Expand description
Bayesian inference primitives.
Small, dependency-free building blocks for exact conjugate updates and for approximate inference when conjugacy runs out:
Beta— conjugate prior for a Bernoulli/Binomial success rate.Gaussian(aliasedNormal) — density, sampling, and the Normal–Normal conjugate update for an unknown mean with known variance.Gamma— conjugate prior for a Poisson rate, and the engine behind Beta sampling.Metropolis— a random-walk Metropolis(–Hastings) sampler for any unnormalised scalar log-target.special— the log-gamma/log-beta/normal-CDF machinery underneath.- Free log-likelihood helpers such as
bernoulli_log_likelihood,normal_log_likelihood, andlog_sum_exp.
Every distribution samples through the Rng and Distribution traits
from tpt-math-prob-core, so a deterministic generator such as
SplitMix64 makes any inference run exactly reproducible.
§Conjugate updates
use tpt_math_prob_bayes::Beta;
// Uniform prior over a coin's bias, then 7 heads and 3 tails.
let posterior = Beta::uniform().update(7, 3);
assert_eq!((posterior.alpha(), posterior.beta()), (8.0, 4.0));
assert!((posterior.mean() - 2.0 / 3.0).abs() < 1e-12);
// Posterior belief with 95% probability mass.
let (lo, hi) = posterior.credible_interval(0.95);
assert!(lo < posterior.mean() && posterior.mean() < hi);§Approximate inference
use tpt_math_prob_bayes::{Beta, Metropolis};
use tpt_math_prob_core::SplitMix64;
let prior = Beta::new(2.0, 2.0);
let (successes, failures) = (30, 70);
// Sample the same posterior with MCMC instead of the conjugate formula.
let target = move |p: f64| prior.log_unnormalized_posterior(p, successes, failures);
let mut sampler = Metropolis::with_gaussian_proposal(target, 0.1);
let mut rng = SplitMix64::seed_from_u64(2024);
let trace = sampler.run_with_burn_in(&mut rng, 0.5, 2_000, 40_000);
let mcmc_mean = trace.iter().sum::<f64>() / trace.len() as f64;
let exact_mean = prior.update(successes, failures).mean();
assert!((mcmc_mean - exact_mean).abs() < 0.01);Modules§
- adaptive
- Adaptive random-walk Metropolis (Haario, Saksman and Tamminen, 2001).
- diagnostics
- Convergence diagnostics for MCMC output and self-normalised importance sampling.
- hmc
- Hamiltonian Monte Carlo with a leapfrog integrator and a unit mass matrix.
- special
- Special functions backing the conjugate-prior distributions.
Structs§
- Beta
- The Beta distribution
Beta(α, β)on the closed unit interval. - Gamma
- The Gamma distribution
Gamma(k, β)in shape/rate parameterisation. - Gaussian
- The Gaussian distribution
N(μ, σ²). - Metropolis
- A random-walk Metropolis sampler over a scalar state.
Enums§
- Parameter
Error - An invalid distribution parameter.
Traits§
- Distribution
- The randomness traits this crate is built on, re-exported for convenience.
A distribution that can produce samples of type
T. - Rng
- The randomness traits this crate is built on, re-exported for convenience. A source of pseudo-random bits.
- Sampler
- The randomness traits this crate is built on, re-exported for convenience.
Bulk sampling for any
Distribution.
Functions§
- bernoulli_
log_ likelihood - Log-likelihood of
successesandfailuresBernoulli trials at ratep. - binomial_
log_ likelihood - Log-likelihood of observing
successesout oftrialsat ratep. - exponential_
log_ likelihood - Log-likelihood of exponential
dataat raterate. - log_
sum_ exp - Numerically stable
ln Σ exp(xᵢ). - normal_
log_ likelihood - Log-likelihood of independent Gaussian
datawith meanmeanand standard deviationstd. - poisson_
log_ likelihood - Log-likelihood of Poisson
countsat raterate.