When building complex computational models—whether for climate forecasting, structural engineering, or financial risk management—we inevitably face the question: Which input parameters actually drive the output variance, and which ones are just noise?
Answering this is the goal of Global Sensitivity Analysis (GSA). The gold standard in GSA is the calculation of Sobol' indices, which decompose the variance of a model's output and attribute fractions of that variance to individual inputs or interactions between them.
However, traditional GSA suffers from a severe computational bottleneck. To isolate the global effect of a single parameter, standard sampling frameworks (such as the "Pick-Freeze" method) require evaluating the model across vast combinatorial grids of the input space. As noted by Fabrice Gamboa et al. in Sensitivity analysis in general metric spaces, achieving statistical convergence often requires millions of model evaluations. For high-fidelity simulations where a single run takes hours, this sampling intensity makes rigorous sensitivity analysis completely impractical.
To bypass this bottleneck, a major research direction focuses on mathematical and algorithmic accelerations. Rather than running millions of brute-force simulations, researchers leverage surrogate modeling, derivative-based bounds, adjoint equations, and advanced statistical estimators to extract global insights at a fraction of the computational cost.
Turning Integration into Algebra via Surrogates
The most intuitive way to accelerate GSA is to replace the expensive "black-box" simulation with a computationally cheap mathematical twin, known as a surrogate model or emulator.
If we simply sample a standard surrogate model millions of times instead of the real model, we save time, but we still inherit sampling errors. The real breakthrough occurs when the surrogate's underlying mathematical structure allows us to calculate global sensitivity indices analytically, bypassing sampling altogether.
Orthogonal Kernels and Chaos Expansions
Consider Gaussian Process (GP) emulators. In ANOVA kernels and RKHS of zero mean functions for model-based sensitivity analysis, David Ginsbourger et al. introduce specialized "centered" ANOVA kernels. Because these kernels are explicitly constructed to produce orthogonal functional components with a mean of zero, the variance decomposition is baked directly into the model. Instead of integrating over the input space via Monte Carlo simulations, the Sobol' indices can be extracted directly through simple matrix algebra on the surrogate's internal coefficients.
A parallel strategy uses Polynomial Chaos Expansions (PCE), which approximate the model using a linear combination of orthogonal polynomials. In Efficient design of experiments for sensitivity analysis based on polynomial chaos expansions, Evgeny Burnaev et al. combine PCE with active learning ($D$-optimality experimental designs). The algorithm intelligently selects the next best point to sample from the expensive model to refine the polynomial coefficients. Once the PCE is fit, the Sobol' indices are calculated instantly from the squared values of the expansion coefficients. To scale this capability to high-dimensional fields, Johannes Exenberger et al. propose Deep Polynomial Chaos Expansion, blending the analytical tractability of classical PCE with the expressive power of deep probabilistic circuits.
Bounding the Global with the Local
Another powerful strategy changes the mathematical question entirely: Can local derivatives tell us something about global variance?
Intuitively, a derivative only tells you the slope of a function at a single, localized point. However, if you square the partial derivative of an input and calculate its expected value (average) across the entire input distribution, you get a Derivative-Based Global Sensitivity Measure (DGSM).
Through the lens of functional analysis—specifically Poincaré and Parseval inequalities—these averaged derivatives act as strict mathematical bounds on global variance indices.
[ Brute-Force GSA ] [ Derivative-Based Proxy (DGSM) ]
Requires millions of runs Averaged local gradients bound
to map entire output variance global variance via inequalities
│ │
▼ ▼
expensive integration cheap local screening
In Derivative-based global sensitivity measures: General links with Sobol' indices and numerical tests, Matieyendou Lamboni et al. formalize this link across a wide variety of probability distributions. They demonstrate that if the average squared derivative of a parameter is small, its total Sobol' index is guaranteed to be small. This allows researchers to rapidly screen out hundreds of uninfluential parameters using cheap local gradients. To complement this, Olivier Roustant et al. show in Parseval inequalities and lower bounds for variance-based sensitivity indices how differential operators can simultaneously establish tight lower bounds, effectively sandwiching the true global sensitivity index without expensive multidimensional integration.
Exploiting Adjoint Gradients for High Dimensions
If we rely on derivative-based measures, we encounter a new problem: calculating gradients for models with thousands of inputs can itself become a bottleneck. Standard finite-difference methods ("bumping" each parameter one by one) scale linearly with the number of inputs, causing a computational explosion.
To solve this, researchers turn to adjoint models and Adjoint Algorithmic Differentiation (AAD).
Instead of propagating changes from the inputs forward to the output, an adjoint model solves the governing equations backwards, propagating sensitivities from the output back to the inputs. The core benefit of this mathematical inversion is efficiency: it computes the gradient of a single target output with respect to every single input parameter simultaneously in a single computational pass.
Forward Sensitivity (Finite Difference):
Input 1 ───► [ Model ] ───► Output Sensitivity 1
Input 2 ───► [ Model ] ───► Output Sensitivity 2 (Repeated N times)
Adjoint Sensitivity:
All Inputs ◄─── [ Adjoint Model ] ◄─── Target Output (Computed in 1 pass)
As demonstrated by Simon C. Warder et al. in Adjoint-based sensitivity analysis for a numerical storm surge model, this approach makes it possible to map how localized physical inputs across vast spatial and temporal domains influence a final prediction without running the simulation thousands of times. Similarly, in the financial sector, Sébastien Geeraert highlights in Mini-symposium on automatic differentiation and its applications in the financial industry that AAD bypasses the unstable and expensive "bump and recalibrate" loops traditionally used for portfolio risk management, turning a massive computational burden into a fixed, manageable overhead.
Smarter Statistics: Ranks and Re-weighting
A final frontier in reducing the cost of GSA involves rethinking how we extract sensitivity indices from data. If we cannot compute derivatives and cannot easily fit a surrogate, we must make our statistical estimators more data-efficient.
Traditional Sobol' estimators require highly structured, paired sampling matrices. If you want to test a new input distribution or a different parameter interaction, you typically have to throw away your old simulation data and generate a brand-new experimental design.
Rank-Based Estimation
In Global sensitivity analysis: A novel generation of mighty estimators based on rank statistics, Pierre A. Gremaud et al. introduce a class of estimators that rely on rank statistics. Instead of computing exact numerical variances from rigid sampling grids, these estimators evaluate how the relative ordering of the outputs changes when inputs are sorted. Because rank-based relationships stabilize far faster than raw variance calculations, these estimators achieve high precision with significantly fewer model evaluations.
Data Recycling via Density Modification
Alternatively, we can recycle existing data using importance sampling. In Density modification-based reliability sensitivity analysis, Pascal Lemaitre et al. address the problem of understanding how changes in input probability distributions affect failure probabilities.
Instead of running new sets of simulations for every proposed distribution change, the authors show that you can mathematically re-weight the outputs of your original simulation run. By applying a density modification factor to the existing dataset, the sensitivity of the system to entirely new input distributions can be evaluated with zero additional model runs.
Key Foundational Work
The Generation of Explanations within Evidential Reasoning Systems (1987) — contributed 1 concept elements to this direction.