Random Number Generation (RNG)

read “Numerical Recipes” on random numbers and

the chi-squared test.

Today we discuss how to generate and test random numbers.

What is a random number?

– A single number is not random. Only an infinite
sequence can be described as random.
– Random means the absence of order. (Negative property).
– Can an intelligent gambler make money by betting on
the next numbers that will turn up?
– All subsequences are equally distributed. This is the
property that MC uses to do integrals and that you will
test for in homework.
“Anyone who considers arithmetical methods of random digits is, of course, in
a state of sin.” John von Neumann (1951)

“Random numbers should not be chosen at random.” Knuth (1986)

1
 Random numbers on a computer
• Truly Random - the result of a physical process such as
timing clocks, circuit noise, Geiger counts, bad memory
– Too slow (we need 1010/sec) and expensive
– Low quality
– Not reproducible. But new INTEL chip does this for security
• Pseudo-random. prng (pseudo means fake)
– Deterministic sequence with a repeat period but with the
appearance of randomness (if you don’t know the
algorithm).
• Quasi-random (quasi means almost random)
– “half way” between random and a uniform grid.
– Meant to fill space with max. distance between space to fill
“holes” sequentially (e.g., [0,100] as 0,1,2,3,…,100).

If doing a low-D integral, QRNG gives up correlation so the sequence will be

more uniform than PRNG and converge to mean value quicker than N-1.2.

2
 Intel’s RNG: Ivy Bridge
• Article describing it: 3 Gbits/sec
• http://spectrum.ieee.org/computing/hardware/behind-intels-
new-randomnumber-generator/0

3
 Desired Properties of RNG on a computer

• Deterministic: easy to debug (repeatable simulations).

• Long Periods: cycle length is long.

• Uniformity: RN generated in space evenly.

– Go through complete cycle, all integers occur once!
– TEST: N occurrences of x. The no. in (a,b) is equal to
N(b-a) ± [N(b-a)]-1/2

• Uncorrelated Sequences:
<f1(xi+1)…fk(xi+k)>= <f1(xi+1)>…<fk(xi+k)>
Monte Carlo can be very sensitive to such
correlations, especially for large k.

• Efficiency: Needs to be fast.

4
Pseudo Random Sequence

S: State and initial seed.

T: Iteration process,
F: Mapping from state to integer RN (I ) or real RN (U).

5
 Period or Cycle Length (L)
• If internal state has M bits, total state space is 2M values.
• If mapping is 1-1, space divides into a finite no. of cycles.
• Best case is a single cycle of length 2M, otherwise 2M/L.
• It is easy to achieve very long cycle length: 32 or 24 bit
generators are not long enough!
• Entire period of the RNG is exhausted in:
• rand (1 processor) ~ 100 second
• drand48 (1 processor) ~ 1 year
• drand48 (100 processor) ~ 3 days
• SPRNG LFG (105 procs) ~ 10375 years
Assuming 107 RN/sec per processor

e.g., In 1997, computer has ~108 RN/proc/sec (or 226).

Hence, it takes 1 sec to exhaust generator with 26 bits and 1 year for 251 states.
e.g., consider 1024x1024 sites (spins) on a lattice. With MC, sweep 107 times to get averages.
You need to generate 1013 RN. A 32-bit RNG from 1980’s dangerous on today’s computers.
6
 Common PRNG Generators
Linear Congruential Generator (LCG) modulo (mod) is remainder after division
64-bit word (or 2 32-bit LCG) give a period of ~1018.
• Multiplicative Lagged zn = zn-k * zn-l
Fibonacci vary
zn = zn-k + zn-l initialization
• Modified Lagged
Fibonacci (modulo 2m )

• 48 bit LCG
zn = a*zn-1 + p vary p
• 64 bit LCG
• Prime Modulus LCG (modulo m )

vary a

• Combined Multiple zn = an-1*zn-1 + ...+

Recursive vary LCG
an-k*zn-k + LCG

Recurrence Parallelization
7
 Uniformity
• Output consists of taking N bits from state and making
an integer in (0,2N-1) “Ik”.

• One gets a uniform real “Uk” by multiplying by 2-N.

• We will study next time how to get other distributions.

• If there is a single cycle, then integers must be uniform

in the range (0,2N-1).

• Uniformity of numbers taken 1 at a time is usually easy to

guarantee. What we need is higher dimensional uniformity!

