Crate salib 

Re-exports

pub use salib_samplers as samplers;

pub use salib_estimators as estimators;

Modules

distribution

Distribution — closed enum of factor distributions, with inverse-CDF (quantile) and support boundaries as the unified extension point.

problem

Problem — the declarative, content-addressable description of an SA campaign’s input space. Vector of factors, each with a name, distribution, and kind (continuous / discrete / categorical / boolean).

reduce

Deterministic f64 reductions for saltelli — fixed-tree pairwise folds whose output is bit-identical regardless of rayon partitioning.

rng

RngState — multi-stream ChaCha20 with deterministic salt-derived forking. The replay-determinism foundation for every salib sampler and estimator.

Structs

Factor

A single factor in the experiment. Name + distribution + kind. #[non_exhaustive] blocks struct-literal construction outside this crate; consumers go through ProblemBuilder.

Group

A named group of factors treated as a single unit in SA.

Problem

The declarative input-space description for an SA campaign.

ProblemBuilder

Builder for Problem. The only public path to a Problem value.

RngState

The serializable RNG state. The single source of truth for any SA campaign’s randomness.

Enums

BuildError

Errors arising from ProblemBuilder::build.

Distribution

Factor distributions saltelli supports. Closed enum, #[non_exhaustive]. Future variants (Truncated, Empirical, Categorical, …) land non-breaking via follow-on ADRs.

FactorKind

The role a factor plays in the experiment. Closed enum, #[non_exhaustive]. Continuous is the default for typical SA applications.

RngAlgorithm

The set of RNG algorithms salib supports. Closed enum + #[non_exhaustive]; future variants land non-breaking.

Constants

BLOCK

Per-chunk size for the parallel block-then-tree reduction. Fixed; changing it would change reduction-tree shape and therefore bit-level results. Future workload-tuned per-Experiment blocks would land via ADR.

Functions

par_tree_dot

Parallel block-then-tree dot product. Same partitioning logic as par_tree_sum; bit-identical to tree_dot.

par_tree_sum

Parallel block-then-tree sum. Partitions xs into BLOCK-sized chunks, reduces each chunk via tree_sum in parallel via rayon, then reduces the per-chunk sums via tree_sum sequentially. Bit-identical to tree_sum(xs) regardless of rayon thread count or worker partition pattern.

par_tree_var

Parallel two-pass unbiased sample variance. Mean via par_tree_sum; centered-square via parallel map (preserves index order on IndexedParallelIterator); then the centered-square reduction via par_tree_sum.

tree_dot

Sequential pairwise tree-fold dot product. Produces Σᵢ aᵢ·bᵢ in tree order. Bit-identical to par_tree_dot.

tree_sum

Sequential pairwise tree-fold sum. O(N) work, O(log N) reduction depth. Bit-identical to par_tree_sum regardless of rayon thread count.

tree_var

Sequential two-pass unbiased sample variance, routed through tree_sum in both passes. Returns 0.0 for inputs of length < 2.