we will instead use the reflections
Pd and Pe about the two perpendicular diagonal axes. All the reflections are defined in
Fig. 47. For the most special momenta, k = (0, 0) and (π , π ), and Lx , Ly arbitrary, we
can also use Pd (or Pe) in addition to Px , Py.
To achieve maximal block-diagonalization for a given momentum (kx , ky ), the opera-
tor Q in (213) can be chosen according to:

1, general k (a)
k = (0, ky ), (π , ky)

(1 + px Px ), (b)



k = (kx , 0), (kx, π )

 (1 + py Py ),

 (c)
Q= (1 + pd Pd ), kx = +ky (Lx = Ly ) (d) (214)


 (1 + p e Pe ), kx = −k y (L x = Ly ) (e)
(1 + py Py )(1 + px Px ), k = (0, 0), (π , π ), px 6= py ( f )



(1 + pd Pd )(1 + py Py )(1 + px Px ), k = (0, 0), (π , π ), px = py (g)


where again all the reflection quantum numbers take the values ±1. Note the restriction
px = py in case (g), i.e., there are no such states for px 6= py . In that case, we can of
course also use option (f), but using (g) splits the blocks further into two.
To prove that the symmetry operators Q defined above indeed do have the correspond-
ing conserved quantum numbers within the different momentum sectors, the following
relationships between the lattice transformations are useful:
Px Tx = Tx−1 Px Py Ty = Ty−1 Py
Pd Tx = Ty Pd Pd Ty = Tx Pd
Pe Tx = Ty−1 Pe Pe Ty = Tx−1 Pe (215)
Pd Px = Py Pd Pd Py = Px Pd
Pe Px = Py Pe Pe Py = Px Pe .
The remaining pairs are commuting operators; [Tx , Ty ] = 0, [Tx , Py ] = 0, [Ty, Px ] = 0,

[Px, Py ] = 0, [Pe, Pd ] = 0. Let us just check the most complicated case, (g) in (214). The
permutations commute with the translations for these special momenta, for the same
reasons as discussed for 1D systems in Sec. 4.1.4. We then only have to investigate the
properties of Q. It is clear that the state is an eigenstate of Pd , that being the left-most
reflection operator. To check whether it is also an eigenstate of Px and Py , we use some of
the relationships in (215) to verify that Px Q = px Q and Py Q = py Q. After a little algebra
we get
Px (1 + pd Pd )(1 + py Py)(1 + px Px ) =
px (1 + px py pd Pd )(1 + py Py )(1 + px Px ), (216)
and an analogous result for Py. Thus, if px = py , the state is indeed, for the special
momenta (0, 0) and (π , π ), an eigenstate also of Px , Py , and Pd .
248
If we want to study all momentum sectors, using the reflection symmetries does not
buy us much, because in most of the blocks we have to use the generic option Q = 1 in
(214)(a). These symmetries are useful when investigating the ground state, which often
is in the maximally blockable sector (g). The most useful aspect of the reflections may
still be just the fact that they give insights into the symmetry-aspects of the states (e.g.,
to classify excitations). Apart from the lattice symmetries, spin-inversion symmetry can
be used in the mz = 0 sector, exactly as in the 1D case.
We have considered combinations of reflection symmetries but could instead have
used a combination of rotations and reflections. This would give us different quantum
numbers, but the same size of the blocks. In group theory, there are irreducible represen-
tations, with standardized names, corresponding to various lattice symmetries. Instead
of working out the applicable symmetries for a new case, one can look up the irreducible
representations and their corresponding character tables and employ these to construct
states with quantum numbers of a chosen irreducible representation. This is very use-
ful when studying more complicated lattices. Ref. [143] has a discussion of this more
formal approach, with examples for chains and square lattices.
We have here used symmetries applicable only to Lx × Ly lattices with the edges along
the x- and y axis on the square lattice. With Lx = Ly = L, which is the most interesting if
we are interested in approaching the infinite 2D lattices, we are then limited to L = 4 and
L = 6 (since odd L introduces undesirable frustration effects for an antiferromagnet). To
increase the set of accessible lattice sizes N, one can also consider “tilted” lattices, with
edges that are not parallel to the square-lattice axes [143].
Implementation. To study a 6 ×6 lattice we have to use long (8-byte) integers for the
bit representation. To save time, it is ten better not to loop over all the 2N possible state-
integers and single out the ones corresponding to a given magnetization (as is done,
e.g., in code {7}), but to use a scheme which from the outset only constructs states
with a given mz (i.e., given number n↑ of ↑ spins). This can of course also be done for
1D systems. Such a more sophisticated generation of specific mz states starts with the
integer a = 2n↑ +1 − 1, in which the first n↑ bits are 1 and the rest are 0, and then execute:
do
call checkstate(a, pass) {26}
if (pass) then M = M + 1; store representative information endif
c=0
do b = 1, N
if (a[b − 1] = 1) then
if (a[b] = 1) then
c = c+1
else
a[0], . . ., a[c − 1] = 1; a[c], . . ., a[b − 1] = 0; a[b] = 1; exit
endif
endif
enddo
if (b = N) exit
enddo
249
TABLE 3. Sizes M(L) of the k = (0, 0) state blocks for L × L lattices
(L = 4, 6) with magnetization mz = 0 and different reflection and spin-
inversion quantum numbers. For px = −py there is no quantum number pd .
For px = −1, py = 1 the block structure is the same as for px = 1, py = −1.
px py pd z M(L = 4) M(L = 6)
+1 +1 +1 +1 107 15,804,956
+1 +1 +1 −1 46 15,761,166
+1 +1 −1 +1 92 15,796,390
+1 +1 −1 −1 38 15,752,772
−1 −1 +1 +1 50 15,749,947
−1 −1 +1 −1 45 15,739,069
−1 −1 −1 +1 42 15,741,544
−1 −1 −1 −1 36 15,730,582
+1 −1 +1 75 31,481,894
+1 −1 −1 108 31,525,574
Here the inner loop is simply searching for the lowest bit position b − 1, for which a set
(1) bit of a can be moved one step to the left (i.e., to a position b where the bit currently
is 0). After such a position has been found, all the previously set bits below this position
(the number of which is kept track of with the counter by c) are moved to the lowest
positions (0, . . ., c − 1). This bit evolution corresponds exactly to how the digits of a
base-2 odometer advance from right to left. The contents of checkstate depend on what
symmetries are used, but it would be very similar to the implementations we discussed
for the 1D case, apart from the fact that the translations [defined in (205)] are more
complicated than just cyclic bit permutations and have to be implemented by hand.
The normalization constant, needed when constructing the hamiltonian matrix ele-
ments, should be delivered by checkstate (if a representative has passed the tests). The
simplest way to compute the normalization (instead of using formal expressions as we
did in the 1D case) is again just to carry out all the symmetry operations of the represen-
tative state and add up the sum Fa of factors (the complex momentum phases as well as
the plus or minus signs from the reflection quantum numbers) in (213) for each symme-
try operation bringing the representative onto itself. That number, along with the number
Da of non-equivalent transformations of |ai, gives the normalization Na = Da |Fa |2 (and
again Na = 0 if the representative is incompatible with the quantum numbers). We do
not discuss further details of how to implement the basis generation and the construc-
tion of the hamiltonian matrix, as these tasks are straight-forward generalizations of
the one-dimensional implementations discussed in Secs. 4.1.3 and 4.1.4. Spin-inversion
symmetry in the mz = 0 sector can be implemented as discussed in Sec. 4.1.5.
Example of block sizes. Table 3 lists block sizes for the square-lattice systems of
interest in Lanczos calculations, for one of the special momenta, k = (0, 0), where the
largest number of symmetries can be exploited. The ground state of the Heisenberg
model is in this block (in the sub-block with all other quantum numbers equal to 1). For
L = 6 even the smallest blocks have more than 15 million states, and the largest blocks
250
(for general k) are about eight times larger. Blocks of this size can still be handled in
Lanczos calculations (the smaller one rather easily on a standard workstation).
To test the basis construction, it is useful to check the sum of the sub-block sizes,
which should equal the total number of states in the block. For L = 4 there are 822 states
with k = (0, 0), which equals the sum of all the reflection-block sizes in Table 3.
4.4.2. The Néel state and its quantum rotor excitations

We now illustrate 2D Lanczos calculations with results for the Heisenberg model. Al-
though the small lattices accessible with this method are not sufficient for quantitatively
accurate extrapolations to the thermodynamic limit, the calculations do illustrate some
important aspects of systems with Néel order (beyond what we discussed in the frame-
work of spin-wave theory in Sec. 2.1). We introduce the quantum rotor mapping of the
low-energy states of finite systems, and based on these discuss the magnetic susceptibil-
ity. We also calculate the sublattice magnetization.
Two-spin model of quantum-rotor states. If there is antiferromagnetic order, the

spins on sublattice A are predominantly oriented in the same direction, and the ones on
sublattice B are predominantly in the direction opposite to those on A. If the number
of spins N is finite and the symmetry is not broken, the over-all direction, defined,
e.g., by the sublattice A spins, is not fixed, however. This situation can be captured
by considering the sum of the spins on the individual sublattices A and B [169],
SA = ∑ Si , SB = ∑ Si , (217)
i∈A i∈B
as two fixed-length spins SA = SB = N/2 (more precisely we would write, SA = SB =

ms N for large N, but the constant is irrelevant), as illustrated in Fig. 48. The two large
spins are assumed to be antiferromagnetically coupled to each other in the simplest
possible rotationally invariant way, which is through an effective Heisenberg interaction;
HAB = JAB SA · SB = 21 (S2 − S2A − S2B ), (218)
where S = SA +SB is the total spin. Here S2A and S2B are just constants proportional to N 2 ,
which can be neglected when we discuss excitation energies. However, these constants
imply hat the coupling constant JAB should be ∝ 1/N, in order for the total ground state
energy to be ∝ N. The ground state of (218) has total spin S = 0 and excitations with
S = 1, 2, . . . at energies JAB S(S + 1)/2 above the ground state.
We define a new N-independent coupling Jeff = NJAB /2 and write the energies as
S(S + 1)
∆S = Jeff . (219)
N
These√excitations are referred to as the tower of quantum rotor states. The states with
S ≪ N become degenerate as N → ∞, and combinations of them can then be formed
which are ground states with fixed direction of the Néel vector (in analogy with the infi-
nite number of momentum states required to localize a particle in quantum mechanics),
251
SA
SB
FIGURE 48. Effective description of the rotationally invariant Néel vector ms in terms of two large
spins, SA , SB , corresponding to the sum of the spins on the two sublattices. There is an effective antiferro-
magnetic coupling between these spins, leading to a singlet ground state and a “tower” of quantum rotor
excitations of total spin S = 1, 2, . . . at energies ∆S ∼ S(S + 1)/N above the ground state.
thus allowing for the symmetry breaking that is the starting point for spin-wave theory. In
the thermodynamic limit, the direction of the ordering vector is fixed (as the time scale
associated with its rotations diverges [169]), and the quantum rotor-states are then in
practice not accessed. They are neglected in standard spin-wave calculations (discussed
in Sec. 2.1) from the outset because the order is by construction locked to the z direction.
One can still access the rotor energies in spin-wave theory, by considering systems in
an external magnetic field, tuned to give a ground state with total magnetization Sz = S
[170, 171]. The rotor states are of great significance in finite clusters.
The effective coupling Jeff in (219) for a given system can be determined if we can
relate it to some physical quantity which depends on the rotor excitations. An obvious
choice is the uniform magnetic susceptibility, χ = dhmz i/dh. Calculating it for the two-
spin model when T → 0 gives χ = 3/Jeff . For the real Heisenberg model on a finite
cluster in dimensions d ≥ 2, χ should be dominated by the quantum rotor states when
T ≪ 1/L, because the lowest spin wave energy scales as ∝ 1/L (while the quantum
rotor states scale as 1/Ld ). Thus, we can write the effective quantum rotor tower for a
Heisenberg model with Néel ground state as
S(S + 1)
∆S = , (220)
3χ N
where χ should be evaluated in the limit N → ∞ (first) and T → 0. Note that I = (3/2)N χ
here plays the role of a moment of inertia, giving an analogy between (220) and the
energy spectrum of a rigid rotor in quantum mechanics.
The relation (220) can also be used as a way to compute the susceptibility of a
Heisenberg models numerically; by extracting the lowest energies as a function of S (for
small S, where the quantum-rotor mapping should apply). More precisely, the small-S
energies gives an estimate for χ as the N → ∞, S → 0 limit of the quantity χ (S, N):
1 3NS (ES − E0 )
= . (221)
χ (S, N) S(S + 1)
Here ES denotes the lowest energy for total spin S. Note that we have to subtract the
ground state energy (S = 0) because in the two-spin effective model we only computed
the excitation energies ∆S with respect to the ground state energy (and the latter is not
given accurately by the two-spin model). One would expect an S-independent behavior
252
16 0.8
L=6 L=6
χ (S,N)=3N∆S/S(S+1)
L=4 L=4
0.6
15
(a) (b)
C(r)
0.4
14
0.2
-1
13
0.0
0 5 10 15 0 1 2 3 4
S r
FIGURE 49. Lanczos results for the Heisenberg model on 4 × 4 and 6 × 6 lattices. (a) The energy
relative to the ground state of the lowest state for each total spin sector, normalized by NS(S + 1), so
that perfect quantum rotor excitations should produce S- and N-independent values. (b) Spin correlation
versus distance for all possible distances r on the periodic lattices. The known r → ∞ value (from the
QMC results in Fig. 5) is indicated by the dashed line.
only for S ≪ L, as the higher rotor states should be influenced by effects not taken into
account in the two-spin model.
In analogy with the low-energy 1D quantum numbers discussed in Secs. 4.1.6 and
4.3.1, on the 2D square lattice (where N = 4n for all even L) the quantum rotor states
correspond to momentum (π , π ) and (0, 0) for odd and even S, respectively. For even
S, the lowest states have reflection quantum numbers px = py = pd = 1 in (214), while
for odd S the appropriate quantum numbers are px = py = −1, pd = 1. The lowest state
for given S can be obtained in the magnetization sector mz = S (and since mz 6= 0 we
cannot use the spin-inversion symmetry here). The (S = 0) ground state is in the fully
symmetric sector; the momentum is (0, 0) and px = py = pd = z = 1.
Lanczos results for L = 4 and L = 6 are shown in Fig. 49(a). There are clearly
large corrections to (220), as there is a significant decrease in χ −1 (S, N) with S and an
increase with N. For fixed N, the difference between S = 1 and S = N/2 is roughly 10%.
Considering the fact that the limit S ≪ L cannot really be studied based on the small
lattices, deviations of this order are not surprising. The rotor states have been studied
using QMC calculations for much larger lattices [171, 33]. The most precise calculation
for small S and large N gives χ −1 (S, N) → 22.8 [33].
To understand the deviations from the rotor picture, one can use the analogy of a
slightly non-rigid quantum rotor, which seems natural considering that the two-spin
model is defined with fixed-length spins, while clearly in the real system the sublattice
spins fluctuate (in a way which can depend on the total spin S). It may be possible to take
these effects into account by adding higher-order terms (SA · SB )2 , etc., in the interaction
(218). Details of this extended two-spin picture have not been worked out, however.
Transverse susceptibility. For an infinite system at T = 0 the spin-rotational symme-

try is broken. One can then consider transverse and longitudinal components (with re-
spect to the Néel vector), χ⊥ and χk , of the susceptibility. In the two-spin model it is clear
that χk = 0 and this is also true in the Heisenberg model. In a large system in which the
253
symmetry is not broken, the components are rotationally averaged, giving χ = (2/3)χ⊥ .
The transverse susceptibility is relevant in practice in real Néel order quasi-2D systems
(such as the high-Tc cuprates) at T > 0, because of anisotropies and/or 3D couplings due
to which the system can have a finite ordering temperature. The energy scales associ-
ated with the ordering are small, however, and the T → 0 value of χ⊥ = (3/2)χ of an
isolated 2D plane (modeled by the Heisenberg antiferromagnet) gives a good estimate
of this quantity in the more complicated system.
Sublattice magnetization. The first attempt to compute the sublattice magnetization

of the 2D Heisenberg model was made using Lanczos calculations by Oitmaa and Betts
in 1978 [172]. At that time, the 6 × 6 lattice was beyond reach, and extrapolations based
on smaller lattices resulted in a value that is too large. Fig. 49 shows the spin correlation
function at all possible separations r on 4 × 4 and 6 × 6 lattices. There is a slow decay
with r, and it at least appears plausible that the results are approaching a value hm2s i > 0
as r → ∞ (and N → ∞). QMC calculations, results of which were already shown in Fig. 5,
give the r → ∞ result indicated with a dashed line in Fig. 49. The existence of long-range
order in the 2D Heisenberg model was, in fact, under debate until the first reliable QMC
simulations were done in the late 1980s [173].
One can attempt to extrapolate Lanczos results to N → ∞ for, e.g., frustrated systems,
were other calculations are very challenging [98, 99]. The results must in general be
viewed with caution, however, because it is assumed that the behavior for very small
lattices are already exhibiting the ultimate asymptotic behavior. Fitting procedures based
on “Betts clusters” [174] of different shapes may then be misleading, even when the size
dependence is smooth. There are examples where cross-overs occurs at larger distances
[82], due to the presence of some finite length-scale in the system, which has to be
exceeded by L before asymptotic behavior can be observed.
5. QUANTUM MONTE CARLO SIMULATIONS AND THE

STOCHASTIC SERIES EXPANSION METHOD
Feynman’s path integral formulation of quantum statistical mechanics [175] has played a
major role in the development of QMC methods. In the case of spin systems and related
lattice models, methods based on the path integral in imaginary time are commonly
referred to as world line methods [176]. These techniques were originally based on
an approximate discretization of imaginary time—the Suzuki–Trotter decomposition
of the Boltzmann operator exp(−β H) [177, 178]. Later, exact algorithms operating
directly in the imaginary time continuum were developed [179, 32]. The first practically
useful QMC method did not, however, use the path integral. Already in the early 1960s,
Handscomb developed an approximation-free method for the Heisenberg ferromagnet
based on the power-series expansion of exp(−β H) and exactly computable traces of
products of permutation operators (in terms of which the Heisenberg exchange can
be written) [180, 181]. Although this scheme was also generalized to the Heisenberg
antiferromagnet [182, 183, 184] and some other systems [185], world line methods
were in general more efficient and dominated the field for a long time. The power-series
approach to QMC calculations was revived with the introduction of a more generally
254
applicable exact stochastic series expansion (SSE) formulation [186, 187], in which
also the traces are sampled (thus circumventing the previous reliance on permutation
operator algebra). In addition to improved computational utility and efficiency, this
generalization of Handscomb’s approach also shows clearly how closely the discrete
power series is related to the path integral in continuous imaginary time. Modern QMC
algorithms based on the two formulations are also in the end rather similar [33, 31]. In
particular, in both cases the spin configurations for some models (such as the Heisenberg
model) can be sampled using highly efficient loop-cluster updates [188, 189, 179, 190],
which are generalizations [191] of the classical Swendsen-Wang [118] cluster algorithm
to quantum systems. Further generalizations of the loop concept to “worms” [32] and
“directed loops” [190, 33] have an even wider applicability (e.g., in the presence of
external fields), and have enabled large-scale studies of a broad range of quantum spin
and boson models. Some of these can now be studied at a level of detail approaching the
state of the art for classical systems.
In the case of frustrated spin systems, QMC methods are in general hampered by
“sign problems”, i.e., a non-positive definite path integral or series expansion [192].
While calculations can still be carried out in principle, and there has been some progress
in controlling the sign fluctuations at high temperatures [193], the statistical errors due to
the mixed signs become uncontrollable at low temperatures. Apart from discussing the
origins of the sign problem for frustrated models, we will here only consider unfrustrated
(bipartite) systems. Despite constituting just a subset of the systems which we are in
principle interested in, such models still exhibit a wealth of interesting physics and
continue to provide important insights in cutting-edge research.
In these notes we will discuss implementations and applications only of the SSE
approach, which for spin systems is normally more efficient and technically simpler
than the continuous-time path integral. It is still useful to understand the relationships
between the two schemes. In Sec. 5.1 we therefore first review path integrals in quantum
statistical mechanics and world line QMC methods, in both discretized and continuous
imaginary time. We introduce the general series expansion formulation of quantum
statistical mechanics in Sec. 5.1.3 and investigate its mathematical relationship to the
path integral. An efficient implementation of the SSE method for the S = 1/2 Heisenberg
model is described in Sec. 5.2. Illustrative calculations and results for several 1D and
2D systems are discussed in 5.3; single chains in 5.3.1, ladder systems in 5.3.2, the
uniform 2D lattice in 5.3.3, dimerized systems in 5.3.4 (focusing on the quantum phase
transition between the Néel state and a non-magnetic state), and J-Q models in 5.3.5
(with examples of both continuous and first-order Néel–VBS transitions).
5.1. Path integral and series expansion formulations

of quantum statistical mechanics
The main technical problem in quantum statistical mechanics is how to deal with the
Boltzmann operator exp(−β H). As we saw with exact diagonalization methods in the
previous section, a direct construction of the corresponding matrix becomes infeasible
for systems with more than a few tens of spins. The path integral method offers a way
255
to transform the trace of this operator (the partition function) into a form that can be
sampled using Monte Carlo methods. An alternative is to start from a power-series
expansion of the exponential. Here we introduce both these approaches as a foundation
for implementing the SSE method for Heisenberg models in Sec. 5.2.
5.1.1. The imaginary-time path integral

The starting point of the path integral formulation is to write the exponential operator
at inverse temperature β as a product of L operators with ∆τ = β /L in the exponent;
( )
L
Z = Tr{e−β H } = Tr ∏ e−∆τ H . (222)
l=1
The trace can be expressed as a sum of diagonal matrix elements in any basis. We
can also insert a complete set of states between each of the exponentials. The partition
function then takes the form of an L-dimensional sum of products of matrix elements;
Z = ∑∑··· ∑ hα0|e−∆τ H |αL−1i · · · hα2|e−∆τ H |α1ihα1|e−∆τ H |α0i. (223)
α0 α1 αL −1
Formally, the exponential operator is equivalent to the Schrödinger time evolution oper-
ator exp(−iHt) (with h̄ = 1) at imaginary (Euclidean) time t = −i∆τ . One can therefore
consider (223) as a mapping of a d-dimensional quantum system to an equivalent sys-
tem in d + 1 dimensions, where the new dimension is imaginary time.6 The state index
l corresponds to a discrete set of imaginary time points τl = l∆τ , with 0 ≤ τl ≤ β and
periodic time-boundary conditions (αL = α0 ) for the states. Often the discrete times are
referred to as “time slices”.
From a technical perspective, the purpose of writing Z in the form (223) is that,
while the matrix elements of exp(−β H) are difficult to evaluate, we can use some
approximation to evaluate the matrix elements of exp(−∆τ H) when ∆τ is small (the
number of time slices is large) . We can then compute the weights for the different time-
periodic “paths” α0 → α1 · · · → αL−1 → α0 over which the system can evolve in the
chosen basis. In QMC calculations, these paths are importance-sampled according to
their weights in (223).
6 Often the equivalent (d + 1)-dimensional system is referred to as an equivalent classical system. This
terminology is, however, appropriate only in cases where all the path weights are positive, which we have
not yet ascertained. It will be true for some classes of systems only, and those are the ones for which
QMC calculations can be performed in practice. In some cases, e.g, the transverse-field Ising model, the
effective model is an anisotropic version of the classical model in d + 1 dimensions. In most cases, the
equivalent classical system is not, however, the same kind of model as the original one with just one more
dimension—typically the path integral corresponds to some completely different statistical mechanics
problem in d + 1 dimensions, with no apparent resemblance to the original d-dimensional system. In some
cases, one can, however, show that the quantum system is equivalent, on large length and time scales, to
the same classical system in d + 1 dimensions, e.g., the 2D quantum Heisenberg antiferromagnet at low
temperatures has the same properties as a 3D classical Heisenberg model [5]. This mapping is normally
carried out using a basis of coherent spin states, as discussed, e.g., in the book by Auerbach [4].
256
To discuss the possible space-time paths and their weights, let us first consider a
seemingly extreme method of approximating the exponential operator by just its Taylor
expansions to linear order in ∆τ H, i.e.,
Z≈ ∑ hα0|1 − ∆τ H|αL−1i · · · hα2|1 − ∆τ H|α1ihα1|1 − ∆τ H|α0i, (224)

{α }
where {α } refers collectively to all the states |α0 i, . . ., |αL−1 i. Since each factor now
has an error of order ∆τ2 and there are L = β /∆τ factors, the relative error in Z at
fixed β is of the order ∆τ . The leading neglected term in each expanded factor also
contains H 2 , and we might therefore suspect that the discretization error should also
scale as N 2 . However, to find exactly how the error scales with β and N requires a more
careful analysis than just naively counting the neglected terms at the level of hH 2 i ∝ N 2
and L = β /∆τ , because the paths contributing to the error are not the same as those
contributing to the approximate Z written as (224). Based on arguments discussed later,
to achieve an error which is independent of β and N we should expect to use at least of
the order of N β time slices in the linear approximation (224), i.e., ∆τ ∝ 1/N.
Far better approximations of the time slice operator can be used to reduce the number
of slices (to independent of N and proportional to β , with an error of ∆2τ or smaller).
However, for the purpose of illustrating many basic aspects of the path integral and
QMC methods based on them, it is convenient to first use the linear approximation.
In fact, as we will see shortly, it is even possible to take the limit ∆τ → 0 within this
approximation and formulate QMC algorithms based on the exact continuous time path
integral [179, 32, 33]. We therefore discuss the linear version in some more detail,
although in practice it is not recommended to implement an actual program using
this scheme with ∆τ > 0. After understanding the properties of the paths in the linear
approximation it will be easy to understand how to take the limit ∆τ → 0 or use a higher-
order discrete approximation, such as the Suzuki-Trotter decomposition discussed in
Sec. 5.1.2.
Boson path integral and world lines. At this stage it is better to continue the dis-
cussion with a particular hamiltonian in mind, in order to have a concrete example of
the paths and how one might go about sampling them. Consider first a boson system
with only kinetic energy. We here work on the lattice,7 and the purely kinetic-energy
hamiltonian can be written as (with an unimportant prefactor set to 1);
H = K = − ∑ Ki j = − ∑ (a+j ai + a+
i a j ), (225)
hi, ji hi, ji
where a+ i and ai are, respectively, boson creation and destruction operators on the
sites i = 1, . . . , N, and hi, ji is a pair of nearest-neighbor sites on an arbitrary lattice
(in any number of dimensions, although for ease of visualization we will explicitly
consider a 1D chain). For simplicity, we will consider hard-core bosons, for which the
7See [194] for a review of boson path integrals in continuous space, and [195] for more recent progress
on efficient QMC algorithms based on them.
257
0 0
19 19
18 18
17 17
16 16
15 15
14 14
13 13
12 12
11 11
10 10
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
FIGURE 50. Graphical representation of terms in the discrete path integral for a system of three bosons
on a chain with 14 sites. There are 20 time slices, labeled by l = 0, . . . , 19, which correspond to the states
|αl i in Eq. (224). The bosons are represented by circles. They are connected to show more clearly the
formation of world lines on the space-time lattice. Note the periodic boundary conditions in the time
(vertical) direction, i.e., the first and last (l = 0) states are the same. Note also that with the linear time
slice operator (1 − ∆H), there cannot be more than one particle hop (diagonal line segment) within a time
slice. With periodic spatial boundary conditions, the identical particles can undergo cyclic permutations
(winding), as in the configuration to the right.
site occupation numbers are restricted to ni = 0, 1 (which can be thought of as arising

