Jon Emil Gudmundsson∗

University of Uppsala

Monte Carlo method and the Ising model

Course: Statistical methods in Physics

Teacher: Gunnar Ingelman

∗
jeg1@hi.is
 Abstract

A Monte Carlo algorithm for a two dimensional Ising model is proposed and
implemented using Matlab. The results of computer simulations agree with other
sources that claim that the critical value of interaction strength is close to 0.44.
Table of contents

1 Introduction 1

2 The Model 1

3 Running the model 4

3.1 A bigger lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

4 Comparing the model with theory 8

References 10

A montecarlo.m 11

B ising.m 13
1 Introduction 1

1 Introduction

The two dimensional Ising Model is very suitable for practising simple Monte Carlo
methods. It describes a lattice with a discrete number of particles that have some
binary propertie such as spin. We can study the evolution of the system over time
depending on a particular variable called the interaction strength. At a certain
value of interaction strength we notice a phase transition, albeit a week one. The
interacation strength at this transition is compared to a theoretical value.

2 The Model

We can dene a Hamiltonian for a system that is dependant on the arrangement

of spins on a lattice and from that deduce properties such as Magnetisation and
susceptibility [1]. Suppose that the Hamiltonian is

(2.1)
X X
H = −J Si Sj − B Si
hiji i

where hiji means that we sum over the nearest-neighbour pair of spins. This
means that the spin at site ij interacts with at i(j ± 1) and (i ± 1)j . We will be
assuming periodic boundary conditions in our model which means that every spin
will interact with four other spins regardless of their position on the nite lattice.
We refer to gure 1 for better understanding of the proposed system. Here J is

Figure 1: The energy of particle on the left is low since all the neighboring particles
have the same alignment of spin. In contrast, the energy of the particle on the right
is at its highes since all the neighboring particles have a dierent spin alignment.

the dimensionless interaction strength and B represents the energy involved in

2 The Model 2

the Magnetisation of the lattice and is also dimensionless. From the Hamiltonian
we can calculate the partition function which is

(2.2)
X
Z= e−Hi
i

where we sum over all the particles in the lattice. Then the probability of nding
the system in a certain state, denoted S, is
e−H(S)
p(S) = (2.3)
Z
We can now calculate quantaties such as Magnetisation and susceptibility using
well known equations from statistical physics. We have
∂ log Z X
(2.4)
X
M= = p(S) Si
∂B S i

for Magnetisation and

!2
∂M
(2.5)
X X
χ= = p(S) Si − M2
∂B S i

for susceptibility. The energy of the system is

(2.6)
X
E= p(S)H(S).
S

and the specic heat is

(2.7)
X
C= p(S)H 2 (S) − E 2 .
S
Considering only a relatively small 20 by 20 particle lattice it would be a compu-
tational feat to calculate these quantaties considering that there are 2400 dierent
congurations. However, we can estimate these quantaties using a Monte Carlo
method. We begin with a random conguration of spins on a 20 by 20 lattice.
We then nd how likely it is for each particle to change its spin using (2.3). The
ratio of probabilties is
p(Sf )
r= = eH(S)−H(Sf ) (2.8)
p(S)
where Sf represent the conguration after the spin has been ipped. The spins of
particles in which r > 1 or r < 1 but greater than a uniformly distributed random
number between 0 and 1 have the potential of being ipped. To dampen the
process we only ip the spin of a randomly selected subgroup which is considerably
smaller then the original group. This makes for a more likely development of the
system. More information about the dampening process and its control variable,
randTol, can be found in chapter 3. Iterating until a stable value for quantaties
2 The Model 3

such as M and E is obtained will give as an estimation of the true value. We must
point out that in the limit of an innitely large lattice, it is possible to solve the
Ising model exactly. If we assume that B = 0 the expressions are simplest and we
see that the energy is [2]
 
