Skip to content

random.md

Latest commit

 

History

History
35 lines (26 loc) · 1.73 KB

random.md

File metadata and controls

35 lines (26 loc) · 1.73 KB
Raw

Random

Rand. Sampler-struct pattern: IfloatSampler { float Next(ref Random rng); }, implemented by floatUniform/floatExponential/floatRayleigh/floatWeibull/floatCauchy/floatLogistic/ floatPareto/floatTriangular/floatGaussian (stateful Box-Muller - pass ref, it caches a spare value between calls).

Filling vectors/matrices

  • Rand.randomInPlace<S>(ref Random rng, ref floatN dest, ref S sampler) - generic over any sampler struct (vector and matrix overloads).
  • Rand.nextUniformInPlace(ref Random, ref dest) / (ref Random, ref dest, min, max) - uniform shortcut.
  • Integer/bool refill: nextUniformInPlace(ref Random, ref intN dest, min, max), nextBernoulliInPlace(ref Random, ref boolN dest, float p), nextBoolInPlace (fair coin).

Picking & shuffling

weightedPick(in weights, ref rng) : int, weightedPickInPlace(in weights, ref Indices dest, ref rng), randomPermutationInPlace(ref Pivot, ref rng), shuffleInPlace(ref Indices, ref rng), sampleKWithoutReplacementInPlace(ref Indices dest, int n, ref rng).

Multivariate normal & structured matrices

multivariateNormalInPlace(ref rng, in cholL, in mean, ref dest) (+ row-batch multivariateNormalRowsInPlace) - caller factors Σ once via CHO.decomp, then draws as many samples as needed. Property-matrix generators (all validate before allocating Temp scratch): Rand.orthogonalInPlace (Haar-uniform, Mezzadri sign-fixed QR), Rand.spdInPlace(..., minEig, maxEig), Rand.conditionedInPlace(..., cond), Rand.withRankInPlace(..., rank), Rand.stochasticInPlace.

Performance

Box-Muller Gaussian sampling calls math.sin and math.cos separately rather than math.sincos, which roughly doubles its trig cost; every other sampler is unaffected.