Reproducible uniforms shared with morie's Python arm (splitmix64)

Returns n pseudo-random numbers on [0, 1) from a 64-bit seed with the splitmix64 generator, the same numbers morie's Python computes for the same seed. Use it where the R and Python arms must draw identical "random" values so that their results agree: the within-day jitter of tied event dates in the TPS Hawkes fits, or a subsample.

Usage

core_uniforms(n, seed)

Arguments

n: Number of uniforms, one non-negative whole number (0 gives numeric(0)).
seed: The seed, one whole number in \([0, 2^{53})\) (so that it is exact as an R double).

Value

A numeric vector of length n, values in [0, 1).

Details

splitmix64 (Steele, Lea & Flood 2014) advances a 64-bit state by the constant 0x9E3779B97F4A7C15 and mixes it with two multiply-xorshift rounds; the top 53 bits of each output divided by \(2^{53}\) give a double in [0, 1). The stream depends only on seed: R's own generator and set.seed() are neither used nor changed. It is not a statistical replacement for R's generators.

Errors: n not one non-negative whole number; seed not one whole number in \([0, 2^{53})\).

References

Steele GL, Lea D, Flood CH (2014). Fast splittable pseudorandom number generators. OOPSLA 2014, 453–472. doi:10.1145/2660193.2660195

Examples

core_uniforms(3, 42)
#> [1] 0.7415649 0.1599104 0.2786011
identical(core_uniforms(5, 1), core_uniforms(5, 1))
#> [1] TRUE

# a stream is a prefix of a longer one with the same seed
all.equal(core_uniforms(3, 7), core_uniforms(10, 7)[1:3])
#> [1] TRUE

# R's own generator is left alone
set.seed(1); a <- runif(1)
set.seed(1); invisible(core_uniforms(100, 9)); b <- runif(1)
identical(a, b)
#> [1] TRUE

# negative counts and seeds past 2^53 are refused
try(core_uniforms(-1, 1))
#> Error : `n` must be a non-negative whole number
try(core_uniforms(2, 2^53))
#> Error : `seed` must be a whole number in [0, 2^53)