Complex Langevin simulations of quantum systems

with the sign problem

Mitsuaki Hirasawa

Department of Particle and Nuclear Physics, School of High Energy Accelerator Science,
The Graduate University for Advanced Studies, SOKENDAI,
1-1 Oho, Tsukuba, Ibaraki 305-0801, Japan
abstract
The non-perturbative dynamics plays important roles in quantum field theories and su-
perstring theory. Monte Carlo calculation of the systems is one of the useful methods
to study non-perturbative dynamics. There are many successful works. However, the
usual Monte Carlo methods are not applicable, for example, in the Lorentzian metric
case, theories with a θ term, theories at finite density, and so on. In such cases, since
the Boltzmann weight becomes complex, it is not possible to interpret the weight as the
probability. This problem is called the sign problem. Some methods to overcome the
sign problem have been proposed. The complex Langevin method (CLM) is one of such
methods. An advantage of the method compared with others is that the numerical cost
for increasing the system size is reasonable.
This method was applied to many toy models, in which it worked well. However, the
applications to physically interesting models are not so many. In our work, we apply the
CLM to the gauge theories with a θ term and the Lorentzian type IIB matrix model.
As a validity test of the CLM for the gauge theories with a θ term, we first apply
the CLM to the 2d U(1) gauge theory on the torus with a θ term which is solvable with
finite lattice spacing and finite volume on an arbitrary manifold. We find that a naive
implementation of the method fails because of the topological nature of the θ term. In
order to circumvent this problem, we simulate the same theory on a punctured torus,
which is equivalent to the original model in the infinite volume limit for |θ| < π. Rather
surprisingly, we find that the CLM works and reproduces the exact results for a punctured
torus even at large θ, where the link variables near the puncture become very far from
being unitary.
Since the application of the CLM to the 2d U(1) case is successful, we next apply
the CLM to the 4d SU(2) case. There are interesting predictions for the phase structure
around θ = π by using ’t Hooft anomaly matching condition. Therefore our aim is to
investigate the phase structure by the first principle calculation. We apply the CLM to
the theory on a torus and find that, in the coarse lattice case, there is no freezing of
topological charge. Although the CLM works well even at large θ, we cannot see the
2π periodicity due to the finite lattice spacing effects. Therefore, we need to decrease
the lattice spacing. However, there is the freezing problem in the fine lattice case. We
try to solve the problem by imposing the periodic boundary conditions We find that
imposing the open boundary condition in all spatial directions alleviates the topology
freezing sufficiently, and the CLM works well at large θ. However, the finite volume effect
occurs as a drawback of the open boundary conditions. We also discuss the possible way
to investigate the phase structure.
The Lorentzian type IIB matrix model was proposed as a non-perturbative formulation
of superstring theory in 1996. The emergence of (3+1)-dimensional expanding space-time
in the model is an intriguing phenomenon which was observed in Monte Carlo studies of

2
this model. In this work we investigate the space-time structure of the matrices generated
by simulating this model and its simplified versions by using the hybrid Monte Carlo
method, and find that the expanding part of the space is described essentially by the
Pauli matrices. We argue that this is due to an approximation used in the simulation to
avoid the sign problem, which actually amounts to replacing eiSb by eβSb (β > 0) in the
partition function, where Sb is the bosonic part of the action.
In order to treat the weight eiSb appropriately, we use the CLM instead of the ap-
proximation to overcome the sign problem. We generalize the model by introducing two
parameters which correspond to the Wick rotation on the world sheet and that in the tar-
get space. This generalized model interpolates among the Lorentzian case, the Euclidean
case, and a model with eβSb . We apply the CLM to this generalized model and find that
the singular structure phase is not stable in the Lorentzian case. We also find that a new
phase appears in the Lorentzian case where the CLM works well although the Boltzmann
weight becomes a pure phase factor. In the bosonic model which is a simplified model of
the type IIB matrix model, we cannot see the spontaneous breaking of SO(9) symmetry
as in the case of the Euclidean bosonic model. We also discuss the possible scenario
for emerging a regular space-time with the (3+1)-dimensional expanding behavior in the
original model.

3
Contents

1 Introduction 7
1.1 Gauge theories with a θ term . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1.1 2D U(1) lattice gauge theory with a θ term . . . . . . . . . . . . . . 8
1.1.2 4D SU(2) lattice gauge theory with a θ term . . . . . . . . . . . . . 9
1.2 The Lorentzian type IIB matrix model . . . . . . . . . . . . . . . . . . . . 10

2 Sign problem 12
2.1 Review of the complex Langevin method . . . . . . . . . . . . . . . . . . . 13
2.1.1 Real Langevin method . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.2 Complex Langevin method . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Some techniques in the complex Langevin method . . . . . . . . . . . . . . 16
2.2.1 Adaptive step size . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.2 Gauge cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.3 Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 2D U(1) lattice gauge theory with a θ term 19

3.1 Lattice formulation of the 2D U(1) gauge theory with a θ term . . . . . . . 19
3.2 Applying the CLM to the 2D U(1) gauge theory . . . . . . . . . . . . . . . 21
3.2.1 The complex Langevin equation for the 2D U(1) . . . . . . . . . . . 21
3.2.2 Gauge cooling for the 2D U(1) . . . . . . . . . . . . . . . . . . . . . 22
3.2.3 Adaptive stepsize . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2.4 Results with the naive implementation . . . . . . . . . . . . . . . . 24
3.2.5 The appearance of large drifts and the topology change . . . . . . . 27
3.3 Introducing a puncture on the 2D torus . . . . . . . . . . . . . . . . . . . . 28
3.3.1 Defining the punctured model on the lattice . . . . . . . . . . . . . 29
3.3.2 Equivalence in the infinite volume limit . . . . . . . . . . . . . . . . 30
3.4 Application of the CLM to the punctured model . . . . . . . . . . . . . . . 32
3.4.1 The drift terms for the punctured model . . . . . . . . . . . . . . . 32
3.4.2 The θ dependence of the partition function . . . . . . . . . . . . . . 33
3.4.3 Validity of the CLM . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.4.4 Results for the observables . . . . . . . . . . . . . . . . . . . . . . . 39

4
4 4D SU(2) gauge theory with a θ term 42
4.1 Lattice formulation of the 4D SU(2) gauge theory . . . . . . . . . . . . . . 42
4.2 Application of the CLM to the 4D SU(2) lattice gauge theory . . . . . . . 44
4.2.1 Gauge cooling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.2 Adaptive step size . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Result for a naive implementation . . . . . . . . . . . . . . . . . . . . . . . 48
4.3.1 Autocorrelation of the topological charge . . . . . . . . . . . . . . . 48
4.3.2 Finite θ simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.4 Modifying the boundary condition . . . . . . . . . . . . . . . . . . . . . . . 50
4.4.1 Open boundary condition for one spatial direction . . . . . . . . . . 50
4.4.2 Open boundary condition for all spatial directions . . . . . . . . . . 51

5 On the emergence of the space-time structure in the type IIB matrix

model 54
5.1 Brief review of the Lorentzian type IIB matrix model . . . . . . . . . . . . 54
5.1.1 Definition of the Lorentzian type IIB matrix model . . . . . . . . . 54
5.1.2 SSB of rotational SO(9) symmetry . . . . . . . . . . . . . . . . . . 56
5.1.3 Expanding behaviors in the simplified models . . . . . . . . . . . . 58
5.2 Space-time structure of the matrix configurations . . . . . . . . . . . . . . 58
5.2.1 Results for the bosonic model . . . . . . . . . . . . . . . . . . . . . 60
5.2.2 Including fermionic contributions . . . . . . . . . . . . . . . . . . . 61
5.2.3 Taking the continuum limit . . . . . . . . . . . . . . . . . . . . . . 63
5.2.4 The Pauli-matrix structure . . . . . . . . . . . . . . . . . . . . . . . 66
5.3 The new interpretation of the simulation . . . . . . . . . . . . . . . . . . . 68
5.3.1 The “derivation” of the partition function (5.1.15) . . . . . . . . . . 68
5.3.2 Subtlety in the derivation and the new interpretation . . . . . . . . 69

6 Complex Langevin simulation of the Lorentzian type IIB matrix model 72

6.1 Generalization of the model . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.2 Complex Langevin simulation of the model . . . . . . . . . . . . . . . . . . 73
6.2.1 Treatment of the cutoffs . . . . . . . . . . . . . . . . . . . . . . . . 73
6.2.2 A way to realize the order of diagonal elements of the temporal
matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
6.2.3 Application of the CLM . . . . . . . . . . . . . . . . . . . . . . . . 75
6.2.4 Hermiticity of the extended matrices . . . . . . . . . . . . . . . . . 76
6.2.5 Adaptive step size . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.3.1 Results for the 10D bosonic model . . . . . . . . . . . . . . . . . . 77
6.3.2 Effects from the fermionic part . . . . . . . . . . . . . . . . . . . . 86

5
7 Summary and discussions 91
7.1 Gauge theories with a θ term . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.1.1 2D U(1) lattice gauge theory with a θ term . . . . . . . . . . . . . . 91
7.1.2 4D SU(2) lattice gauge theory with a θ term . . . . . . . . . . . . . 92
7.2 The Lorentzian type IIB matrix model . . . . . . . . . . . . . . . . . . . . 93
7.2.1 On the emergence of the space-time structure in the type IIB matrix
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2.2 Complex Langevin simulation of the Lorentzian type IIB matrix
model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

A Derivation of the exact result 97

A.1 The K-functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
A.2 Partition function for the non-punctured model . . . . . . . . . . . . . . . 100
A.3 Partition function for the punctured model . . . . . . . . . . . . . . . . . . 101
A.4 Evaluation of the observables . . . . . . . . . . . . . . . . . . . . . . . . . 103

B The punctured model with the sine definition Qsin 105

C The determination of the parameter p 108

6
Chapter 1

Introduction

Sign problem is the main obstacle in studying the non-perturbative dynamics of some
systems by using Monte Carlo simulations. The sign problem occurs in the Lorentzian
metric cases, theories with a θ term, theories at finite density, and so on. In such cases,
since the Boltzmann weight becomes complex, it is not possible to interpret the weight as
the probability. Therefore, usual Monte Carlo methods are not applicable. This problem
is called the sign problem.
Some methods to overcome the sign problem have been proposed. For example, the
complex Langevin method [1, 2, 3, 4, 5, 6], the tensor renormalization method [7, 8,
9, 10, 11], the generalized Lefschetz thimble method [12, 13, 14, 15, 16, 17], the path
optimization method[18, 19, 20, 21], and so on. Each method has its pros and cons. In
this thesis, we mainly focus on the complex Langevin method (CLM). An advantage of
the method compared with others is that the numerical cost for increasing the system size
is reasonable. This method was applied to many toy models, in which they worked well.
However, the applications for physically interesting models are not so many.
In this thesis, we apply the CLM to gauge theories with a θ term and the Lorentzian
type IIB matrix model.

1.1 Gauge theories with a θ term

The θ term provides an interesting avenue of research in quantum field theories. Due
to its topological nature, its effects on physics should be genuinely nonperturbative, if
present at all. In particular, it does not affect the equation of motion, which implies that
θ is a parameter that does not exist in the corresponding classical theory. For instance,
the θ term in QCD is given by Sθ = −iθQ, where Q is the topological charge defined by
1
Z
Q= µνρσ d4 x trFµν Fρσ , (1.1.1)
32π 2
which takes integer values on a compact space. The θ term breaks parity and time-reversal
symmetries, and hence the CP symmetry. This leads to a non-vanishing electric dipole

7
moment of a neutron, which is severely restricted by experiments. The current upper
bound on θ thus obtained is |θ| . 10−10 [22, 23], which is extremely small although there
is no reason for it theoretically. This is a naturalness problem known as the strong CP
problem.
A popular solution to this problem is the Peccei-Quinn mechanism [24, 25, 26, 27],
which introduces axions as a pseudo Nambu-Goldstone boson of a hypothetical global
U(1)PQ symmetry. In this mechanism, the potential for the axions induced by QCD
chooses a CP invariant vacuum automatically. Recently, gauge theories with a θ term have
attracted attention also from the viewpoint of the ’t Hooft anomaly matching condition
[28, 29, 30] and the gauge-gravity correspondence [31, 32, 33, 34]. In particular, there
is an interesting prediction for a phase transition at θ = π, which claims that either
spontaneous CP breaking or deconfinement should occur there [28, 29, 35]. In order to
investigate gauge theories from first principles in the presence of a θ term motivated either
by the physics related to axions or by the recent predictions, one needs to perform non-
perturbative calculations based on Monte Carlo methods. However, this is known to be
extremely difficult because the θ term appears as a purely imaginary term in the Euclidean
action S. The Boltzmann weight e−S becomes complex, and one cannot interpret it as
the probability distribution as one does in Monte Carlo methods.
One can still use the reweighting method by treating the phase of the complex weight
as a part of the observable. In the case at hand, this amounts to obtaining the histogram
of the topological charge at θ = 0 and taking an average over the topological sectors
characterized by the integer Q with the weight eiθQ . Various results obtained in this way
are nicely reviewed in Ref. [36]. Clearly, the calculation becomes extremely difficult due to
huge cancellations between topological sectors when topological sectors with |Q|  π/|θ|
make significant contributions to the partition function, which occurs either for |θ| ∼ π
or for smaller |θ| with sufficiently large volume.

1.1.1 2D U(1) lattice gauge theory with a θ term

In chapter 3, we apply the CLM to the 2D U(1) lattice gauge theory with a θ term on
the torus [37]. The model can be solved analytically with finite lattice spacing and finite
volume on an arbitrary manifold [38, 39, 40]. Therefore this model is a useful testing
ground for new methods [38, 41, 42, 43, 44] aiming at solving the sign problem. By using
the reweighting method [42], for instance, one can only reach θ ∼ 2.2 with a 16 × 16
lattice, and in particular, it seems almost impossible to approach θ = π by this method.
Note also that the region of θ that can be explored by this method shrinks to zero as one
increases the lattice size.
We find that a naive implementation of the CLM fails. The reason for this is that the
configurations that appear when the topology change occurs during the Langevin process
necessarily result in a large drift term. Due to this fact, the criterion [6] for the validity

8
of the method based on the histogram of the drift term cannot be satisfied. If one tries to
suppress the appearance of the problematic configurations by approaching the continuum
limit, the criterion can be satisfied, but the topology change does not occur during the
Langevin process, hence the ergodicity is lost.
In order to cure this problem, we introduce a puncture on the torus, which makes the
base manifold noncompact. We have obtained exact results for this punctured model as
well. Even in the continuum limit, the topological charge is no longer restricted to integer
values and the 2π periodicity in θ does not hold. However, if we take the infinite volume
limit with |θ| < π, one cannot distinguish the model from the original non-punctured
model as far as the observables that make sense in that limit are concerned. Note that in
that limit, the topological charge can take arbitrarily large values and therefore it does
not really matter whether it is an integer or not.
On the other hand, the situation of the complex Langevin simulation changes drasti-
cally for the punctured model. The topology change occurs freely and the appearance of
the problematic configurations can be suppressed by simply approaching the continuum
limit. Thus the criterion for the validity of the CLM is met without losing the ergodicity,
and we are able to reproduce the exact results for the punctured model.
The most striking aspect of our results is that the CLM works even if the link variables
close to the puncture become very far from being unitary. This can happen because the
direct effect of the θ term on the complex Langevin dynamics is actually concentrated
on these link variables. While the link variables are allowed to be non-unitary in the
CLM in general in order to include the effects of the complex action, all the previous
work suggested that the condition for the validity cannot be satisfied unless the non-
unitarity is sufficiently suppressed. Precisely for this reason, the gauge cooling [45, 5, 6]
was invented as a crucial technique in applying the CLM to gauge theories. In fact, we
also use the gauge cooling in our simulation, but the link variables close to the puncture
nevertheless become far from being unitary when θ or the physical volume gets large.
Yet the criterion for the drift term is not violated and the exact results are perfectly
reproduced.

1.1.2 4D SU(2) lattice gauge theory with a θ term

In chapter 4, we study the 4D SU(2) gauge theory with a θ term. In pure SU(2) gauge
theory, there is a mixed ’t Hooft anomaly between CP and the center symmetry at θ = π.
Due to the ’t Hooft anomaly matching condition, there are some constraints on the phase
structure. Therefore, some possible phase diagrams are expected depending on the nature
of the vacuum at θ = π [28].
Assuming that the theory is in the confined phase at the low temperature, the condition
indicates that the CP symmetry should be spontaneously broken in the confined phase
because the center symmetry is preserved. Therefore the following relation should be

9
satisfied
Tdec (θ = π) ≤ TCP , (1.1.2)
where Tdec (θ = π) is the critical temperature for the confinement-deconfinement phase
transition at θ = π.
The holographic Yang-Mills model with N  1 supports Tdec (θ = π) = TCP scenario
[33]. On the other hand, softly-broken N = 1 supersymmetric SU(2) Yang-Mills theory
supports Tdec (θ = π) < TCP scenario [46]. Therefore, our aim is to determine the phase
diagram by the first principle calculation. Since there is the sign problem, we apply the
CLM to the 4D SU(2) lattice gauge theory with a θ term.
First, we naively apply the CLM to the theory, and we find that, unlike the case of the
2D U(1) model, there is a region where the simulations are free from both the topology
freezing problem and the wrong convergence problem for |θ| < 2π. However, we cannot
see the 2π periodicity in this region due to the large UV fluctuations. Since we cannot
use a cooling procedure such as the gradient flow in the CLM, we need to decrease the
lattice spacing to see the 2π periodicity. We find that, in the fine lattice case, the topology
freezing problem also occurs in the 4D SU(2) case.
It is known that the open boundary conditions alleviate the topology freezing. There-
fore, we impose the boundary conditions and find that the open boundaries for all spatial
directions alleviate the problem sufficiently. The criterion is satisfied for |θ| < 2π also in
this case. However, we find that the finite volume effects occur as a drawback of the open
boundary condition. We will discuss a possible way to investigate the phase diagram.

1.2 The Lorentzian type IIB matrix model

Superstring theory is one of the promising candidates for quantum gravity. The theory is
defined on 10d instead of 4d space-time due to the consistency. One of the most amazing
things for the theory is that all of the four fundamental interactions are treated in one
quantum theory. However, there are still many open problems that must be answered.
One of the problems in superstring theory is the relation between the 4d universe and
the 10d space-time. Compactification is a procedure to obtain the 4d space-time from
the 10d space-time. However, there is an enormous number of perturbative vacua. This
circumstance is called the string landscape. It is difficult to determine the vacuum which
actually describes our universe.
The type IIB matrix model was proposed as a non-perturbative formulation of su-
perstring theory [47]. The action of the model is formally obtained by the dimensional
reduction [48] of the action for 10d N = 1 SYM theory to 0d [49]. Therefore, in this
model, the space-time does not exist a priori, and it emerges from dynamical degrees
of the matrices. There is evidence that the model is considered as one of the promis-
ing candidates for the non-perturbative formulation. One reason is that this model can

10
be interpreted as the matrix regularization of the type IIB matrix model in the Schild
gauge [50]. Another reason is that this model describes the well known interaction be-
tween D-branes. In addition, this model reproduces the light-cone string filed theory [51]
in the large N limit [52]. Therefore, this model has the potential to clarify a possible
non-perturbative mechanism for dynamical compactification in superstring theory.
The Euclidean version of the theory was investigated, and the SSB of SO(10) symmetry
was suggested by some approaches [53, 54, 55, 56, 57, 58, 59]. However, latest calculation
based on the CLM suggested that the SO(10) symmetry spontaneously breaks to SO(3)
instead of SO(4).
These results provided a strong motivation to study the Lorentzian type IIB matrix
model. In the previous works, the Monte Carlo simulation of this model was performed,
and interesting results are obtained. At some point in time, the SO(9) symmetry sponta-
neously breaks to SO(3) [60]. After the SSB, the 3d space expands exponentially in early
time [61]. And the expansion law changes from exponential to power-law as time proceeds
[62]. These results are interesting also from the viewpoint of cosmology. However, the
structure of the space has not been investigated in detail yet.
In chapter 5, we study the space-time structure in detail [63]. In order to study the
space-time structure, we investigate how the space spreads in the radial direction. We
find that the space is essentially described by the Pauli matrices. Namely, the space
is actually more like a fuzzy sphere. We call this structure the Pauli matrix structure.
We also observed the situation remains in the late time or in the continuum limit. This
structure is different from the space-time we observe. We consider that the cause of the
problem is an approximation that was used to avoid the sign problem. The approximation
corresponds to replacing eiS by eβS (β is a real positive value) in the partition function.
It is expected that, when the complex-valued weight is appropriately treated, the smooth
structure dynamically appears without losing the (3+1)d expanding behavior.
In chapter 6, we use the CLM to overcome the sign problem instead of the approxi-
mation which was used so far. We generalize the model by introducing two parameters
that correspond to the Wick rotation on the world sheet and that in the target space.
This generalized model interpolates among the Lorentzian case, the Euclidean case, and
a model with eβSb . We applied the CLM to the generalized model, and find that a new
phase appears at the Lorentzian case. In this phase, continuous space appears, however,
there is no clear expansion and the SSB either yet in this preliminary study. We also find
that the new phase is smoothly connected to a phase that appears in the Euclidean model.
We also discuss, the possible scenario for the emergence of expanding (3+1)d space-time.

11
Chapter 2

Sign problem

First of all, we consider the case of a real-valued action S(x) ∈ R where x is a set of real
variables x = (x1 , x2 , ..., xN ) ∈ RN . The partition function is given by
Z
Z = dxe−S(x) . (2.0.1)

In usual Monte Carlo case, we consider the Boltzmann weight e−S(x) as a probability and
generate configurations under the probability:

P (x) ∝ e−S(x) . (2.0.2)

The expectation values of a observable O(x) is calculated by the ensemble average as

Nconfig
1 X
hO(x)i = On (x), (2.0.3)
Nconfig n=1

where Nconfig is the number of configurations.

However, in a complex-valued action S(x) ∈ C case, since the Boltzmann weight
becomes complex, it is impossible to interpret the weight as the probability. This problem
is called “sign problem”. In principle, we can calculate the expectation value of the
observable O(x) by using the reweighting method,

hO(x)e−i Im(S(x)) ie−Re(S(x))

hO(x)i = (2.0.4)
he−i Im(S(x)) ie−Re(S(x))

where the brackets of the right-hand side are expectation value evaluated by using config-
urations which are obtained under the weight e−Re(S(x)) . This method is useful when the
phase of the weight does not highly oscillate. However, when the phase highly oscillates
the denominator and the numerator become very small, and we need to obtain the finite
value from two very small values. In order to do that, we need a very large number of
configurations. Thus the reweighting method is not useful in that situation.

12
2.1 Review of the complex Langevin method
The real Langevin method was introduced by Parisi and Wu in [64] as a stochastic quan-
tization approach to quantum field theories. This method widely used in particle physics.
However, in general, we cannot use this method for systems with the sign problem.
In 1983, Klauder and Parisi proposed the complex Langevin method independently as a
method to overcome the sign problem [1, 2]. After the proposals, the method applied many
systems, and a problem that the method sometimes gives wrong results was reported. This
problem is called as the wrong convergence problem. The mechanism of this problem was
not understood completely. Due to this situation, results obtained by this method was
not reliable and the application of the method was limited.
Recently, the conditions for the correct convergence was studied in [3], and a practical
criterion for correct convergence was proposed in [6]. After the breakthrough, the method
applied many systems.

2.1.1 Real Langevin method

First, we review the real Langevin method. Therefore, we consider a system with a real-
valued action S(x) ∈ R. In this method, the configuration is generated by the Langevin
equation as
∂S(x) p
xi (t + ∆t) = xi (t) − ∆t + (∆t)ηi (t), (2.1.1)
∂x
where t is a fictitious time so-called the Langevin time, ∆t is the Langevin step size, and
ηi (t) is the Gaussian noise. The noise satisfies

hηi (s)ηj (t)iη = 2δi,j δs,t , (2.1.2)

where the expectation value h· · ·iη is defined by

− 41 ∆t k η 2 (tk )
RQ P
k dη(tk ) · · · e
h· · ·iη = R Q − 14 ∆t k η 2 (tk )
P . (2.1.3)
k dη(tk )e

Here we consider an arbitrary f (x) and its expectation value

Z


f x(t) η = dxf (x)P (x; t), (2.1.4)

where P (x; t) is the probability distribution of x(t) defined by

* +
Y 
P (x; t) = δ xi − xi (t) . (2.1.5)
i η

13
If we consider the evolution of f (x), one obtains

1 ∂ 2f √
   
∂f ∂S
( ∆t) ηi (t)ηj (t) + O ∆t2
2




f x(t + ∆t) η − f x(t) η = −∆t +
∂xi ∂xi 2 ∂xi ∂xj η
2
 
∂f ∂S ∂ f
Z
+ 2 P (x; t) + O ∆t2


= ∆t dx −
∂xk ∂xk ∂xk
 
∂ ∂S ∂
Z
P (x; t) + O ∆t2 .


= ∆t dxf (x) +
∂xk ∂xk ∂xk
(2.1.6)

Using (2.1.4), this quantity should be rewritten as

Z  



f x(t + ∆t) η − f x(t) η = dxf (x) P (x; t + ∆t) − P (x; t) . (2.1.7)

Comparing (2.1.6) and (2.1.7), one obtains

 
∂ ∂S ∂
P (x; t) + O ∆t2 .


P (x; t + ∆t) − P (x; t) = ∆t + (2.1.8)
∂xk ∂xk ∂xk

(2.1.8) is the discretized version of the Fokker-Planck equation 1 . After the thermalization,
the probability distribution is independent from t, therefore, the distribution is obtained
by solving  
∂ ∂S ∂
P (x) + O ∆t2 = 0,


∆t + (2.1.9)
∂xk ∂xk ∂xk
where P (x) is the time independent probability distribution. The solution is

P (x) = e−S(x)+O(∆t) . (2.1.10)

Therefore, in the Langevin process, the configurations are generated under this probabil-
ity.

2.1.2 Complex Langevin method

Next we review the complex Langevin method. In the CLM, a real variable xi is promoted
to a complex variable zi ∈ C:
xi → zi = xi + iyi (2.1.11)
This variable is updated by the complex Langevin equation:
∂S(z) p
zi (t + ∆t) = zi (t) − ∆t + (∆t)ηi (t), (2.1.12)
∂zi
1
The Fokker-Plank equation
 
∂P (x; t) ∂ ∂S ∂
= + P (x; t)
∂t ∂xk ∂xk ∂xk

is obtained by taking ∆t → 0 limit of (2.1.8).