2
2
E = −N J coth(2J) 1 + ξK1 (κ) , (2.9)
π

the Magnetisation is
(1 + z 2 )1/4 (1 − 6z 2 + z 4 )1/8
M = ±N 2 (2.10)
(1 − z 2 )1/2

when J > Jc but vanishes when J < Jc and the specic heat is
2  hπ i
C = N 2 (J coth(2J))2 2K1 (κ) − 2E1 (κ) − (1 − ξ) + ξK1 (κ) . (2.11)
π 2
Here
sinh(2J)
κ=2 ≤ 1, (2.12)
cosh2 (2J)
ξ = 2 tanh2 (2J − 1) (2.13)
and z = e−2J . The critical value of J for which κ = 1 is Jc = 0.4406868. The
complete eliptical integrals of the rst and second kind are
Z π/2
dφ
K1 (κ) ≡ , (2.14)
0 (1 − sin2 φ)1/2
κ2
Z π/2
E1 (κ) ≡ dφ(1 − κ2 sin2 φ)1/2 . (2.15)
0
These thermodynamic parameters can be calculated using numerical methods and
compared to the results of the Monte Carlo model.
3 Running the model 4

3 Running the model

We wrote a code in Matlab that implements the model, a part of the code is
borrowed from particularly well written Matlab code by Tobin Fricke [3]. The
code can be seen in the appendix. We choose to begin by creating a square lattice
with 400 particles and assign random spin orientation. We also choose a random
value between 0 and 1 for the interaction strength and then watch how the system
evolves over 1000 steps. Note that a positive value for interaction strength means
that the energy is lower if the spin is in the same direction as that of its neighbors.
This corresponds to ferromagnetism while a negative value of J means that the
spins will tend to be anti-aligned corresponding to anti-ferromagnetism. The speed
of evolution is controlled by the variable randTol and in this case we decided that
randTol = 0.1. A value of 0.1 means that only 10% percent of the originally
selected group will have its spin ipped. In essence this parameter tries to mimic
the evolution of real systems. Even though a certain particle will have a smaller
energy with its spin ipped it doesn't mean that the all the particles in the lattice
that follow that criterion will have their spins ipped immediatly. We watched the
evolution of 12,000 systems and took note of some of the important parameters.
We then plot the total energy of the system as a function of interaction, see gure
2. Although it isn't very distinctive we can see some transition at an interaction

Figure 2: Energy as a function of interaction strength for the 202 particle lattice.

strength between 0.4 and 0.5. This can be seen more clearly when we look at the
change in Magnetisation as a function of interaction strength, see gure 3. When
the strength of interaction is low the spin of particles tends to be randomly oriented
which results in a zero net Magnetisation. This corresponds to a high temperature
(the crossing of Curie's temperature) since J = −E/kT which means that J is
inversely proportional to T . However, as we increase the interaction strength the
spins tend to align. When looking at the relation between Magnetisation and
3 Running the model 5

Figure 3: Magnetisation as a function of interaction strength for the 202 particle

lattice.

energy, see gure 4 we see that there are two regions of low energy corresponding
to spin up orientation and spin down. There are also low energy cases that do
not show a high net Magnetisation. This happens when we have relatively large
clusters of either spin up and spin down. This is analogous to the domains of
ferromagnets. It is good to notice that a net Magnetisation or the formation of

Figure 4: Magnetisation as function of energy for the 202 particle lattice.

domains leads to decreased energy. This is in agreement with observations of the

behaviour of real materials.
3.1 A bigger lattice 6

3.1 A bigger lattice

We would now like to see if the phase transition at an interaction strength of
around J = 0.44 is more apparent for lattices with higher number of particles.
We therefore initialize a square lattice with 1002 particles and watch how the
system evolves over 10,000 steps. The energy versus interaction strength relation
can be seen in gure (5). Note that in the case of the smaller lattice we only needed

Figure 5: Energy as function of interaction strength for the 1002 particle lattice.

