Generating Uniform Random Numbers
Christos Alexopoulos and Dave Goldsman
Georgia Institute of Technology, Atlanta, GA, USA
June 7, 2009
Alexopoulos and Goldsman June 7, 2009 1 / 38
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 2 / 38
Introduction
Introduction
Uniform(0,1) random numbers are the key to random variategeneration in simulation.
Goal: Give an algorithm that produces a sequence of pseudo-randomnumbers (PRN’s) R1, R2, . . . that “appear” to be iid Unif(0,1).
Desired properties of algorithm
output appears to be iid Unif(0,1)
very fast
ability to reproduce any sequence it generates
References: Banks, Carson, Nelson, and Nicol (2005); Bratley, Fox,and Schrage (1987); Knuth (2) (1981); Law (2007).
Alexopoulos and Goldsman June 7, 2009 3 / 38
Introduction
Classes of Unif(0,1) Generators
output of random device
table of random numbers
midsquare (not very useful)
Fibonacci (not very useful)
linear congruential (most commonly used in practice)
Tausworthe (linear recursion mod 2)
hybrid
Alexopoulos and Goldsman June 7, 2009 4 / 38
Some Generators We Won’t Use
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 5 / 38
Some Generators We Won’t Use
Some Generators We Won’t Use
a. Random Devices
Nice randomness properties. However, Unif(0,1) sequence storagedifficult, so it’s tough to repeat experiment.
Examples:
flip a coin
particle count by Geiger counter
least significant digits of atomic clock
b. Random Number Tables
List of digits supplied in tables.Cumbersome and slow — not very useful.Once tabled no longer random.
Alexopoulos and Goldsman June 7, 2009 6 / 38
Some Generators We Won’t Use
c. Mid-Square Method (J. von Neumann)
Idea: Take the middle part of the square of the previous randomnumber. John von Neumann was a brilliant and fun-loving guy, butmethod is lousy!
Example: Take Ri = Xi/10000, ∀i, where the Xi’s are positive integers< 10000.
Set seed X0 = 6632; then 66322 → 43983424;So X1 = 9834; then 98342 → 96707556;So X2 = 7075, etc,...
Unfortunately, positive serial correlation in Ri’s.
Also, occasionally degenerates; e.g., consider Xi = 0003.
Alexopoulos and Goldsman June 7, 2009 7 / 38
Some Generators We Won’t Use
d. Fibonacci and Additive Congruential Generators
These methods are also no good!!
TakeXi = (Xi−1 + Xi−2)mod m, i = 1, 2, . . . ,
where Ri = Xi/m, m is the modulus, X0,X1 are seeds, anda = bmod m iff a is the remainder of b/m, e.g., 6 = 13mod 7.
Problem: Small numbers follow small numbers.
Also, it’s not possible to get Xi−1 < Xi+1 < Xi or Xi < Xi+1 < Xi−1
(which should occur w.p. 1/3).
Alexopoulos and Goldsman June 7, 2009 8 / 38
Linear Congruential Generators
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 9 / 38
Linear Congruential Generators
Linear Congruential Generators
LCG’s are the most widely used generators. These are pretty goodwhen implemented properly.
Xi = (aXi−1 + c) mod m, where X0 is the seed.
Ri = Xi/m, i = 1, 2, . . .
Choose a, c,m carefully to get good statistical quality and long periodor cycle length, i.e., time until LCG starts to repeat itself.
If c = 0, LCG is called a multiplicative generator.
Alexopoulos and Goldsman June 7, 2009 10 / 38
Linear Congruential Generators
Trivial Example: For purposes of illustration, consider the LCG
Xi = (5Xi−1 + 3)mod 8
If X0 = 0, we have X1 = (5X0 + 3)mod 8 = 3; continuing,
i 0 1 2 3 4 5 6 7 8 9
Xi 0 3 2 5 4 7 6 1 0 3
Ri 0 3
8
2
8
5
8
4
8
7
8
6
8
1
80 3
8
so that the sequence starts repeating with X8 = 0.