14
where t is a fictitious time so-called the Langevin time, ∆t is the Langevin step size, and
ηi (t) is the Gaussian noise. ηi (t) is construct from two independent real-valued Gaussian
noise as
√ (R) √ (I)
ηi (t) = cR ηi (t) + i cI ηi (t), (2.1.13)
where ηi,R (t) and ηi,I (t) satisfy
(R) (R)
hηi (s)ηj (t)iη = 2δi,j δs,t ,
(I) (I)
hηi (s)ηj (t)iη = 2δi,j δs,t , (2.1.14)
(R) (I)
hηi (s)ηj (t)iη = 0 ,

where the expectation value h· · ·iη is defined by

− 41 ∆t k {η (R) (tk )2 +η (I) (tk )2 }
RQ P
(R) (I)
k dη (tk )dη (tk ) · · · e
h· · ·iη = R Q 1 . (2.1.15)
(R) (t )dη (I) (t )e− 4 ∆t k {η (R) (tk )2 +η (I) (tk )2 }
P
k dη k k
√ √
The coefficients cR and cI in (2.1.13) should be satisfy cI ≥ 0 and cR − cI = 1.
The time evolution of the expectation value of a holomorphic observable O(x + iy) is
given by
Z


O x(t + ∆t) + iy(t + ∆t) η = dxdyO∆t (x + iy)P (x, y; t), (2.1.16)

where O∆t (x + iy) is defined as

1
Z
1 (R) 2 (I) 2 √
O∆t (x + iy) = dη (R) dη (I) e− 4 {(η ) +(η ) } O(z + ∆tv(z) + ∆tη), (2.1.17)
2π
where v(z) is the drift term defined by
∂S(z)
v(z) = . (2.1.18)
∂z
Note that we assume v(z) is a holomorphic quantity. We expand (2.1.17) with respect to
∆t and we perform the integration for η, then we obtain
∞
X 1
O∆t (z) = (∆t)n : Ln : O(z), (2.1.19)
n=0
n!

where L is defined by
   
∂ ∂ ∂ ∂
L = Re vi (z) + cR + Im vi (z) + cI , (2.1.20)
∂xi ∂xi ∂yi ∂yi
and : · · · : means that the operators are ordered in such a way that derivative operators
appear on the right. Since the observable is holomorphic quantity, we obtain
 
∂ ∂O
LO(z) = vi (z) + (cR − cI )
∂zi ∂zi (2.1.21)
= L̃O(z)

15
 Using (2.1.19) and (2.1.21) in (2.1.16), we obtain
∞
1
X Z  
n n


O x(t + ∆t) + iy(t + ∆t) η
= (∆t) dxdy : L̃ : O(z) P (x, y; t). (2.1.22)
n 0
n!

If (2.1.22) is valid, we can neglect higher order terms for sufficiently small ∆t as


O x(t + ∆t) + iy(t + ∆t) η
Z   (2.1.23)
= O x(t) + iy(t) η + ∆t dxdy L̃O(z) P (x, y; t) + O (∆t)2 .

We consider the condition for the validity of the expansion (2.1.22) to have a finite
convergence radius. Here we consider the magnitude of the drift term defined as

u(z) = max |vi (z)| (2.1.24)

i

which has the largest contribution to the radius. The integral in (2.1.22) involves
Z Z ∞
n
dxdyu(z) P (x, y; t) = duun p(u, t), (2.1.25)
0

where the probability distribution of the magnitude of the drift term is defined as
Z
p(u, t) = dxdyδ(u(z) − u)P (x, y; t). (2.1.26)

If the p(u, t) is suppressed faster than exponential, the expansion (2.1.22) has a finite
convergence radius. Therefore, it is needed for the correct convergence that p(u, t) falls
off faster than exponential in the CLM.
√
It is better to set cI = 0, because the fluctuation for the imaginary direction some-
times causes a large excursion from real-value, that is one of the reasons for the wrong
convergence [3, 4, 65]. Therefore, the rest of this thesis, we set cR = 1 and cI = 0.

2.2 Some techniques in the complex Langevin method

There are two main reasons for the failure of the CLM. One is that when the variables
have a large excursion in the imaginary direction, the CLM sometimes fails. This problem
is called the excursion problem. The other is that when the drift term has singularities
if the complexifyed variables approach the singularities frequently the CLM fails. This
problem is called the singular drift problem.
There are some techniques to avoid these problems. Here we explain the adaptive step
size algorithm, the gauge cooling technique, and the deformation technique.

16
2.2.1 Adaptive step size
The large drift term is a trigger for that the variables to go away from the real value. If
the large drift terms appear frequently the CLM fails. Choosing a small Langevin step
size is one of the ways to avoid the excursion problem, however, this small Langevin step
makes the simulation inefficiency. The adaptive step size algorithm [66] is a method to
avoid the excursion problem efficiently.
In the adaptive step size algorithm, the step size is decreased when the large drift term
appears. Therefore, we choose the step size as
(
∆t0 for u < v0 ,
∆t = v0 (2.2.1)
∆t0 otherwise ,
u
where ∆t0 is the default stepsize, u is the magnitude of drift term, and v0 is the threshold
for the magnitude of drift term.

2.2.2 Gauge cooling

Instead of the real variable x, here we consider a dynamical variable U which belongs to
a gauge group G, namely U ∈ G. In the CLM, we “complexify” the variable U ∈ G to
U ∈ H because the drift term belongs to H when the system has the sign problem. For
example, if G is SU(3) then H becomes SL(3, C). When U has a large excursion from the
original group G large drift term may appear frequently, as the result, the CLM fails.
The idea of gauge cooling [45] is to reduce the deviation of the complexifyed variables
from the original gauge group as much as possible by making gauge transformations
corresponding to the complexified group H after each Langevin step. The details of the
technique are shown in 3.2.2 and 4.2.1.
It is proofed that this procedure does not affect the holomorphic observables. More-
over, this procedure can be added without affecting the argument for justifying the CLM
as demonstrated explicitly in Refs. [5, 6]. Recently, the mechanism of the gauge cooling
for stabilizing the complex Langevin simulation has been investigated [67].

2.2.3 Deformation
When the theory has fermion Tr log M is included in the effective action which is obtained
after the path integration for fermionic variables. In the Langevin simulation, we need to
evaluate the drift term for Tr log M which is Tr (M −1 ∂M
∂x
) where x is a dynamical variable.
Therefore, if the Dirac operator M has near-zero eigenvalues the drift term becomes large.
If such drift terms appear frequently the CLM fails, that problem is called the singular
drift problem.
In order to avoid the near-zero eigenvalues, we introduce the additional term like a
fictitious mass term to the action of the system. This additional term is called the defor-

17
mation term. Since this term modifies the theory, extrapolation for vanishing deformation
term. The details of the algorithm are shown in 6.3.2.

18
Chapter 3

2D U(1) lattice gauge theory with a

θ term

3.1 Lattice formulation of the 2D U(1) gauge theory

with a θ term
In this section, we review pure 2D U(1) gauge theory with a θ term and discuss how to
define it on a lattice.
In the continuum 2D U(1) gauge theory on a Euclidean space, the action for the gauge
field Aµ (x) (µ = 1, 2) is given by

1
Z
Sg = 2 d2 x (Fµν )2 , (3.1.1)
4g
where g is the gauge coupling constant and Fµν is the field strength defined as

Fµν = ∂µ Aν (x) − ∂ν Aµ (x) . (3.1.2)

We add a θ term
Sθ = −iθQ (3.1.3)
in the action, where Q is the topological charge defined by
1
Z
Q= d2 x µν Fµν , (3.1.4)
4π
which takes integer values if the space is compact.
We put this theory on a 2D torus, which is discretized into an L × L periodic lattice
with the lattice spacing a. On the lattice, we define the link variables Un,µ ∈ U(1), where
n = (n1 , n2 ) labels the lattice site as xµ = anµ . We also define the plaquette

−1
Pn = Un,1̂ Un+1̂,2 Un+ U −1 ,
2̂,1 n,2
(3.1.5)

19
where µ̂ is a unit vector in the µ direction. This plaquette is invariant under the gauge
transformation:
−1
Un,µ̂ 7→ gn Un,µ̂ gn+µ̂ . (3.1.6)
−1 †
Here we write Un,µ instead of Un,µ , which will be important later in applying the CLM,
where we complexify the dynamical variables respecting holomorphicity.
The lattice counterpart of the field strength (3.1.2) can be defined as
1
Fn,12 = log Pn , (3.1.7)
ia2
where we take the principal value for the complex log; namely log z = log |z| + i arg z with
−π < arg z ≤ π. Since the plaquette can then be written in terms of Fn,µν as
2F
Pn = eia n,12
, (3.1.8)

the lattice counterpart of the gauge action (3.1.1) can be defined as

βX X
Pn + Pn−1 = −β cos a2 Fn,12 ,



Sg = − (3.1.9)
2 n n

which approaches
1 X 2
Sg ' 2
a (Fn,µν )2 , (3.1.10)
4g n
in the continuum limit up to an irrelevant constant with the identification
1
β= . (3.1.11)
(ga)2

In the present 2D U(1) theory, the topological charge can be defined as

1 X 2 µν
Qlog = a  Fn,µν
4π n
1 X 2
= a Fn,12 (3.1.12)
2π n
i X
=− log Pn ,
2π n

which gives an integer value even at finite a. This can be proved easily by noting that
Q
n Pn = 1 since each link variable appears twice in this product with opposite directions.
We call this definition (3.1.12) the “log definition”. As an alternative definition, we
consider
i X 1 X
Pn − Pn−1 = sin a2 Fn,12 ,



Qsin = − (3.1.13)
4π n 2π n

which approaches (3.1.4) in the continuum limit recalling (3.1.8). Note, however, that the
topological charge defined on the lattice in this way can take non-integer values in general

20
before taking the continuum limit. We call this definition (3.1.13) the “sine definition”.
Thus the lattice theory is given by

S = Sg + Sθ , (3.1.14)

where Sg is given by (3.1.9) and Sθ is given by (3.1.3) with Q defined either by (3.1.12)
or by (3.1.13).
Since this theory is superrenormalizable, we can take the continuum limit a → 0 with
fixed g, which is set to unity throughout this work without loss of generality. In this unit,
the physical volume of the 2D torus is given by
L2
Vphys = (La)2 = , (3.1.15)
β
where we have used (3.1.11).

3.2 Applying the CLM to the 2D U(1) gauge theory

3.2.1 The complex Langevin equation for the 2D U(1)
The first step of the CLM is to complexify the dynamical variables. In the present case
of U(1) gauge theory, we extend the link variables Un,µ ∈ U(1) to Un,µ ∈ C \ {0}, which
corresponds to extending the gauge field Aµ (x) ∈ R to Aµ (x) ∈ C in the continuum
theory. Then we consider a fictitious time evolution Un,µ (t) of the link variables governed
by the complex Langevin equation
hn √ oi
Un,µ (t + ∆t) = Un,µ (t) exp i − ∆t Dn,µ S + ∆t ηn,µ (t) , (3.2.1)

where ηn,µ (t) is a real Gaussian noise normalized by hηn,µ (s)ηk,ν (t)i = 2δn,k δµ,ν δs,t . The
term Dn,µ S is the drift term defined by

S(ei Un,µ ) − S(Un,µ )

Dn,µ S = lim , (3.2.2)
→0 
first for the unitary link variables Un,µ (t), and then it is defined for the complexified link
variables Un,µ (t) by analytic continuation in order to respect holomorphicity. Using the
action (3.1.14), we obtain Dn,µ S = Dn,µ Sg + Dn,µ Sθ , where the first term is given as
β −1
Dn,1 Sg = − i (Pn − Pn−1 − Pn−2̂ + Pn− ),
2 2̂
β −1
Dn,2 Sg = − i (−Pn + Pn−1 + Pn−1̂ − Pn− ). (3.2.3)
2 1̂

The second term Dn,µ Sθ depends on the definition of the topological charge. If one uses
the log definition (3.1.12), Eq. (3.2.2) for the θ term becomes a δ-function, which vanishes
identically except for configurations with Pn = −1 for some n, reflecting the topological

21
nature of the definition. Such configurations are precisely the ones that appear when the
topology change occurs within the configuration space of Un,µ . It is not straightforward
to extend such a term to a holomorphic function of Un,µ .
On the other hand, if one uses the sine definition (3.1.13), the drift term becomes

θ −1
Dn,1 Sθ = − i (Pn + Pn−1 − Pn−2̂ − Pn− ),
4π 2̂
θ −1
Dn,2 Sθ = − i (−Pn − Pn−1 + Pn−1̂ + Pn− ), (3.2.4)
4π 1̂

which may be viewed as an approximation of the δ-function mentioned above. Moreover,

it can be readily extended to a holomorphic function of Un,µ . For this reason, we use the
sine definition for the non-punctured model.
The criterion [6] for the validity of the CLM states that the histogram of the drift
term should fall off exponentially or faster. There are two cases in which this criterion
cannot be met. The first case occurs when the configuration comes close to the poles
of the drift terms (3.2.3), (3.2.4), which correspond to configurations with Pn = 0 for
some n. If this happens during the Langevin process, there is a possibility of violating
the criterion. This problem is called the singular-drift problem [68, 69], which was found
first in simple models [70, 71]. In the present model, the same problem is caused also by
approaching configurations with |Pn | = ∞ for some n, which are related to the poles by
the parity transformation.
The second case occurs when the dynamical variables make large excursions in the
imaginary directions [3]. This problem is called the excursion problem. In the present
model, this corresponds to the situation in which the link variables have absolute values
|Un,µ | far from unity.
Both the singular-drift problem and the excursion problem can occur because the
link variables Un,µ are not restricted to be unitary in the CLM. In order to avoid these
problems, it is important to perform the gauge cooling, which we explain in the next
section.

3.2.2 Gauge cooling for the 2D U(1)

The idea of gauge cooling [45] is to reduce the non-unitarity of link variables as much as
possible by making gauge transformations corresponding to the complexified Lie group
after each step (3.2.1) of the Langevin process.
The deviation of the link variables from U(1) can be defined by the unitarity norm
1 Xn ∗ ∗ −1
o
N = U U
n,µ n,µ + (U U
n,µ n,µ ) − 2 . (3.2.5)
2L2 n,µ

The gauge cooling reduces this quantity by a complexified gauge transformation, which
is determined as follows.

22
 First we consider an infinitesimal gauge transformation

δUn,µ = (n − n+µ̂ ) Un,µ , (3.2.6)

where n ∈ R. The change of the unitarity norm due to the transformation is given by
1 Xn ∗ ∗ −1
o
δN = 2( n − n+µ̂ )U n,µ U n,µ − 2(n − n+µ̂ )(U n,µ U n,µ )
2L2 n,µ
1 X
= 2n Gn , (3.2.7)
2L2 n

where Gn is defined as
Xn o
∗ ∗ ∗ −1 ∗ −1
Gn = Un,µ Un,µ − Un−µ̂,µ Un−µ̂,µ − (Un,µ Un,µ ) + (Un−µ̂,µ Un−µ̂,µ ) . (3.2.8)
µ

Therefore, we find that the unitarity norm is reduced most efficiently by choosing n ∝
−Gn .
Using this result, we consider a finite gauge transformation
−1
Un,µ 7→ gn Un,µ gn+µ̂ ; gn = e−αGn , (3.2.9)

which makes the unitarity norm

1 Xn ∗ −2α(Gn −Gn+µ̂ ) ∗ −1 2α(Gn −Gn+µ̂ )
o
N (α) = U U
n,µ n,µ e + (U U
n,µ n,µ ) e − 2 , (3.2.10)
2L2 n,µ

depending on α in (3.2.9). We search for an optimal α that minimizes (3.2.10). Note here
that it is typically a small number since the gauge cooling is performed after each step
of the Langevin process. We therefore expand Eq. (3.2.10) with respect to α up to the
second order and obtain the value of α that minimizes it as
P 2
1 n Gn
α= P ∗ U ∗ −1
. (3.2.11)
2 n,µ [(Gn − Gn+µ̂ ) {Un,µ
2
n,µ + (Un,µ Un,µ ) }]

We repeat this procedure until the unitarity norm changes by a fraction less than 10−5 .

3.2.3 Adaptive stepsize

When we solve the complex Langevin equation in its discretized version (3.2.1), it occa-
sionally happens that the drift term becomes extremely large, in particular during the
thermalization process. This causes a large discretization error, which either makes the
thermalization slow or destabilizes the simulation. We can avoid this problem by using a
small stepsize ∆t, but the computational cost for a fixed Langevin time increases propor-
tionally to (∆t)−1 and the calculation becomes easily unfeasible. The adaptive stepsize
[66] is a useful technique, which amounts to reducing the stepsize only when the drift
term becomes large.

23
 (β, L) = (12, 20)
(β, L) = (3, 10)
0
10 (β, L) = (12, 20)
10

8
10-2
6

-4
10 4

2
10-6
0
1 10 100 -3 -2 -1 0 1 2 3
u Re Qsin

Figure 3.1: The results obtained by the CLM for the non-punctured model using the sine
definition Qsin of the topological charge. (Left) The histogram of the magnitude u of the
drift term defined by (3.2.12) is shown for (β, L) = (3, 10), (12, 20) with θ = π. (Right)
The histogram of Re Qsin is shown for (β, L) = (12, 20) with θ = π. The exact result
obtained for (β, L) = (12, 20) with θ = 0 is shown by the solid line for comparison.

In our simulation, we measure the magnitude of the drift term defined as

u = max |Dn,µ S| (3.2.12)

n,µ

at each step, and choose the Langevin stepsize ∆t in (3.2.1) as

(
∆t0 for u < v0 ,
∆t = v0 (3.2.13)
∆t0 otherwise ,
u
where ∆t0 is the default stepsize, and v0 is the threshold for the magnitude of drift term.
In the present work, the default stepsize is set to ∆t0 = 10−5 , and the threshold is set
to v0 = 2β, considering a bound u ≤ 2β for θ = 0, where the CLM reduces to the real
Langevin method. The measurement of the observables should be made with the same
interval in terms of the Langevin time but not in terms of the number of steps.

3.2.4 Results with the naive implementation

In this section, we present our results obtained by the CLM, which is implemented naively
using the non-punctured model explained above as opposed to the punctured model, which
we use later. As for the definition of the topological charge, we adopt the sine definition
(3.1.13) for the reason given in Section 3.2.1.
We have performed simulations at various θ for (β, L) = (3, 10), (12, 20) corresponding
to a fixed physical volume Vphys ≡ L2 /β = 102 /3. Below we show our results only for
θ = π, where the sign problem becomes severest, but the situation is the same for all
values of θ.

24
 In Fig. 3.1(Left), we show the histogram of the magnitude u of the drift term. The
distribution falls off rapidly for (β, L) = (12, 20), but it decays slowly with a power law for
(β, L) = (3, 10). Thus the criterion for correct convergence is satisfied for (β, L) = (12, 20)
but not for (β, L) = (3, 10) due to the large drifts.
In Fig. 3.1(Right), we plot the histogram of Re Qsin obtained by the CLM for (β, L) =
(12, 20) with θ = π, which has a sharp peak at Re Qsin ∼ 0. In the same figure, we also
plot the exact result for (β, L) = (12, 20) with θ = 0 for comparison, which exhibits a
few sharp peaks at integer values within the range −2 . Re Qsin . 2. From these two
plots, we conclude that the transitions between different topological sectors are highly
suppressed in the simulation, which causes a problem with the ergodicity.
This occurs also at θ = 0 for large β, and it is called the “topology freezing problem” in
the literature. In fact, the results one obtains by simulations suffering from this problem
correspond to the expectation values restricted to the topological sector specified by the
initial configuration. This is true for both θ = 0 and θ 6= 0. In this case, however, the
effect of the θ term cancels between the numerator and the denominator of the expectation
values, and the calculation essentially reduces to that of the real Langevin method at
θ = 0.
For (β, L) = (3, 10) with θ = π, on the other hand, the histogram of Re Qsin obtained
by the CLM has broad peaks that overlap with each other, which looks similar to the
exact result for (β, L) = (3, 10) with θ = 0. This implies that the topology freezing
problem does not occur for (β, L) = (3, 10). See also Fig. 3.3.
Below we define the observables we investigate. First, we define the average plaquette
by
1 ∂
w= log Z . (3.2.14)
V ∂β
Hereafter, V denotes the number of plaquettes in the action, which is V = L2 for the
non-punctured model and V = L2 − 1 for the punctured model we define in Section 3.3.1.
The topological charge density is defined by
1 1 ∂
hQi = −i log Z , (3.2.15)
V V ∂θ
which is zero at θ = 0 and purely imaginary for θ 6= 0. Finally, the topological suscepti-
bility is defined by
1 1 ∂2
hQ2 i − hQi2 = −


χ= log Z , (3.2.16)
V V ∂θ2
which is real for all θ. In fact, the topological susceptibility χ is related to the topological
charge density (3.2.15) through
1 ∂
χ = −i hQi . (3.2.17)
V ∂θ
Note, however, that this relation can be violated if the CLM fails to calculate the expec-
tation values correctly.

25
 (β, L) = (3, 10) (β, L) = (12, 20)
0.9584

0.84

0.9580
0.83
w

w
0.82
0.9576

0.81
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.006
0.02

0.003
Im 〈 Qsin 〉 / V

Im 〈 Qsin 〉 / V

0.01

0.00 0.000

-0.01
-0.003

-0.02
-0.006
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.02 0.004

0.000
0.00

-0.004
-0.02
χ

-0.008
-0.04
-0.012

-0.06
-0.016
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

Figure 3.2: The results for various observables obtained by the CLM for the non-punctured
model with the sine definition Qsin . The average plaquette (Top), the imaginary part of the
topological charge density (Middle), the topological susceptibility (Bottom) are plotted
against θ for (β, L) = (3, 10) (Left) and (12, 20) (Right). The exact results for the same
(β, L) are shown by the dashed lines for comparison.

26
 In Fig. 3.2, we show the results obtained by the CLM for the non-punctured model.
We also plot the exact results for comparison, which are derived in Appendix A.4. In the
left column, we present our results for (β, L) = (3, 10), which suffer from the incorrect
convergence, whereas in the right column, we present our results for (β, L) = (12, 20),
which suffer from the topology freezing problem. In either case, our results do not repro-
duce the exact results as anticipated. Note that our results at θ = 0 agree with the exact
results for (β, L) = (3, 10) but not for (β, L) = (12, 20). This is because the topology
freezing problem occurs for large β even at θ = 0, where the sign problem is absent.
Thus we find that the CLM with the naive implementation fails for both (β, L) =
(3, 10) and (β, L) = (12, 20) for different reasons. For (β, L) = (3, 10), the topology
change occurs but the criterion for correct convergence is not satisfied due to the large
drifts. For (β, L) = (12, 20), the criterion for correct convergence is satisfied, but the
ergodicity is violated due to the topology freezing problem. We have searched for a
parameter region in which neither of the problems occur, but we could not find one. In
fact, we will see in the next section that these problems are related to each other at least
in the present model.

3.2.5 The appearance of large drifts and the topology change

In this section, we provide more in-depth discussions on the relationship between the
appearance of large drifts and the topology change in the non-punctured model. Let us
first recall that the drift terms are given by (3.2.3) and (3.2.4), which depend on Pn .
When β is large, the gauge action Sg favors configurations with Pn ∼ 1 for all n, which
implies that the drift terms are small.
On the other hand, the notion of topological sectors can be defined by the real part of
(3.1.12), which takes integer values, even for complexified configurations that are gener-
ated in the CLM. In order for a transition between different topological sectors to occur,
one of the plaquettes has to cross the branch cut; namely the phase of the plaquette has
to jump from −π to π or vice versa. When this occurs, large drift terms can appear as
can be seen from Fig. 3.3, where we plot the histories of Re Qlog and the magnitude of
the drift term (3.2.12). We observe clear correlation between the large drift term and the
topology change. We have also confirmed that the large drift term appears for the link
variables composing the plaquette that crosses the branch cut.
In order to understand this observation better, we focus on a particular link variable
Uk,1 , and consider the corresponding drift term, which depends on the plaquettes Pk and
Pk−2̂ sharing the link. For simplicity, we set Pk−2̂ = 1 and consider the drift term v as a
function of Pk
θ
v = β sin φ − i (cos φ − 1) , (3.2.18)
2π
where we have defined a complex parameter φ by φ = −i log Pk . A large drift appears

27
 2

Re Qlog
0

-1

-2
100
u

10

0 5 10 15 20 25
t

Figure 3.3: The results obtained by the CLM for the non-punctured model with the sine
definition Qsin for (β, L) = (3, 10) with θ = π. The upper plot shows the history of the
topological charge Qlog with the log definition, whereas the lower plot shows the history
of the magnitude u of the drift term in the log scale.

when |Im φ| → ∞. In Fig. 3.4(Left), we plot the drift term as a flow diagram for β = θ = 1.
Considering that the contribution of the drift term v to the change of φ at a Langevin
step is given by ∆φ = −v∆t, we actually plot (−v) in the complex φ plane.
In what follows we assume that β > θ/2π. Then we find from Eq. (3.2.18) that
there are two fixed points corresponding to v = 0. One is φ = 0 and the other is
φ = i log[(θ/2π + β)/(θ/2π − β)], which is close to ±π for β  θ/2π. As one can see from
Fig. 3.4(Left), the fixed point φ = 0 is attractive, which confirms that Pk tends to become
unity when β is large. The other fixed point φ ∼ ±π is repulsive, and the magnitude |v|
grows exponentially as one flows away in the imaginary direction; See Fig.3.4(Right). As
we mentioned above, when the transition between topological sectors occurs, one of the
plaquettes crosses the branch cut, which corresponds to Re φ = ±π in the flow diagram.
When this happens, the configuration can flow in the imaginary direction, which causes
a large drift.

3.3 Introducing a puncture on the 2D torus

Since the problem we encounter in the previous section occurs due to the topological
nature of the θ term, a simple remedy would be to change the topology of the base
manifold to a noncompact one. Here we consider introducing a puncture on the 2D
torus. Once we introduce a puncture, the drift term Dn,µ Sθ with the log definition of
the topological charge has nonzero contributions for the link variables surrounding the
puncture, which enable us to include the effect of the θ term correctly in the CLM as

28
 4

100
2

80
Imϕ

0
60

| v (φ) |
-2 40

20
-4

0
-π - π2 0 π
π -4 -2 0 2 4
2

Reϕ Im φ

Figure 3.4: (Left) A flow diagram representing −v defined by (3.2.18) is shown as a

function of φ for β = θ = 1. (Right) The absolute value |v(φ)| is plotted against Im φ for
Re φ = π.

we will see in Section 3.4. Therefore, for the rest of this study, we basically use the log
definition to simplify our discussions. Unlike the non-punctured model, the topological
charge is no more restricted to integer values, and it can be changed freely.
Since the puncture affects the theory only locally, its effect is expected to die out in
the infinite volume limit for |θ| < π as we demonstrate explicitly in this section using
the exact results. Thus unless we are interested in a theory with a finite volume, the
punctured model is as good as the original model, the difference simply being a different
choice of “boundary conditions”. In fact, we will see that the non-punctured model has
slow convergence to the infinite volume limit for θ ∼ π, which is not the case in the
punctured model.

3.3.1 Defining the punctured model on the lattice

There are various ways to introduce a puncture on the periodic lattice. Here we consider
removing a plaquette as a simple choice. More precisely, we define the punctured model
by removing one plaquette, let say PK , from the sum appearing in the gauge action (3.1.9)
and the topological charge (3.1.12) when we define the action (3.1.14).
As an alternative method, we have also tried introducing a slit at a particular link,
which amounts to duplicating the corresponding link variable and including each of them
in the plaquettes that share the link. The results turn out to be qualitatively the same
as the ones obtained by removing a plaquette. There are, of course, many others, but in
any case, one can obtain exact results for a finite lattice as we explain in Appendix A,
and using them, one can demonstrate explicitly that the punctured model is equivalent
to the original non-punctured model in the infinite volume limit for |θ| < π.

