negative distance returned in IndexFlatL2 range_search query

[joshim5]

Summary

This issue was raised in #297 for search. Here I report the same issue still occurs in range_search.

Platform

OS: Ubuntu 18.04

Faiss version: e8c8db8

Faiss compilation options: CUDA 10.1

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

Perhaps this was fixed for index.search(Data, 10), but I get negative distances for index.range_search. Here's an example:

lims, D, I = index.range_search(data, 555)
>>> min(D)
-3145728.0
>>> max(D)
512.0

Is this numerical overflow?

[mdouze]

Could you post some code to reproduce?

[mdouze]

Seems to be an issue related to badly conditioned data, see
https://github.com/facebookresearch/faiss/wiki/FAQ#why-do-i-get-weird-results-with-brute-force-search-on-vectors-with-large-components