Pseudo-random numbers: z }| { mostly of code
| {z } a line at a time
Nelson H. F. Beebe
Research Professor
University of Utah
Department of Mathematics, 110 LCB
155 S 1400 E RM 233
Salt Lake City, UT 84112-0090
USA
Email: beebe@math.utah.edu
, beebe@acm.org
, beebe@computer.org
(Internet)
WWW URL: http://www.math.utah.edu/~beebe
Telephone: +1 801 581 5254
FAX: +1 801 581 4148
24 September 2011
Nelson H. F. Beebe (University of Utah)
24 September 2011 1 / 61
What are random numbers good for?
o Decision making (e.g., coin flip).
o Generation of numerical test data.
o Generation of unique cryptographic keys.
o Search and optimization via random walks.
o Selection: quicksort (C. A. R. Hoare, ACM Algorithm 64:
Quicksort , Comm. ACM.
4 (7), 321, July 1961) was the first widely-used divide-and-conquer algorithm to reduce an O ( N
2
) problem to (on average) O ( N lg ( N )) . Cf. Fast Fourier Transform
(Clairaut (1754), Lagrange (1759), Gauss (1805 unpublished, 1866)
[Latin], Runge (1903), Danielson and Lanczos [crystallography]
(1942), Cooley and Tukey (1965)).
Nelson H. F. Beebe (University of Utah)
24 September 2011 2 / 61
Historical note: al-Khwarizmi
Abu ’Abd Allah Muhammad ibn Musa al-Khwarizmi (ca. 780–850) is the father of algorithm and of algebra , from his book Hisab Al-Jabr wal
Mugabalah (Book of Calculations, Restoration and Reduction) . He is celebrated in a 1200-year anniversary Soviet Union stamp:
Nelson H. F. Beebe (University of Utah)
24 September 2011 3 / 61
What are random numbers good for? . . .
o Simulation.
o Sampling: unbiased selection of random data in statistical computations (opinion polls, experimental measurements, voting,
Monte Carlo integration, . . . ). The latter is done like this ( x k random in ( a , b ) ): is
Z b f a
( x ) dx ≈
( b − a )
N k
N
∑
= 1 f ( x k
)
!
+ O ( 1 /
√
N )
Nelson H. F. Beebe (University of Utah)
24 September 2011 4 / 61
Monte Carlo integration
Here is an example of a simple, smooth, and exactly integrable function, and the relative error of its Monte Carlo integration: f(x) = 1/sqrt(x
2
+ c
2
) [c = 5]
0.200
0.150
0.100
0.050
0.000
-10 -8 -6 -4 -2 0 2 4 6 8 10
N
Convergence of Monte Carlo integration
0
-1
-2
-3
-4
-5
-6
-7
0 20 40 60 80 100
N
Convergence of Monte Carlo integration
0
-1
-2
-3
-4
-5
-6
-7
0 1 2 log(N)
3 4 5
Nelson H. F. Beebe (University of Utah)
24 September 2011 5 / 61
When is a sequence of numbers random?
o Computer numbers are rational, with limited precision and range.
Irrational and transcendental numbers are not represented.
o Truly random integers would have occasional repetitions, but most pseudo-random number generators produce a long sequence, called the period , of distinct integers: these cannot be random.
o It isn’t enough to conform to an expected distribution: the order that values appear in must be haphazard.
o Mathematical characterization of randomness is possible, but difficult.
o The best that we can usually do is compute statistical measures of closeness to particular expected distributions.
Nelson H. F. Beebe (University of Utah)
24 September 2011 6 / 61
Distributions of pseudo-random numbers o Uniform (most common).
o Exponential.
o Normal (bell-shaped curve).
o Logarithmic: if ran () is uniformly-distributed in ( a , b ) , define randl ( x ) = exp ( x ran ()) . Then a randl ( ln ( b / a )) is logarithmically distributed in ( a , b ) . [Important use: sampling in floating-point number intervals.]
Nelson H. F. Beebe (University of Utah)
24 September 2011 7 / 61
Distributions of pseudo-random numbers . . .