This is a full-period generator, since it has cycle length m = 8.Generally speaking, full-period is a good thing. 2
Alexopoulos and Goldsman June 7, 2009 11 / 38
Linear Congruential Generators
Better Example: Here’s a portable FORTRAN implementation(Bratley, Fox, and Schrage 1987) that we’ve seen before. It works fine,is fast, and is full-period with cycle length > 2 billion,
Xi = 16807Xi−1 mod (231 − 1).
Input: IX is an integer seed between 1 and 231 − 1.
FUNCTION UNIF(IX)K1 = IX/127773 (integer division leaves no remainder)IX = 16807*(IX - K1*127773) - K1*2836IF (IX.LT.0) IX = IX + 2147483647UNIF = IX*4.656612875E-10RETURNEND
Output: UNIF is real-valued in (0,1). 2
Stay tuned for various ways to assess the quality of PRN generators.
Alexopoulos and Goldsman June 7, 2009 12 / 38
Linear Congruential Generators
So what can go wrong with LCG’s?
a. Something like Xi = (4Xi−1 + 2)mod 8 is not full-period, since itonly produces even integers.
b. Something like Xi = (Xi−1 + 1)mod 8 is full-period, but itproduces very non-random output: X1 = 1, X2 = 2, X3 = 3, etc.
c. In any case, if m is small, you’ll get quick cycling whether or notthe generator is full period. “Small” could mean anything less than2 billion or so!
d. And just because m is big, you still have to be careful. In additionto a. and b. above, some subtle problems can arise. Take a look atRANDU. . . .
Alexopoulos and Goldsman June 7, 2009 13 / 38
Linear Congruential Generators
Example: The infamous RANDU generator,
Xi = 65539Xi−1 mod231,
was popular during the 1960’s.
Here’s what (Ri−2, Ri−1, Ri) look like if you plot them in 3-D (stolenfrom Wikipedia). If they were truly iid Unif(0,1), you’d see dotsrandomly dispersed in the unit cube. But instead, the random numbersfall entirely on 15 hyperplanes (not good).
Alexopoulos and Goldsman June 7, 2009 14 / 38
Tausworthe Generator
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 15 / 38
Tausworthe Generator
Tausworthe Generator
Define a sequence of binary digits B1, B2, . . ., by
Bi =
( q∑
j=1
cjBi−j
)
mod2,
where ci = 0 or 1.
Looks a bit like a generalization of LCG’s.
Usual implementation (saves computational effort):
Bi = (Bi−r + Bi−q)mod 2 (0 < r < q).
Obtain
Bi = 0, if Bi−r = Bi−q or Bi = 1, if Bi−r 6= Bi−q.
To initialize the Bi sequence, specify B1, B2, . . . , Bq.Alexopoulos and Goldsman June 7, 2009 16 / 38
Tausworthe Generator
Example (Law 2007):r = 3, q = 5;B1 = · · · = B5 = 1Bi = (Bi−3 + Bi−5) mod 2 = Bi−3 XOR Bi−5, i > 5B6 = (B3 XOR B1) = 0, B7 = (B4 XOR B2) = 0, etc. 2
Turns out period of 0-1 bits is 2q − 1 = 31.
How do we go from Bi’s to Unif(0,1)’s?
Easy way: Use (ℓ-bit binary integers)/2ℓ.
Example: Set ℓ = 4 in previous example and get:15
16, 8
16, . . . → 1111, 1000, . . .. 2
Lots of potential for Tausworthe generators. Nice properties, includinglong periods, fast calculation. Theoretical issues still beinginvestigated.
Alexopoulos and Goldsman June 7, 2009 17 / 38
Generalizations of LCGs
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 18 / 38
Generalizations of LCGs
Generalizations of LCGs
A Simple Generalization:
Xi =(∑q
j=1aiXi−j
)
modm, where the ai’s are constants.
Extremely large periods possible (up to mq − 1 if parameters arechosen properly). But watch out! — Fibonacci is a special case.
