I got a little bit of help off-forum and wanted to post the results here. Apparently sampled statistics over estimate uniqueness, so the algorithm fudges the numbers slightly to show higher duplicates in an attempt to compensate (there are exceptions for columns that have constraints that enforce uniqueness of course). This would explain why the statistics I saw were outside the bounds of anything stored in the table. Unfortunately, I don't have a blog or white paper to reference - if anyone has a source I can link to, please let me know.

Thanks!

Chad