Sample logarithmic distribution:
% hoc a = 1 b = 1000000 for (k = 1; k <= 10; ++k) printf "%16.8f\n", a*randl(ln(b/a))
664.28612484
199327.86997895
562773.43156449
91652.89169494
34.18748767
472.74816777
12.34092778
2.03900107
44426.83813202
28.79498121
Nelson H. F. Beebe (University of Utah)
24 September 2011 8 / 61
Uniform distribution
Here are three ways to visualize a pseudo-random number distribution, using the Dyadkin-Hamilton generator function rn01() , which produces results uniformly distributed on ( 0, 1 ] :
Uniform Distribution
1.0
0.8
0.6
0.4
0.2
0.0
0 2500 5000 output n
7500 10000
Uniform Distribution
1.0
0.8
0.6
0.4
0.2
0.0
0 2500 5000 sorted n
7500 10000
Uniform Distribution Histogram
150
100
50
0
0.0
0.2
0.4
x
0.6
0.8
1.0
Nelson H. F. Beebe (University of Utah)
24 September 2011 9 / 61
Disorder and order
The Swiss artist Ursus Wehrli dislikes randomness:
Nelson H. F. Beebe (University of Utah)
24 September 2011 10 / 61
Exponential distribution
Here are visualizations of computations with the Dyadkin-Hamilton generator rnexp() , which produces results exponentially distributed on
[ 0,
∞
) :
10
8
6
4
2
0
0
Exponential Distribution
2500 5000 output n
7500 10000
10
8
6
4
2
0
0
Exponential Distribution
2500 5000 sorted n
7500 10000
1000
800
600
400
200
0
0
Exponential Distribution Histogram
1 2 3 x
4 5 6
Even though the theoretical range is [ 0,
∞ ) , the results are practically always modest: the probability of a result as big as 50 is smaller than
2 × 10
− 22
. At one result per microsecond, it could take 164 million years of computing to encounter such a value!
Nelson H. F. Beebe (University of Utah)
24 September 2011 11 / 61
Normal distribution
Here are visualizations of computations with the Dyadkin-Hamilton generator rnnorm() , which produces results normally distributed on
( − ∞
, + ∞ ) :
4
3
2
1
0
-1
-2
-3
-4
0
Normal Distribution
2500 5000 output n
7500 10000
4
3
2
1
0
-1
-2
-3
-4
0
Normal Distribution
2500 5000 sorted n
7500 10000
Normal Distribution Histogram
400
350
300
250
200
150
100
50
0
-4 -3 -2 -1 0 1 2 3 x
4
Results are never very large: a result as big as 7 occurs with probability smaller than 5 × 10
− 23
. At one result per microsecond, it could take
757 million years of computing to encounter such a value.
Nelson H. F. Beebe (University of Utah)
24 September 2011 12 / 61
Logarithmic distribution
Here are visualizations of computations with the hoc generator randl(ln(1000000)) , which produces results normally distributed on
( 1, 1000000 ) :
1000000
800000
600000
400000
200000
0
0
Logarithmic Distribution
2500 5000 7500 10000 output n
1000000
800000
600000
400000
200000
0
0
Logarithmic Distribution
2500 5000 7500 10000 sorted n
500
400
300
200
100
0
0
Logarithmic Distribution Histogram
50 100 150 200 250 x
The graphs are similar to those for the exponential distribution, but here, the result range is controlled by the argument of randl() .
Nelson H. F. Beebe (University of Utah)
24 September 2011 13 / 61
Goodness of fit: the
χ
2 measure
Given a set of n independent observations with measured values M k expected values E fit. So is ∑ n k = 1
( E k k
, then ∑
− M k
) 2 n k = 1
| ( E k
− M k
) | and is a measure of goodness of
. Statisticians use instead a measure introduced in 1900 by one of the founders of modern statistics, the English mathematician Karl Pearson (1857–1936):
χ
2 measure = k n
∑
= 1
( E k
− M k
) 2
E k
Equivalently, if we have s categories expected to occur with probability p k
, and if we take n samples, counting the number Y k in category k , then
χ
2 measure = k s
∑
= 1
( np k
− Y k
)
2 np k
(1880)
Nelson H. F. Beebe (University of Utah)
24 September 2011 14 / 61
Goodness of fit: the
χ
2 measure . . .
The theoretical
χ
2 distribution depends on the number of degrees of freedom, and table entries look like this (highlighted entries are referred to later):
D.o.f.
ν
ν
= 1
= 5
ν
ν
= 10
= 50 p = 1% p = 5% p = 25% p = 50% p = 75% p = 95% p = 99%
0.00016 0.00393
0.1015
0.4549
1.323
0.5543
1.1455
2.675
4.351
6.626
2.558
3.940
6.737
29.71
34.76
42.94
9.342
49.33
12.55
56.33
3.841
11.07
18.31
67.50
6.635
15.09
23.21
76.15
For example, this table says:
For
ν
= 10 , the probability that the
χ
2 measure is no larger than 23.21 is 99%.
In other words,
χ
2 measures larger than 23.21
should occur only about 1% of the time.
Nelson H. F. Beebe (University of Utah)
24 September 2011 15 / 61
Goodness of fit: coin-toss experiments
Coin toss has one degree of freedom,
ν then it must be tails.
= 1 , because if it is not heads,
% hoc for (k = 1; k <= 10; ++k) print randint(0,1), ""
0 1 1 1 0 0 0 0 1 0
This gave four 1s and six 0s:
χ
2 measure =
( 10 × 0.5
− 4 )
2
+ ( 10 × 0.5
− 6 )
2
10 × 0.5
= 2 / 5
= 0.40
Nelson H. F. Beebe (University of Utah)
24 September 2011 16 / 61
Goodness of fit: coin-toss experiments . . .
From the table, for
ν
= 1 , we expect a
χ
2 measure no larger than
0.4549
half of the time, so our result is reasonable.
