Sample-Efficient Global Sensitivity Analysis

31 concepts15 papers1 proposal

When dealing with complex computational models—whether they are climate simulations, financial risk engines, or deep neural networks—we often need to know which inputs drive the output variance. This is the domain of Global Sensitivity Analysis (GSA), traditionally quantified using Sobol indices.

The mathematical definition of a first-order Sobol index is elegant: it measures the fraction of output variance explained by a single input variable $X_i$:

$$S_i = \frac{\text{Var}(\mathbb{E}[Y \mid X_i])}{\text{Var}(Y)}$$

However, computing this expectation exactly is impossible for black-box models. Historically, practitioners relied on brute-force Monte Carlo or "pick-freeze" sampling designs. To evaluate $d$ inputs, these methods require generating multiple independent sampling matrices and swapping columns, scaling the required model evaluations to thousands or tens of thousands of runs. If a single model evaluation takes minutes or hours, standard GSA becomes computationally intractable.

To overcome this bottleneck, a modern research direction focuses on sample-efficient sensitivity estimation. By combining advanced statistical estimators, derivative-based bounds, and rigorous surrogate frameworks, researchers have found ways to map input importance with a fraction of the computational overhead.


Bounding Variance with Derivatives and Poincaré Inequalities

One of the most intuitive ways to save sample evaluations is to substitute global sampling with local gradient information. If a model's derivative with respect to an input is consistently small across the domain, that input cannot heavily influence the final output variance.

To formalize this intuition, Lamboni et al. established a theoretical bridge between derivative-based global sensitivity measures (DGSM) and traditional Sobol indices. By leveraging the Poincaré inequality, the authors demonstrated that local derivatives can provide a mathematically rigorous upper bound on total variance-based indices:

$$\text{Total Sobol Index} \le C \times \int \left(\frac{\partial f}{\partial X_i}\right)^2 d\mu(X)$$

where $C$ is a constant depending on the input probability distribution. This inequality means that by computing cheap gradient information, practitioners can definitively screen out non-influential variables using up to twenty times fewer samples than traditional variance calculations.

Conversely, we can also compute lower bounds to guarantee a minimum level of influence. Olivier Roustant et al. utilized Poincaré differential operators to construct orthogonal Hilbert bases from their eigenfunctions. By applying Parseval inequalities to these bases, the authors derived derivative-based lower bounds that closely approximate true Sobol indices even when using a highly restricted number of basis functions. Together, these upper and lower bounds allow for rapid variable screening without ever computing the full, expensive indices.


Squeezing More from Less: Advanced Statistical Estimators

If we must estimate the exact indices via sampling, we can radically optimize how we handle the data. Traditional pick-freeze schemes are wasteful, often discarding intermediate combinations or relying on simple sample covariances.

Information-Augmented and U-Statistics

Fabrice Gamboa et al. introduced an information-augmented estimation framework that reuses every available data point within a pick-freeze sample matrix. By exploiting all cross-combinations of the samples, their specialized estimators achieve faster statistical concentration and lower variance, reducing the number of black-box evaluations needed to reach convergence.

Taking data reuse a step further, Fabrice Gamboa et al. reframed global sensitivity analysis within the architecture of U-statistics. By structuring the sensitivity index as a generalized U-statistic, the authors proved that the required number of black-box code calls can be cut down to exactly twice the sample size ($2N$), eliminating the triple or quadruple evaluation overhead ($3N$ or $4N$) required by classical Monte Carlo approaches while maintaining strict asymptotic normality.

Single-Sample Rank Statistics

The ultimate frontier in sampling efficiency is eliminating the need for specialized experimental designs altogether. Traditional methods require structured pairs of inputs (e.g., swapping the $i$-th column between two matrices). What if we only have a single, unstructured observational sample of input-output pairs?

Pierre A. Gremaud et al. solved this by introducing rank-based statistical estimators. By sorting the observed outputs based on the order of a given input variable, the authors adapted empirical correlation coefficients to estimate first-order Sobol indices, Cramér-von Mises indices, and higher-order moments.