8
 Sequential RNG Problems
• Correlations non-uniformity in higher dimensions

Uniform in 1-D but

non-uniform in 2-D

This is the important property

to guarantee:
< f1 (xi+1 ) f2 (xi+2 ) >=< f1 (xi+1 ) >< f2 (xi+2 ) >
MC uses numbers many at a time--they need to be uniform.
e.g., 128 atom MC moving 3N coords. at random needs 4 PRNG (x,y,z, i) or
128 x 4=1024. So every 1024 moves may be correlated!
Compare error for each PRNG, if similar OK. (Use χ2-test.)
12
 LCG Numbers fall in planes.
Good LCG generators have the planes close together.
For a k-bit generator, do not use them more than k/2 together.

13
 SPRNG: originally a NCSA project

• A library of many well tested excellent parallel PRNGs

• Callable from FORTRAN.C, C++, and JAVA
• Ported to popular parallel and serial platforms

SPRNG Functions
• int *init_sprng(int streamnum, int nstreams, int seed, int param)
• double sprng(int *stream)
• int isprng(int *stream)
• int print_sprng(int *stream)
• int make_sprng_seed()
• int pack_sprng(int *stream, char **buffer)
• int *unpack_sprng(char *buffer)
• int free_sprng(int *stream)
• int spawn_sprng(int *stream, int nspawned, int ***newstreams)

14
 What is a chisquare test?
• Suppose { x1,x2,….} are N Gaussian random numbers with
mean zero, and variance 1
N
• What is the distribution of y = ∑ x2 ?
k =1
• Q( χ | N )
2
is the probability that y>c2

• The mean value of y is N, its variance is 2N.

• We use it to test hypotheses. Is the fluctuation we see, what

we expect? N
( x −x )
2
fit

χ =∑
2 k k

σ k2
k=1

• Roughly speaking we should have N − 2N < χ 2 < N + 2N

• More precisely look up value of Q( χ 2 | N ) “chisquare

function with N degrees of freedom”
15
 Chi-squared test of randomness
• HW exercise: do test of several generators
– Divide up “cube” into “N” bins.
– Sample many “P” triplets.
N=100
– Number/bin should be n=P/N ±(P/N)1/2
In example: expected n = 10 ±(10)1/2 P=1000

GOOD BAD

16
 See “Numerical Recipes” on reserve or on-line
(link on class webpages)

(N i − ni )2
• Chi-squared statistic is χ = ∑i
2

ni
– ni (Ni) is the expected (observed) number in bin i.
– Ni is an integer! (Omit terms with 0 = ni = Ni)
– Term with ni = 0 and Ni ≠ 0, correctly give χ2 = ∞.)
– Large values of χ2 indicate the “null” hypothesis, i.e. unlikely.

• The chi-squared probability is Q(χ2|N-1) The incomplete Γ-fct

(N-1) because of “sum rule”. N.R. Sec. 6.2 and 14.3

• Table may help with interpreting the chi-squared probability.

Probability Interpretation
<1% reject
1-5% suspect
5 - 10 % almost suspect
10 - 90 % pass
90 - 95 % almost suspect
95 - 99 % suspect
17
> 99 % reject
Recommendation:
For careful work use several generators!
• For most real-life MC simulations, passing existing
statistical tests is necessary but not sufficient.

• Random number generators are still a black art.

• Rerun with different generators, since algorithm

may be sensitive to different RNG correlations.

• Computational effort is not wasted, since results

can be combined to lower error bars.

• In SPRNG, relinking is sufficient to change RNGs.

18
 Parallel Simulations

• Parallel Monte Carlo is easy? Or is it?

• Two methods for easy parallel MC:

– Cloning: same serial run, different random numbers.
– Scanning: different physical parameters (density,…).

• Any parallel method (MPI, ..) can be used.

• Problems:
– Big systems require excessive wall clock time.
– Excessive amounts of output data generated.
– Random number correlation?

20
 Parallelization of RNGs

• Cycle Division: Leapfrog or Partition. Correlation problems

• Cycle Parameterization: If PRNG has several cycles,

assign different cycles to each stream

Set of possible states Each cycle gets a disjoint subset

• Iteration Parameterization: Assign a different iteration

function to each stream

21
 Examples of Parallel MC codes and Problems

• Embarrassingly parallel simulations. Each processor has its

own simulation. Scanning and cloning.
Unlikely to lead to problems unless they stay in phase.

• Lots of integrals in parallel (e.g. thousands of Feynman

diagrams each to an accuracy of 10-6).
Problem if cycle length is exhausted.

• Particle splitting with new particles forking off new

processes. Need lots of generators.
Problem if generators are correlated initially.

• Space-time partitioning. Give each local region a

processor and a generator.
Problem if generators are correlated.

23