from a very strong on-site repulsion). Hard-core bosons also have a simple relationship
to S = 1/2 quantum spins. Occupied and empty sites correspond directly to ↑ and ↓
spins, and with no interactions the boson hamiltonian is equivalent to the XY model,
with Hi j = −(Si+ S−j + Si− S+j ). There is no chemical potential in (225)—we use the
canonical ensemble with fixed number of bosons NB (density ρ = NB /N or, in the
spin language, magnetization m = ρ − 1/2). Later, we will see that it is easy to also
incorporate interactions.
We work in the basis of site occupation numbers ni , i = 1, . . ., N. Consider a ma-
trix element hαl+1|1 + ∆τ Ki j |αl i, which appears in Eq. (224) when we also write each
instance of H as a sum over all the individual hopping terms Ki j (and note the cancel-
lation of the minus signs). There are then two possible relationships between the states
|αl i and |αl+1i resulting in non-vanishing matrix elements; either |αl+1i = |αl+1 i or
|αl+1i = Ki j |αl i. In the latter case, there must initially be a particle at site i, which is
moved by Ki j to a previously empty site j, or vice versa. Since these conditions must
hold for all consecutive matrix elements, l = 0, . . . , L − 1, 0, each particle has to fol-
low a “world line” (a term borrowed from the similar concept of a space-time path in
relativity theory), which at each step of the imaginary time propagation either stays at
the same spatial position or moves (jumps) by one lattice spacing. Two such world line
configurations for a 1D system are illustrated in Fig. 50.
Note again the periodic boundary conditions in the imaginary time direction, which
follow from the trace over α0 in (224). These boundary conditions apply to the boson
states (occupation numbers), but not strictly to the world lines. Since the bosons are
identical particles, they can be permuted in the course of their evolution from τ = 0 to
τ = β and still fulfill the required time-periodicity. The only possible permutation in a
258
1D hard-core system with nearest-neighbor hopping is a cyclic permutation involving
all the particles on a periodic chain, resulting in a net particle current around the
ring. An example of such a “winding” configuration is shown to the right in Fig. 50.
Other permutations can take place with soft-core bosons (i.e., with no restriction on the
occupation numbers ni ), and also hard core bosons in two or three dimensions (and even
in one dimension if hopping beyond nearest neighbors is included), but only winding
results in a net current. The net number of times world lines wrap around the system
(the total current divided by the system length) is called the winding number. It is a
topological quantity, characterizing a non-local property of the configuration. In higher
dimensions, there can be winding in each of the spatial directions, with corresponding
winding numbers. An interesting aspect of winding in bosons systems is that it (a
properly normalized variance of the winding number) corresponds to superfluidity [175,
196]. In spin systems, winding is related to the spin stiffness, or helicity modulus, as we
will discuss further below.
We now denote by W ({α }) the weight of a proper world line configuration {α } (i.e.,
one satisfying the constraints discussed above) in the partition function;
Z= ∑ W ({α }). (226)
{α }
The weight has a very simple form in the case considered here: Each kinetic operation
Ki j is multiplied by the time step ∆τ in (224), and the matrix elements of the boson
operators is always 1 in the hard-core case. The path weight is therefore
W ({α }) = ∆τnK , (227)
where nK is the total number of kinetic jumps, corresponding to diagonal world line
segments in Fig. 50. This weight can be simply generalized to the case of soft-core √
bosons,
√ where the creation and annihilation operators are associated with factors ni
and ni − 1. Below we will also consider interactions between the bosons, which lead
to a more complicated factor multiplying the kinetic contribution (227).
In a QMC simulation, one makes changes in the world lines in such a way that
detailed balance is satisfied with the configurations distributed according to W ({α }).
Local updates of world lines are illustrated in Fig. 51. Note that while these local
moves are ergodic within a sector of fixed winding number, they cannot change the
topological winding number. In higher dimensions, local moves of individual world lines
also cannot lead to any of the other permutations that need to be included. Local updates
involving two world lines can be used to sample permutations, but are not enough to
change the winding numbers associated with periodic boundaries. Algorithms based
on simple local updates were historically successful in studies of some systems, but
more sophisticated and powerful loop updates [31], and related updating methods in an
extended configuration space [32, 33], have been developed more recently. These are
much more efficient in evolving the world lines in an ergodic way (instead of “getting
stuck” or spending long times in some restricted part of the contributing configuration
space) and can also lead to winding number changes.
In this section we are primarily concerned with the path integral representation,
and we we will not yet address the practical implementation of world line sampling
259
(a) (b)
FIGURE 51. Local updates of world lines, which change the number of kinetic jumps (diagonal world
line segments) by ±2. In (a) the diagonal segments are directly following each other, whereas in (b) they
are separated by one time slice. One can consider moves with arbitrary separation of the jumps.
(which we will do in detail only for the rather similar configurations arising in the
series expansion formulation, although we will later in this section return to some more
quantitative discussion of world line sampling as well). Next, we consider estimators for
expectation values that we may want to compute. This will also give us some deeper
insights into the properties of the world lines themselves.
Expectation values and their world line estimators. Consider an arbitrary operator
O and its thermal expectation value;
1
hOi = Tr{Oe−β H }. (228)
Z
Proceeding with the numerator as we did for Z in Eq. (223), the exact time-sliced form
of the expectation value can be written as
1
Z {∑
hOi = hα0 |e−∆τ H |αL−1 i · · ·hα2 |e−∆τ H |α1 ihα1|e−∆τ H O|α0 i. (229)
α}
We would like to express this expectation value in the form appropriate for Monte Carlo
importance sampling;
∑{α } O({α })W ({α })
hOi = . (230)
∑{α } W ({α })
This is not always possible, however, because the paths contributing to the numerator in
(229) may be different from those contributing to Z. If W ({α }) = 0 for a configuration
contributing to the expectation value, then no estimator O({α }) can be defined in (230).
One then has to proceed with the calculation in a different way. We begin by discussing
classes of expectation values for which (230) does apply.
Diagonal operators. For quantities that are diagonal in the occupation numbers, e.g.,
a density correlator hni n j i, the form (230) is trivially valid, because O|α0 i = O(α0 )|α0 i
in (229), where O(α0 ) denotes an eigenvalue of O. The estimator can then be taken
as O({α }) = O(α0 ). This form of the estimator for a diagonal operator remains valid
regardless of how the time-slice evolution operator e−∆τ H is approximated. Because of
the cyclic property of the trace, by which the operator O can be inserted anywhere in the
product of time slice operators in (229), one can also average over all time slices and use
1 L−1
O({α }) = ∑ O(αl ),
L l=0
(231)
260
which normally improves the statistics of a simulation. In practice, to save time without
significant loss of statistics, one may only perform a partial summation in (231) over,
e.g., every Nth time slice (normally L > N), because eigenvalues O(αl ) and O(αl+1) of
nearby states differ very little (at most corresponding to one single world line jump).
States separated by ∝ N slices typically differ significantly and contribute independent
statistics, at least to some degree (exactly how much can be statistically quantified in
terms of an imaginary-time dependent correlation function hO(τ )O(0)i).
The kinetic energy. Expectation values of off-diagonal operators are in general more
complicated. The kinetic energy hKi is an exception. In addition to being easy to
evaluate, it is also a quantity of particular interest for understanding the properties of the
path integral. At the level of the linear approximation of e−∆τ H in Z, we can approximate
Ke−∆τ H in the form (229) with O = K by just K, because this is done only at a single
time-slice and leads to a relative error of order ∆τ . This is of the same order as the total
error from the linear approximation made at all the other time slices in both (224) and
(229). Given a configuration {α } that contributes to Z, the estimator for a specific kinetic
operator Ki j is then
hα1 |Ki j |α0 i
Ki j ({α }) = . (232)
hα1 |1 − ∆τ K|α0 i
Here the numerator is non-zero only if there is a world line jump between sites i and j at
the first time slice. In that case Ki j ({α }) = 1/∆τ = L/β . In all other cases the estimator
vanishes. We can again average over all time-slice locations of the operator Ki j in (229),
which results in
hni j i
hKi j i = , (233)
β
where ni j denotes the number of kinetic jumps in the world line configuration between
sites i and j. Thus, the total kinetic energy is given by the average of the total number of
kinetic jumps nK ; hKi = −hnK i/β .
We could have derived the expression for the kinetic energy in a simpler way, by ap-
plying the thermodynamic formula for the internal energy (here just the kinetic energy);
E = ∂ ln(Z)/∂ β , with Z given by Eqs. (226) and (227). The more complicated derivation
illustrates explicitly how the form (230) involves matching the configuration spaces of
the numerator and denominator. This matching is not possible for generic off-diagonal
operators; only for ones that are part of the hamiltonian. Before considering other cases,
let us discuss another important aspect of the kinetic energy estimator.
The utility of the expression (233) is not just that it enables us to compute the kinetic
energy. It also carries with it a fundamental message about the path integral and the
nature of the world lines. It is natural to ask how typical world lines will evolve as we
increase the number of time slices. Specifically, how many kinetic jumps can we expect
in a typical world line configuration?
At low temperatures the kinetic energy should be almost temperature independent
(approaching a constant as β → ∞). It should be proportional to the lattice size N. Using
Eq. (233), we can therefore deduce the expected size and temperature scaling of the
number of kinetic jumps as
hnK i = −hKiβ ∼ N β . (234)
261
β
(a) (b)
0
FIGURE 52. Continuous-time world line configurations. Here the kinetic events (jumps) occur at
arbitrary imaginary times 0 ≤ τ < β . Configuration (a) contributes to the partition function, whereas
(b) includes a pair of creation (solid circle) and annihilation (open circle) operators separated by more
than one lattice spacing and does not contribute to Z or diagonal expectation values. Such a configuration
instead contributes to an off-diagonal expectation value ha+i a j i.
This tells us that there is a typical scale of roughness of the world lines; for large L, the
typical time separation between jumps is ∼ 1/N. As mentioned in the beginning of the
section, we should expect to use of the order N β time slices in a discrete path integral
with the linear approximation of the time slice operator to avoid errors growing with N
and β . We now have the explanation for that, because it is clear from (234) that a smaller
number of time slices cannot accommodate the number of kinetic jumps necessary for
the correct behavior of the kinetic energy (and then indirectly any other quantity).
The relationship between hnK i and hKi is also important from another perspective;
it indicates that it should be feasible to formulate simulation algorithms directly in the
continuum limit—even if we let the number of time slices L → ∞, the number of kinetic
jumps (“events”) in the contributing configurations stays finite. The continuous-time
world lines can then be stored in the form of just a single state |α0 i and the nK events;
the times τi , i = 1, . . . , nK at which they occur and the direction of each jump. With hnK i
scaling as N β , it should be possible to construct Monte Carlo sampling algorithms with
this scaling in system size and temperature of the time and memory requirements.
Examples of world line configurations in continuous time is shown in Fig. 52. The
configuration to the left contributes to the partition function, whereas the one to the right
does not, because it does not satisfy all time-periodicity constraints. It instead contributes
to the expectation value of an operator a+ i a j , the Fourier transform of which is the
momentum distribution function ρ (k) = ha+ k ak i. We briefly discuss such off-diagonal
quantities next.
Off-diagonal operators and broken world lines. If we consider a path {α } contribut-

ing to Z and proceed to treat the expectation value of a general off-diagonal operator
a+
i a j in the same way as we did for the kinetic terms in (232), it is clear that we always
get zero, unless i, j are nearest neighbors (in which case the operator is part of the kinetic
energy). On the other hand, we can construct other paths, which do contribute to ha+ i a ji
but not to Z, like the one in Fig. 52(b). If the sites i, j are separated by more than one
lattice spacing, the expectation value ha+ i a j i cannot be written in the standard Monte
Carlo sampling form (230), because W ({α }) = 0 for all O({α }) 6= 0. It is still possi-
ble to evaluate ha+k ak i, as well as the corresponding imaginary-time dependent function
262
k (τ )ak (0)i (the single-particle Green’s function) by working in the combined space
ha+
of the periodic world line configurations contributing to Z and those with two defects
corresponding to the presence of a creation and annihilation operator [32].
The path integral including interactions. It is not difficult to generalize the bosonic
path integral discussed above to hamiltonians including potential-energy terms in ad-
dition to the kinetic energy K in (225). Let us denote by V any interactions that are
diagonal in the occupation number basis, e.g., attractive (vi j < 0) or repulsive (vi j > 0)
terms of the form vi j ni n j . With H = K +V , we can decompose exp(−∆τ H) as
e−∆τ H = e−∆τ K e−∆τ V + O(∆2τ ), (235)

where the error is due to the commutator [K,V ] 6= 0. With V diagonal, we can for each
time slice in (223) write
hαl+1 |e−∆τ H |αl i ≈ e−∆τ Vl hαl+1|e−∆τ K |αl i, (236)
where Vl denotes the potential energy evaluated at the lth time slice. The weight (227)
of a world line configuration is therefore modified in the presence of interactions as
!
L−1
W ({α }) = ∆nτ K exp −∆τ ∑ Vl . (237)
l=0
The interaction part of the weight can be easily taken into account in elementary world
line moves of the kind shown in Fig. 51, and also in more sophisticated treatments in
continuous imaginary time [32].
The continuum limit. We have already discussed the fact that the continuum limit
of the path integral can be used directly in QMC algorithms. However, examining the
configuration weight, (227) for the purely kinetic hamiltonian or (237) in the presence
of interactions, we have an apparent problem; the weight vanishes as ∆τ → 0. This must
clearly be compensated in some way by the number of configurations increasing, in such
a way that the partition function remains finite. In Monte Carlo calculations we do not
deal with the partition function directly and only need ratios of configuration weights
to compute acceptance probabilities for world-line updates. Looking at such ratios for
the simplest kinds of updates, the insertions and removals of opposite kinetic jumps
illustrated in Fig. 51, they are also problematic in the continuum limit: Formally the
probability for insertions and removals is zero and infinity, respectively. This is not just
a problem in the continuum, but also for small ∆τ , where the probability of accepting an
insertion would be non-zero but very small.
Proper ways to handle the continuum limit were introduced in the context of “worm”
[32] and loop algorithms [179], twenty years after simulations based on discrete path
integrals were first introduced [177, 178]. Here we only consider local world line moves
for a toy model, as an illustration of how Monte Carlo sampling can in fact rather easily
be formulated in the continuum.
Fig. 51 shows two different ways to introduce two events moving a world line to
and from a neighboring site. Consider a line segment which is initially straight over m
263
(a) (b)
τ2 τ2 τ2 τ2 τ2
τb τb τ’b
τa τa
τ’a
τ1 τ1 τ1 τ1 τ1
FIGURE 53. Local world line moves in continuous time. In (a), two opposite kinetic jumps are inserted
at randomly chosen times τa and τb between two existing jumps at τ1 and τ2 . In (b), two opposite jumps
are either removed (←) or the times of the two events are randomly changed to new arbitrary times τa′ and
τb′ between τ1 and τ2 (→). The acceptance probabilities are given in Eq. (238).
time slices, between times τ1 and τ2 , with τ2 = τ1 + m∆τ . Assuming that there is no
world line on the neighboring site within the range [τ1 , τ2 ], the weight change when
introducing the two events is ∆τ2 (for a purely kinetic hamiltonian), regardless of where
in this range the events occur. There are a total of m(m − 1)/2 different ways of inserting
events, which gives a total relative weight ∆τ2 m(m − 1)/2 for the subspace of such
modified configurations, versus 1 for the original configuration with two less events. In
the continuum limit, the weight for all the possible updated configurations of this kind
becomes (τ2 − τ1 )2 /2. As we have discussed above, when ∆τ → 0, the average total
number of kinetic events remains finite. The typical length τ2 − τ1 of a straight world
line segments then also remains finite, and we should be able to construct a scheme with
finite insertion probabilities. The key is that we do not just consider a single specific
updated configuration, but a continuous range of possible updates.
We also need to consider the removal of two events. In order to simplify the discus-
sion, we here consider just a single boson on two sites. This avoids the complications
of having to consider also constraints imposed by world lines on neighboring sites. The
world line moves for which we will construct probabilities satisfying detailed balance
are illustrated in Fig. 53. The event insertion discussed above is illustrated in (a), where
it should be noted that the times τ1 and τ2 correspond to two consecutive existing events
(and in case there are no events, we take τ1 = 0 and τ2 = β ). We call the times of the two
new events τa and τb , and these are chosen at random anywhere between τ1 and τ2 . In
the opposite update of removing two events, we consider the total weight of two existing
events τa and τb within two surrounding events at τ1 and τ2 . Then we can again compute
the total weight of the existing events, not only at the fixed current times τa and τb but
at any times τa′ and τb′ within [τ1, τ2 ]. That relative weight is again (τ2 − τ1 )2 /2, versus
1 for the single configuration with those two events absent.
Thus, once τ1 and τ2 have been identified by inspecting the current configuration
(which can be stored in the form of a list of events), the acceptance probabilities for
two-event insertion and removal are
(τ2 − τ1 )2 /2 1
Pinsert = , Premove = . (238)
1 + (τ2 − τ1 )2 /2 1 + (τ2 − τ1 )2 /2
264
For an accepted insertion, we generate τa and τb at random, while in a rejected removal
we also generate new times τa′ and τb′ for the existing events.
In a larger system, with more than one boson, the above scheme would have to be
modified to take into account the constraints of other world lines when a chosen world
line is moved. The times τ1 and τ2 should then reflect those constraints. In the presence
of interactions, the acceptance probabilities would be modified to take into account the
potential-energy factor in (237). In practice, one would not use these local updates,
however, as loop and worm updates [32, 179, 33] are not much more complicated to
implement, but much more efficient. The purpose of the discussion here was to show
how the apparent problems of the continuum limit can be overcome in principle.
5.1.2. The Suzuki-Trotter decomposition

While modern world line QMC algorithms for lattice models are normally based on
the path integral in the continuum limit, historically time-discretized variants based on
a Suzuki-Trotter approximation of the time-slice operator exp(−∆τ H) were developed
first. The discrete approach is still some times used and is the most practical option in
some cases [197]. We here discuss the main features of Suzuki-Trotter based methods
and how they are applied to boson and spin systems.
One example of a Suzuki-Trotter (or split operator) approximation [177, 198] was al-
ready written down in Eq. (235). This is just one among many possible decompositions
of an exponential of two (or more) non-commuting operators into a product of exponen-
tials [199]. The most commonly used approximants are, for any pair of operators A and
B and a small factor ∆;
∆A ∆B
e e + O(∆2 )
e∆(A+B) = (239)
e∆B/2 e∆A e∆B/2 + O(∆3 ).
Here the errors are also proportional to the commutator [A, B]. Using a larger number
of judiciously chosen exponentials of functions of A and B, the remaining error can
be further reduced, in principle to an arbitrary high power of ∆ [200]. High-order
approximants are often too complicated to work with in practice, however.
If we are interested in the trace of a product of exponentials, as in the path integral, it
is easy to see that the two low-order approximants (239) are actually equivalent, because
( ) ( )
L L
Tr ∏ e∆B/2 e∆A e∆B/2 = Tr ∏ e∆A e∆B , (240)
l=1 l=1
due to the cyclic property of the trace. Thus, although the world line method is often
discussed based on the first line of Eq. (239), where the error from each factor is ∝ ∆τ2 ,
the error is in effect smaller; ∝ ∆τ3 . However, since the number of exponential factors in
the path integral is L = β /∆τ , the total error made in a product such as (240) for fixed N
and β is ∝ ∆2τ . In addition, one can show that for sufficiently large L, the error also does
not grow with N at fixed ∆τ . In most cases, one can therefore keep ∆τ independent of
265
N and β (although for some classes of expectation values, care has to be taken to avoid
error divergences when β → ∞ [201, 202]).
In the context of path integrals, the utility of the Suzuki-Trotter approximation is
that even though exp(−∆τ H) in the partition function (223) cannot be easily evaluated,
if we find a suitable decomposition of H into two terms (or some small number of
terms), H = HA +HB , we may be able to evaluate exp(−∆τ HA ) and exp(−∆τ HB ) exactly.
The prototypical example is a one-dimensional system containing only nearest-neighbor
kinetic energy and interactions, which we collectively denote Hi,i+1 for a site pair i, i +1.
We then decompose H = HA + HB according to
HA = ∑ Hi,i+1, HB = ∑ Hi,i+1, (241)
odd i even i
With this arrangement, all terms in each of the sums HA and HB are mutually commuting;
[Hi,i+1, Hi+2m,i+2m+1 ] = 0 for any i and m. We can therefore split exp(−∆τ HA ) and
exp(−∆τ HB ) into two-body operators without making any further errors, e.g.,
e−∆τ HA = e−∆τ H1,2 e−∆τ H3,4 · · · e−∆τ HN−1,N . (242)

Using products like this in the time-sliced partition function (223), we can insert com-
plete sets of states between all the exponentials. Since each operator now involves only
two sites, the matrix elements reduce to the form
hβ |e−∆τ Hi, j |α i = hnβ i , nβ j |e−∆τ Hi j |nα i , nα j i, (243)
which for hard-core bosons are the elements of a 4 × 4 matrix. Since the bond operator
Hi, j conserves the number of bosons on the two sites involved, this matrix is bock-
diagonal, with two single-element blocks (for nα i = nα j = nβ i = nβ j ) and a 2 × 2 matrix
with both diagonal and off-diagonal elements connecting the states with a single boson
on the bond. It is thus easy to evaluate all the matrix elements, also if interactions are
included (as long as they do not extend beyond nearest neighbors; otherwise a more
complicated decomposition of H has to be used).
Pictorially, each matrix element (243) corresponds to a four-site plaquette with zero,
one, or two world line segments going through it, according to the occupation num-
bers. Considering all time slices and all site pairs i, i + 1 in both exp(−∆τ HA ) and
exp(−∆τ HB ), these plaquettes form a checkerboard pattern, with two adjacent rows cor-
responding to one time slice, as illustrated in Fig. 54. For obvious reasons, the hamilto-
nian decomposition (241) is also often called the checkerboard decomposition.
For a hamiltonian conserving the total number of particles, the constraints on allowed
world line configurations are similar to those in the linear approximation of the time-
slice operator, with a few important modifications. Kinetic jumps are allowed only on
the shaded plaquettes in Fig. 54, but there is no further constraint on the number of jumps
within a time slice, unlike in the linear time-slicing approximation illustrated in Fig. 50
(where there can be at most one jump in each time slice). The world line configurations
still look very similar. They become equivalent when ∆τ → 0, in the physically relevant
sense of their evolution from some time τ to some later τ ′ , with τ ′ − τ ≫ ∆τ .
In a Monte Carlo simulation, the world lines can be updated using simple moves
of the kind illustrated in Fig. 51, with the constraint that the diagonal loop segments
266
β
∆τ
0
FIGURE 54. A 1D world line configuration based on the checkerboard decomposition with the Suzuki-
Trotter approximation. Kinetic jumps of the bosons (or flips of a pair of ↑ and ↓ spins) are allowed only
across the shaded squares (plaquettes). A time slice of width ∆τ consists of two consecutive rows of pla-
quettes. The six isolated plaquettes shown to the right correspond to the non-zero matrix elements, which
in the case of a spin model with Heisenberg interactions (for world lines and empty sites corresponding
to ↑ and ↓ spins, respectively) are given by Eq. (244).
(kinetic jumps) are allowed only on the shaded plaquettes in Fig. 54. More complicated
“loop” and “directed loop” updates, in which large segments of several world lines can
be moved simultaneously, are used in modern algorithms [31, 191, 33] (which we will
discuss in detail below in the context of the stochastic series expansion method).
Application to the Heisenberg model. It is useful to consider a particular example

of the path weights in the Suzuki-Trotter approach. Let us compute the plaquette matrix
elements for the antiferromagnetic Heisenberg interaction; Hi,i+1 = Si · Si+1 . In this case
the boson occupation numbers in (243) are replaced by spin states ↑ and ↓. We can
consider the world lines forming between the ↑ spins (and note that we could also draw
world lines for the ↓ spins in pictures such as Fig. 54; they occupy all sites not covered
by ↑ world lines and cross those lines at each diagonal segment). The calculation just
involves straight-forward algebra and we just list the results for the six allowed (non-
zero) matrix elements;
h↑i ↑ j |e−∆τ Hi j | ↑i ↑ j i = h↓i ↓ j |e−∆τ Hi j | ↓i ↓ j i = +e−∆τ /4

h↑i ↓ j |e−∆τ Hi j | ↑i ↓ j i = h↓i ↑ j |e−∆τ Hi j | ↓i ↑ j i = +e∆τ /4 cosh(∆τ /2) (244)
−∆τ Hi j −∆τ Hi j ∆τ /4
h↓i ↑ j |e | ↑i ↓ j i = h↑i ↓ j |e | ↓i ↑ j i = −e sinh(∆τ /2)
The weight of a world line configuration is a product of these matrix elements, all of
which are pictorially represented in the right part of Fig. 54. Note the minus sign in
front of the off-diagonal matrix elements. For an allowed world line configuration, all
the signs cancel out due to the periodicity constraint on the world lines. This is true
also for world line methods applied to bipartite lattices in higher dimensions, but for
frustrated systems there is a “sign problem” because of the presence of both negative and
positive weights (as we will discuss further in Sec. 5.1.3). In practice, world line methods
and similar QMC approaches are therefore useful primarily for studies of bipartite spin
systems and bosons models. For a fermion system, permutation of world lines also
lead to sign problems, except in one dimension where only global cyclical permutations
(winding) are possible (with associated signs that can be avoided by choosing periodic
or anti-periodic boundary conditions [176, 187]).
267
In higher dimensions one can use the Suzuki-Trotter approach with various decom-
positions analogous to the one discussed above. Local updates such as those we have
discussed for 1D systems can be adapted to higher dimensions as well. It is worth read-
ing Refs. [79, 203], which are two pioneering world line studies that firmly established
the basic properties of the 2D S = 1/2 Heisenberg antiferromagnet.
Although the number of time slices used in the Suzuki-Trotter approach does not have
to depend on the system size (because ∝ N events can take place within a single time
slice), the computational effort of Monte Carlo sampling the world-line configurations
scales in the same way as in the continuous time formulation, as N β . This scaling is
achieved in the exact continuum formulation by only storing and manipulating the times
of the events, whereas in the Suzuki-Trotter approach one normally works with the full
space-time lattice (although one could also in principle only store the events there). For
this reason, the continuous-time methods actually normally run faster on the computer,
in addition to not being affected by discretization errors. The only reason to use a discrete
path integral would be in problems where it is difficult to take the continuum limit in
practice, e.g., in effective models including dissipation, which leads to time-dependent
interactions among the world lines [197].
5.1.3. The series expansion representation

As an alternative to the discrete time-slicing approach or continuum limit of the path
integral, one can also construct a configuration space suitable for Monte Carlo sampling
by using the Taylor expansion of the Boltzmann operator;
∞
(−β )n n
e−β H = ∑ H . (245)
n=0 n!
This approach was pioneered by Handscomb [180, 181], who developed a method for
studying the S = 1/2 ferromagnet. The Taylor expansion was later considered more
generally as a starting point for exact QMC algorithms for a wide range of models [182,
185, 186, 187]. The power-series expansion of the exponential operator is convergent
for a finite lattice at finite β (which technically is due to the act that the spectrum of H
is bounded). In effect, as we will see below, the series expansion allows for a discrete
representation of the imaginary time continuum, thus avoiding approximations but fully
retaining the advantages of a finite enumerable configuration space.
Choosing a basis, the partition function can be written as
∞
(−β )n
Z= ∑ ∑ hα0|H|αn−1i · · · hα2|H|α1ihα1|H|α0i, (246)
n=0 n! {α } n
where the subscript on {α }n indicates that there are n states to sum over. This expression
can be compared with the linear-order discrete path integral (224). Taking H to be the
bosonic kinetic energy in (225), we can clearly draw world line pictures very similar
to those in Fig. 50 to represent the contributing terms. The main difference is that
the number of “slices”, the expansion power n, is varying and for given n there are
268
particle jumps at each slice. The new “propagation” dimension we have introduced
in this representation is different from imaginary time, but it is clear that the label
p = 0, . . . , n − 1 of the states is closely related to imaginary time in the path integral.
We will discuss the exact relationship further below.
In the case of only kinetic energy, all matrix elements in (246) equal −1 for an allowed
configuration, the minus signs cancel and the weight is
βn
W ({α }n) = . (247)
n!
In more general cases, this weight will be modified by the product of matrix elements
in (246). As in the path integral, the weights are positive definite (or can be made so
with some simple tricks) for boson systems and quantum spins without frustration in the
off-diagonal terms. We will here proceed under the assumption of positive definiteness
and discuss the precise conditions for this further below.
It is useful to derive a general expression for the total internal energy U = hHi. For
any hamiltonian, we can write it as
1 ∞ (−β )n
E= ∑ n! ∑ hα0|H|αni · · ·hα2|H|α1ihα1|H|α0i,
Z n=0
(248)
{α } n+1
where the presence of an additional matrix element of H in each term should be noted;
for given n, the summation is over n + 1 states instead of the n states in the partition
function (246). All configurations that contribute in (248) contribute also to the partition
function, but the weights differ. We can match the configurations explicitly by writing
1 ∞ (−β )n n
E=− ∑ n! β
Z n=1 ∑ hα0|H|αni · · · hα2|H|α1ihα1|H|α0i. (249)
{α }n
We can extend the sum over n to include also n = 0, because this term vanishes.
The terms in (246) and (249) then match exactly, and n/β can be identified as the
energy estimator. Thus, with the configurations sampled according to their weights in
the partition function, the energy is simply given by
hni
E=− . (250)
β
Writing H as a sum, H = − ∑i Hi , and using this for all instances of H in (246) and
(248), we can derive a similar expression for the expectation value of an individual term
Hi . The result for this estimator is the average number of times the operator appears in
the operator string in the expansion of the partition function;
hni i
hHi i = . (251)
β
In the case of a kinetic-energy term, this is identical to the path-integral expression
(233). This shows that there is a one-to-one correspondence between the paths in the two
269
formulations, if we by path mean just the order in which kinetic events occur, without
reference to the times τl of the events in the discrete or continuous path integral. This
correspondence holds quite generally. The difference is that in the series expansion, the
potential-energy terms are treated in the same way as the kinetic operators (i.e., they
are not re-exponentiated, as they are in the path integral). Their presence corresponds
to “non-events”, which do not affect the world lines but are associated with diagonal
matrix elements that modify the weight of the paths. In the path integral, there is instead
the potential-energy factor in Eq. (237), which depends on the actual time-points of the
kinetic events.
Eq. (250) shows that although the sum over expansion orders n in (246) extends up
to n = ∞, it can in practice be truncated at some order nmax ∝ N β . To further motivate
why this will only lead to an exponentially small and completely negligible error, one
can make use of the estimator for the specific heat, obtained by taking the temperature
derivative of (250);
C = hn2 i − hni2 − hni. (252)
When the temperature T → 0, C should vanish, and (252) then shows that the variance
of the distribution of n equals hni. It is therefore clear that the distribution vanishes ex-
ponentially beyond some power n ∝ N β . In practice, QMC (stochastic series expansion,
SSE) algorithms based on this representation automatically sample the n-distribution
according to the relative weights of the different n sectors.
By explicitly truncating the Taylor expansion, one can make the relationship between
the series and path integral representations even clearer. Truncating at n = L, we can
formally construct a fixed-size sampling space by augmenting all powers H n with n < L
by L − n unit operators I. Allowing for all possible locations of the unit operators in
the product of L operators, we can define a modified operator string S = S1 , S2, . . . , SL ,
in which Si ∈ {H, I}, and do a summation over all these sequences. We then have to
compensate the weight by the number Ln of equivalent terms, giving
(−β )n (L − n)!
Z=∑ ∑ ∑ hα0|SL |αL−1i · · · hα2|S2|α1ihα1|S1|α0i, (253)
S L! {α } L {Si }
where n now refers to the number of elements Si = H in the operator string (and we
no longer need an explicit sum over n). If we now consider the boson kinetic energy
(225) and take the limit L → ∞, the weight reduces to (β /L)n , which is the same as
the path integral weight (227) for L time slices. In this limit, it is clear that the index
p = 0, . . ., L − 1 is related to imaginary time according to τ = pβ /L. The difference
is that the series expansion with the full weight, β n (L − n)!/L!, is in practice exact
for L ∝ β N, whereas in the path integral we have to take the continuum limit to avoid
a discretization error. For finite L, the series index p does not correspond exactly to
imaginary time, but represents a distribution of imaginary times [187]. Because of the
close relationship between the propagation index p and imaginary time, it is appropriate
to refer to the p space as the “time” dimension also in the series expansion.
The practical advantage of the series expansion is that it provides an exact but discrete
representation of the imaginary time continuum. Algorithms based on it are normally
easier to implement than continuous-time world lines and can be more efficient compu-
tationally. On the other hand, for a hamiltonian with large potential-energy terms, the
270
number of operators hni in the expansion can be much larger than the number of ki-
netic events hnK i in the path integral. The latter approach should then be more efficient,
although an exact treatment necessitates manipulation of floating-point time variables.
For quantum spin systems, there is in general a favorable balance between kinetic (off-
diagonal) and potential (diagonal) energies, and the series expansion is then typically
preferable. Next, we proceed to develop a QMC algorithm based on it.
5.2. SSE method for the S = 1/2 Heisenberg model

.
We now discuss in detail how to implement the SSE method for the S = 1/2 Heisen-
berg model. Initially we do not have to specify the lattice and consider the hamiltonian
written as a sum of bond operators,
Hb = Jb Si(b) · S j(b) , (254)
as in (204), with the lattice encoded as a list of sites [i(b), j(b)] connected by the bonds,
b = 1, . . . , Nb . An example of a labeling scheme in 2D was given in Fig. 46.
A positive-definite SSE can be constructed for any bipartite lattice, i.e., when all sites
i(b) and j(b) belong to sublattice A and B, respectively, in any number of dimensions.
This constitutes a large class of interesting and important systems. Initially we will con-
sider just a single antiferromagnetic coupling constant Jb = J > 0, but the modifications
needed for non-uniform systems are very simple.
5.2.1. Configuration space

It is useful to subdivide the Heisenberg interaction (254) into its diagonal and off-
diagonal parts in the standard basis of diagonal z spin components. We then define
operators with two indices, Ha,b , with a = 1, 2 referring to diagonal and off-diagonal,
respectively, and b = 1, . . ., Nb is the bond index as before;
1 z z
H1,b = 4 − Si(b) S j(b) , (255)
1 + − − +
H2,b = 2 (Si(b) S j(b) + Si(b) S j(b) ). (256)
Here we have also introduced a minus sign and a constant in the diagonal operator, so
that the full hamiltonian can be written as
Nb
JNb
H = −J ∑ (H1,b − H2,b) + 4
. (257)
b=1
The reason for including the constant in H1,b is to make the series expansion positive-
definite, as we will see shortly. The constant JNb /4 is irrelevant in the algorithm but we
will include it when calculating the energy.
271
Series expansion of the partition function. The general form of the partition function
in the SSE approach can be written as (246), but here we will initially not write out all the
complete sets of states inserted between the operators. We instead focus on the operators,
expanding all instances of H as sums over all the bond operators in (257). The starting
point of the SSE algorithm for the Heisenberg hamiltonian is thus
∞
* +
n2 β
n n−1
Z = ∑ ∑ (−1) ∑ ∏ a(p),b(p) α .

α H (258)

n!

