12/9/2024

12. MONTE-CARLO METHODS

LECTURER: DR. NGUYEN THI THANH SANG
REFERENCES:
[4].C.14

[12] FRED S KLEINER, E CHENEY, E WARD CHENEY, DAVID KINCAID, DAVID R KINCAID, "NUMERICAL
MATHEMATICS AND COMPUTING", 7TH EDITION, CENGAGE LEARNING, 2013, EISBN-13: 9781133712350

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 1

MONTE-CARLO INTEGRATION

• Overview
1. Generating Psuedo-Random Numbers
2. Multidimensional Integration
a) Handling complex boundaries.
b) Handling complex integrands.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 2

1
 12/9/2024

PSEUDO-RANDOM NUMBERS

• Definition of random from Merriam-Webster:

• Main Entry: random
Function: adjective
Date: 1565
1 a : lacking a definite plan, purpose, or pattern b : made, done, or chosen at random <read
random passages from the book>
2 a : relating to, having, or being elements or events with definite probability of occurrence
<random processes> b : being or relating to a set or to an element of a set each of whose
elements has equal probability of occurrence <a random sample>; also : characterized by
procedures designed to obtain such sets or elements <random sampling>

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 3

RANDOM NUMBER

• What is random number ? Is 3 ?

• There is no such thing as single random number
• Random number
• A set of numbers that have nothing to do with the other numbers in the sequence
• In a uniform distribution of random numbers in the range [0,1] , every
number has the same chance of turning up.
• 0.00001 is just as likely as 0.5000
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 4

2
 12/9/2024

RANDOM V. PSEUDO-RANDOM

• Random numbers have no defined sequence or formulation. Thus,

for any n random numbers, each appears with equal probability.
• If we restrict ourselves to the set of 32-bit integers, then our
numbers will start to repeat after some very large n. The numbers
thus clump within this range and around these integers.
• Due to this limitation, computer algorithms are restricted to
generating what we call pseudo-random numbers.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 5

MONTE-CARLO METHODS

• 1953, Nicolaus Metropolis

• Monte Carlo method refers to any method that makes use of random numbers
• Simulation of natural phenomena
• Simulation of experimental apparatus
• Numerical analysis

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 6

3
 12/9/2024

HOW TO GENERATE RANDOM NUMBERS ?

• Use some chaotic system (Balls in a barrel – Lotto)

• Use a process that is inherently random
• Radioactive decay
• Thermal noise
• Cosmic ray arrival
• Tables of a few million random numbers
• Hooking up a random machine to a computer.
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 7

PSEUDO RANDOM NUMBER GENERATORS

• The closest random number generator that can be obtained by
computer algorithm.
• Usually, a uniform distribution in the range [0,1]
• Most pseudo random number generators have two things in common
• The use of large prime numbers
• The use of modulo arithmetic

• Algorithm generates integers between 0 and M, map to zero and

one.
X n  In / M
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 8

4
 12/9/2024

AN EARLY EXAMPLE (JOHN VON

NEUMANN,1946)
• To generate 10 digits of integer
• Start with one of 10 digits integers
• Square it and take middle 10 digits from answer
• Example: 57721566492 = 33317792380594909291

• The sequence appears to be random, but each number is determined from the previous  not random.
• Serious problem : Small numbers (0 or 1) are lumped together, it can get itself to a short loop. For example:
• 61002 = 37210000
• 21002 = 04410000
• 41002 = 16810000
• 51002 = 65610000

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 9

LINEAR CONGRUENTIAL METHOD

• Lehmer, 1948
• Most typical so-called random number generator
• Algorithm : I  (aI  c) mod( m )
n 1 n

• a,c >=0 , m > I0,a,c

• Advantage :
• Very fast
• Problem :
• Poor choice of the constants can lead to very poor sequence
• The relationship will repeat with a period no greater than m (around
m/4)
• C complier RAND_MAX : m = 32767

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 10

10

5
 12/9/2024

RANDU GENERATOR

• 1960’s IBM
• Algorithm

I n 1  (65539  I n ) mod( 2 31 )

• This generator was later found to have a serious

problem
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 11

11

1D AND 2D DISTRIBUTION OF RANDU

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 12

12

6
 12/9/2024

RANDOM NUMBER ALGORITHMS

• The class of multiplicative congruential random-number generators has the
form: The choice of the coefficients is critical. An example:
ln
xn 
231  1
ln 1   75 ln  mod  231  1
l0  1
l1  75  x1  0.0000078263692594256108903445354152213e-6
l2  710 mod  231  1  710  x2  0.13153778814316624223402060672362
l3  7 mod  2  1  1622650073
15 31
 x3  0.7556053221950332271843372039424
l4   75 *1622650073  mod  231  1  984943658  x4  0.45865013192344928715538665985474
x5  0.53276723741216922058359217857178

