math.log domain error

[bencrts]

For the following params:

param = LWE.Parameters(
     n = 2**17, #32768
     q = 2**(3522),
     Xs = ND.Uniform(-1, 1),
     Xe = ND.DiscreteGaussian(3.19),
     m = oo
)

est = LWE.primal_bdd(param, red_cost_model = RC.MATZOV)

there is an error thrown by math.log():

[...]
~/Desktop/TBA/Github/lattice-estimator/estimator/lwe_primal.py in svp_dimension(cls, r, D)
    300
    301         d = len(r)
--> 302         r = [log(x) for x in r]
    303
    304         if d > 4096:

~/Desktop/TBA/Github/lattice-estimator/estimator/lwe_primal.py in <listcomp>(.0)
    300
    301         d = len(r)
--> 302         r = [log(x) for x in r]
    303
    304         if d > 4096:

ValueError: math domain error

this is caused by the math.log() function throwing a domain error for very small inputs.

[ludopulles]

Alternatively, you might be able to avoid this error and speedup the code if the simulator outputs log(GS norm) instead.

GSA, ZGSA and LGSA internally model the log-norms base 2, e, 2 respectively.
However, CN11 takes the GS-norms from fplll, so if that is causing the issue, this doesn't help.

[ludopulles]

It seems that the simulator from FPyLLL uses log_2(GS-norms) to simulate the profile. Perhaps it's easiest here, if you add a log2_simulator function to the FPyLLL code, and let simulator call the log2_simulator, while the lattice-estimator then directly calls the log2_simulator?

If you roll out your custom simulator in the lattice-estimator, this gives redundancy (although the FPyLLL code hasn't changed a lot the last years).

[malb]

I like this solution.

[bencrts]

I'll give it a shot and report back here :) thanks for the suggestions/discussion!