29
3.3.2 Equivalence in the infinite volume limit
In this section, we show the equivalence of the non-punctured model and the punctured
model in the infinite volume limit. Here we use the log definition of the topological charge,
but a similar statement holds as far as the same definition is used for the two models.1
The partition function for the non-punctured model is given by (See Appendix A.2
for derivation.)
+∞
X
Znonpunc = [I(n, θ, β)]V (3.3.1)
n=−∞

for finite V = L2 , where the function I(n, θ, β) is defined by

Z π
1
dφ eβ cos φ+i( 2π −n)φ .
θ
I(n, θ, β) = (3.3.2)
2π −π

Let us take the infinite volume limit V → ∞, in which the sum over n in (3.3.1) is
dominated by the term that gives the largest absolute value |I(n, θ, β)|. This corresponds
θ
to the n that minimizes | 2π − n|. Thus in the infinite volume limit, the free energy is
obtained as
1
lim log Znonpunc = log I(0, θ̃, β) , (3.3.3)
V →∞ V

where θ̃ is defined by θ̃ = θ − 2πk with the integer k chosen so that −π < θ̃ ≤ π.

On the other hand, the partition function for the punctured model is given by (See
Appendix A.3 for derivation.)

Zpunc = [I (0, θ, β)]V (3.3.4)

for finite V = L2 − 1, which implies that the free energy

1
log Zpunc = log [I (0, θ, β)] (3.3.5)
V
is actually V independent. Hence all the observables that can be derived from it has
no finite size effects. Note also that this model does not have the 2π periodicity in θ.
By comparing (3.3.3) and (3.3.5), one can see that the two models are equivalent in the
infinite volume limit for |θ| < π.
The observables defined in Section 3.2.4 can be calculated for the two models us-
ing (3.3.1) and (3.3.4) by numerical integration (See Appendix A.4 for the details.). In
Fig. 3.5, we plot the average plaquette (Top) defined by (3.2.14), the imaginary part of the
topological charge density (Middle) defined by (3.2.15) and the topological susceptibility
(Bottom) defined by (3.2.16) for L = 10 (Left) and L = 20 (Right), respectively, with the
same β = 12. The results for the two models tend to agree as L increases for |θ| < π.
1
In the case of the sine definition, the equivalence of the two models in the infinite volume limit holds
for |θ| < θc (β), where θc (β) ∼ π{1 + 1/(2β)} for large β.

30
 (β, L) = (12, 10) (β, L) = (12, 20)
nonpunctured L = 10 nonpunctured L = 20
0.961 nonpunctured L → ∞ 0.961 nonpunctured L → ∞
punctured punctured

0.960 0.960
w

w
0.959 0.959

0.958 0.958

0.957 0.957
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

(β, L) = (12, 10) (β, L) = (12, 20)

0.015 0.015
nonpunctured L = 10 nonpunctured L = 20
nonpunctured L → ∞ nonpunctured L → ∞
punctured punctured
0.010 0.010
Im 〈 Qlog 〉 / V

Im 〈 Qlog 〉 / V

0.005 0.005

0.000 0.000

-0.005 -0.005

-0.010 -0.010
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

(β, L) = (12, 10) (β, L) = (12, 20)

0.004 0.004

0.000 0.000

-0.004 -0.004
χ

-0.008 -0.008

-0.012 -0.012

nonpunctured L = 10 nonpunctured L = 20
-0.016 nonpunctured L → ∞ -0.016 nonpunctured L → ∞
punctured punctured
0 0.5 1 1.5 2 0 0.5 1 1.5 2
θ/π θ/π

Figure 3.5: The exact results for various observables obtained by using the log definition
Qlog of the topological charge. The average plaquette (Top), the imaginary part of the
topological charge density (Middle), the topological susceptibility (Bottom) obtained for
the non-punctured (solid line) and punctured (dashed line) models are plotted against θ
for L = 10 (Left) and L = 20 (Right) with the same β = 12. Note that the results for the
punctured model are actually independent of L. For the non-punctured model, we also
plot the results in the infinite volume limit L → ∞ with β = 12 by the dash-dotted lines
for comparison.

31
 We can evaluate the free energy (3.3.5) for the punctured model more explicitly for
large β, which is relevant in the continuum limit. By integrating over φ in Eq. (3.3.2) as
1 2
eβ− 2β ( 2π −n) ,
1 θ
I(n, θ, β) ' √ (3.3.6)
2πβ
we get
1 1 θ2
log Zpunc ' β − log 2πβ − 2 . (3.3.7)
V 2 8π β
From this, we can obtain various observables for the punctured model as
1 θ2
w '1− + 2 2 , (3.3.8)
2β 8π β
hQi iθ
' 2 , (3.3.9)
V 4π β
1
χ' 2 (3.3.10)
4π β
for finite V , which explains the θ dependence observed in Fig. 3.5.
From Fig. 3.5, we also find that the results for the non-punctured model have sizable
finite volume effects, in particular around θ ∼ π, which is absent in the punctured model.
While the volume independence of the punctured model may well be peculiar to the
present 2D gauge theory case, the advantage of the punctured model compared with the
non-punctured model from the viewpoint of finite volume effects may hold more generally.

3.4 Application of the CLM to the punctured model

In this section, we apply the CLM to the punctured model using the log definition Qlog of
the topological charge. Our results reproduce the exact results discussed in the previous
section as long as we are close enough to the continuum limit. We also show that the
topology freezing problem is circumvented without causing large drifts thanks to the
puncture.

3.4.1 The drift terms for the punctured model

We have discussed the drift terms in the non-punctured model in Section 3.2.1. For
the punctured model, we only have to modify the drift terms for the four link variables
surrounding the puncture; i.e., UK,1 , UK+2̂,1 , UK,2 and UK+1̂,2 . Thus we obtain
β −1
 −1
 −i 2 (Pn − Pn − Pn−2̂ + Pn−2̂ ) for n 6= K, K + 2̂ ,

−1
Dn,1 S = −i β2 (−PK−2̂ + PK− 2̂
θ
) + i 2π for n = K , (3.4.1)
−1
 β θ
−i 2 (PK+2̂ − PK+2̂ ) − i 2π for n = K + 2̂ ,

β −1
 −1
 −i 2 (−Pn + Pn + Pn−1̂ − Pn−1̂ ) for n 6= K, K + 1̂ ,

−1
Dn,2 S = −i β2 (PK−1̂ − PK− 1̂
) − i 2πθ
for n = K , (3.4.2)
−1
 β θ
−i 2 (−PK+1̂ + PK+ ) + i 2π for n = K + 1̂ ,

1̂

32
 (β, L) = (3, 10) (β, L) = (12, 20)

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0.0 0.0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Re Qlog Re Qlog

Figure 3.6: The topological charge distribution for θ = 0 obtained by the CLM for the
punctured model using the log definition Qlog is plotted for (β, L) = (3, 10) (Left) and
(β, L) = (12, 20) (Right). The solid lines represent the exact results obtained by evaluating
(3.4.4) using the partition function (3.3.4).

where we have ignored the issue of δ-function discussed in Section 3.2.1. This is justified
if all the plaquettes in the action never cross the branch cut; i.e., |Im log Pn | ≤ π −  for
∀n 6= K with a strictly positive  during the Langevin simulation. We will see that this
assumption is justified at sufficiently large β in Section 3.4.3.
Note that the drift term from the θ term appears only for the link variables surrounding
the puncture, and it is actually a constant independent of the configuration. While these
properties are peculiar to the log definition Qlog , similar properties hold also for the sine
definition Qsin at large β, where all the plaquettes Pn approach unity except for PK , which
corresponds to the puncture. We discuss the case with the sine definition in Appendix B,
where we see that the obtained results are qualitatively the same as those obtained with
the log definition.

3.4.2 The θ dependence of the partition function

As we have seen in Section 3.3.2, the punctured model is equivalent to the non-punctured
model in the infinite volume limit for |θ| < π, beyond which the equivalence ceases to
hold. In particular, the punctured model does not have the 2π periodicity in θ, which
exists in the non-punctured model.
In order to understand this point better, we discuss the θ dependence of the partition
function in this section. Let us first note that the partition function for arbitrary θ is re-
lated to the topological charge distribution ρ(q) for θ = 0 through Fourier transformation

33
as
Z
Z(θ) = dU e−Sg [U ]+iθQ[U ]
Z Z
−Sg [U ]
= dU e dq eiθq δ(Q[U ] − q)
Z
= Z(0) dq eiθq ρ(q) . (3.4.3)

Therefore, the absence of the 2π periodicity in θ in the punctured model is directly related
to its property that the topological charge can take non-integer values even if we use the
log definition Qlog . Going beyond the fundamental region −π < θ ≤ π simply amounts to
probing the fine structure of the topological charge distribution ρ(q), which is irrelevant
in the infinite volume limit.
By making an inverse Fourier transform, we can obtain the topological charge distri-
bution ρ(q) for θ = 0 as Z ∞
1 dθ
ρ(q) = Z(θ) e−iθq . (3.4.4)
Z(0) −∞ 2π
We calculate this quantity for the punctured model by the CLM for θ = 0. In Fig. 3.6, we
show the results for (β, L) = (3, 10) (Left) and (β, L) = (12, 20) (Right), which agree well
with the exact results obtained by evaluating (3.4.4) using the partition function (3.3.4).
Note that the calculation actually reduces to that of the real Langevin method due to the
absence of the sign problem for θ = 0. We therefore have no concerns about the criterion
for correct convergence here.
While the sign problem is absent for θ = 0, the topology freezing problem can still
be an issue for large β. The agreement we see for (β, L) = (12, 20) confirms that this
problem is resolved in the punctured model at least for θ = 0.

3.4.3 Validity of the CLM

In this section, we discuss the validity of the CLM for the punctured model. Fig. 3.7(Left)
shows the histogram of the drift term for (β, L) = (3, 10) and (β, L) = (12, 20) with θ = π,
which are the parameters used in Section 3.2.4 for the non-punctured model. We find that
the criterion is satisfied for (β, L) = (12, 20) but not for (β, L) = (3, 10), similarly to the
situation in the non-punctured model. The difference from the non-punctured model is
seen, however, in Fig. 3.7(Right), where we show the histogram of Re Qlog obtained by
the CLM for (β, L) = (12, 20) with θ = π. (The result for (β, L) = (3, 10) looks quite
similar to this plot.) It is widely distributed within the range −3 . Re Qlog . 3, which is
in sharp contrast to the plot in Fig. 3.1(Right) for the same (β, L) = (12, 20) in the case
of the non-punctured model. In fact, it turns out to be close2 to the exact result obtained
2
Note, however, that precise agreement is not expected here since the histogram of Re Qlog is a non-
holomorphic quantity, for which the CLM does not allow a clear interpretation.

34
 (β, L) = (12, 20)

(β, L) = (3, 10)

0 0.5
10 (β, L) = (12, 20)

0.4
-2
10
0.3

-4
10 0.2

0.1
10-6
0.0
1 10 100 -3 -2 -1 0 1 2 3
u Re Qlog

Figure 3.7: The results obtained by the CLM for the punctured model using the log
definition Qlog of the topological charge. (Left) The histogram of the magnitude u of
the drift term defined by (3.2.12) is shown for (β, L) = (3, 10) and (12, 20) with θ = π.
(Right) The histogram of Re Qlog is shown for (β, L) = (12, 20) with θ = π. The exact
result obtained for (β, L) = (12, 20) with θ = 0 is shown by the solid line for comparison.

for the same (β, L) = (12, 20) with θ = 0, which is plotted in the same figure. Thus we
find that the topology freezing problem at large β is circumvented in the punctured model
and yet the CLM remains valid.
Next we discuss the reason why the punctured model can avoid the topology freezing
problem without causing large drifts. The difference from the non-punctured model is
that one of the plaquettes, PK , is removed from the action. Note that the topological
charge Qlog for the punctured model is given by

i X i
Qlog = − log Pn + log PK , (3.4.5)
2π n 2π

where the first term is nothing but the topological charge defined for the non-punctured
model, whose real part takes integer values. The second term has a real part which lies
within the interval [− 21 , 12 ). Therefore it makes sense to define the “topology change” in
the punctured model as the situation in which the real part of the first term changes by
±1. As we discussed in Section 3.2.5 for the non-punctured model, one of the plaquettes
should inevitably cross the branch cut in order for the topology change to occur in the
above sense. When β is large, this process is highly suppressed for all the plaquettes that
are included in the action. In the non-punctured model, the topology freezing problem
occurs precisely for this reason. However, in the punctured model, the particular plaquette
PK is removed from the action, and therefore it can change freely even for large β.
This is demonstrated in Fig. 3.8, where we plot the probability distribution of the phase
of the plaquette PK as well as that of the other plaquettes Pn (n 6= K) for (β, L) = (3, 10)
(Left) and (β, L) = (12, 20) (Right). We find that the phase of the removed plaquette
PK is almost uniformly distributed for both (β, L). On the other hand, the distribution

35
 (β, L) = (3, 10) (β, L) = (12, 20)

0 n=K n=K
10
n≠K 0 n≠K
10

10-1
10-2

10-2

10-4
-3
10

-1 -0.5 0 0.5 1 10-6-1 -0.5 0 0.5 1

Im( log Pn ) / π Im( log Pn ) / π

Figure 3.8: The distribution of the phase of the plaquettes is plotted for the punctured
model with the log definition (3.1.12) of the topological charge for (β, L) = (3, 10) (Left)
and (β, L) = (12, 20) (Right) with θ = π. We show the results for the plaquette (n = K)
removed from the action and those for all the other plaquettes (n 6= K) separately.

of the phase of the other plaquettes depends on (β, L). It has a compact support for
(β, L) = (12, 20) but not for (β, L) = (3, 10). In the former case, there is no distribution
at the branch cut, which implies that the branch cut crossing of the plaquettes Pn (n 6= K)
does not occur at all. In the latter case, there is a small but finite distribution at the
branch cut, which means that the value of β is not large enough to suppress the branch
cut crossing of the plaquettes Pn (n 6= K) completely.
This is consistent with the fact that the histogram of the drift term has fast fall-off for
(β, L) = (12, 20) but not for (β, L) = (3, 10) considering the discussion given in Section
3.2.5. While the flow diagram in Fig. 3.4(Left) is obtained for the sine definition of the
topological charge, it looks similar for the log definition, which simply corresponds to
setting θ = 0 in (3.2.18). Therefore, large drifts can appear when one of the plaquettes
Pn (n 6= K) crosses the branch cut, which indeed occurs for (β, L) = (3, 10) also for the
punctured model. For (β, L) = (12, 20), on the other hand, the topology change is made
possible by allowing the removed plaquette PK to cross the branch cut freely, but all
the plaquettes that are included in the action are forced to stay close to unity because
of large β. This justifies our assumption that the issue of δ-function can be neglected in
deriving the drift terms (3.4.1) and (3.4.2). Since the plaquette PK does not appear in the
drift terms, it does not cause large drifts even if it crosses the branch cut. This makes it
possible for the punctured model to avoid the topology freezing problem without causing
large drifts.
Let us next discuss how the unitarity norm (3.2.5) behaves in our complex Langevin
simulations. As we can see from (3.4.1) and (3.4.2), the link variables surrounding the
puncture have a drift term in the imaginary direction coming from the θ term. At each
Langevin step, two of the link variables are multiplied by eθ∆t/2π and the other two
are multiplied by e−θ∆t/2π so that the removed plaquette is multiplied by e2θ∆t/π due

36
 4
10

102

100
N

10-2 θ = 1.0 π
θ = 0.8 π
θ = 0.6 π
10-4
θ = 0.4 π
θ = 0.2 π
10-6 0 60 120 180 240 300 360
t

Figure 3.9: The history of the unitarity norm N is plotted for the punctured model with
the log definition (3.1.12) of the topological charge for various θ with (β, L) = (5, 16).

to this drift term. Therefore, there is a danger that the magnitude of these four link
variables increases or decreases exponentially and hence the unitarity norm (3.2.5) grows
exponentially with the Langevin time.
In Fig. 3.9, we plot the history of the unitarity norm (3.2.5) for various θ with (β, L) =
(5, 16). Similar results are obtained for other (β, L). (Here and for the rest of this
subsection, we restrict ourselves to the parameter sets, for which the histogram of the
drift term has fast fall-off.) Indeed we observe an exponential growth at early stage, but
the unitarity norm actually saturates to a constant depending on θ at sufficiently long
Langevin time. This saturation occurs since the non-unitarity of the four link variables
surrounding the puncture propagates to all the other link variables on the lattice due to
the interaction caused by the gauge action Sg , which tries to make each plaquette except
the removed one close to unity. We find that thermalization of various observables can
be achieved only after the saturation of the unitarity norm.
In fact, we find that the unitarity norm is not distributed uniformly on the lattice due
to the existence of the puncture, as is also expected from the above discussion. In order
to see this, we define the “local unitarity norm” by
1 X n ∗ ∗ −1
o
N (n) = Uk,µ Uk,µ + (Uk,µ Uk,µ ) − 2 , (3.4.6)
4
(k,µ)∈Pn

which is an average of the unitarity norm for the four link variables surrounding each
plaquette Pn . The unitarity norm defined by (3.2.5) is simply an average of N (n) over all
the plaquettes including the removed one; namely N = L12 n N (n). In Fig. 3.10(Left),
P

we plot this quantity N (n) against n = (n1 , n2 ) for (β, L) = (12, 20) with θ = π, where
the puncture is located at n = K = (10, 10). We observe a sharp peak at the puncture,

37
 (β, L) = (5, 16)
108 (β, L) = (10, 16)
N(n) (β, L) = (6, 20)
104 x p exp( c1x + c2 )
106
102

N(K)
100 104

10-2
2
15 10
10-4 0 10
5 n2
10 5
n1 15
0 100 0 50 100 150 200
|θ| L2 / β

Figure 3.10: (Left) The local unitarity norm N (n) for each plaquette Pn defined by
(3.4.6) is plotted against n = (n1 , n2 ) for the punctured model with the log definition
(3.1.12) of the topological charge for (β, L) = (12, 20) with θ = π. The removed plaquette
corresponds to n = K = (10, 10) in this figure. (Right) The local unitarity norm N (K) for
the removed plaquette PK obtained for various β, L and θ is plotted against x = |θ|L2 /β.
The solid line represents a fit to xp exp(c1 x + c2 ) with c1 = 0.079(5), c2 = −3(1) and
p = 0.8(4).

which goes up to N (K) ∼ 6 × 103 . The plaquettes adjacent to the puncture have a local
unitarity norm ∼ 1.5 × 103 . This implies that the unitarity norm is mostly dominated by
the four link variables surrounding the puncture.
The local unitarity norm N (K) at the puncture depends not only on θ but also on β
and L. In Fig. 3.10(Right), we plot this value against x = |θ|Vphys = |θ|L2 /β for various
θ, β and L. All the data can be fitted to a single curve N (K) = xp exp(c1 x + c2 ), which
reveals an exponential behavior at large x.
What actually matters for the validity of the CLM is not so much the local unitarity
norm N (n) as the absolute value of each plaquette Pn , which we plot in Fig. 3.11 against
n = (n1 , n2 ) for the same parameters as in Fig. 3.10(Left). The absolute value of PK
p
corresponding to the removed plaquette is close to ( N (K))4 ∼ 3.6 × 107 , which implies
−1 −1
p
that |UK,1 |, |UK+1̂,2 |, |UK+ 2̂,1
| and |U K,2 | are close to N (K). Except for this removed
plaquette, the absolute value of the plaquette deviates only slightly from unity due to
large β.
In fact, this deviation of |Pn | from unity for n 6= K has a physical meaning since
1
P
Im hQlog i = − 2π n6=K hlog |Pn |i as one can see from (3.4.5). From the exact result (3.3.9)
obtained at large β, we find that |Pn | ∼ e−θ/(2πβ) for n 6= K, which is ∼ 0.96 for θ = π
and β = 12 in agreement with the value observed in Fig. 3.11. If we flip the sign of θ,
which corresponds to the parity transformation, we find that |Pn | 7→ |Pn |−1 for all n.
Note also that PK does not appear in the drift term, which implies that its absolute
value can become large without causing large drifts. We have confirmed that the criterion
for correct convergence is satisfied for sufficiently large β, and indeed the exact results

38
 | Pn |
1.00

0.98

0.96

0.94
15
0.92 0 10
5 n2
10 5
n1 15
0

Figure 3.11: The absolute value of the plaquette Pn is plotted against n = (n1 , n2 )
for the punctured model with the log definition (3.1.12) of the topological charge for
(β, L) = (12, 20) with θ = π. The removed plaquette corresponds to n = K = (10, 10) in
this figure.

for various observables can be reproduced correctly as we will see in the next section.
This remains to be the case even for large θ and/or large Vphys , where the unitarity
norm becomes large.3 Thus the present model provides a counterexample to the common
wisdom that the CLM fails when the unitarity norm becomes large.

3.4.4 Results for the observables

In this section, we calculate the observables for the punctured model by the CLM and
compare our results with the exact results derived in Appendix A.4. Let us recall that,
in the definitions (3.2.14), (3.2.15) and (3.2.16), V denotes the number of plaquettes in
the action, which is V = L2 − 1 for the punctured model. In contrast, we define the
physical volume Vphys by Eq. (3.1.15) not only for the non-punctured model but also for
the punctured model, which simplifies the relationship between β and L for fixed Vphys .
In Fig. 3.12, we show our results for the average plaquette w (Top), the topological
charge (Middle) and the topological susceptibility χ (Bottom) against θ for (β, L) = (3, 10)
and (12, 20) in the left and right columns, respectively, which correspond to a fixed physical
volume Vphys ≡ L2 /β = 102 /3. The exact results obtained for the same model with the
same parameter sets are also shown for comparison. We find from our results for the
average plaquette that the exact results are reproduced for (β, L) = (12, 20), but there
3
We find, however, that the fluctuation of the local unitarity norm is small even for the one N (K) at
the puncture, which implies that the distribution of each link variable has fast fall-off. Therefore, it is
suggested that the problem due to the boundary terms discussed in Refs. [3, 4, 72, 73] does not occur.

39
 (β, L) = (3, 10) (β, L) = (12, 20)

0.86
0.959

0.84
w

w
0.958

0.82

0.957
0 0.5 1 0 0.5 1
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.05 0.010

0.04 0.008
Im 〈 Qlog 〉 / V

Im 〈 Qlog 〉 / V

0.03 0.006

0.02 0.004

0.01 0.002

0.00 0.000
0 0.5 1 0 0.5 1
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.020 0.004

0.015 0.003

0.010 0.002
χ

0.005 0.001

0.000 0.000
0 0.5 1 0 0.5 1
θ/π θ/π

Figure 3.12: The results for various observables obtained by the CLM for the punctured
model with the log definition Qlog . The average plaquette (Top), the imaginary part of the
topological charge density (Middle), the topological susceptibility (Bottom) are plotted
against θ for (β, L) = (3, 10) (Left) and (12, 20) (Right). The exact results for the same
(β, L) are shown by the dashed lines for comparison.

40
is slight deviation for (β, L) = (3, 10). This is consistent with our observation in Section
3.4.3 that the condition for correct convergence is met for (β, L) = (12, 20) but not for
(β, L) = (3, 10).
For the topological charge, we find that our results reproduce the exact results not
only for (β, L) = (12, 20) but also for (β, L) = (3, 10). The same holds for the topological
susceptibility. We consider that the agreement observed here for (β, L) = (3, 10) is acci-
dental, though, since the condition for correct convergence is not satisfied. The fact that
the results of the CLM for the punctured model with (β, L) = (3, 10) is not as bad as
those for the non-punctured model with the same (β, L) shown in Fig. 3.2(Left) can be
understood by considering that the effect of the θ term is included correctly by the drift
terms for the link variables composing the removed plaquette, but it is only the infrequent
branch cut crossing of the other plaquettes that spoils the validity of the CLM.

41
Chapter 4

4D SU(2) gauge theory with a θ term

There are some predictions about the phase structure around θ = π by using the ’t Hooft
anomaly matching condition. The aim of this work is to investigate the phase structure
around θ = π by using the CLM.

4.1 Lattice formulation of the 4D SU(2) gauge theory

In this section, we review 4D SU(2) gauge theory with a θ term and discuss how to define
it on a lattice.
In the continuum 4D SU(N ) gauge theory, the kinetic term for gauge field Aµ = Aaµ ta
is given by
1 1
Z Z
Sg = 2 d x Tr [Fµν Fµν ] = 2 d4 xFµν
4 a a
Fµν , (4.1.1)
2g 4g
where µ, ν run from 1 to 4, an index a labels the generators of SU(N ), g is the gauge
coupling constant, and Fµν is the field strength defined by
a a
Fµν = ∂µ Aν − ∂ν Aµ + i [Aµ , Aν ] = Fµν t . (4.1.2)

We use the Hermitian generators ta which satisfy the following relations:

(ta )† = ta , (4.1.3)
Tr ta = 0, (4.1.4)
[ta , tb ] = if abc tc , (4.1.5)
Tr ta tb = 12 δ ab . (4.1.6)

In 4D gauge theories, the topological charge is defined by

1 1
Z Z
4
Q= d xµνρσ Tr [Fµν Fρσ ] = d4 xµνρσ Fµν
a a
Fρσ , (4.1.7)
32π 2 64π 2
which becomes some integer when the base manifold is compact. Therefore, the total
action is composed by the kinetic term and the θ term as

S = Sg + Sθ , (4.1.8)

42
 Sθ = −iθQ, (4.1.9)
where Sg is given by (4.1.1).
We put this theory on the L1 × L2 × L3 × L4 lattice. In the rest of this section, we
set L1 = L2 = L3 = Ls and L4 = Lt . For the lattice formulation of gauge theories, we
introduce link variables Un,µ which are related to gauge fields An,µ as

Un,µ = eiaAn,µ , (4.1.10)

where n = (n1 , n2 , n3 , n4 ) represents the position on the lattice, and a is a lattice spacing.
The simplest gauge action is the plaquette action which is sometimes called as the
Wilson action. The action is defined as
β XX
Sglat = Tr (I − Pnµν )
2N n µ6=ν
(4.1.11)
β XX
= 6βV − Tr [Pnµν + (Pnµν )−1 ],
2N n µ<ν

where V = L3s × Lt is the lattice volume, β is related to the gauge coupling as

2N
β= , (4.1.12)
g2
and Pnµν is a plaquette which is defined by

−1 −1
Pnµν = Un,µ Un+µ̂,ν Un+ν̂,µ Un,ν . (4.1.13)

Next, we consider the lattice formulation of the topological charge. There are some
definitions of the topological charge on the lattice. Comparisons between different defini-
tions have been discussed in [74, 75].
In this work, we use the most commonly used definition, namely the cloverleaf defi-
nition which is introduced in [76]. The cloverleaf definition of the topological charge is
given by
±4
1 X 1 X
QL := − ˜µνρσ Tr [Pnµν Pnρσ ] , (4.1.14)
32π 2 n 24 µ,ν,ρ,σ=±1

where ˜µνρσ is defined as

˜µνρσ = sgn(µνρσ)|µ||ν||ρ||σ| , (4.1.15)
where µνρσ is the Levi-Civita symbol. The plaquettes for the negative directions are
defined as
−1
Pn−µν = Un−µ̂,µ −1
Un−µ̂,ν Un−µ̂+ν̂,µ Un,ν ,
−1 −1
Pnµ−ν = Un,µ Un+µ̂−ν̂,ν Un−ν̂,µ Un−ν̂,ν , (4.1.16)
−1 −1
Pn−µ−ν = Un−µ̂,µ Un−µ̂−ν̂,ν Un−µ̂−ν̂,µ Un−ν̂,ν .

