add docs on FracMinHash downsampling

[ctb]

@drtamermansour asked some questions on slack about how FracMinHash signatures with different scaled values are handled in practice, and I took a look in the docs and couldn't find anything that was clearly written. We should add that somewhere.

(On the plus side, it's pretty well tested, I think?)

Off the top of my head,

• for most purposes, when there is a difference between the query and a subject signature, the query and signature are downsampled to the same scaled, i.e. scaled is increased to the same value. This results in a loss of resolution in situations where the signature gets modified for further searching.
• this loss of resolution can be ...problematic when searching multiple databases with gather, in particular; if there's a match to a low rez signature, then the query will be appropriately downsamples and will forevermore be low rez.
• also, there are some databases that cannot be downsampled properly, like SBTs in particular.

This was all actually written up internally in the code base - see #407 and PR #1420 - but the details didn't make it into the docs. Oops!

[drtamermansour]

In the current implementation, when there is a difference between the query and a subject signature, sourmash rescale the DB but not the sample.

I tried:

• sample scale=500 & DB scale=1000 ==> runtime error (ValueError: new scaled 500 is lower than current sample scaled 1000)
• sample scale=2000 & DB scale=1000 ==> works fine

[ctb]

Yep, SBTs work that way. (but that will depend to some extent on the database type in question - there are several kinds, including SBTs, LCAs, and collections of signatures)

…

On Jan 20, 2022, at 2:08 PM, Tamer Mansour ***@***.***> wrote: In the current implementation, when there is a difference between the query and a subject signature, sourmash rescale the DB but not the sample. I tried: • sample scale=500 & DB scale=1000 ==> runtime error (ValueError: new scaled 500 is lower than current sample scaled 1000) • sample scale=2000 & DB scale=1000 ==> works fine — Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.