Transcript Random Number Generation

Random Number Generation
Fall 2013
By Yaohang Li, Ph.D.
Review
• Last Class
– Variance Reduction
• This Class
–
–
–
–
Random Number Generation
Uniform Distribution
Non-uniform Distribution Random Number Generation
Assignment 3
• Next Class
– Quasi-Monte Carlo
Random Numbers
• Application of Random Numbers
– Simulation
• Simulate natural phenomena
– Sampling
• It is often impractical to examine all possible cases, but a
random sample will provide insight into what constitutes
typical behavior
– Numerical analysis
– Computer programming
– Decision making
• “Many executives make their decisions by flipping a coin…”
– Recreation
Natural Random Number
• Natural Random Numbers
– No two snowflakes are the same
– Sources
• White Noise
• Water Molecule Distribution
• etc.
– Generation
• Measurement
• Irreproducible
• Errors
Pseudorandom Number Generators
• Pseudorandom Numbers
– Using a Mathematical Formula
– Deterministic
– Behave like real random numbers
• Comments
– There is no “perfect” pseudorandom number generator
– We should never completely trust results from a single
pseudorandom number generator
– Good random number generators are hard to find
Middle-square method
• Developed by von Neumann
• Procedure
– 4 digit starting value is created
– Square and produce 8-digit number
– Get the middle 4 digits as the result and seed for next number
• Problems
– What if the middle 4 digits are 0
– Forsythe found that the sequence may stuck in 6100, 2100,
4100, 8100, 6100, …
• Not a good generator
Quality of Pseudorandom Numbers
•
•
•
•
•
•
•
Uniformity
Randomness
Independence
Reproducibility
Portability
Efficiency
A sufficiently long period
Generating Uniform Random Numbers
• Generating Uniform Random Numbers
– Uniform distribution on [0,1)
– Generation
• Un=Xn/m
– Xn: Random number Integer
– m: Max(Xn)+1: Usually the word size of a computer
– Un: Uniform real random number at [0,1)
Linear Congruential Method
• Linear Congruential Method
– Most commonly used generator for pseudorandom numbers
xn  axn1  b (mod m)
• m: modulus
• a: multiplier
• b: additive constant
– Period
• m constrains the period
– max period: 2m-1
– m is usually chosen to be either prime of a power-of-two
Shift-Register Generators (SRG)
• Shift-Register Generators
– based on the following recursion
k 1
xn k   ai xni (mod 2),
i 0
– ai and xi are either 0 or 1
– Comments
• The recursion produces only bits
• Incorporate these bits into integers
Lagged-Fibonacci Generators
• Lagged-Fibonacci Generators (LFG)
– Additive Lagged-Fibonacci Generators
xn  xn j  xnk (mod 2m ), j  k
– Multiplicative Lagged-Fibonacci Generators
xn  xn j  xnk (mod 2m ), j  k
• Comments
– LFG has a much longer period than LCGs
– (2k-1)2m-1
Spectral Tests
Spectral Tests
Inversive Congruential Generators
• Inversive Congruential Generators (ICGs)
– Recursive ICGs
xn  a x n1  b (mod m),
– Explicit ICGs
xn  an  b (mod m)
cc  1 (mod m)
– Advantage of ICGs
• ICGs do not fall in hyperplanes
Combined Generators
• Combined Generators
– Combining different recurrences can increase the period length
– Improve the structural properties of pseudorandom generators
– Construct a new random sequence
z n  xn  y n ,
– 
• exclusive-or operator
• addition modulo
• addition of floating-point random numbers modulo 1
– x, y
• Different random number sources
Parallel Random Number Generators
•
Requirements of Parallel Random Numbers
–
Every random number sequence generated on each processor
should satisfy the requirements of a good sequential generator.
– The parallel generator must be reproducible both on different
machines and on the same machine with a different
partitioning of the processing resources.
– The parallelly generated random streams must be
uncorrelated and must not overlap.
– The parallel generator should work for an arbitrary, but
perhaps bounded, number of processors.
