Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more ➡
Standard view
Full view
of .
×
0 of .
Results for:
P. 1
Introduction to Markov Chain Monte Carlo Simulations and their Statistical Analysis

# Introduction to Markov Chain Monte Carlo Simulations and their Statistical Analysis

Ratings:

4.0

(1)
|Views: 1,072|Likes:

### Availability:

See More
See less

04/16/2014

pdf

text

original

arXiv:cond-
mat/0410490
v119Oct2004
March 21, 2006
Master Review Vol. 9in x 6in \u2013 (for Lecture Note Series, IMS, NUS)
article
INTRODUCTION TO MARKOV CHAIN MONTE CARLO
SIMULATIONS AND THEIR STATISTICAL ANALYSIS
Bernd A. Berg

Department of Physics
Florida State University
Tallahassee, Florida 32306-4350, USA

and

School of Computational Science
Florida State University
Tallahassee, Florida 32306-4120, USA

This article is a tutorial on Markov chain Monte Carlo simulations and their statistical analysis. The theoretical concepts are illustrated through many numerical assignments from the author\u2019s book [7] on the subject. Computer code (in Fortran) is available for all subjects covered and can be downloaded from the web.

Contents
1. Introduction

Markov chain Monte Carlo (MC) simulations started in earnest with the 1953 article by Nicholas Metropolis, Arianna Rosenbluth, Marshall Rosen- bluth, Augusta Teller and Edward Teller [18]. Since then MC simulations have become an indispensable tool with applications in many branches of science. Some of those are reviewed in the proceedings [13] of the 2003 Los Alamos conference, which celebrated the 50th birthday of Metropolis simulations.

The purpose of this tutorial is to provide an overview of basic concepts, which are prerequisites for an understanding of the more advanced lectures of this volume. In particular the lectures by Prof. Landau are closely related.

The theory behind MC simulations is based on statistics and the analy-
1
\ue01d
March 21, 2006
Master Review Vol. 9in x 6in \u2013 (for Lecture Note Series, IMS, NUS)
article
2
B.A. Berg

sis of MC generated data is applied statistics. Therefore, statistical concepts are reviewed \ufb01rst in this tutorial. Nowadays abundance of computational power implies also a paradigm shift with respect to statistics: Computation- ally intensive, but conceptually simple, methods belong at the forefront. MC simulations are not only relevant for simulating models of interest, but they constitute also a valuable tool for approaching statistics.

The point of departure for treating Markov chain MC simulations is the Metropolis algorithm for simulating the Gibbs canonical ensemble. The heat bath algorithm follows. To illustrate these methods our systems of choice are discrete Potts and continuousO(n) models. Both classes of models are programmed for arbitrary dimensions (d = 1, 2, 3, 4,... ). On the advanced side we introduce multicanonical simulations, which cover an entire tem- perature range in a single simulation, and allow for direct calculations of the entropy and free energy.

In summary, we consider Statistics, Markov Chain Monte Carlo simu- lations, the Statistical Analysis of Markov chain data and, \ufb01nally, Multi- canonical Sampling. This tutorial is abstracted from the author\u2019s book on the subject [7]. Many details, which are inevitably ommitted here, can be found there.

2. Probability Distributions and Sampling
Asample space is a set of points or elements, in natural sciences called
measurementsor observations, whose occurrence depends on chance.
Carrying out independent repetitions of the same experiment is calledsam-
pling. The outcome of each experiment provides an event called data point.
InN such experiments we may \ufb01nd the eventA to occur withfrequency
n, 0\u2264 n\u2264 N. Theprobability assigned to the event Ais a number P(A),
0\u2264P (A)\u2264 1, so that
P(A) = lim
N\u2192\u221e
n
N.
(1)
This equation is sometimes called thefrequency de\ufb01nition of probabil-
ity.Let us denote byP (a, b) the probability thatxr\u2208 [a, b] wherexr is a
continuousrandom variable drawn in the interval (\u2212\u221e, +\u221e) with the
probability densityf(x). Then,
P(a, b) =
b
af(x) dx.
(2)
\ue01d
x
\u2212\u221ef(x) dx .
(4)
A particularly important case is theuniform probability distribution
for random numbers between [0, 1),
u(x) =
\ue00f
1 for 0\u2264x < 1;
0 elsewhere.
(5)
Remarkably, the uniform distribution allows for the construction of general
probability distributions. Let
y= F(x) =
\ue01d
x
\u2212\u221ef(x\u2032) dx\u2032
and assume that the inversex =F\u22121(y) exists. Foryr being a uniformly
distributed random variable in the range [0, 1) it follows that
xr= F\u22121(yr)
(6)
is distributed according to the probability densityf (x).
TheGaussian ornormal distribution is of major importance. Its
probability density is
g(x) =1
\u03c3\u221a2\u03c0e\u2212x2/(2\u03c32)
(7)
where\u03c32 is thevariance and\u03c3 > 0 thestandard deviation. The Gaus-
sian distribution functionG(x) is related to that of variance\u03c32 = 1 by
G(x) =
\ue01d
March 21, 2006
Master Review Vol. 9in x 6in \u2013 (for Lecture Note Series, IMS, NUS)
article
Markov Chain Monte Carlo Simulations and their Statistical Analysis
3
Knowledge of all probabilitiesP (a, b) implies
f(x) = lim
y\u2192x\u2212
P(y, x)
x\u2212 y\u22650.
(3)
The(cumulative) distribution function of the random variablexr is
de\ufb01ned as
F(x) = P(xr\u2264 x) =
\ue01d
x/\u03c3
\u2212\u221ee\u2212(x\u2032\u2032 )2/2 dx\u2032\u2032=12 +1
2erf
\ue006
x
\u03c3\u221a2
\ue007
.

(8) In principle we could now generateGaussian random numbers according to Eq. (6). However, the numerical calculation of the inverse error function is slow and makes this an impractical procedure. Much faster is to express

x
\u2212\u221eg(x\u2032) dx\u2032=1
\u221a2\u03c0

## Activity (15)

### Showing

AllMost RecentReviewsAll NotesLikes