to look at the evolution of the system over 1,000 steps to reach an equilibrium.
We see that there are two reasons for an increased computational time in a larger
lattice. One, since it is bigger the ipping of spins for more particles needs to be
decided. Two, it takes greater time for a larger system to reach equilibrim which
means that you have to let the system evolve over a larger number of steps. In this

Figure 6: Magnetisation as function of interaction strength for the 1002 particle

lattice.

case, since the calculations are more demanding, we only look at the evolution of
3.1 A bigger lattice 7

1,000 systems. Unlike before the value of the parameter randTol was 0.5 which
meant that the speed of evolution was much faster. The computer time needed to
nish these computations was approximately 43 thousand seconds on a Pentium
4, 3GHz processor. As can be seen in gure 5 the phase transition is much more
apparent. The splitting of the line which becomes more apparant as interaction
strength increases is probably due to dierent energy for the lattice with domains,
i.e. zero net magnetisation, and the lattice with a total magnetisation of -1 or
1. We also notice, see gure 6 and 7, that more systems seem to populate the

Figure 7: Magnetisation as function of Energy for the 1002 particle lattice.

M = −1 and M = 1 branches than for the case of a 400 particle lattice. It turns
out that this is in agreement with theory. We will end by comparing these results
with theoretical model proposed in chapter 2.
4 Comparing the model with theory 8

4 Comparing the model with theory

It is easy to calculate the theoretical values for energy and Magnetisation ac-
cording to equations (2.9) and (2.10). We can then compare them to the values
obtained using our model. The theoretical model assumes an innite lattice which
aects the meaning of all our comparisons. It is however a reasuring thing that
our Monte Carlo model gives better results for the bigger lattice. This proba-
bly implies that the greater lattice you model the more agreement you will nd
with the theoretical value. We can see that the theoretical model for the energy
is quite dierent from what we obtain using our model, see gure 8. But like

Figure 8: Energy as a function of interaction strength for the 1002 particle lattice.

we said before, the theoretical model was in the limit of an innite lattice and
therefore we cannot be too surprised that theory and our model do not agree. We

Figure 9: Magnetisations as a function of interaction strength for the 1002 particle

lattice.
would also like to point out that according to theory the energy per site goes to
4 Comparing the model with theory 9

innity when the interaction strength goes to zero. This is in agreement with the

Figure 10: A random distribution of spins at low interaction strength.

fact that the interaction strength is reversly proportional to temperature and as

temperature goes to innity the energy of each particle should too. In gure 9 we
see a much nicer agreement between theory and the Monte Carlo model. However,
like we mentioned before, there are high interaction strength cases where the net
Magnetisation is close to zero and these correspond to the formation of domains
in ferromagnetic materials. We must then ask, why does theory not predict for-
mation of domains in real material? This is a deep question. Could it be as a
result of the fact that real life materials arent composed of innite lattices but
instead are nite and in many cases have impurities that cause a more caotic
development of the lattice? Of course our model is very primative and compar-
ison with real materials isn't very meaningful. We will end by showing gures
from the optional visulisation part of the Monte Carlo program. There we see
how small interaction strength results in a random distribution of spins which in
turn gives a net zero magnetisation and relatively high energy, see gure 10. In
contrast, high interaction strength will either result in the formation of domains
or an average magnetisation with an absolute value of one, see gure 11 for case
of domain formation.
REFERENCES 10

Figure 11: The formation of domains at relatively high interaction strength.

References

[1] S. E. Koonin and D. C. Meredith, 1990. Computational Physics . Addison-

Wesley Publishing Company, California.
[2] K. Huang, 1963. Statistical Mechanics . John Wiley and Sons, New York.
[3] Tobin Fricke. Monte Carlo investigation of the Ising model. Taken on the 9th
of March 2008.
http://web1.pas.rochester.edu/∼tobin/notebook/2006/12/27/ising-paper.pdf
A montecarlo.m 11

Appendix
A montecarlo.m

Contents
ˆ Initial Conguration
ˆ Setting up the theoretical model
ˆ The Monte Carlo Loop
ˆ Figure Generation