Combinations of Generators:
Can combine two generators X1,X2, . . . and Y1, Y2, . . . to constructZ1, Z2, . . .. Some suggestions:
Set Zi = (Xi + Yi) mod m
Shuffling
Set Zi = Xi or Zi = Yi
Sometimes desired properties are improved (difficult to prove).Alexopoulos and Goldsman June 7, 2009 19 / 38
Generalizations of LCGs
A Really Good Combined Generator due to L’Ecuyer (1999) (anddiscussed in Law 2007).
Initialize X1,0, X1,1, X1,2, X2,0, X2,1, X2,2. For i ≥ 3, set
X1,i = (1,403,580 X1,i−2 − 810,728 X1,i−3)mod(232 − 209)
X2,i = (527,612 X2,i−1 − 1,370,589 X2,i−3)mod(232 − 22,853)
Yi = (X1,i − X2,i)mod(232 − 209)
Ui = Yi/(232 − 209)
As crazy as this generator looks, it’s actually pretty simple, works well,and has an amazing cycle length of about 2191!
It is of interest to note that Matsumoto and Nishimura have developedthe “Mersenne Twister” generator, which has period of 219937 − 1 (yes,that’s a prime number).
Alexopoulos and Goldsman June 7, 2009 20 / 38
Choosing a Good Generator — Some Theory
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 21 / 38
Choosing a Good Generator — Some Theory
Choosing a Good Generator — Some Theory
Here are some miscellaneous results due to Knuth and others that arehelpful in determining the quality of a PRN generator.
Theorem: The generator Xi+1 = aXi mod 2n (n > 3) can have cyclelength of at most 2n−2. This is achieved when X0 is odd and a = 8k + 3or a = 8k + 5 for some k.
Theorem: Xi+1 = (aXi + c) mod m, c > 0 has full cycle if (i) c and mare relatively prime; (ii) a − 1 is a multiple of every prime which dividesm; and (iii) a − 1 is a multiple of 4 if 4 divides m.
Corollary: Xi+1 = (aXi + c) mod 2n (c, n > 1) has full cycle if c is oddand a = 4k + 1 for some k.
Lots of cycle length results like these.
Alexopoulos and Goldsman June 7, 2009 22 / 38
Choosing a Good Generator — Some Theory
Example (Banks et al.): Xi = 13Xi−1 mod(64).
i Xi Xi Xi Xi
0 1 2 3 4
1 13 26 39 52
2 41 18 56 36
3 21 42 63 20
4 17 34 51 4
......
......
...
8 33 2 35...
......
...
16 1 3
The minimum period = 4... terrible random numbers!Why does cycling occur so soon? See first theorem.
Alexopoulos and Goldsman June 7, 2009 23 / 38
Choosing a Good Generator — Some Theory
And here’s a theorem that gives a condition for multiplicativegenerators to be full period.
Theorem: The multiplicative generator Xi = aXi−1 modm, with primem has full period (m − 1) if and only if
(a) m divides am−1 − 1.
(b) For all integers i < m − 1, m does not divide ai − 1.
How many such multipliers exist?
For m = 231 − 1, it can be shown that 534,600,000 multipliers yield fullperiod.
Remark: The “best” multiplier with m = 231 − 1 is a = 950,706,376(Fishman and Moore 1986).
Alexopoulos and Goldsman June 7, 2009 24 / 38
Choosing a Good Generator — Some Theory
Geometric Considerations
Theorem: The k-tuples (Ri, . . . , Ri+k−1), i ≥ 1, from multiplicativegenerators lie on parallel hyperplanes in [0, 1]k .
The following geometric quantities are of interest.
Minimum number of hyperplanes (in all directions). Find themultiplier that maximizes this number.
Maximum distance between parallel hyperplanes. Find themultiplier that minimizes this number.
Minimum Euclidean distance between adjacent k-tuples. Find themultiplier that maximizes this number.
Remark: The RANDU generator is particularly bad since it lies on only15 hyperplanes.
Alexopoulos and Goldsman June 7, 2009 25 / 38
Choosing a Good Generator — Some Theory
Can also look at one-step serial correlation.