On the other hand, if we got nine 1s and one 0, then we have
χ
2 measure =
( 10 × 0.5
− 9 ) 2 + ( 10 × 0.5
− 1 ) 2
10 × 0.5
= 32 / 5
= 6.4
This is close to the tabulated value 6.635
at p = 99%. That is, we should only expect nine-of-a-kind about once in every
100 experiments.
If we had all 1s or all 0s, the
χ
2 measure is 10 (probability
[ twice in 1000 experiments ].
If we had equal numbers of 1s and 0s, then the
χ
2 p = 0.998) measure is 0, indicating an exact fit.
Nelson H. F. Beebe (University of Utah)
24 September 2011 17 / 61
Goodness of fit: coin-toss experiments . . .
Let’s try 100 similar experiments, counting the number of 1s in each experiment:
% hoc for (n = 1; n <= 100; ++n) { sum = 0 for (k = 1; k <= 10; ++k) \ sum += randint(0,1) print sum, ""
}
4 4 7 3 5 5 5 2 5 6 6 6 3 6 6 7 4 5 4 5 5 4
3 6 6 9 5 3 4 5 4 4 4 5 4 5 5 4 6 3 5 5 3 4
4 7 2 6 5 3 6 5 6 7 6 2 5 3 5 5 5 7 8 7 3 7
8 4 2 7 7 3 3 5 4 7 3 6 2 4 5 1 4 5 5 5 6 6
5 6 5 5 4 8 7 7 5 5 4 5
Nelson H. F. Beebe (University of Utah)
24 September 2011 18 / 61
Goodness of fit: coin-toss experiments . . .
The measured frequencies of the sums are: k
Y k
100 experiments
0 1 2 3 4 5 6 7 8 9 10
0 1 5
1
2
1
9
3
1
1
6
1
2 3 1 0
Notice that nine-of-a-kind occurred once each for 0s and 1s, as predicted.
Nelson H. F. Beebe (University of Utah)
24 September 2011 19 / 61
Goodness of fit: coin-toss experiments . . .
A simple one-character change on the outer loop limit produces the next experiment: k
Y k
1000 experiments
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
1 2 3 3 8 7
1
6
1
4
2
9
5
1
4
3
6
2
6
2
7
9
8
4
9
3
8
4
7
6
5
4
5
9
2
9
4
3
3
1
2
1
1
8
1
0 7 6 1 1 0
Nelson H. F. Beebe (University of Utah)
24 September 2011 20 / 61
Goodness of fit: coin-toss experiments . . .
Another one-character change gives us this: k
Y k
10 000 experiments
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70
0 0 3 1 7 7
1
2
2
7
3
0
8
5
9
9
1
6
8
2
2
4
2
9
5
4
2
0
4
8
4
5
8
8
6
6
3
7
6
6
7
9
9
8
0
4
7
5
5
7
6
6
6
2
8
5
5
9
4
7
0
4
1
3
2
9
8
2
0
7
1
5
0
9
3
7
0
4
8
2
0
1
5
1
2 8 4 1 0 0
Nelson H. F. Beebe (University of Utah)
24 September 2011 21 / 61
Goodness of fit: coin-toss experiments . . .
A final one-character change gives us this result for one million coin tosses: k
Y k
100 000 experiments
30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70
1 4
1
2
3
4
4
7
7
0
1
1
8
2
4
2
4
1
8
8
0
2
8
6
1
0
1
6
3
3
2
2
5
9
3
1
1
2
3
9
8
7
4
8
8
0
5
6
0
9
6
5
8
7
7
3
2
0
8
1
1
3
8
2
2
7
7
8
2
8
7
1
7
1
6
6
0
7
5
6
0
4
4
7
4
0
3
9
6
2
3
0
2
9
2
2
1
2
1
5
4
4
9
9
6
6
5
4
4
7
4
2
5
7
1
4
1
1
0
7
4
3
3
7
2
1 5 5
Nelson H. F. Beebe (University of Utah)
24 September 2011 22 / 61
Are the digits of
π random?
Here are
χ
2 results for the digits of
π from recent computational records
(
χ
2 (
ν
= 9, p = 0.99
) ≈ 21.67
):
π
Digits
6B
Base
χ
2
10 9.00
50B
200B
10
10
5.60
8.09
1T
1T
10 14.97
16 7.94
p(
χ
2
)
0.56
0.22
0.47
0.91
0.46
1 /
π
Digits
6B
Base
χ
2
10 5.44
50B
200B
10
10
7.04
4.18
p(
χ
2
)
0.21
0.37
0.10
Whether the fractional digits of
π
, and most other transcendentals, are normal ( ≈ equally likely to occur) is an outstanding unsolved problem in mathematics.
Nelson H. F. Beebe (University of Utah)
24 September 2011 23 / 61
Are the first 1000 fractional digits of
π random?