43
By introducing the cloverleaf P̄nµν whose definition is

P̄nµν = Pnµν − Pn−µν − Pnµ−ν + Pn−µ−ν , (4.1.17)

then the definition (4.1.14) can be rewritten as

4
1 X 1 X
µνρσ Tr P̄nµν P̄nρσ .
 
QL = − (4.1.18)
32π 2 n 24 µ,ν,ρ,σ=1

Moreover, the lattice definition of topological charge (4.1.18) can be rewritten in a

simpler form as
1 X  12 34
Tr Rn Rn + Rn13 Rn42 + Rn23 Rn14 ,

QL = − 2
(4.1.19)
256π n

where Rnµν is defined as

Rnµν := P̄nµν − P̄nνµ . (4.1.20)
Rnµν is anti-symmetric under a swapping of indices µ and ν as Rnµν = −Rnνµ .
This topological charge is not an integer when the lattice spacing a is finite. In the
continuum limit, the lattice topological charge becomes some integer.

4.2 Application of the CLM to the 4D SU(2) lattice

gauge theory
We apply the CLM to the 4D SU(2) lattice gauge theory with a θ term. The link variables
are extended from Un,µ ∈ SU(2) to Un,µ ∈ SL(2, C). Then we consider a fictitious time
evolution of the link variables Un,µ (t) governed by the complex Langevin equation
hn
a a
√ oi
Un,µ (t + ∆t) = Un,µ (t) exp i − ∆t Dn,µ S t + ∆t ηn,µ (t) , (4.2.1)

where t is the fictitious time, ∆t is a step size, and ηn,µ (t) is the Gaussian noise. The drift
a
term Dn,µ S is defined by
a
a S(eit Un,µ ) − S(Un,µ )
Dn,µ S = lim , (4.2.2)
→0 
first for the unitary link variables Un,µ (t), and then it is defined for the complexified link
variables Un,µ (t) by analytic continuation in order to respect holomorphicity.
a
The drift term Dn,µ S is consist of two part
a a a
Dn,µ S = Dn,µ Sg + Dn,µ Sθ , (4.2.3)

where the first term is defined as

" #
X
a
Dn,µ Sg = −iβTr ta (Pnµν + Pnµ−ν − Pn−νµ − Pnνµ ) , (4.2.4)
ν6=µ

44
and the second term is defined as

a iθ a
X
Dn,µ Sθ = D Tr [Rn12 Rn34 − Rn13 Rn24 + Rn14 R23 ]
256π 2 n,µ n
θ
=− Tr [ta {δµ1 (Kn,12
34 24
− Kn,13 23
+ Kn,14 )
256π 2
34 13 23
+ δµ2 (−Kn,21 − Kn,24 + Kn,14 ) (4.2.5)
12 24 23
+ δµ3 (Kn,34 + Kn,31 − Kn,14 )
12 13 23
+ δµ4 (−Kn,43 + Kn,24 − Kn,41 )}]
θ
=: − 2
Tr [ta Jn,µ
(θ)
],
256π
ρσ ρσ ρσ −1
Kn,µν := Un,µ fn,µν (U ) + gn,µν (U )Un,µ , (4.2.6)
ρσ −1 −1 ρσ −1 −1 ρσ
fn,µν (U ) : = Un+µ̂,ν Un+ν̂,µ Un,ν Rn − Un+µ̂−ν̂,ν Un−ν̂,µ Rn−ν̂ Un−ν̂,ν
−1 ρσ −1 −1 ρσ −1
+ Un+µ̂,ν Un+ν̂,µ Rn+ν̂ Un,ν − Un+µ̂−ν̂,ν Rn+µ̂−ν̂ Un−ν̂,µ Un−ν̂,ν
ρσ −1 −1 −1 −1
(4.2.7)
+ Rn+ν̂ Un+µ̂,ν Un+ν̂,µ Un,ν − Un+µ̂−ν̂,ν Un−ν̂,µ Un−ν̂,ν Rnρσ
ρσ −1 −1 ρσ −1 −1
+ Un+µ̂,ν Rn+µ̂+ν̂ Un+ν̂,µ Un,ν − Rn+µ̂ Un−µ̂+ν̂,ν Un−ν̂,µ Un−ν̂,ν ,
ρσ −1 −1 ρσ
gn,µν (U ) : = Rnρσ Un,ν Un+ν̂,µ Un+ν̂,ν − Un−ν̂,ν Rn−ν̂ Un−ν̂,µ Un+µ̂−ν̂,ν
−1 ρσ −1
+ Un,ν Un+ν̂,µ Un+µ̂,ν Rn+µ̂ − Rnρσ Un−ν̂,ν Un−ν̂,µ Un+µ̂−ν̂,ν
ρσ −1 −1 ρσ
(4.2.8)
+ Un,ν Rn+µ̂ Un+ν̂,µ Un+µ̂,ν − Un−ν̂,ν Un−ν̂,µ Rn+µ̂−ν̂ Un+µ̂−ν̂,ν
−1 −1 ρσ
+ Un,ν Un+ν̂,µ Rn+µ̂+ν̂ Un+µ̂,ν − Un−µ̂,ν Un−µ̂,µ Un+µ̂−ν̂,ν Rn+µ̂ .
Therefore, the total drift term is

a
Dn,µ (Sg + Sθ ) = iTr [Jn,µ ta ], (4.2.9)

(g) iθ
Jn,µ := −βJn,µ + J (θ) . (4.2.10)
256π 2 n,µ
In this work, we use the second-order Runge-Kutta algorithm [77, 78] to reduce the
finite step size effects. When we evolve the gauge configuration by (4.2.1), the probability
for a gauge configuration P (U ) which is realized after the thermalization is given by

P (U ) ∝ exp (−S̃), (4.2.11)

where S̃ is
 
∆t CA ∆t X a 2 a 2

2

S̃ = 1+ S+ 2(Dn,µ ) S − (Dn,µ S) + O (∆t) , (4.2.12)
12 4 n,µ

and CA is the Casimir invariant for the adjoint representation. Note that, in the SU(N )
case, CA = N . In other words, when we evolve the gauge configuration by (4.2.1) the
observables have O(∆t) systematic errors.

45
 In the second-order Runge-Kutta algorithm, the Langevin evolution is described by
the following equations:
h n ∆t  √

∆t CA a
oi
Un,µ (t + ∆t) = Un,µ (t) exp i − 1+ vnµ ta + ∆t ηn,µ (t) ,
2 6 (4.2.13)
 
a a a 0



vnµ = Dn,µ S U (t) + Dn,µ S U (t ) ,

where a gauge configuration U (t0 ) is obtained by a tentative update as

hn √ oi
Un,µ (t0 ) = Un,µ (t) exp i − ∆t Dn,µ
a
S U (t) ta + ∆t ηn,µ (t) .


(4.2.14)

By this algorithm, the first order of ∆t in (4.2.12) is removed, therefore the probability
for a gauge configuration becomes


2
P (U ) ∝ exp −S + O (∆t) . (4.2.15)

Thus, the systematic errors are reduced from O(∆t) to O((∆t)2 ). As a result, we can use
a larger ∆t, which means that the simulations are more efficient.

4.2.1 Gauge cooling

We define a semi-positive definite norm so called the unitarity norm as
1 X †
N = Tr [Un,µ Un,µ − I]. (4.2.16)
4N V n,µ

This norm describes how the link variables are far from being unitary. If and only if all
link variables are unitary Un,µ ∈ SU(2), their Hermite conjugates are equivalent to their
† −1
inverse matrices Un,µ = Un,µ , therefore the norm is equal to zero. We reduce the norm by
a gauge transformation for the extended gauge group as
−1
Un,µ −→ gn Un,µ gn+µ gn ∈ SL(2, C). (4.2.17)

This procedure is called the gauge cooling. The gauge cooling does not affect the expec-
tation values for holomorphic quantities [3].
Next we consider how to choose the gauge transformation. Here we consider the
SL(2, C) gauge transformation for an infinitesimal parameter ∆an ∈ R such as

gn = e∆n t ∼ 1 + ∆an ta ,
a a

−1
gn Un,µ gn+µ = Un,µ + ∆an Un,µ − ∆an+µ̂ Un,µ ta , (4.2.18)
−1 † † † †
gn+µ̂ Un,µ gn = Un,µ − ∆an+µ̂ Un,µ + ∆an Un,µ ta .

The change of the unitarity norm is given by

1 X
∆N = Tr [∆an Gˆn ta ], (4.2.19)
2N V n

46
where Gˆn is defined as

Gˆn :=
X †
†
(Un,µ Un,µ − Un−µ̂,µ Un−µ̂,µ ), (4.2.20)
µ

†
which satisfies Gˆn = Gˆn . Therefore, we find that the unitarity norm is reduced most
efficiently by choosing ∆an ∝ −Gˆn ta .
Using this result, we consider following gauge transformation

−1
Un,µ −→ gn Un,µ gn+µ ; gn = e−αGn , (4.2.21)

where Gn is defined as
Gn := Tr [Gˆn ta ]ta . (4.2.22)
After this gauge transformation, the unitarity norm becomes
1
N 0 (α) := †
Tr [Un,µ e−2αGn Un,µ e2αGn+µ̂ − 1]. (4.2.23)
4N V
We search for an optimal value for α that minimize N 0 (α). Since we perform the gauge
cooling after each Langevin step, the α is typically a small number. Therefore, we expand
Eq. (4.2.23) with respect to α up to first order, and we estimate optimal value for α. The
optimal value is given by
2
P
n ||Gn ||
α= P , (4.2.24)
n,µ ||Gn Un,µ − Un,µ Gn+µ̂ ||
2

where || · || is the Frobenius norm which is defined as ||A||2 = Tr (AAT ). We repeat this
procedure until the change of the unitarity norm becomes less than 10−4 .

4.2.2 Adaptive step size

In order to avoid the excursion problem, we also use the adaptive step size algorithm. We
measure maximum of the magnitude of the drift term defined as

a
u = max |Dn,µ S| (4.2.25)
n,µ

at each Langevin step. When the magnitude of drift term is larger than a threshold u0 ,
we modify the step size ∆t as
(
∆t0 for u < v0 ,
∆t = v0 (4.2.26)
∆t0 otherwise ,
u
where ∆t0 is a default step size. In our simulation, we choose the threshold as u0 = 3β/Nc
because when θ = 0 the drift term is bounded by this value.

47
 (L, β) = (16, 3), original (L, β) = (16, 6), original
10 10

5 5
Re Q

Re Q
0 0

-5 -5

-10 -10
0 10 20 30 40 50 0 10 20 30 40 50
t t

Figure 4.1: We plot the histories of the topological charge for θ = 0. Here we consider
Ls = Lt = L where Ls is a number of the lattice point in spatial directions and Lt is that
in the temporal direction. The simulations start from random SU(2) configurations (hot
start). Different lines correspond to different initial configurations. In (L, β) = (16, 3)
case (Left), the topological charge seems to change frequently. On the other hand, in
(L, β) = (16, 6) case (Right), the topology does not change during a simulation.

4.3 Result for a naive implementation

First, we impose periodic boundary for all directions.

4.3.1 Autocorrelation of the topological charge

First of all, we study the behaviors of the topological charge at θ = 0 where the CLM
turns out to be the real Langevin method. We show the histories of the topological
charge in Fig. 4.1. In the small β case, the topological charge seems to change during a
simulation. On the other hand, in the large β case, the topology freezing problem occurs.
This situation resembles the 2D U(1) case.
We also measure the distributions of the topological charge, and we show them in
Fig. 4.2. The width becomes large as the β becomes small. Note that, unlike in the
2D U(1) with the sine definition case, we cannot find the region where the comb-shaped
distribution of the charge appears because of the large UV fluctuations.
We can reduce the UV fluctuations by a cooling procedure such as the gradient flow.
In the gradient flow technique, the gauge configurations are evolved by the following
equation:
hn oi
a a
Vn,µ (τ + ∆τ ) = Vn,µ (τ ) exp i − ∆τ Dn,µ S t ,
(4.3.1)
Vn,µ (0) = Un,µ ,

where τ is called as the flow time. Eq. (4.3.1) is equivalent to the Langevin evolution
(4.2.1) without the noise term. After the gradient flow, the comb-shaped distribution

48
 60 160

140
50
120
40
100

30 80

60
20
40
10
20

0 0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Q Q

Figure 4.2: We plot the histograms of the topological charge for θ = 0. Here, we set
Ls = Lt = 16. In β = 3 case (Left), the topological charge is distributed in (−3, 3). On
the other hands, in β = 6 case (Right), the topological charge is distributed in (−1, 1).

0.3
before G.F.
after G.F.

0.2

0.1

0
-5 -4 -3 -2 -1 0 1 2 3 4 5
Q

Figure 4.3: The charge distributions before and after the gradient flow. The parameters
are Ls = Lt = 8, β = 1.0 and θ = 0. We use ∆t = 10−3 for generating the configurations
by real Langevin method, and we use ∆τ = 10−1 for the gradient flow. The total flow
time is 10.

49
 ( β, V ) = ( 3.25, 164 ) ( β, V ) = ( 3.25, 164 )
6 4
10
θ=0
5
10 θ = 0.5 π
θ=π 3
104
θ = 1.5 π

Im 〈 Q 〉
3 θ=2π
10
2
2
10

101 1
0
10
0
10-1 0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2
log10 u θ/π

Figure 4.4: The results obtained by the CLM for various θ with Ls = Lt = 16 and
β = 3.25. (Left) The histogram of the magnitude of the drift term defined by (4.2.25).
(Right) The imaginary part of the topological charge.

of the topological charge appears (see Fig. 4.3). However, the gradient flow technique
cannot be justified in the CLM, therefore we cannot use this technique at θ 6= 0.

4.3.2 Finite θ simulation

We turn on the θ at small β where the topology freezing problem does not occur. In
Fig. 4.4 (Left), we plot the histogram of the drift terms for various θ. We find that
the criterion for the correct convergence of the CLM is satisfied up to θ = 2π. In Fig.
4.4 (Right), the imaginary part of the topological charge whose gradient becomes the
topological susceptibility is plotted against the θ. We cannot see the 2π periodicity which
the theory has. We consider the reason for this problem is the large UV fluctuations. Since
we cannot use the gradient flow in the CLM, a possible way to reduce the UV fluctuations
is increasing β, which corresponds to decreasing the lattice spacing a. However, in the
large β region, there is the topology freezing problem. Therefore, we need to solve this
problem.

4.4 Modifying the boundary condition

4.4.1 Open boundary condition for one spatial direction
Some methods to overcome the topology freezing problem have been studied [79, 80, 81, 82,
83, 84]. In particular, Luscher and Schaefer suggested modifying the boundary condition.
In [79], they impose the open boundary condition in the imaginary time direction focusing
on the zero temperature case. They found that the topology freezing problem is alleviated
by this method. They also found that a drawback of this method is a strong finite volume
effect.

50
 In this study, since we want to study the theory at finite temperature, we need to
impose the periodic boundary condition for the imaginary time direction. Thus, we
impose the open boundary condition in the spatial directions. Since we want to avoid
large finite volume effects from the boundaries, first we try to impose the open boundary
condition for only one of the spatial directions.
Here we impose the open boundary condition for n1 direction. In this case the Wilson
gauge action is rewritten as
β X X µν
Sg(lat, obc)
= wn Tr (I − Pnµν ), (4.4.1)
2N n µ6=ν

where wnµν are weights. Except for at the boundary (n1 = 1 or n1 = L1 ), these weights
are unity. At the boundary these weights are

0 (n1 , µ) = (L1 , 1)|ν6=µ ,



wnµν = 1
2
(n1 , µ, ν) = (1, 2, 3), (1, 2, 4), (1, 3, 4), (L1 , 2, 3), (L1 , 2, 4), (L1 , 3, 4), (4.4.2)


1

otherwise,

where we consider only for µ < ν case because these weights are symmetric wnµν = wnνµ .
In the periodic boundary case, each of the plaquettes belongs to 4 unit hypercubes. On
the other hands, in the open boundary for one spatial direction case, at the boundary,
half of the hypercubes are absent for some plaquettes. Therefore, some plaquettes at the
boundary belong to only 2 unit hypercubes. In this reason we use wnµν = 24 = 12 for such
plaquettes.
In the open boundary case, we removed the cloverleaves at the boundary from the
summation in the definition of the topological charge (4.1.18). In Fig. 4.5, we plot the
histories of the topological charge at β = 6 in the periodic boundary case (Left) and that
in the open boundary case (Right). We find that the topology freezing problem seems to
be alleviated. However, the autocorrelation time is still long even if we impose the open
boundary condition. We expect that the autocorrelation time decreases if we impose the
open boundary condition for all spatial directions.

4.4.2 Open boundary condition for all spatial directions

We impose the open boundary conditions for all spatial directions to solve the topology
freezing problem. The weights in (4.4.1) are determined by the same way as we explained
in the previous subsection.
In Fig. 4.6, we plot the histories of the topological charge in the case of the open
boundaries for all spatial directions. We find that the topology freezing is more milder
than that for the case of open boundary for one of the spatial directions.

51
 (L, β) = (16, 6), original (L, β) = (16, 6), 1 open
10 10

5 5
Re Q

Re Q
0 0

-5 -5

-10 -10
0 10 20 30 40 50 0 10 20 30 40 50
t t

Figure 4.5: We plot the histories of the topological charge for θ = 0 with (L, β) =
(16, 6). The simulations start from random SU(2) configurations (hot start). Different
lines correspond to different initial configurations. We plot the histories in the periodic
boundary case in (Left), on the other hand, that in the open boundary for only one spatial
direction is plotted in (Right). We cannot see the qualitative difference between them.

(L, β) = (16, 6), 3 open

10

5
Re Q

-5

-10
0 10 20 30 40 50
t

Figure 4.6: We plots the histories of the topological charge for θ = 0 with (L, β) = (16, 6).
We impose the open boundary condition for all spatial directions. The simulations start
from random SU(2) configurations (hot start). Different lines correspond to different
initial configurations.

52
 ( LT, β ) = ( 4, 6.0 ), 3 open ( LT, β ) = ( 4, 6.0 ), 3 open
0.9
1.4 LS = 16
0.8 LS = 32
1.2
LS = 64

Im 〈 Q 〉 105 / V
0.7 1.0 LS = 80
0.6 0.8
〈P〉

0.5 0.6
LS = 16 0.4
0.4
LS = 32
LS = 64 0.2
0.3
LS = 80 0.0
0.2
0 1 2 0 1 2
θ/π θ/π
( LT, β ) = ( 4, 6.0 ), 3 open
0.2
0.18
Im 〈 Q 〉 105 / ( θ V )

0.16
0.14
0.12
0.1
0.08
θ = 0.5 π
0.06
θ = 1.0 π
0.04
θ = 1.5 π
0.02
θ = 2.0 π
0
0 1e-05 2e-05 3e-05 4e-05 5e-05
1 / VS

Figure 4.7: We plot the results for Lt = 4, β = 6 and various Ls at finite θ. Here we
impose the open boundary condition for all spatial directions. In all cases, the criterion for
the correct convergence of the CLM is satisfied. (Top-Left) The Polyakov loop is plotted
against θ for various Ls . (Top-Right) We plot the density of the imaginary part of the
topological charge against θ/π. (Bottom) We plot ImQ/(θV ) against 1/Vs for Ls = 32, 64
and 80 cases.

In Fig. 4.7 (Top-Left), the Polyakov loop is plotted against θ for various Ls . The
Polyakov loop is defined by
1 X Y
P = 3 Un,4 , (4.4.3)
Ls n ,n ,n n
1 2 3 4

which is an order parameter to distinguish between the confined and deconfined phases.
hP i = 0 corresponds to the confined phase, while, hP i = 6 0 corresponds to the deconfined
phase. This figure shows that the Polyakov loop has a nonzero value for all cases. This
result implies that the theory is in the deconfined phase. In Fig. 4.7 (Top-Right), we plot
the density of the imaginary part of Q against θ for various Ls . We find that Im(Q) linearly
increases. We also find that the density of Im(Q) gradually decreases as Ls increases. In
Fig. 4.7 (Bottom), we plot Im(Q)/(θV ) against 1/Vs , where Vs = L3s is a spatial lattice
volume. We cannot see the finite volume scaling. Therefore, we need to increase Vs more
to take Vs → ∞ extrapolation.

53
Chapter 5

On the emergence of the space-time

structure in the type IIB matrix
model

5.1 Brief review of the Lorentzian type IIB matrix

model
In this section, we define the Lorentzian type IIB matrix model and its simplified versions,
and review some results obtained by Monte Carlo simulations.

5.1.1 Definition of the Lorentzian type IIB matrix model

The action of the type IIB matrix model is given as [47]

S = Sb + Sf , (5.1.1)
1  µ ν

Sb = − Tr [Aµ , Aν ] [A , A ] , (5.1.2)
4
1  µ

Sf = − Tr Ψα (CΓ )αβ [Aµ , Ψβ ] , (5.1.3)
2
where Aµ (µ = 0, 1, · · · , 9) and Ψα (α = 1, · · · , 16) are bosonic and fermionic N × N
traceless Hermitian matrices. The indices µ and ν are contracted with the Lorentzian
metric ηµν = diag(−1, 1, . . . , 1). The 16 × 16 matrices Γµ and C are the 10-dimensional
gamma matrices and the charge conjugation matrix, respectively, obtained after the Weyl
projection. The action (5.1.1) has a manifest SO(9,1) Lorentz symmetry, under which Aµ
and Ψα transform as a Lorentz vector and a Majorana-Weyl spinor, respectively.
This model is invariant under the following transformations

δ (1) Aµ = i¯1 Γµ Ψ,
i (5.1.4)
δ (1) Ψ = Γµν [Aµ , Aν ]1 ,
2

54
 δ (2) Aµ = 0,
(5.1.5)
δ (2) Ψ = 2 1,

δT Aµ = cµ 1,
(5.1.6)
δT Ψ = 0,

δG Aµ = i[λ, Aµ ],
(5.1.7)
δG Ψ = i[λ, Ψ],
where 1 and 2 are Majorana-Weyl spinors, cµ is a 10D vector, 1 is N × N unit matrix,
and λ is a N × N Hermitian matrix. The transformation (5.1.7) is the zero volume limit
of the 10-dimensional SU(N) gauge transformation.
Here we write the generators of (5.1.4), (5.1.5), and (5.1.6) as Q(1) , Q(2) , and Pµ
respectively. And we define Q̃(1) and Q̃(2) as

Q̃(1) = Q(1) + Q(2) ,

(5.1.8)
Q̃(2) = i(Q(1) − Q(2) ).

These generator satisfies the following relation

[¯1 Q̃(i) , ¯2 Q̃(j) ] = −2δ ij ¯1 Γµ 2 Pµ , (5.1.9)

up to the gauge symmetry (5.1.7) and the equation of motion for fermionic matrices as

Γµ [Aµ , Ψ] = 0. (5.1.10)

If we identify Pµ as the momentum, the relation corresponds to the algebra of 10 dimen-

sional N = 2 supersymmetry. Therefore (5.1.6) corresponds to translation, which means
that we can identify the eigenvalues of the matrices as the space-time coordinates.
There is evidence that the model is considered a promising candidate for the non-
perturbative formulation of superstring theory. One is that the action (5.1.1) is recognized
as a matrix regularization for the action of the type IIB string theory. Another is that this
model can realize the correct interaction between D-branes. Other is that the light-cone
string field theory is derived from this model.
The partition function of the Lorentzian type IIB matrix model is defined as [60]
Z Z
iS[A,Ψ]
Z = dA dΨ e = dA PfM(A) eiSb , (5.1.11)

where the “i” in front of the action is motivated from the fact that the string worldsheet
metric has a Lorentzian signature. Note that the bosonic action Sb can be written as
1 1
Sb = Tr (Fµν F µν ) = −2Tr (F0i )2 + Tr (Fij )2 , (5.1.12)
4 4
where we have introduced the Hermitian matrices Fµν = i [Aµ , Aν ]. Hence Sb is not
positive semi-definite unlike in the Euclidean case. Note also that, unlike in the Euclidean

55
version [85, 86], the matrix integral in (5.1.11) is divergent because eiSb is a pure phase
factor and the Pfaffian PfM(A) obtained by integrating out the fermionic matrices is a
polynomial in Aµ .
In order to make the partition function (5.1.11) finite, we need to introduce the IR
cutoffs both in the temporal and spatial directions, for instance, as
1  p 1  p
Tr (A0 )2 ≤ κp Tr (Ai )2 , (5.1.13)
N N
1  p
Tr (Ai )2 ≤ L2p . (5.1.14)
N
The power p is a parameter, which can be used to test how much the obtained results
depend on the way the IR cutoff is introduced [87]. While p = 1 would be a natural
choice, it was proposed that p should be chosen to be a slightly larger value in order
to make the results almost independent of p. Too large values of p lead to pathological
behaviors, however.
The Pfaffian PfM(A) in (5.1.11) is real in the Lorentzian version unlike in the Eu-
clidean version, where it becomes complex due to the replacement A0 = iA10 . However,
the phase factor eiSb causes the sign problem when one tries to investigate the Lorentzian
model by Monte Carlo methods. Here, we avoid this problem1 following previous work
[60, 88, 62] by rewriting the partition function (5.1.11) as
Z 1  1   1 
µν 2 p p 2 p
Z = dA PfM(A) δ TrFµν F − C δ Tr{(Ai ) } − 1 θ κ − Tr{(A0 ) } ,
N N N
(5.1.15)
where θ(x) is the Heaviside step function. This can be obtained by integrating out the
overall scale factor of the bosonic matrices Aµ first and using certain approximation
as discussed in section 5.3. The parameter C should be set to zero according to the
“derivation”, but we generalize the model by choosing C 6= 0, which allows us to obtain
results for larger matrices in the original C = 0 model by using smaller matrices [89, 88].
See Appendix B of ref. [88] for the details of the Monte Carlo simulation of the model
(5.1.15).

5.1.2 SSB of rotational SO(9) symmetry

Next we discuss how one can extract the time-evolution from a given matrix configuration
generated by Monte Carlo simulation [60]. Since the eigenvalues of the temporal matrix
A0 represents time, we work in an SU(N ) basis which diagonalizes A0 as

A0 = diag(α1 , . . . , αN ) , where α1 < · · · < αN . (5.1.16)

1
Strictly speaking, the model (5.1.15) is not completely free of sign-problem because the Pfaffian is
real but not positive semi-definite. However, configurations with positive Pfaffian dominates the path
integral (5.1.15) at large N , and therefore one can safely replace the Pfaffian by its absolute value in the
simulation.