Serial Correlation of LCG’s (Greenberger 1961):
Corr(R1, R2) ≤ 1
a
(
1 − 6c
m+ 6(
c
m)2
)
+a + 6
m
This upper bound is very small for m in the range of 2 billion and, say,a = 16807.
Lots of other theory considerations that can be used to evaluate theperformance of a particular PRN generator.
Alexopoulos and Goldsman June 7, 2009 26 / 38
Choosing a Good Generator — Statistical Tests
Outline
1 Introduction
2 Some Generators We Won’t Use
3 Linear Congruential Generators
4 Tausworthe Generator
5 Generalizations of LCGs
6 Choosing a Good Generator — Some Theory
7 Choosing a Good Generator — Statistical Testsχ2 Goodness-of-Fit TestRuns Tests for Independence
Alexopoulos and Goldsman June 7, 2009 27 / 38
Choosing a Good Generator — Statistical Tests
Choosing a Good Generator — Statistical Tests
We’ll look at two classes of tests:
Goodness-of-fit tests — are the PRN’s approximately Unif(0,1)?
Independence tests — are the PRN’s approximately independent?
If a particular generator passes both types of tests (in addition to othertests that we won’t tell you about), we’ll be happy to use the PRN’s itgenerates.
All tests are of the form H0 (our null hypothesis) vs. H1 (the alternativehypothesis).
Alexopoulos and Goldsman June 7, 2009 28 / 38
Choosing a Good Generator — Statistical Tests
We regard H0 as the status quo, so we’ll only reject H0 if we have“ample” evidence against it.
In fact, we want to avoid incorrect rejections of the null hypothesis.Thus, when we design the test, we’ll set the level of significance
α ≡ P (Reject H0|H0 true) = P (Type I error)
(typically, α = 0.05 or 0.1). We won’t worry about Type II error at thispoint.
Alexopoulos and Goldsman June 7, 2009 29 / 38
Choosing a Good Generator — Statistical Tests χ2 Goodness-of-Fit Test
χ2 Goodness-of-Fit Test
Test H0 : R1, R2, . . . Rn ∼ Unif(0,1).
Divide the unit interval into k cells (subintervals). If you chooseequi-probable cells [0, 1
k ), [ 1
k , 2
k ), . . . , [k−1
k , 1], then a particularobservation Rj will fall in a particular cell with prob 1/k.
Tally how many of the n observations fall into the k cells. IfOi ≡ # of Rj ’s in cell i, then (since the Rj ’s are iid), we can easily seethat Oi ∼ Bin(n, 1
k ), i = 1, 2, . . . , k.
Thus, the expected number of Rj ’s to fall in cell i will beEi ≡ E[Oi] = n/k, i = 1, 2, . . . , k.
Alexopoulos and Goldsman June 7, 2009 30 / 38
Choosing a Good Generator — Statistical Tests χ2 Goodness-of-Fit Test
We’ll reject the null hypothesis H0 if the Oi’s don’t match well with theEi’s.
The χ2 goodness-of-fit statistic is
χ20 ≡
k∑
i=1
(Oi − Ei)2
Ei.
A large value of this statistic indicates a bad fit.
In fact, we reject the null hypothesis H0 (that the observations areuniform) if χ2
0 > χ2α,k−1
, where χ2α,k−1
is the appropriate (1− α) quantilefrom a χ2 table, i.e., P (χ2
k−1< χ2
α,k−1) = 1 − α.
If χ20 ≤ χ2
α,k−1, we fail to reject H0.
Alexopoulos and Goldsman June 7, 2009 31 / 38
Choosing a Good Generator — Statistical Tests χ2 Goodness-of-Fit Test
Usual recommendation from baby stats class: For the χ2 g-o-f test towork, pick k, n such that Ei ≥ 5 and n at least 30. But. . .
Unlike what you learned in baby stats class, when we test PRNgenerators, we usually have a huge number of observations n (at leastmillions) with a large number of cells k. When k is large, we can usethe approximation
χ2α,k−1 ≈ (k − 1)
[
1 − 2
9(k − 1)+ zα
√
2
9(k − 1)
]3
,
where zα is the appropriate standard normal quantile.