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 13

13

USE OF PRIME NUMBERS

• The number 231 – 1 is a prime number, so the remainder when a

number is divided by a prime is rather, well random.
• Notes on the previous algorithm:
• The l’s can reach a maximum value of the prime number.
• Dividing by this number maps the integers into reals within the open
interval (0,1.0).
• Why open interval?
• l0 is called the seed of the random process. We can use anything here.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 14

14

7
 12/9/2024

OTHER ALGORITHMS
• Multiply by a large prime and xn 1   2736731631558xn  138  mod 2 48
take the lower-order bits. x0  1
x1  2736731631696
• Here, we use higher-bit integers
to generate 48-bit random x2  216915228954218
numbers. x3  44664858844294
x4  123276424030766
• Drand48() x5  162415264731678
• Ex. 29961701459390
51892741493630
251715685692926
37108576904446
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA 163500647628542
December 9, 2024 15

15

OTHER ALGORITHMS
• Many more such algorithms.
un 1   8t  3  un t is any large number

un
xn  What is this operation?
2q
• Some do not use integers. Integers were just more efficient on old computers.

xn 1    xn  mod 1
5

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 16

16

8
 12/9/2024

OTHER ALGORITHMS

• One way to improve the behavior of

random number generator
I n  ( a  I n 1  b  I n  2 ) mod( m )
Has two initial seed and can have a period greater
than m

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 17

17

THE RANMAR GENERATOR

• Available in the CERN Library

• Requires 103 initial seed
• Period : about 1043
• This seems to be the ultimate random number generator

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 18

18

9
 12/9/2024

PROPERTIES OF PSEUDO-RANDOM NUMBERS

• Three key properties that you should remember:

1. These algorithms generate periodic sequences (hence not random). To see this,
consider what happens when a random number is generated that matches our initial
seed.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 19

19

PROPERTIES OF PSEUDO-RANDOM NUMBERS

2. The restriction to quantized numbers (a finite-set), leads to problems in high-

dimensional space. Many points end up to be co-planar. For ten-dimensions, and 32-
bit random numbers, this leads to only 126 hyper-planes in 10-dimensional space.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 20

20

10
 12/9/2024

3D DISTRIBUTION FROM RANDU

Problems seen when

observed at the right
angle

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 21

21

THE MARSAGLIA EFFECT

• 1968, Marsaglia
• Random numbers fall mainly in the planes
• The maximum number of hyperplanes will be less in the space of higher-
dimensions.
• In the RANDU generator, the replacement of the multiplier from 65539 to
69069 improves performance significantly.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 22

22

11
 12/9/2024

PROPERTIES OF PSEUDO-RANDOM NUMBERS

3. The individual digits in the random number may not be independent. There may be a
higher probability that a 3 will follow a 5.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 23

23

AVAILABLE FUNCTIONS

• Standard C Library
• Type in “man rand” on your CIS Unix environment.
• Rather poor pseudo-random number generator.
• Only results in 16-bit integers.
• Has a periodicity of 2**31 though.
• Type in “man random” on your CIS Unix environment.
• Slightly better pseudo-random number generator.
• Results in 32-bit integers.
• Used rand() to build an initial table.
• Has a periodicity of around 2**69.
• #include <stdlib.h>
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 24

24

12
 12/9/2024

AVAILABLE FUNCTIONS

• Drand48() – returns a pseudo-random number in the range from zero to one,

using double precision.
• Pretty good routine.
• May not be as portable.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 25

25

INITIALIZING WITH SEEDS

• Most of the algorithms have some state that can be initialized. Many times this
is the last generated number (not thread safe).
• You can set this state using the routines initialization methods (srand, srandom
or srand48).
• Why would you want to do this?

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 26

26

13
 12/9/2024

INITIALIZING WITH SEEDS

• Two reasons to initialize the seed:

1. The default state always generates the same sequence of random numbers. Not really
random at all, particularly for a small set of calls. Solution: Call the seed method with
the lower-order bits of the system clock.
2. You need a deterministic process that is repeatable.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 27

27

INITIALIZING WITH SEEDS

• We do not want the mountain to change as the camera moves.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 28

28

14
 12/9/2024

MAPPING RANDOM NUMBERS

• Most computer libraries supporting for random numbers only provide random
numbers over a fixed range.
• You need to map this to your desired range.
• Two common cases:
• Random integers from zero to some maximum.
• Random floating-point or double-precision numbers mapped to the range zero to one.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 29

29

NON-RECTANGULAR AREAS

• In 2D, we may want points randomly distributed over some region.

• Square – independently determine x and y.
• Rectangle - ???
• Circle - ???
• Wrong way – independently determine r and .