α n=0 S p=0 n
Here J has been absorbed into β = J/T , Sn refers to the products (strings) of the
Heisenberg bond operators (255) and (256) originating from H n ,
Sn = [a(0), b(0)], [a(1), b(1)], . . ., [a(n − 1), b(n − 1)], (259)
and n2 is the total number of off-diagonal operators, i.e., the number of elements with
b(i) = 2 in the string. Note that the label p referring to the position of an operator in
the string here takes the values 0, . . . , n − 1; this labeling will be more convenient than
p = 1, . . ., n in the program implementation.
When the operator string acts on the state |α i = |Siz , . . ., SNz i we get a succession of
other basis states, with no branching into superpositions of more than one state. We will
refer to these as propagated states |α (p)i,
p−1
|α (p)i ∝ ∏ Ha(i),b(i) |α i . (260)
i=0
We have not written down the normalization of these states explicitly but will consider
all |α (p)i as properly normalized. These states of course correspond to those in the
summations over complete sets in (246), but in practice the SSE approach is framed
around the summation over the operator strings in (258), and we do not always have to
write out the states explicitly.
With the constant 1/4 in (255), all operations on parallel spins destroy the states;
H1,b | ↑i(b) ↑ j(b) i = 0, H2,b | ↑i(b) ↑ j(b) i = 0,
H1,b | ↓i(b) ↓ j(b) i = 0, H2,b | ↓i(b) ↓ j(b) i = 0. (261)
An operator-state configuration (α , Sn) contributing to Z thus has to involve only op-

erations on anti-parallel spins, and the propagated states (260) are defined under this
assumption. The corresponding matrix elements are
h↑i(b) ↓ j(b) |H1,b | ↑i(b) ↓ j(b) i = 21 , h↓i(b) ↑ j(b) |H2,b | ↑i(b) ↓ j(b) i = 21 ,
h↓i(b) ↑ j(b) |H1,b | ↓i(b) ↑ j(b) i = 12 , h↑i(b) ↓ j(b) |H2,b | ↓i(b) ↑ j(b) i = 21 . (262)
The fact that all these are equal will be very useful in the sampling algorithm, and this
was the reason for including the constant 1/4 in the diagonal operators.
In addition to the local constraints of only operations on anti-parallel spins, for the
matrix element of the full operator product in (258) to be non-zero, the propagation also
has to satisfy the periodicity |α (n)i = |α (0)i, where |α (0)i = |α i.
272
Although one can formulate Monte Carlo sampling procedures in the space of oper-
ator strings of fluctuating length n [180, 187, 204], it is normally in practice somewhat
easier to work within the fixed-size truncated space discussed in the previous section
(and mathematically it is also easier to prove detailed balance in such a space).8 In ac-
tual simulations, the cut-off L will be determined automatically by the program, as we
will discuss below, such that L safely exceeds the largest n ever sampled. Note again that
the truncation then does not cause any detectable errors and should not be considered as
an approximation. Defining the unit operators used for augmenting strings with n < L as
H0,0 = I and including [a(p), b(p)] = [0, 0] as an allowed element in the index list (259),
the partition function can be written as
* +
n2 β (L − n)!
n L−1
Z = ∑ ∑(−1) α ∏ Ha(p),b(p) α ,

(263)

α S L
L! p=0

where n refers to the number of non-[0, 0] elements in the fixed-length operator string SL
(and n is summed implicitly by the sum over SL ). Since all matrix elements equal 1/2,
the weight of an allowed configuration is given by
β
n
(L − n)!
W ( α , SL ) = . (264)
2 L!
Here we have left out the sign factor (−1)n2 , which, as we will discuss further below,
always is positive for a bipartite (unfrustrated) system. Note also that L! is an irrelevant
normalization factor and can be left out as well.
With ↑ and ↓ spins represented by the presence and absence of a hard-core boson,
respectively, we could draw pictures like Fig. 50 to illustrate the rules for the contributing
SSE configurations. However, we now also have diagonal operators that do not change
the world lines but modify the path weight. In the path integral approach the diagonal
operators are all brought together into a common exponential factor, Eq. (237), but
there is no simple way to do this in the series expansion. As mentioned above in
Sec. 5.1, the SSE approach should not be used when the diagonal energy dominates,
but in the case considered here we have hH1,b i ≈ hH2,bi and the diagonal operators do
not dominate. As we will see below, the presence of a significant number of diagonal
operators is in fact exploited in the SSE algorithm. When discussing the SSE method,
it is useful to draw pictures showing explicitly the diagonal and off-diagonal operators.
An example is given in Fig. 55, which includes also illustrations of the data structures
that we will later use in a computer implementation of the method. The states will be
represented by integers σ (i) = ±1 corresponding to Siz = ±1/2. For compact storage of
the operator string, it is also useful to combine the indices [a(p), b(p)] into a single list
of integers s(p) = 2b(p) + a(p) − 1. The full propagated states do not have to be stored
simultaneously. They can be generated as needed from the single stored state |α (0)i and
8 A case where it is clearly better to work with sequences of fluctuating length n is when the SSE approach
is combined with a multi-canonical ensemble, in which a range of temperatures is sampled and therefore
n can fluctuate over a very wide range [204].
273
i = 1 2 3 4 5 6 7 8
σ(i) = -1 +1 -1 -1 +1 -1 +1 +1
p a(p) b(p) s(p)
11 1 2 4
10 0 0 0
9 2 4 9
8 2 6 13
7 1 3 6
6 0 0 0
5 0 0 0
4 1 2 4
3 2 6 13
2 0 0 0
1 2 4 9
0 1 7 14
FIGURE 55. An SSE configuration for an 8-spin chain, with all the propagated states shown. Open and
solid bars indicate diagonal H1,i and off-diagonal H2,i operators respectively, while no bar between states
corresponds to a “fill-in” unit operator H0,0 . The ↑ and ↓ spins of the state |α i are stored as σ (i) = ±1,
and the operator string SL is encoded using even and odd integers for diagonal and off-diagonal operators,
respectively, according to s(p) = 2b(p) + a(p) − 1.
the operator string. We will later introduce a different compact storage involving some
spins of the propagated states as well.
Frustrated interactions and the “sign problem”. At first sight, it appears that we
have a sign problem—a non-positive definite expansion—because of the factor (−1)n2
in (258). Actually, all the terms are positive for a bipartite lattice. This is because an
even number n2 of off-diagonal operators are required in every allowed configuration, in
order to satisfy the “time” periodicity |α (L)i = |α (0)i. We already discussed this in the
context of the world line method, where the off-diagonal matrix elements in (244) are
negative, but the continuity of the world lines require an even number of these. This is
yet another example of the close relationship between the two approaches.
For frustrated systems, the series expansion is not positive-definite (and neither is the
path integral, for exactly the same reason). This can be easily demonstrated for a system
of three spins on a triangle. As shown in Fig. 56, an allowed configuration can in this
case contain three off-diagonal operators, resulting in an over-all minus sign. This is
true for any system in which loops with an odd number of sites can be formed between
antiferromagnetically interacting spins—this can be used as the definition of frustration.
Positive-definiteness for a bipartite system can also be proved in a different way, by
carrying out a unitary transformation of the spin operators on one of the sublattices, say
B, such that S+j → −S+j and S−j → −S−j (and no change in the diagonal operators Szj ) for
j ∈ B. This does not affect the spectrum of the model (since the commutation relations
among all spin operators remain unchanged), but the sign in front of the off-diagonal
terms in the hamiltonian (257) changes to +. The factor (−1)n2 in (258) is then absent.
274
FIGURE 56. An example of three off-diagonal operations (indicated by bars) bringing all spins on a
triangle back to their original states. Each spin flip is associated with a minus sign, resulting in a negative
path weight and a “sign problem” (due to cancellations of configurations with different signs) in QMC
simulations of this and other frustrated systems.
For a frustrated system, no such transformation can remove all the signs.
Note that only the off-diagonal part of the interaction causes a sign problem. One can
study systems with frustration in the diagonal part, but then the interaction is no longer
spin-isotropic. Just neglecting the sign for a frustrated Heisenberg system corresponds to
antiferromagnetic diagonal couplings but ferromagnetic off-diagonal couplings. Here we
only consider isotropic interactions and restrict the discussion to unfrustrated systems.
For a bipartite system with periodic boundaries, the absence of sign problems holds
strictly only if the lattice lengths Lγ are even in all directions γ = 1, . . ., d. Winding
configurations, like the one shown for the path integral in Fig. 50(b), are negative if
there is an odd number of windings around an odd number of sites. Clearly, in such
cases there is also frustration, due to the boundaries. We will therefore consider only
systems of even length (in all periodic lattice directions).
Linked-vertex storage of the configurations. We have discussed computer storage of

the SSE configuration as a state |α (0)i and an operator string SL . By acting sequentially
with the operators, it is easy to generate all the propagated state |α (p)i illustrated in
Fig. 55. However, during the Monte Carlo sampling we will need to access operators
and some properties of the propagated states also in non-sequential order—given an
operator and the spins it acts on, we will need to to know which operators act on those
spins next (when moving up as well as down in the operator sequence). It would be
prohibitively time consuming to propagate a single state back and forth to extract this
information, and also it would not be practical to store all the propagated states. We
therefore use also another kind of data structure, in which the “connectivity” of the
operators is explicit and represented as a network in a compact way. This linked vertex
structure is illustrated in Fig. 57, using the same SSE configuration as in Fig. 55. Here,
in the pictorial representation, the constant state of each spin between operators have
been replaced by straight lines, which in a computer program will correspond to links
(pointers). The operators are shown along with only the two spin states before and after
each operator has acted. We will call these spin-operator objects vertices and refer to the
four spins as the legs of the vertices. One stage of the Monte Carlo sampling procedures
will involve making changes to the vertices (while their locations are changed at another
stage). The links will allow us to quickly move between the vertices and make a series
of changes maintaining all the constraints.
The allowed vertices are dictated by the hamiltonian. In the case of the isotropic
Heisenberg system considered here there are four of them, depicted in Fig. 58. They
of course correspond to the non-zero matrix elements (262), and, again, the constant 41
in the diagonal operator is the reason why there is no vertex with all four spins equal.
Anisotropic interactions or an external magnetic field would necessitate inclusions of
275
p v X(v) v X(v) v X(v) v X(v)
11 44 18 45 30 46 16 47 17
10 40 - 41 - 42 - 43 -
9 36 31 37 7 38 4 39 5
8 32 14 33 15 34 12 35 0
7 28 19 29 6 30 45 31 36
6 24 - 25 - 26 - 27 -
5 20 - 21 - 22 - 23 -
4 16 46 17 47 18 44 19 28
3 12 34 13 2 14 32 15 33
2 8 - 9 - 10 - 11 -
1 4 38 5 39 6 29 7 37
0 0 35 1 3 2 13 3 1
l=0 l=1 l=2 l=3
FIGURE 57. Linked vertex storage of the configuration in Fig. 55. In the graphical representation to
the left, constant spin states between operators have been replaced by lines (links) connecting the spins
just before and after the operator acts. The links can be stored in a list X(v), where the four elements
v = 4p + l, l = 0, 1, 2, 3, correspond to the legs (with the numbering convention shown in Fig. 58) of the
vertex at position p in the sequence SL . For two linked legs v and v′ , X(v) = v′ and X(v′ ) = v.
2 3 2 3 2 3 2 3
0 1 0 1 0 1 0 1
FIGURE 58. Allowed vertices for the isotropic S = 1/2 Heisenberg model. The numbering l = 0, 1, 2, 3
of the vertex legs corresponds to the position v = 4p + l, in linked-list storage illustrated in Fig. 57.
additional vertices [190, 33] (and the algorithm discussed here would then also have to
be modified). Although the spin states at the four legs uniquely identify the vertices, we
will continue to use also the open and solid bars in pictures, to indicate diagonal and
off-diagonal vertices, respectively, for added clarity.
For a given position p in the operator sequence SL , the corresponding list element
s(p) tells us the operator type (diagonal or off-diagonal) and the bond b on which it
acts (as explained in Fig. 55). As will become clear below, along with this information,
we only have to store the connectivity of the vertices, not their spin states. The links
allowing us to jump between connected vertex legs are stored as a list X (v), as explained
in Fig. 57. For clarity of the illustration, the one dimensional list has here been arranged
in four columns, with elements labeled v = 4p + l, corresponding to each type of leg,
l = 0, 1, 2, 3, with the labeling specified in Fig. 58. We will later describe an efficient
way to construct this linked list, given the operator sequence. For now, it is sufficient to
know that for a given operator at location p in the sequence, the position of its l:th leg
in the linked vertex list is v = 4p + l. This leg is linked to another vertex leg with list
address v′ = X (v). This kind of structure constitutes a doubly-linked (bi-directional) list,
276
(a) (b)
FIGURE 59. (a) A pair of off-diagonal operators, indicated by solid-line boxes, which can be replaced
by diagonal ones if the spins between the two operators are flipped as well. Such an update cannot be
done with the operators enclosed by dashed boxes, because of the illegal spin configuration (a vertex with
all four legs in the same spin state) that would result at the operator acting between the two boxed vertices
on their left spin. However, as shown in (b), this operator pair can be changed if instead the spins on the
opposite sides of the operators are flipped. This also forces a change in the stored state |α i.
with X [X (v)] = v, in which we can move both “up” and “down”. From a position v in the
list we can extract the corresponding operator location in SL , p = v/4 (its integer part)
and leg index l = mod[v, 4]. We can move “sideways” to the other leg on the same vertex
by changing the leg label 0 ↔ 1 or 2 ↔ 3, which in both cases can be expressed as a
simple rule of an even↔odd change in the list address v. These movements will allow
us to construct closed loops of changed spins and operators.
5.2.2. Monte Carlo sampling procedures

When Monte Carlo sampling the partition function (258), we have to make changes
in the operator sequence SL as well as in the stored state |α i. It is clear from Figs. 55 and
57, however, that updates in these two data structures are not necessarily independent.
Allowed changes (i.e., ones maintaining all configuration constraints) in operators at
locations p ∈ {p1 , . . . , p2 } lead to corresponding changes in the propagated states in
the range |α (p1 + 1)i, . . ., |α (p2 )i. Fig. 59 illustrates such an update of two operators,
where only the types of the operators are changed, diagonal↔off-diagonal (which leads
to changes in the states similar to the world line update in Fig. 51). As the changed
operator at p1 acts on |α (p1 )i, the state |α (p1 + 1)i and subsequent states also change
(by two flipped spins in the example), but when the last updated operator at p2 has acted
the modifications are “healed” and the resulting state |α (p2 + 1)i is the same as before
the update. Since the propagated states form a cyclically permutable periodic structure,
there is nothing special with the stored state |α i = |α (0)i and it can change as well,
depending on the locations of the updated operators.
Because of the time (p) periodicity of |α (p)i, the way state changes are forced by
operator updates is not unique. In the above example, where we assumed p2 > p1 ,
we could also have acted first on |α (p2 )i with the new operator at p2 . Then the states
277
|α(p+1)
|α(p)
FIGURE 60. Diagonal update. An operator is either inserted, [0, 0] p → [1, b] p , at a randomly selected
bond or removed, [1, b] p → [0, 0] p . In the former case, the update is canceled if the spins at the chosen
bond are parallel, and therefore the state |α (p − 1)i must be known. The acceptance probabilities depend
on the number of operators n in the sequence before the update attempt, according to Eq. (265).
|α (p2 + 1)i, . . ., |α (L − 1)i as well as |α (0)i, . . ., |α (p1 )i would be affected. There are
always two ways of changing the states in this way, but only one may be allowed in
any given case because of the constraints, as illustrated in Fig. 59. In most cases, such
updates done at random locations would not at all satisfy the constraints, and one has to
specifically look for two (or more) operators that can be updated [205].
As seen in Figs. 55 and 57, the state |α i can some times be updated without any
changes to the operator sequence—since there is no operator acting on the spin σ (1),
its state is arbitrary. Such free spins are very rare at low temperatures, when the average
number of operators on each spin is large.
Focusing on the operator sequence, we will sample the number n of non-unit operators
[a, b] p 6= [0, 0], their positions p ∈ {0, . . . , L − 1} in SL , bond indices b ∈ {1, . . ., Nb }, and
types induces a ∈ {1, 2}. The general strategy is to let the expansion order n change only
in exchanges of diagonal operators with the fill-in unit operator; [0, 0] p ↔ [1, b] p (where
n → n±1). Then, keeping all lattice locations b of the operators fixed, we change the type
index, [1, b] p ↔ [2, b] p for some set of operators (as in the example with two operators
above, but we will do it in a much more efficient way involving an arbitrary number
of operators). With these off-diagonal replacements carried out properly, the sampling
is ergodic in combination with the diagonal n → n ± 1 updates. We next consider the
details of these updating procedures. We first assume that the expansion cut-off L has
been properly determined and then discuss how this can be ensured in practice.
Diagonal updates. Updates of single diagonal operators, illustrated in Fig. 60, can
be carried out sequentially at the locations p = 0, . . . , L−1 in SL . There are no constraints
involved in such an update in the direction [1, b] p → [0, 0] p (removal of a diagonal
operator), whereas an insertion of a diagonal operator, [0, 0] p → [1, b] p, is allowed only
if the spins on bond b are antiparallel, σ (i(b)) 6= σ ( j(b)), in the propagated state |α (p)i
on which the operator acts. We therefore have to generate these states (storing only
the one currently needed), which is simply done by flipping the two spins σ (i(b)) and
σ ( j(b)) each time an off-diagonal operator [2, b] p is encountered (in which case no
single-operator update can be carried out at p).
The weight ratio to use in the Metropolis acceptance probability is easily obtained
from (264). Note, however, that we also need to correct for the inherent imbalance of
these update attempts—there is only a single unique way of carrying out the update in
the direction [1, b] p → [0, 0] p, whereas when changing [0, 0] p → [1, b] p the bond b should
be generated at random among the Nb possible locations [and if, for the chosen b, the
spins σ (i(b)) = σ ( j(b)), the update is immediately rejected and we proceed to the next
p]. Therefore, following the criteria for detailed balance discussed in Sec. 3.2, specif-
ically the transition probability (43) written as a product of selection and acceptance
278
probabilities, we have to include the ratio Nb of the selection probabilities for updates
involving a specific bond b. We then obtain the following acceptance probabilities;
β Nb

Paccept ([0, 0] → [1, b]) = min ,1 , (265)
2(L − n)

2(L − n + 1)
Paccept ([1, b] → [0, 0]) = min ,1 , (266)
β Nb
where n is the number of operators before the update (and after n → n + 1 or n → n − 1).
To prove detailed balance for the diagonal updates, it may seem necessary to carry
them out at random positions p, not sequentially. Otherwise, after having carried out an
update at some position p, there is zero probability of carrying out the reverse update as
the next step. However, if we consider the whole sequence (sweep) of updates from p = 0
to P = L − 1 and follow this by a reverse updating sequence, starting from P = L − 1 and
ending at p = 0, then detailed balance holds for the sweeps. In practice, one does not
have to switch between the two directions of updating the operator sequence.
Off-diagonal updates. Updates involving off-diagonal operators clearly have to in-

volve at least two operators in order to maintain the periodicity constraint on the prop-
agated sates. The simplest kind of pair update was already discussed and illustrated in
Fig. 59. For a 1D system, such updates are ergodic with open boundary conditions, but
with periodic boundaries local updates cannot change the topological winding number
(which in clear from Fig 50, where no local deformations can affect cyclic permutations).
This may still not be a very serious problem in principle, because for any T > 0 only the
sector with zero winding number contributes when the system size N → ∞ (in practice
for some large N, larger for lower T ). If we are interested in ground state properties, we
also can obtain the correct result when T → 0 for fixed N, because in this limit an ef-
fective winding number defined within a finite range p, . . . , p + D p , with D p ≫ N, of the
propagated states |α (p)i (corresponding to a finite range of imaginary times) can still
fluctuate freely, irrespective of the constraint of zero global winding number. Physical
quantities are then insensitive to the winding number [206]. To converge to the ground
state, much lower T has to be used, however.
In addition to their inability to change the winding number in one dimension, in higher
dimensions the kind of operator substitutions shown in Fig. 59 cannot lead to any of
the permutations of same spins (permutation of particles in the boson language) that
should be included (while in one dimension, cyclical permutations of all spins are the
only possible permutations in the kind of system we consider here). To remedy this
problem, one can construct other kinds of local updates, e.g., involving two operators
acting on the corners of a plaquette of 2 × 2 sites on a 2D square lattice [205], which can
sample among all such permutations within a given sector of winding numbers. Updates
changing the topological winding number has to involve lines of sites wrapping around
the boundaries, but for lattices of length larger than ≈ 20 these targeted updates become
very unlikely (i.e., it becomes too difficult to satisfy the constraints when a large number
of operators are changed simultaneously) [186, 205].
Fortunately, instead of dealing with updates specifically targeting permutations and
winding number sectors, there is a very efficient and simple type of loop update that
279
p i =1 2 3 4 5 6 7 8 s(p) i =1 2 3 4 5 6 7 8 s(p)
11 4 5
10 0 0
9 9 8
8 13 13
7 6 6
6 0 0
5 0 0
4 4 5
3 13 13
2 0 0
1 9 8
0 14 14
FIGURE 61. A linked-vertex SSE configuration with one loop traced out and shown in both of its
“orientations”, along with the corresponding operator-index sequences. All spins covered by the loop are
flipped, and operators are changed, diagonal ↔ off-diagonal, each time the loop passes by (with no net
change of an operator visited twice). Every vertex leg (spin) belongs uniquely to one loop, and spins not
acted upon by any operator (here the one at i = 1) can also be regarded as forming their own loops.
accomplishes all these things automatically. This class of updates was initially intro-
duced as a generalization of a cluster algorithm for the Ising model to a model where the
flipped clusters take the form of loops; the classical six-vertex model [191]. The effec-
tive world line system for the S = 1/2 Heisenberg model constructed using the discrete
Suzuki-Trotter decomposition is exactly equivalent to an anisotropic six-vertex model,
and the loop update for it was therefore at the same time a generalization of the clas-
sical cluster update to a quantum mechanical system. These ideas were subsequently
applied also to continuous-time world lines [179] as well as to the off-diagonal updates
in the SSE method [190]. The improvements in performance relative to local updates
are enormous (as in the classical case, leading to a much reduced dynamic exponent)
and brought simulations of quantum spin systems to an entirely new level. Like classical
cluster algorithms, the loop updates are in practice limited to certain classes of models,
of which the isotropic Heisenberg systems is one. Generalizations of the loop concept to
worms [32] and directed loops [33] (both of which can be regarded as loops that are al-
lowed to self-intersect during their construction, unlike the original loop updates where
no self-intersection is allowed) are applicable to a wider range of systems.
For the S = 1/2 model considered here, there is no reason to even discuss local off-
diagonal updates in any greater detail, and we will just focus on how to implement the
much more powerful loop updates. In the case of the SSE method, the operator string is
again the main focus, and the loop update corresponds to constructing a loop of operators
(vertices) connected by the links in the linked-list representation.
Operator-loop updates. An example of an operator-loop and how it is flipped is

shown in Fig. 61. Here “flipping” refers to the spins along the loop (explicitly those on
the vertex legs and implicitly in all propagated states covered by the loop) as well as
280
the operators themselves when we map the changes back to the sequence SL ; a diagonal
vertex with two legs attached to a loop changes to off-diagonal, and vice versa. In the
case of all four legs of a vertex belonging to the same loop (and example of which is
present in the figure), the operator type does not change.
The key aspect of the loop update is that the weight (264) only depends on the number
of operators, which does not change when the operator-loop is flipped. Therefore,
according to the detailed balance rules, such a flip can always be accepted. It is also clear
that loops can be large, and a loop flip can therefore lead to changes in the global spin
permutations. The previously discussed pair substitutions correspond to flipping loops
of two operators on the same bond. Any operator substitution specifically constructed
to sample permutations can also be formulated as an operator-loop, including ones that
change a winding number.
Note that the loops are completely deterministic once the operator locations have been
specified and each spin in the full space-time configuration belongs uniquely to one
loop. The diagonal update is the de facto mechanism by which the loops are changed,
and the purpose of the loop update is just to identify and flip some of the loops. It is
then best to construct all the loops and to flip each of them with probability 1/2 (which
also maintains detailed balance), instead of constructing loops at random and always
flipping them (in which case some loops would be constructed more than once and
flipped unnecessarily). Free spins, on which there are no vertices (such as the first spin
in Fig. 61), can also be considered as loops, since flipping such a spin implicitly flips a
whole line of spins in the full configuration of time-periodic propagated state.
5.2.3. Computer implementation

A convenient definition of a Monte Carlo sweep in the SSE method is a full sequence
of diagonal updates, followed by construction (and flip with probability 1/2) of all loops.
From the loop illustration in Fig. 61, it is clear why the linked-vertex storage X () of the
SSE configuration is useful; constructing a loop corresponds to moving in this list using
very simple rules (following a link or moving laterally to the adjacent spin on the same
vertex). On the other hand, in the diagonal update the original storage illustrated in
Fig. 55, using a single spin state σ () and the operator string s(), is more convenient.
The configuration will always be stored in this way, and the linked vertices will be
constructed before each set of loop updates. The loop updates are carried out by moving
in the linked list X (), updating the corresponding operators in s() at the same time (as
indicated in Fig. 57). The stored spins σ () can also be affected by the loop updates,
and this can be taken care of after all loop flips have been carried out. We now discuss
pseudocode implementations of all the main steps involved in a Monte Carlo sweep.
Diagonal update. We assume that we currently have a valid configuration stored

(which in the beginning of the simulation can be a random spin state and an operator
sequence with only [0, 0] elements), using σ (i) = ±1 for the individual spins and the
compact storage of [a, b] p as single integers s(p), with even and odd numbers 2b and
2b + 1 corresponding to diagonal and off-diagonal operators, respectively (as illustrated
281
in Fig. 55). A sweep of diagonal updates can then be implemented in the following way;
do p = 0 to L − 1
if (s(p) = 0) then {27}
b = random[1, . . ., Nb ]; if σ (i(b)) = σ ( j(b)) skip to next p
if (random[0 − 1] < Pinsert (n)) then s(p) = 2b; n = n + 1 endif
elseif (mod[s(p), 2] = 0) then
if (random[0 − 1] < Premove (n)) then s(p) = 0; n = n − 1 endif
else
b = s(p)/2; σ (i(b)) = −σ (i(b)); σ ( j(b)) = −σ ( j(b))
endif
enddo
Here Pinsert (n) and Premove (n) are the acceptance probabilities (265) and (266) for insert-
ing and removing a diagonal operator when the current expansion order is n. These prob-
abilities can be precalculated or evaluated on the fly. One can also just precalculate the n-
independent constant β Nb /2 and accept a removal if random[0 − 1] × (L − n) < β Nb /2
and an insertion if random[0 − 1] × β Nb /2 < L − n + 1. This avoids the more time
consuming divisions required with the full acceptance probabilities (265) and (266), or
storage of a large number of the n-dependent probabilities.
The sites belonging to bond b can be stored in a list [i(b), j(b)]. With the exception of
this list, the SSE sampling algorithm is lattice independent. It is also easy to modify the
above code for non-uniform (b-dependent) Heisenberg coupling strengths. The diagonal
update is the only stage at which the interactions appear explicitly, and no modifications
at all are required in the loop update.
Construction of the linked vertex list. We will construct the linked list by traversing
the operators s(p), starting from p = 0. Recall that for vertex leg number l = 0, 1, 2, 3 of
an operator at p, the corresponding links are stored as X (v) with v = 4p + l, as illustrated
in Figs. 57 and 58. To construct the list, we will make use of two other data structures.
The position v in X () of the first vertex leg for spin i will be stored as Vfirst (i), while
Vlast (i) will be the position of the last leg on i encountered so far. Note that the first
leg on a given spin is always below the operator (before the operator has acted, l = 0, 1),
whereas the last leg is above the operator (l = 2, 3). Initially all Vfirst (i) and Vlast (i) are set
to −1. Later, when encountering an operator s(p) acting on spin i and Vlast (i) ≥ 0, then
Vlast (i) is the previous list element corresponding to an operation on i. We then use the
corresponding lower leg (l = 0 or 1, depending on which of the two spins acted on equals
i) of the new operator and set the links between v = 4p + l and Vlast (i); X (v) = Vlast (i)
and X (Vlast (i)) = v. On the other hand, if Vlast (i) = −1, then we have found the first
operation on i and set Vfirst (i) = v. In both cases, we set the element for the last operation
on i as Vlast (i) = v + 2, where the addition of 2 corresponds to the vertex leg on spin i
after the operation (leg index l = 2 or 3). For each operator, we have to examine both
the spins i1 and i2 on which it acts. In the following pseudocode, the vertex list positions
to be filled for the operator at p are v0 + l, with v0 = 4p, where the leg pairs l = 0, 2 and
l = 1, 3 correspond to spins i1 and i2 , respectively. We use v1 and v2 to denote the list
elements of the last (previous) operation on the two spins. All links, except those across
282
the boundary p = L − 1, 0, can be constructed according to;
do p = 0 to L − 1
if (s(p) = 0) skip to next p {28}
v0 = 4p; b = s(p)/2; i1 = i(b); i2 = j(b); v1 = Vlast (i1 ); v2 = Vlast (i2 )
if (v1 6= −1) then X (v1) = v0 ; X (v0) = v1 else Vfirst (i1) = v0 endif
if (v2 6= −1) then X (v2) = v0 ; X (v0) = v2 else Vfirst (i2) = v0 + 1 endif
Vlast (i1) = v0 + 2; Vlast (i2 ) = v0 + 3
enddo
To connect the links across the time boundary we can use the arrays containing the first
and last list positions;
do i = 1 to N
f = Vfirst (i) {29}
if ( f 6= −1) then l = Vlast (i); X ( f ) = l; X (l) = f endif
enddo
Since there are also positions in X () containing no links, corresponding to fill-in oper-
ators s(p) = 0, we should have some way to distinguish these. We can set all X (v) to a
negative number, e.g., X (v) = −1, before constructing the links. Then any v for which
X (v) ≥ 0 can be used as a starting point to trace a loop.
Implementation of the operator-loop update. We want to trace all the loops and flip
each of them with probability 1/2. In order to make sure that no loop is considered
more than once, we should always start a new loop from a position in X () not previously
visited. We then need some flag to indicate whether a position has been visited or not.
Instead of allocating separate storage for such flags, we can actually use the list X ()
itself—the links stored in it only have to be used once, and as soon as a link at v has
been used we can set X (v) to a negative number, to indicate that this position should
not be used again. To start a new loop, we can look for the first position v0 for which
X (v0) ≥ 0. Before traversing the loop, we make the random decision of whether or not to
flip it. If we do not flip, the loop tracing still has to be carried out, in order to flag all the
corresponding X (v) as visited. For a flipped loop, we should also change the operators
in s() to reflect the change diagonal↔off-diagonal for each visited vertex. We will take
care of possible flips of free spins in the stored state |α i later. The structure of a code
for tracing and flipping the loops is;
do v0 = 0 to 4L − 1 step 2
if (X (v0) < 0) skip to next v0 {30}
v = v0
if (random[0 − 1] < 1/2) then
• traverse the loop; for all v in loop, set X (v) = −1
else
• traverse the loop; for all v in loop, set X (v) = −2
• flip the loop (change operator types s(p = v/4) while loop is traversed)
endif
enddo
283
Here v0 is a tentative starting point for a new loop, and we actually start a new loop
only if X (v0 ) is a non-negative number. Note that we only have to consider even starting
points (the step 2 on the first line of {30}), because of the structure of the vertex list
and the loops, i.e., two consecutive (even, odd) elements in X (v) always correspond to
the same loop. To not visit the same loop more than once, we set X (v) = −1 for loops
that are only visited, and X (v) = −2 for loops that are also flipped. We will need the
distinction between flipped and only visited loops later, when updating the stored state
|α i to reflect the flipped loops.
To trace a loop starting at some v0 , we can use the leg index l0 = mod[v0 , 4] and move
to the adjacent leg l0′ on the same vertex. With the leg labeling convention in Fig. 58, the
rule for the leg adjacent to a leg l can be summarized as l = (0, 1, 2, 3) → l ′ = (1, 0, 3, 2).
This rule can be very efficiently implemented using a bit-level operation on l, as l ′
corresponds to flipping (0 ↔ 1) the lowest bit of l. Here we just assume that such
functionality is available in the programming language used [e.g., in Fortran 90 one
can use the exclusive-or operation of the integer l with 1; l ′ = ieor(l, 1)] and assign
l0′ = flipbit(l0 , 0), where 0 is the bit flipped. We actually do not need to extract the leg
index itself, but can do the corresponding move to the adjacent leg just by manipulating
its location v0 = 4p0 + l0 in X (). We can use the same bit flip method to find it;
v′0 = flipbit(v0 , 0). To move to the next vertex, we use the link; v1 = X (v′0). This
completes one step of the loop tracing procedure. We next proceed in the same way
to find the location v′1 of to the adjacent leg, and from there we move to the leg v2 linked
to it. This continues until at some step k we find vk = v0 . We have then completed a full
loop. In pseudocode form the traversal of a loop and flipping it can be accomplished by
(where no subscripts or primes are needed on the visited list locations v);
v = v0
do {31}
p = v/4; s(p) = flipbit(s(p), 0); X (v) = −2
v = flipbit(v, 0); X (v) = −2
v = X (v); if (v = v0 ) exit
enddo
Here the flip of the operator type diagonal↔off-diagonal is also accomplished using
the flipbit procedure, which corresponds exactly to the changing the operator code s(p)
from 2b (diagonal operator) to 2b + 1 (off-diagonal operator) or vice versa. In the case
in code {30} where the loop is not to be flipped, the only difference is that the two
statements involving p and s(p) in {31} are absent, and we mark all visited locations as
X (v) − 1 instead if −2 (information which will be used to update the state |α i).
After all loops have been traced, we also have to update the spins in the stored state
|α i. For each site i which has a flipped loop passing through it, the stored spin σ (i)
should be flipped. We should also flip with probability 1/2 any spin that has no loop
connected to it (i.e., spins on which no operator acts). Both these tasks can be taken care
of using the information stored in Vfirst (i) during the construction of the linked vertex list.
If v = Vfirst (i) = −1, that implies that there is no operator acting on spin i. Otherwise,
we can use the flag −1 or −2 now stored in X (v) to determine whether the loop passing
284
through i has been flipped [with X (v) = −2 for flipped loops]. In pseudocode form:
do i = 1 to N
v = Vfirst (i) {32}
if (v = −1) then
if (random[0-1]< 1/2) σ (i) = −σ (i)
else
if (X (v) = −2) σ (i) = −σ (i)
endif
enddo
This completes the loop update, and we can repeat the cycle consisting of: (i) perform-
ing a sweep of diagonal updates, (ii) constructing the linked vertex list, and (iii) trac-
ing/flipping all the loops and updating |α i accordingly. These procedures constitute one
Monte Carlo sweep in the SSE method.
Equilibration and expansion truncation. We now discuss the expansion cut-off L

