[cfe-commits] [libcxx] r103983 - in /libcxx/trunk: include/ test/numerics/rand/rand.dis/rand.dist.norm/rand.dist.norm.cauchy/

[Howard Hinnant]

On May 17, 2010, at 6:03 PM, Anton Korobeynikov wrote:

> Hello, Howard
> 
>> [rand.dist.norm.cauchy].  I'm having trouble testing the output as all statistical properties are undefined.  They do not converge upon any one value as the number of samples increases.  Suggestions for tests welcome.
> Several suggestions here:
> 
> 1. If you remove the smallest and the largest observation, then the
> resulting random variable will have some mean. If you remove 2
> observations more, then there will be some variance as well
> 2. Why don't use some omnibus goodness-of-fit test? E.g.
> Kolmogorov-Smirnov or Cramer-von Mises? You need to simulate the
> sample of p-levels, which should be uniformly distributed.

Thanks, this helped.  I didn't do exactly this, but reading up on your suggestions helped me to realize a fairly simple but good test.

-Howard