Global sensitivity analysis asks which uncertain inputs of a simulator actually drive its output. The standard answer is the Sobol' index, the share of output variance explained by each input. Estimating it by plain Monte Carlo can take millions of simulator runs. The usual workaround is to fit a cheap surrogate model and read the indices off the surrogate. That moves the problem rather than solving it: which few expensive simulator runs should we spend so that the indices, rather than the surrogate's overall fit, come out precise? This note proposes an answer for one surrogate family where the question has a closed-form solution. The family is Gaussian processes (probabilistic regression models that return a mean prediction and an uncertainty) built with "centered ANOVA" kernels, whose predictions split exactly into per-input and per-interaction pieces.
The core idea is to pick each new simulation by D-optimal design applied to the vector of variance components the indices are made of. D-optimal design is the classical rule of choosing the experiment that most shrinks the volume of the uncertainty ellipsoid around the quantities of interest. With a centered ANOVA kernel, each variance component is a quadratic form in the Gaussian process posterior. The means and covariances of Gaussian quadratic forms are known exactly. So the determinant of the covariance of the variance components, and how a candidate run would change it, can be computed without any sampling. The resulting design rule depends on the data seen so far, handles observation noise natively, and has no fixed polynomial truncation. It is a one-step extension of existing work, but the step has not been taken.
Core Claim
Efficient design of experiments for sensitivity analysis based on polynomial chaos expansions (Burnaev, Panin and Sudret) showed that designing for Sobol' index precision beats designing for surrogate accuracy. They derived an asymptotic covariance of the indices under a polynomial chaos expansion (a regression on orthogonal polynomials) and chose points by Fedorov exchange to minimize its determinant. Their own stated limitations frame our starting point. They fix the polynomial degree and truncation in advance, need a non-degenerate initial design, and lose most of their advantage under observation noise.
ANOVA kernels and RKHS of zero mean functions for model-based sensitivity analysis (Durrande, Ginsbourger and Roustant) supplies the surrogate. Their centered kernels make the posterior mean split into orthogonal main effects and interactions, with Sobol' indices extracted in closed form from the Gram matrix. They also report that the index estimates "remained remarkably robust" on the Sobol' g-function even when noise variance was four times the signal variance. However, they use fixed space-filling designs and never ask where to sample.
Our claim is that these two pieces fit together exactly. Write $f_u$ for the ANOVA component of the surrogate on input subset $u$. Its variance $V_u=\int f_u^2\,d\mu$ is a quadratic form $V_u = z^\top A_u z$ in the jointly Gaussian posterior $z$ of the latent function at the design points. Here $A_u$ is a matrix built from integrals of the kernel against the input distribution $\mu$, available in closed form for Gaussian and Matérn kernels under uniform or Gaussian inputs. For posterior mean $m$ and covariance $\Sigma$, the standard Gaussian moment identity gives
$$\operatorname{Cov}(V_u, V_v) = 2\,\operatorname{tr}(A_u \Sigma A_v \Sigma) + 4\, m^\top A_u \Sigma A_v\, m .$$
Stack the first-order components $V_1,\dots,V_d$, the total-order components (also closed form for product kernels), and the total variance $V$ into a vector $\mathbf{V}$. Call its posterior covariance $C(\xi)$ under design $\xi$. We choose the next run $x$ to maximize
$$\Delta(x) = \log\det C(\xi) - \mathbb{E}_{y_x}\!\left[\log\det C(\xi\cup\{x\})\right],$$
where $y_x$ is the not-yet-observed output at $x$.
The term with $m$ in the covariance formula is what makes the criterion data-dependent. In linear-Gaussian design the optimal design ignores the observed outputs. Here, once the data show that an input carries signal, the rule spends runs refining exactly the components whose ratios decide the ranking. The log-determinant also has a concrete advantage over trace-based (A-optimal) rules: rescaling any single variance component shifts every candidate's score by the same constant. A rule targeting $\mathbf{V}$ is therefore insensitive to how small or large individual components are. It still accounts for the correlations among components that determine whether two inputs can be told apart.
Research Program
The first step builds the criterion. That means closed-form $A_u$ matrices for centered Gaussian and Matérn-5/2 kernels under uniform and Gaussian input measures, plus total-order components through the product structure $\prod_i(1+k_0^i)$. Adding a point is a rank-one update to the posterior, so we update $\Sigma$ with the Sherman–Morrison–Woodbury identity, the same device Gradient Descent-based D-optimal Design for the Least-Squares Polynomial Approximation uses for $(A^\top A)^{-1}$. Scoring one candidate then costs roughly $O(d^2N^2)$ rather than $O(N^3)$ per component pair, where $N$ is the number of runs so far. That cost is negligible next to one simulator call in the regimes we target.
Because $y_x$ is unknown, the expectation in $\Delta(x)$ is taken by Gauss–Hermite quadrature over its posterior predictive, using three to five nodes. We will check this choice against plugging in the posterior mean.
The second step handles the nonlinear catch-22 that Optimal experimental design: Formulations and computations identifies: the criterion depends on kernel hyperparameters that are themselves estimated from the data. We average the criterion over a small set of hyperparameter samples, drawn either from a Laplace approximation or from short MCMC. The aim is to avoid the degenerate locally-optimal designs that Burnaev et al. warn about. We also test whether this removes their need for a carefully chosen initial design.
The third step is analytical and asks how points are placed. Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes observed that learning slopes calls for sampling adjacent pairs of points, whereas classical D-optimality pushes points apart. Variance components are integrals of squared effects, not pointwise slopes. We therefore predict that our rule will spread points along influential coordinates and stay coarse along inert ones, giving neither pure space-filling nor pairing. We will measure this with per-coordinate nearest-neighbour statistics of the selected designs.
