Comment by archi42

I loved this, since this resonates with my own experience. If the problem space is small enough, just test against a known good (but slower) implementation. With custom algorithms this can also be very interesting for the person writing the test: Their reference code needs to be correct beyond doubt, but can forfeit mostly every manual optimization. Also, for larger input spaces I like to do a nested for-loop over vectors of "interesting" values, plus add a large set of random values.