Remarks: (1) 16807 PRN generator usually passes the g-o-f test justfine. (2) We’ll show how to do g-o-f tests for other distributions later on— just doing uniform PRN’s for now. (3) Other g-o-f tests:Kolmogorov-Smirnov test, Anderson-Darling test, etc.
Alexopoulos and Goldsman June 7, 2009 32 / 38
Choosing a Good Generator — Statistical Tests χ2 Goodness-of-Fit Test
Illustrative Example (Banks et al.): n = 100 observations, k = 10intervals. Thus, Ei = 10 for i = 1, 2, . . . , 10. Further, suppose thatO1 = 13, O2 = 8, . . . , O10 = 11. (In other words, 13 observations fell inthe cell [0,0.1), etc.)
Turns out that
χ20 ≡
k∑
i=1
(Oi − Ei)2
Ei= 3.4.
Let’s take α = 0.05. Then from χ2 tables, we have
χ2α,k−1 = χ2
0.05,9 = 16.9.
Since χ20 < χ2
α,k−1, we fail to reject H0, and so we’ll assume that the
observations are approximately uniform. 2
Alexopoulos and Goldsman June 7, 2009 33 / 38
Choosing a Good Generator — Statistical Tests Runs Tests for Independence
Runs Tests for Independence
Now we consider H0 : R1, R2, . . . , Rn are independent.
First look at Runs Tests .
Consider some examples of coin tossing:
A. H, T, H, T, H, T, H, T, H, T,. . . (negative correlation)
B. H, H, H, H, H, T, T, T, T, T,. . . (positive correlation)
C. H, H, H, T, T, H, T, T, H, T,. . . (“just right”)
Alexopoulos and Goldsman June 7, 2009 34 / 38
Choosing a Good Generator — Statistical Tests Runs Tests for Independence
A run is a series of similar observations.
In A above, the runs are: “H”, “T”, “H”, “T”,. . . . (many runs)
In B, the runs are: “HHHHH”, “TTTTT”, . . . . (very few runs)
In C: “HHH”, “TT”, “H”, “TT”,. . . . (medium number of runs)
A runs test will reject the null hypothesis of independence if there are“too many” or “too few” runs, whatever that means.
Alexopoulos and Goldsman June 7, 2009 35 / 38
Choosing a Good Generator — Statistical Tests Runs Tests for Independence
Runs Test “Up and Down”. Consider the following sequence ofuniforms.
.41 .68 .89 .84 .74 .91 .55 .71 .36 .30 .09. . .
If the uniform increases, put a +; if it decreases, put a − (like H’s andT’s). Get the sequence
+ + − − + − + − − − . . .
Here are the associated runs:
++, −−, +, −, +, −−−, . . .
So do we have too many or two few runs?
Alexopoulos and Goldsman June 7, 2009 36 / 38
Choosing a Good Generator — Statistical Tests Runs Tests for Independence
Let A denote the total number of runs “up and down” out of nobservations. (A = 6 in the above example.)
Amazing Fact: If n is large (at least 20) and the Rj ’s are actuallyindependent, then
A ≈ Nor(
2n − 1
3,
16n − 29
90
)
.
So if n = 100, we would expect around 67 runs!
We’ll reject the null hypothesis if A is too big or small. Thestandardized test statistic is
Z0 =A − E[A]√
Var(A).
Thus, we reject H0 if |Z0| > zα/2. E.g., if α = 0.05, we reject if|Z0| > 1.96.
Alexopoulos and Goldsman June 7, 2009 37 / 38
Choosing a Good Generator — Statistical Tests Runs Tests for Independence
Illustrative Example: Suppose that n = 100. Then
A ≈ Nor(66.7, 17.5).
So we could expect to see 66.7 ± 1.96√
17.5 ≈ [58.5, 74.9] runs.
If we see anything out of that range, we’ll reject. 2
Lots of other tests available for independence: Other runs tests,correlation tests, gap test, poker test, birthday test, etc.
Alexopoulos and Goldsman June 7, 2009 38 / 38