Baselines
The naive reference point is a Sobol' sequence (a quasi-Monte Carlo design that covers the input space evenly) of the same size, fed to the same centered ANOVA Gaussian process with the same hyperparameter treatment. This is exactly the setup of Durrande et al. It is the most important comparison, because it isolates whether choosing the points matters at all once the surrogate is right. We count it as beaten only if our rule reaches a given index error with clearly fewer runs, averaged over at least twenty random restarts per benchmark.
The strongest existing method for direct Sobol' estimation is Burnaev et al.'s Adaptive-for-SI polynomial chaos design. We will reimplement it from the paper, using the authors' Fedorov-exchange variant, the same candidate pools as our method, and polynomial degree selected by cross-validation on a held-out initial batch so it is not handicapped by an arbitrary truncation. We found no reported numeric result for Adaptive-for-SI on Ishigami, Borehole or Wing Weight on Lacuna, only the qualitative statement that it attains lower index error than Latin hypercube, random and generic D-optimal designs. We will therefore regenerate its curves rather than quote them. Beating it means lower mean squared index error at equal budgets of 30 to 200 runs on those benchmarks, and in particular holding that advantage as noise is added, where Burnaev et al. report their method degrades.
The strongest derivative-based competitor is Belakaria et al.'s information-gain and maximum-variance acquisition on Gaussian process gradients. We run it with the same kernel family (uncentered) and the same budgets, and score its output by the Sobol' indices of the fitted model, as well as by its native DGSM targets. No reported numeric result was found on Lacuna for it on Ishigami, the g-function or Morris. We also include plain predictive-variance active learning. This is the "fill the holes" strategy that Belakaria et al. report flatlines on 15- and 20-dimensional problems.
Two ablations isolate the key ingredients. The first replaces the log-determinant with the trace of $C$ (A-optimality on the same components). The second replaces $C$ with the posterior covariance of the function values themselves, which is classical function-targeted D-optimality on the same surrogate. If neither ablation loses to the full method, the index-targeting idea is not doing the work.
Why not just…?
Why not just use a Sobol' sequence with a good surrogate? This is the real threat. Durrande et al.'s robustness result suggests that, at low dimension and low noise, the surrogate already does most of the work. Burnaev et al.'s evidence that targeted designs beat Latin hypercube and random designs comes from a polynomial surrogate, not from this one. Our naive baseline settles the question directly. If the gap is negligible on Ishigami and Borehole, we will report that and restrict the claim to the regimes where it holds, namely higher dimension (Morris, Wing Weight) and heavy noise.
Why not just extend Burnaev et al.'s polynomial chaos design with adaptive degree? That would address the fixed-truncation limitation, but not the noise vulnerability they themselves report. It would also still require choosing a polynomial basis whose size grows combinatorially with dimension. The Gaussian process formulation has no truncation, and treats noise as part of the posterior rather than as a nuisance. The cross-validated-degree version of their method is exactly our second baseline, so this alternative is tested, not argued away.
Why not just learn gradients actively and bound the total indices? The derivative-based bounds of Derivative-based global sensitivity measures: General links with Sobol' indices and numerical tests are cheap; that paper reports actionable screening with about 20 times fewer runs than standard Sobol' sampling on Morris. However, the bounds are one-sided, can be conservative (sometimes exceeding 1 for some input distributions), and give no first-order indices. Belakaria et al. also note that their information-gain rule loses to maximum variance on 20-dimensional Morris. If DGSM active learning recovers the correct ranking of influential inputs as quickly as our rule, the cheaper method wins for screening. Our ranking metric reports that comparison separately from index accuracy.
Proposed Experiments
The benchmarks are those in the file with analytical or cheaply computed reference indices: Ishigami (3 inputs), the Sobol' g-function (5 and 20 inputs), Piston (7), Borehole (8), Wing Weight (10) and Morris (20). Reference indices come from analytical formulas or from a $10^6$-run pick-freeze estimate, following Statistical inference for Sobol pick-freeze Monte Carlo method. Each benchmark is run noise-free and with additive Gaussian noise at variance ratios of 0.1, 1 and 4 relative to the signal, the last matching Durrande et al.'s hardest setting.
We measure four things: mean squared error of the first- and total-order indices against budget; Kendall rank correlation and top-$k$ recovery of influential inputs; coverage of 90% posterior intervals on the indices, which tests whether the quadratic-form moments are honest; and the wall-clock cost of designing each point.
Three outcomes would settle the proposal. The first is whether the full method beats the Sobol'-sequence baseline by a meaningful margin in run count at 10 to 20 inputs or noise ratios of 1 or more. The second is whether it beats both ablations, which shows that targeting the variance components with a determinant, rather than the function or a trace, is what helps. The third is whether its advantage over the polynomial chaos design widens as noise grows, as the analysis predicts. A null result on the first check, with positive results only in the noisy regime, would still be informative: it would show that design for sensitivity analysis matters mainly when the simulator is stochastic.
References
- Efficient design of experiments for sensitivity analysis based on polynomial chaos expansions
- ANOVA kernels and RKHS of zero mean functions for model-based sensitivity analysis
- Gradient Descent-based D-optimal Design for the Least-Squares Polynomial Approximation
- Optimal experimental design: Formulations and computations
- Active Learning for Derivative-Based Global Sensitivity Analysis with Gaussian Processes
- Derivative-based global sensitivity measures: General links with Sobol' indices and numerical tests
- Statistical inference for Sobol pick-freeze Monte Carlo method