3rd problem set on Computer Simulations, 2023

Hand-in: see schedules for your group, linked in the TeachCenter.

Please upload/submit your program files in the TeachCenter before the conversation
with the tutor.

This exercise is an example for a typical application of computer simulations. There are several
phases, namely (after initial problem definition, model building and choice of method) it encom-
passes writing and verifying a program, simulations, error analysis, and physics analysis. The
main emphasis is on producing reliable results, with reliable error estimates, and on how to find
the limits of applicability of a given method.

When one performs calculations carefully in this way (and often with better MC update meth-
ods), Monte Carlo simulations are the method which can produce reliable results for the largest
systems, by far, (e.g. millions of sites) even in very difficult cases like Quantum Spin systems.

Simulation of the Ising model

Write and apply a program to simulate the two-dimensional Ising model on a square
lattice. The Hamilton function is
X
H = −J si sj
hiji

The "critical temperature" of the two-dimensional Ising model is known exactly. One
finds tanh2 (2Jβc ) = 12 ⇒ Jβc = 0.44068 . . . , i.e. kBJTc ' 2.269 . . . .
At T < Tc the Ising model is in a ferromagnetic phase, at T > Tc it is paramagnetic. You
should perform simulations in the range 0 ≤ βJ ≤ 0.6. (The unit of energy „J = 1” is
often not written.) Read the whole problem set before beginning.

1 Program
(1.5 points) Use the Metropolis algorithm, as explained in the lecture notes. In order
to facilitate writing an efficient program, the following guide is quite detailed. The program
should consist of the following major parts, called from a main program.
1. Lattice geometry: The lattice has periodic boundary conditions, i.e. the topology of
a torus, and size L × L. First you should generate an array neighbor(i,nu), where
i = 1, 2, . . . , L2 enumerates the lattice sites, and nu = 1, 2, 3, 4 denotes the direc-
tions to the nearest neighbors of site i. Then j = neighbor(i, nu) is the respective
nearest neighbor site. You can use the matlab–script get_neighbor() supplied on the
web site of this class: neighbor = get_neighbor(L). This way one separates the lattice
geometry from the update algorithm ! On the one hand, this approach saves time,
since the location of neighbors does not need to be calculated every time. On the
other hand, it permits to simulate other lattice geometries very easily, basically by
changing the array neighbor(i,nu).
 2. Update routine: Metropolis-Hastings algorithm, as described in the lecture notes.
Write a subroutine "sweep", which repeats L2 times: select a site at random, and
call a subprogram „Metropolis-Single-Spin-Flip” for that site.
Note: since only a single spin si is proposedPto be flipped, in „Metropolis-Single-
P
Spin-Flip” you do not need the whole sum hiji si sj , but only the sum sj over
the spins of nearest neighbors of site i, which you can obtain easily by using the
array "neighbor". This sum can only take a few discrete values, for which you can
tabulate
P the corresponding Metropolis probabilities beforehand. (The expression
hiji means that every geometric edge of the lattice appears once in the sum. For
a periodic L × L lattice, the sum then contains 2L2 terms.)

3. Measurements: Measure energy E and magnetization M of the whole lattice after

each sweep, then normalize by 1/L2 , i.e. per lattice site. During measurement,
also calculate averages and the "naive" variance of averages, i.e. the variance which
would be correct in the absence of autocorrelations. This is the variance of indi-
vidual measurements, divided by the number of measurements. Its square root is
the "naive" error. Note that in order to calculate this variance on the fly, you only
need a sum of measured values and a sum of their squares, both of which can be
updated cheaply.

4. Visualisation: If possible, display the spin configurations graphically during simu-

lation, initially after each complete sweep for illustration. More rarely for longer
runs, in order to save time.

(plus 0.5 points) Note that the CPU time per sweep for updates and measurements nat-
urally grows only linearly with system size for large systems. You must make sure that
you do not accidentally write a program with a quadratic (or worse) behavior.

Examples of actual run-times (without graphics): On an 8×8 lattice the run-time for one
million sweeps (of L2 = 64 single spin update attempts each) in some matlab program
was about 90 seconds, and in a compiled fortran program about 8 seconds. The run-
time for the autocorrelation analysis was less than a second.

2 Verification
(1 point) Calculate the expectation value of the energy by hand for a 2 × 2 lattice, by
adding the contributions from all spin configurations. Note the periodic boundary con-
ditions. On the 2 × 2 lattice they imply that all nearest neighbor pairs of sites are dou-
bly connected. Compare to the average energy from simulations, for several values of
βJ. The energy should agree to about three decimal digits (relative error of about 10−3 ),
within suitably small error bars. One value as a reference: the energy per site at βJ = 0.4
is about −1.602168.
3 Statistical tools and their application
1. Time series: (0.5 points) Generate graphical time series of energy and magnetiza-
tion for each simulation, and analyze them by eye. For each simulation: what is
roughly the longest time scale on which you can recognize autocorrelations ? (You
may have to zoom into the time series to see correlations).

(0.5 points) Note that the length of the simulation must be much longer (rule of
thumb: 100 times) than the largest time scale visible in an observable. You will
likely encounter (and you must recognize!) cases in which the run cannot be made
long enough. Then the average value of that observable is not converged well,
i.e. it is wrong. The autocorrelation coefficients are then also incorrect, and the
variance cannot be calculated correctly. Such a ”short” run (however expensive)
can not be used to calculated reliable averages. In an actual application it would
have to be extended or thrown away.