(when the fixed-length scheme is used). We should make sure that L exceeds the max-
imum expansion order n that will be sampled, so that the truncation in practice is not
an approximation. Since new diagonal operators are inserted through exchanges with
the fill-in unit operators, we also want there to be a reasonably large number of these
s(p) = 0 elements present. But since the memory and CPU time scales with L, we should
not make L excessively large. One could in principle define an optimal L, for which the
acceptance rate of the diagonal updates is maximized. In practice, it does not matter
much, however, whether L is really optimal, since the off-diagonal updates are anyway
more important in determining the autocorrelation times. We will therefore just make
sure that some significant fraction of the elements s(p) = 0, so that the sampled n never
reaches L and that many insertions of operators can be attempted in the diagonal update.
We can aim √for L ≈ (1 + a)hni with, e.g., a = 1/3. Since the relative fluctuations of n
are small, ∼ n, we can do this by letting L → a × n and augmenting SL with additional
s(p) = 0 elements after each Monte Carlo sweep when a × n > L [setting s(p) = 0
for Lold + 1 < p < Lnew or distributing the zeros randomly. Fig. 62 shows an example
of how this procedure works in practice, during equilibration of a 16 × 16 system at
inverse temperature β = 16. Initially L was set to N/2 = 128. The adjustment of L
quickly converges to an acceptable value, and, as shown in the inset, in a subsequent
long simulation n never comes close to L.
5.2.4. Estimators for physical observables

The autocorrelation time in SSE simulations with loop updates is typically very
short; often just a few Monte Carlo sweeps or less [33, 31]. Calculations of physical
observables (“measurements”) should therefore normally be performed after every (or
every few) Monte Carlo sweep. As discussed in Sec. 3.2, the autocorrelation times do
not dictate how often measurements may be done, only with what frequency independent
data are generated. Autocorrelations, thus, affect the error bars but have no effect on the
285
L
6000 nmax
n
5000
4000 0.006
n
0.005
3000 0.004
P(n)
0.003
2000
0.002
0.001
1000
0
4600 4800 5000 n
0
0 50 100 150 200 250
MC sweeps
FIGURE 62. Evolution of the expansion cut-off L at the initial stage of an equilibration run of a 16 × 16
Heisenberg system at β = 16. The number of operators n in the string after each Monte Carlo sweep is
also shown, along with the maximum n reached so far. The final cut-off after 5000 sweeps was L = 6764.
The inset shows the distribution of n in a subsequent run of 107 MC sweeps.
correctness of computed expectation values (i.e., it is not wrong to measure correlated

data, although it may be wasteful if the measurements take a long time to evaluate)
provided that the total simulation time is much longer than the autocorrelation time. Data
binning should be used to compute error bars reliably, as also discussed in Sec. 3.2. We
here discuss several types of observables of interest in SSE calculations. More details
and derivations of the expressions can be found in Refs. [186, 187, 207].
Energy and specific heat. We already discussed some observables in Sec. 5.1.3; the
internal energy (250) and the related expectation value (251) of an individual operator
in the hamiltonian, as well as the specific heat (252). Given that hHi = −hni/β , it may
seem surprising that the specific heat C = (hH 2 i − hHi2 )/T 2 is not given just by the
fluctuation in n, but there is also a term −hni. This is because hH 2 i = hn(n − 1)i/β 2 ,
which can be easily shown using the same procedures leading to hHi = −hni/β in
Sec. 5.1.3. Note that this expression includes the constant 1/4 subtracted from each bond
operator in (256), but that this constant cancels out in the specific heat expression (252).
The specific heat in practice becomes difficult to compute reliably (i.e., its statistical
error is large) at low temperatures (where it becomes small), because it is the difference
of two large numbers (∼ N 2 β 2 ).
Diagonal operators. Expectation values of operators diagonal in the z-component

basis are also easy to evaluate, using averages over the propagated states;
1 L−1 1 L−1
hOz i = ∑
L p=0
h α (p)|O z | α (p)i = ∑ Oz (p).
L p=0
(267)
286
It is not necessary to evaluate this average using all p, as successive propagated states
differ by at most two flipped spins. One can instead use a partial summation over p,
e.g., Oz (0) + Oz(N) + Oz (2N) + . . ., to save time. In some cases, however, one may as
well just propagate Oz (p) from p = 0 to p = L − 1 by first computing Oz (0) and then
update the value Oz (p) according to the spin flips occurring when propagating |α (p)i.
For example, the fully averaged contribution to the expectation value of the squared
staggered magnetization from an SSE configuration can be computed as follows:
m = (1/2) ∑i φi σ (i); ms2 = 0
do p = 0 to L − 1 {33}
if (mod(s(p),2)=1) then
b = s(p)/2; σ (i(b)) = −σ (i(b)); σ ( j(b)) = −σ ( j(b))
m = m + 2φi σ (i(b))
endif
if (s(p) 6= 0) ms2 = ms2 + m2
enddo
ms2 = ms2 /(nN 2 )
Here φi = ±1 is the staggered phase factor for site i and m contains the staggered
magnetization, which evolves as the operator string is traversed. The change in m when
two antiparallel spins are flipped can be expressed as 2φi σ (i(b)), because φ j σ ( j(b)) =
φi σ (i(b)). The average of m squared is accumulated as ms2 , in this example only using
the n propagated states generated by the original index sequence without the fill-in unit
operators [i.e., skipping the steps where the operator s(p) = 0]. One can also sum over
all L instances of the states, in which case the result would be divided by L instead of n
on the last line. As written above, the special case n = 0 is not treated correctly, but the
code is easy to modify by just using ms2 = m2 /N 2 in that case.
The computational effort of the measurement in {33} scales as β N, i.e., the same as
the SSE sampling algorithm. To compute the Fourier transform of the spin correlations
(the static structure factor) at some arbitrary momentum [noting that m2s corresponds to
S(π ), with π denoting the staggered wave-vector, e.g., π = (π , π ) in two dimensions],
e.g., for use in the correlation-length definition (70) or (73), one can use a similar
procedure for the real and imaginary parts of m(q),
m(q) = ∑ Srz cos(q · r) + i ∑ Srz sin(q · r), (268)

r r
with φi in {33} replaced by the corresponding sine and cosine factors [and explicitly
considering the changes from both the flipped spins at i(b) and j(b)]. The structure
factor is then accumulated according to S(q) = S(q) + Re2 {m(q)} + Im2 {m(q)}.
For computing the full correlation function,
Cz (ri j ) = hSizSzj i, (269)
averaged over all i, j [or the Fourier transforms for all q, which can be written as
hm(q)m(−q)i], it would not be practical to update all these ∝ N different functions after
each spin flip as in {33}. It is then better to compute all the correlations from scratch
in, e.g., every N:th propagated state. This still leads to a rather expensive scaling ∝ β N 2
287
of the effort to measure all the correlations—a factor N worse than the sampling. One
should then judge whether it is better to compute only a subset of the correlations [e.g.,
along some lines in the (x, y) plane] or just measure them less frequently, in order for
the measurements not to completely dominate the calculation.
Susceptibilities. Another important class of observables are generalized susceptibil-

ities, i.e., linear response functions of the form
∂ hA(b)i
χAB = , (270)
∂b
where b is the prefactor in a field term bB added to the hamiltonian and A is the operator
whose response to this perturbation we want to compute. For example, we may be
interested in the response χi j at site j when a magnetic field acts only on site i, in which
case A = Szj and B = Siz . Such a susceptibility is given by the Kubo formula [186]
Zβ
χAB = d τ hA(τ )B(0)i − β hAihBi, (271)
0
where A(τ ) = e−τ H Aeτ H . If both A and B are diagonal, this Kubo integral can be
evaluated in SSE simulations using the generic formula
Zβ
* ! !+ * +
β n−1 n−1
β n−1
d τ hA(τ )B(0)i =
n(n + 1) ∑ A(p) ∑ B(p) + ∑ A(p)B(p) .
n + 1 p=0
p=0 p=0
0
(272)
The sums over A(p) and B(p) can be computed using code similar to {33}.
We are often interested in susceptibilities for which A = B, e.g., the magnetic response
χ (q) at wave-vector q, in which case A = B = m(q), with m(q) the Fourier transform of
the spin configuration given by (268). Since the exact value of χ (q) must be real-valued,
one only has to compute the real part of (272). If all local response functions χi j =
χ (ri j ) are computed in a simulation, the Fourier transform χ (q) can later be evaluated
using these. These response functions are directly accessible in NMR experiments; see
Ref. [208] for an example.
Note that if A and B commute with the hamiltonian, then A(τ ) = A(0) = A [and
A(p) = A and B(p) = B are independent of p in (272)] and the susceptibility reduces to
the classical expression χAB = β (hABi − hAihBi). In practice, for the Heisenberg model
the only case where this form applies is the uniform magnetic susceptibility;
β 2 N
χ = χ (0) = hM i, Mz = ∑ Siz . (273)
N z i=1
Here the term β hMz i2 in (271) vanishes, since H does not include a magnetic field.
If the operators A and B in (271) are not diagonal, the susceptibility is more com-
plicated in general. Here we will only discuss an important special case, in which the
288
estimator actually is very simple. If the operators involved are two terms of the hamil-
tonian, as defined in Eqs. (255) and (256) and here referred to just as HA and HB for
any two instances of those (any two diagonal or off-diagonal bond operators), then the
susceptibility measurement just involves counting the numbers N(A) and N(B) of those
operators in the sampled SSE operator sequences;
1
χHa HB = hN(A)N(B)i − δA,BhN(A)i .

(274)
β
The most important example of this type is the current susceptibility χIx Ix , where the
spin current operator Ix (here in the lattice x direction, for definiteness) is defined by
N
Ix = ∑ [S−(ri )S+(ri + x̂) − S+ (ri )S− (ri + x̂)], (275)
i=1
where we assume, for simplicity, a hamiltonian with only nearest-neighbor interactions.

For longer-range interactions, there would be corresponding current terms between the
same site pairs as in the hamiltonian. Although (275) is not exactly a sum of the off-
diagonal operators (256) used in the SSE sampling, those operators can be written as
sums of two parts,
H2,b = Hb+ + Hb− , Hb+ = S+j(b) Si(b)

−
, Hb− = S−j(b) Si(b)
+
, (276)
where the site pairs [i(b), j(b)] in each case are assumed to be ordered in such a way
that the + and − terms transport one unit of spin in the positive and negative direction
r j −ri of the bond, respectively. A key point here is that, although the SSE configurations
contain the full off-diagonal operators H2,b , only the + or − part in (276) contributes in
each instance (with the other part destroying the state it acts on). One can therefore use
(274) to evaluate local current-current response functions of the form
Zβ
Λb1 b2 = d τ hIb1 (τ )Ib2 (0)i, (277)
0
where Ib here denotes the current operator at bond b. Using (274) one obtains
Λb1 b2 = h[N + (b1 ) − N − (b1 )][N + (b2 ) − N − (b2 )]i − δb1 b2 hN + (b1 ) − N − (b1 )i, (278)
where N + (b) and N − (b) denote the number of operators in the SSE operator sequence
transporting spin in the positive and negative direction, respectively, across the bond
b. The meaning of this becomes clear when looking at a graphical representation of an
SSE configuration, such as Fig. 55. There one just has to sum the number of off-diagonal
events in which an ↑ spin is moved to the right (the positive direction) minus the number
of left-moving events. The term N + (b) − N − (b) is zero, unless the winding number in
the lattice direction defined by bond b is not zero, which can be seen in the illustration of
winding numbers in Fig. (50). It can be noted that the form (278) of the current response
is identical in SSE and path integral (world line) methods.
289
Spin stiffness. We discussed the spin stiffness in the context of the classical XY
model in Sec. 3.5. The basic definitions of this quantity at T = 0, Eq. (83), and at
T > 0, Eq. (86), are identical for quantum spin systems (XY or Heisenberg models)
as well as bosonic systems more broadly (where the analogous quantity is the superfluid
phase stiffness, which is proportional to the superfluid density). Deriving Monte Carlo
estimators for SSE or world line methods, one finds that the term β hIx2 i in (97), which
is a classical response function, should be replaced by the corresponding quantum
mechanical Kubo formula, giving the spin stiffness in the form
Zβ
1 1
ρs = hHx i − d τ hIx (τ )Ix (0)i, (279)
N N
0
where we assume that the phase twist is imposed in the lattice x direction. The Kubo
integral here consists of a sum of bond-current response functions of the form (277),
and using the result (278) one finds that the energy term hHx i is exactly canceled by the
δ -function terms coming from (278). This leads to a very simple expression for the spin
stiffness in terms of the SSE (or world line) configurations:
1
ρs = h(Nx+ − Nx− )2 i. (280)
βN
Here Nx+ and Nx− denote the total number of operators transporting spin in the positive
and negative x direction, respectively. For a spatially isotropic system in two or three di-
mensions, this can of course be averaged over the dimensions, whereas for an anisotropic
system the stiffness depends on the direction.
The spin stiffness is often expressed as the fluctuation of the winding number (here in
the x direction, with analogous expressions for other directions),
1 +
Wx = (N − Nx− ), (281)
Lx x
i.e., the size normalized current (Wx = 0, ±1, ±2, . . .). For a d-dimensional system with
N = Ld the stiffness is just ρs = hWa2 i/β for any direction a, but for other shapes (e.g.,
N = Lx × Ly with Lx 6= Ly ) the lengths also enter and (280) looks simpler.
In derivations of the spin stiffness, it is normally for simplicity assumed that the spins
order in the XY plane (in spin space), whereas in a Heisenberg model the direction of
the order parameter in spin space is not restricted to this plane (the symmetry is not
broken) in an SSE simulation of a finite system. A simple rotational averaging argument
shows that the expression (280) should be multiplied by 3/2 in order to obtain the correct
stiffness of a symmetry-broken state (i.e., when considering the thermodynamic limit).
The expression (280) applies also to the quantum XY model [209], where the result
should not be multiplied by 3/2.
A winding number estimator completely analogous to (280) was originally derived in
the context of the superfluid phase stiffness of bosons in continuous space [210]. The
role of global cyclical permutations of particles in superfluidity dates back to Feynman’s
pioneering work on path integrals [175].
290
6 1 4 2 6 1 4 2
1 1 4 2 1 1 4 2
5 3 1 1 5 3 1 1
5 3 3 3 5 3 3 3
FIGURE 63. Example of clusters formed by space-time loops passing through a propagated state |α (p)i
(for arbitrary fixed p). Here there are six such clusters, labeled 1, . . . , 6. Open and solid circles correspond
to ↑ and ↓ spins, respectively, in |α (p)i. When a loop is flipped, all spins in the corresponding cluster are
also flipped, as indicated here with two different configurations corresponding to the two states of cluster
1 (the sites enclosed by larger circles). Note that the spins within each cluster are always in one of the two
staggered configurations.
5.2.5. Improved estimators

The operator-loop update in the SSE method (as well as loop updates more broadly
[31, 191]) is an example of a cluster update. Such non-local updates were first developed
for classical Monte Carlo simulations of the Ising model [118]. One aspect of cluster
methods is that it is possible to take averages of estimators for physical quantities over
all orientations of the clusters, because the configuration weight does not change upon
flipping a cluster. This is immediately clear in SSE simulations of S = 1/2 Heisenberg
models, because the weight (264) only depends on the number of operators n in the
sequence, which does not change when a loop is flipped. If the number of clusters (here
operator-loops) is m, then the total number of equal-weight configurations is 2m , and
the average over all of these configurations can provide a much less noisy estimator
than one depending on just a single configuration. The crucial point here is that, for
many important quantities, this average can be computed analytically, and the resulting
improved estimator is of a simple form that can be evaluated rapidly in simulations. Here
we only discuss the rather simple cases of the the static (equal-time) structure factor and
the uniform magnetic susceptibility. For improved estimators for some other quantities,
see the review article by Evertz [31].
Consider a propagated state |α (p)i, e.g., the stored |α (0)i. In the linked vertex
representation of the SSE configuration, illustrated in Fig. 61, there is a loop passing
through each of the spins in this state (with spins without operators acting on them also
considered as individual loops). The same loop can go through many spins in |α (0)i, and
all spins belonging to the same loop form a cluster, in the sense that if the loop is flipped
all the spins in the clusters are flipped simultaneously. Note that the loops are objects
in space-time, while the clusters discussed here are defined on a cut at fixed time (here
propagation index p). A cluster can consist of several parts that appear disconnected in
space, since such pieces can be connected in the larger space-time volume where the
loops exists. An example of clusters on a 2D lattice is shown in Fig. 63
Since we are dealing with a bipartite lattice, and because the loop structure is such
that the spin on a vertical loop segment (referring to pictures such as Fig. 61) changes
each time one changes direction when moving along a loop, the spins within a cluster
formed at a given state |α (p)i always have a staggered structure. The staggered magne-
tization ms ( j) of a cluster labeled j, with j = 1, . . .,C, where C is the total number of
291
clusters, is then ms ( j) = ±n j , where n j is the number of spins in cluster j. For a given
configuration, the total staggered magnetization Ms = ∑Cj=1 ms ( j). When averaging the
square of this sum over all the different realizations of cluster orientations, the cross
terms hms (i)ms( j)i = 0 (for i 6= j). One is then left with just the i = j contributions, and
the staggered structure factor is simply given by
* +
C
1
S(π ) =
4N j=1 ∑ nj .
2
(282)
Structure factors at other wave-vectors q are only marginally more complicated, de-
manding in place of the cluster sizes n j a summation over each cluster of the phases
φr exp(iq · r), where r refers to sites on a given cluster and φr = ±1 is the staggered
phase factor, which takes care of the staggered spin structure within the clusters (and the
denominator 4 corrects for the fact that the spin values are ±1/2). One can here also use
the fact that the true structure factor must be real-valued for any q.
In principle, equal-time correlation functions such as the structure factor can also
be averaged (fully or partially) over the propagation index p, as in Eq. (267). This,
however, requires more work for the improved estimator than in code {33} for the simple
estimator, because it takes some book keeping during the loop update to construct the
clusters for several fixed p, and doing so may not always pay off. Without this averaging,
however, a simple p-averaged estimator, such as the one implemented in code {33}, may
actually give better results at low-temperatures, where the gain due to averaging can be
very significant. The case q = 0 is special in this regard, because this corresponds to the
total squared magnetization, which is a conserved quantity (i.e., independent on the SSE
propagation index p), and no further averaging over p can then be done to improve the
statistics further. The optimal estimator for the uniform susceptibility (273) is therefore
* nj
!2 +
β C
χ=
4N j=1∑ ∑ φi . (283)
i=1
Susceptibilities at other wave-vectors involve the full space-time loop structure, not just
the clusters (cut through the loops) formed at a fixed state. For example, the staggered
susceptibility is given by the sum of the squares of all the loop sizes [31].
5.2.6. Program verification

QMC programs should always be verified by comparing results for small systems
with exact diagonalization data. When correctly implemented, the SSE method should
be exact, which means that the deviation of a computed quantity from its true value
should be purely statistical, due to the finite number of sampled configurations. We
have discussed how to quantify the statistical fluctuations in terms of “error bars” in
Sec. 3.2. Deviations beyond the error bars are due either to programming errors of flaws
in the random number generator used. While most programming errors would lead to
obviously wrong results, there are also possible subtle errors that may only lead to
292
0.08
improved estimator
standard estimator
0.06 -6
4×10
χ
-6
2×10
χSSE-χexact
0.04
0
-6
-2×10
0.02
-6
-4×10
0 0.5 1 1.5 2
0
0 0.5 1 1.5 2
T/J
FIGURE 64. The uniform susceptibility of a 4 × 4 Heisenberg system versus the temperature. The curve
is the exact result from a full diagonalization. The points are SSE results based on 1010 updating sweeps
for each T . The error bars cannot be resolved on this scale. The inset shows the deviation of χ from the
exact result with error bars, for both simple and improved (loop) estimators. The data points have been
slightly shifted off their actual T values in order for the error bars of the two estimators not to overlap.
minute deviations from exact results for small lattices. Such systematical errors may
grow with the system size, with potentially grave consequences. The same may be said
about random number generators; bias effects due to imperfect random numbers may
be very small for small lattices, but can become more significant for larger systems. It
is therefore important to make comparisons with exact diagonalization results based on
very long SSE runs, to detect possible small deviations. While it is impossible to strictly
prove that a program is correct in all respects, agreement to within very small error bars
with exact data makes this very likely.
Test results. We now discuss some test results for 1D and 2D Heisenberg models.
The random number generator used in these calculations (and most other SSE calcula-
tions discussed in these lecture notes) was a simple 64-bit linear congruential generator
with multiplier 2862933555777941757 [211] and addition of 1013904243.
Fig. 64 shows the susceptibility per spin of a 4 × 4 system. For each temperature
1010 updating sweeps and measurements were carried out, with the data subdivided into
100 bins for computing the statistical errors. For T = J/10, this required approximately
100 CPU hours on a mid-range PC workstation. The error bar (defined as one standard
deviation of the estimated fluctuation of the average, as discussed in Sec. 3.2) is typically
≈ 10−6 when the standard estimator for χ is used, and even smaller with the improved
estimator (283). The relative error (the error bar divided by χ ) is ≈ 5 ×10−6 for T /J ≈ 1.
The deviations of the averages are completely consistent with the size of the error bars.
Recall that for correctly computed statistical errors, one should expect about 2/3 of the
data points to bracket the true data within one error bar. The gain in precision with the
improved estimator can be much larger than in Fig. 64 for larger systems.
While it is essential to confirm the unbiased nature of the SSE calculations on small
293
-E/N
0.443148
0.443146
N = 1024 (SSE)
0.443144
N = 1024 (T=0, exact)
N = 4096 (SSE)
0.443142 N = 4096 (T=0, exact)
0.443140
8 9 10 11 12 13 14
m
m (β = 2 )
FIGURE 65. SSE results for the internal energy per spin at several inverse temperatures β = 2m for
Heisenberg chains of length N = 1024 and 4096 compared with the corresponding exact (Bethe ansatz)
ground state energies. The SSE results were obtained using 1 − 2 × 106 updating sweeps for each m, which
required several hundred CPU hours for the largest N, β .
lattices, one might still worry about potential problems with the random number gen-
erator for larger systems. In one dimension, we can also test SSE calculations against
exact Bethe ansatz results for very long chains in the ground state. To approach the
ground state in SSE calculations, it is convenient to use inverse temperatures of the form
β = 2m and go to sufficiently large m for calculated quantities to become m independent.
This approach is illustrated in Fig. 65, which shows the internal energy, E = hHi, com-
puted using the simple expansion-order estimator (250) for chains of length N = 1024
and 4096. Numerically computed Bethe ansatz results for these chain lengths are listed
in [212]. The agreement between the calculations is perfect within error bars, with no
detectable temperature dependence for the last three or four points for each N. This good
agreement, to within relative statistical errors as low as 5 × 10−7 , shows quite convinc-
ingly that the calculation is for practical purposes completely unbiased and that ground
states of even quite large systems can be studied.
5.3. Applications of SSE to 1D and 2D systems

We next discuss several illustrative results for 1D and 2D Heisenberg models obtained
with the SSE method. Results showing clearly the logarithmic corrections to critical
behavior in the 1D chain are presented in 5.3.1. The qualitatively different ground states
of n-chain ladder systems with even and odd n are discussed in 5.3.2. The sublattice
magnetization and some other low-energy parameters associated with the Néel state are
computed for the standard 2D Heisenberg model in 5.3.3, and in 5.3.4 dimerization is
introduced in this model, to drive a quantum phase transition into a plain non-magnetic
state. Quantum critical finite-size scaling behaviors of various quantities are analyzed.
The more complex case of a Néel–VBS transition is investigated in 5.3.5, using two
types of J-Q models that exhibit, respectively, continuous and first-order transitions.
294
-1
10
-2
10
C(r)
-3
10
N = 1024
N = 4096
-4
10
0 5 10
m
m (r = 2 )
FIGURE 66. Spin correlation function at distances r = 2m for chains of length N = 1024 and 4096.
The error bars are smaller than the symbols. The solid curve is of the expected form Ar−1 ln(r/r0 )1/2 ,
with A = 0.21 and r0 = 0.08. The dashed curve shows the form ∝ r−1 for comparison. These results
were obtained using inverse temperatures β = 213 and 214 for N = 1024 and 4096, respectively, which is
sufficient for T → 0 convergence.
5.3.1. The Heisenberg chain

Spin correlations at T = 0. In Sec. 4.3.1 we discussed Lanczos results for the spin
correlation function of the Heisenberg chain and saw some hints of the expected loga-
rithmic correction to the ∼ 1/r critical behavior (Fig. 34). The system sizes accessible
with the Lanczos method are not sufficient for studying these scaling corrections quanti-
tatively, however. As we saw above, with the SSE method unbiased studies of the ground
state is possible for chains of several thousand spins (with careful checks of the conver-
gence to the T → 0 limit). Fig. 66 shows the spin correlations for N = 1024 and 4096
at distances r = 2m , graphed on a log-log scale. To save time, only the correlations at
these distances were computed [for a scaling N log(N) of the time to carry out spatially
averaged measurements]. The results for the two system sizes coincide closely for r up
to 27 , indicating convergence to the infinite size values up to this distance for N = 1024
(and therefore up to r ≈ 29 for N = 4096, since the convergence behavior should scale
approximately linearly with N).
The expected form |C(r)| = A ln1/2 (r/r0 )r−1 [57, 58, 59] is very well reproduced up
to r = 29 for N = 4096. The parameters A and r0 obtained from a fit are listed in the
figure caption. If one leaves the exponent σ = 1/2 of the logarithm as a free parameter
to be obtained from the data based on a fit, the exponent indeed comes out close to 0.5,
but with a rather large error bar of, roughly, ±0.1. To really investigate the exponent
carefully, one should further increase the chain length (which is possible in principle).
If one did not know about the existence of a log correction and tried to extract the
form of the spin correlations on the basis of numerical calculations alone, one might at
first sight conclude that the decay is ∝ 1/rα with α ≈ 0.85, based on the data in Fig. 66.
There are, however, small but significant deviations from a pure power-law, which can
be detected only if the relative statistical errors are sufficiently small. In the data shown
295
0.15 N = 16
m
N = 2 , m=4,5,...,11
N = 4096
0.10
Jχ 0.05
0.00 -4 -3 -2 -1 0
10 10 10 10 10
T/J
FIGURE 67. The circles show SSE results for the uniform susceptibility of the Heisenberg chain at
temperatures T of the form 2−m , m = 0, 1, 2, ... for different power-of-2 chain lengths up to N = 4096.
Error bars are much smaller than the circles. The exact N = 16 result is shown as the solid curve and the
dashed curve shows the low-T form (284). The asymptotic T → 0 value, J χ = 1/π 2, is indicated by the
short horizontal line segment on the left side of the graph.
in the figure, the error bars are too small to discern. While the absolute errors are small,
typically ≈ 2 × 10−6 for N = 4096 (based on approximately 106 Monte Carlo sweeps at
β = 214 and using an improved estimator), the relative error is much larger, about 0.002
for r = 27 , growing to 0.02 for the longest distance r = 211 .
Low-temperature magnetic susceptibility. Logarithmic corrections are also impor-

tant in the Heisenberg chain at T > 0. The most prominent example is the uniform mag-
netic susceptibility, for which a renormalization-group study of the low-energy field-
theory predicted the form [7, 213]
1 1
χ (T ) = + , (284)
2π c 4π c ln(T0 /T )
where the T → 0 value x(0) = 1/2π c, where c = J π /2 is the spinon velocity, agrees
with the Bethe ansatz solution of the ground state of the Heisenberg chain. The Bethe
ansatz can also be extended to T > 0. Good agreement with the asymptotic expression
(284), with the parameter T0 = 7.7 (adjusted to fit the Bethe ansatz results), was found
at low temperatures (T /J < 0.02) [7]. Note that the logarithmic correction in (284)
implies a very slow convergence to the T = 0 limit, and it would therefore be difficult
to extrapolate numerical results without knowing about the presence of this kind of
correction. A higher-order logarithmic correction to (284) is also known [213], but here
we will just consider the leading-order correction.
Fig. 67 shows SSE results for the susceptibility obtained using chain lengths of the
form N = 2n and temperatures T /J = 2−m . Due to the excitation gap in chains with
finite N (discussed in Sec. 4.3.1), χ decays exponentially to zero below a temperature
T ∼ 1/N. The log-lin scale used in the figure makes these finite-size effects very clear
and also shows how the results converge rapidly to the thermodynamic limit form once
296
the finite-size gap becomes smaller than T . The agreement with the form (284), using
the value T0 = 7.7 determined in Ref. [7], is very good for T /J below 0.05 (in fact, for
some unknown reason, the agreement appears to be even better than in Ref. [7]).
The main utility of the form (284), beyond its role in establishing the correctness of
the low-energy field-theory, is that it is valid not only for the simple Heisenberg chain
considered here, but for any 1D spin system (at sufficiently low temperatures) which
is in the same phase as the Heisenberg chain. This includes the frustrated J1 -J2 chain
discussed in Sec. 4.3.2, for J2 /J1 less than the dimerization point (J2 /J1 )c ≈ 0.241.
Exactly at the dimerization point the leading logarithmic correction should vanish [164]
(but there are still other, higher-order corrections), while above the transition point a
spin gap opens and χ → 0 even for N = ∞. The asymptotic form (284) should be valid
throughout the QLRO(π ) phase of the chain including longer-range interactions [the
model discussed in Sec. 4.3.3], as well as in many other systems. The parameters c
and T0 depend on the model parameters and fitting of numerical data provides a way to
extract, in particular, the velocity c. As we will see in the next section, ladder systems
consisting of an odd number of chains also are in the same critical phase as the single
chain, and (284) applies also there.
5.3.2. Ladder systems