3.14159265358979323846264
33 83279502 88 41971693 99 37510
58209749 44 592307816406286208998628034825342 11 70679
8214808651328230 66 47093844609 55 058 22 31725359408128
48111745028410270193852110555964462294895493038196
44288109756659334461284756482337867831652712019091
45648566923460348610454326648213393607260249141273
724587 00 660631558817488152092096282925409171536436
78925903600113305305488204665213841469519415116094
33057270365759591953092186117381932611793105118548
07446237996274956735188575272489122793818301194912
98336733624406566430860213949463952247371907021798
60943702 77 0539217176293176752384674818467669405132
00056812714526356082778577134275778960917363717872
14684409012249534301465495853710507922796892589235
42019956112129021960864034418159813629774771309960
51870721134 999999 837297804995105973173281609631859
50244594553469083026425223082533446850352619311881
71010003137838752886587533208381420617177669147303
5982534904287554 6873115956286388235378759375195778
1857780532171226806613001927876611195909216 42 01989
Nelson H. F. Beebe (University of Utah)
24 September 2011 24 / 61
Are the first 1000 fractional digits of
π random?
In the first 1000 fractional digits of
π
:
83 digit pairs (81 expected)
77-digit sequence without a 4 (probability: ( 9 / 10 )
77 ≈ 0.0003) six consecutive 9 digits (probability: 1/1,000,000) last five digits are a calendar year (probability: 1/100,000)
Conclusion: for a finite sequence of digits, the answer is no!
See Aaldert Compagner, Definitions of randomness , American Journal of
Physics 59 (8) 700–705 (1991).
URL http://m.ajp.aapt.org/resource/1/ajpias/v59/i8/p700 s1
Nelson H. F. Beebe (University of Utah)
24 September 2011 25 / 61
Is your name in
π
?
From http://www.dr-mikes-maths.com/pisearch.html
:
NELSON was not found, but I searched 31 415 929 digits of
π
, and found
BEEBE 4 times. The first occurrence was at position 846 052. What this means is that
π
= 3 + . . .
+
27
B
278246052
+
27
E
278246053
+
27
E
278246054
+
27
B
278246055
+
27
E
278246056
+ . . .
where A = 1, B = 2, C = 3, and so on.
Nelson H. F. Beebe (University of Utah)
24 September 2011 26 / 61
The Central-Limit Theorem
The famous Central-Limit Theorem (de Moivre (1718), Laplace
(1810), and Cauchy (1853)), says:
A suitably normalized sum of independent random variables is likely to be normally distributed, as the number of variables grows beyond all bounds. It is not necessary that the variables all have the same distribution function or even that they be wholly independent.
— I. S. Sokolnikoff and R. M. Redheffer
Mathematics of Physics and Modern Engineering, 2nd ed.
Nelson H. F. Beebe (University of Utah)
24 September 2011 27 / 61
The Central-Limit Theorem . . .
In mathematical terms, this is
P ( n
µ
+ r
1
√ n ≤ X
1
+ X
2
+ · · · + X n
≤ n
µ
+ r
2
√ n )
≈
σ
√
1
2
π
Z r
2 exp ( − t
2
/ ( 2
σ
2
)) dt r
1 where the X k are independent, identically distributed, and bounded random variables,
µ is their mean value ,
σ is their standard deviation , and
σ
2 is their variance .
Nelson H. F. Beebe (University of Utah)
24 September 2011 28 / 61
The Central-Limit Theorem . . .
The integrand of this probability function looks like this:
The Normal Distribution
2.0
1.5
1.0
σ
= 0.2
σ
= 0.5
σ
= 1.0
σ
= 2.0
σ
= 5.0
0.5
0.0
-10.0
Nelson H. F. Beebe (University of Utah)
-5.0
0.0
x
5.0
10.0
24 September 2011 29 / 61
The Central-Limit Theorem . . .
The normal curve falls off very rapidly. We can compute its area in
[ − x , + x ] with a simple midpoint quadrature rule like this: func f(x) { global sigma; return (1/(sigma*sqrt(2*PI)))* exp(-x*x/(2*sigma**2))
} func q(a,b) { n = 10240 h = (b - a)/n area = 0 for (k = 0; k < n; ++k) \ area += h*f(a + (k + 0.5)*h); return area
}
Nelson H. F. Beebe (University of Utah)
24 September 2011 30 / 61
The Central-Limit Theorem . . .
sigma = 3 for (k = 1; k < 8; ++k) \ printf "%d %.9f\n", k, q(-k*sigma,k*sigma)
1 0.682689493
2 0.954499737
3 0.997300204
4 0.999936658
5 0.999999427
6 0.999999998
7 1.000000000
In computer management, 99.999% (five 9’s) availability is five minutes downtime per year .
In manufacturing, Motorola’s 6
σ reliability with 1.5
σ drift is about three defects per million (from q ( − ( 6 − 1.5
) ∗
σ
, +( 6 − 1.5
) ∗
σ
) / 2).
Nelson H. F. Beebe (University of Utah)
24 September 2011 31 / 61
The Central-Limit Theorem . . .