2. Autocorrelation function: (0.5 points) Provide a program which calculates the auto-
correlation coefficient (=autocorrelation function) for E and M , and which plots
it in semilogarithmic fashion as a function of distance t. You may be able to use
your program from the previous exercise.
For each simulation and observable (E and M ): determine approximately the time
scale τ of the autocorrelation coefficient (inverse slope, when plotted semilogarith-
mically). Since you need to do this step often, you may want to automate it, for
example by a fit. As an approximation, it is ok for this exercise to just take the
inverse slope between two times in a sensible range (determined separately for
each case), instead of a fit. To determine the range, note that the actual autocorre-
lation coefficient ρ(t) îs positive and decays monotonically. You can set a time t∗
just before ρ(t) violates these conditions, and then calculate the inverse slope τ of
ρ between 0.3t∗ and 0.6t∗ .
How does τ compare to the time scale which you can see in the times series ?

3. Error estimate: (1 point) In order to calculate an estimate of the true error of your
observables (energy, magnetization), you can choose one of the following ap-
proaches:
a) In the present example you will see more or less a single exponential, and
you can therefore take the inverse slope τ as an estimate for the integrated
autocorrelation time, and use it to calculate approximately the true variance
of your observables (see lecture notes).
b) Alternatively you can integrate over the autocorrelation function according
to the definition of τint (but do not integrate too far into noise!),
c) or use the binning procedure to obtain error estimates (and thus also esti-
mates of τint ).
(plus 0.5 points) Make sure that you specify average values always together with an esti-
mate of their error!
4 Exploring the limits of applicability
In order to produce reliable results, calculations must begin on very small lattices, and
be extended slowly to larger lattices, while monitoring autocorrelations (at least by eye).
After verification for the 2 × 2 system, begin by simulating on a 4 × 4 lattice.

1. Freezing at low temperatures: (1 point) Look at the time series (including thermal-
ization) of measurements of M and E at β = 0.2, β = 0.4, and β = 0.6. You will
find that the magnetization at the larger values of β exhibits long autocorrelations
already on the 4 × 4 system. Attention: This is about time series of individual
measurements, not about time series of running averages, because in those, auto-
correlations are more and more hidden at later times.
Repeat the calculations for a 6 × 6 and an 8 × 8 lattice. Leave out initial measure-
ments for thermalization now. Calculate the autocorrelation function, the auto-
correlation time, and an error estimate of the averages of E and M .
At the larger values of β and larger lattices, the magnetization may not converge
anymore, depending on the actual length of your simulations. (Reminder: the
expectation value of M at vanishing magnetic field is zero). The largest time scale
can even become too long to be recognized.
N.B. At low temperatures the magnetization "freezes" and autocorrelation times
are particularly large, growing even exponentially in β and lattice volume. In
order to simulate such cases, one needs better Monte Carlo methods than we are
using here.

2. Autocorrelations at the critical temperature: (1 point) Autocorrelations are also im-

portant around the critical temperature, which is the physically most interesting
region.
Perform calculations on lattices L = 4, 8, 16, 32 at β = βc . Look at the time series
and autocorrelations of E and M and perform error estimates.
You can estimate the largest autocorrelation time to be expected here beforehand:
The correlation length is bounded by L, thus the maximum autocorrelation time
is roughly L2 (and thus much smaller than at low temperatures). The time scales
visible in time series are roughly of this size; the actual integrated autocorrelation
time of the energy is lower by about a factor of 10. Using the expected autocorre-
lation time you can estimate how many sweeps you need to perform in order to
obtain reliable averages.
(N.B. Away from the critical temperature, the correlation length becomes smaller
in both directions. The autocorrelation time becomes smaller at higher tempera-
tures).
5 Phase transition
1. (0.5 points) Look at a few configurations for systems of size L = 8 and L = 32 at
β = 0.2, β = βc β = 0.6. What are typical features at low and at high temperatures
? How can one interprete this physically ?

2. (1 point) Plot the average of E and of the modulus |M | (modulus taken for each
measured spin configuration) as a function of β for lattices of sizes L = 4, 8, 16
(and perhaps 32), in the range 0 < β < 0.6. For each L, choose sensible values
for β, such that one can recognize the occurence of a phase transition. It should
become sharper in β on larger lattices.
N.B. Such data, and even the precise dependence on lattice size ("Finite Size Scal-
ing") can be used to obtain detailed information about phase transitions in in-
finitely large systems, i.e. in the thermodynamic limit. More on this in the lectures
on "Phase transition and critical phenomena".

6 Handing in
You can work on this problem set in groups of two students and produce a joint solu-
tion. Each student still needs to hand in the solution individually and fulfill the usual
requirements.
Some parts of this problem set require long Monte Carlo calculations; you might run
your program up to a few hours for individual simulations. Prepare the presentation of
corresponding results in advance. For simple simulations, the program should also run
during the presentation.
Upload: If you work in a team, then please upload the files only for one of the par-
ticipants. The other participant should only upload a file called "See_nnnnn", where
nnnnn is the name of the first participant.