A ladder lattice consists of a fixed number Ly of coupled 1D chains (often referred to
as the legs of the ladder) with the chain length Lx taken to infinity (or, in practice, Lx /Ly
sufficiently large to give results converged to this limit). Strongly-correlated quantum
systems in this geometry [9, 214, 215] play an important role as a means of interpolating
between one and two dimensions. Many materials exhibit a structure of weakly coupled
ladders with small Ly , which motivates studies of Ly = 2, 3, etc. On a more fundamental
level, it is very interesting to see how the special properties of 1D systems evolve with
increasing Ly and approach the 2D limit [216]. Here we will investigate some of the
most essential properties of Heisenberg ladders with Ly up to 6.
The most important aspect of the physics of S = 1/2 Heisenberg ladders is that they
have completely different low-energy properties for even and odd Ly (the number of sites
on each rung of the ladder). For even Ly there is always a spin gap (which vanishes as
Ly → ∞), while odd-Ly systems are gapless and have properties similar to a single 1D
chain (below an energy or temperature scale which vanishes as Ly → ∞). This difference
can be understood roughly based on a simple picture of valence bonds [217], illustrated
in Fig. 68. If there is no coupling between the rungs (Jx = 0, Jy > 0), the ground state of
the 2-leg ladder is just the product of rung singlets—a unique (non-degenerate) state with
a gap Jy to the first excited state (in which one of the singlets is promoted to a triplet).
For the 3-leg ladder, on the other hand, the ground state of an individual rung is two-fold
degenerate, with Sz = ±1/2. When coupling the rungs (Jx > 0), this degeneracy is lifted,
but, for any ratio Jx /Jy , there is a remnant of the degeneracy, which can be understood
as arising from Lx interacting S = 1/2 degrees of freedom. This leads to low-energy
properties similar to those of a single S = 1/2 chain. In the 2-leg ladder, the singlet-
triplet excitation gap remains non-zero for any Jx /Jy , and the low-energy excitations
297
Jx = 0
FIGURE 68. Illustration of the ground states of the 2-leg and 3-leg Heisenberg ladders at inter-rung
coupling Jx = 0. The vertical bars indicate valence bonds (singlets), and the circles show unpaired ↑ and
↓ spins. The ground state of the 2-leg ladder is a unique singlet-product, whereas in the 3-leg case each
z
rung-state is two-fold degenerate, with Srung = ±1/2. Each of these states is a symmetric combination of
the two states with the unpaired spin on the upper and lower chain. Here a random configuration of the
location and Sz of the unpaired spin is shown for each rung.
above this gap form a band of propagating rung triplets [218]. The spin correlations are
exponentially decaying with distance, with the correlation length diverging as Ly → ∞.
In the valence bond basis (discussed in Sec. 2.2), the 2-leg ladder with Jx > 0 is still
dominated by short bonds (one can say that it is an RVB spin liquid, but, due to the
constraints of the ladder geometry, it is also appropriate to call it a VBS), whereas the
critical state of the 3-leg ladder requires bond probabilities that decay with the bond
length as a power-law [217] (and such a state may be called a critical RVB state). The
same pictures remain valid also for the ground states of ladders of larger width, but the
energy scale associated with ladder behavior vanishes as Ly → ∞, with cross-overs into
2D behavior at higher energies. The relevant energy scale is the spin gap for even Ly and
the spin stiffness for odd Lx —as we discussed in connection with the KT transition in
Sec. 3.5, power-law correlations can be sufficient to sustain an non-zero spin stiffness,
and this is the case with the ∼ 1/r correlations in the single chain and odd-leg ladders.
Here we will discuss results only for spatially isotropic couplings; Jx = Jy . For even
Ly , the behavior is similar for periodic and open boundary conditions in the y direction
(and for Ly = 2 periodic boundaries only corresponds to doubling Jy ), while for odd Ly
the gapless nature of the system for all Jx /Jy applies only for open y boundaries. In the
“tube” geometry (periodic y boundaries), odd Ly leads to frustration, and the behavior
is then much more complex, with several possible ground state phases [219]. In the
SSE calculations discussed below, open y boundaries were used in all cases, while the x
boundaries were periodic.
Spin correlations. In the preceding section we studied the spin correlations of the
single chain and confirmed the presence of a logarithmic correction to the ∝ 1/r critical
behavior. We now test this behavior for odd Ly > 1, and also look at the exponentially
decaying correlations for even Ly .
With open boundary conditions in the y direction, the correlator hSi · S j i is not a
function just of the separation ri j between the two spins, but depends on the y coordinates
of both spins. Here we use the maximal reflection-symmetric (in the y-direction) distance
between two spins, i.e., for a given spin at (x, y), with x = 1, . . . , Lx and y = 1, . . ., Ly ,
the correlation is computed with the spin at (x + Lx /2, Ly − y + 1), with the periodic
boundary condition taken into account in the x direction. Averages are then taken over
all (x, y). This is just one convenient choice, and it is not important exactly how the
separation in the y direction is treated, because the correlations anyway depend only
weakly on it when the x separation is large.
298
-1 Ly = 6
10
Ly = 5
Ly = 4
Ly = 3
Ly = 2
-2
|C(x)|
10 Ly = 1
-3
10
2 4 6 8 10
log2(x) (Lx=2x)
FIGURE 69. Spin correlation function at the longest x distance (with a certain averaging over y
separations, as discussed in the text) in Lx × Ly Heisenberg ladders with different Ly as a function of
the length Lx . The dashed line has the form C(x) ∝ 1/x.
Fig. 69 shows the long-distance correlation function versus x = Lx /2 for ladders of

width Ly = 1 − 6. The qualitative difference between even and odd Ly is clear, and in the
case of odd Lx > 1 the behavior for large Lx is very similar to the Lx = 1 system, with
deviations from the form C(x) ∝ 1/x that can be explained by a logarithmic correction
(which we will not analyze in more detail here). For even Lx , there is an exponential
decay for large x, but for Ly = 6 a different short-distance behavior can already be seen
emerging. For very large Ly and x up to ∼ Ly , one should expect the correlation function
to be of the 2D form, for both even and odd Ly , i.e., C(x) should approach the value
m2s ≈ 0.095 of the 2D squared staggered magnetization (illustrated with data in Fig. 5).
For x ≫ Ly , there should be a cross-over into either an exponential fall-off (for even Ly )
or the 1/x form with a logarithmic correction (for odd Ly ). In Fig. 69 one can see that
the correlation at x = 8 is almost the same for Ly = 4 − 6, but for larger x there is still
an increase with Ly . To clearly observe 2D behavior followed by cross-overs into either
even- or odd-leg ladder asymptotic behavior, much larger Ly would be required.
Susceptibility. The temperature dependence of the susceptibility provides a conve-

nient way to extract the spin gap ∆ for even-Ly ladders, and it is also interesting to
investigate the applicability of the expected asymptotic form (284) for odd Ly . Fig. 70
shows SSE results for Ly = 1 − 6, with Lx sufficiently large for each temperature to con-
verge accurately to the thermodynamic limit. For the lowest temperatures Lx = 1024 was
used. For fixed T , convergence as a function of Ly can be seen for T /J > 0.2, where the
results for Ly = 4, 5, and 6 are almost the same. The converged curve corresponds to the
2D limit (for which results for large L × L lattices are shown in Fig. 73). The qualitative
difference between ladders of even and odd width is clear at low temperatures, with χ
for even Lx decaying to zero exponentially below an Ly dependent temperature. For odd
Ly , it can be noted that the T → 0 value decreases with Ly . One can argue that the T → 0
299
0.15
0.08
0.10 0.06
Jχ Ly = 1
Ly = 2 0.04
Ly = 3
0.05
Ly = 4
Ly = 5 0.02
Ly = 6
0.00 0.00
0.0 0.5 1.0 1.5 0 0.1 0.2 0.3 0.4
T/J T/J
FIGURE 70. Temperature dependence of the susceptibility of Heisenberg ladders of different width Ly .
The length Lx was sufficiently large for each T to represent accurately the Lx → ∞ limit. The right panel
shows the low-temperature behavior for Ly = 2 − 6 on a more detailed scale. The statistical errors are not
discernible, thanks to the use of the improved susceptibility estimator (283).
susceptibility per rung should be roughly independent of Ly for small Ly [216], and, thus,
the susceptibility per spin should scale approximately as 1/Ly . We will not do any fitting
to the asymptotic form (284) here, because it is valid only for very low temperatures for
Ly ≥ 3 (and a higher-order logarithmic correction may be necessary unless extremely
low temperatures are used [216]). Note that the χ (0) ∼ 1/Ly form is not inconsistent
with the 2D behavior χ2D (0) > 0, because the ladder behavior (for both even and off Ly )
applies only below some T ∗ (Ly ) which goes to zero when Ly → ∞.
Let us now analyze the Ly dependence of the spin gap ∆ for the even-width ladders.
One expects roughly an exponential low-temperature susceptibility, χ ∼ exp(−∆/T ),
due to the fact that the ground state is a singlet (non-magnetic, with hMz2 i = 0) and the
first excited state is a triplet (magnetic, with hMz2 i = 2/3). This behavior is, however,
modified by the fact that there is a whole continuum of magnetic states (for Lx → ∞)
above ∆. A low-temperature form of the uniform susceptibility of the 2-leg ladder,
a
χ (T ) = √ e−∆/T , (285)
T
has been obtained by analyzing the limit of weak inter-rung couplings perturbatively
[218]. This form can be expected to hold asymptotically for T → 0 for any Jy /Jx and
even Ly (for Lx →√∞). The gap can be extracted from numerical χ (T ) data by considering
the logarithm of T χ in (285), giving
√
−T ln( T χ ) = ∆ − T ln(a). (286)
√
Thus, the intercept of a line fitted to −T ln( T χ ) versus T equals the gap ∆. Fig. 71
shows data analyzed in this way. In all cases, a linear behavior obtains at low tempera-
tures, and line fits give the gaps listed in the figure. A field theoretical treatment shows
that the gap for a spin-S ladder system should decrease with Ly as ∆ ∼ exp(−π SLy )
[220], which is not in good agreement with the results obtained here (and less precise
300
0.8
0.6
-Tln(T χ)
1/2
0.4
Ly = 2, ∆ = 0.505
0.2
Ly = 4, ∆ = 0.164
Ly = 6, ∆ = 0.056
0
0 0.05 0.1 0.15 0.2
T/J
FIGURE 71. The susceptibility of even-width ladders analyzed according to Eq. (286). Extrapolations
in the linear low-T regime, shown as lines extending through the points used in line fits, give the spin gaps
indicated in the figure. The statistical precision is roughly the number of digits shown.
earlier calculations [216]). Calculations for larger Ly would have to be carried out (which
can certainly be done) to investigate this issue in more detail.
It would also be interesting to study the odd-leg ladders for larger Ly and lower
temperatures than what has been done until now, to extract the Ly dependence of the
parameters c and T0 in (284). An alternative approach is to map the low-energy properties
of a ladder onto a single chain with longer-range interactions [221], which arise because
the S = 1/2 degrees of freedom of the isolated rungs are not completely localized to
individual rungs once they are coupled. The localization length increases with Ly , and
eventually, for Ly → ∞, an effective model with sufficiently long-ranged interactions to
produce Néel order (as discussed in Sec. 4.3.3) should obtain. The mapping procedure
has not yet been tested for ladders with large Ly , however, and one should therefore also
study the full ladder systems at lower temperatures and for larger Ly .
5.3.3. Long-range order in two dimensions

We already discussed the nature of the antiferromagnetically ordered (Néel) ground
state of the 2D Heisenberg model in Secs. 2.1 and 4.4.2. Extrapolation of the best
currently available QMC results [50] for the squared sublattice magnetization, shown in
Fig. 5 gives ms = 0.30743(1) for the infinite system, where (1) denotes the statistical
error in the preceding digit. This result deviates only by about 1% from the linear
spin wave result ms = 0.3034. Higher-order spin wave calculations [40, 41, 42] give
ms ≈ 0.3070. As discussed in Sec. 2.1, this good agreement with the actual value can be
traced to the fact that the quantum fluctuations are not that strong, reducing the sublattice
magnetization by only about 40% from the classical value. Other important ground state
quantities, such as the spin stiffness [222], the transverse magnetic susceptibility, and
the spin wave velocity, are in similar good agreement with spin wave theory (but we will
301
not make any detailed comparisons of values here).
After discussing some general data fitting issues, we will here compute the energy
and the spin stiffness using finite-size extrapolations of T → 0 converged SSE results
(with proper convergence confirmed using checks such as those shown in Fig. 65 for
all quantities of interest). In addition to the L × L lattices normally used, we will also
consider rectangular Lx × Ly lattices with Lx = 2Ly . This enables a consistency check of
the extrapolations. We also discuss the finite-temperature susceptibility and extrapolate
it to zero temperature, using L × L lattices sufficiently large to completely eliminate
finite-effects for the range of temperatures considered. We finally discuss the divergence
of the correlation length as T → 0, which is another important manifestation of the Néel
ordered ground state [5].
Extrapolations of the ground state parameters of the 2D Heisenberg model have been
presented in a large number of papers previously, e.g., Refs. [223, 224, 205, 225, 226].
For most practical purposes, the precision already achieved is quite sufficient. Since the
model is one of the most important prototypical systems in quantum magnetism, it is,
however, useful to continue to establish more precise benchmark calculations (which
can be useful, e.g., for testing other methods [94]). Some of the results presented below,
in particular the ground state energy, represent the most precise calculations to date.
In addition to the extrapolated N = ∞, T = 0 values of the various physical quantities,
their finite-size and temperature corrections are also of interest, because they have been
predicted in great detail based on field-theoretical methods [227], and numerical test are
useful to establish the range of validity of the low-energy theories. The corrections are
statistically much noisier and require very long simulations to establish precisely. The
discussion here is only intended to give a flavor of what can be done.
Data fitting issues. When extrapolating finite-size result to the thermodynamic limit,
it is useful to know the expected form of the size corrections based on analytical calcu-
lations. Away from a critical point, the size corrections for systems of dimensionality
d > 1 normally take the form of a polynomial in the inverse system length 1/L, but
some times the leading correction is ∝ 1/La with a an integer larger than 1. As dis-
cussed in Sec. 2.4.1, the leading size correction to the sublattice magnetization of the
2D Heisenberg model can be obtained from the spin wave theory of the Néel state and is
∝ 1/L. This result is also a more general consequence of the fact that the order parameter
is a vector, as discussed in Sec. 3.3. The ground state energy E/N per site has a leading
correction ∝ 1/L3 , and thanks to this high power it is relatively easy to obtain the energy
to high precision even based on rather small lattices [205]. The leading correction to the
spin stiffness is ∝ 1/L [228]. Even if the leading power is not known, one can normally
find it empirically based on data, provided that the numerical precision is sufficiently
high. When fitting data to a polynomial one may find that the coefficient of the linear
term (and possibly higher-order terms) is very small, which then makes it likely that it
actually should be exactly zero. It is best to use a fitting program which allows one to
specify exactly what powers in 1/L should be included.
For a given set of data (here a quantity computed for a set of lattice sizes, but we will
use the same technique also for fitting temperature dependent quantities as powers in T )
one should include high enough powers of 1/L for the fit to be statistically sound, i.e.,
χ 2 per degree of freedom (χ 2 /dof) should be close to 1 for a large data set (as discussed
302
0.674 R=1
R=1 0.15 R = 2 (av)
0.673 R=2 R = 2 (x)
R = 2 (y)
-E/(NJ)
0.672 0.14
ρs/J
0.671
0.13
0.670
0.669 0.12
0 0.05 0.1 0 0.05 0.1
-1/2 -1/2
N N
FIGURE 72. Size dependence of the ground state energy (left) and spin stiffness (right) of 2D Heisen-
berg lattices with aspect ratio R = 1 (where N = L2 ) and 2 (where N = 2L2 ). For R = 2, the spin stiffness
on a finite lattice is different in the x and y direction. Both values are shown here along with the average.
The curves are polynomial fits discussed in the text.
in standard texts on data analysis). The statistical errors of the fitting parameters can
be computed based on the error bars of the data. The safest way, which does not rely
on any assumptions of the errors being small (although normally they should be), is to
carry out a large number of fits with Gaussian noise added to the data (with the standard
deviation of the noise equal to the corresponding error bar) and compute the standard
deviation of the distribution of the resulting parameter values. Note that when increasing
the number of fitting parameters (here powers of 1/L), the error bars of the best-fit
parameters normally increase (and χ /dof may also get worse). One should therefore not
include many more powers than needed to bring χ 2 /dof close to 1. To safeguard against
possible remaining effects of higher-order corrections, one may nevertheless want to
include one more parameter than strictly needed for and acceptable χ 2 /dof.
Ground state energy and spin stiffness. The energy per site E/N and the spin
stiffness ρs were evaluated using the estimators (250) and (280), respectively, with the
energy adjusted for the constant 1/4 subtracted from each bond operator in the SSE
simulations. As discussed in Sec. 5.2.4, the value of the spin stiffness extrapolated to the
thermodynamic limit should be multiplied by a factor 3/2, to account for the fact that
the spin-rotational symmetry is not broken in the SSE simulations.
Comparing extrapolations for lattices of different shapes is a good way to check for
detectable consequences of finite-size corrections beyond those included in the data fits
√ N = L × L (aspect ratio R = 1) and N = 2L × L
[82, 94]. Fig. 72 shows results for both
(R = 2) lattices, graphed versus 1/ N (= 1/L for R = 1 and ∝ 1/L for R = 2). These
results were computed using inverse temperatures β as high as 32 × L for the largest
systems (L up to 40 for R = 1 and up to 32 for R = 2). For the largest systems several
hundred CPU hours were used. The error bars are too small to be visible in the figure.
An example of the results produced, for the 32 × 32 system E/N = −0.6695115(8)
and ρs = 0.12606(5), based 5 × 106 Monte Carlo sweeps. The error bars are of similar
magnitude for the other systems as well.
303
For R = 1 using the systems with L ≥ 6 and a 5th-order polynomial (without the
linear and quadratic terms, which are predicted not to be present, as discussed above)
gives E/N = −0.6694421(5), while a fit of the same order to the R = 2 data with
Ly ≥ 6 delivers E/N = −0.6694422(6). These results are in perfect statistical agreement
with each other, and it is then permissible to use their statistically weighted average,
E/N = −0.6694421(4), as a final estimate of the ground state energy. It should be noted
that the numerical values depend slightly on what lattices are included in the fit and the
order of the polynomial used. Once the fit is statistically sound, these fluctuations should
be consistent with the statistical errors.
Turning now to the stiffness, note first that for a finite lattice with aspect ratio R 6= 1,
there are two stiffness constants, for phase twists imposed in the x and y direction. Both
of them are graphed for R = 2 in the right panel of Fig. 72, along with the arithmetic
average of the two and the R = 1 values. It can be seen here that the R = 1 stiffness
and the average of the R = 2 stiffnesses are better behaved for extrapolations than the
individual x and y values for R = 2, and no fits are therefore included for the latter.
Quadratic fits for R = 1 and R = 2 (using L ≥ 6 and Lx ≥ 8 data) gives ρs = 0.12065(4)
and ρs (∞) = 0.12070(6), respectively. Taking the average of these statistically consistent
values, and including the rotational factor 3/2, gives the final estimate ρs = 0.18100(5)
for the infinite system.
Properties at T > 0. Since the Mermin-Wagner theorem rules out magnetic order
in the 2D Heisenberg model for any T > 0, the correlation length must diverge as
T → 0 for the behavior to be consistent with the ordered ground state. The behavior
is similar in the classical 2D Heisenberg model, where the correlation length diverges
exponentially. The long-distance correlations and fluctuations of the quantum system
can in fact formally be mapped onto a classical system with renormalized couplings
[5], and the low-temperature regime in which such a mapping holds is referred to as
the renormalized classical regime. Many predictions in this regime have resulted from
field-theoretical treatments [5, 227, 84]. Remarkably detailed results have been derived
for the temperature dependence of, e.g., the correlation length and the susceptibility. The
predicted forms, which are believed to be asymptotically (T → 0) exact, depend only on
the ground state parameters (e.g., the spin stiffness and the spin wave velocity c).
Let us first investigate the uniform magnetic susceptibility, which should have the
following form for T → 0 (in an infinite system) [227] :
" 2 #
T T
χ (T ) = χ (0) 1 + + . (287)
2πρs 2πρs
The only model dependent parameter here is the ground state spin stiffness, which we
determined above. To compare numerical data with this result, we first have to make
sure that we can achieve the thermodynamic limit. For fixed temperature, we can study
the behavior as a function of the system size, where we here use quadratic lattices with
N = L2 . The left panel of Fig. 73 shows results for several lattice sizes. For any finite
system the susceptibility vanishes as T → 0 due to the singlet ground state and finite-
size gap in the spectrum. As we discussed in Sec. 4.4.2, the lowest excitations are the
quantum rotor states, with gaps scaling as 1/N. The T /J < 1/L susceptibility for a finite
304
0.060
0.08 L = 128
L = 256
0.055
χ 0.06
L=4
0.04 0.050
L=8
L = 16
0.02 L = 256
0.045
0.00
0 0.5 1 1.5 2 0 0.05 0.1 0.15 0.2 0.25 0.3
T/J T/J
FIGURE 73. Uniform susceptibility of the 2D Heisenberg model computed using several L × L lattices.
The right panel shows low-T results for L = 128 and 256 on a more detailed scale (with the two data
sets coinciding within statistical errors, indicating convergence to the thermodynamic limit). The solid
curve in the right panel is a fit to the L = 256 data (using only the T /J ≤ 0.18 data points) of the form
χ = a + bT + cT 2 with the parameters discussed in the text. Error bars are too small to discern but are
typically ≈ 2 × 10−5 for the low-temperature L = 256 results.
system is dominated by these excitations. Once T > 1/L spin waves can also be excited,
and one may therefore suspect that a system size of at least L > 1/T will be required
for convergence to the thermodynamic limit. Here we will not study the convergence in
detail, but just conclude based on the data in Fig. 73 that the L = 256 results are safely
converged down to T /J = 0.03 and can be used to test the form (287).
Tests of analytical predictions can be carried out in different ways. We could here
use the value of ρs extracted from the finite-size extrapolations of the T = 0 data above
and check the agreement between the SSE results for χ (T ) and Eq. (73), with only
χ (0) adjusted as a fitting parameter to obtain the best agreement. Another way would
be to also adjust ρs , the best-fit value of which can be compared with the result of the
ground state extrapolations. Here we will proceed in a different way, which tests both
the T and T 2 corrections in (287). Fitting SSE data to a form χ = a0 + a1 T + a2 T 2 , we
can use the form of the coefficients in (287) to extract corresponding estimates ρs (1)
and ρs (2) for the spin stiffness from the linear and quadratic coefficients a1 and a2 .
These estimates should agree with each other and with the result of the ground state
calculation if the temperatures used in the fit are sufficiently low. At higher temperatures,
corrections including higher powers of T will be important and should lead to a bad fit
to the quadratic form and disagreements between the different stiffness estimates.
The quadratic fit is good and consistent values of the exponents are obtained if only
data for T /J ≤ 0.18 are included. The fit is shown in the right panel of Fig. 73, and the
stiffness constants extracted from the parameters are ρs(1) = 0.180(3) from slope and
ρs(1) = 0.179(6) from the quadratic correction. These values are in good agreement
with (but noisier than) the ground state result ρs = 0.1811(1). Fixing the form (287) and
extracting the best-fit values of χ (0) and ρs gives a less noisy estimate [225], but since
we are still dealing with corrections to a T = 0 quantity it is not easy to achieve the same
precision as we did by extrapolating the finite-size values of ρs at T = 0.
305
L = 256
100
L = 16, 32, 64, 128
L=8
ξ
10
0 1 2 3 4 5 6 7
J/T
FIGURE 74. The temperature dependent correlation length (graped versus the inverse temperature) of
the 2D Heisenberg model p computed for different lattice sizes using the spin structure factor definition
(70), divided by the factor 15/16 coming from Eq. (72). The dashed curve is the expected asymptotic
(L = ∞, T → 0) form (289) with the parameters ρs = 0.1811 and c = 1.660 determined previously.
The T → 0 susceptibility extracted from the above fit is a0 = χ (0) = 0.04378(3). As

we discussed in Sec. 4.4.2, the transverse susceptibility χ⊥ in the symmetry-broken state
at T = 0 is this number multiplied by 3/2, giving χ⊥ = 0.06567(4). In combination with
the spin stiffness, this value gives us access to another ground state parameter; the spin
wave velocity. We discussed the spin stiffness constant as an elastic modulus in Sec. 3.5.
This elastic-medium approach to the long-wavelength properties of quantum magnets is
also referred to as the hydrodynamic description [229]. The transverse susceptibility is
there analogous to a mass density, and one can relate the spin wave velocity to the spin
stiffness and the susceptibility according to
ρs
r
c= . (288)
χ⊥
Using the values extracted above for ρs and χ⊥ gives√the velocity c = 1.660(1). Recall
that linear spin wave theory (Sec. 2.1.1) gives c = 2. The difference between these
two values is captured very well by spin wave theory including 1/S corrections, and
the “renormalization factor” of c can also be computed quite precisely using a series
expansion around the Ising model [230].
The temperature dependence of the correlation length of the infinite system is also
known, including a correction to the leading exponential divergence [5, 227]:
4πρs 2πρs

e c
ξ (T ) = 1− exp . (289)
8 2πρs T T
We have already determined the two parameters involved here, the ground state spin
stiffness and velocity, and we will use their values quoted above to directly test the
306
analytical prediction against SSE results based on the “second moment” structure factor
definition (70) of the correlation length.
Since the correlation length is a rapidly divergent quantity, very large lattices are re-
quired to converge the calculations to the thermodynamic limit at low temperatures. The
lattice length L should be several times larger than ξ to achieve complete convergence.
Fig. 74 shows results for several lattices sizes up to L = 256. While larger systems have
been studied in the past [225, 226], already these results, which are well converged for
T /J ≥ 0.25, show that the form (289) describes the numerical data very well. The small
deviations have been discussed in the literature [225, 226] and can be understood as
arising from higher-order corrections to the form (289) .
5.3.4. Quantum phase transition in a dimerized system

We now study the quantum phase transition in the 2D columnar dimerized Heisenberg
model illustrated in Fig. 4(b). The phase transition of the ground state takes place as a
function of the coupling ratio g = J2 /J1 > 1 and is caused by quantum fluctuations,
which here correspond to an increasing density of singlets on the dimers and at some
point lead to the loss of the Néel order existing when g ≈ 1. We already examined some
SSE results for the sublattice magnetization of this system in Fig. 5, and the behavior
indicated a phase transition between a Néel-ordered and a nonmagnetic ground state
at coupling ratio g ≈ 1.9. In this section we apply the machinery of finite-size scaling
at criticality, which we discussed in the context of classical systems in Sec. 3.3 and
generalized to quantum systems in Sec. 3.6. Here we first analyze the critical behavior
of several quantities calculated at sufficiently low temperatures to access the ground state
critical behavior, and then discuss consequences of the T = 0 quantum critical point at
non-zero temperatures. Before that, let us spend a few words on the implementation of
the SSE algorithm for a model with non-uniform couplings.
SSE method. An appealing feature of the SSE algorithm for the Heisenberg model
is that the coupling strengths only enter in the acceptance probabilities (265) and (266)
in the diagonal updates, where the coupling J is absorbed in the inverse temperature
β = J/T . For non-constant couplings, we only have to define βb = Jb /T for each bond
b. In the implementation in pseudocode {27}, some scheme has to be used to identify
the coupling corresponding to a generated bond b in an insertion attempt. In a system
with a large number of different couplings (which could even be random, different for
every bond) the couplings (in the form of βb ) should be stored as a list. For systems with
just two different couplings, such as the dimerized lattice considered here, it is better
to just order the bonds in such a way that the coupling can be quickly determined, e.g.,
with b ≤ N/2 and b > N/2 corresponding to J2 and J1 , respectively. Note that for an
operator removal, unlike the case of uniform couplings in code {27}, we now also need
to extract the current bond, b = S(p)/2, in order to determine the probability.
Finite-size scaling for T → 0. When analyzing the ground state, we can proceed as
in the case of the standard uniform 2D Heisenberg model in the previous section, simu-
lating systems at sufficiently low temperatures to achieve T = 0 convergence of all the
307
0.75 1.2 L=8
L = 16
0.70 L = 32
1.1
L = 64
0.65
ρsL
L=8
U2
L = 128
L = 16 1.0
0.60
L = 32
0.55 L =64 0.9
L = 128
0.50
0.8
1.8 1.85 1.9 1.95 1.88 1.90 1.92 1.94
g g
FIGURE 75. Binder cumulant (left) and spin stiffness (in the x direction) multiplied by the system
length (right) of the dimerized Heisenberg model. The crossing points of these curves for different L tend
toward the critical value of the coupling ratio g. Error bars are much smaller than the symbols.
quantities of interest. This approach is discussed for various dimerized systems in, e.g.,
Refs. [85, 88] (as well as in many older works). Another approach is to study systems
at inverse temperature β = Lz , where z is the dynamic critical exponent (which we dis-
cussed in Sec. 3.6) [231]. This is motivated in the following way, by a generalization of
the finite-size scaling hypothesis (64): In a quantum system the scaling function f (ξ /L)
should be replaced by a function with two arguments, f (ξ /L, ξτ /Lτ ), where the correla-
tion length in the imaginary time dimension depends on the spatial correlation length ξ
according to ξτ ∼ ξ z (which defines the dynamic exponent) and the length of the system
in the imaginary time direction is Lτ = c/T ∼ β (where c is a velocity). If we choose
β ∝ Lz , then the scaling function can be written as f [ξ /L, (ξ /L)z ], which is a function
of the single argument ξ /L. Thus, the finite-size scaling procedures can be used exactly
as in the classical systems discussed in Sec. (3.3.2). This is the case also if we take the
limit β → ∞ for each L (in practice finite β large enough for convergence to this limit),
because then ξτ /Lτ → 0, and there is again only one argument ξ /L left in the scaling
function.
There is plenty of evidence already that z = 1 in dimerized Heisenberg models, and
we will here use systems with β = L. This allows for studies of larger systems than in
the β → ∞ limit, although it is not a priori clear which approach is in the end better,
since the corrections to the leading finite-size scaling behavior can be different. Here we
use L up to L = 128. We will also test explicitly that systems with β = L exhibit behavior
consistent with z = 1, by studying quantities which depend on z.
We first locate the critical coupling by examining quantities that should be size
independent at gc . Fig. 75 shows the g dependence of both the Binder cumulant and the
spin stiffness, with the latter multiplied by L to compensate for the expected quantum
critical scaling form ρs ∼ 1/L, obtained the classical form (99) with d → d + z = 3.
The Binder cumulant is defined according to (77), with the number of components
n = 3. Note, however, that (77) is defined with the full scalar product m2 = m · m in
(75), whereas with the SSE method we here only compute the z component expectation
values hm2z i and hm4z i (the off-diagonal components being more difficult to evaluate
308
1.92
1.91
1.90
gc
U2
1.89 ρs (x)
ρs (y)
1.88 ξx/L
ξy/L
1.87
0.00 0.02 0.04 0.06 0.08 0.10