56
In this basis, the spatial matrices Ai turn out to have an approximate band-diagonal
structure. By this, we mean that there exists2 some integer n such that the elements of
the spatial matrices (Ai )ab for |a − b| > n are much smaller than those for |a − b| < n.
Thanks to this structure, we can naturally consider the n × n submatrices Āi


Āi IJ
(t) ≡ (Ai )ν+I,ν+J (5.1.17)

representing the state at time t defined by

n
1X
t≡ αν+I , (5.1.18)
n I=1

where I, J = 1, . . . , n and ν = 0, 1, . . . , N − n. For example, we can define the extent of

the 9d space at time t using Āi (t) as
* 9 +
X1 2
R2 (t) =


tr Āi (t) , (5.1.19)
i=1
n

where the symbol “tr” represents a trace over the n × n submatrix. We can also define
the “moment of inertia tensor”
1

Tij (t) = tr Āi (t)Āj (t) , (5.1.20)
n
which is a 9 × 9 real symmetric tensor. The eigenvalues of Tij (t) represent the spatial
extent in each of the nine directions at time t, and we denote them by λi (t) with the
ordering
λ1 (t) > λ2 (t) > · · · > λ9 (t) . (5.1.21)
Note that R2 (t) and λi (t) are related as
9
X
2
R (t) = htr T i = hλi (t)i . (5.1.22)
i=1

The expectation values hλi (t)i can be used as the order parameters for the spontaneous
breaking of the rotational SO(9) symmetry of the model. If the nine eigenvalues do not
approach a common value in the large-N limit, we conclude that the SO(9) symmetry is
spontaneously broken. From the Monte Carlo simulations of the model (5.1.15), it was
found [60] that the three eigenvalues hλi (t)i (i = 1, 2, 3) start to grow with t after a critical
time tc , which implies that the SO(9) symmetry is spontaneously broken down to SO(3)
for t > tc . (See refs. [88, 62] for a precise definition of the critical time tc , which we use
in this work.)
2
In practice, the integer n can be determined by observing the scaling behavior for i |(Ai )ab |2 with
P

(a + b)/2 fixed to different values corresponding to different time slices. See section 5 of ref. [62] for the
details.

57
5.1.3 Expanding behaviors in the simplified models
It is interesting to investigate how the 3d space expands with time. For that, one clearly
needs to increase the matrix size, which is very time-consuming due to the existence of
the Pfaffian in (5.1.15). This led to the proposal of the simplified models, the VDM model
[88] and the bosonic model [62], which amounts to replacing the Pfaffian as

∆(α)16 for the VDM model ,
PfM(A) =⇒ (5.1.23)
1 for the bosonic model ,

where ∆(α) ≡ N
Q
a>b (αa −αb ) is the van der Monde (VDM) determinant. This replacement
reduces the computational cost from O(N 5 ) to O(N 3 ), which enables simulations with
considerably large matrix size. These two models are expected to describe the qualitative
behaviors of the original model at early times and at late times, respectively.
In both these models, the spontaneous breaking of the SO(9) rotational symmetry
to SO(3) was observed after some critical time as in the original model, and the rate of
expansion at late times was investigated. In the VDM model, the extent of space R(t)
defined in (5.1.19) exhibits an exponential growth [88]

R(t) ∼ eΛt , (5.1.24)

which is reminiscent of inflation3 , and this behavior does not seem to change with in-
creasing t. In the bosonic model, on the other hand, the exponential expansion observed
at early times changes into a power-law expansion [62]

R(t) ∼ t1/2 (5.1.25)

at later times, which is reminiscent of the Friedmann-Robertson-Walker Universe at the

radiation dominated era. Based on these results, it has been speculated that the extent
of space R(t) in the original model shows an exponential growth at early times and a
power-law expansion at later times. If true, it implies that the e-folding or the duration
of the cosmic inflation may be determined dynamically in the original model.

5.2 Space-time structure of the matrix configurations

In this section, we investigate the space-time structure of the matrix configurations gen-
erated by the Monte Carlo simulation of the model (5.1.15) and the simplified models
(5.1.23).

58
 800

2000 700
600

〈λi(t)〉 / R2(tc)
R(t)2 / R(tc)2

1500 500
400
1000
300
200
500
100
0 0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
(t-tc) / R(tc) (t-tc) / R(tc)

20000 100
18000
16000
50
〈ak(1)(t)〉 / R(tc)
〈qk(t)〉 / R2(tc)

14000
12000
10000 0
8000
6000
-50
4000
2000
0 -100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
(t-tc) / R(tc) (t-tc)/R(tc)

15 15

10 10
〈ak(1)(t)〉 / R(tc)

〈ak(4)(t)〉 / R(tc)

5 5

0 0

-5 -5

-10 -10

-15 -15
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
(t-tc) / R(tc) (t-tc) / R(tc)

Figure 5.1: The extent of space R2 (t)/R2 (tc ) (Top-Left) and the normalized eigenvalues
hλi (t)i/R2 (tc ) of Tij (t) (Top-Right) are plotted against time (t − tc )/R(tc ) for the bosonic
model with N = 256, C = 100, κ = 1, p = 1.5 and the block size n = 18. Similarly,
the eigenvalues of Q(t)/R2 (tc ) (Middle-Left), the eigenvalues of Ā(1) (t)/R(tc ) (Middle-
Right, Bottom-Left, the latter being the zoom-up version of the former), the eigenvalues
of Ā(4) (t)/R(tc ) (Bottom-Right) are plotted against time (t − tc )/R(tc ).

59
5.2.1 Results for the bosonic model
In this subsection, we consider the bosonic model, which is a simplified model for the late
time behaviors. Let us first look at the basic quantities such as the extent of space R2 (t)
and the eigenvalues hλi (t)i of Tij (t). In Fig. 5.1 we plot the extent of space R2 (t)/R2 (tc )
(Top-Left) and the normalized eigenvalues hλi (t)i/R2 (tc ) of Tij (t) (Top-Right) against
(t − tc )/R(tc ) for N = 256, C = 100, κ = 1.0 with the block size n = 18 in (5.1.19).
Here and for all the other plots in Fig. 5.1, we only present the results in the t < 0 region
since the results are symmetric4 under the time reflection t 7→ −t. The power p in the IR
cutoff (5.1.13) and (5.1.14) is chosen to be p = 1.5, which is found to be large enough to
make the results almost independent of p (See Appendix C.). Let us recall that R2 (t) is
related to hλi (t)i through (5.1.22). While the extent of space R2 (t)/R2 (tc ) grows with t
for t > tc , it is only three out of nine eigenvalues of Tij (t) that grow with t, which suggests
that the rotational SO(9) symmetry is broken spontaneously to SO(3). These results are
analogous to the previous results obtained for p = 1 [62].
The simplest way to probe the space-time structure is to define an n × n matrix
9
X
Q(t) ≡ (Āi (t))2 , (5.2.1)
i=1

which is invariant under SO(9) rotations. Let us denote its eigenvalues as qk (t) (k =
1, · · · , n) with the ordering
q1 (t) < · · · < qn (t) . (5.2.2)
These eigenvalues tell us how the space spreads in the radial direction at each time t.
In Fig. 5.1 (Middle-Left), we plot the eigenvalues qk (t)/R2 (tc ) against (t − tc )/R(tc ).
We find that the two largest eigenvalues grow with t, but not the others. Let us note that
the eigenvalues of Q(t) are related to the extent of space R2 (t) as
  * X n
+
1 1
R2 (t) = tr Q(t) = qk (t) . (5.2.3)
n n k=1

This implies that the time-dependence of R2 (t) seen in the Top-Left panel is caused only
by the two largest eigenvalues of Q(t).
Let us next discuss the space-time structure in the three extended directions and
the six shrunken directions separately. Since we are dealing with spontaneous symmetry
breaking, we need to choose the frame properly in order to distinguish these directions.
(i)
Suppose vj (t) (j = 1, · · · , 9) are the normalized eigenvectors of the “moment of inertia
3
This behavior was observed also in the original model [61] although the matrix size used was not
large enough to confirm the long-time behavior.
4
This does not mean that the Big Crunch occurs in this model because the time difference between
the symmetric point t = 0 and the critical time t = tc seems to diverge in physical units in an appropriate
large-N limit. See section 5.2.3.

60
tensor” (5.1.20) corresponding to the eigenvalues λi (t) with the ordering (5.1.21). Then,
we can define the n × n matrix corresponding to the spatial direction with the extent λi
as
X9
(i) (i)
Ā (t) = vj (t) Āj (t) (5.2.4)
j=1
(i)
and its eigenvalues ak (t) (k = 1, · · · , n) with the ordering
(i)
a1 (t) < · · · < a(i)
n (t) . (5.2.5)
(1)
In Fig. 5.1 (Middle-Right), we plot the eigenvalues ak (t)/R(tc ) against (t − tc )/R(tc ).
(1) (1)
We find that only two eigenvalues a1 (t) and an (t) grow in magnitude with time t, and
all the others remain close to zero. Similar behaviors are seen also for the eigenvalues
(2) (3)
ak (t) and ak (t) obtained for the other extended directions. In Fig. 5.1 (Bottom-Left),
we zoom up the same plot to make visible the eigenvalues close to zero. In Fig. 5.1
(4)
(Bottom-Right), we plot the eigenvalues ak (t)/R(tc ) against (t − tc )/R(tc ). We find
that all the eigenvalues remain close to zero. Similar behaviors are seen also for the
(5) (9)
eigenvalues ak (t), · · · , ak (t) obtained for the other shrunken directions. Comparing the
two plots at the bottom of Fig. 5.1, we notice that the eigenvalue distribution of Ā(i) is
almost identical for the extended directions and the shrunken directions except for the
two eigenvalues with large magnitude.
Similarly to (5.2.3), the eigenvalues of Ā(i) (t) are related to the extent of space λi (t)
in the ith direction as n
1 X  (i) 2
λi (t) = a (t) . (5.2.6)
n k=1 k
Our observation implies that the spontaneous symmetry breaking of the SO(9) rotational
symmetry seen in the Top-Right panel is caused only by the two eigenvalues of Ā(i) (t)
with large magnitude.

5.2.2 Including fermionic contributions

In order to seek for the possibility to obtain a regular space-time, we repeat the analysis
in the previous subsection in the case of the original model (5.1.15) including fermionic
contributions. Since the cost of Monte Carlo simulations increases from O(N 3 ) to O(N 5 ),
here we restrict ourselves to a rather small matrix size N = 16.
In Fig. 5.2 we plot the same quantities as in Fig. 5.1 for the original model with
N = 16, C = 3.91, κ = 0.38 and the block size n = 6. The power p in the IR cutoff
(5.1.13) and (5.1.14) is chosen to be p = 1.6, which is found to be large enough to make
the results almost independent of p (See Appendix C.). These results are qualitatively the
same as those obtained for the bosonic model. While the fermionic matrices are expected
to play an important role in the properties of the model such as the expanding behavior,
they do not seem to affect the singular space-time structure.

61
 4
14
3.5
12
3

〈λi(t)〉 / R2(tc)
R2(t) / R2(tc)

10
2.5
8
2
6 1.5
4 1
2 0.5
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(t-tc) / R(tc) (t-tc) / R(tc)

40
35 4

30
〈ak(1)(t)〉 / R(tc)
〈qk(t)〉 / R2(tc)

2
25
20 0
15
-2
10
5 -4
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(t-tc) / R(tc) (t-tc) / R(tc)

4
〈ak(4)(t)〉 / R(tc)

-2

-4

0 0.2 0.4 0.6 0.8 1

(t-tc) / R(tc)

Figure 5.2: The extent of space R2 (t)/R2 (tc ) (Top-Left) and the normalized eigenvalues
hλi (t)i/R2 (tc ) of Tij (t) (Top-Right) are plotted against time (t − tc )/R(tc ) for the original
model with N = 16, C = 3.91, κ = 0.38, p = 1.6 and the block size n = 6. Similarly, the
eigenvalues of Q(t)/R2 (tc ) (Middle-Left), the eigenvalues of A(1) (t)/R(tc ) (Middle-Right)
and the eigenvalues of A(4) (t)/R(tc ) (Bottom) are plotted against time (t − tc )/R(tc ).

62
 N C κ n ∆ 
64 8.81 0.14 24 1.0990(16) 0.0550(1)
96 0 2.00 14 1.3811(41) 0.1151(3)
64 0 2.00 10 1.2726(63) 0.1591(8)
64 0 4.00 7 1.3762(87) 0.2752(17)

Table 5.1: The parameter sets (N , C, κ) used for the simulation of the VDM model are
listed. We also present the block size n, the “volume” ∆ and the “lattice spacing” 
determined from the data for each parameter set.

5.2.3 Taking the continuum limit

As yet another possibility to obtain a regular space-time, let us consider taking the con-
tinuum limit. Here we use the VDM model, which is a simplified model for the early time
behaviors. In Fig. 5.3 (Top-Left), we plot the extent of space R2 (t)/R2 (tc ) against time
(t − tc )/R(tc ) for various N , C and κ with the block size n listed in table 5.1. The power
p in the IR cutoff (5.1.13) and (5.1.14) is chosen as p = 1.4 following ref. [87]. From this
plot, we observe a clear scaling behavior for (t − tc )/R(tc ) . 0.40.
In Fig. 5.3 (Top-Right), we plot the normalized eigenvalues hλi (t)i/R2 (tc ) of Tij (t) for
the VDM model with N = 96, C = 0 and κ = 2. Similar behaviors are obtained for the
other parameter sets. We find that three out of nine eigenvalues of Tij (t) grow with time,
which suggests that the rotational SO(9) symmetry is broken spontaneously to SO(3) for
t > tc . These results are similar to those obtained in refs. [88, 87].
In order to discuss the continuum limit, let us define the “volume” and the “lattice
spacing” in the temporal direction as [88]
tpeak − tc ∆
∆≡ , ≡ , (5.2.7)
R (tc ) ν

where tpeak represents the position of the peak in R2 (t) and ν is the number of data points
of R2 (t) contained within tc < t ≤ tpeak . Roughly speaking, the lattice spacing  represents
the average horizontal spacing between the adjacent data points of R2 (t)/R2 (tc ). In
table 5.1, we present the volume ∆ and the lattice spacing  obtained for each parameter
set (N, C, κ) used in Fig. 5.3. The deviation from the scaling behavior for (t − tc )/R(tc ) >
0.40 seen in Fig. 5.3 can be understood either as the finite volume effects or as the finite
lattice spacing effects depending on the parameter set.
In what follows, we focus on the point (t − tc )/R(tc ) ∼ 0.40, at which the results for
R (t)/R2 (tc ) with the four parameter sets agree with each other. In Fig. 5.3 (Middle-
2

Left), we plot the normalized eigenvalues hqk (t)i/R2 (tc ) (k = 1, · · · , n) of Q(t) against
their label (k −1)/(n−1) for the four parameter sets. This reveals a clear scaling behavior
except for the two largest eigenvalues, which grow as the lattice spacing  decreases. Note
that the time dependence of R2 (t)/R2 (tc ) is caused by the two largest eigenvalues of Q(t)

63
 4
N=64, C=8.81, κ=0.14 1.4
3.5 N=96, C=0, κ=2
N=64, C=0, κ=2 1.2
3 N=64, C=0, κ=4

〈λi(t)〉 / R2(tc)
R2(t) / R2(tc)

1
2.5
0.8
2
1.5 0.6

1 0.4

0.5 0.2

0 0
-2 -1.5 -1 -0.5 0 0.5 1 1.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5
(t-tc) / R(tc) (t-tc) / R(tc)

8 1.5
N=64, C=8.81, κ=0.14 N=64, C=8.81, κ=0.14
7 N=96, C=0, κ=2 N=96, C=0, κ=2
N=64, C=0, κ=2 1 N=64, C=0, κ=2
6 N=64, C=0, κ=4 〈ak(1)(t)〉 / R(tc) N=64, C=0, κ=4
〈qk(t)〉 / R2(tc)

0.5
5
4 0
3
-0.5
2
-1
1
0 -1.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
(k-1) / (n-1) (k-1) / (n-1)

1.5
N=64, C=8.81, κ=0.14
N=96, C=0, κ=2
1 N=64, C=0, κ=2
N=64, C=0, κ=4
〈ak(4)(t)〉 / R(tc)

0.5

-0.5

-1

-1.5
0 0.2 0.4 0.6 0.8 1
(k-1) / (n-1)

Figure 5.3: (Top-Left) The extent of space R2 (t)/R2 (tc ) is plotted against time (t −
tc )/R(tc ) for the VDM model with the parameter sets (N , C, κ) and the block size n
listed in table 5.1. The power p in the IR cutoff (5.1.13) and (5.1.14) is chosen as p = 1.4.
(Top-Right) The normalized eigenvalues hλi (t)i/R2 (tc ) of Tij (t) are plotted against time
(t−tc )/R(tc ) for N = 96, C = 0, κ = 2. The eigenvalues of Q(t)/R2 (tc ) (Middle-Left), the
eigenvalues of Ā(1) (t)/R(tc ) (Middle-Right) and the eigenvalues of Ā(4) (t)/R(tc ) (Bottom)
obtained at (t − tc )/R(tc ) ∼ 0.40 are plotted against their labels (k − 1)/(n − 1) for the
four parameter sets listed in table 5.1.

64
 0.6 0.5

0.5

n
0.4

(〈an(1)(t)〉 / R(tc)) / √
(〈qn(t)〉 / R2(tc)) / n

0.4
0.3
0.3
0.2
0.2

0.1
0.1

0 0
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2
1/n 1/n

Figure 5.4: (Left) The largest eigenvalue qn (t) of the matrix Q(t) obtained at (t −
tc )/R(tc ) ∼ 0.40 and normalized by R2 (tc ) and n is plotted against 1/n. (Right) The
(1)
largest eigenvalue an (t) of the matrix Ā(1) (t) obtained at (t − tc )/R(tc ) ∼ 0.40 and
√
normalized by R(tc ) and n is plotted against 1/n.

as we have seen in the previous subsections. Therefore, the scaling of R2 (t)/R2 (tc ) implies
that the two largest eigenvalues of Q(t) should grow linearly in n in the continuum limit.
This is confirmed numerically in Fig. 5.4 (Left) assuming the presence of 1/n corrections.
Let us next consider the space-time structure in the extended directions and the
shrunken directions separately. In Fig. 5.3 (Middle-Right), we plot the eigenvalues of
Ā(1) (t)/R(tc ) obtained at (t − tc )/R(tc ) ≈ 0.40 against the label (k − 1)/(n − 1). Here
again we observe a clear scaling behavior except for the ones at both ends of the spectrum.
Similar behaviors are obtained for the other extended directions. According to the same
argument as in the previous paragraph, we can deduce that the normalized eigenvalues
√
at both ends of the spectrum grow in magnitude as O( n) in the continuum limit, which
is confirmed in Fig. 5.4 (Right) assuming the presence of 1/n corrections.
In Fig. 5.3 (Bottom), we plot the eigenvalues of Ā(4) (t)/R(tc ) obtained at (t−tc )/R(tc ) ≈
0.40 against the label (k − 1)/(n − 1). We observe a clear scaling behavior here as well. In
fact, the eigenvalues are almost the same as those for the extended directions except for
the ones at both ends. Similar behaviors are obtained for the other shrunken directions.
Thus we find in the VDM model that the singular space-time structure becomes even
more pronounced in the continuum limit instead of getting milder. It is surprising that
the two eigenvalues of Ā(i) (t)/R(tc ) (i = 1, 2, 3 ) actually diverges in the continuum
limit although the extent of space defined by R2 (t)/R2 (tc ) remains finite. It is these
two eigenvalues that cause the spontaneous breaking of the SO(9) rotational symmetry
and the expansion of space. All the other eigenvalues of Ā(i) (t)/R(tc ) remain finite and
contribute only to the time-independent SO(9) symmetric part of the “moment of inertia
tensor” Tij (t).

65
5.2.4 The Pauli-matrix structure
In this subsection, we provide deeper understanding of the singular space-time structure
observed in the previous subsections. Let us work in the SU(n) basis which diagonalizes
Q(t) at each time t with the ordering (5.2.2), and consider the 2 × 2 submatrix Xi (t) in
the bottom-right corner of !
∗ ∗
Ā(i) (t) = (5.2.8)
∗ Xi (t)
for the extended directions i = 1, 2, 3. Here we use the VDM model with the parameter
sets given in table 5.1 and take the continuum limit focusing on the time (t − tc )/R(tc ) ≈
0.40 as we did in section 5.2.3.
We show below that the three matrices Xi in (5.2.8) tend to satisfy the SU(2) Lie
algebra
[Xi , Xj ] = icijk Xk (5.2.9)
for some real constant c in the continuum limit. In order to determine the optimal value
of c, we consider a quantity

S(c) ≡ tr(iijk [Xi , Xj ] + 2cXk )2 , (5.2.10)

which represents the violation of the relation (5.2.9). The value of c that minimizes S(c)
can be readily obtained as
iijk tr(Xk [Xi , Xj ])
c̃ = − . (5.2.11)
2tr(Xl2 )
Using c = c̃ as the optimal value for each configuration, we investigate to what extent the
relation (5.2.9) is satisfied.
In Fig. 5.5, we show a scatter plot for the real part (Left) and the imaginary part
(Right) of each side of (5.2.9). The quantities on both sides are normalized by tr(Xl2 ) so
that they become invariant under the scale transformation Xi 7→ const.Xi . We observe
that the data points tend to converge to the line y = x as one goes from the top to the
bottom corresponding to decreasing the lattice spacing  (See table 5.1.). This shows that
the 2 × 2 matrices Xi (i = 1, 2, 3) tend to satisfy (5.2.9) in the continuum limit.
Thus we conclude that the singular space-time structure observed for the matrix con-
figurations generated by simulations is essentially associated with the Pauli matrices. The
Pauli matrices may be regarded as the simplest matrix configuration that has SO(3) sym-
metry in the sense that their SO(3) rotation can be absorbed by an appropriate SU(N )
transformation. Given the situation characterized by the two large eigenvalues of Q(t),
the appearance of the Pauli-matrix structure may not be that surprising.

66
 real part imaginary part
0.4 N=64, C=0, κ=4 0.4 N=64, C=0, κ=4

0.2 0.2

0 0
y

y
-0.2 -0.2

-0.4 -0.4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
x x

real part imaginary part

0.4 N=64, C=0, κ=2 0.4 N=64, C=0, κ=2

0.2 0.2

0 0
y

y
-0.2 -0.2

-0.4 -0.4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
x x

real part imaginary part

0.4 N=96, C=0, κ=2 0.4 N=96, C=0, κ=2

0.2 0.2

0 0
y

-0.2 -0.2

-0.4 -0.4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
x x

real part imaginary part

0.4 N=64, C=8.81, κ=4 0.4 N=64, C=8.81, κ=4

0.2 0.2

0 0
y

-0.2 -0.2

-0.4 -0.4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
x x

Figure 5.5: (Left) A scatter plot for the real part x = Re(ic̃ijk (Xk )ab )/tr(Xl2 ) and
y = Re([Xi , Xj ]ab )/tr(Xl2 ) of each side of (5.2.9) with (5.2.11) is shown for (i, j) =
(1, 2), (2, 3), (3, 1) and (a, b) = (1, 1), (1, 2), (2, 2) using 10 configurations obtained by simu-
lating the VDM model with the parameter sets given in table 5.1. The solid line represents
y = x. (Right) A scatter plot for the imaginary part x = Im(ic̃ijk (Xk )ab )/tr(Xl2 ) and
y = Im([Xi , Xj ]ab )/tr(Xl2 ) of each side of (5.2.9) with (5.2.11) is shown in the same way.

67
5.3 The new interpretation of the simulation
In this section, we attribute the observed Pauli-matrix structure to the approximation
involved in deriving the partition function (5.1.15), which was used in Monte Carlo simu-
lation. We point out a subtlety in the approximation, and argue that the approximation
amounts to replacing eiSb by eβSb in the original partition function (5.1.11). This new
interpretation of the simulation provides us with a natural understanding of the (3+1)d
expanding behavior with the Pauli-matrix structure discussed in section 5.2. We also
speculate on a possible scenario for the original model with the correct eiSb factor.

5.3.1 The “derivation” of the partition function (5.1.15)

Let us first review how one can obtain the partition function (5.1.15) used in Monte Carlo
simulation from the original partition function (5.1.11). (This was done in Appendix A
of ref. [88] for p = 1, but here we generalize it to arbitrary p.)
Note that the integrand of the partition function (5.1.11) involves a phase factor eiSb .
As is commonly done in integrating oscillating functions, we introduce the convergence
factor e−|Sb | and take the  → 0 limit after the integration.
The partition function can then be rewritten as
Z L2p    
1  1
Z
2 p
Z = dA p
d(r ) δ Tr (Ai ) p p p
− r θ κ r − Tr (A0 ) 2p
eiSb −|Sb | PfM ,
0 N N
(5.3.1)
where κ and L are the cutoff parameters introduced in (5.1.13) and (5.1.14), respectively.
Rescaling the variables Aµ 7→ r1/2 Aµ in the integrand, we get
   
1 1
Z
2 p p 2p
Z = dA PfM(A) f (Sb ) δ Tr {(Ai ) } − 1 θ κ − Tr (A0 ) . (5.3.2)
N N
Here we have defined the function f (Sb ) by
Z L2p
2 2
f (Sb ) ≡ d(rp ) r9(N −1)−1 er (iSb −|Sb |) , (5.3.3)
0

which is a complex-valued function with the property f (−Sb ) = f (Sb )∗ .

For |Sb |  L14 , the function can be well approximated by
p 2 9(N 2 −1)+p−1
f (Sb ) ≈ (L ) . (5.3.4)
9(N 2 − 1) + p − 1
For |Sb | & L14 , on the other hand, the phase of the integrand in (5.3.3) starts to oscillate
p
violently in the region r & 1/ |Sb |, and hence the integral decreases rapidly in magnitude
for increasing |Sb |. In particular, the asymptotic behavior of f (Sb ) for Sb  L14 can be
estimated as
 9(N 2 −1)+p−1
|f (Sb )| 9(N 2 − 1) + p + 1
 
1 2
4
=Γ + O(e−L |Sb | ) (5.3.5)
f (0) 2 L |Sb |
4

68
by deforming the integration contour in (5.3.3).
Recalling eq. (5.1.12), the condition |Sb |  L14 for (5.3.4) can be rewritten as

1 4
Tr (Fµν F µν )  . (5.3.6)
N N L4

Therefore, assuming that the right-hand side N4L4 of (5.3.6) becomes small at large N , we
may make a replacement
 
1 µν
f (Sb ) =⇒ δ Tr (Fµν F ) (5.3.7)
N
up to a normalization constant. For the bosonic model and the VDM model, one simply
has to replace the Pfaffian in (5.3.1) and (5.3.2) as (5.1.23).

5.3.2 Subtlety in the derivation and the new interpretation

The only step in the derivation that may go wrong is the replacement (5.3.7). The subtlety
in this replacement can be seen as follows. Note that the phase factor eiSb in the partition
function (5.1.11) favors configurations at which the bosonic action Sb is stationary. On
the other hand, the above approximation essentially replaces the phase factor eiSb by the
delta function δ(Sb ), which amounts to picking up configurations at which Sb is stationary
only under rescaling Aµ 7→ const.Aµ . While it is true that |f (Sb )| is sharply peaked at
Sb = 0, the function f (Sb ) is actually a complex-valued function, whose phase rotates
violently around Sb = 0. This effect of the phase should be responsible for favoring the
configurations at which Sb is stationary. The approximation ignores this effect completely,
and hence it cannot be justified.
If the model (5.1.15) is not equivalent to the original model (5.1.11), what kind of
model does it actually correspond to? Here we point out that the constraint on Sb that
appears in (5.1.15) may be regarded as the constraint one uses in defining a microcanonical
ensemble. From this viewpoint, we consider that the model (5.1.15) is actually equiva-
lent to the corresponding canonical ensemble with the Boltzmann weight eβSb . The real
parameter β depends on the parameter C in the constraint5 . As we will see below, we
consider that the model (5.1.15) corresponds essentially to replacing eiSb by eβSb with
β > 0.
For β > 0, the first term in (5.1.12) that appears in eβSb favors configurations in which
A0 and Ai commute. This means that the spatial matrices Ai tend to become diagonal in
the SU(N ) basis which diagonalizes A0 . On the other hand, the second term in (5.1.12)
favors configurations in which the noncommutativity among the spatial matrices Ai is
large. The band-diagonal structure, which plays a crucial role in extracting the real-time
5
This connection also provides clear justification of the renormalization-group-like method [88, 89],
which amounts to tuning the parameter C in order to obtain the late-time behaviors with smaller matrix
size.

