What are the gory details behind DBMS_STATS.AUTO_SAMPLE_SIZE?

I am curious how the algorithms behind DBMS_STATS.AUTO_SAMPLE_SIZE actually pick a sample size (10.2.0.1). We have witnessed a case where the sample size for a particular column was reduced 91.6%, causing the number of distinct values to be inaccurate, further causing a query to choose an FTS rather than an index. As far as I can tell, the skew has not changed that much, if at all, and the overall volume has increased by ~1%.

I am aware of workarounds (compute stats, lock stats, use an outline, etc), but like I said, I am really curious why the algorithm made such a drastic jump in the first place. I could not easily locate an appropriate white paper on metalinks, and I get too many hits on Google. Can anyone help satiate my burning desire to know? =)

charles schultz
oracle dba
aits - adsd
university of illinois

--
http://www.freelists.org/webpage/oracle-l