Usenet 1994 January/usenetsourcesnewsgroupsinfomagicjanuary1994.iso/sources/unix/volume25/mrandom/README

Text File | 1991-12-12 | 1.6 KB | 33 lines

I wrote this package to overcome some troubles I had with the random()
package.  I had been saving random()'s state table to a disk file, then
restarting the random sequence in the next program run.  I discovered some
seriously non-random behavior in the numbers resulting from this practice.
Further investigation (including examination of the object code for
random() -- painful!) showed me that it is necessary to count the number of
calls to random() in order to safely restart it.  Hence the enclosed code.

I believe random(), with my modifications, is superior in many respects
to its competition within 4.3bsd Unix, namely rand() and rand48().
Those generators are based on a multiplicative congruential scheme,
which makes them difficult to use in applications where one is generating
points uniformly distributed on the unit square.  Certainly rand() has
the problem that a high-resolution 2-d plot of the x-y points generated by
    for (i=0; i<n; i++) {
      x[i] = rand()/MAXLONG;
      y[i] = rand()/MAXLONG;
    }
will show stripes.  Using rand()%m is suicidal, since its low-order bits
are known to be cyclic.  Perhaps I'm selling rand48() short, since I haven't
examined it in detail: quite possibly its constants were chosen to minimize
the separation between the stripes in a 2-d plot.

To help you get started, I have also enclosed a test-driver called mrtest.

Please let me know if you find any defects in the sequence generated
by the default seed built into mrtest.  I picked that seed for
sentimental reasons: it was the last number generated by my old,
and now discredited, random() sequence.

                        Clark Thomborson
                        [e-mail]