69
evolution as in section 5.1.2, can be understood as a consequence of the balance of these
two effects.
We can also understand the reason for the (3+1)d expanding behavior with the Pauli-
matrix structure. Here we assume that the first term in (5.1.12) is not important except
in realizing the band-diagonal structure and focus on the effect of the second term in
(5.1.12), which favors large Tr (Fij )2 , where Fij = i [Ai , Aj ]. We also have to take into
p
account the constraint N1 Tr (Ai )2 = 1, where we set p = 1 in what follows.


Simplifying the band-diagonal structure of the spatial matrices Ai (i = 1, · · · , 9), we

consider the block-diagonal structure given as
 (1) 
Āi
(2)
Āi
 
 
Ai = 

..
 , (5.3.8)
 . 

(B)
Āi
where n is the common block size and B is the number of blocks satisfying N = nB.
Within this ansatz, we would like to maximize Tr (Fij )2 under the constraint N1 Tr (Ai )2 =
1. Note that we have
B
1 2 1 X1 (b)
Tr (Ai ) = Tr (Āi )2 , (5.3.9)
N B b=1 n
B
1 1 X1 (b)
Tr (Fij )2 = Tr (F̄ij )2 , (5.3.10)
N B b=1 n
(b) (b) (b)
where we have defined F̄ij = i[Āi , Āj ] for each block b.
Let us solve the maximization problem in two steps. First we fix
1 (b)
Tr (Āi )2 = (rb )2 , (5.3.11)
n
B
1 X
(rb )2 = 1 , (5.3.12)
B b=1

and maximize Tr (Fij )2 under this constraint. Following the discussion given in ref. [60],
the solution to this first maximization problem can be written in terms of the Pauli
matrices σi as
(b) 1
Āi = √ rb (σi ⊕ 0n−2 ) , (5.3.13)
6
(b)
for i = 1, 2, 3 and Āi = 0 otherwise, up to the symmetries of the problem such as the
SO(9) rotational symmetry and the SU(n) symmetry within each block. The value of
Tr (Fij )2 for (5.3.13) is given as
B
2X
Tr (Fij )2 = (rb )4 . (5.3.14)
3 b=1

70
As the second step of the maximization, we maximize (5.3.14) under the constraint
(5.3.12). The maximum is given when all but one of the rb ’s are zero.
In reality, one should also take into account the entropic factor due to quantum fluc-
tuations, which is expected to favor certain distribution of rb . Due to the time-reversal
symmetry A0 7→ −A0 of the model, the most natural distribution would be that rb is
large around t = 0 and decreases with |t|. Thus we can understand the appearance of the
(3+1)d expanding behavior with the Pauli-matrix structure.

71
Chapter 6

Complex Langevin simulation of the

Lorentzian type IIB matrix model

We use the CLM to overcome the sign problem instead of the approximation which was
used so far. Since the sign problem is severe for the Lorentzian model, we generalize the
model by introducing two parameters which are related to the Wick rotation on the world
sheet and that in the target space. We apply the CLM to the generalized model.

6.1 Generalization of the model

The partition function of the generalized model is given by
   
1 1
Z
−S̃ 2 2
Z = dAdψe δ Tr (A0 ) − κ δ Tr (Ai ) − 1 , (6.1.1)
N N

S̃ = S˜b + S̃f , (6.1.2)

where β = 1/(g 2 N ). S˜b and S̃f are the bosonic and fermionic part of the action respec-
tively. The bosonic part and the fermionic part are defined as
 
˜ isπ/2 1 −ikπ 2 1 2
Sb = −iN βe − e Tr (F0i ) + Tr (Fij ) ,
2 4
  (6.1.3)
isπ/2 1 −ikπ/2 0 1 i
S̃f = −iN βe − e ψ̄Γ [A0 , ψ] + ψ̄Γ [Ai , ψ] ,
2 2

where Fµν = i[Aµ , Aν ], eisπ/2 comes from the Wick rotation on the world sheet, and e−ikπ/2
comes from the Wick rotation in the target space as A0 → e−ikπ/2 A0 .
After the path integration for the fermionic matrices, the partition function is rewritten
as
   
1 1
Z
−S̃eff 2 2
Z = dAdψe δ Tr (A0 ) − κ δ Tr (Ai ) − 1 ,
N N (6.1.4)
isπ/2 ˜ −ikπ/2
S̃eff = −iN βe Sb − log PfM(e A0 , Ai ),

72
where M(e−ikπ/2 A0 , Ai ) is the Dirac operator which acts on the fermionic matrices ψ.
The Pfaffian PfM is complex in general.
The Lorentzian model corresponds to setting (s, k) = (0, 0), whereas the Euclidean
model corresponds to setting (s, k) = (1, 1). Note that the approximation (5.3.7) corre-
sponds to setting (s, k) = (−1, 0) where the sign problem is absent.

6.2 Complex Langevin simulation of the model

In the CLM, since it is difficult to impose the cutoffs by using the usual way, we choose
an alternative way to impose the cutoffs which are explained in 6.2.1. In our simulations,
we choose an SU(N ) basis as (5.1.16) which diagonalizes the temporal matrix. Therefore,
in 6.2.2, we explain how to realize the order of diagonal elements of the temporal matrix
in the simulations. At the end of this section 6.2.3, we explain the application of CLM to
the generalized model.

6.2.1 Treatment of the cutoffs

In the previous works, quadratic functions are used to impose two cutoffs as
   
1 1
δ 2
Tr (A0 ) − κ δ Tr (Ai ) − 1 → e−Spot ,
2
N N (6.2.1)
1 1 1 1
Spot = cs ( Tr (A0 )2 − κ)2 + ct ( Tr (Ai )2 − 1)2 ,
2 N 2 N
where cs and ct are the some constants. In order to impose the cutoffs correctly, we
need to use very large coefficients for the quadratic functions. However, these very large
coefficients sometimes cause very large drift terms. If these large drifts appear frequently
in the complex Langevin simulations, the criterion for the correct convergence is not
satisfied. Therefore, we treat the cutoffs by using another way.
Actually, by the change of variables, we can impose the cutoffs correctly. First, we
introduce two auxiliary variables u and v, and rewrite the partition function as
Z ∞ Z
Z= dudv dAdψup v q e−f (u)−g(v) e−S̃eff (A0 ,Ai )
0
    (6.2.2)
1 2 1 2
×δ Tr (A0 ) − κ δ Tr (Ai ) − 1 ,
N N

where p and q are some positive constants, and f (u) and g(v) should be chosen so that
the integrations for u and v converge.
Next, we make a change the variables as

√
r
u
X0 = A0 , Xi = vAi . (6.2.3)
κ

73
The measure becomes
  21 N 2
u 1 2
dX = v 2 (D−1)N dA, (6.2.4)
κ
and the constraints become
   
1 2 u 1 2
δ Tr (A0 ) − κ = δ Tr (X0 ) − u ,
N κ N
    (6.2.5)
1 2 1 2
δ Tr (Ai ) − 1 = v δ Tr (Xi ) − v .
N N

When we choose p = 12 N 2 − 1 and q = 21 (D − 1)N 2 − 1 the exponents of u and v vanish.

Therefore the partition function becomes
Z ∞ Z
1 2
√ √
Z= dudv dXdψκ 2 N −1 e−f (u)−g(v) e−S̃eff ( κ/uX0 , 1/vXi )
0
    (6.2.6)
1 2 1 2
×δ Tr (X0 ) − u δ Tr (Xi ) − v .
N N
After Integrations of u and v, we obtain
Z
Z = dXdψe−S ,
√ !
(6.2.7)
κX0 Xi 1  1 
S = S̃ q ,q +f Tr (X0 )2 + g Tr (Xi )2
1
Tr (X0 )2 1
Tr (Xi )2 N N
N N

where we neglect an irrelevant overall factor. In this work, we choose f (x) = g(x) = 12 N 2 x
because we can obtain the analytic results for this choice. We will show the analytic results
later.

6.2.2 A way to realize the order of diagonal elements of the

temporal matrix.
We use SU(N ) symmetry to diagonalize the temporal matrix X0 as

X0 = diag(x1 , x2 , ..., xN ), where x1 < x2 < · · · < xN . (6.2.8)

By using “the gauge fixing”, we can rewrite the partition function as

N
Z Y Z
Z= dxa ∆(x) 2
dAi e−S ,
a=1
N
(6.2.9)
Y
∆(x) = (xa − xb ),
a>b

where ∆(x) is the van der Monde determinant.

74
 In order to to realize the order (6.2.8), we use a method which is proposed in [90]. We
perform the change of variable by introducing new variables τa as
N
X −1
τ1 τ1 τ2
x1 = 0, x2 = e , x3 = e + e , ..., xN = eτa , (6.2.10)
a=1

where we use the shift symmetry1 to setting x1 = 0. By treating the τa as dynamical

variables, the order of the diagonal elements of X0 is automatically realized. The effective
action becomes
 
 2 2 
−i π2 (1−s)
 1 −ikπ κ Tr [X0 , Xi ] 1 Tr [Xi , Xj ] 
Seff = N βe e  2 −
2 4 1 Tr (Xi )2 2 


1 1 2

N
Tr X̃ 0 N
Tr (X i ) N
(6.2.11)
N −1
1   2 1 X
+ N Tr X̃0 + N Tr (Xi )2 − 2 log ∆(x) − τa ,
2 2 a=1

where the last term comes from the change of the variables (6.2.10), and X̃0 = X0 −
1
N
Tr (X0 ). We can obtain the exact results for X0 and Xi as
 
1 2 1
Tr (X̃0 ) = 1 − 2 ,
N N
  (6.2.12)
1 2
Tr (Xi ) = D − 1.
N
These analytic results can be used as a validity check for complex Langevin simulations.
When we calculate observables, we should use Aµ which are obtained from Xµ by
rescaling as
√
κX̃0 Xi
A0 = q , Ai = q . (6.2.13)
1 1 2
N
Tr (X̃ 0 )2
N
Tr (X i )

6.2.3 Application of the CLM

First of all we extend the degrees of freedom of the dynamical variables which correspond
to Xi ∈ SU(N ) → Xi ∈ SL(N, C) and τa ∈ R → τa ∈ C. These variables are updated by
the following complex Langevin equations:
dτa ∂Seff
=− + ηa (tL ),
dtL ∂τa
(6.2.14)
d(Xi )ab ∂Seff
=− + (ηi )ab (tL ),
dtL ∂(Xi )ba
where tL is a fictitious time for the Langevin simulations, and η are Gaussian noises. The
drift terms ∂S eff
∂τa
∂Seff
and ∂(X i )ba
are defined first for the real variables τa and the Hermitian
1
The action (6.1.2) is invariant under the shift of A0 : A0 → A0 + c1, where c ∈ R is some constant.

75
matrices Xi , and then it is defined for the complexified variables τa and Xi by analytic
continuation in order to respect holomorphicity.
The drift terms are given explicitly as
∂Seff π κ
=βe−i 2 (1−s) e−ikπ eτa
∂τa L
( N
2N X X
− (xb − xc ) (Xi )bc (Xi )cb
K b=a+1 c6=b
N
! ) (6.2.15)
1 X 1 X X 2
+ 2 xb − xc (xd − xe ) (Xi )de (Xi )ed
K b=a+1 N c de
N X N
!
X 2 X 1 X
− e τa − 1 + N e τa xb − xc ,
b=a+1 c6=b
x b − x c
b=a+1
N c

∂Seff π κ
= − βe−i 2 (1−s) e−ikπ
∂ (Xi )ba K
( )
N Tr [X0 , Xj ]2
[X0 , [X0 , Xi ]]ab + (Xi )ab
L L2
( ) (6.2.16)
2
π N [X j , [X j , Xi ]] Tr [X j , Xk ]
+ βe−i 2 (1−s) ab
+ (Xi )ab
L2 L3
+ N (Xi )ab ,
where we define
N  X N 2
1 2 1 X 2 1
K = Tr (X̃0 ) = (xa ) − xa , (6.2.17)
N N a=1 N a=1
1
L= Tr (Xi )2 . (6.2.18)
N

6.2.4 Hermiticity of the extended matrices

In the simulations, we calculate a quantity that measures how the extended matrices Xi
are far from Hermitian matrices as an indication for the excursion problem. Here, we
define the Hermiticity norm as

−Tr (Xi − Xi† )2

H= , (6.2.19)
4Tr (Xi† Xi )

whose lower bound is 0 and the upper bound is 1. This quantity becomes 0 for Hermitian
matrices and 1 for anti-Hermitian matrices. Note that, in our simulation, since we fix the
gauge to diagonalize X0 , we cannot use the gauge cooling technique.

76
6.2.5 Adaptive step size
In order to avoid the excursion problem, we use the adaptive step size algorithm in the
simulations. We modify the Langevin step size depending on the magnitude for drift
terms as (
∆t0 for u < v0 ,
∆t = v0 (6.2.20)
∆t0 otherwise ,
u
where ∆t0 is a default step size, v0 is a threshold for the algorithm, and u is defined as

u = max (uτ , us ),
v
u 1 N
u −1 2
X ∂Seff
uτ = t 3 ,
N a=1 ∂τa (6.2.21)
v
u D−1 N X
N 2
u 1 XX ∂Seff
us = t .
N 3 (D − 1) i=1 a=1 b=1 ∂(Xi )ab

In this work, we set v0 = 10 because we found that the typical value is O(1).

6.3 Results
In this section, we will show the results for a simplified model and the original model.
Note that the observables which we will show later are snapshots obtained by the final
configuration for each simulation.

6.3.1 Results for the 10D bosonic model

Here, we study the bosonic model, which is a simplified model obtained by omitting the
fermionic part S̃f from the action (6.1.2). In the simulations for the bosonic model, we
use the 2nd-order RK algorithm to reduce the finite step size effects.
First we set k = 0 which means that the signature in the target space is the Lorentzian.
We find that the complex Langevin simulations starting from the Pauli-matrix configura-
tions which are obtained at s = −1 seem to be unstable at s 6= −1. In Fig. 6.1, we plot
the results for N = 32 case. The Hermiticity norm gradually increases and finally reaches
0.5. In this case, the drift histogram has a long tail, which means that the criterion for
the correct convergence of the CLM is not satisfied. In Fig. 6.2, we plot the results for
N = 256 case. Although the Hermiticity norm has not reached 0.5 so far, the norm is
gradually increasing. Moreover, there is no tendency for the norm to be stable at some
small value which is less than 0.5. It is possible that if one continues the simulation for
a long time, the norm finally reaches 0.5. Therefore, we expect that this simulation is
unstable.

77
 5
0.5 10

0.45

0.4 104

0.35

0.3 103
H

0.25

0.2 102

0.15

0.1 101

0.05

0 100
0 5 10 15 20 25 30 35 40 0 50 100 150 200 250 300 350 400
tL u

Figure 6.1: We plot the results for the bosonic model with N = 32, β = 2.5, κ = 0.4
and (s, k) = (−0.8, 0). The simulation starts from the Pauli-matrix configuration which
is obtained from a simulation at (s, k) = (−1, 0) with the same parameter setting. (Left)
A Hermiticity norm is plotted against the Langevin time. The norm gradually increases
and finally reaches 0.5. (Right) We plot a histogram of the drift terms. We can see a long
tail in the histogram.

0.16

0.14

0.12

0.1
H

0.08

0.06

0.04

0.02

0
0 20 40 60 80 100 120
tL

Figure 6.2: We plot the results for the bosonic model with N = 256, β = 2.5, κ = 1.0
and (s, k) = (−0.8, 0). The simulation starts from the Pauli-matrix configuration which
is obtained from a simulation at (s, k) = (−1, 0) with the same parameter setting. A
Hermiticity norm is plotted against the Langevin time. We cannot see a tendency for the
norm to be stable at some small value which is less than 0.5.

78
 6
0.035 10

0.03

0.025
105

0.02
H

0.015

104
0.01

0.005

0 103
0 1 2 3 4 5 6 7 8 9 10 0 0.5 1 1.5 2
tL u

0.4 0.65
Re[R22(t)]
Im[R (t)]
0.6
0.35

0.55
0.3
0.5

0.25
0.45

0.2
R (t)

0.4
qk(t)
2

0.15 0.35

0.3
0.1

0.25
0.05
0.2

0
0.15

-0.05 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.08

0.07

0.06

0.05
λi(t)

0.04

0.03

0.02

0.01
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 6.3: We plot the results for the bosonic model with N = 32, β = 2.5, κ = 0.8 and
(s, k) = (−0.8, 0). (Top-Left) A history of the Hermiticity norm where the horizontal axis
is the Langevin time. The Hermiticity seems to be controlled. (Top-Right) A histogram of
the drift term. This histogram falls off faster than exponential which means the criterion
for the correct convergence is satisfied. (Middle-Left) We plot the real and imaginary part
of R2 (t) against the time. (Middle-Right) The eigenvalues of Q(t) are plotted against time.
This plot shows that the space is continuous. (Bottom) We plot the eigenvalues of Tij (t)
which shows that there is no SSB of SO(9).

79
 On the other hand, when simulations start from scratch at s = −0.8, the CLM works
well. In Fig. 6.3 (Top-Left), we plot a history of the Hermiticity norm which is stable
for a long time at a small value. Therefore, the excursion problem does not occur in this
simulation. In Fig. 6.3 (Top-Right), we plot the histogram of the drift terms which falls
off very quickly. This means that the criterion for correct convergence is satisfied. We
show the extent of space R2 (t) whose definition is (5.1.19) in Fig. 6.3 (Middle-Left). The
space slightly expands as time proceeds. In Fig. 6.3 (Middle-Right) and (Bottom), we
also show the eigenvalues of the moment of inertia tensor Tij (t) defined by (5.1.20) and
the eigenvalus of Q(t) defined by (5.2.1). When we calculate Tij (t) and Q(t), instead of
Āi (t), we use Hermitian matrices ĀHerm
i (t) which are defined as
 
Herm 1 †
Āi (t) = Āi (t) + Āi (t) , (6.3.1)
2

where Ai are obtained by rescaling Xi as (6.2.13), and Āi (t) is the block matrices defined
like (5.1.17). Since the Hermiticity norm is very small, using these Hermitian matrices is
a good approximation. The spectrum for the eigenvalues of Q(t) looks very continuous,
therefore the space continuously spreads in the radial direction. However, the 9 eigenvalues
of Tij (t) are almost the same value, therefore the SO(9) symmetry is preserved. We also
found that this new phase appears at N = 128 (see Fig. 6.4).
We increase s to approach the Lorentzian model. In Fig. 6.5, we show the results for
various values of s. We found that the CLM works well even at the Lorentzian model.
We also found that the shape of R2 (t) changes from a bell-shaped curve to a parabola like
curve which we have seen in the classical solutions [91]. In Fig. 6.6, we plots the results
of the s = 0 case. At least, in this parameter choice (β = 2.5 and κ = 0.8), we cannot see
the SSB of SO(9) even if we reach the Lorentzian model.
It is possible that the reason for no SSB of the SO(9) is the parameter choice for β and
κ. When we decrease β or κ, although the criterion for correct convergence is satisfied,
the SSB does not occur. Therefore, we try to increase β or κ at N = 128. However, in
the large β or the large κ region, the criterion is not satisfied.
This no SSB behavior reminds us of results for the Euclidean bosonic model case in
which there is no SSB of SO(10) [59]. Therefore, we study whether this new phase is
smoothly connected to (s, k) = (1, 1) which corresponds to the Euclidean model with the
constraints2 .
In Fig. 6.7 and Fig. 6.8, we plot the results at k = s line which lineally connects the
Lorentzian case and the Euclidean case. We find that there are no qualitative differences
between them. Therefore this new phase which we found at (s, k) = (0, 0) is smoothly
connected to the Euclidean case.
2
Actually, the Euclidean model is well defined without the cutoffs. Therefore, strictly speaking, the
generalized model at (s, k) = (1, 1) is not the Euclidean type IIB matrix model.

80
 0.03
106

0.025
105

104
0.02
H

0.015
103
0.01

102
0.005

101
0
0 10 20 30 40 50 60 70 80 90 100
tL 1000.8 0.9 1 1.1 1.2 1.3 1.4
0.4 0.7
Re[R22(t)]
Im[R (t)]

0.6
0.3

0.5

0.2
R2(t)

qk(t)

0.4

0.1

0.3

0
0.2

-0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.048

0.046

0.044

0.042

0.04
λi(t)

0.038

0.036

0.034

0.032

0.03

0.028
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 6.4: We plot the results for the bosonic model with N = 128, β = 2.5, κ = 0.8 and
(s, k) = (−0.8, 0). (Top-Left) The history of the Hermiticity norm where the horizontal
axis is the Langevin time. The Hermiticity seems to be controlled. (Top-Right) The
histogram of the drift term. This histogram falls off faster than exponential which means
the criterion for the correct convergence is satisfied. (Middle-Left) We plot the real and
imaginary part of R2 (t) against the time. There is a little expansion. (Middle-Right) The
eigenvalues of Q(t) are plotted against time. This plot shows that the space is continuous.
(Bottom) We plot the eigenvalues of Tij (t) which shows that there is no SSB of SO(9).

81
 0.4 0.4
2 2
Re[R2(t)] Re[R2(t)]
Im[R (t)] Im[R (t)]

0.3 0.3

0.2 0.2
R (t)

R2(t)
2

0.1 0.1

0 0

-0.1 -0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.4 0.4
Re[R22(t)] Re[R22(t)]
Im[R (t)] Im[R (t)]

0.3 0.3

0.2 0.2
R2(t)

R2(t)

0.1 0.1

0 0

-0.1 -0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.5: We plot the R2 (t) against time for the bosonic model with N = 128, β =
2.5, κ = 0.8, and k = 0. (Top-Left), (Top-right), (Bottom-Left), and (Bottom-right)
correspond to s = −0.8, −0.6, −0.4, and 0 respectively.

82
 0.08
105
0.07

0.06
104

0.05

103
H

0.04

102
0.03

0.02

0.01 101

0
0 10 20 30 40 50 60
tL 1000.8 0.9 1 1.1 1.2 1.3 1.4
0.4 0.55
Re[R22(t)]
Im[R (t)]
0.5

0.3
0.45

0.4

0.2
0.35
R2(t)

qk(t)

0.3
0.1

0.25

0.2
0

0.15

-0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.042

0.04

0.038

0.036

0.034
λi(t)

0.032

0.03

0.028

0.026

0.024
-1.5 -1 -0.5 0 0.5 1 1.5

Figure 6.6: We plot the results for the bosonic model with N = 128, β = 2.5, κ = 0.8 and
(s, k) = (0, 0). (Top-Left) A history of the Hermiticity norm where the horizontal axis is
the Langevin time. The Hermiticity seems to be controlled. (Top-Right) A histogram of
the drift term. This histogram falls off faster than exponential which means the criterion
for the correct convergence is satisfied. (Middle-Left) We plot the real and imaginary part
of the R2 (t) against the time. There is a little expansion. (Middle-Right) The eigenvalues
of Q(t) are plotted against time. This plot shows that the space is continuous. (Bottom)
We plot the eigenvalues of Tij (t) which shows that there is no SSB of SO(9).

83
 0.5 0.5

0.45 0.45

0.4 0.4

0.35 0.35
qk(t)

qk(t)
0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.55 0.55

0.5 0.5

0.45 0.45

0.4 0.4

0.35 0.35
qk(t)

qk(t)

0.3 0.3

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.7: We plot the eigenvalues of Q(t) against time for the bosonic model with
N = 32, β = 1.4, and κ = 1.0. (Top-Left), (Top-right), (Bottom-Left), and (Bottom-
right) correspond to s = k = 0, 0.3, 0.7, and 1 respectively.

84
 0.065 0.07

0.06

0.06
0.055

0.05

0.05
0.045

0.04
λi(t)

λi(t)
0.04
0.035

0.03
0.03

0.025

0.02
0.02

0.015

0.01 0.01
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.07 0.07

0.06 0.06

0.05 0.05
λi(t)

λi(t)

0.04 0.04

0.03 0.03

0.02 0.02

0.01 0.01
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.8: We plot the eigenvalues of Tij (t) against time for the bosonic model with
N = 32, β = 1.4, and κ = 1.0. (Top-Left), (Top-right), (Bottom-Left), and (Bottom-
right) correspond to s = k = 0, 0.3, 0.7, and 1 respectively.

85
6.3.2 Effects from the fermionic part
In the Euclidean type IIB matrix model, the phase of the Pfaffian plays an essential role
in the SSB of the rotational SO(10) symmetry [59]. Therefore, it is expected that the
Pfaffian plays an important role also in the Lorentzian type IIB matrix model.
When some of the eigenvalues of the Dirac operator approach 0 the drift term becomes
very large. The appearance of such drift terms causes the failure of the CLM. This
problem is known as the singular drift problem. In order to avoid this problem, usually,
the deformation term is added to the action. The suitable deformation term depends on
the model, especially the spectrum of the Dirac operator.
Here we focus on the 6d type IIB matrix model due to the numerical cost. In the
6d model, after the integration for the fermionic matrices, the determinant of the Dirac
operator appears instead of Pfaffian which appears in the 10d case. In the Lorentzian
type IIB matrix model, the Dirac operator is defined as

M(e−ikπ/2 A0 , Ai ) := −Γ0 [e−ikπ/2 A0 , · ] + Γi [Ai , · ], (6.3.2)

where matrices Γµ are constructed from Pauli matrices σi as

Γ0 = 1 ⊗ 1, Γ1 = σ1 ⊗ σ2 , Γ2 = σ2 ⊗ σ2
(6.3.3)
Γ3 = σ3 ⊗ σ2 , Γ4 = 1 ⊗ σ2 , , Γ5 = 1 ⊗ σ3 .