It is remarkable that the Central-Limit Theorem applies also to nonuniform distributions. Here is a demonstration with sums from exponential and normal distributions:
700
600
500
400
300
200
100
0
5
Sums from Exponential Distribution
10 15
Sum of 10 samples
20
700
600
500
400
300
200
100
0
5
Sums from Normal Distribution
10 15
Sum of 10 samples
Superimposed on the histograms are rough fits by eye of normal distribution curves 650 exp ( − ( x − 12.6
) 2
/ 4.7
) and
550 exp ( − ( x − 13.1
)
2
/ 2.3
) .
20
Nelson H. F. Beebe (University of Utah)
24 September 2011 32 / 61
The Central-Limit Theorem . . .
Not everything looks like a normal distribution. Here is a similar experiment, using differences of successive pseudo-random numbers, bucketizing them into 40 bins from the range [ − 1.0, + 1.0
] : k
Y k
10 000 experiments (counts scaled by 1/100)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
1
3
3
5
6
1
8
8
1
1
3
1
3
8
1
6
3
1
8
7
2
1
1
2
3
6
2
6
2
2
9
0
3
1
2
3
3
9
3
6
1
3
8
7
4
1
4
4
3
7
4
6
4
4
8
7
4
8
7
4
6
7
4
3
7
4
1
4
3
8
5
3
6
5
3
3
7
3
1
2
2
8
8
2
6
1
2
3
6
2
1
2
1
8
8
1
6
2
1
3
7
1
1
3
8
7
6
3
3
6
1
2
This one is known from theory: it is a triangular distribution. A similar result is obtained if one takes pair sums instead of differences.
Nelson H. F. Beebe (University of Utah)
24 September 2011 33 / 61
Digression: Poisson distribution
The Poisson distribution arises in time series when the probability of an event occurring in an arbitrary interval is proportional to the length of the interval, and independent of other events:
P ( X = n ) =
λ n e
−
λ n !
In 1898, Ladislaus von Bortkiewicz collected Prussian army data on the number of soldiers killed by horse kicks in 10 cavalry units over 20 years:
122 deaths, or an average of 122 / 200 = 0.61 deaths per unit per year.
Deaths
2
3
0
1
4
λ
= 0.61
Kicks Kicks
(actual) (Poisson)
109
65
22
3
1
108.7
66.3
20.2
4.1
0.6
Cavalry deaths by horse kick (1875--1894)
120
100 lambda = 0.61
80
60
40
20
0
-1 0 1 2
Deaths
3 4 5
Nelson H. F. Beebe (University of Utah)
24 September 2011 34 / 61
The Central-Limit Theorem . . .
Measurements of physical phenomena often form normal distributions:
1250
Chest girth of Scottish soldiers (1817)
1000
750
500
250
0
32 34 36 38 40 42 44 46 48
Inches
2000
Height of French soldiers (1851--1860)
1500
1000
500
0
56 58 60 62 64 66 68 70
Inches
Nelson H. F. Beebe (University of Utah)
Weights of 10,000 gold sovereigns (1848)
4000
3000
2000
1000
0
-0.3 -0.2 -0.1
0.0
0.1
0.2
0.3
Grains from average
24 September 2011 35 / 61
The Central-Limit Theorem . . .
Error in erf(x)
1.0
0.5
0.0
-0.5
-1.0
-5 -4 -3 -2 -1 0 1 2 3 4 5 x
Error in gamma(x)
20
15
10
5
0
-5
-10
-15
-20
0 1 2 3 4 5 6 7 8 9 10 x
Nelson H. F. Beebe (University of Utah)
Error in erf(x), x on [-5,5]
800
600
400
200
0
-1.0
σ
= 0.22
-0.5
0.0
0.5
Units in the last place
1.0
Error in gamma(x), x on [0..10]
2500
2000
1500
σ
= 3.68
1000
500
0
-15 -10 -5 0 5
Units in the last place
10 15
24 September 2011 36 / 61
The Central-Limit Theorem . . .
Error in log(x)
1.0
0.5
0.0
-0.5
-1.0
0 1 2 3 4 5 6 7 8 9 10 x
Error in sin(x)
1.0
0.5
0.0
-0.5
-1.0
0 1 2 3 x
4 5 6
Nelson H. F. Beebe (University of Utah)
700
600
500
400
300
200
100
0
-1.0
Error in log(x), x on (0..10]
-0.5
0.0
σ
= 0.22
0.5
Units in the last place
1.0
400
300
200
100
0
-1.0
Error in sin(x), x on [0..2
π
)
σ
= 0.19
-0.5
0.0
0.5
Units in the last place
1.0
24 September 2011 37 / 61
The Normal Curve and Carl-Friedrich Gauß (1777–1855)
Nelson H. F. Beebe (University of Utah)
24 September 2011 38 / 61
The Normal Curve and the Quincunx
~
~
~
~
~ quincunx , n.
2 . An arrangement or disposition of five objects so placed that four occupy the corners, and the fifth the centre, of a square or other rectangle; a set of five things arranged in this manner.
b . spec. as a basis of arrangement in planting trees, either in a single set of five or in combinations of this; a group of five trees so planted.
