gsax

Sobol Indices

The standard variance-based method — first-order (S1), total-order (ST), and second-order (S2) shares of output variance. Pick it when you can choose where to evaluate your model.

RS-HDMR

Fits a spline surrogate to any existing (X, Y) data and reads Sobol-compatible indices off it, with a built-in emulator. Pick it when your model runs already exist and a structured design isn't an option.

PCE

Fits an orthogonal-polynomial surrogate and computes Sobol indices exactly from its coefficients. Very sample-efficient for smooth models.

Shapley Effects

Splits output variance fairly among inputs — each interaction shared equally by its participants — so the effects sum to exactly 1. Computed analytically from a fitted PCE (default) or RS-HDMR surrogate, with no extra model runs.

eFAST

Estimates S1 and ST from the Fourier spectrum of the output along sinusoidal search curves. A classic alternative to Saltelli sampling that needs only a plain N × D design.

DGSM

Bounds the total Sobol index from model derivatives, obtained through JAX autodiff. A cheap screening option when your model is differentiable.

Morris Screening

Coarse one-at-a-time screening that ranks inputs (mu_star) and flags nonlinearity or interactions (sigma) from very few model runs. Use it to discard unimportant inputs before a full Sobol study.

HSIC

Kernel-based measure that detects any statistical dependence — nonlinear, non-monotone, or heteroscedastic — from any (X, Y) pairs, with permutation p-values. Works even with correlated inputs.

PAWN

Measures how much the whole output distribution (its CDF) shifts when an input is fixed, capturing effects that variance-based indices miss. Useful when tails and extremes matter, not just spread.

Borgonovo Delta

Measures the shift of the entire output density when an input is fixed — moment-independent like PAWN, but density-based. The given-data estimator works on any (X, Y) pairs and includes bootstrap confidence intervals.

Save & Reload Samples

Persist a sample set with SamplingResult.save(), evaluate your model elsewhere, and reload with gsax.load() — the analysis metadata travels with the samples.

Multi-Output & Time-Series

Pass scalar, (N, K), or (N, T, K) outputs to any of the ten methods and get indices for every output and timestep in a single vectorized pass. Set output_names and gsax disambiguates 2-D layouts, fixing obvious transposes with a warning.

Up to 668× Faster than SALib

Fused JIT kernels and vectorized execution replace Python loops. Sobol up to 15.8× faster, HDMR up to 668× on multi-output workloads.

gsaxGlobal Sensitivity Analysis in JAX