1/L
FIGURE 76. Size dependent critical coupling for the dimerized Heisenberg model extracted from
(L, 2L) crossing points of the Binder cumulant, the spin stiffness constants, and the correlation lengths.
The curves show fits to the form gc (L) = gc (∞) + a/Lω .
[232]). One can easily find the geometrical factors relating these by integrating the z
component cos(Θ) of a classical 3D unit vector over the the angles, giving hm2 i = 3hm2z i
and hm4 i = 5hm2z i. For locating the critical point, these factors play no role, and we could
also use the plain Binder ratio defined as R2z = hm4z i/hm2z i2 .
Since the dimerized lattice does not have 90◦ rotational symmetry, the stiffness con-
stants in the x and y directions are different. Although the numerical values are indeed
quite different, their scaling behaviors close to the critical point is very similar, how-
ever [the x stiffness is approximately a factor 2 larger—the dimers are oriented in the x
direction as in Fig. 4(b)]. Only the x stiffness is shown in Fig. 75(b).
Curve crossings are indeed seen in Fig. 75 for both U2 and ρs L, and after some
significant drift of the crossing points (e.g., for systems of size L and 2L) for small
L, they seem to converge to roughly the same value in both cases. Note that the crossing
points for U2 and ρs approach gc from opposite directions, which can be useful for
bracketing the critical value [85, 88]. Crossing points can be located numerically by
fitting a polynomial of suitable order to some of the data points, repeating the procedure
several times with added Gaussian noise to compute error bars. Fig. 76 shows results
of such procedures for the Binder ratio, the x and y stiffness constants, as well as
the correlation lengths [computed using the definition (70)] in both the x and the y
direction. Fits to the data points of the form gc (L) = gc (∞) + a/Lω are also shown.
This form describes well all the data for L ≥ 10 (the sizes shown in the figure). All the
extrapolated values of gc fall within the range [1.9094, 1.9096], and if the L = 10 data are
excluded the range narrows even further. The exponent ω is in the range 2 ∼ 2.5 for all
quantities (being largest for U2 ). Treating all five values obtained in these extrapolations
as independent statistical data gives gc = 1.90948(4) as a final estimate for the critical
point. This is in good agreement with (but with smaller error bar than) a recent estimate
gc = 1.9096(2) obtained using T → 0 data for the same quantities on lattices with L up
to 64. The crossing point shift exponents ω are also in good agreement.
309
The existence of a limiting value of the location gc at which the Lρs curves cross does
not prove that z = 1. One also should check that the values of Lρs are well behaved, i.e.,
that a crossing point in the plane (g, ρsL) really forms. A plot such as Fig. 76 for ρs L at
the crossing points for lattices of size L and 2L confirms that this is the case. Examining
the Binder cumulant in a larger window of couplings, one can not find any indications of
negative values, which would be a sign of phase coexistence at a first-order transition (as
discussed in Sec. 3.4). Thus, the scaling behavior so far supports at continuous quantum
critical point with z = 1.
We could now proceed to perform data collapse fits in order to find the correlation
length exponent ν , as we discussed for classical systems in Sec. 3.3. This has been done
for dimerized Heisenberg model, including also scaling corrections [85, 88], and the
results are in good agreement with the expected 3D classical Heisenberg universality
class. Here we instead just discuss the exponent η appearing in the critical correlation
function (53), where we should again replace d by d + z = 3. The staggered structure
factor S(π , π ) is the spatial integral (sum on the 2D lattice) of the correlation function
(269), while the Kubo integral (271) for the staggered susceptibility χ (π , π ) corresponds
to a 3D space-time integral (a 2D lattice sum and an imaginary time integral). Performing
these integrals with the above critical form of the correlation function, with a cut-off
equal to the system size L, gives S(π , π ) ∼ Ld−z−η and χ (π , π ) ∼ Ld−η . We will test this
behavior of the critical system and extract the exponent η , using the value of gc found
above. Instead of performing new simulations at this point (which is known only to
within a statistical error), one can perform polynomial interpolations within the existing
data sets. One can then also easily check the behavior for points slightly off the best
estimate of gc (plus and minus one error bar), to check the sensitivity of the fitted η to
the location of the critical point. The simplest way to analyze the critical behavior is to
fit a straight line to ln(S) and ln(χ ) versus ln(L) (as was done for S(π , π ) of this model
in [88]). Some corrections to scaling are always expected, and if the data have small
error bars a straight line can only be fitted to large lattices. With the data sets used here,
statistically acceptable linear fits are only obtained when using system sizes L ≥ 48. We
will therefore also include subleading corrections and assume the following forms
S(π , π ) = aL1−η + bLω , χ (π , π ) = aL2−η + bLω , (290)
where one would expect the subleading exponent ω to be much smaller (possibly even
negative) than the leading exponents. One could in principle perform a combined fit with
η fixed to be the same for the two quantities (with ω and the constants different), but
here we will fit the data sets independently.
In order to more clearly see the role of the subleading correction, S(π , π ) and χ (π , π )
are graphed in Fig. 77 with the dominant L and L2 factors divided out. The asymptotic
L → ∞ behavior is then in both cases ∼ L−η , where η is expected to be small. The cur-
rently best estimate for the classical 3D Heisenberg universality class is η = 0.0375(5)
[113]. The fits to (290) give η = 0.029(2) for S(π , π ) and η = 0.020(3) for χ (π , π ).
In principle the forms (290) should of course work only exactly at gc , but in practice,
for a finite range of system sizes, they fit the data well in some window around the
true critical point. As seen in Fig. 77, the interpolated values of the two quantities at
gc plus and minus one error bar deviate visibly from those at the midpoint, but the fits
310
0.200
0.330
g=1.90944
g=1.90948
0.195
0.325 g=1.90952
2
S(π,π)/L
χ(π,π)/L
0.320 0.190
g = 1.90944
0.315 0.185 g = 1.90948
g = 1.90952
0.310
0.180
0 20 40 60 80 100 120 0 20 40 60 80 100 120
L L
FIGURE 77. Finite-size scaling of the staggered structure factor (left) and susceptibility (right) in close
proximity of the estimated critical coupling ratio gc = 1.90948(4). The powers of L corresponding to z = 1
and η = 0 have been divided out. The remaining asymptotic size dependence should then be governed by
the actual value of η . The fits to the forms in (290) give η = 0.029(3) from S(π , π ) and η = 0.020(4)
from χ (π , π ). The fits were based on L ≥ 8 data, but the resulting curves also go through the L = 4 and 6
points. The dashed curves shows the behavior at g = 1.90948 without the scaling corrections (i.e., with η
and a kept at their values obtained in the fit including the corrections).
are statistically acceptable in all three cases. The statistical errors quoted above arise
predominantly from the uncertainty in the critical coupling. The subleading exponents
in (290) are ω = −0.2(2) for S(π , π ) and ω = 0.6(2) for χ (π , π ).
Fig. 77 also shows the fitted functions with the subleading corrections disregarded
(with the other parameters kept at their values obtained in the fit with corrections).
Clearly the corrections are quite significant, being completely responsible for the maxi-
mums in both curves at L ≈ 10. One can of course obtain much better fits to the larger
lattices without subleading corrections. As mentioned above, with the rather small er-
ror bars of the data used here, only systems with L ≥ 48 can be included in such a fit.
Even then, there must be some influence of the neglected corrections. The values of η
do in fact come out somewhat lower if no corrections are included. With the corrections,
all the data (L ≥ 4) can be included, but to be on the safe side only L ≥ 8 data were
included in the fits quoted here and graphed in the figure. These fits still do pass very
closely through the L = 4 and 6 data points, which further reinforces the quality of the
functional form used.
The η values obtained here, and also in Ref. [88], are a bit lower (by a few error
bars) than the best available classical 3D Heisenberg value [113] quoted above. Most
likely, these small discrepancies are due to further scaling corrections, but it would still
be good to push QMC calculations for various dimerized Heisenberg models to even
higher precision (using larger lattices, a denser grid of lattices sizes, and also reducing
the error bars of the computed quantities) in order to establish the agreement with the
classical exponents for sure. This is particularly interesting and important in light of
the fact that the staggered dimer model illustrated in Fig.4(c) (and also some other
dimerization patterns) seems to show small but statistically significant deviations from
the expected 3D Heisenberg exponents [89, 233, 234]. These systems may still not be in
311
T
QC
RC QD
g
FIGURE 78. Cross-overs (indicated by thick curves) in the coupling-temperature plane of a generic
dimerized 2D Heisenberg model. The quantum-critical point (circle) controls the behavior in the T ≥ 0
quantum critical (QC) regime, up to some temperature where lattice effects become important (i.e., when
ξ ∼ 1). For couplings away from the critical point there is a low-temperature cross-over into either
renormalized classical (RC) or quantum disordered behavior. For a 2D system there is Néel order only
exactly at T = 0. The thickness of the curves separating the different regimes reflect the fact that the
cross-overs are not sharp, but can take place over significant temperature windows.
a different universality class, as originally proposed [89], because the deviations could
also originate from anomalously large scaling corrections [233, 234]. Regardless of the
underlying reason for the deviations, their origin remains unclear and should be further
examined (and such work is in progress [234]).
Quantum critical scaling at T > 0. One of the most remarkable and important as-
pects of quantum criticality is that the properties (the universality class) of a T = 0
critical point also can strongly influence a system at T > 0, often up to rather high tem-
peratures [84] and also if the system does not have exactly the couplings corresponding
to a T = 0 critical point (but is near such a point, in a sense which we will make more
precise below). Critical fluctuations and scaling behavior can therefore be much more
generic features of quantum systems than classical systems [142], where critical behav-
ior often can be observed only very close to the critical point.
In a quantum system, critical behavior can be expected when the temperature is the
dominant energy scale. In dimerized 2D Heisenberg models, the energy scales on the
Néel and nonmagnetic side of the transition are, respectively, the spin stiffness ρs and
the singlet-triplet excitation gap ∆. One can therefore expect manifestations of quantum
criticality for T ≫ ρs and T ≫ ∆, for g < gc and g > gc , respectively. In continuum field
theories there is no upper-bound temperature for the quantum critical behavior, while
for a lattice hamiltonian there has to be a cross-over to a different behavior above some
temperature where the spins become essentially independent due to thermal fluctuations
(which cannot be captured by continuum field theories, where independent discrete spins
do not exist). The cross-over to the high-temperature limit could be defined, e.g., on the
basis of the Curie (independent spin) susceptibility χ → 1/(4T ) obtaining for T ≫ J.
As we will see below, the quantum critical susceptibility is completely different.
Putting all this together, for a 2D dimerized Heisenberg model one expects a generic
“cross-over diagram” [5, 84] of the type shown in Fig. 78. For temperatures below the
high-T cross-over, there are three different “regimes”, in which the behavior is controlled
312
be the corresponding largest energy scale in the problem; ρs , ∆, or T . In the renormalized
classical (RC) regime, the spin stiffness dominates, and the correlation length diverges
as T → 0 according to (289), as we saw explicitly for the standard uniform Heisenberg
model in Sec. 5.3.3. In the quantum disordered (QD) regime, the correlation length
saturates at some finite value as T → 0, with ξ → constant roughly below T ≈ ∆. In the
quantum critical (QC) regime, the correlation length should diverge at gc when T → 0
according to
ξ ∼ 1/T z . (291)
This behavior can be understood based on the path integral mapping of the quantum
system in d dimensions onto a classical (d + z)-dimensional system. There the coupling
g of the quantum system corresponds to T of the (d + z)-dimensional system. We we can
then think of temperature scaling in the quantum system as a kind of finite-size scaling
in the imaginary time dimension, in which the thickness of the effective system is finite;
Lτ ∝ 1/T (i.e., we keep T constant and just change the thickness of the effective system).
By definition of the dynamic exponent, at gc we have
1/z
ξ ∼ ξτ . (292)
If the time dimension thickness Lτ = ∞, both the correlation lengths ξ and ξτ diverge at
this point, but with finite Lτ (but infinite spatial size L) we should replace ξτ by Lτ ∝ 1/T
in scaling formulas (exactly as we do in classical finite-size scaling when L < ξ ), whence
(292) gives (291).
Quantum critical scaling forms have been derived for many quantities based on these
ideas worked out quantitatively and in great detail [84] using renormalization group
methods and large-N expansions [with N here being the number of components of
SU(N) symmetric spins, with N = 2 for the physical S = 1/2 spins] within the continuum
field theory [the (2 +1)-dimensional nonlinear sigma-model]. The main point is that one
can observe scaling, with some corrections (the size of which depends on the quantity),
also when g 6= gc , in the QC regime illustrated in Fig. 77. Away from gc , there is a cross-
over into either RC or QD behavior. Here we only look at two important quantities;
in addition to the correlation length we also analyze the uniform susceptibility, which
is perhaps the quantity for which QC behavior away from gc is the most robust. The
predicted form is linear in T exactly at gc , with a constant shift away from gc ;
a
χ= T + b, (b = 0 at g = gc ), (293)
c2
where c is the spin wave velocity. The constant a is not known exactly, but its value
computed based on the leading terms in an 1/N expansion is believed to be close to the
actual value [84]. The linear behavior should apply strictly only when T → 0, since there
are higher powers of T , with unknown prefactors, which are not included in (293).
We begin by investigating the finite-size convergence of the susceptibility at gc (using
gc = 1.9095, within the error bars of the critical point determined above), in order to
make sure that we can reliably obtain results reflecting the thermodynamic limit at low
temperatures. The left panel in Fig. 79 shows the temperature dependence of χ /T (which
is easier to look at in graphs than χ itself) for several power-of-two system sizes up to
313
0.12 0.092
L=8 L = 32
L = 16 L = 64
0.11 L = 32 0.090 L = 128
L = 64 L = 256
L = 128 L = 512
χ/T
0.10 L = 256 0.088
0.09 0.086
0.08 0.084
0 0.2 0.4 0.6 0 0.05 0.1 0.15 0.2 0.25 0.3
T/J1 T/J1
FIGURE 79. Susceptibility divided by temperature for the dimerized system at the critical point (g =
1.9095) computed using different lattice sizes. The peaks seen in the left panel for small systems is a finite-
size feature which moves toward T = 0 as 1/L. The solid curve in the right panel is a cubic polynomial
fit to the L = 512 data (T /J1 ≤ 0.20), which gives the prefactor a/c2 = 0.0916(1) in Eq. (293). For
comparison, in the right panel, size-converged results for for g = 1.9090 are also shown (the lowest curve,
marked with +). In this temperature range, quantum critical behavior is seen also at this coupling, which
is slightly on the nonmagnetic side of the transition (and hence χ /T → 0 eventually as T → 0)
L = 256. As expected, the results converge quickly at high temperatures, with finite-size
effects entering at approximately T ∼ 1/L (which can be expected on account of the
dynamic exponent z = 1). The right panel of Fig. 79 shows the low-T results on a more
detailed scale, including also results for L = 512. Based on this comparison of results
for different sizes, one can conclude that the thermodynamic limit can be studied with
L ≤ 512 lattices down to T /J1 = 0.03 (and probably even a bit below that).
The right panel of Fig. 79 also shows a fit of χ /T to a cubic polynomial at low
temperatures. The corrections to the asymptotic linear form (293), is quite prominent,
leading to an ≈ 8% increase in χ /T [which can be considered as effective temperature
dependent prefactor a/c2 in (293)] from the minimum around T /J1 ≈ 0.3 to the eventual
T → 0 value. In principle the extrapolated T = 0 value a/c2 = 0.0916 can be used to
extract the spin wave velocity, but since the prefactor a is not known exactly [84], this
estimate would likely not be very precise. Another use of the result obtained here would
be to use it in combination with an accurate estimate of c obtained in some other way,
which would allow a test of the approximate calculation [84] of the prefactor. This is
beyond the scope of the discussion here, however.
Fig. 80(a) shows the temperature dependence of χ at g = gc and at two values some
distance away on either side of the critical point. There is a broad maximum at T ≈ J1 ,
which corresponds to the cross-over into the eventual high-temperature Curie behavior.
Below the maximum, these near-critical systems all exhibit an approximately T -linear
susceptibility, in accord with the form (293) with b < 0 and b > 0 for g > gc and
g < gc , respectively. At still lower temperatures, there will be cross-overs into RC or
QD behavior, which are not seen clearly here because the temperatures are still too high.
In the QD regime χ → 0 exponentially for T below the gap ∆. The RC form is also linear,
like the QC form, but the slope changes in the cross-over region. Such a cross-over can
314
0.08
(a) 25 (b)
0.06 20
15
χJ1
ξx
0.04
g = 1.8500 10
0.02 g = 1.9095
g = 2.0000 5
0.00 0
0 0.5 1 1.5 0 2 4 6 8 10 12
T/J1 J1/T
FIGURE 80. Temperature dependence of the susceptibility (left) and the correlation length (right) in
critical (g = 1.9095) and near-critical systems. The solid lines show the asymptotic quantum-critical T -
linear behavior of χ (from the T = 0 value of the fit in Fig. 79) and the 1/T -linear behavior of χx [following
the form (292) with a small negative constant correction].
actually be seen in the standard 2D Heisenberg model, corresponding to g = 1, which

may appear to be too far away from the critical point. In the left panel of Fig. 73 one can
nevertheless see an approximately linear behavior of χ in the range T /J ∈ [0.3 ∼ 0.5],
before the RC behavior sets in at lower temperatures. The slope of χ in the narrow
window is in very good agreement with that obtained with the known spin wave velocity
and the approximately calculated prefactor a in (79), which indicates that this behavior
really is a manifestation of quantum critical behavior far from a quantum critical point
[84, 225] (although the very good agreement may be to some extent fortuitous, since, as
we concluded above, there are significant corrections to the purely linear form exactly
at gc , at much lower temperatures than the QC window at g = 1).
Fig. 80(b) shows the correlation length in the x direction (which is about 30% larger
than the y correlation length at gc ) at the same near-critical couplings as in Fig. 80(a).
While the behavior is very linear, with a small constant correction to the asymptotic
form (292) with z = 1, the results for the systems slightly off the critical point deviate
significantly from linearity below T /J1 ≈ 0.3. Being a divergent quantity for g ≤ gc , the
correlation length has much larger corrections to the critical form than the non-divergent
uniform susceptibility.
5.3.5. Néel–VBS transitions in J-Q models

With the J-Q models introduced in Sec. 2.4.3, one can study quantum phase transi-
tions at which not only the antiferromagnetic long-range order vanishes, but a different
symmetry is broken in the nonmagnetic state as a VBS forms. Superficially, a VBS may
seem rather similar to the nonmagnetic state of a dimerized Heisenberg model, because
in both cases the system exhibits a pattern of strongly and weakly correlated nearest-
neighbor pairs. These states are fundamentally very different, however, because in a
“manually” dimerized Heisenberg model the hamiltonian itself breaks the translational
315
symmetry, and the strongly correlated bonds also correspond to the strongly coupled
spin pairs. In contrast, in J-Q models, the hamiltonian obeys all the symmetries of the
lattice, and the translational symmetry is spontaneously broken in the VBS state. The
quantum fluctuations of this VBS, especially close to the phase transition, are therefore
very different (and much more interesting).
On the square lattice, which we will consider here, and with the types of interactions
we choose, the VBS can form in four equivalent patterns; hence the broken symmetry is
Z4 . We will first discuss the standard J-Q model with four-spin couplings [17] (and the
term J-Q model will then refer just to that particular case). In this case the Néel–VBS
transition appears to be continuous and shows many of the hallmarks of the proposed
[16, 235] (and still controversial [236]) deconfined quantum critical point. The VBS
state in this case is most likely columnar, but the exact nature of the state is masked
by an emergent U(1) symmetry in the VBS state (which in this model is always near-
critical, with large fluctuations of the VBS order parameter). This is one of the salient
features of the putative deconfined quantum critical point. Despite the many ways in
which the behavior of the J-Q model agrees with the theory, there are also aspects of
the Néel–VBS transition that remain unexaplained—although not necessarily in conflict
with the theory, because detailed analytical calculations are very difficult [235, 237]
and many properties of the deconfined quantum critical point in SU(2) spin systems
are simply not known quantitatively based on the field theory proposed to describe the
transition. We will here look at one example of strong corrections (possibly logarithmic)
to the quantum critical scaling behavior, which had not been predicted by the theory of
deconfined quantum critical points.
We will also discuss some results for a different kind of multi-spin coupling, which
leads to a staggered VBS pattern [like the dimer pattern in Fig. 4(c)]. The phase transi-
tion in this case is clearly first-order [109]. We will discuss the reasons for the qualita-
tively different nature of the transitions in the two models.
SSE implementation for the J-Q model. We first discuss some implementation issues
for studies of the J-Q model with the SSE method. The model was first studied with a
projector QMC technique formulated in the valence bond basis (which we will briefly
discuss in Sec. 6). It is, however, also easy to implement the SSE method for it [238],
which was initially done using “directed loop” updates (a generalized loop algorithm,
which we will also summarize Sec. 6). An SSE algorithm for the J-Q model can also be
devised which is almost identical to the standard algorithm with operator-loops for the
isotropic Heisenberg model. To see this, we can write the J-Q hamiltonian with four-spin
interactions, Eq. (23), in terms of the diagonal and off-diagonal bond operators (255) and
(256) used in the SSE algorithm for the Heisenberg model;
H = −J ∑(H1b − H2b ) − Q ∑ (H1b − H2b )(H1c − H2c ). (294)

[b] [bc]
As in the Heisenberg model [b] is a bond with two interacting spins [i(b), j(b)], while [bc]
denotes two parallel bonds [i(b), j(b); k(c), l(c)], corresponding to the singlet projector
product Si j Skl in (23). We will consider these bonds arranged as in Fig. 8, with the
summation in (294) taken over all translations on the lattice and including horizontal
316
J J Q Q Q Q
FIGURE 81. Examples of vertices in the J-Q model. Only some of the allowed spin states of the
vertex legs (open and solid circles) corresponding to the operators are shown. Open and solid bars
indicate diagonal and off-diagonal bond operators, respectively. The J-vertices are identical to those in the
Heisenberg model. The Q-terms in Eq. (294) are products of two bond operators, which when expanded
out include all combinations of diagonal and off diagonal factors. Allowed loops pass only through one of
the operator factors in the case of Q-vertices, as illustrated here with loop segments at all the permissible
leg pairs. Flipping loops can lead to any combination of allowed operators and spin states.
as well as vertical bond orientations. The scheme is, however, independent on how the
singlet projectors are arranged, and also the generalization to an arbitrary number of
bonds in the Q term is trivial. The only constraint is that we have to avoid sign problems,
which we do if both J > 0 and Q > 0 [with the minus signs in (294), which corresponds
to energetically favoring singlets on the bonds included in both the J and Q terms]. The
absence of sign problems was discussed based on a sublattice rotation in [238], and it
can also be demonstrated using the simple operator counting arguments used for bipartite
Heisenberg models in Sec. 5.2.
We now have J-vertices with four legs as well as Q-vertices with eight legs, as
illustrated in Fig. 81. The Q-vertices can be considered as two J-vertices joined together,
in all possible combinations of diagonal and off-diagonal parts arising from the four
operators in the Q term of (294). It is then clear that we can proceed in the same way as
we did for the Heisenberg model, updating the operator string and a stored state using a
combination of diagonal and loop updates. The key is here again that the matrix elements
are the same for all J- and Q-vertices (the values being J/2 and Q/4), which means that a
loop update in which the type of vertex (J or Q) is not changed can always be accepted.
In the case of the Q-vertices, the loops satisfying this constraint enter and exit at the
same operator factor, as illustrated with loop segments in Fig. 81.
Both J and Q diagonal operators are inserted and removed in the diagonal update. The
simplest way to insert diagonal operators is to choose completely randomly among all
the possible single-bond [b] and double-bond [bc] instances in (294). There are N each
of horizontal and vertical bonds [b] and also N each of horizontal and vertical bond pairs
[bc], for a total of 4N cases to choose from. The spins in the current state have to be
antiparallel on the bond or bonds acted on by the chosen operator, and if that is the case
the acceptance probability is a simple modification of Eq. (265), with either βJ = J/(2T )
or βQ = Q/(4T ) replacing β , depending on the type of operator inserted. The number
of bonds Nb is replaced by the total number of bonds and bond pairs, i.e., 4N. The same
modifications apply to the removal probability (265) as well. If the ratio Q/J is much
different from 1, which is the case in the VBS state and at the phase transition, it is better
to take this ratio into account already when generating the bonds or bond pairs (but even
the trivial random operator generation actually works very well). Note that Q is much
larger than J in the parameter regime we are interested in, and it is therefore best to
define the temperature in units of Q, i.e., setting Q = 1 in the program.
317
1.0
0.6
0.8
0.6 L = 256 0.4

U2
L =128
0.4 L = 64
L = 32 0.2
0.2 L = 16
0.0 0.0
0 0.2 0.4 0.6 0.8 1 0 0.02 0.04 0.06
J/Q J/Q
FIGURE 82. Binder cumulant of the staggered magnetization of the J-Q model. The left panel shows
results up to J/Q = 1, where U2 approaches 1, indicating a Néel state. For smaller J/Q the cumulant
vanishes with increasing size, as shown in greater detail in the right panel, demonstrating a nonmagnetic
state for J/Q < 0.045 (below the crossing points, which accumulate at the critical point). The cumulant
remains positive for all J/Q, showing that there is no phase coexistence, i.e., based on these results the
transition is continuous.
The loop update is identical to the one we developed for the Heisenberg model, with
the simple extension that the linked vertex list (illustrated in Fig. 57 for the Heisenberg
model) now has eight elements for each vertex, instead of only four in the case of the
Heisenberg model. Although the J-vertices have only four legs, the most practical way
is to allocate eight storage elements for all vertices, using only the first four of them for
the J-vertices (and filling the unused ones with a negative number, so that they are never
visited in the loop update).
Continuous quantum phase transition in the J-Q model. The initial study of the J-Q
model [17] established the existence of a VBS ground state for small J/Q and a Néel–
VBS transition compatible with a z = 1 quantum critical point. These results were soon
thereafter confirmed in SSE studies of the T > 0 quantum critical regime [238, 239]. A
later world line QMC study claimed evidence for a first-order transition [240], and this
scenario was argued for also based on studies of other systems [236]. Recent calculations
on very large lattices (up to L = 256) do not, however, find any indications of first-
order behavior [110]. This is in violation of the “Landau” rule, according to which a
direct transition between states breaking unrelated symmetries should be generically
discontinuous, but it agrees with the theory of deconfined quantum critical points [16].
There are, however, significant unexpected scaling corrections in some quantities (which
may have been mistaken for signs of a first-order transition in [240]). Here we will first
discuss some of the evidence for a single continuous Néel–VBS transition.
The Binder cumulant is a useful quantity for analyzing both continuous and first-
order phase transitions. As we discussed in Sec. 3.4, with an example in Fig. 20, phase
coexistence at a first-order transition leads to a negative divergent peak in the cumulant
at the transition point. Fig. 82 shows results for the antiferromagnetic Binder cumulant
U2 of the J-Q model, calculated with the SSE method using β = Q/T = L (motivated
by the previous work showing that z = 1). For J/Q > 0.05, U2 increases with L and
318
0.30 L = 16 0.30
L = 32
L = 64
0.25 L =128 0.25
L = 256
-1
-1
ξdL
ξsL
0.20 0.20
0.15 0.15
0.10 0.10
0.02 0.03 0.04 0.05 0.06 0.02 0.03 0.04 0.05 0.06
J/Q J/Q
FIGURE 83. Finite-size scaling of the spin (left panel) and dimer (right panel) correlation lengths of
the J-Q model. Note that the crossing points for ξs and ξd move with increasing lattice size in different
directions, toward each other.
tends toward 1, showing that the system has long-range Néel order. For smaller coupling
ratios the cumulant decreases to 0, with no indications of a negative peak developing.
The crossing points are well behaved and indicate a critical point at (J/Q)c ≈ 0.045.
Next, we examine the correlation length, computed using the structure factor defi-
nition (70) for both the spin-spin and dimer-dimer correlations. While the wave-vector
for the spin (Néel) order parameter is (π , π ), a columnar VBS corresponds to (π , 0) or
(0, π ), which then is used as the reference point q = 0 in (70). To compute the dimer
correlations, we use the definition (25) but only include the diagonal correlations in the
bond operator (26) i.e.,
Bxx (r) = Sz (r)Sz(r + x̂). (295)
Although the finite-size scaling properties of the correlation functions based on the
rotational-symmetric and the z component definitions may be different [238], the corre-
lation lengths extracted from these functions should still both exhibit finite-size scaling
∼ L (up to scaling corrections) at the critical point.
We can use the scaled correlation lengths ξs /L (spin) and ξd /L (dimer) to test whether
the two order parameters vanish at the same point, or whether there could possibly be two
different transitions. Fig. 83 shows the raw SSE data, which again was generated using
β = L. The curves for different system sizes cross each other, although there is some
drift, both horizontally and vertically, even for the largest lattices. While the crossing
J/Q values for ξs /L and ξd /L are quite far apart for small lattices, they move toward
each other with increasing system size, in a way consistent with a single critical point.
Fig. 84 shows crossing points extracted from L and 2L data for several different
quantities (those discussed above as well as the spin stiffness). It is interesting to
compare this graph with the corresponding one for the dimerized Heisenberg model,
Fig. 76. The convergence of the critical point is clearly much slower for the J-Q model.
While the behavior for all quantities is almost linear in 1/L over some range of system
sizes, for the largest systems the behavior starts to change, in a way which suggests
an eventual convergence with some higher power of L to a common critical point
(J/Q)c ≈ 0.045. Recall that for the dimerized Heisenberg model the behavior is ∼ 1/Lω
319
0.055 U2
ρs
ξs
0.050 ξd
(J/Q)c
0.045
0.040
0.035
0 0.01 0.02 0.03 0.04 0.05 0.06
1/L
FIGURE 84. Size dependent critical coupling estimates based on curve crossings for systems of size L
and 2L, using several different quantities—the Binder cumulant U2 , the size-scaled spin stiffness ρs L, as
well as the spin and dimer correlation lengths ξs /L and ξd /L. The average of the ξs /L and ξd /L estimates,
shown with solid squares, exhibits almost no size dependence. The average of these points for the four
largest L give the critical-point estimate (J/Q)c = 0.04498(3), which is shown with the horizontal line.
with ω in the range 2 ∼ 2.5, and this behavior applies already for rather small lattices
(L ≈ 8). Because of the slow convergence, it is difficult to estimate the critical point
of the J-Q model precisely based on these individual crossing point estimates. There
is, however, a remarkable feature seen in Fig. 84: The ξs /L and ξd /L crossing points
form curves that look very symmetric, approaching the apparent asymptotic value at the
same rate but from different sides. Therefore, the average of these two estimates exhibits
almost no size dependence at all (as also shown in the figure). Based on these average
values one can therefore obtain a much better estimate of the critical point than what
might initially have been expected. The result based on the four largest-L points (which
agree completely withn statistical errors) is (Q/J)c = 0.04498(3).
The slow convergence of the crossing points indicates large scaling corrections in
the underlying physical quantities. Note that the Binder cumulant crossings have the
weakest size dependence of the quantities analyzed in Fig. 84, which is also clear
when comparing the raw data in Figs. 82 (cumulant) and 83 (correlation lengths).
Since the Binder cumulant involves a ratio of two similar quantities, anomalous scaling
corrections in these may partially cancel each other. It is not at all clear why the finite-
size corrections in the crossing points for the two correlation lengths cancel each other
almost completely. It may indicate some hidden symmetry between the near-critical
VBS and Néel states—perhaps some kind of duality inherent to deconfined quantum
criticality [which is the case for the U(1) version of the theory [16], but not explicitly in
the SU(2) variant, which is the one of relevance for the J-Q model]. Note that it is not
just the locations of the crossing points that are symmetric, but in fact the correlation
length curves themselves look rather symmetric in Fig. 83. By reflecting one of the
graphs about the critical point, J/Q − (J/Q)c → −[J/Q − (J/Q)c], the spin and dimer
correlation lengths for the largest system fall almost on top of each other, i.e., even the
scaling functions for these correlation lengths appear to be the same.
320
0.050 J/Q
J/Q =
= 0.050
0.050 0.050
J/Q
J/Q =
= 0.045
0.045
J/Q
J/Q =
= 0.040
0.040
χ/T
0.045 0.045
0.040 0.040
0 0.1 0.2 0.3 0.03 0.05 0.1 0.2 0.3
T/Q T/Q
FIGURE 85. Uniform susceptibility χ /T of the J-Q model graphed on linear (left) and logarithmic
(right) temperature scales, for three coupling ratios in the neighborhood of the critical value, (J/Q)c =
0.04498(3). The solid curve is a fit to the form χ /T = a + b ln(Q/T ) + cT 2 at J/Q = 0.045 (which is
within the error bar of the critical value). The SSE calculations were carried out using sufficiently large
lattices (up to L = 512 at the lowest temperatures) to obtain results representing the thermodynamic limit.
Quantum critical susceptibility scaling at T > 0. At a conventional quantum critical

