Reversible random number generator

Implementation of the Mersenne-Twister pseudo-random number generator of Matsumoto and Nishimura (1997) with reversible updates to the generator internal state.

That is following a sequence of samples from a generator, given only the generator internal state the exact reverse of that sequence of samples can be generated.

The reverse updates are based on James Roper's excellent blog post: https://jazzy.id.au/2010/09/25/cracking_random_number_generators_part_4.html

Both a C implementation and numpy compatible Python wrapper are provided.

Example usage

import revrng
import numpy as np

seed = 12345
n_iter = 10
rng = revrng.ReversibleRandomState(seed)
us = []
ns = []

# sample set of random uniform and normal ndarrays of increasing length
for i in range(n_iter):
    us.append(rng.standard_uniform(shape=(i,)))
    ns.append(rng.standard_normal(shape=(i,)))
    
# reverse direction of random number generator
rng.reverse()

# sample same random uniform and normal vectors in reverse
for i in range(n_iter - 1, -1, -1):
    assert np.all(ns.pop(-1) == rng.standard_normal(shape=(i,)))
    assert np.all(us.pop(-1) == rng.standard_uniform(shape=(i,)))

Alternatives

A header-only C implementation of a reversible linear congruential generator by Johan Helsing is available here, with the readme.md also giving some notes on other possible alternatives.