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 30

30

15
 12/9/2024

MONTE-CARLO TECHNIQUES

• Problem: What is the probability that 10 dice throws add up exactly to 32?

• Exact Way. Calculate this exactly by counting all possible ways of making 32
from 10 dice.

• Approximate (Lazy) Way. Simulate throwing the dice (say 500 times), count
the number of times the results add up to 32, and divide this by 500.

• Lazy Way can get quite close to the correct answer quite quickly.
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 31

31

MONTE-CARLO TECHNIQUES

• Sample Applications
• Integration
• System simulation
• Computer graphics - Rendering.
• Physical phenomena - radiation transport
• Simulation of Bingo game
• Communications - bit error rates
• VLSI designs - tolerance analysis

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 32

32

16
 12/9/2024

b
SIMPLE EXAMPLE: . a
p ( x)dx
• Method 1: Analytical Integration
• Method 2: Quadrature
• Method 3: MC -- random sampling the area enclosed by a<x<b and 0<y<max (p(x))
b  # 
 a
p ( x)dx  max( p ( x))(b  a )
 # #


P(x) P(x)

a b x a b
IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 33

33

b
SIMPLE EXAMPLE: . a
p ( x)dx

• Intuitively:
b  # 

a
p ( x)dx  max( p ( x))(b  a )
 # #



 AreaBox  Probability  y  f  x 

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 34

34

17
 12/9/2024

Shape of High Dimensional Region

• Higher dimensional shapes can be complex.
• How to construct weighted points in a grid
that covers the region R ?

Problem :
mean-square distance from the origin

  ( x  y )dxdy
2 2

 r 2 
  dxdy

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 35

35

Integration over simple shape ?

1 inside R
s
0 outside R

0.5 0.5
 dx  dy( x  y ) s( x, y )
2 2
r 2
 0.5 0.50.5 0.5
0.5 dx 0.5 dys( x, y )
Grid must be fine enough !

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 36

36

18
 12/9/2024

MONTE-CARLO INTEGRATION
D’: rectangular
• Integrate a function over a complicated
domain
• D: complicated domain.
• D’: Simple domain, superset of D. D
• Pick random points over D’:
D’: circle
• Counting: N: points over D
• N’: points over D’
VolumeD N

VolumeD N 
D

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 37

37

EXERCISES

• Generate random numbers in an open interval (0,6) whose probability density

function is f(x) = |x-|, using MC integration . Simulate their
distribution.
• Generate random numbers in an open interval (-1,2) whose probability
density function is f(x) = x2, using MC integration . Simulate
their distribution.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 38

38

19
 12/9/2024

ESTIMATING  USING MONTE CARLO

• The probability of a random point lying
inside the unit circle: (x,y)

M
N
• If pick a random point N times and M of
those times the point lies inside the unit
circle:


• If N becomes very large, P=P0

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 39

39

ESTIMATING  USING MONTE CARLO

• Results:
• N= 10,000 Pi= 3.104385
• N= 100,000 Pi= 3.139545
• N= 1,000,000 Pi= 3.139668
• N= 10,000,000 Pi= 3.141774
• …

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 40

40

20
 12/9/2024

ESTIMATING  USING MONTE CARLO

double x, y, pi;
const long m_nMaxSamples = 100000000;
long count=0;
for (long k=0; k<m_nMaxSamples; k++) {
x=2.0*drand48() – 1.0; // Map to the range [-1,1]
y=2.0*drand48() – 1.0;
if (x*x+y*y<=1.0) count++;
}
pi=4.0 * (double)count / (double)m_nMaxSamples;

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 41

41

STANDARD QUADRATURE

• We can find numerical value of a definite integral by the definition:

f lim f Δ
→

where points xi are uniformly spaced.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 42

42

21
 12/9/2024

ERROR IN QUADRATURE

• Consider integral in d dimensions:

 f( X )dx 1dx 2 dx d   f( Xi ) x d

• The error with N sampling points is

|  f(X )dX   f(X )x d |  N 1/ d

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 43

43

MONTE CARLO ERROR

• From probability theory one can show that the Monte Carlo error decreases
with sample size N as

1

N
independent of dimension d.

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 44

44

22
 12/9/2024

GENERAL MONTE CARLO

• If the samples are not drawn uniformly but with some probability distribution
P(X), we can compute by Monte Carlo:

1 N

 f( X )P( X ) d X 
N
 f(X )
i 1
i

Where P(X) is normalized,  P(X ) d X 1

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 45

45

EXERCISES

What is
the output?

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 46

Book: Introduction to Numerical Programming: A Practical Guide for Scientists and Engineers Using Python and
C/C++

46

23
 12/9/2024

THANK YOU FOR LISTENING

IT142IU - ANALYTICS FOR OBSERVATIONAL DATA December 9, 2024 47

47

24