Parallel Random Numbers Generations
(Leapfrog)
• Leapfrog
Parallel Random Numbers Generations
(Sequence Splitting)
• Sequence Splitting
Parallel Random Numbers Generations
(Sequence Splitting)
• Random Tree Method
– Also called
parameterization method
SPRNG
• http://sprng.cs.fsu.edu
Random Choices from a Finite Set
•A random integer X between 0 and k-1
X  kU 
– U is a random number uniformly distributed in [0,1)
•A more general case
x1 , if 0  U  p1


X   x2 , ifp1  U  p1  p2
 x , ifp  p  ...  p  U  1
2
k
 k 1
Inverse Function Method
• Cumulative Distribution Function
– Most real-valued distribution may be expressed in terms of its
distribution function F(x)
• Inverse Function Method
– X=F-1(U)
– Now the problem reduces to how to evaluate the inverse
function F-1()
Interesting Trick
• Generating the random samples of F(x)=x2
• Inverse Function Method
– X=U-1/2
• A short cut method
– If X1 is a random variable having the distribution F1(x) and if
X2 is a random variable having the distribution F2(x)
• max(X1, X2) has the distribution F1(x)F2(x)
• min(X1, X2) has the distribution F1(x)+F2(x)-F1(x)F2(x)
– Then X=max(U1, U2) has the distribution of F(x)=x2
– Hard to believe that max(U1, U2) and U-1/2 have the same
distribution
Normal Distribution
•
Polar Method
1.
2.
3.
4.
5.
Generate two independent random variables, U1 and U2
Set V1=2U1-1, V2=2U2-1
Set S=V1*V1+V2*V2
If S>=1, return to Step 1
Set X1 and X2 according to the following two equations
X 1  V1
 2 ln S
S
X 2  V2
 2 ln S
S
Acceptance-Rejection Method
•
Desired pdf
–
•
Suppose we bound the desired
probability distribution function to
sample from a box
Algorithm
1. Generate a random variable x from
U(0,1)
2. Generate another random variable y
from U(0,1)
3. If x<f(y)/fmax then return y
4. else repeat from step 1
Acceptance-Rejection Method Example
• Determine an algorithm for generating random variates
for a random variable that take values 1, 2, …, 10 with
probabilities 0.11, 0.12, 0.09, 0.08, 0.12, 0.10, 0.09, 0.09.
0.10, 0.10 respectively
• Acceptance-Rejection Method
– u1=U(0,1), u2=U(0,1), c=max(p())=0.12
– Y=floor(10*u1+1)
– while (u2>p(Y)/c)
• u1=U(0,1), u2=U(0,1)
• Y=floor(10*u1+1)
– output Y
Analysis of Acceptance-Rejection
Method
• Advantage
– Acceptance-Rejection Method can fit in different pdfs
– popularly used in complicated probability geometry
• Disadvantage
– Inefficient if the volume of the region of interest is small
relative to that of the box
• most of the darts will miss the target
Exponential Distribution
• F(x)=1-e-x/
– Logarithm method (inverse function method)
– X=-lnU
Shuffling Algorithm
•
Let X1, X2, …, Xt be a set of t numbers to be shuffled
1.
2.
3.
4.
5.
j=t
Generate U
Set k=floor (jU)+1
Exchange Xk with Xj
Decrease j by 1. If j>1, return to step 2
• Random Numbers
Summary
– Uniform Random Numbers
• Generation of Uniform Random Numbers
– Natural Random Number Generators
– Pseudorandom Number Generators
• Pseudorandom Number Generators
–
–
–
–
–
–
Requirement of Pseudorandom Number Generators
LCG
LFG
SRG
ICG
Combined Random Number Generators
• Parallel Random Number Generators
– Requirement of Parallel Random Number Generators
– Techniques for Parallel Random Number Generators
• Leapfrog
• Sequence Splitting
• Random Tree
Summary
• Numerical Distribution
– Random Choices from a finite set
– General methods for continuous distributions
• inverse function method
• acceptance-rejection method
– Distributions
• Normal distribution
– Polar method
• Exponential distribution
– Shuffling
What I want you to do?
• Review Slides
• Review basic probability/statistics concepts