Nelson H. F. Beebe (University of Utah)
Oxford English Dictionary
24 September 2011 39 / 61
The Normal Curve and the Quincunx . . .
For simulations and other material on the quincunx (Galton’s bean machine ), see: http://www.ms.uky.edu/~mai/java/stat/GaltonMachine.html
http://www.rand.org/statistics/applets/clt.html
http://www.stattucino.com/berrie/dsl/Galton.html
http://teacherlink.org/content/math/interactive/ flash/quincunx/quincunx.html
http://www.bun.kyoto-u.ac.jp/~suchii/quinc.html
Nelson H. F. Beebe (University of Utah)
24 September 2011 40 / 61
Remarks on random numbers
Any one who considers arithmetical methods of producing random numbers is, of course, in a state of sin.
— John von Neumann (1951)
[ The Art of Computer Programming, Vol. 2,
Seminumerical Algorithms , 3rd ed., p. 1]
He talks at random; sure, the man is mad.
— Queen Margaret
[William Shakespeare’s 1 King Henry VI ,
Act V, Scene 3 (1591)]
Nelson H. F. Beebe (University of Utah)
A random number generator chosen at random isn’t very random.
— Donald E. Knuth (1997)
[ The Art of Computer Programming, Vol. 2,
Seminumerical Algorithms , 3rd ed., p. 384]
24 September 2011 41 / 61
How do we generate pseudo-random numbers?
o Linear-congruential generators (most common): r n + 1
0
= ( ar n
+ c ) mod m , for integers a , c , and m , where 0 < m ,
≤ a < m , 0 ≤ c < m , with starting value 0 ≤ r
0
< m .
o Fibonacci sequence (bad!): r n + 1
= ( r n
+ r n − 1
) mod m .
o Additive (better): r n + 1
= ( r n −
α
+ r n −
β
) mod m .
o Multiplicative (bad): r n + 1
= ( r n −
α
× r n −
β
) mod m .
o Shift register: r n + k
= ∑ i k − 1
= 0
( a i r n + i
( mod 2 )) ( a i
= 0, 1 ) .
Nelson H. F. Beebe (University of Utah)
24 September 2011 42 / 61
How do we generate pseudo-random numbers? . . .
Given an integer r ∈ [ A , B ) , x = ( r − A ) / ( B − A + 1 ) is on [ 0, 1 ) .
However, interval reduction by A + ( r − A ) mod s to get a distribution in
( A , C ) , where s = ( C − A + 1 ) , is possible only for certain values of s .
Consider reduction of [ 0, 4095 ] to [ 0, m ] , with m ∈ [ 1, 9 ] : we get equal distribution of remainders only for m = 2 q − 1: m counts of remainders k mod
OK 1 2048 2048
2 1366 1365 1365
OK 3 1024 1024 1024 1024
( m + 1 ) , k ∈ [ 0, m ]
6
OK 7
4
5
8
9
820 819 819 819 819
683 683 683 683 682 682
586
512
585
512
585
512
585
512
585
512
585
512
585
512 512
456 455 455 455 455 455 455 455 455
410 410 410 410 410 410 409 409 409 409
Nelson H. F. Beebe (University of Utah)
24 September 2011 43 / 61
How do we generate pseudo-random numbers? . . .
Samples from other distributions can usually be obtained by some suitable transformation. Here is the simplest generator for the normal distribution, assuming that randu() returns uniformly-distributed values on ( 0, 1 ] : func randpmnd() \
{ ## Polar method for random deviates
## Algorithm P, p. 122, from Donald E. Knuth,
## The Art of Computer Programming, vol. 2, 3/e, 1998 while (1) \
{ v1 = 2*randu() - 1 # v1 on [-1,+1] v2 = 2*randu() - 1 # v2 on [-1,+1] s = v1*v1 + v2*v2 # s on [0,2] if (s < 1) break # exit loop if s inside unit circle
} return (v1 * sqrt(-2*ln(s)/s))
}
Nelson H. F. Beebe (University of Utah)
24 September 2011 44 / 61
Period of a sequence
All pseudo-random number generators eventually reproduce the starting sequence; the period is the number of values generated before this happens.
Widely-used historical generators have periods of a few tens of thousands to a few billion, but good generators are now known with very large periods:
> 10
14
> 10
57
> 10
18
> 10
38
> 10
449
> 10
2894
> 10
6001
> 10
14100
> 10
16736
POSIX drand48() LCG (2
48
) (1982),
Marsaglia short and fast xorshift() (2
192
) (2003),
Numerical Recipes ran2() (1992),
NIST Advanced Encryption Standard (AES) (2
128
) (2003),
Matlab’s rand() ( ≈ 2
1492
Columbus generator),
Marsaglia’s Monster-KISS (2000),
Matsumoto and Nishimura’s Mersenne Twister (1998),
Deng and Xu (2003),
Berdnikov, Trutia, & Compagner MathLink (1996).