point, one would expect an asymptotic linear temperature dependence of the uniform
magnetic susceptibility, as we saw in the dimerized Heisenberg model in Fig. 79, where
corrections in the form of higher powers of T are also visible in the ratio χ /T . In the
J-Q model there are much stronger corrections to the expected quantum critical form.
As shown in Fig. 85, there appears to be a multiplicative logarithmic correction, i.e., the
asymptotic T → 0 critical behavior may be of the form χ ∼ T ln(1/T ). The results are
graphed both on linear and logarithmic temperature scales, in order to clearly convey
how the behavior differs from that of the dimerized model in Fig. 79. The behavior in
the J-Q model does not appear to be compatible with a conventional higher power-law
correction, while a logarithm describes the behavior very nicely.
Logarithmic corrections have not been suggested based on the present theory of
deconfined quantum criticality. All analytical calculations within the non-compact CP1
field theory proposed to describe the transition have been performed using large-N
expansions of the SU(N) generalization of the system (the field theory generalized to
CPN−1 , where N = 2 is the physical number of spin components) [16, 16, 237]. It
is possible that different features appear for N = 2. On the other hand, Monte Carlo
calculations of lattice versions of the field theory [241, 242, 236] have not pointed
to any anomalous scaling corrections (although one only found first-order transitions
[236]). Since these calculations are in disagreement with each other, the situation for the
N = 2 theory remains unsettled. Logarithmic corrections in the susceptibility and other
quantities do appear in related fermionic gauge field theories [243]. As we discussed
in Sec. 4.3 and 5.3.1, in field theory language logarithmic corrections can appear as a
consequence of marginal operators, but other explanations are also possible. In principle
the corrections may also not really be logarithmic—they could also be due to almost
marginal (but in the end irrelevant) operators. Corresponding power-law corrections
with a very small exponents may not be distinguishable from logs for the system
321
(a) (b) (c) (d)
FIGURE 86. Illustration of how local fluctuations of valence bond pairs can lead to a gradual change
of a columnar state (a) into a plaquette state (d). In (b) and (c) a small number of bond pairs have been
flipped. Local horizontal ↔ vertical fluctuations of the maximum number of all bond pairs can take place
with the plaquettes (bond superpositions) arranged as in (d) [one of four equivalent translated patterns]. A
fraction p ∈ [0, 1] of vertical bond pairs on these plaquettes corresponds to a VBS angle φ ∈ [0, π ], defined
using the x and y dimer order parameters (296). In these figures, angles φ ∈ [π , 2π ] correspond to shifting
the vertical bonds up by one step up.
sizes accessible. While this scenario is possible for some quantities, in the case of the
uniform susceptibility that would also be highly unusual, since normally all corrections
to quantum critical behavior in 2D antiferromagnets are integer powers of T [84].
Emergent U(1) symmetry in the VBS. A prominent feature of the theory of decon-
fined quantum criticality is that, in addition to the normal correlation length ξ (which
can be taken as the dimer correlation length ξd in the Néel state and the spin correlation
length ξs in the VBS state), there is another length scale Λ in the VBS phase. This is the
distance scale on which excited S = 1/2 spinons (which are defects, a kind of vortices,
in the VBS [245]) are confined into S = 1 “triplons” (i.e., Λ is the size of the spinon
bound state). The same length scale should also be associated with an emergent U(1)
symmetry, related to a particular kind of quantum fluctuations connecting two different
kinds of VBS ordering patterns—the columnar state and the plaquette state. These “an-
gular” VBS fluctuations and the two kinds of VBS states are illustrated in Fig.86. The
angular fluctuations occur on length scales less than Λ (i.e., on length scales less than
Λ, the system is in a superposition of the two different VBS states). The excited spinons
can move freely away from each other up to distance Λ, but at larger distances they
become confined, due to an effective potential mediated by the VBS long-range order
(and in this regard one can note interesting connections to confinement/deconfinement
in certain gauge theoreis in particle physics [18]). As the critical point is approached, Λ
should diverge as a power of the correlation length, Λ ∼ ξ a , with the exponent a > 1.
The existence of an approximate U(1) symmetry in the VBS order parameter can
be tested by examining the distribution P(Dx , Dy) of order parameters Dx and Dy ,
corresponding to columnar order with x- and y-oriented bonds and defined as
N N
Dx = ∑ S(ri ) · S(ri + x̂)(−1)xi , Dy = ∑ S(ri ) · S(ri + ŷ)(−1)yi . (296)
i=1 i=1
For a given Monte Carlo configuration, Dx and Dy can be evaluated at fixed imaginary
time (propagated state in the SSE), using the improved estimator for the correlation func-
tion discussed in Sec. 5.2.5. The improved estimator is spin-rotational invariant, which
can best be understood by considering the equivalence between the loop formulation
322
J/Q = 0 J/Q = 0.045
FIGURE 87. VBS order parameter distributions P(Dx , Dy ) calculated for L = 128 systems at J/Q = 0
(left) and 0.045 (right). The histograms are shown in regions |Dx |, |Dx | ≤ Dcut , with Dcut chosen for the two
coupling ratios to cover the region where the weight is significant. Brighter features correspond to higher
probability density. The absolute scale is not important here, only the circular shapes of the distributions.
The weigh is not distributed completely uniformly in these histograms [obeying neither perfect Z4 nor
U(1) symmetry], due to the inefficiency of the simulations to fluctuate the VBS angle for large systems.
and the valence bond basis [244, 31, 50] (which we briefly discuss in Sec. 6). In fact, the
improved loop estimator for an equal time correlation function is exactly equivalent to
an expectation value of the operator in a particular valence bond state, provided that the
temperature is sufficiently low for the simulation only to sample the singlet sector.
While the calculation of P(Dx , Dy ) could be done with the SSE method, the results to
be discussed below were instead generated with a ground state method in the valence
bond basis (which in its most recent formulation [50] actually is very similar to the
SSE method, as discussed in Sec. 6). The rotationally invariant estimator is not strictly
needed here [109], and one can also just use the z spin components in (296), as we did
in the correlation function Eq. (295). In the valence bond basis it is, however, natural to
compute rotation-invariant quantities. Note that P(Dx , Dy ) is not an operator expectation
value (i.e., it is not a normal physical observable), just a distribution of individual
measurements in a simulation carried out in a particular basis. Its symmetry properties
nevertheless reflect the actual symmetries of the VBS order parameter of the system.
We have already looked at a classical analogy of P(Dx , Dy ); the distribution P(Mx, My )
of stripe order parameters in the frustrated Ising model, with results shown in Fig. 19.
In that case the four-fold symmetry of the order parameter at and below the transition
temperature is obvious, while going to very high temperatures (not shown in the fig-
ure) the distribution turns into a central peak with full rotational symmetry [i.e., U(1)
symmetry] because of the independent Gaussian fluctuations of the short-range order in
different parts of the system. In a similar way, in a VBS one should expect P(Dx , Dy )
to be U(1) symmetric in the Néel state (where there is only short-range VBS correla-
tions), and become four-fold (Z4 ) symmetric inside the VBS state. As seen in Fig. 87
(which shows recent results for larger systems than in [17, 240, 108]), this is not quite
the case, however. Inside the VBS phase, at J/Q = 0, the distribution is ring shaped,
with no weight at the center and no signs of a four-fold symmetry. This shows that the
323
magnitude D, with D2 = D2x + D2y , of the VBS order parameter has formed, hD2 i > 0,
but its angular symmetry has not yet been established, i.e., while the magnitude of the
VBS order parameter has relatively small fluctuations, the fluctuations in the individual
components Dx and Dy [which define an angle φ of the VBS in the (Dx , Dy ) plane] are
very large. Close to the critical point, at J/Q = 0.045 in the figure, the distribution is
peaked at the center, as expected, but again there is no Z4 symmetry.
In the continuum field theory, the operator responsible for the VBS formation is
dangerously irrelevant [16, 245, 246], which means that although it is a perturbation
with Z4 symmetry, which leads to a Z4 symmetry-broken VBS state, this perturbation
does not affect the critical point. The theory of deconfined quantum critical point has
built-in U(1) symmetry, and this is unchanged in the presence of the VBS operator.
The J-Q model has no explicit U(1) symmetry—what is built in here is instead the Z4
discrete rotational symmetry of the square lattice, and naively one would expect the
VBS forming on this lattice to exhibit Z4 symmetry, exactly as the stripe order of the
Ising model discussed in Sec. 3.4. Instead, we see a U(1) symmetric order parameter
distribution P(Dx , Dy ), which in light of the theory of deconfined quantum critical points
should be expected if the system is smaller than the deconfinement length scale Λ (which
therefore has to be large, Λ > 128 even at J/Q = 0 according to the results in Fig. 87).
If we increase the system size, we should eventually, for L > Λ, obeserve a four-fold
symmetric distribution (which has been seen in some related systems [108, 247, 248]).
A columnar state corresponds to peaks at (0, ±D) and (±D, 0).
As discussed above and illustrated in Fig. 86, the emergent U(1) symmetry is a
consequence of local fluctuations between columnar and plaquette VBS patterns [245],
in such a way that all possible values of Dx and Dy compatible with a magnitude D
(weakly fluctuating) are sampled equally (or almost equally—perfect U(1) symmetry
only applies at the critical point). There is an analogy in elementary quantum mechanics;
a particle in a slightly assymmetric double well. If the barrier between the wells is very
high, then the particle is localized in the deeper well, but if the barrier is not very high
(relative to the kinetic energy), then the particle fluctates between both wells and the
wave function also has weight inside the barrier. In the VBS, the wells correspond to the
columnar and plaquette states, at angles nπ /2 (columnar) and nπ /2 + π /4 (plaquette),
with n = 0, 1, 2, 3, and fluctuations between them correspond to non-zero probability in
the continuum of angles between these points. In the deconfinement theory, the columnar
and plaquette VBS states are almost degenerate, becoming exactly degenerate at the
critical point. Approaching the critical point, the barrier between them also is reduced.
In the double-well analogy, the wells are becoming more degenerate and the barrier
between them is reduced. There will then be significant fluctuations between columnar
and plaquette states in finite systems (up to L ∼ Λ), and this implies P(Dx , Dy ) with
weight at all the angles and an U(1) to Z4 cross-over of the symmetry when L > Λ.
This applies also to the order parameter computed on subsystems of finite length L in an
infinite (or very large) system (which has been studied in classical models with a q-fold
order parameter which exhibits U(1) to Zq symmetry cross-over [249, 250]).
Since the Z4 symmetry is not yet clearly manifested in the J-Q model up to the largest
lattices studied so far, the issue of the nature of the VBS remains open. However, a J-Q
model in which the Q term consists of products of three parallel singlet projectors, the
VBS order is more robust and a Z4 symmetric distribution corresponding to a columnar
324
state is seen clearly [108]. Simulations of SU(N) generalizations of the standard J-Q
model, with N = 3 and 4, also show columnar VBS ground states [108]. Most likely,
the standard SU(2) J-Q model also has a columnar state, and there is indeed some weak
hints of this in recent (ongoing) long simulations for L = 64.
For system sizes L ≫ Λ, the VBS distribution should be four-fold symmetric, and by
investigating how the change from U(1) to Z4 symmetry takes place one can in principle
determine the exponent a relating the deconfinement length scale Λ and the correlation
length ξ . Classical examples of the cross-over phenomenon, which also inspired the
theory [16], were studied in Refs. [249, 251, 250, 252]. A finite-size scaling method
for a quantity sensitive to Z4 symmetry can be used to extract the exponent a [252]. In
the standard J-Q model, such an analysis cannot yet be performed, because the cross-
over into a Z4 symmetric order parameter is not seen clearly enough in the currently
accessible lattices. The analysis has, however, been carried out in the J-Q model with
columnar six-spin interactions, with the result that a ≈ 1.2.
Other critical exponents. In addition to the dynamic exponent z = 1, other critical

exponents have been analyzed in several QMC studies [17, 238, 239, 108, 110]. The
most noteworthy result is that ηs (the exponent of the critical spin-spin correlation
function) is anomalously large, ηs ≈ 0.3, which is in qualitative agreement with the
theory of deconfined quantum critical points. A “large” ηs was argued for [16] based on
the N = ∞ value ηs = 1 in the 1/N expansion of the CPN−1 field theory (i.e., different
from the conventional mean-field value ηs = 0). The 1/N corrections are difficult to
compute, but in general support an unusually large exponent [237] (although it is not
possible to extend the results reliably to N = 2). The behavior of the dimer-dimer
exponent ηd obtained in QMC calculations for SU(N) systems [108] seems to follow
the scaling with N predicted in the theory [246, 253] (where ηd diverges with N).
In light of the presence of significant scaling corrections, possibly logarithmic, in
the susceptibility analyzed in Fig. 85, and also in the spin stiffness [110], an important
question is whether the calculated exponents also are affected. This has so far not
been apparent, because standard data collapse procedures with reasonable values of the
subleading exponents work very well [108]. The best current estimates for ηs , based on
both T > 0 [238] and T = 0 [108] QMC calculations agree well with each other. On
the other hand, the results discussed here and in [110] put the critcal point at slightly
higher J/Q than previously, and also the correlation length exponent ν analyzed with
a logarithmic correction included [110] is significantly smaller than other estimates
[238, 108]. It would therefore be good to repeat the studies of all the exponent, using
several different quantities and larger lattices (which is work in progress).
There is also evidence for logarithmic corrections in the J-Q model from studies of its
response to impurities. Introducing a vacancy into the system and studying the spatial
distribution of the resulting magnetization imbalance, one can normally, at conventional
quantum critical points in dimerized models, observe scaling with the system size in
this distribution [254, 255]. In the J-Q model, such scaling analysis again requires the
introduction of a multiplicative logarithmic correction [255].
Applications of other methods to the J-Q model. One interesting and puzzling as-
pects of the J-Q model is that it seems very difficult to capture its properties correctly
325
FIGURE 88. In the J-Q3 model studied here, three singlet projectors are arranged in a staggered pattern.
All distinct orientations (as shown) and translations of the projector products are included.
using analytical many-body techniques, even ones that can rather accurately locate the
Néel–VBS transition in the 2D frustrated J1 -J2 Heisenberg model. A mean-field treat-
ment starting from a columnar dimer state gives a critical J/Q very far from the QMC
result [256]. This approach can be improved to better take into account some of the local
fluctuations on plaquettes, which improves the value of the critical point but seems to
results in a strongly first-order transition [256]. Cluster mean-field calculations converge
very poorely with the cluster size [257]. These results point to unusually strong non-local
quantum fluctuations [256], which cannot be easily captured with local approaches start-
ing from small clusters or conventional fluctuations around a fixed dimer pattern. The
reason for these difficulties to capture the VBS state may be the emergent U(1) sym-
metry, which makes it difficult to obtain both the correct long-distance behavior (likely
columnar order) as well as the strong fluctuations between columnar and plaquette order
on shorter length scales.
First-order transition in a staggered J-Q model. One way to test the link between
emergent U(1) symmetry and a continuous Néel–VBS transition is to construct a model
in which the local fluctuations responsible for rotating the VBS angle are suppressed.
Intuition for how to accomplish this comes from the Rokhsar-Kivelson (RK) quantum
dimer model [258, 259], which can be regarded as an effective model for an extreme
nonmagnetic system dominated by short valence bonds (for which the internal singlet
structure is also neglected—the bond configurations are regarded as orthogonal states).
The RK hamiltonian on the square lattice can be written as HRK = vV − kK, where
V is the diagonal (potential-energy) operator, which counts the number of flippable
plaquettes [parallel bond pairs, exactly as in Fig. 86(b,c)], and K is an off-diagonal
(kinetic) term which flips such a pair. This model has a critical point at k = v which
separates a plaquette VBS state [similar to the one in Fig. 86(d)] for v < k and a staggered
VBS state [with the bond pattern exactly as in Fig. 4(c)]. While the plaquette state is
destroyed continuously by quantum fluctuations as v → k− , the staggered state (of which
there are four symmetry-related equivalent ones) has no fluctuations, because it has no
flippable plaquettes. The transition upon v → k+ is therefore first-order.
This simple picture of the RK model suggests that an actual staggered VBS in a spin
model also should have strongly suppressed local fluctuations, and therefore should
not be associated with an emergent U(1) symmetry. Due to the suppression of local
fluctuations (and therefore also of large-scale fluctuations), the transition between it
and the Néel state should be first-order. The picture is not complete, however, because
clearly there must be some fluctuations in the staggered VBS state, considering that
a reasonable spin hamiltonian will be quite far from a dimer model and the valence
326
1.0
-1.60
L = 36
0.5
-1.65
0.0
E/N
U2
-1.70
-0.5 L=6
L=8
-1.0 L = 10 -1.75
decreasing Q3/J
L = 12
increasing Q3/J
-1.5 L = 14 -1.80
1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.10 1.15 1.20 1.25 1.30
Q3/J Q3/J
FIGURE 89. Evidence for a first-order transition in the saggered J-Q3 model [109]. The Binder cumu-
lant of the sublattice magnetization (a) exhibits a negative peak, which grows with the system size. (b)
Hysteresis effects are observed: simulations in which the coupling ratio is slowly increased and decreased
give different properties in large systems, with the energies crossing each other at a point which can
be taken as a size-dependent transition point. Extrapolating these points for different system sizes gives
(Q3 /J)c = 1.1933(1), which is marked with the vertical dashed line in the left panel. All these results
were obtained with SSE simulations at inverse temperature Q3 /T = L.
bond configurations making up the gound state should also include some fraction of
longer bonds (in the presence of which some local fluctuations are always possible). It
is therefore worth testing this picture in simulations of a model whose ground state can
be tuned from a Néel antiferromagnet to a staggered VBS. Such a model can easily be
constructed within the J-Q framework, by arranging the singlet projectors of the Q term
in a staggered fashion, instead of within a 2 ×2 plaquette (or in columns of three or more
projectors, as in [108]). It turns out, however, that a Q term consisting of two staggered
projectors is not sufficient for destroying the Néel state. With three projectors, arranged
as in Fig. 88 (and the coupling of which we call Q3 ), a staggered VBS is stabilized,
however. We here briefly summarize the evidence for a first-order Néel—VBS found in
a recent SSE study of the staggered J-Q3 model [109] .
Some results are shown in Fig. 89. The Binder cumulant of the Néel order parameter
exhibits a negative peak, which grows with increasing system size. This is a strong
indication of a first-order transition, as discussed in Sec. 3.4 (and the results can be
compared with those of the classical frustrated Ising model in Fig. 20). The data become
very noisy for L > 12, because of the inefficiency of the simulations in transitioning
between the two coexisting states. For L > 24 the system can get completely trapped in
a metastable state. On the VBS side close to the transition, if the simulation happens to
enter a Néel like configuration during the equilibration, it will stay in a metastable Néel
state for a very long time (and vice versa on the other side of the transition). One can
use this trapping to investigate the transition in other ways. By doing a series of runs
at closely spaced values of Q3 /J, starting either far inside the VBS or Néel phase, and
starting each new run from the last configuration of the previous run, one can follow the
ground state into its corresponding metastable state at the “wrong” side of the transition.
The energies computed in two such runs, with Q3 /J either increasing or decreasing,
327
are shown in Fig. 89. The two branches cross each other, which clearly shows the first-
order nature of the transition. The level crossing is most likely an avoided one, between
two states with the same quantum numbers, but the spacing between the levels should
be extremely small here. This spacing should be inversely related to the tunneling time
between the two states at the coexistence point (but note that the relationship between
simulation time and the real time dynamics of the system is not known).
The crossing point of the two energy branches can be taken as a finite-size definition
of the critical coupling. It shifts with the system size approximately as 1/L3 , which is
consistent with a first-order transition with dynamic exponent z = 1 [260]. The extrap-
olated transition point is Q3 /J = 1.1933(1). At the transition, the VBS order parameter
is about 75% of its maximum value (in a perfect staggered valence bond state), which is
rather large and motivates the classification of the transition as strongly first-order.
6. SURVEY OF RELATED COMPUTATIONAL METHODS

In these lecture notes we have discussed exact diagonalization methods and the SSE
QMC method in some detail and also looked at some illustrative calculations with results
for S = 1/2 models. There are many issues that were left untouched, regarding methods
as well as physics, by restricting the discussion to S = 1/2 systems, and within that class
also to spin-isotropic systems. The exact diagonalization approach can be easily used
for any spin model, but the rapidly growing size of the Hilbert space with S imposes
even more severe restrictions on the lattices sizes for S > 1/2 systems. The world line
and SSE QMC methods can be generalized to any spin model, as long as there is no
frustration leading to sign problems. In this section we briefly summarize some QMC
schemes for more general spin models, as well as some further recent developments for
isotropic S = 1/2 systems.
There are several other important computational methods that go beyond the scope of
these lecture notes, such as series expansions techniques (high-temperature expansions
and expansions around some solvable limit of the ground state, e.g., the Ising limit of an
anisotropic Heisenberg model) [261, 262] and the DMRG method [28, 29]. While series
expansions can in principle be applied to any model, there are issues with extrapolating
the series in a controllable manner, and one cannot, in most cases of interest, expect to
reach the same level of accuracy as in QMC simulations of sign problem free models
(e.g., in studies of quantum phase transitions). Series expansion methods are neverthe-
less one of the most powerful classes of methods currently available for 2D frustrated
quantum spin systems [100], where QMC methods are limited by the sign problem (but
note recent progress in controlling the sign problem at high temperatures [193]). More
powerful series expansion schemes are also still being actively developed [263], and one
can expect progress to continue on this front. The DMRG method is very powerful for
1D systems and can also be used for 2D systems of moderate size [94]. Extensions of the
DMRG method and the related matrix-product states [26, 264, 21] to higher-dimensional
“tensor-network” states [21, 106, 265] are currently being explored very intensely. This
is an exciting line of research, that may eventually lead to viable schemes for unbiased
studies of frustrated spins (and even fermionic systems). It appears unlikely, however,
that these methods will, for the foreseeable future, be able to probe quantum spin systems
328
(a) (b) (c) (d)
FIGURE 90. Examples of the four possible paths through vertices in the anisotropic S = 1/2 Heisenberg
model (which may include also an external magnetic field) and the new vertices generated when moving
along the paths in the directed loop algorithm. The vertical lines at the vertices to the left in each box
indicate the entrance and exit legs (with either direction of movement being possible). Given an entrance
leg, the exit must be chosen such that a new allowed vertex is generated when the leg spins are flipped (the
vertices to the right, without the path indicated). In the “bounce” process (d), the exit is at the same leg as
the entrance and the vertex is not modified. There are six allowed vertices in the generic anisotropic
Heisenberg model—those in which the spin (z component) is conserved. In the “deterministic” loop
algorithm discussed in Sec. 5, there are only four vertices and only paths of type (a) are allowed
at the level of precision possible with existing QMC methods for unfrustrated systems.
In light of the above advantages and disadvantages of different methods, the author
advocates a two-pronged approach in research on quantum spin models and related
systems: (i) Explore new and improved schemes (including also QMC schemes) that
may be useful to study models beyond the reach of currently available QMC techniques.
(ii) Study interesting physics with state-of-the-art QMC methods for sign problem free
models. It is far from true that everything that can be done has already been done with
sign problem free models (although one can some times hear such claims)! With the
QMC methods available now, such as the SSE algorithm described in Sec. 5, and with
the power of modern computers (which still is increasing at an amazing pace, e.g., with
a steadily increasing numbers of cores per CPU), it is possible to access very interesting
physics that was beyond reach only a few years ago. The discussion of the dimerized and
J-Q classes of models in Sec. 5.3.5, in particular, has hopefully convinced the reader that
there is much to explore in such systems, and beyond (e.g., random systems exhibiting
unusual states and quantum phase transitions [266, 267, 268, 278]).
Directed loop QMC algorithms. The directed loop algorithm is a generalization of

the SSE operator-loop scheme, introduced in [33] (building on a previous less efficient
formulation [190]) and applied there to S = 1/2 Heisenberg systems with interaction
anisotropies (Ising and XY) and external magnetic field. The main difference between
loops and directed loops (in both SSE and world line formulations) is that the path
through the vertices is not unique in the directed loop scheme. The path (the exit leg,
given an entrance leg to a vertex) is chosen according to certain probabilities, constructed
to maintain detailed balance in a space with two defects—the open loop ends (in a
way similar in spirit to the “worm” algorithm [32]). For S = 1/2 systems, there are
four types of vertex paths, corresponding to the location of the exit relative to the
entrance, as illustrated and discussed in more detail in Fig. 90. The path probabilities,
which should be solutions of the corresponding directed loop equations (for detailed
balance) are in general not unique. For some models, eliminating the “bounce” process
in Fig. 90(d) reduces the directed loop scheme to one of the standard loop updates
previously developed for S = 1/2 (as in the isotropic system discussed in Sec. 5)
329
and higher-S models [191, 31, 189]. In other cases the directed loops (and worm)
algorithms allow for efficient simulations where the standard loop algorithms are not
applicable [33, 269, 270, 271]. Here it can also be mentioned that it is some times useful
in SSE directed loop algorithms to work with operators defined on cells larger than
those containing the elementary operators of the Hamiltonian (e.g., the two-spin bond
operators of the Heisenberg model) [272]. This allows for more options in choosing how
operators are reconfigured when vertices are traversed by the loops, which some times
can make simulations much more efficient.
QMC algorithms in the valence-bond basis. In Sec. 2.2 we briefly discussed the
valence bond basis, in which a basis state is a product of two-spin singlets, as in
Eq. (20), and in which any total singlet singlet state can be expanded. The basis is
overcomplete and non-orthogonal, which implies that such an expansion is not unique.
One way to work with the valence bond basis is to construct and optimize variational
states, the simplest type of which is an amplitude product states [273]. Such a state is a
superposition
|Ψi = ∑ Ψα |Vα i, (297)
α
where |Vα i is a valence bond product state of the form (20) and the expansion coeffi-
cients are products of amplitudes h(rα ,i ) corresponding to the “shapes” of the bonds (the
bond lengths in the x and y direction in the case of a 2D system), with i = 1, . . . , N/2
referring to the N/2 valence bonds in the configuration (bond tiling) labeled by α ;
N/2
Ψα = ∏ h(rα ,i ). (298)
i=1
Amplitude product states of this kind can closely reproduce the ground states of many √
bipartite Heisenberg systems, for which each singlet (a, b) = (| ↑a ↓b i − | ↓a ↑b i)/ 2
should be defined with site a on sublattice A and site b on sublattice B. With all positive
expansion coefficients Ψα in (297), this convention for the singlet sign corresponds to
Marshall’s sign rule for the ground state wave function of a bipartite system [274],
which in the standard ↑, ↓ spin basis can be written as
sign[Ψα ] = Ψα /|Ψα | = (−1)nA↓ , (299)
where nA↓ is the number of ↓ spins on sublattice A.

The properties of these bipartite amplitude product states can be studied by Monte
Carlo sampling of the valence bonds [273, 275, 50]. For the 2D Heisenberg model,
the state with all the amplitudes h(x, y) variationally optimized is a Néel state with
properties in very good agreement with the true ground state [275], e.g., the energy is
within 0.1% and the sublattice magnetization within 1% of the values obtained in QMC
calculations. For frustrated systems, appropriate sign rules are not known (and may be
too complicated to write down in practice).
The idea to use amplitude product states as a starting point for projector QMC
simulations is quite old [276, 277], but was only recently developed into a generic
and efficient tool [47, 49, 50]. The general idea of a projector scheme is the same as
330
FIGURE 91. Example (from Ref. [50]) of a loop configuration in the valence-bond projector QMC
method formulated in a combined space of spins (open and solid circles) and valence bonds (the arcs
capping the loops on the left and right boundaries). The bars with four spins are diagonal and off-diagonal
vertices with the same meaning and function as in the SSE operator-loop method (illustrated in Fig. 61).
Loops form according to the connectivity of the vertices, and also through the valence bonds at the
boundaries, and can be flipped without changing the configuration weight. The sum over all possible
loop configurations corresponds exactly to the formulation of the projector scheme in the pure valence
bond basis (i.e., without using the spins at all) [47]. Using the spins allows for a more efficient sampling
of the configurations. Expectation values are evaluated using loop estimators at the mid-point indicated
with a vertical dashed line, and this is analogous to the improved estimators (discussed in Sec. 5.2.5) in
the finite-temperature versions of the loop algorithm.
we discussed in the context of Krylov space methods in Sec. 4.2: By acting with a
high power H Λ of the hamiltonian on an arbitrary state |Ψi, only the component of
that state with the largest energy eigenvalue—normally the ground state (and if not this
can be arranged by subtracting a constant from H)—survives in the limit Λ → ∞, as
demonstrated by Eq. (170). Expectation values of the form
hΨ|H Λ AH Λ |Ψi
hAi = , (300)
hΨ|H 2Λ |Ψi
where |Ψi is a valence bond state or superposition (e.g., an amplitude product state), can
be sampled using Monte Carlo simulations. In the original formulations of this approach
[62, 47], H is written as a sum of singlet projector operators (as in the Heisenberg
antiferromagnet), or products of singlet operators (as in J-Q models), and strings of such
operators are sampled (along with the valence bond configurations of |Ψi). The singlet
projectors Si j , defined in Eq. (21), lead to simple reconfigurations of bond pairs when
acting on valence bond states. When acting on a valence bond, this operator is diagonal
with eigenvalue 1,
Sab (a, b) = (a, b), (301)
while acting on a pair of different valence bonds leads to a simple reconfiguration of
those bonds, with matrix element 1/2;
Sbc (a, b)(c, d) = 12 (c, b)(a, d), (302)

