Extended statistics
===================

When estimating various quantities (e.g. condition selectivities) the default
approach relies on the assumption of independence. In practice that's often
not true, resulting in estimation errors.

Extended statistics track different types of dependencies between the columns,
hopefully improving the estimates and producing better plans.

Currently we only have one type of extended statistics - ndistinct
coefficients, and we use it to improve estimates of grouping queries. See
README.ndistinct for details.


Size of sample in ANALYZE
-------------------------
When performing ANALYZE, the number of rows to sample is determined as

    (300 * statistics_target)

That works reasonably well for statistics on individual columns, but perhaps
it's not enough for extended statistics. Papers analyzing estimation errors
all use samples proportional to the table (usually finding that 1-3% of the
table is enough to build accurate stats).

The requested accuracy (number of MCV items or histogram bins) should also
be considered when determining the sample size, and in extended statistics
those are not necessarily limited by statistics_target.

This however merits further discussion, because collecting the sample is quite
expensive and increasing it further would make ANALYZE even more painful.
Judging by the experiments with the current implementation, the fixed size
seems to work reasonably well for now, so we leave this as a future work.