Tests

[vkhodygo]

We employ some data and techniques provided by NIST for existing tests. However, the problem is still there:

• GSL uses some magic numbers to check how accurate the functions are, still, there is no justification for that. In addition, higher accuracy is possible (say 1e-16 instead of 1e-15);
• There is no simple implementation of a generator of ill-conditioned sum.