Cesar199999
commented
5 months ago Hey @Pratyush, @DimitrisPapac and I have been looking at the SparseMultilinearExtension::evaluate method in poly/src/evaluations/multivariate/multilinear/sparse.rs and we are not sure what you meant by "linear in the size of the hypercube" vs "linear in the number of non-zero coefficients". In SparseMultilinearExtension, the polynomial is represented as a list of non-zero evaluations, which is also the definition of sparsity adopted throughout Arkworks. So, did you mean "linear in the size of the hypercube" vs "linear in the number of non-zero evaluations"?
If that's the case, then the current implementation is already linear in the number of non-zero evaluations. The method fix_variables chooses the window size dynamically as log2 of the number of non-zero evals. When the number of non-zero evals is small/large, it behaves like the naive sparse/dense algorithm. It does so in a pretty optimal way from the looks of it. For reference, we ran some benchmarks using the naive sparse algorithm, i.e., the following code snippet which is linear in the number of non-zero evaluations and the number of variables:
self.evaluations.iter()
.map(|(&i, &v): (&usize, &F)| v * (0..self.num_vars)
.map (|k| (i >> k) & 1)
.enumerate()
.map(|(j, bit)| if bit == 1 { point[j] } else { F::ONE - point[j] })
.product::<F>()
)
.sum()and we observed that the current implementation is much more efficient:
SparseMultilinear/Eval with num_vars = 12/64
time: [27.815 µs 28.005 µs 28.236 µs]
change: [+183.54% +187.82% +192.19%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 13/64
time: [31.129 µs 31.659 µs 32.155 µs]
change: [+172.20% +179.54% +186.73%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 14/128
time: [67.307 µs 68.197 µs 69.064 µs]
change: [+214.55% +222.07% +229.83%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 15/128
time: [69.144 µs 69.678 µs 70.292 µs]
change: [+220.91% +226.30% +231.55%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 16/256
time: [153.67 µs 155.13 µs 156.80 µs]
change: [+247.99% +260.19% +272.32%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 17/256
time: [168.49 µs 170.16 µs 171.91 µs]
change: [+320.87% +326.18% +331.39%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 18/512
time: [361.08 µs 365.49 µs 370.68 µs]
change: [+358.08% +369.54% +382.02%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 19/512
time: [387.14 µs 392.74 µs 398.55 µs]
change: [+342.45% +354.34% +365.11%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 20/1024
time: [813.39 µs 825.67 µs 838.82 µs]
change: [+387.20% +400.18% +415.22%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 21/1024
time: [886.92 µs 898.19 µs 911.98 µs]
change: [+413.20% +424.94% +436.74%] (p = 0.00 < 0.05)
Performance has regressed.
SparseMultilinear/Eval with num_vars = 22/2048
time: [1.7714 ms 1.7876 ms 1.8048 ms]
change: [+424.87% +434.76% +443.92%] (p = 0.00 < 0.05)
Performance has regressed.Did you have any other algorithm in mind? If so, could you please point us to a paper or any other resource where we can read about it? Thanks!
It takes time that is linear in the size of the hypercube, as opposed to linear in the number of non-zero coefficients.