When the matrices Aµ are Hermitian A†µ = Aµ this Dirac operator satisfies the following
relation:
M(e−ikπ/2 A0 , Ai ) = M† (e−ikπ/2 A0 , Ai ). (6.3.4)
Therefore the eigenvalues of the Dirac operator are real.
In this case, a suitable deformation term is

imf Γ0 ψ̄ψ (6.3.5)

where mf ∈ R is called as the deformation parameter. By adding this term to the action,
the eigenvalues of the Dirac operator shift to the imaginary direction. As a result, we
can avoid the singular drift problem. Since this term modifies the theory, taking mf → 0
limit is needed.
In Fig. 6.9 and Fig. 6.10, we plot the results at s = k line for the mf = 5 case. We
find that there is no qualitative difference between them. Moreover, we cannot see the
qualitative differences between the results obtained for the bosonic model and that for
the fermionic model. We expect that the reason is the large deformation parameter. We
plot the results for the mf = 2 and mf = 1 cases in Fig. 6.11, the results are almost the
same as the results for mf = 5 case. Therefore, we consider that mf = 1 is still large to
see the fermionic effects.
In Fig. 6.12, we plot the eigenvalues of the Dirac operator. We find that the average
of the imaginary part of the eigenvalues is close to the values of mf . Note that, since

86
 0.8 0.7

0.7 0.6

0.6 0.5

0.5 0.4
qk(t)

qk(t)
0.4 0.3

0.3 0.2

0.2 0.1

0.1 0
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.7 0.7

0.6 0.6

0.5 0.5
qk(t)

qk(t)

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.9: We plot the eigenvalues of Q(t) against time for the 6d model with N = 32
and mf = 5. (Top-Left), (Top-right), (Bottom-Left), and (Bottom-right) correspond to
s = k = 0, 0.3, 0.7, and 1 respectively.

87
 0.13 0.12

0.12 0.11

0.11
0.1

0.1
0.09

0.09
0.08
0.08
λi(t)

λi(t)
0.07
0.07
0.06
0.06

0.05
0.05

0.04
0.04

0.03 0.03

0.02 0.02
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.12 0.12

0.11 0.11

0.1 0.1

0.09 0.09

0.08 0.08
λi(t)

λi(t)

0.07 0.07

0.06 0.06

0.05 0.05

0.04 0.04

0.03 0.03

0.02 0.02
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.10: We plot the eigenvalues of Tij (t) against time for the 6d model with N = 32
and mf = 5. (Top-Left), (Top-right), (Bottom-Left), and (Bottom-right) correspond to
s = k = 0, 0.3, 0.7, and 1 respectively.

88
 0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5
qk(t)

qk(t)
0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

0.13 0.13

0.12 0.12

0.11 0.11

0.1 0.1

0.09 0.09

0.08 0.08
λi(t)

λi(t)

0.07 0.07

0.06 0.06

0.05 0.05

0.04 0.04

0.03 0.03

0.02 0.02
-1.5 -1 -0.5 0 0.5 1 1.5 -1.5 -1 -0.5 0 0.5 1 1.5

t t

Figure 6.11: The results for the 6d model with N = 32. We plot the eigenvalues of
Q(t) in (Top) and the eigenvalues of Tij (t) against time in (Bottom). (Left) and (Right)
correspond to mf = 2, and 1 respectively.

89
 2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
-6 -5 -4 -3 -2 -1 0 1 2 3 4 5 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5

0.2 0.25

0.18

0.16 0.2

0.14

0.12 0.15
H

H
0.1

0.08 0.1

0.06

0.04 0.05

0.02

0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
tL tL

Figure 6.12: The results for the 6d model with N = 32. (Top) The eigenvalues of
M(e−ikπ/2 A0 , Ai ) where the horizontal and vertical axes correspond to real and imaginary
part of the eigenvalues respectively. (Bottom) We plot the Hermiticity norm against the
Langevin time. (Left) and (Right) correspond to mf = 2, and 1 respectively.

the matrices Aµ are deviated from being Hermitian, the eigenvalues are distributed with
some width in the imaginary direction. Even at mf = 1, the near-zero eigenvalues do not
appear. Therefore it is possible to decrease mf more.

90
Chapter 7

Summary and discussions

In this thesis, we applied the CLM to gauge theories with a θ term and the Lorentzian
type IIB matrix model.

7.1 Gauge theories with a θ term

7.1.1 2D U(1) lattice gauge theory with a θ term
In this work, we have made an attempt to apply the CLM to gauge theories with a θ
term. As a first step, we applied the CLM to the 2D U(1) case, which is exactly solvable
on a finite lattice with various boundary conditions. We find that a naive implementation
of the method fails due to the topological nature of the θ term.
While the gauge configurations are complexified in the CLM, one can still define
the notion of topological sectors by Re Qlog ∈ Z. When a transition between different
topological sectors occurs, one of the plaquettes has to cross the branch cut inevitably,
which causes the appearance of large drift terms. This indeed happens at small β, where
we find that the criterion for correct convergence of the CLM is not satisfied. Increasing
β makes all the plaquettes close to unity. The large drift terms do not appear in this
case, and the criterion for correct convergence of the CLM is satisfied. However, the
topology change does not occur during the simulation and the ergodicity is violated.
This is analogous to the topology freezing problem, which is known to occur for θ = 0.
The results obtained in this case correspond to the expectation values for an ensemble
restricted to a particular topological sector specified by the initial configuration.
In order to avoid this problem, we have considered the punctured model, which can
be obtained by removing one plaquette from the action, both from the gauge action and
from the θ term. While the quantity Re Qlog is no more restricted to integer values, we
can still formally classify the complexified configurations into “topological sectors” by
adding back the contribution of the removed plaquette to Re Qlog . Even for large β, the
removed plaquette can cross the branch cut easily, which results in frequent transitions

91
between different “topological sectors”. Note also that, as far as β is sufficiently large, all
the other plaquettes are close to unity, and hence large drift terms do not appear. Thus
the criterion for correct convergence of the CLM can be satisfied by simply approaching
the continuum limit without causing the topology freezing problem. Indeed our results
obtained by the CLM for the punctured model reproduce the exact results even at large
θ.
In the case of the punctured model, the drift term from the θ term appears only for
θ
the link variables composing the removed plaquette, and it is given by ±i 2π , which causes
θ
∓∆t 2π
multiplication by a constant factor e to these link variables at each Langevin step.
The local unitarity norm of these link variables grows exponentially at early Langevin
times, but it saturates at some point to some constant, which increases exponentially for
large |θ|Vphys . We have seen that the CLM works perfectly even in this situation as far
as β is sufficiently large. This provides a counterexample to the common wisdom that
the CLM fails when the unitarity norm becomes large. Thus our results also give us new
insights into the method itself.
The punctured model is actually equivalent to the non-punctured model in the infinite
volume limit for |θ| < π. In that limit, the topological charge can take arbitrarily large
values, so the discretization of Q to integers is no more important. This equivalence has
been confirmed explicitly by obtaining exact results for the punctured model. In fact, the
exact results also reveal the absence of finite volume effects in the punctured model as
opposed to the non-punctured model, which exhibits sizable finite volume effects around
θ ∼ π. It is conceivable that the smearing of the topological charge somehow results in the
reduction of finite volume effects. If so, a similar conclusion should hold more generally.

7.1.2 4D SU(2) lattice gauge theory with a θ term

Since the application of the CLM to the 2D U(1) case was successful, we next applied
the CLM to the SU(2) gauge theory with a θ term. First, we naively applied the CLM
to the theory. At very small β, the topological charge freely changes during a simulation,
however, the criterion for the correct convergence of the CLM is not satisfied. In 4D
SU(2) case, unlike in the 2D U(1) case, there is a region for β where the simulations are
free from both the topology freezing problem and the wrong convergence. At β = 3.25,
the criterion is satisfied for |θ| ≤ 2π, however we cannot see the 2π periodicity which the
theory has. A possible reason for the absence of 2π periodicity is the UV fluctuations.
The fluctuations hinder the appearance of the comb-shaped distribution of the topological
charge, as a result, the 2π periodicity does not appear.
One of the ways to reduce the UV fluctuations is using the cooling procedure such
as the gradient flow. We applied the gradient flow for θ = 0 case, and then we saw
the appearance of the comb-shaped distribution of the topological charge. However, we
cannot use the gradient flow in the CLM because this procedure is cannot be justified in

92
the method. Another way to reduce the UV fluctuations is decreasing the lattice spacing
a, which corresponds to increasing β. However, we found that the topology freezing
problem occurs at large β.
It is known that the open boundary conditions alleviate the topology freezing problem,
however, as a drawback, the strong finite volume effects appear. Therefore, we tried to
solve the problem by imposing the open boundary condition for only one of the spatial
directions. At β = 6, we found that the topology freezing is alleviated, however, the
alleviation is not enough. Thus, we imposed the open boundary condition for all spatial
directions, and we found that the topology freezing is milder than that for the case of
open boundary for a spatial direction.
We apply the CLM to the theory with the open boundary condition for all spatial
directions and found that the criterion is satisfied for |θ| < 2π. In order to study phase
structure, we measured the Im Q which is an order parameter for the SSB of the CP
symmetry, and the result shows that the Im Q 6= 0 at θ = π in the deconfined region. It
seems that this result naively implies the CP symmetry is spontaneously broken there.
There are two possibilities. One is that critical temperatures satisfy Tdec (θ = π) < TCP ,
and the other is that we cannot see the CP restoration due to the finite volume effects
and the finite lattice spacing effects.
In order to study the finite volume effects, we performed the simulation for various
spatial volume up to L3s = 803 . We cannot see the finite volume scaling so far, therefore,
we need to increase Ls more to take the extrapolation for infinite volume limit. Since
the CLM works well at sufficiently large θ, we expect that the phase diagram can be
determined by the complex Langevin simulations if we take the infinite volume and the
continuum limits.

7.2 The Lorentzian type IIB matrix model

7.2.1 On the emergence of the space-time structure in the type
IIB matrix model
In this work, we have investigated the space-time structure of the matrix configurations
obtained in Monte Carlo studies of the Lorentzian type IIB matrix model and the sim-
plified models. In these models, the time-evolution can be extracted from the matrix
configurations by working in the SU(N ) basis which diagonalizes the temporal matrix A0 .
The n × n spatial submatrices Āi (t) (i = 1, · · · , 9) at each time t show that only three out
of nine directions expand after some critical time suggesting the SSB of rotational SO(9)
symmetry to SO(3). By calculating the eigenvalues of Āi (t) at each t, however, we have
found that only two of them increase in magnitude with t in the extended directions, while
the rest are independent of t and SO(9) symmetric. This implies that the SSB is caused

93
only by the two eigenvalues. In the continuum limit, the magnitude of the two eigenvalues
diverges in physical units and the spatial matrices Āi (t) approach a configuration which
is essentially described by the Pauli matrices.
We have attributed this problem to the approximation used in Monte Carlo simulation
to avoid the sign problem, which actually amounts to replacing eiSb by eβSb in the partition
function (5.1.11) of the Lorentzian type IIB matrix model. This new interpretation of the
Monte Carlo simulation enables us to understand the interesting aspects of the obtained
results such as the band-diagonal structure of the spatial matrices Ai as well as the
appearance of the (3+1)d expanding behavior with the Pauli-matrix structure.

7.2.2 Complex Langevin simulation of the Lorentzian type IIB

matrix model
In order to treat the weight eiSb appropriately, we applied the CLM instead of the ap-
proximation. We generalized the model by introducing two parameters s and k which
are related to the Wick rotation on the world sheet and that in target space respectively.
This generalized model interpolates among the Lorentzian case, the Euclidean case, and
a model with eβSb , which correspond to setting (s, k) = (0, 0), (1, 1), and (−1, 0) respec-
tively.
In the complex Langevin simulation, when we use quadratic functions for the IR
cutoffs, we need to use large coefficients, which sometimes results in the appearance of
large drift terms. If these large drift terms appear frequently, the criterion for the correct
convergence of the CLM is not satisfied. Therefore, we improved the treatment of the
cutoffs by changes of the variables (6.2.3).
We first applied the CLM to the bosonic version of the generalized model which is
obtained by neglecting the fermionic part in S̃. We found that, in (s, k) 6= (−1, 0) case, if
the simulation starts from a configuration obtained from a simulation at (s, k) = (−1, 0)
then the excursion problem occurs, as a result, the criterion for the correct convergence
of the CLM was not satisfied for small matrix size case (N = 32). In the N = 256
case, although the excursion has not been horrible so far, however, if one continues the
simulation for a long time the excursion problem might occur. In other words, the singular
structure phase which we observed in a model with the weight eβSb seems to be unstable
at the Lorentzian case.
On the other hand, when the complex Langevin simulation started from scratch the
CLM worked well and a new phase appeared. In a new phase, the space is continuous
but the spontaneously breaking of the SO(9) symmetry doesn’t occur in our preliminary
studies. The behaviors in the new phase remind us of the Euclidean bosonic case in which
the SO(9) does not break spontaneously. In order to clarify this point, we studied the
model on the s = k line, and found the phase in the Lorentzian case is smoothly connected
to that in the Euclidean case.

94
 According to the study of the Euclidean type IIB matrix model, the fermionic part
plays an important role in the SSB of SO(9). We expect that the same is true in the
Lorentzian case. Therefore we applied the CLM to see the 6D version of the generalized
model to study the effects of the fermion. When the action includes the fermionic part,
in order to avoid the singular drift problem, we add a “mass term” mf which is called
a deformation term to the action. We simulated the model at various values of the
deformation parameter, and we found that the values which we studied are too large to
see the fermionic effects. Fortunately, according to the spectrum of the eigenvalues of
the Dirac operator, it may be possible to decrease the value of the parameter more. We
expect that the SSB of the SO(9) occurs in the Lorentzian model after taking both the
large N and the mf → ∞ limit as in the case of the Euclidean model.

95
Acknowledgment
I would like to thank my supervisor Prof. Jun Nishimura for valuable discussions,
advices, encouragements, and supports. I also would like to my collaborators, Prof.
Asato Tsuchiya, Prof. Konstantinos N. Anagnostopoulos, Prof. Takehiro Azuma, Prof.
Masazumi Honda, Prof. Yuta Ito, Dr. Kohta Hatakeyama, Dr. Toshihiro Aoki, Dr.
Stratos Kovalkov Papadoudis, Mr. Akira Matsumoto, and Mr. Atis Yosprakob for
valuable discussions through collaborations. Numerical computation was carried out on
K computer (Project ID : hp170229, hp180178), Oakbridge-CX in University of Tokyo
(Project ID : hp120281, hp200106), TSUBASA in KEK (Program No. 2020-009), FX10
in University of Tokyo, the PC clusters in KEK Computing Research Center and KEK
Theory Center, and XC40 at YITP in Kyoto University.

96
Appendix A

Derivation of the exact result

In 2D lattice gauge theory, we can obtain the partition function explicitly on any manifold
at finite lattice spacing and finite volume [40], from which various observables can be
obtained. In this section, we review the derivation using the so-called K-functional [39].

A.1 The K-functional

Let us consider a lattice gauge theory with a θ term on a 2D lattice manifold M. Here
we take the gauge group to be U(N ), which is a generalization of U(1) considered so far.
Note that the topology of the gauge field becomes trivial for SU(N ) in 2D gauge theories.
As a building block for evaluating the partition function, we define the K-functional
KA for the region A ⊂ M defined by [39]
 
Z Y
KA (Γ) =  dUi  e−SA , (A.1.1)
Ui ∈A\C

where the integral goes over the link variables inside A leaving out those on the boundary
C. (See Fig. A.1.) The action SA in Eq. (A.1.1) is given by
X  β θ

−1


SA = Tr − Pi + Pi − log Pi , (A.1.2)
P ∈A
2 2π
i

where the sum goes over the plaquettes Pi included in the region A. Here we use the log
definition (3.1.12) of the topological charge, but the results for the sine definition (3.1.13)
can be obtained in a similar manner as we mention at the end of Section A.4.
The K-functional depends on the link variables on the boundary C = ∂A, but due to
the gauge invariance, it actually depends only on
Y
Γ= Ui , (A.1.3)
Ui ∈C

97
 A
C

Figure A.1: An example of the region A, which has a boundary C = ∂A. The K-functional
for this region is defined by integrating out the link variables represented by the dashed
lines. The result depends on Γ defined by (A.1.3) for the loop C represented by the solid
line with arrows.

which is a consecutive product of link variables along the loop C. The choice of the
starting point of the loop C does not matter since a different choice simply corresponds
to making a gauge transformation of Γ, which leaves the K-functional invariant.
We can calculate the K-functional for any A by gluing the K-functional for a single
plaquette P , which is nothing but
 
β −1

θ
K(P ) = exp Tr P +P + log P . (A.1.4)
2 2π
Note here that (A.1.4) is a function of the group element P ∈ U(N ), which is invariant
under
P → gP g −1 ; g ∈ U(N ) . (A.1.5)
It is known that any function having this property can be expressed by the so-called
character expansion
X
K(P ) = λr χr (P ) , (A.1.6)
r

which is analogous to the Fourier expansion for periodic functions. Here χr (P ) is the
group character, which is defined by the trace of P for an irreducible representation r,
and it satisfies the orthogonality relation
Z
dU χr1 (U −1 )χr2 (U ) = δr1 ,r2 . (A.1.7)

Using this relation, the coefficient λr in the expansion (A.1.6) can be readily obtained as
Z
λr = dU χr (U −1 )K(U ) . (A.1.8)

As an example, let us obtain the K-functional K2×1 for a 2 × 1 rectangle by gluing

two neighboring plaquettes P1 = U1 Ω and P2 = Ω−1 U2 as shown in Fig. A.2. The group

98
 U1 P1 Ω P2 U2

Figure A.2: The K-functional K2×1 for a 2 × 1 rectangle is obtained by considering the
K-functional for the two plaquettes P1 = U1 Ω and P2 = Ω−1 U2 , which are glued together
by integrating out the shared link variable Ω.

elements U1 and U2 are the products of three link variables, and Ω represents the link
variable shared by P1 and P2 . Integrating out the shared link variable Ω, we get
Z
K2×1 (U1 U2 ) = dΩ K(P1 )K(P2 )
X Z
= λr1 λr2 dΩ χr1 (U1 Ω)χr2 (Ω−1 U2 )
r1 ,r2
 2
X λr
= dr χr (U1 U2 ) , (A.1.9)
r
dr

where dr = χr (1) is the dimension of the representation r and we have used a formula
1
Z
dΩ χr1 (U1 Ω)χr2 (Ω−1 U2 ) = χr (U1 U2 )δr1 ,r2 . (A.1.10)
dr1 1
Iterating this procedure, we obtain the K-functional for any simply connected region A
as
X  λr |A|
KA (Γ) = dr χr (Γ) , (A.1.11)
r
d r

where |A| is the number of plaquettes in A, and Γ is defined by (A.1.3).

In the U(1) case, the representation can be labeled by the charge n ∈ Z, and the
dimension of the representation is dn = 1 for ∀n ∈ Z. Since the character for the plaquette
P = eiφ is given by χn (P ) = einφ , the K-functional for a single plaquette (A.1.6) reduces
to
+∞
X
K(P ) = λn einφ , (A.1.12)
n=−∞

where the coefficient λn is a function of θ and β given explicitly as

λn ≡ I(n, θ, β)
Z π
1
dφ e−inφ K P = eiφ


=
2π −π
Z π    
1 θ
= dφ exp β cos φ + i −n φ (A.1.13)
2π −π 2π

99
using (A.1.4) with P = eiφ . This function reduces to the modified Bessel function of the
first kind for θ = 0.
The character expansion in the U(N ) case is more complicated, so we only show the
end results referring the reader, for instance, to the appendix of Ref. [92] for the details.
The representation of the U(N ) group is labeled by N integers

ρ = (ρ1 , ρ2 , · · · , ρN ) ∈ ZN (A.1.14)

satisfying ρi ≥ ρi+1 , and the dimension of the representation ρ can be calculated by

N 
ρi − ρj
Y 
dρ = χρ (1) = 1− . (A.1.15)
i>j
i−j

The coefficient λρ in (A.1.6) that corresponds to the representation ρ is expressed as a

determinant
λρ = det M(ρ, θ, β) , (A.1.16)
where the matrix M(ρ, θ, β) is given as
Z π    
1 θ
Mjk (ρ, θ, β) = dφ exp β cos φ + i + ρk + j − k φ , (A.1.17)
2π −π 2π

which may be viewed as a generalization of (A.1.13).

A.2 Partition function for the non-punctured model

Let us evaluate the partition function for the 2D U(N ) lattice gauge theory on a torus. For
that, we first consider the K-functional KL1 ×L2 for a rectangle composed of V = L1 × L2
plaquettes, which can be expressed as (A.1.11). As is shown in Fig. A.3, we identify the
top and bottom sides represented by U −1 and U , respectively, and identify the left and
right sides represented by W −1 and W , respectively. Integrating out the group elements
U and W , we obtain the partition function for the non-punctured model as
Z
Znonpunc = dU dW KL1 ×L2 (U W U −1 W −1 )
X  λr V Z
= dr dU dW χr (U W U −1 W −1 )
r
d r

X  λr V Z
= dU χr (U )χr (U −1 )
r
dr
X λr V

= , (A.2.1)
r
d r

100
 U −1

W −1 W

Figure A.3: The partition function for the 2D U(N ) gauge theory on a torus is obtained
from the K-functional for the rectangle by integrating out the group elements U and W
corresponding to the identified sides.

where we have used the orthogonality relation (A.1.7) and a formula

1
Z
dΩ χr (U ΩW Ω−1 ) = χr (U )χr (W ) . (A.2.2)
dr
In the U(1) case, the partition function (A.2.1) reduces to
+∞
X
Znonpunc = [I(n, θ, β)]V . (A.2.3)
n=−∞

As one can see from (A.1.13), the integral I(n, θ, β) has a property I(n, θ + 2πk, β) =
I(n − k, θ, β) for ∀k ∈ Z, which guarantees the 2π periodicity of (A.2.3) in θ.
Let us consider taking the V → ∞ and β → ∞ limits simultaneously with fixed
Vphys ≡ V /β, which corresponds to the continuum limit. In this limit, the integral (A.1.13)
can be evaluated as
1 2
eβ− 2β ( 2π −n) .
1 θ
I(n, θ, β) ' √ (A.2.4)
2πβ
Plugging this into (A.2.3), we obtain
 β V X +∞
"  2 #
e V θ
Znonpunc ' √ exp − −n
2πβ n=−∞
2β 2π
+∞
"  2 #
X 1 θ
∼ exp − Vphys −n , (A.2.5)
n=−∞
2 2π
omitting the divergent constant factor.

A.3 Partition function for the punctured model

Let us extend the calculation in the previous section to the punctured model. First, we
calculate the K-functional for a rectangle with a hole shown in Fig. A.4, which we divide

101
 Ω1
A1

U1 ω1 ω2 U2

A2
Ω2

Figure A.4: The K-functional for a rectangle with a hole is obtained by gluing the two
regions A1 and A2 . From this, the K-functional for the punctured torus is obtained
similarly to what we did in Fig. A.3. Integrating out the link variables surrounding the
puncture, we obtain the partition function for the 2D U(N ) gauge theory on a punctured
torus.

into two regions A1 and A2 by cutting along two segments Ω1 and Ω2 . The outer and
inner boundaries of the rectangle are divided into two segments (U1 , U2 ) and (ω1 , ω2 ),
respectively. Then, the K-functional for each region is given, respectively, as
 |A1 |
X λr1
KA1 (U1 Ω2 ω1 Ω1 ) = dr1 χr1 (U1 Ω2 ω1 Ω1 ) , (A.3.1)
r1
dr1
 |A2 |
X λr2
KA2 (Ω−1 −1
1 ω2 Ω2 U2 ) = dr2 χr2 (Ω−1 −1
1 ω2 Ω2 U2 ) . (A.3.2)
r2
dr2

By gluing the two regions A1 and A2 together at Ω1 and Ω2 , we obtain the K-functional
for the rectangle with a hole as
Z
KA1 ∪A2 = dΩ1 dΩ2 KA1 (U1 Ω2 ω1 Ω1 )KA2 (Ω−1 −1
1 ω2 Ω2 U2 )
 |A1 |  |A2 | Z
X λr1 λr2
= dr1 dr2 dΩ1 dΩ2 χr1 (U1 Ω2 ω1 Ω1 )χr2 (Ω−1 −1
1 ω2 Ω2 U2 )
r1 ,r2
dr1 d r2
 |A1 |  |A2 | Z
X λr λr
= dr dΩ2 χr (U1 Ω2 ω1 ω2 Ω−1
2 U2 )
r
dr dr
X  λr V
= χr (U1 U2 )χr (ω1 ω2 ) , (A.3.3)
r
d r

where we have defined V = |A1 ∪ A2 | = |A1 | + |A2 |.

Let us introduce the group elements U and W for the outer boundary as we did
in Fig. A.3 so that U1 U2 = U W U −1 W −1 , and define ω = ω1 ω2 for the inner boundary.
Integrating out the group elements U and W , we obtain the K-functional for the punctured

102
torus as
X  λr V Z
Kpunc (ω) = χr (ω) dU dW χr (U W U −1 W −1 )
r
dr
X  λr V Z
1
= χr (ω) dU χr (U )χr (U −1 )
r
dr dr
X 1  V
λr
= χr (ω) . (A.3.4)
r
dr dr

Finally, we integrate out the link variables surrounding the puncture to get the partition
function for the punctured model as
Z X 1  λr  V
Zpunc = dω Kpunc (ω) = δr,0 = (λ0 )V , (A.3.5)
r
dr dr

where r = 0 corresponds to the trivial representation, which has d0 = 1.

In the U(1) case, the partition function reduces to

Zpunc = [I(0, θ, β)]V , (A.3.6)

which does not have the 2π periodicity in θ.

Let us consider taking the V → ∞ and β → ∞ limits simultaneously with fixed
Vphys ≡ V /β, which corresponds to the continuum limit. Similarly to the case of the
non-punctured model discussed in Section A.2, we obtain
 β V "  2 #
e V θ
Zpunc ' √ exp −
2πβ 2β 2π
"  2 #
1 θ
∼ exp − Vphys , (A.3.7)
2 2π

omitting the divergent constant factor. This coincides with (A.2.5) in the Vphys → ∞
limit for |θ| < π. Note, however, that the equivalence between the punctured and non-
punctured models does not hold for finite Vphys .

