Monte Carlo Method

Rashi Agarwal
(Dated: November 5, 2018)
From risk analysis in business to optimising control parameters in industry, computer algorithms
based on Monte Carlo methods are now a part of problem solving tools widely used in varied
sectors. Using random sampling, the method provides numerical results for problems which might
be probabilistic or deterministic in nature. Starting with a general description of the method, the
present paper introduces the applicability of stochastic approach to obtain estimates of physical
observables of a magnetic system. In particular 2-D Ising model for a ferromagnet is solved using
Metropolis algorithm for importance sampling and estimates of spontaneous magnetisation and
susceptibility are obtained. Temperature variation of the obtained parameters is then employed to
illustrate phase transition at critical temperature.

I. INTRODUCTION motion of neutrons in it [1] .

The advent of computational methods has bridged the
gap between theory and experimentation. On one hand
we can now attempt to study complex, real world sys-
tems, especially biological systems which once were be-
yond the scope of theoretical treatment, on the other
hand theoretical results can now be tested where exper-
imental realisations are not feasible, for example nuclear
weapons. The Monte Carlo method is one such com-
putational method which has been significantly useful
to mathematicians, scientists, engineers and economists
alike.[2]
The essence of Monte Carlo method(MCM) is to make
use of inferential statistics and estimate numerical val-
ues. Though the term is relatively recent, the method
has now been used for over a century under the name of
”statistical sampling”. The earliest applications were in
the fields of mathematics - to integrate poorly behaved
functions etc., and research - by Enrico Fermi, on neu-
tron diffusion in the early 1930s.[5] Today, the method is
being extensively used across all branches of contempo-
rary sciences, finance and engineering to design and test
optimal models, strategies and designs.
The method first caught attention of the physics com-
munity as a result of developments that took place dur-
ing the Manhattan project due to von Neumann and
Ulam followed by the publication in 1949 by Metropolis
and Ulam.[4] Being a part of a secret project , the term
’Monte Carlo’ was supposedly the code coined by Nico-
las Metropolis to refer to the work on neutron diffusion
problem.
In the following sections we will apply MCM to study
a deterministic system under the framework of statisti-
cal mechanics. The efficiency of the applied MC algo-
rithm depends upon the stochastic sampling techniques
implemented. The real world systems are represented by
sampling a relatively smaller number of possible forms
that the system can take. The sample should, however,
be selected such that it mimics the properties of the real
system and hence can be used to draw to inferences.
Two sampling methods are:
Simple sampling: If all possible states of the system
are equally likely then we can randomly choose a
representative sample and use it to infer proper-
ties of the system by assign equal probabilities of
occurrence to each state of the sample.
Importance Sampling: If of all accessible states, the
system spends relatively larger time in only a few,
then simple sampling is not a good option. In this
case each state of the chosen sample is assigned
a weight factor depending upon its accessibility to
the system.
III. APPLICATION
The methods of Statistical Mechanics deal with esti-
mating any physical observable, eg. temperature, of mat-
ter in bulk on the basis of dynamics of the constituent
particles.[6]
The average value of desired observable is calculated
as follows,

II. METHOD
Xi exp−Hi /kB T
P
allstates
hXi = (1)
Z
MCM is essentially concerned with experiments on
generating and treating random numbers. The problems where, X is the physical observable, Xi is its value in
which are solved can be either deterministic - such as state i, H is the Hamiltonian for the system. Expression
estimating observables in statistical physics where theo- of H contains the resultant energies due to interaction of
retical models can be formulated but not solved exactly, constituent particles and that of the system with its envi-
or probabilistic - such the performance of nuclear reac- ronment, T is the temperature, and kB is the Boltzmann
tor by choosing random number to represent the random constant.

