1. Convex hull sampled databases problem:

The convex hull polyhedral databases described in 6_subsec:polyhedral_databases contain fewer tests than the other databases, thus sometimes small database histograms like 6_fig:SPEC95_convex_hull_dimension_C3C64_C3P64_100 do not make a good distribution. Exceptionally, the comparisons between the partial factorization and New POLKA only counts three tests in 6_fig:PerfectClub_convex_hull_dimension_C3C64_PK64_100, hence their evaluation is meaningless . Therefore the full databases are used instead, though it requires much longer execution times since convex hull is an expensive operator.

Figure 35: SPEC95: Dimension C3C64 vs C3P64 in sampled database - too few tests

Figure 36: PerfectClub: Dimension C3C64 vs PK64 in sampled database - too few tests