Distributions

One shot

ransu.normal(170, 6); // 168.42…
ransu.poisson(4); // 3
ransu.binomial(100, 0.3); // 28
ransu.exponential(2);
ransu.zipf(1.2, 1000);

Reusable samplers

Same name, two shapes: positional arguments give a number, an options object gives a sampler. The samplers live in ransu/distribution.

import { normal, categorical } from "ransu/distribution";

const height = normal({ mean: 170, sd: 6 });
height.sample(); // one draw
height.samples(10_000); // a Float64Array
height.mean; // 170
height.variance; // 36

const loot = categorical({ weights: [70, 25, 5] }); // O(1) per draw

Build the sampler when you draw more than once: it validates its parameters and precomputes its constants up front.

What is available

Continuous — uniform, normal, logNormal, exponential, gamma, beta, chiSquared, studentT, fisherF, cauchy, laplace, logistic, gumbel, pareto, weibull, rayleigh, triangular, irwinHall, bates, vonMises, dirichlet.

Discrete — bernoulli, discreteUniform, categorical, binomial, poisson, geometric, negativeBinomial, hypergeometric, zipf, multinomial.

Algorithms

Distribution	Method
normal	128-level ziggurat, tables built numerically at load
gamma	Marsaglia–Tsang, with a boost for shapes below 1
binomial	CDF inversion when n·p < 10, Hörmann’s BTRS above
poisson	Knuth when λ < 30, Hörmann’s PTRS above
categorical	Vose’s alias method
hypergeometric	Sequential urn sampling — exact
Everything else	Inverse transform, or composition of the above

The rejection methods compare against the exact log-pmf rather than a Stirling approximation. The rejection rate is low enough that the extra cost does not matter, and it removes any chance of a mis-transcribed constant.

How they are checked

Matching the mean does not prove a distribution is right. CI runs:

• Kolmogorov–Smirnov against the exact CDF for the continuous families.
• Chi-square against the exact pmf for the discrete ones — separately for each algorithm branch, so both the inversion and the rejection paths are covered.
• For the normal, the 3σ and 4σ tail rates, plus skewness and kurtosis. The ziggurat only enters its tail routine beyond 3.44σ, so a moment test alone would never notice a broken tail.

Passing an engine

const r = new Random(42);
normal({ mean: 0, sd: 1, engine: r.engine }).sample();

Without engine, a sampler draws from the global stream, so seed() makes it reproducible too.