False positives about bloomfilter.js

[jasondavies]

Perhaps @Gopiandcode would be interested in this. He is the author of Bloom filters debunked: Dispelling 30 Years of bad math with Coq!

[jasondavies]

Interesting. The only way I was able to fix this was by using k independent hash functions, instead of the current version which uses 2 independent hash functions a and b internally, and combines them to generate k hash functions using:

h_i(d) = a(d) + i * b(d)

As far as I know this is common practice, and the idea is mentioned in Building a Better Bloom Filter. I'm guessing that this optimisation has some effect on the false positive rate as the k hash functions are not completely independent.

[nponeccop]

I'd suggest you try the benchmarking using a CSRNG from crypto instead of Math.random(). With such complex bit patterns, the randomness provided by simple RNGs is often not enough.