Lots of ways to do it, depending on needed quality. The one I use for cross-platform stability in testing is the Middle-square method. It's actually a pretty bad RNG if you need any real rigor or a very large number of points, but it's fast, intuitive and insensitive to round-off.

my $seed = 21324556;
sub _prng {
    # von Neumann middle-square method
    return 1e-8*($seed = int(1e-4 * $seed**2) % 1e8);
}
[download]

This seed converges to zero after 16348 iterations, with an average in that population of 0.5002 and a standard deviation of 0.2878. If you go for more than 8 digits of precision, you will start seeing machine dependence.