A.4 Evaluation of the observables

We can evaluate the expectation values of various observables defined in Section 3.2.4
from the partition function derived above, namely (A.2.3) for the non-punctured model
and (A.3.6) for the punctured model. Since the latter case is easier due to the absence of
an infinite sum, we only discuss the former case in what follows.
The average plaquette w defined by (3.2.14) is given as
+∞
1 X
w= A(n, θ, β) [I(n, θ, β)]V , (A.4.1)
Znonpunc n=−∞

103
where we have defined
∂
A(n, θ, β) = log I(n, θ, β)
∂β
Z π    
1 1 θ
= dφ cos φ exp β cos φ + i −n φ
I(n, θ, β) 2π −π 2π
I(n − 1, θ, β) + I(n + 1, θ, β)
= . (A.4.2)
2I(n, θ, β)

Similarly, the topological charge density defined by (3.2.15) can be obtained from
+∞
V X
hQi = −i B(n, θ, β) [I(n, θ, β)]V , (A.4.3)
Znonpunc n=−∞

where we have defined

1 ∂
B(n, θ, β) = I(n, θ, β)
I(n, θ, β) ∂θ
Z π    
i 1 θ
= dφ φ exp β cos φ + i −n φ . (A.4.4)
I(n, θ, β) 4π 2 −π 2π

Finally, the topological susceptibility defined by (3.2.16) can be obtained from

+∞
V X
2
C(n, θ, β) + (V − 1)B(n, θ, β)2 [I(n, θ, β)]V ,
 
hQ i = − (A.4.5)
Znonpunc n=−∞

where we have defined

1 ∂2
C(n, θ, β) = I(n, θ, β)
I(n, θ, β) ∂θ2
Z π    
1 1 2 θ
=− dφ φ exp β cos φ + i −n φ . (A.4.6)
I(n, θ, β) 8π 3 −π 2π

Note that I(n, θ, β) and the functions (A.4.2), (A.4.4) and (A.4.6) derived from it
are all real-valued, and we can calculate them by numerical integration with sufficient
precision. Also, when we evaluate the infinite sum in the expressions (A.4.1), (A.4.3) and
(A.4.5), we have to truncate it at some n. Note here that |I(n, θ, β)| vanishes quickly as
|θ/2π − n| increases. We can therefore evaluate the infinite sum with sufficient precision
by keeping only a few terms when the lattice volume V is sufficiently large.
In this section, we have derived the exact results for the log definition (3.1.12) of the
topological charge. As is clear from the derivation, we can obtain the exact results for
the sine definition (3.1.13) by simply replacing I(n, θ, β) with
Z π  
1 θ
Ĩ(n, θ, β) = dφ exp β cos φ + i sin φ − inφ . (A.4.7)
2π −π 2π

104
Appendix B

The punctured model with the sine

definition Qsin

In Sections 3.3 and 3.4, we have discussed the punctured model with the log definition
(3.1.12) of the topological charge for simplicity. In fact, we can also use the sine definition
(3.1.13) in the punctured model. Here we discuss what happens in this case.
The drift terms for the sine definition are given already for the non-punctured model
in Section 3.2.1. When we consider the punctured model, the only modification from the
non-punctured model appears in the drift terms for the four link variables surrounding
the puncture; i.e., UK,1 , UK+2̂,1 , UK,2 and UK+1̂,2 . Thus the drift terms are given as

−1 −1


 −i β2 (Pn − Pn−1 − Pn−2̂ + Pn− 2̂
θ
) − i 4π (Pn + Pn−1 − Pn−2̂ − Pn− 2̂
)

 for n 6= K, K + 2̂ ,
Dn,1 S = β −1 θ −1

 −i 2 (−PK−2̂ + PK−2̂ ) + i 4π (PK−2̂ + PK−2̂ ) for n = K ,

 −i β (P −1 θ −1
−P ) − i (P +P ) for n = K + 2̂ ,

2 K+2̂ K+2̂ 4π K+2̂ K+2̂
(B.0.1)

−1 −1


 −i β2 (−Pn + Pn−1 + Pn−1̂ − Pn− 1̂
θ
) − i 4π (−Pn − Pn−1 + Pn−1̂ + Pn− 1̂
)

 for n 6= K, K + 1̂ ,
Dn,2 S = β −1 θ −1

 −i 2 (PK−1̂ − PK−1̂ ) − i 4π (PK−1̂ + PK−1̂ ) for n = K ,

 −i β (−P −1 θ −1
+P ) + i (P +P ) for n = K + 1̂ .

2 K+1̂ K+1̂ 4π K+1̂ K+1̂
(B.0.2)

At large β, all the plaquettes except PK , namely the one that is removed, approach unity.
The drift term from the θ term therefore vanishes for all the link variables except for
θ
those surrounding the puncture, which have constant drifts ±i 2π . Thus in the continuum
limit, the drift terms for the sine definition agree with those for the log definition given
by (3.4.1) and (3.4.2). This connection makes it easier to understand why we can safely
ignore the issue of δ-function in the drift term for the log definition described in Section
3.2.1.

105
 (β, L) = (12, 20)

(β, L) = (3, 10)

0 0.5
10 (β, L) = (12, 20)

0.4
-2
10
0.3

-4
10 0.2

0.1
10-6
0.0
1 10 100 -3 -2 -1 0 1 2 3
u Re Qsin

Figure B.1: The results obtained by the CLM for the punctured model using the sine
definition of the topological charge. (Left) The histogram of the magnitude u of the drift
term is shown for (β, L) = (3, 10) and (12, 20) with θ = π. (Right) The histogram of
Re Qsin for the punctured model is shown for (β, L) = (12, 20) with θ = π. The exact
result obtained for (β, L) = (12, 20) with θ = 0 is shown by the solid line for comparison.

It is therefore expected that the results of the CLM for the sine definition are essentially
the same as those for the log definition for large β. In Fig. B.1, we show our results for
the punctured model with the sine definition for the same (β, L) as those in Fig. 3.7 with
the log definition. For (β, L) = (12, 20), we find that the histogram of the magnitude u
of the drift term falls off rapidly, and that the histogram of Re Qsin obtained by the CLM
is widely distributed within the range −3 . Re Qsin . 3. Hence the topology freezing
problem is circumvented without causing large drifts similarly to the situation with the
log definition.
On the other hand, for (β, L) = (3, 10), we find that the histogram of the magnitude
u of the drift term falls off fast and that the condition for the validity of the CLM is
satisfied unlike the case of the log definition. As a result, all the observables are in
complete agreement with the exact results for all values of θ even with (β, L) = (3, 10).
This can be seen from Fig. B.2, where we show our results for the punctured model
with the sine definition for the same values of (β, L) as the ones used in Fig. 3.12. For
(β, L) = (1.92, 8) corresponding to the same Vphys ≡ L2 /β, however, we actually find that
the histogram has a power-law tail similarly to the case of the log definition. Therefore,
the difference between the two definitions is merely a small shift in the validity region of
the CLM.
We also show the exact results for the punctured model with the log and sine defini-
tions, which tend to agree as β is increased with fixed Vphys ≡ L2 /β, which corresponds
to the continuum limit.

106
 (β, L) = (3, 10) (β, L) = (12, 20)

log log
0.86
sin sin
0.959

0.84
w

w
0.958

0.82

0.957
0 0.5 1 0 0.5 1
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.05 0.010
log log
sin sin
0.04 0.008
Im 〈 Q 〉 / V

Im 〈 Q 〉 / V

0.03 0.006

0.02 0.004

0.01 0.002

0.00 0.000
0 0.5 1 0 0.5 1
θ/π θ/π

(β, L) = (3, 10) (β, L) = (12, 20)

0.020 0.004
log log
sin sin

0.015 0.003

0.010 0.002
χ

0.005 0.001

0.000 0.000
0 0.5 1 0 0.5 1
θ/π θ/π

Figure B.2: The results for various observables obtained by the CLM for the punctured
model with the sine definition Qsin . The average plaquette (Top), the imaginary part
of the topological charge density (Middle), the topological susceptibility (Bottom) are
plotted against θ for (β, L) = (3, 10) (Left) and (12, 20) (Right). The exact results for the
punctured model with the log and sine definitions are shown for the same (β, L) by the
dashed lines and the dash-dotted lines, respectively, for comparison.

107
Appendix C

The determination of the parameter

p

In this appendix, we explain how we determine the parameter p in the IR cutoff (5.1.13)
and (5.1.14). While a naive choice would be p = 1, it was proposed in ref. [87] that one
should choose a slightly larger value so that the results become almost independent of
p. There it was found in the VDM model that the results for the extent of space R2 (t)
become independent of p when p is larger1 than pc = 1.2 ∼ 1.3. Based on this observation,
we used p = 1.4 when we simulate the VDM model in section 5.2.3.
Here we repeat the same analysis in the case of the bosonic model and the original
model. In Fig. C.1, we plot the extent of space R2 (t)/R2 (tc ) against time (t − tc )/R(tc )
for the bosonic model (Left) and the original model (Right), respectively, with various
values of p. For all values of p, we find that only three directions start to expand at some
critical time tc . In the bosonic model, the results scale for p = 1.3, 1.4, 1.5 except for the
data around the peak of R2 (t). Similar scaling behavior is observed for the original model
for p = 1.4, 1.5, 1.6. Based on these results, we use p = 1.5 for the bosonic model and
p = 1.6 for the original model in sections 5.2.1 and 5.2.2, respectively.

1
For the values of p in this region, it was also observed [87] from the analysis of the Schwinger-Dyson
equations that the effect of the IR cutoff decreases as one takes the infinite volume limit.

108
 2000
p=1.0 2.4 p=1.4
p=1.3 p=1.5
p=1.4 2.2 p=1.6
1500 p=1.5 y=a+(1-a)exp(bx)
y=a+(1-a)exp(bx) 2
R2(t)/R2(tc)

R2(t)/R2(tc) 1.8
1000 1.6
1.4
500 1.2
1
0 0.8
-1 -0.5 0 0.5 1 1.5 2 0 0.1 0.2 0.3 0.4 0.5 0.6
(t-tc)/R(tc) (t-tc)/R(tc)

Figure C.1: (Left) The extent of space R2 (t)/R2 (tc ) obtained for the bosonic model is
plotted against x = (t − tc )/R(tc ) for various values of p with N = 256, C = 100, κ = 1.0.
The block size is chosen as n = 32, 24, 20, 18 for p = 1.0, 1.3, 1.4, 1.5, respectively. The
solid line represents a fit to the p = 1.4 data with R2 (t)/R2 (tc ) = a+(1−a) exp(bx), which
gives a = 0.92(5), b = 7.3(6). (Right) The extent of space R2 (t)/R2 (tc ) obtained for the
original model is plotted against x = (t − tc )/R(tc ) for various values of p with N = 16,
C = 5, κ = 0.46. The block size is chosen as n = 7, 6, 6 for p = 1.4, 1.5, 1.6, respectively.
The solid line represents a fit to the p = 1.6 data with R2 (t)/R2 (tc ) = a + (1 − a) exp(bx),
which gives a = 0.83(4), b = 5.3(7).

109
Bibliography

[1] John R. Klauder. Coherent State Langevin Equations for Canonical Quantum Sys-
tems With Applications to the Quantized Hall Effect. Phys. Rev. A, 29:2036–2047,
1984.

[2] G. Parisi. ON COMPLEX PROBABILITIES. Phys. Lett. B, 131:393–395, 1983.

[3] Gert Aarts, Erhard Seiler, and Ion-Olimpiu Stamatescu. The Complex Langevin
method: When can it be trusted? Phys. Rev. D, 81:054508, 2010.

[4] Gert Aarts, Frank A. James, Erhard Seiler, and Ion-Olimpiu Stamatescu. Complex
Langevin: Etiology and Diagnostics of its Main Problem. Eur. Phys. J. C, 71:1756,
2011.

[5] Keitaro Nagata, Jun Nishimura, and Shinji Shimasaki. Justification of the complex
Langevin method with the gauge cooling procedure. PTEP, 2016(1):013B01, 2016.

[6] Keitaro Nagata, Jun Nishimura, and Shinji Shimasaki. Argument for justification of
the complex Langevin method and the condition for correct convergence. Phys. Rev.
D, 94(11):114515, 2016.

[7] Michael Levin and Cody P. Nave. Tensor renormalization group approach to 2d
classical lattice models. Phys.Rev.Lett., 99(12):120601, 2007.

[8] Z. Y. Xie, J. Chen, M. P. Qin, J. W. Zhu, L. P. Yang, and T. Xiang. Coarse-

graining renormalization by higher-order singular value decomposition. Phys. Rev.
B, 86:045139, Jul 2012.

[9] G. Evenbly and G. Vidal. Tensor network renormalization. Phys. Rev. Lett.,
115:180405, Oct 2015.

[10] Daiki Adachi, Tsuyoshi Okubo, and Synge Todo. Anisotropic tensor renormalization
group.

[11] Daisuke Kadoh and Katsumasa Nakayama. Renormalization group on a triad net-
work.

110
[12] Edward Witten. Analytic continuation of Chern-Simons theory. AMS/IP Stud. Adv.
Math., 50:347–446, 2011.

[13] Marco Cristoforetti, Francesco Di Renzo, and Luigi Scorzato. New approach to the
sign problem in quantum field theories: High density qcd on a lefschetz thimble.
Phys.Rev.D, 86:074506, 2012.

[14] H. Fujii, D. Honda, M. Kato, Y. Kikukawa, S. Komatsu, and T. Sano. Hybrid monte
carlo on lefschetz thimbles - a study of the residual sign problem. JHEP, 10:147,
2013.

[15] Andrei Alexandru, Gokce Basar, Paulo F. Bedaque, Gregory W. Ridgway, and
Neill C. Warrington. Sign problem and monte carlo calculations beyond lefschetz
thimbles. JHEP, 05:053, 2016.

[16] Masafumi Fukuma and Naoya Umeda. Parallel tempering algorithm for integration
over lefschetz thimbles. PTEP, 2017(7):073B01, 2017.

[17] Masafumi Fukuma, Nobuyuki Matsumoto, and Naoya Umeda. Implementation of

the hmc algorithm on the tempered lefschetz thimble method.

[18] Yuto Mori, Kouji Kashiwa, and Akira Ohnishi. Toward solving the sign problem
with path optimization method. Phys.Rev.D, 96(11):111501, 2017.

[19] Yuto Mori, Kouji Kashiwa, and Akira Ohnishi. Application of a neural network to
the sign problem via the path optimization method. PTEP, 2018(2):023B04, 2018.

[20] Kouji Kashiwa, Yuto Mori, and Akira Ohnishi. Controlling the model sign problem
via the path optimization method: Monte carlo approach to a qcd effective model
with polyakov loop. Phys.Rev.D, 99(1):014033, 2019.

[21] Kouji Kashiwa, Yuto Mori, and Akira Ohnishi. Application of the path optimization
method to the sign problem in an effective model of qcd with a repulsive vector-type
interaction. Phys.Rev.D, 99(11):114005, 2019.

[22] C.A. Baker et al. An improved experimental limit on the electric dipole moment of
the neutron. Phys.Rev.Lett., 97:131801, 2006.

[23] J. M. Pendlebury et al. Revised experimental upper limit on the electric dipole
moment of the neutron. Phys. Rev. D, 92(9):092003, 2015.

[24] R. D. Peccei and Helen R. Quinn. CP conservation in the presence of instantons.

Phys. Rev. Lett., 38:1440–1443, 1977. [,328(1977)].

[25] R.D. Peccei and Helen R. Quinn. Constraints imposed by cp conservation in the
presence of instantons. Phys.Rev.D, 16:1791–1797, 1977.

111
[26] Steven Weinberg. A new light boson? Phys.Rev.Lett., 40:223–226, 1978.

[27] Frank Wilczek. Problem of strong p and t invariance in the presence of instantons.
Phys.Rev.Lett., 40:279–282, 1978.

[28] Davide Gaiotto, Anton Kapustin, Zohar Komargodski, and Nathan Seiberg. Theta,
time reversal, and temperature. JHEP, 05:091, 2017.

[29] Ryuichiro Kitano, Takao Suyama, and Norikazu Yamada. θ = π in su(n)/Z n gauge
theories. JHEP, 09:137, 2017.

[30] Naoto Kan, Ryuichiro Kitano, Shimon Yankielowicz, and Ryo Yokokura. From 3d
dualities to hadron physics.

[31] Andrei Parnachev and Ariel R. Zhitnitsky. Phase transitions, theta behavior and
instantons in qcd and its holographic model. Phys.Rev.D, 78:125002, 2008.

[32] Sergei Dubovsky, Albion Lawrence, and Matthew M. Roberts. Axion monodromy in
a model of holographic gluodynamics. JHEP, 02:053, 2012.

[33] Francesco Bigazzi, Aldo L. Cotrone, and Roberto Sisca. Notes on theta dependence
in holographic yang-mills. JHEP, 08:090, 2015.

[34] Daniel Arean, Ioannis Iatrakis, Matti Jarvinen, and Elias Kiritsis. Cp-odd sector
and θ dynamics in holographic qcd. Phys.Rev.D, 96(2):026001, 2017.

[35] Ryuichiro Kitano, Norikazu Yamada, and Masahito Yamazaki. Is N = 2 Large? 10

2020.

[36] Ettore Vicari and Haralambos Panagopoulos. Theta dependence of su(n) gauge
theories in the presence of a topological term. Phys.Rept., 470:93–150, 2009.

[37] Mitsuaki Hirasawa, Akira Matsumoto, Jun Nishimura, and Atis Yosprakob. Complex
Langevin analysis of 2D U(1) gauge theory on a torus with a θ term. JHEP, 09:023,
2020.

[38] U.J. Wiese. Numerical simulation of lattice θ vacua: The 2-d u(1) gauge theory as a
test case. Nucl.Phys.B, 318:153–175, 1989.

[39] B. E. Rusakov. Loop averages and partition functions in U(N) gauge theory on
two-dimensional manifolds. Mod. Phys. Lett., A5:693–703, 1990.

[40] Claudio Bonati and Paolo Rossi. Topological susceptibility of two-dimensional U (N )

gauge theories. Phys. Rev., D99(5):054503, 2019.

112
[41] Ahmed S. Hassan, Masahiro Imachi, Norimasa Tsuzuki, and Hiroshi Yoneyama.
Character expansion, zeros of partition function and theta term in u(1) gauge theory.
Prog.Theor.Phys., 94:861–872, 1995.

[42] Jan C. Plefka and Stuart Samuel. Monte Carlo studies of two-dimensional systems
with a theta term. Phys. Rev., D56:44–54, 1997.

[43] Yoshinobu Kuramashi and Yusuke Yoshimura. Tensor renormalization group study
of two-dimensional U(1) lattice gauge theory with a θ term. JHEP, 04:089, 2020.

[44] Christof Gattringer and Oliver Orasch. Density of states approach for lattice QCD
with a θ-term. 4 2020.

[45] Erhard Seiler, Denes Sexty, and Ion-Olimpiu Stamatescu. Gauge cooling in complex
Langevin for QCD with heavy quarks. Phys. Lett., B723:213–216, 2013.

[46] Shi Chen, Kenji Fukushima, Hiromichi Nishimura, and Yuya Tanizaki. Deconfine-
ment and CP breaking at θ = π in Yang-Mills theories and a novel phase for SU(2).
Phys. Rev. D, 102(3):034020, 2020.

[47] N. Ishibashi, H. Kawai, Y. Kitazawa, and A. Tsuchiya. A Large N reduced model as

superstring. Nucl. Phys. B, 498:467–491, 1997.

[48] Tohru Eguchi and Hikaru Kawai. Reduction of Dynamical Degrees of Freedom in
the Large N Gauge Theory. Phys. Rev. Lett., 48:1063, 1982.

[49] David J. Gross and Yoshihisa Kitazawa. A Quenched Momentum Prescription for
Large N Theories. Nucl. Phys. B, 206:440–472, 1982.

[50] Alfred Schild. Classical Null Strings. Phys. Rev. D, 16:1722, 1977.

[51] Michael B. Green, John H. Schwarz, and Lars Brink. Superfield Theory of Type II
Superstrings. Nucl. Phys. B, 219:437–478, 1983.

[52] Masafumi Fukuma, Hikaru Kawai, Yoshihisa Kitazawa, and Asato Tsuchiya. String
field theory from IIB matrix model. Nucl. Phys. B, 510:158–174, 1998.

[53] Jun Nishimura and Fumihiko Sugino. Dynamical generation of four-dimensional

space-time in the IIB matrix model. JHEP, 05:001, 2002.

[54] H. Kawai, Shoichi Kawamoto, Tsunehide Kuroki, T. Matsuo, and S. Shinohara.

Mean field approximation of IIB matrix model and emergence of four-dimensional
space-time. Nucl. Phys. B, 647:153–189, 2002.

113
[55] T. Aoyama and H. Kawai. Higher order terms of improved mean field approximation
for IIB matrix model and emergence of four-dimensional space-time. Prog. Theor.
Phys., 116:405–415, 2006.

[56] Jun Nishimura, Toshiyuki Okubo, and Fumihiko Sugino. Systematic study of the
SO(10) symmetry breaking vacua in the matrix model for type IIB superstrings.
JHEP, 10:135, 2011.

[57] Konstantinos N. Anagnostopoulos, Takehiro Azuma, and Jun Nishimura. Monte

Carlo studies of the spontaneous rotational symmetry breaking in dimensionally re-
duced super Yang-Mills models. JHEP, 11:009, 2013.

[58] Konstantinos N. Anagnostopoulos, Takehiro Azuma, Yuta Ito, Jun Nishimura, and
Stratos Kovalkov Papadoudis. Complex Langevin analysis of the spontaneous sym-
metry breaking in dimensionally reduced super Yang-Mills models. JHEP, 02:151,
2018.

[59] Konstantinos N. Anagnostopoulos, Takehiro Azuma, Yuta Ito, Jun Nishimura,

Toshiyuki Okubo, and Stratos Kovalkov Papadoudis. Complex Langevin analysis
of the spontaneous breaking of 10D rotational symmetry in the Euclidean IKKT
matrix model. 2 2020.

[60] Sang-Woo Kim, Jun Nishimura, and Asato Tsuchiya. Expanding (3+1)-dimensional
universe from a Lorentzian matrix model for superstring theory in (9+1)-dimensions.
Phys. Rev. Lett., 108:011601, 2012.

[61] Yuta Ito, Sang-Woo Kim, Jun Nishimura, and Asato Tsuchiya. Monte Carlo studies
on the expanding behavior of the early universe in the Lorentzian type IIB matrix
model. PoS, LATTICE2013:341, 2014.

[62] Yuta Ito, Jun Nishimura, and Asato Tsuchiya. Power-law expansion of the Universe
from the bosonic Lorentzian type IIB matrix model. JHEP, 11:070, 2015.

[63] Toshihiro Aoki, Mitsuaki Hirasawa, Yuta Ito, Jun Nishimura, and Asato Tsuchiya.
On the structure of the emergent 3d expanding space in the Lorentzian type IIB
matrix model. PTEP, 2019(9):093B03, 2019.

[64] G. Parisi and Yong-shi Wu. Perturbation Theory Without Gauge Fixing. Sci. Sin.,
24:483, 1981.

[65] Gert Aarts, Pietro Giudice, and Erhard Seiler. Localised distributions and criteria
for correctness in complex Langevin dynamics. Annals Phys., 337:238–260, 2013.

114
[66] Gert Aarts, Frank A. James, Erhard Seiler, and Ion-Olimpiu Stamatescu. Adaptive
stepsize and instabilities in complex Langevin dynamics. Phys. Lett. B, 687:154–159,
2010.

[67] Zhenning Cai, Yana Di, and Xiaoyu Dong. How does Gauge Cooling Stabilize Com-
plex Langevin? 2019.

[68] Jun Nishimura and Shinji Shimasaki. New Insights into the Problem with a Singular
Drift Term in the Complex Langevin Method. Phys. Rev. D, 92(1):011501, 2015.

[69] Gert Aarts, Erhard Seiler, Denes Sexty, and Ion-Olimpiu Stamatescu. Complex
Langevin dynamics and zeroes of the fermion determinant. JHEP, 05:044, 2017.
[Erratum: JHEP 01, 128 (2018)].

[70] A. Mollgaard and K. Splittorff. Complex Langevin Dynamics for chiral Random
Matrix Theory. Phys. Rev. D, 88(11):116007, 2013.

[71] Jeff Greensite. Comparison of complex Langevin and mean field methods applied to
effective Polyakov line models. Phys. Rev. D, 90(11):114507, 2014.

[72] Manuel Scherzer, Erhard Seiler, Dénes Sexty, and Ion-Olimpiu Stamatescu. Complex
Langevin and boundary terms. Phys. Rev. D, 99(1):014512, 2019.

[73] M. Scherzer, E. Seiler, D. Sexty, and I.-O. Stamatescu. Controlling Complex

Langevin simulations of lattice models by boundary term analysis. Phys. Rev. D,
101(1):014501, 2020.

[74] Yusuke Namekawa. Comparative study of topological charge. PoS, LAT-

TICE2014:344, 2015.

[75] Constantia Alexandrou, Andreas Athenodorou, Krzysztof Cichy, Arthur Dromard,

Elena Garcia-Ramos, Karl Jansen, Urs Wenger, and Falk Zimmermann. Comparison
of topological charge definitions in Lattice QCD. Eur. Phys. J. C, 80(5):424, 2020.

[76] P. Di Vecchia, K. Fabricius, G.C. Rossi, and G. Veneziano. Preliminary Evidence for
U(1)-A Breaking in QCD from Lattice Calculations. pages 426–442, 5 1981.

[77] G.G. Batrouni, G.R. Katz, Andreas S. Kronfeld, G.P. Lepage, B. Svetitsky, and K.G.
Wilson. Langevin Simulations of Lattice Field Theories. Phys. Rev. D, 32:2736, 1985.

[78] M. Fukugita, Y. Oyanagi, and A. Ukawa. Langevin Simulation of the Full {QCD}
Hadron Mass Spectrum on a Lattice. Phys. Rev. D, 36:824, 1987.

[79] Martin Luscher and Stefan Schaefer. Lattice QCD without topology barriers. JHEP,
07:036, 2011.

115
[80] Christopher Czaban, Arthur Dromard, and Marc Wagner. Studying and removing
effects of fixed topology. Acta Phys. Polon. Supp., 7(3):551, 2014.

[81] Simon Mages, Balint C. Toth, Szabolcs Borsanyi, Zoltan Fodor, Sandor D. Katz,
and Kalman K. Szabo. Lattice QCD on nonorientable manifolds. Phys. Rev. D,
95:094512, 2017.

[82] Wolfgang Bietenholz, Christopher Czaban, Arthur Dromard, Urs Gerber,

Christoph P. Hofmann, Héctor Mejı́a-Dı́az, and Marc Wagner. Interpreting Nu-
merical Measurements in Fixed Topological Sectors. Phys. Rev. D, 93(11):114516,
2016.

[83] Wolfgang Bietenholz, Krzysztof Cichy, Philippe de Forcrand, Arthur Dromard, and
Urs Gerber. The Slab Method to Measure the Topological Susceptibility. PoS,
LATTICE2016:321, 2016.

[84] Martin Hasenbusch. Fighting topological freezing in the two-dimensional CPN −1

model. EPJ Web Conf., 175:02004, 2018.

[85] Werner Krauth, Hermann Nicolai, and Matthias Staudacher. Monte Carlo approach
to M theory. Phys. Lett. B, 431:31–41, 1998.

[86] Peter Austing and John F. Wheater. Convergent Yang-Mills matrix theories. JHEP,
04:019, 2001.

[87] Yuta Ito, Jun Nishimura, and Asato Tsuchiya. Universality and the dynamical space-
time dimensionality in the Lorentzian type IIB matrix model. JHEP, 03:143, 2017.

[88] Yuta Ito, Sang-Woo Kim, Yuki Koizuka, Jun Nishimura, and Asato Tsuchiya. A
renormalization group method for studying the early universe in the Lorentzian IIB
matrix model. PTEP, 2014(8):083B01, 2014.

[89] Takehiro Azuma, Yuta Ito, Jun Nishimura, and Asato Tsuchiya. A new method for
probing the late-time dynamics in the Lorentzian type IIB matrix model. PTEP,
2017(8):083B03, 2017.

[90] Jun Nishimura and Asato Tsuchiya. Complex Langevin analysis of the space-time
structure in the Lorentzian type IIB matrix model. JHEP, 06:077, 2019.

[91] Kohta Hatakeyama, Akira Matsumoto, Jun Nishimura, Asato Tsuchiya, and Atis
Yosprakob. The emergence of expanding space–time and intersecting D-branes from
classical solutions in the Lorentzian type IIB matrix model. PTEP, 2020(4):043B10,
2020.

[92] Jean-Michel Drouffe and Jean-Bernard Zuber. Strong coupling and mean field meth-
ods in lattice gauge theories. Phys. Rept., 102:1, 1983.

116