salib

Global sensitivity analysis for Rust, implemented from the primary literature. Bit-deterministic by construction.

crates.io · API reference · source

Start here

• Quickstart — Ishigami function, Saltelli sampling, Sobol’ indices. Five minutes.
• Choosing a method — which estimator for which question.

Methods

Variance-based (Sobol’)

Decompose output variance into contributions from each input. The workhorse of global SA.

• Saltelli 2010 — improved estimator for first-order and total-effect indices. The default choice.
• Jansen 1999 — total-effect estimator with better convergence for small samples.
• Janon 2014 — asymptotically efficient first-order estimator.
• Owen 2013 — three-matrix design for improved second-order estimates.
• Given-data Sobol’ — Sobol’ indices from observational data without designed experiments. Plischke et al. 2013.

Elementary effects

OAT trajectories through the input space. Screening: which factors matter, cheaply.

• Morris 1991 — mean and standard deviation of elementary effects. The original screening method.
• Grouped Morris — Morris with factor groups. Campolongo et al. 2007.

Frequency-based

Probe model response at characteristic frequencies per factor.

• FAST / eFAST — Fourier Amplitude Sensitivity Test. First-order via spectral decomposition; extended variant adds total-effect. Cukier 1973, Saltelli 1999.
• RBD-FAST — Random Balance Designs. Reuses a single random sample for all factors. Tarantola et al. 2006.

Distribution-based

Sensitivity beyond variance: shift in the entire output distribution.

• Borgonovo δ — moment-independent importance measure. Captures any distributional shift. Borgonovo 2007.
• PAWN — CDF-based sensitivity via Kolmogorov–Smirnov statistic. Pianosi et al. 2015.
• QOSA — quantile-oriented sensitivity analysis. Fort et al. 2016.

Derivative-based

• DGSM — Derivative-based Global Sensitivity Measures. Upper bounds on total-effect indices from gradients. Sobol’ & Kucherenko 2009.

Regression

• SRC / SRRC / PCC / PRCC — standardized regression and partial correlation coefficients. Linear and rank-transformed. Saltelli & Marivoet 1990.

Surrogate

Build a cheap approximation, extract indices analytically.

• Polynomial Chaos Expansion — full OLS and sparse LARS/OMP. Analytic Sobol’ indices from coefficients. Xiu & Karniadakis 2002, Blatman & Sudret 2011.
• HDMR — High-Dimensional Model Representation. Cut-HDMR with PCE component functions. Li et al. 2002.
• Active Subspaces — gradient-based dimension reduction. Eigendecomposition of the uncentered gradient covariance. Constantine 2015.

Game-theoretic

• Shapley Effects — cost-sharing of output variance via coalitional game theory. Song, Nelson & Staum 2016.

Experimental design

• ANOVA — two-way and three-way analysis of variance. Fisher 1925.
• G-Theory — generalizability theory D-study. Variance components for measurement designs. Brennan 2001.
• Discrepancy — L2-star discrepancy of point sets. Hickernell 1998.
• Fractional Factorial — main effects and interactions from designed experiments. Box, Hunter & Hunter 1978.

Reference

• Bibliography — annotated references for every method.
• Internals — bit-determinism, tree-structured reductions, the rayon contract.
• Crate map — which crate owns what, dependency graph, feature flags.