This can be shown easily by going back to the basis of ↑ and ↓ spins. Note the order of
the indices within the singlets, which reflects consistently the convention corresponding
to Marshall’s sign rule discussed above; a, c are on sublattice A and b, d on sublattice B.
These rules allow for a well-defined path integral (or SSE) like propagation of the states,
and these propagation paths can be sampled using a Monte Carlo scheme.
331
In a more recent formulation of the valence bond projector method [50], the ↑ and
↓ spin basis is reintroduced, by expressing all singlets in terms of their sums over
antiparallel spin pairs. This leads to an algorithm very similar to SSE and world line
loop methods. Essentially, the periodic time boundaries used when simulating a system
at a fixed temperature are cut open and replaced by two separate boundaries at which the
valence bond states live. This is illustrated and discussed in more detail with a simple
configuration for a four-site Heisenberg chain in Fig. 91.
The valence bond basis (or its translation to loop methods [244, 50]) has some unique
aspects which makes it possible to access observables that are normally out of reach
or difficult to calculate. In an extended valence bond basis for the triplet sector, one
can study some properties of excited states [47, 49, 278]. One can also generalize the
valence bond basis to include one [279] or several [278] unpaired spins, which is useful
for studies of, e.g., the magnetization distribution in systems with unequal sublattice
occupation of the spins. Simulations in the valence bond basis have also recently found
applications in studies of entanglement entropy [23, 280, 281, 282, 283]. One can also
extend valence bond projector methods to SU(N) spins [108] (including even non-integer
N generalizations [248]) and other related symmetry groups [282].
ACKNOWLEDGMENTS
These lecture notes build largely on material to be published as a book, Quantum Spin
Systems—A Computational Perspective, by Cambridge University Press. The opportu-
nity to lecture at the XIV Training Course in the Physics of Strongly Correlated Systems
(Vietri sul Mare, Salerno, Italy, in October 2009) prompted me to prepare a condensed
version of this as of yet unfinished work (which will discuss a larger set of models and
computational methods). These lecture notes cover the topics I discussed at Vietri sul
Mare, but I have added more background and details. Some of the applications have
been complemented with recent results. As in my lectures, the target audience for these
notes is graduate students and postdocs who are entering the field of quantum mag-
netism, or some related area in where quantum spin models are of interest, and need a
source for elementary concepts and calculations. I believe that this material could also
form the core of a one-semester advanced topics course in quantum magnetism (focused
on concepts and numerical calculations).
I would like to thank the organizers of the Salerno Training Course, Professors
Ferdinando Mancini and Adolfo Avella, for the invitation to lecture and publish these
notes. I also thank the students at Vietri sul Mare for their enthusiasm. I am grateful
to Ying Tang for her careful reading of the manuscript. I thank Cambridge University
Press and the American Institute of Physics for agreeing on the copyright issues related
to dual publication of most of the material presented here. Excerpts are reprinted with
the permission of Cambridge University Press. Some of this work was supported by the
NSF under Grant No. DMR-0803510.
332
REFERENCES
1. S. Sachdev, Quantum Phase Transitions (Cambridge University Press, 1999).
2. S. Sachdev, Nature Physics 4, 173 (2008).
3. S. Sachdev, arXiv:0901.4103.
4. A. Auerbach, Quantum Magnetism (Springer-Verlag, 1994).
5. S. Chakravarty, B. I. Halperin, and D. R. Nelson, Phys. Rev. B 39, 2344 (1989).
6. E. Manousakis, Rev. Mod. Phys. 63, 1 (1991).
7. S. Eggert, I Affleck, and H. Takahashi, Phys. Rev. Lett. 73, 332 (1994).
8. S. Eggert, Phys. Rev. B 53, 5116 (1996).
9. E. Dagotto and T. M. Rice, Science 271, 618 (1996).
10. J. S. Miller and A. J. Epstein, MRS Bulletin 25, 21 (2000).
11. U. Schollwöck, J. Richter, D. J. J. Farnell, and R. F. Bishop (Editors), Quantum Magnetism, Lecture
Notes in Physics, 645 (Springer, 2004).
12. H. T. Diep (Editor), Frustrated Spin Systems (World Scientific, Singapore, 2004).
13. J. S. Gardner, M. J. P. Gingras, and J. E. Greedan, Rev. Mod. Phys. 82, 53 (2010).
14. L. Balents, Nature 464, 199 (2010).
15. J. Cardy, Scaling and Renormalization in Statistical Physics (Cambridge University Press, 1996).
16. T. Senthil, A. Vishwanath, L. Balents, S. Sachdev, and M. P. A. Fisher, Science 303, 1490 (2004).
17. A. W. Sandvik, Phys. Rev. Lett. 98, 227202 (2007).
18. S. Sachdev and X. Yin, Annals of Physics, 325, 2 (2010).
19. I. Bloch, J. Dailbard, and W. Zwerger, Rev. Mod. Phys. 80, 885 (2008).
20. K. Kim et al., Nature 465, 590 (2010).
21. F. Verstraete, V. Murg, and J. I. Cirac, Adv. Phys. 57, 143, (2008).
22. M. Levin and X.-G. Wen, Phys. Rev. Lett 96, 110405 (2006).
23. M. B. Hastings, I. Gonzt’alez, A. B. Kallin, and R. G. Melko, Phys. Rev. Lett. 104, 157201 (2010).
24. E. H. Lieb and D. C. Mattis (Editors), Many-Body Problem: An Encyclopedia of Exactly Solved
Models in 1d (World Scientific, 1994).
25. B. S. Shastry and B. Sutherland, Physica B 108, 1069 (1981).
26. I Affleck, T. Kennedy, E. Lieb, and H. Tasaki, Commun. Math. Phys. 115, 477 (1988).
27. L. A. Fernandez, V. Martin-Mayor, S. Perez-Gaviro, A. Tarancon, and A. P. Young, Phys. Rev. B
80, 024422 (2009).
28. S. R. White, Phys. Rev. Lett. 69, 2863 (1992).
29. U. Schollwöck, Rev. Mod. Phys. 77, 259 (2005).
30. I. P. McCulloch, arXiv:0804.2509.
31. H. G. Evertz, Adv. Phys. 52, 1 (2003).
32. N. V. Prokofév, B. V. Svistunov, and I. S. Tupitsyn, Zh. Eks. Teor. Fiz. 114, 570 (1998) [JETP 87,
310 (1998)]; arXiv:cond-mat/9703200.
33. O. F. Syljuåsen and A. W. Sandvik, Phys. Rev. E 66, 046701 (2002).
34. Links to codes and instructions are available on-line at: http://physics.bu.edu/˜sandvik.
35. P. W. Anderson, Phys. Rev. B 115, 2 (1959).
36. D. C. Mattis, The Theory of Magnetism I (Springer-Verlag, Berlin 1988).
37. J.-Y. P. Delannoy, M. J. P. Gingras, P. C. W. Holdsworth, and A.-M. S. Tremblay, Phys. Rev. B 79,
235130 (2009).
38. D. Mermin and H. Wagner, Phys. Rev. Lett. 17 , 1133 (1966).
39. P. C. Hohenberg, Phys. Rev. 158 , 383 (1967).
40. C. M. Canali and S. M. Girvin, Phys. Rev. B 45, 7127 (1992).
41. C. J. Hamer, Z. Weihong, and P. Arndt, Phys. Rev. B 46, 6276 (1992).
42. J. Igarashi, Phys. Rev. B 46, 10763 (1992).
43. E. Dagotto, Rev. Mod. Phys. 66, 763 (1994).
44. T. Aimi and M. Imada, J. Phys. Soc. Jpn. 76, 113708 (2007).
45. S. R. White and D. J. Scalapino, Phys. Rev. B 79, 220504 (2009).
46. S. Sachdev, Rev. Mod. Phys. 75, 913 (2003).
48. K. S. D. Beach and A. W. Sandvik, Nucl. Phys. B 750, 142 (2006).
49. A. W. Sandvik and K. S. D. Beach, arXiv:0704.1469.
333
50. A. W. Sandvik and H. G. Evertz, Phys. Rev. B 82, 024407 (2010).
51. H. A. Bethte, Z. Phys. 71, 205 (1931).
52. L. Hulthén, Ark. Mat. Astron. Fys. 26A, 1 (1938).
53. J.-S. Caux and R. Hagemans, J. Stat. Mech. (2006) P12013.
54. J.-M. Maillet, in Quantum Spaces, Edited by B. Duplantier and V. Rivasseau (Birkhäuser Verlag,
Berlin 2007).
55. F. D. M. Haldane, Phys. Rev. Lett. 50, 1153 (1983).
56. I. Affleck, Phys. Rev. Lett. 55, 1355 (1985).
57. I. Affleck, D. Gepner, H. J. Schulz, and T. Ziman, J. Math. Phys. A: Math. Gen. 22, 511 (1989).
58. R. R. P. Singh, M. E. Fisher, and R. Shankar, Phys. Rev. B 39, 2562 (1989).
59. T. Giamarchi and H. J. Schulz, Phys. Rev. B 39, 4620 (1989).
60. C. K. Majumdar and D. K. Ghosh, J. Math. Phys. 10, 1388 (1969), ibid., 1399 (1969).
61. K. Nomura, Phys. Rev. B 40, 2421 (1989).
62. S. Liang, Phys. Rev. Lett. 64, 1597 (1990).
63. U. Schollwöck, O. Golinelli, and T. Jolicoeur, Phys. Rev. B 54, 4038 (1996).
64. D. A. Tennant, T. G. Perring, R. A. Cowley, and S. E. Nagler, Phys. Rev. Lett. 70, 4003 (1993).
65. L. P. Regnault I. Zaliznyak, J. P. Renard, and C. Vettier, Phys. Rev. B 50, 9174 (1994).
66. G. E. Granroth, M. W. Meisel, M. Chaparala, Th. Jolicoeur, B. H. Ward, and D. R. Talham, Phys.
Rev. Lett. 77 1616 (1996).
67. R. G. Pereira, J. Sirker, J.-S. Caux, R. Hagemans, J. M. Maillet, S. R. White, and I. Affleck Phys.
Rev. Lett. 96, 257202 (2006).
68. J.-C. Caux and J. M. Maillet, Phys. Rev. Lett. 95, 077201 (2005).
69. A. Läuchli, G. Schmid, and S. Trebst, B 74, 144426 (2006).
70. A. W. Sandvik and D. K. Campbell, Phys. Rev. Lett. 83, 195 (1999).
71. F. Michel and H. G. Evertz, arXiv:0705.0799.
72. N. Laflorencie, I. Affleck, and M. Berciu, J. Stat. Mech. (2005) P12001.
74. Ch. Rügg et al., Phys. Rev. Lett. 100, 205701 (1008).
75. R. R. P. Singh, M. P. Gelfand, and D. A. Huse, Phys. Rev. Lett. 61, 2484 (1988).
76. X.-G. Wen, Quantum Field Theory of Many-body Systems (Oxford University Press, 2004).
77. A. V. Chubukov and D. K. Morr, Phys. Rev. B 52, 3521 (1995).
78. D. A. Huse, Phys. Rev. B 37, 2380 (1988).
79. J. D. Reger and A. P. Young, Phys. Rev. B 37, 5978 (1988).
80. A. J. Millis and H. Monien, Phys. Rev. B 50, 16606 (1994).
81. A. W. Sandvik and D. J. Scalapino, Phys. Rev. Lett. 72, 2777 (1994).
82. A. W. Sandvik, Phys. Rev. Lett. 83, 3069 (1999)
83. P. V. Shevchenko, A. W. Sandvik, and O. P. Sushkov, Phys. Rev. B 61, 3475 (2000).
84. A. V. Chubukov, S. Sachdev, and J. Ye, Phys. Rev. B 49, 11919 (1994).
85. L. Wang, K. S. D. Beach, and A. W. Sandvik, Phys. Rev. B 73, 014431 (2006).
86. M. Matsumoto, C. Yasuda, S. Todo, and H. Takayama, Phys. Rev. B 65, 014407 (2001).
87. M. Troyer, H. Kontani, and K. Ueda, Phys. Rev. Lett. 76, 3822 (1996).
88. S. Wenzel. and W. Janke, Phys. Rev. B 79, 014410 (2009).
89. S. Wenzel, L. Bogacz, and W. Janke, Phys. Rev. Lett. 101, 127202 (2008).
90. J. Stephenson, J. Math. Phys. 11, 413 (1970).
91. H. W. J. Blöte and H. J. Hilhorst, J. Phys. A: Math. Gen. 15, L631 (1982).
92. P. Fazekas and P. W. Anderson, Philos. Mag. 30, 23 (1974).
93. B. Bernu, P. Lecheminant, C. Lhuillier, and L Pierre, Phys. Rev. B 50, 10048 (1994).
94. S. R. White and A. L. Chernyshev, Phys. Rev. Lett. 99, 127004 (2007).
95. J. Yin and D. P. Landau, Phys. Rev. E 80, 051117 (2009).
96. A. Kalz, A. Honecker, S. Fuchs, T. Pruschke, J. Phys.: Conf. Ser. 145, 012051 (2009).
97. S. Jin, A. Sen, and A. W. Sandvik (manuscript in preparation).
98. E. Dagotto and A. Moreo, Phys. Rev. Lett. 63, 2148 (1989).
99. H. J. Schulz, T. Ziman, and D. Poilblanc, J. Phys. I 6, 675 (1996).
100. R. R. P. Singh, Z. Weihong, C. J. Hamer, and J. Oitmaa, Phys. Rev. B 60, 7278 (1999).
101. O. P. Sushkov, J. Oitmaa, and Z. Weihong, Phys. Rev. B 63, 104420 (2001).
102. F. Krüger and S. Scheidl, Europhys. Lett. 74, 896 (2006).
334
103. K. S. D. Beach, Phys. Rev. B 79, 224431 (2009).
104. L. Isaev, G. Ortiz, and J. Dukelsky, Phys. Rev. B 79, 024409 (2009).
105. L. Capriotti, F. Becca, A. Parola, and S. Sorella, Phys. Rev. Lett. 87, 097201 (2001).
106. V. Murg, F. Verstraete, J. I. Cirac, Phys. Rev. B 79, 195119 (2009).
107. N. Read and S. Sachdev, Phys. Rev. Lett. 62, 1694 (1989).
108. J. Lou, A. W. Sandvik, and N. Kawashima, Phys. Rev. B 80, 180414 (2009).
109. A. Sen and A. W. Sandvik, arXiv:1009.0691.
111. L. Onsager, Phys. Rev. 65, 117 (1944).
112. R. J. Baxter, Exactly Solved Models in Statistical Mechanics (Academic Press, New York, 1982).
113. M. Campostrini, M. Hasenbusch, A. Pelissetto, P. Rossi, and E. Vicari, Phys. Rev. B 65, 144520
(2002).
114. R. Guida and Jean Zinn-Justin, J. Phys. A: Math. Gen. 31, 8103 (1998).
115. D. Landau and K. Binder, A Guide to Monte Carlo Simulations in Statistical Physics (Cambridge
University Press, 2005).
116. M. E. J. Newman and G. T. Barkema, Monte Carlo Methods in Statistical Physics (Oxford Univer-
sity Press, 2002).
117. N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller, J. Chem. Phys. 21,
1087 (1953).
118. R. H. Swendsen and J.-S. Wang, Phys. Rev. Lett. 58, 86 (1987).
119. U. Wolff, Phys. Rev. Lett. 62, 361 (1989).
120. M. E. Fisher and M. N. Barber, Phys. Rev. Lett. 28, 1516 (1972).
121. K. Binder, Phys. Rev. Lett. 47, 693 (1981).
122. K. Binder and D. P. Landau, Phys. Rev. B 30, 1477 (1984).
123. M. E. Fisher and A. N. Berker, Phys. Rev. B 26, 2507 (1982).
124. K. Vollmayr, J. D. Reger, M. Scheucher, and K. Binder, Z. Phys. B 91, 113 (1991).
125. J. Lee and J. M. Kosterlitz, Phys. Rev. B 43, 3265 (1991).
126. E. Marinari, in Advances in Computer Simulation, edited by J. Kertsz and I. Kondor, Lecture Notes
in Physics 501 (Springer-Verlag, Berlin, 1998).
127. K. Hukushima and K. Nemoto, J. Phys. Soc. Jpn. 65, 1604 (1996).
128. B. A. Berg and and T. Neuhaus, Phys. Rev. Lett. 68, 9 (1992).
129. F. Wang and D. P, Landau, Phys. Rev. Lett. 86, 2050 (2001); Phys. Rev. E 64, 056101 (2001).
130. S. Trebst, D. A. Huse, and M. Troyer, Phys. Rev. E 70, 046701 (2004).
131. M. E. Fisher, M. N. Barber, and D. Jasnow, Phys. Rev. A 8, 1111 (1973).
132. B. D. Josephson, Phys. Lett. 21, 608 (1966).
133. N. Schultka and E. Manousakis, Phys. Rev. B 49, 12071 (1994).
134. N. Schultka and E. Manousakis, Phys. Rev. B 51, 11712 (1995).
135. J. M. Kosterlitz and D. J. Thouless, J. Phys. C 6, 1181 (1973).
136. J. M. Kosterlitz, J. Phys. C 7, 1046 (1974).
137. Y. Tomita and Y. Okabe, Phys. Rev. B 65, 184405 (2002).
138. H. Weber, P. Minnhagen, Phys. Rev. B 37 (1987) 5986.
139. D. R. Nelson, J. M. Kosterlitz, Phys. Rev. Lett. 39 1201 (1977).
140. T. Vojta and R. Sknepnek, Phys. Stat. Sol. 241, 2118 (2004).
141. B. I. Halperin, P. C. Hohenberg and E. D. Siggia, Phys. Rev. B 13, 1299 (1976).
142. S. L. Sondhi, S. M. Girvin, J. P. Carini, and D. Shahar, Rev. Mod. Phys. 69, 315 (1997).
143. See, e.g., the chapter by N. Laflorencie and D. Poilblanc in Ref. [11]; Lecture Notes in Physics 645,
227 (2004) [arXiv:cond-mat/0408363].
144. M. Mambrini, A. Läuchli, D. Poilblanc, and F. Mila, Phys. Rev. B 74, 144422 (2006).
145. H. Q. Lin Phys. Rev. B 42, 6561 (1990).
146. J. C. Bonner and M. E. Fisher, Phys. Rev. 135, A640 (1964).
147. W. E. Estes, D. P. Gravel, W. E. Hatfield, and D. J. Godjson, Inorg. Chem. 17, 1415 (1978).
148. C. Lanczos, J. Res. Nat. Bur. Stand. 49 33 (1952).
149. J. des Cloizeaux and J. J. Pearson, Phys. Rev. 128, 2131 (1962).
150. F. Woynarovich and H. -P. Eckel, J. Phys. A 20 (1987) L97; A 20 L443 (1987).
151. H. J. Schulz, Phys. Rev. Lett. 64, 2831 (1990).
152. J. Voit, Rep. Prog. Phys. 58 977 (1995).
335
153. I Affleck, J. Phys. A: Math. Gen. 22, 511 (1989).
154. J. Lorenzana and R. Eder, Phys. Rev. B 55 3358 (1997).
155. M. Takigawa, O. A. Starykh, A. W. Sandvik, and R. R. P. Singh, Phys. Rev. B 56, 13681 (1997).
156. B. Lake, D. A. Tennant, C. D. Frost, and S. E. Nagler, Nature Materials 4, 329 (2005).
157. K. Nomura and K. Okamoto, Phys. Lett. A 169, 433 (1992).
158. R. Bursill, G. A. Gehring, D. J. J. Farnell, J. B. Parkinson, T. Xiang, and C. Zeng, J. Phys. Condens.
Matter 7, 8605 (1995).
159. M. Kumar, Z. G. Soos, D. Sen, and S. Ramasesha, Phys. Rev. B 81, 104406 (2010).
160. T. Hikihara, L. Kecke, T. Momoi, and A. Furusaki, Phys. Rev. B 78, 144404 (2008).
161. P. M. van den Broek, Phys. Lett. A 77, 261 (1980).
162. B. S. Shastry and B. Sutherland, Physica B 108, 1069 (1981).
163. D. J. Klein, J. Phys. A 15, 661 (1982).
164. S. Eggert, Phys. Rev. B 54, R9612 (1996).
165. F. D. M. Haldane, Phys. Rev. Lett. 60, 635 (1988).
166. B. S. Shastry, Phys. Rev. Lett. 60, 639 (1988).
167. F. H. L. Essler, Phys. Rev. B 51, 13357 (1995).
168. A. W. Sandvik, Phys. Rev. E 68, 056701 (2003).
169. P. W. Anderson, Phys. Rev. B 115, 2 (1959).
170. T. Oguchi, Phys. Rev. 117, 117 (1960).
171. C. Lavalle, S. Sorella, and A. Parola, Phys. Rev. Lett. 80, 1746 1998.
172. J. Oitmaa and D. D. Betts, Can. J. Phys. 56, 897 (1978).
173. J. D. Reger and A. P. Young, Phys. Rev. B 37, 5978 (1988).
174. D. D. Betts, H.-Q. Lin, and J. S. Flynn, Can. J. Phys. 77, 353 (1999).
175. R. P. Feynman, Phys. Rev. 91, 1291 (1953).
176. J. E. Hirsch, R. L. Sugar, D. J. Scalapino, and R. Blankenbecler, Phys. Rev. B 26, 5033 (1982).
177. M. Suzuki, Prog. Theor. Phys. 56, 1454 (1976).
178. M. Suzuki, S. Miyashita and A. Kuroda, Prog. Theor. Phys. 58, 1377 (1977).
179. B. B. Beard and U.-J. Wiese, Phys. Rev. Lett. 77, 5130 (1996).
180. D. C. Handscomb, Proc. Cambridge Philos. Soc. 58, 594 (1962).
181. D. C. Handscomb, Proc. Cambridge Philos. Soc., 60 115 (1964).
182. J. W. Lyklema, Phys. Rev. Lett. 49, 88 (1982).
183. D. H. Lee, J. D. Joannopoulos and J. W. Negele, Phys. Rev. B 30, 1599 (1984).
184. E. Manousakis and R. Salvador, Phys. Rev. B 39, 575 (1989).
185. S. C. Chakravarty and D. B. Stein, Phys. Rev. Lett. 49, 582 (1982).
186. A. W. Sandvik and J. Kurkijärvi, Phys. Rev. B 43, 5950 (1991).
187. A. W. Sandvik, J. Phys. A 25, 3667 (1992).
188. H.-P. Ying, U.-J. Wiese, and D.-R. Jia, Phys. Lett. A 183, 441 (1993).
189. N. Kawashima and J. E. Gubernatis, Phys. Rev. Lett. 73, 1295 (1994).
190. A. W. Sandvik, Phys. Rev. B 59, R14157 (1999).
191. H. G. Evertz, G. Lana, and M. Marcu, Phys. Rev. Lett. 70, 875 (1993).
192. P. Henelius and A. W. Sandvik, Phys. Rev. B 62, 1102 (2000).
193. M. Nyfeler, F.-J. Jiang, F. Kämpfer, and U.-J. Wiese, Phys. Rev. Lett. 100, 247206 (2008).
194. D. M. Ceperley, Rev. Mod. Phys. 67, 279 (1995).
195. M. Boninsegni, N. V. Prokof’ev, and B. V. Svistunov, Phys. Rev. E 74, 036701 (2006).
196. D. M. Ceperley and E. L. Pollock, Phys. Rev. Lett. 56, 351 (1986).
197. P. Werner, K. Völker, M. Troyer, and S. Chakravarty, Phys. Rev. Lett. 94, 047201 (2005).
198. H. F. Trotter, Proc. Am. Math. Soc. 10, 545 (1959).
199. H. De Raedt and B. De Raedt, Phys. Rev. A 28, 3575 (1983).
200. N. Hatano and M. Suzuki, in Quantum Annealing and Other Optimization Methods, Edited by B.
K. Chakrabarti and A. Das (Springer-Verlag, 2005).
201. R. M. Fye, Phys. Rev. B 33, 6271 (1986).
202. R. M. Fye and R. T. Scalettar, Phys. Rev. B 36, 3833 (1987).
203. M, S. Makivić and H.-Q. Ding, Phys. Rev. B 43, 3562 (1991).
204. S. Wessel, N. Stoop, E. Gull, S. Trebst, and M. Troyer, J. Stat. Mech., P12005 (2007).
205. A. W. Sandvik, Phys. Rev. B 56, 11678 (1997).
206. P. Henelius, S. M. Girvin, and A. W. Sandvik, Phys. Rev. B 57, 13382 (1998).
336
207. A. W. Sandvik, R. R. P. Singh, and D. K. Campbell, Phys. Rev. B 56, 14510 (1997).
208. A. W. Sandvik and D. J. Scalapino, Phys. Rev. B 51, 9403-9406 (1995).
209. K. Harada and N. Kawashima, Phys. Rev. B 55 11949 (1997).
210. E. L. Pollock and D. M. Ceperley, Phys. Rev. B 36, 8343 (1987).
211. D. E. Knuth, The Art of Computer Programming, Vol. 2: Seminumerical Algorithms (Addison-
Wesley, Menlo Park, 1998).
212. M. Karbach, K. Hu, and G. Müller, Computers in Physics 12, 565 (1998).
213. K. Nomura, Phys. Rev. B 48, 16814 (1993).
214. E. Dagotto, J. Riera, and D. Scalapino, Phys. Rev. B 45, 5744 (1992).
215. D. N. Sheng, O. I. Motrunich, and M. P. A. Fisher, Phys. Rev. B 79, 205112 (2009).
216. B. Frischmuth B. Ammon, and M. Troyer, Phys. Rev. B 54, R3714 (1996).
217. S. R. White, R. M. Noack, and D. J. Scalapino, Phys. Rev. Lett. 73, 886 (1994).
218. M. Troyer, H. Tsunezugu, and D. Würtz, Phys. Rev. B 50, 13515 (1994)
219. T. Sakai, M. Sato, K. Okunishi, Y. Otsuka, K. Okamoto, and C. Itoi, Phys. Rev. B 78, 184415
(2008).
220. S. Dell’Aringa, E. Ercolessi, G. Morandi, P. Pieri, and M. Roncaglia, Phys. Rev. Lett. 78, 2457
(1997).
221. B. Frischmuth, S. Haas, G. Sierra, and T. M. Rice, Phys. Rev. B 55, R3340 (1997).
222. C. J. Hamer, Z. Weihong, and J. Oitmaa, Phys. Rev. B 50, 6877 (1994).
223. K. J. Runge, Phys. Rev. B 45, 7229 (1992); Phys. Rev. B 45, 12292 (1992).
224. U.-J. Wiese and H.-P. Ying, Phys. Lett. A 168, 143 (1992).
225. J.-K. Kim and M. Troyer, Phys. Rev. Lett. 80, 2705 (1998).
226. B. B. Beard, R. J. Birgeneau, M. Greven, and U.-J. Wiese, Phys. Rev. Lett. 80, 1742 (1998).
227. P. Hasenfratz and F. Niedermayer, Z. Phys. B 92, 91 (1993).
228. T. Einarsson and H. J. Schulz, Phys. Rev. B 51, 6151 (1995).
229. B. I. Halperin and P. C. Hohenberg, Phys. Rev. 188, 898 (1969).
230. R. R. P. Singh and D. A. Huse, Phys. Rev. B 40, 7247 (1989).
231. W. Krauth, N. Trivedi, and D. Ceperley, Phys. Rev. Lett. 67, 2307 (1991).
232. A. Dorneich and M. Troyer, Phys. Rev. E 64, 066701 (2001).
233. F.-J. Jiang and U. Gerber, arXiv:0906.2539.
234. S. Wessel and S. Wenzel, Bull. Am. Phys. Soc. 55, P38.7 (2010); S. Wessel (private communica-
tion).
235. T. Senthil, L. Balents, S. Sachdev, A. Vishwanath, and M. P. A. Fisher, Phys. Rev. B 70, 144407
(2004).
236. A. B. Kuklov, M. Matsumoto, N. V. Prokof’ev, B. V. Svistunov, and M. Troyer, Phys. Rev. Lett.
101, 050405 (2008).
237. F. S. Nogueira, S. Kragset, and A. Sudbø, Phys. Rev. B 76, 220403(R) (2007).
238. R. G. Melko and R. K. Kaul, Phys. Rev. Lett. 100, 017203 (2008).
239. R. K. Kaul and R. G. Melko, Phys. Rev. B 78, 014417 (2008).
240. F.-J. Jiang, M. Nyfeler, S. Chandrasekharan, and U.-J. Wiese, J. Stat. Mech., P02009 (2008).
241. O. I. Motrunich and A. Vishwanath, Phys. Rev. B 70, 075104 (2004); arXiv:0805.1494.
242. S. Takashima, I. Ichinose, and T. Matsui, Phys. Rev. B 72, 075112 (2005).
243. D. H. Kim, P. A. Lee, and X.-G. Wen, Phys. Rev. Lett. 79, 2109 (1997).
244. M. Aizenman and B. Nachtergaele, Comm. Math. Phys. 164, 17 (1994).
245. M. Levin and T. Senthil, Phys. Rev. B 70, 220403 (2004).
246. M. A. Metlitski, M. Hermele, T. Senthil, and M. P. A. Fisher, Phys. Rev. B 78, 214418 (2008).
247. J. Lou and A. W. Sandvik, Phys. Rev. B 80, 212406 (2009).
248. K. S. D. Beach, F. Alet, M. Mambrini M, and S. Capponi, Phys. Rev. B 80, 184401 (2009).
249. M. Oshikawa, Phys. Rev. B 61, 3430 (2000).
250. J. Hove and A. Sudbø, Phys. Rev. E 68, 046107 (2003).
251. J. M. Carmona, A. Pelissetto, and E. Vicari, Phys. Rev. B 61, 15136 (2000).
252. J. Lou, A. W. Sandvik, and L. Balents, Phys. Rev. Lett. 99, 207203 (2007).
253. G. Murthy and S. Sachdev, Nucl. Phys. B 344, 557 (1990).
254. K. H. Höglund, A. W. Sandvik, and S. Sachdev, Phys. Rev. Lett. 98, 087203 (2007).
255. A. Banerjee, K. Damle, and F. Alet, arXiv:1002.1375.
256. V. N. Kotov, D. X. Yao, A. H. Castro Neto, and D. K. Campbell, Phys. Rev. B 80, 174403 (2009).
337
257. L. Isaev, G. Ortiz, and J. Dukelsky, J. Phys. Cond. Mat. 22, 016006 (2009)
258. D. Rokhsar and S. Kivelson, Phys. Rev. Lett. 61, 2376 (1988).
259. E. Fradkin, D. A. Huse, R. Moessner, V. Oganesyan and S. L. Sondhi, Phys. Rev. B 69, 224415
(2004)
260. M. A. Continentino and A. S. Ferreira, Physica A 339, 461 (2004).
261. C. Domb and M. S. Green, Phase Transitions and Critical Phenomena, Vol. 3 (Academic Press,
New York, 1974).
262. J. Oitmaa, C. Hamer, and W-H. Zheng, Series Expansion Methods for Strongly Interacting Lattice
Models (Cambridge University Press, Cambridge, 2006).
263. M. Rigol, T. Bryant, and R. R. P. Singh, Phys. Rev. E 75, 061118 (2007).
264. S. Östlund and S. Rommer, Phys. Rev. Lett. 75, 3537 (1995).
265. G. Vidal, Phys. Rev. Lett. 99, 220405 (2007).
267. R. Yu, T. Roscilde, and S. Haas, Phys. Rev. B 73, 064406 (2006).
268. T. Roscilde and S. Haas, Phys. Rev. Lett. 99, 047205 (2007).
269. O. F. Syljuasen, Phys. Rev. E 67, 046701 (2003).
270. N. Kawashima and K. Harada, J.Phys. Soc. Jpn. 73, 1379 (2004).
271. R. G. Melko and A. W. Sandvik, Phys. Rev. E 72, 026702 (2005).
272. K. Louis and C. Gros, Phys. Rev. B 70, 100410 (2004).
273. S. Liang, B. Doucot, and P. W. Anderson, Phys. Rev. Lett. 61, 365 (1988).
274. W. Marshall, Proc. Roy. Soc. A 232, 48 (1955).
275. J. Lou and A. W. Sandvik, Phys. Rev. B 76, 104432 (2007).
276. S. Liang, Phys. Rev. B 42, 6555 (1990).
277. G. Santoro et al., Phys. Rev. Lett. 83, 3065 (1999).
278. L. Wang and A. W. Sandvik, Phys. Rev. B 81, 054417 (2010).
279. A. Banerjee and K. Damle, J. Stat. Mech. (2010) P08017.
280. F. Alet, S. Capponi, N. Laflorencie, and M. Mambrini, Phys. Rev. Lett. 99, 117204 (2007).
281. R. W. Chhajlany, P. Tomczak, and A. Wójcik, Phys. Rev. Lett. 99, 167204 (2007).
282. H. Tran and N. E. Bonesteel, arXiv:0909.0038.
283. A. B. Kallin, I. González, M. B. Hastings, and R. G. Melko, Phys. Rev. Lett. 103, 117203 (2009)
338

Computational Studies

Uploaded by

Copyright:

Available Formats

Computational Studies

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computational Studies

Uploaded by

Copyright:

Available Formats

Computational Studies of Quantum Spin Systems

Citation: AIP Conf. Proc. 1297, 135 (2010); doi: 10.1063/1.3518900

Additional information on AIP Conf. Proc.

Topics. Two classes of computational methods will be discussed in these lecture

2.1. The Néel state and its quantum fluctuations

2.1.1. spin wave theory

γk = 21 [cos(kx ) + cos(ky )]. (10)

Using the Bogolubov transformation, this becomes

2.1.2. Destruction of the Néel order

Valence-bond states. The above intuitive picture of a state of fluctuating singlets

|φ i = N −1/4 |(i1, j1 )(i2 , j2) · · · (iN/2 , jN/2 )i, (20)

2.3. One-dimensional systems

Tc [gc] T [g] Tc [gc] T [g]

where the spin correlations decay exponentially with distance [60].

2.4. Models with quantum phase transitions in two dimensions

2.4.1. Dimerized systems

renormalization-group treatments of one such field theory—the nonlinear σ -model in

2.4.2. Frustrated systems

(a) (b) (c)

On the square lattice, frustration effects can be investigated by adding interactions

2.4.3. The J-Q class of models

FIGURE 8. Graphical representation of possible arrangements of products of singlet-projector oper-

sonable microscopic hamiltonian. Answering this question requires large-scale compu-

Order parameters. As an alternative to the square of the full spatially averaged

C(ri j ) = hSi · S j i, (24)

Dxx (ri j ) = hBx (ri )Bx (r j )i, (25)

where the bond operator is given by

Bx (ri ) = S(ri ) · S(ri + x̂). (26)

D∗xx (r) = Dxx (r) − 12 [Dxx (r − x̂) + Dxx (r + x̂)]. (27)

3. CLASSICAL PHASE TRANSITIONS, MONTE CARLO

3.1. Mean-field theory of the Ising model

and we assume a translationally invariant system, so that this sum is independent of i.

m = tanh[(Jsm + h)/T ]. (36)

3.2. Monte Carlo simulations of the Ising model

P(σ → σ −i ) = Pselect (i)Paccept (σ −i ), (43)

Symmetry-breaking and finite systems. To obtain a qualitative understanding of how

Autocorrelations and statistical errors. Before calculating expectation values, we

3.3. Finite-size scaling and critical exponents

C(ri j ) = hσi σ j i, (49)

3.3.1. Scaling and critical exponents

where α can be positive or negative. In mean-field theory α = 0. When α < 0, there is

|t| ∼ ξ −1/ν , (60)

Qmax (L) ∼ Lκ /ν . (62)

|tmax (L)| ∼ L−1/ν . (63)

which, using ξ ∼ |t|−1/ν , we can also write as

Q(t, L) = Lσ g(tL1/ν ). (65)

Q(t, L) = Lκ /ν g(tL1/ν ). (66)

yL = Q(t, L)L−κ /ν , xL = tL1/ν , (67)

S(q) = hσ−q σq i = ∑ e−iq·rC(r) = ∑ cos(q · r)C(r), (68)

where σq is the Fourier transform of an individual spin configuration,

We denote by Q the wave-vector of the dominant correlations—for a ferromagnet Q = 0,

Thus, for d = 1 and 3, ξa = ξ , whereas the 2D case is special, with ξa = ξ (15/16)1/2

S(q) = hRe{σq }2 i + hIm{σq }2 i. (74)

Computing the full correlation function C(r) and Fourier-transforming it post-simulation

the Binder ratio [121, 122]:

In an ordered state, U2 → 1 for N → ∞, because the magnetization distribution P(m)

properties of U2 one therefore has to define it for a general order parameter as

Corrections to scaling . As a check, it is often a good idea to include some correc-

yL = Q(t, L)L−κ /ν (1 + aL−ω )−1 , xL = tL1/ν , (78)

where ω is a subleading exponent and a is a constant. Here there is no correction to