% File montecarlo.m

Initial Conguration
n_grid = 20; % Size of grid
Ms = [];
Js = [];
Ns = [];
Es = [];

Setting up the theoretical model

J = linspace(0,1,1000);
kappa = 2.*sinh(2*J)./(cosh(2*J)).^2;
kappam = 2*(tanh(2*J)).^2-1;
z = exp(-2*J);
% Calculating Energy & Magnetisation as a function of interaction strength
E = -J.*(coth(2*J)).*(1+(2/pi)*kappam.*K1(kappam));
M1 = ((1+z.^2).^(1/4).*(1-6*z.^2+z.^4).^(1/8))./(1-z.^2).^(1/2);
M2 = -((1+z.^2).^(1/4).*(1-6*z.^2+z.^4).^(1/8))./(1-z.^2).^(1/2);
M1(1,1:441) = 0; %When J < J_c, M = 0
M1(1,1:441) = 0; %When J < J_c, M = 0

The Monte Carlo Loop

for i = 1:12000
A montecarlo.m 12

%while (1),
% Choose a random value between 0 and 1 for the interaction strength
J = rand()+1e-10;

% Perform a simulation
[M, N, E] = ising2(n_grid, J);

% Records the results

Ms = [Ms M/(n_grid^2)];
Es = [Es E/(n_grid^2)];
Ns = [Ns N];
Js = [Js J];
i = i+1;
end
Error using ==> evalin
Undefined command/function 'ising2'.

Figure Generation
Energy per site, versus interaction strength

figure(1)
plot(Js, Es, 'ro');
hold on
plot(J,E);
ylabel('energy per site');
xlabel('interaction strength');
pbaspect([2 1 1]);
print(gcf, '-depsc2', 'ising-energy');

% Magnetisation per site, versus interaction strength

figure(2)
plot(Js, Ms, 'bo');
hold on
plot(J,M1)
hold on
plot(J,M2)
ylabel('Magnetisation per site');
xlabel('interaction strength');
B ising.m 13

ylim([-1.1 1.1]);
pbaspect([2 1 1]);
print(gcf, '-depsc2', 'ising-Magnetisation');

% Magnetisation per site, versus Energy per site

figure(3)
plot(Es, Ms, 'o', 'Color', [0 0.5 0]);
xlabel('Energy per site');
ylabel('Magnetisation per site');
pbaspect([2 1 1]);
print(gcf, '-depsc2', 'ising-mvse');

B ising.m

function [M, num, E] = ising(N,J)

B = 0;
M = []; % The total magnetic field of the system
E = []; % The total energy of the system

randTol = 0.1; % The tolerance, dampens the spin flip process

% First we generate a random initial configuration

spin = (-1).^(round(rand(N)));

% Then we let the system evolve for a fixed number of steps

for i=1:1000,
% Calculating the total spin of neighbouring cells
neighbours = circshift(spin, [ 0 1]) + ...
circshift(spin, [ 0 -1]) + ...
circshift(spin, [ 1 0]) + ...
circshift(spin, [-1 0]);

% Calculate the change in energy of flipping a spin

DeltaE = 2 * (J*(spin .* neighbours) + B*spin);

% Calculate the transition probabilities

p_trans = exp(-DeltaE);
B ising.m 14

% Decide which transitions will occur

transitions = (rand(N) < p_trans ).*(rand(N) < randTol) * -2 + 1;

% Perform the transitions

spin = spin .* transitions;

% Sum up our variables of interest

M = sum(sum(spin));
E = -sum(sum(DeltaE))/2; % Divide by two because of double counting

% Display the current state of the system (optional)

image((spin+1)*128);
xlabel(sprintf('J = %0.2f, M = %0.2f, E = %0.2f', J, M/N^2, E/N^2));
set(gca,'YTickLabel',[],'XTickLabel',[]);
axis square; colormap bone; drawnow;

end

% Count the number of clusters of 'spin up' states

[L, num] = bwlabel(spin == 1, 4);