The essence of this formalism lies in the calculation of
partition function Z. The general form of Z for a classical
system is
X bility and the required observables - energy, spontaneous
Z= exp−H/kB T (2)
allstates
As the number of possible states non-degenerate micro-
states that a system can take increases, analytical calcu-
lations become difficult. Computer simulations over all
possible states also become impractical as the number in-
creases because of the constraints on speed and memory.
An an example consider a square lattice of 25 spins in a 5
x 5 arrangement. If each spin can take either of two value
±1, our 25 spin system has a total of 225 = 33554432 pos-
sible states. This is when we assume that the constituent
spins do not interact among themselves.
Despite a large number of possible states, the proba-
bility of the system being in one of the all possible states
follows a Boltzmann distribution which makes a small
fraction of states more accessible as compared to other,
especially at low temperatures. This is the idea that is
used to select an optimum set of sample states through
MC simulations.
The main utility of Monte Carlo methods is to study
the behaviour of the given system over a range external
parameters such as temperature and magnetic field and
deal with problems pertaining to phenomena like phase
transitions where interactions play a key role and the
determination of partition function becomes even more
difficult.
We will use MCM to study equilibrium phase transi-
tions in a 2-D ferromagnet at a finite temperature using
Ising Model.
The Ising model is the simplest model to study phase
transitions in a magnetic system as a consequence of in-
teractions among the constituent spins. Following are
the important considerations of Ising model:
• Each constituent dipole is localized at a site of the
square lattice.
• Each spin can take only two value +1 or -1.
• Each spin interacts only with its nearest neighbors.
• Hamiltonian for Ising model is given as:
X
H = −J si sj (3)
<i,j>
where, J is the interaction energy between nearest
neighbors and the summation runs over all nearest
neighbors.
2
A. Algorithm
A randomly allocated lattice configuration will make
transition from one state to other with a certain proba-
magnetization and susceptibility, will be averaged over
the MC time steps.
A continuous transition from paramagnetic to ferro-
magnetic phase takes place at the critical temperature,
Tc and is marked by a discontinuity in the quantities
like susceptibility and specific heat which depend upon
a second order derivation of the free energy. While the
order parameter, the first order derivative of free energy
changes continuously. Order parameter comprises of any
physical observable which is non-zero below critical tem-
perature and zero above it, spontaneous magnetization
servers as the order parameter in the present case.
In order to simulate the random thermal fluctuation of
the system from state to state over the course of an ex-
periment Metropolis method is used. The system makes
transition from an initial state A to a final state B with
a probability, PAB given as:[3]
( −(EB −EA )
exp kB T
, if ∆E>0
PAB =
1, if ∆E<0
where, pA and pB are the boltzmann probabilities of
the system being in states A and B respectively and
∆E = EB − EA .
Thus, the transition takes place if the energy of final
state is less than that of initial state. However, we also
accept transition to higher energy states with a proba-
bility PAB , the relative Boltzmann probability.
For simplicity we set kB Tc /J = C, some dimensionless
constant and study the dynamics of lattice spins at
temperatures above and below Tc .
B. Metropolis importance sampling Monte Carlo
scheme
[3]
1. Choose a randomly allocated initial lattice state
2. Choose a site i
3. Calculate the energy change ∆E which results if
the spin at site is overturned
4. Generate a random number r such that 0 < r < 1
− k∆ET
5. If r < exp B , flip the spin
6. Go to the next site and go to (3)
For a square lattice, each spin has 4 nearest neighbors.
To maximize interaction consider that the spins at oppo-
site geometric ends interact.