Traditional Pick-Freeze:
Matrix A:  [X1, X2, X3] -> Y_A
Matrix B:  [Z1, Z2, Z3] -> Y_B
Swap i=1:  [X1, Z2, Z3] -> Y_Swap  (Requires targeted resampling)

Rank-Based GSA:
Single Sample: [X1, X2, X3] -> Y   (Sort by X_i, compute rank correlation)

Because this framework operates on a single, static dataset, it bypasses the need to rerun the simulation model for specific input combinations, offering a massive computational cost reduction for high-dimensional problems.


Rigorous Surrogates and Metamodeling

When the underlying model is too expensive for even optimized estimators, the standard engineering solution is to build a fast surrogate model (or "metamodel"), such as a Gaussian Process or a Polynomial Chaos Expansion (PCE). However, surrogates introduce approximation errors, which can invalidate downstream sensitivity rankings.

To bridge the gap between computational efficiency and statistical rigor, Massimo Aufiero proposed the Floodgate framework. Floodgate allows practitioners to leverage fast, imperfect surrogate models to perform GSA while still outputting mathematically guaranteed, valid confidence intervals for the true model's sensitivity indices. This ensures that even if the surrogate is biased, the resulting variable importance ranking remains reliable.

Other approaches optimize how the surrogate is constructed specifically for GSA. Evgeny Burnaev et al. combined PCE with a $D$-optimality experimental design. By analyzing the asymptotic statistical distribution of the estimated indices, their algorithm adaptively selects new sample points that actively minimize the variance and covariance of the sensitivity estimates, reducing the required simulation count by orders of magnitude.

Similarly, the authors of Support Vector Regression-Based Active Subspace (SVR-AS) Modeling of High-Speed Links for Fast and Accurate Sensitivity Analysis demonstrated that combining active subspaces with support vector regression allows high-fidelity design spaces to be explored in hours rather than weeks.


Beyond Scalar Variance: Distributional Sensitivity

Many real-world models do not output a single scalar; they output curves, images, or full probability distributions. Extending variance-based GSA to these spaces typically compounds the sample-size explosion.

To address this, Fabrice Gamboa et al. extended sensitivity indices to multidimensional and functional outputs, proving their consistency and asymptotic normality. To capture how an input alters the entire shape of an output distribution (rather than just its variance), Fabrice Gamboa et al. introduced distance-based sensitivity indices using the Cramér-von Mises distance. By utilizing a clever pick-and-freeze estimation scheme that reuses fixed samples, they bypassed the need for expensive non-parametric density estimation, making distribution-based GSA computationally comparable to standard scalar analysis.

In the context of structural reliability, where we care specifically about the probability of a rare failure event, Pascal Lemaitre et al. introduced Density Modification-Based Reliability Sensitivity Indices (DMBRSI). Instead of running thousands of new simulations to see how a change in an input's probability distribution affects failure rates, their approach cleverly re-weights existing simulation data using importance sampling weights derived from the Central Limit Theorem:

Original Inputs (Dist P)  ---> [ Black-Box Model ] ---> Failure Probability
                                       |
                                (Re-weight Samples)
                                       v
Modified Inputs (Dist Q)  ---> [ Analytical Weighting ] -> New Failure Probability

This density-modification strategy allows practitioners to perform comprehensive reliability sensitivity analysis without executing a single additional model simulation.

Key Foundational Work

Derivative-based global sensitivity measures: General links with Sobol' indices and numerical tests (2013) — contributed 3 concept elements to this direction.
Sensitivity analysis for multidimensional and functional outputs (2014) — contributed 1 concept elements to this direction.
Density modification based reliability sensitivity analysis (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_sample_efficient_sensitivity_estimation_8d31af63,
  title = {sample-efficient sensitivity estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sample-efficient-sensitivity-estimation/rdr_8d31af6314474cf58de1151debc10de2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_sample_efficient_sensitivity_estimation_8d31af63_v1,
  title = {sample-efficient sensitivity estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/sample-efficient-sensitivity-estimation/rdr_8d31af6314474cf58de1151debc10de2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake