Benoit Jacob 95d88e1327 Big reworking of ColPivQR and its unit test, which now passes even with thousands of repetitions and correctly tests matrices of all sizes. Several surprises along the way: for example, a major cause of trouble was the optimized "table of column squared norms" where the accumulation of imprecision was a serious issue; another surprise is that tests like "x!=0" before dividing by x really benefit from being replaced by fuzzy tests, as i hit real cases where i got wrong results in 1/epsilon.

2009-12-01 13:26:29 -05:00

5.6 KiB

Raw Blame History

View Raw