Nelson H. F. Beebe (University of Utah)
24 September 2011 45 / 61
Reproducible sequences
In computational applications with pseudo-random numbers, it is essential to be able to reproduce a previous calculation. Thus, generators are required that can be set to a given initial seed :
% hoc for (k = 0; k < 3; ++k) \
{ setrand(12345) for (n = 0; n < 10; ++n) print int(rand()*100000),"" println ""
}
88185 5927 13313 23165 64063 90785 24066 37277 55587 62319
88185 5927 13313 23165 64063 90785 24066 37277 55587 62319
88185 5927 13313 23165 64063 90785 24066 37277 55587 62319
Nelson H. F. Beebe (University of Utah)
24 September 2011 46 / 61
Reproducible sequences . . .
If the seed is not reset, different sequences are obtained for each test run.
Here is the same code as before, with the setrand() call disabled: for (k = 0; k < 3; ++k) \
{
## setrand(12345) for (n = 0; n < 10; ++n) print int(rand()*100000),"" println ""
}
36751 37971 98416 59977 49189 85225 43973 93578 61366 54404
70725 83952 53720 77094 2835 5058 39102 73613 5408 190
83957 30833 75531 85236 26699 79005 65317 90466 43540 14295
In practice, software must have its own source-code implementation of the generators : vendor-provided ones do not suffice.
Nelson H. F. Beebe (University of Utah)
24 September 2011 47 / 61
The correlation problem
Random numbers fall mainly in the planes
— George Marsaglia (1968)
Linear-congruential generators are known to have correlation of successive numbers: if these are used as coordinates in a graph, one gets patterns, instead of uniform grey:
Good Bad
0.5
0.4
0.3
0.2
0.1
0
0
1
0.9
0.8
0.7
0.6
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2
0.4
0.6
0.8
1 0.2
0.4
0.6
0.8
The number of points plotted is the same in each graph.
1
Nelson H. F. Beebe (University of Utah)
24 September 2011 48 / 61
The correlation problem . . .
The good generator is Matlab’s rand() . Here is the bad generator:
% hoc func badran() { global A, C, M, r; r = int(A*r + C) % M; return r }
M = 2^15 - 1; A = 2^7 - 1 ; C = 2^5 - 1 r = 0 ; r0 = r ; s = -1 ; period = 0 while (s != r0) {period++; s = badran(); print s, "" }
31 3968 12462 9889 10788 26660 ... 22258 8835 7998 0
# Show the sequence period println period
175
# Show that the sequence repeats for (k = 1; k <= 5; ++k) print badran(),""
31 3968 12462 9889 10788
Nelson H. F. Beebe (University of Utah)
24 September 2011 49 / 61
The correlation problem . . .
Marsaglia’s (2003) family of xor-shift generators: y ^= y << a; y ^= y >> b; y ^= y << c; l-003
4e+09
3e+09
2e+09
1e+09
0e+00
0e+00 1e+09 2e+09 3e+09 4e+09 l-028
4e+09
3e+09
2e+09
1e+09
0e+00
0e+00 1e+09 2e+09 3e+09 4e+09 l-007
4e+09
3e+09
2e+09
1e+09
0e+00
0e+00 1e+09 2e+09 3e+09 4e+09 l-077
4e+09
3e+09
2e+09
1e+09
0e+00
0e+00 1e+09 2e+09 3e+09 4e+09
Nelson H. F. Beebe (University of Utah)
24 September 2011 50 / 61
Generating random integers
When the endpoints of a floating-point uniform pseudo-random number generator are uncertain, generate random integers in [ low , high ] like this: func irand(low, high) \
{
# Ensure integer endpoints low = int(low) high = int(high)
# Sanity check on argument order if (low >= high) return (low)
# Find a value in the required range n = low - 1 while ((n < low) || (high < n)) \ n = low + int(rand() * (high + 1 - low)) return (n)
} for (k = 1; k <= 20; ++k) print irand(-9,9), ""
-9 -2 -2 -7 7 9 -3 0 4 8 -3 -9 4 7 -7 8 -3 -4 8 -4 for (k = 1; k <= 20; ++k) print irand(0, 10^6), ""
986598 580968 627992 379949 700143 734615 361237
322631 116247 369376 509615 734421 321400 876989
940425 139472 255449 394759 113286 95688
Nelson H. F. Beebe (University of Utah)
24 September 2011 51 / 61
Generating random integers in order
% hoc func bigrand() { return int(2^31 * rand()) }
# select(m,n): select m pseudo-random integers from (0,n) in order proc select(m,n) \
{ mleft = m remaining = n for (i = 0; i < n; ++i) \
{ if (int(bigrand() % remaining) < mleft) \
{ print i, "" mleft--
} remaining--
} println ""
}
See Chapter 12 of Jon Bentley, Programming Pearls , 2nd ed.,
Addison-Wesley (2000), ISBN 0-201-65788-0. [ACM TOMS 6 (3),
359–364, September 1980].
Nelson H. F. Beebe (University of Utah)
24 September 2011 52 / 61
Generating random integers in order . . .