IV. RESULTS
The results obtained for a 100 by 100 square lattice us-
ing single spin flip Metropolis algorithm in zero external
field condition are presented here.
A. Lattice configurations
Starting with an initial configuration, the lattice dis-
tribution of spins at different MC time steps,t for T > Tc
and T < Tc is shown below.
(a) t=0 (b) t=6
(c) t=128
FIG. 1: T > Tc
Initially all spins are in the same direction(i.e., have
value +1 or -1) as the system reaches its equilibrium
state by maximizing entropy, disorder in the system in-
creases and the opposite spins are uniformly distributed
over the lattice sites. Hence, the system should behave
as a paramagnet with no net magnetization when it is in
equilibrium with the environment.
3
(a) t=0 (b) t=64
(c) t=8192
FIG. 2: T < Tc .
The system starts with a random distribution of op-
posite spins. As it proceeds towards equilibrium it forms
clusters of like spins and as t increases the cluster of one
type of spin grows at the expense of other. At equilib-
rium, the internal energy is minimized and almost all
the spins are aligned in the same direction. The system
should, therefore, be in ferromagnetic phase with a non-
zero magnetization as t → ∞.
Once the system is in equilibrium state, we can measure
any macroscopic parameter of interest. The question,
however, arises - How do we know that the equilibrium
state has been reached? The next subsection aims to
present the answer.
B. Equilibrium State
For any physical system, the equilibrium condition is
marked by no time variation of the parameters under
observation or rather a small fluctuation about an aver-
age value. Here we will examine the variation of average
spontaneous magnetization, m with t to identify the equi-
librium state. The value of m can be obtained easily from
the average spin of the system.

FIG. 3: Plot of m vs t for T < Tc
Starting with a randomly allocated lattice, initially m
is zero and the system is in paramagnetic state. But
since T is below Tc , the disordering tendency of entropy
is overcome by the ordering tendency of internal energy
and in equilibrium all spins are in parallel alignment,
thereby, minimizing internal energy in accordance with
equation 3.
C. Phase Transition
The time variation of m can help us identify whether
the system is in para or ferromagnetic state. Temper-
ature dependence of m and average susceptibility, χ is
what we need to visualize the transition itself. For a
particular lattice configuration, the average value of spin
gives the magnetization, m and susceptibility per lattice
2 2
site, χ = <m >−<m>
kB T . At a particular temperature, the
equilibrium value(t→ ∞) of parameters is averaged over
many random initial configurations for more reliable re-
sults.
[1] John Hammersley. Monte carlo methods. Springer Science
& Business Media, 2013.
[2] Dirk P Kroese, Tim Brereton, Thomas Taimre, and
Zdravko I Botev. Why the monte carlo method is so im-
portant today. Wiley Interdisciplinary Reviews: Compu-
tational Statistics, 6(6):386–392, 2014.
[3] David P Landau and Kurt Binder. A guide to Monte Carlo
simulations in statistical physics. Cambridge university
4
FIG. 4: Plots of m vs T and χ vs T(in units J/kB T ).
The solid line represents the exact results where as the
circles represent the MCS results.
The continuous change in m and the discontinuity in
susceptibility at the transition temperature illustrates
that the system undergoes a continuous transition from
para to ferromagnetic phase at a critical temperature.
Accuracy of the results obtained can be improved by in-
creasing the lattice size. The constraints on computa-
tional speed and memory are, however, to be taken into
consideration.
V. CONCLUSIONS
Computer algorithms based on Monte Carlo methods
are of significant importance for developing theoretical
model for complex real world systems or for virtual real-
izations of available theoretical models. As seen for the
Ising model, simulations for lattice of size much smaller
than the systems otherwise encountered in nature, were
in good agreement to the exact solutions. Hence, the
method provide an alternative to tedious and approxi-
mate numerical treatments. Being inherently paralleliz-
able, MC algorithms can be used to improve the existing
models by simultaneously testing different independent
interaction parameters. In order to widen the class of
solvable problems or to improve the existing approaches
efforts need to made to develop new algorithms to over-
come the limitations due to computational speeds and
available memory spaces.
press, 2014.
[4] Nicholas Metropolis and Stanislaw Ulam. The monte carlo
method. Journal of the American statistical association,
44(247):335–341, 1949.
[5] M Newman and G Barkema. Monte carlo methods in sta-
tistical physics chapter 1-4. Oxford University Press: New
York, USA, 1999.
[6] RK Patharia. Statistical mechanics, 2nd edi, 1996.