[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: fun with random numbers

In article <38BD4DB6.E9F60A17@phys.ucalgary.ca>,
  Brian Jackel <bjackel@phys.ucalgary.ca> wrote:
> Greetings
> We've just spent a couple hours figuring out why a Monte-Carlo
> simulation was giving peculiar (ie. wrong) results.
> The "test" procedure creates two random variables and prints
> them. Called twice, you might think that the results would be
> totally different.  Here's an example result:
> x     0.120838
> y     0.649213     0.577647     0.139354     0.745442
> x     0.649213
> y     0.577647     0.139354     0.745442     0.942317
showing how the "random" numbers are merely being shifted by one


This bug has existed since IDL V5.1.1 and continues into IDL V5.3.    It
turns out that there were *two* RANDOMU bugs introduced into V5.1.1.
The first one was that the SEED variable was being initialized to the
same value at the start of each session.   This bug was fixed in V5.2.1.
But the bug described above by Brian Jackel appears to continue into

Below I update my epic on the history of RANDOMU problems.    Note that
I also give the wrapper RANDOM() routine suggested by Pat Broos to fix
the problem with multiple RANDOMU calls described above.

--Wayne Landsman                    landsman@mpb.gsfc.nasa.gov


 V4.0.1:    No problems?  (But the algorithm was rumored to be far from
state of the art.)

 V5.0:    RANDOMU could yield a non-random distribution if two programs
using RANDOMU are interleaved.    For example, in the program  demo.pro
given at the bottom of this message, the command demo,/breakit will show
a significant excess in the distribution of random  numbers between 0
and 0.03.

 V5.1:      A *negative* seed value must be specified if you want to
            the same "random" sequence

               IDL> seed = 2 & print, randomu(seed, 3)
                     0.0594004    0.982075    0.358593
               IDL> seed = 2 & print, randomu(seed, 3)
                     0.831999    0.303037    0.506712


               IDL>  seed = -2 & print, randomu(seed, 3)
                       0.342299    0.402381    0.307838
               IDL> seed = -2 & print, randomu(seed, 3)
                       0.342299    0.402381    0.307838

        This isn't necessarily a bug, but it means that RANDOMU works
        differently in V5.1 than in all other IDL versions.

V5.1.1 and V5.2:

 (1) The seed variable is now initialized to the same value at the start
of each session rather than the system clock.  Thus, Monte  Carlo
simulations from different IDL sessions, might yield decidedly unrandom

 (2) Perhaps more insidious, only the first call to RANDOMU is
initialized inside a program.    Thus, if one calls the following
program test.pro multiple times, you will see that the "random" vector
is simply the vector on the  previous call, shifted by one.

 PRO test
 print, randomu(seed)
 print, randomu(seed,6)

V5.2.1 and V5.3

  Problem (1) in V5.1.1 has been fixed -- the seed variable is correctly
initialized.   But problem (2) concerning multiple RANDOMU calls inside
a program remains.  For this last problem, Pat Broos suggests using the
following wrapper program to RANDOMU to store the seed value in a common

 FUNCTION random, n1, n2, n3, NEW_SEED=new_seed, _EXTRA=extra

 COMMON random_seed, seed

 if keyword_set(new_seed) then seed = long(new_seed)

 case n_params() of
     0: return, randomu(seed,          _EXTRA=extra)
     1: return, randomu(seed,n1,      _EXTRA=extra)
     2: return, randomu(seed,n1,n2,    _EXTRA=extra)
     3: return, randomu(seed,n1,n2,n3, _EXTRA=extra)



 FUNCTION lib_random

 return, randomu(other_seed,1)

 PRO demo, x, BREAK_IT=break_it

 ; Type demo,/breakit to see the "non-random" distribution that can
result in ;
 V5.0.  Works correctly in earlier and later IDL versions

 x = fltarr(100000)

 for ii = 0L, n_elements(x)-1 do begin
   x(ii) = randomu(seed,1)

   if keyword_set(break_it) then dummy = lib_random()

 h = histogram( x, MIN=0.0, BIN=0.01 )
 plot, h, PSYM=10
 print, h

Sent via Deja.com http://www.deja.com/
Before you buy.