Here is how the select() function works: select(3,10)
5 6 7 select(3,10)
0 7 8 select(3,10)
2 5 6 select(3,10)
1 5 7 select(10,100000)
7355 20672 23457 29273 33145 37562 72316 84442 88329 97929 select(10,100000)
401 8336 41917 43487 44793 56923 61443 90474 92112 92799
24 September 2011
5604 8492 24707 31563 33047 41864 42299 65081 90102 97670
53 / 61
Testing pseudo-random number generators
Most tests are based on computing a
χ
2 theoretical values.
If one gets values p < 1% or p > measure of computed and
99% for several tests, the generator is suspect.
Marsaglia Diehard Battery test suite (1985): 15 tests.
Marsaglia/Tsang tuftest suite (2002): 3 tests.
All produce p values that can be checked for reasonableness.
These tests all expect uniformly-distributed pseudo-random numbers.
Nelson H. F. Beebe (University of Utah)
24 September 2011 54 / 61
Testing nonuniform pseudo-random number generators
How do you test a generator that produces pseudo-random numbers in some other distribution? You have to figure out a way to use those values to produce an expected uniform distribution that can be fed into the standard test programs.
For example, take the negative log of exponentially-distributed values, since − log ( exp ( − random )) = random.
For normal distributions, consider successive pairs
2-dimensional vector, and express in polar form ( r ,
(
θ x
)
,
: y
θ
) as a is then uniformly distributed in [ 0, 2
π
) , and
θ
/ ( 2
π
) is in [ 0, 1 ) .
Nelson H. F. Beebe (University of Utah)
24 September 2011 55 / 61
The Marsaglia/Tsang tuftest tests
Just three tests instead of the fifteen of the Diehard suite: o b’day test (generalization of Birthday Paradox).
o Euclid’s (ca. 330–225BC) gcd test.
o Gorilla test (generalization of monkey’s typing random streams of characters).
Nelson H. F. Beebe (University of Utah)
24 September 2011 56 / 61
Digression: The Birthday Paradox
The birthday paradox arises from the question How many people do you need in a room before the probability is at least half that two of them share a birthday?
The answer is just 23, not 365 / 2 = 182.5.
The probability that none of n people are born on the same day is
P ( 1 ) = 1
P ( n ) = P ( n − 1 ) × ( 365 − ( n − 1 )) / 365
The n -th person has a choice of 365 − ( n − 1 ) days to not share a birthday with any of the previous ones. Thus, ( 365 − ( n − 1 )) / 365 is the probability that the n -th person is not born on the same day as any of the previous ones, assuming that they are born on different days.
Nelson H. F. Beebe (University of Utah)
24 September 2011 57 / 61
Digression: The Birthday Paradox . . .
Here are the probabilities that n people share a birthday (i.e., 1 − P ( n ) ):
% hoc128
PREC = 3 p = 1 for (n = 1;n <= 365;++n) \
{p *= (365-(n-1))/365; println n,1-p}
1 0
2 0.00274
3 0.00820
4 0.0164
...
22 0.476
23 0.507
24 0.538
...
100 0.999999693
...
P ( 365 ) ≈ 1.45
× 10
− 157
[cf. 10
80 particles in universe].
Nelson H. F. Beebe (University of Utah)
24 September 2011 58 / 61
Digression: Euclid’s algorithm (ca. 300BC)
This is the oldest surviving nontrivial algorithm in mathematics.
func gcd(x,y) \
{ ## greatest common denominator of integer x, y r = abs(x) % abs(y) if (r == 0) return abs(y) else return gcd(y, r)
} func lcm(x,y) \
{ ## least common multiple of integer x,y x = int(x) y = int(y) if ((x == 0) || (y == 0)) return (0) return ((x * y)/gcd(x,y))
}
Nelson H. F. Beebe (University of Utah)
24 September 2011 59 / 61
Digression: Euclid’s algorithm . . .
Complete rigorous analysis of Euclid’s algorithm was not achieved until
1970–1990!
The average number of steps is
A ( gcd ( x , y )) ≈ ( 12 ln 2 ) /
π
2
≈ 1.9405 log
10 y ln y and the maximum number is
M ( gcd ( x , y )) = b log
φ
(( 3 −
φ
) y ) c where
φ
= ( 1 +
≈ 4.785 log
10 y + 0.6723
√
5 ) / 2 ≈ 1.6180 is the golden ratio.
Nelson H. F. Beebe (University of Utah)
24 September 2011 60 / 61
Hardware generators
Cryptographic accelerator boards (IBM, Sun/Oracle, . . . )
Quantis PCI card (2004) ( http://www.randomnumbers.info/ )
863 328 405 985 310 188 300 795 5 886
84 210 411 664 264 438 221 561 756 152
617 652 112 316 551 102 682 2 851 425
New Intel hardware: RdRand for 16-, 32-, and 64-bit random values
[IEEE Spectrum September 2011]
Nelson H. F. Beebe (University of Utah)
24 September 2011 61 / 61