I have following proposal. Currently the ts_typanalyze function accumulates frequency statistics for ts_vector using lossy counting technique. But no frequency statistics is collecting over arrays. I'm going to generalize ts_typanalyze to make it collecting statistics for arrays too. ts_typanalyze internally uses lexeme comparison and hashing. I'm going to use functions from default btree and hash opclasses of array element type in this capacity. Collected frequency statistics for arrays can be used for && and @> operators selectivity estimation.