Derivative-Based Global Sensitivity Indices
- Derivative-based global sensitivity indices are measures that integrate squared local derivatives to quantify each input variable's overall impact on a model's output.
- They leverage efficient estimation techniques such as adjoint, surrogate, and spectral methods, offering significant computational advantages in high-dimensional scenarios.
- DGSM underpin advanced practices like reduced-order modeling and active learning, facilitating robust uncertainty quantification and effective variable screening.
Derivative-based global sensitivity indices (DGSM) quantify the global influence of each input variable on model outputs by integrating (typically squared) local partial derivatives over the input space. Unlike classical variance-based Sobol’ indices, which decompose variance through ANOVA expansions, DGSM leverage gradient information to provide computationally efficient screening and upper bounds—especially vital in high-dimensional or computationally intensive models. DGSM extend to arbitrary distributions via Poincaré-type inequalities and admit efficient estimation via adjoint, surrogate, and spectral methods. Their generalization to functional outputs, active learning, and their rigorous connection to total-effect Sobol’ indices and entropy-based GSA underpin their centrality in contemporary uncertainty quantification.
1. Mathematical Definition and Theoretical Foundations
Let be a vector of independent input variables with probability measure , and let be a square-integrable response function with square-integrable partial derivatives. The derivative-based global sensitivity measure for variable is defined as
Optionally, normalization by model variance yields
which is analogous in interpretation to a first-order or total-effect Sobol’ index (Cleaves et al., 2019).
The generalization to functional outputs , where is a space or time variable and a vector of parameters, involves integrating squared sensitivities over the output domain: 0 and normalizing by the trace of output covariance 1 to enable direct ranking (Cleaves et al., 2019).
2. Links with Sobol’ Indices and Bounding Results
Sobol’ total-effect indices 2 satisfy rigorous upper bounds in terms of DGSM and Poincaré-type constants: 3 with the optimal constant 4 defined for general Boltzmann measures (densities of the form 5) as
6
and specializing to 7 for uniform and 8 for Gaussian marginals (Lamboni et al., 2012). These inequalities are sharp for linear responses. Lower bounds via Cauchy–Schwarz and Hardy–Littlewood–Polya techniques further bracket 9 in terms of DGSM. For arbitrary measures, Cheeger or log-Sobolev constants enter the bounds (Kucherenko et al., 2016, Kucherenko et al., 2014).
In settings with functional outputs, the bound extends pointwise and is aggregated over the functional domain: 0 where 1 is the Poincaré constant for 2, specific to its marginal law (Cleaves et al., 2019).
3. Numerical Estimation, Surrogate, and Spectral Methods
DGSM estimation exploits both finite-difference and automatic/adjoint differentiation for gradient evaluation, yielding practical computational costs:
- Monte Carlo: 3 model evaluations per 4 samples for all 5 variables.
- With adjoint-mode differentiation: all gradients at a single point can be computed for 6 model runs—essentially dimension-independent cost (Kucherenko et al., 2014, Kucherenko et al., 2016).
Polynomial chaos expansions (PCE) offer analytical computation of DGSM via post-processing: 7 where 8 are PCE coefficients, 9 the scaling, and 0 the univariate polynomial derivative norm (Sudret et al., 2014). Poincaré chaos expansions ("PoinCE") generalize this by constructing bases such that derivatives are orthogonal with respect to the input measure. Parseval-type equalities enable closed-form expressions for both Sobol’ and DGSM indices within this framework, allowing gradient-enhanced PoinCE regression to produce highly accurate and robust sensitivity estimates, even in data-scarce settings (Roustant et al., 3 Oct 2025, Lüthen et al., 2021).
4. Active Learning and Gaussian Process Surrogates for DGSM
Gaussian process (GP) surrogates allow estimation of DGSM and their uncertainty through the GP posterior gradient's mean and covariance: 1 Specialized active learning strategies utilize acquisition functions targeting uncertainty reduction in DGSM, optimizing sample efficiency. Criteria include gradient variance, variance reduction, and information gain, extendable to both local and global gradient structures, and are computationally accelerated via chunking and K-means approximations (Lambert et al., 16 Jan 2026, Belakaria et al., 2024). Empirical studies demonstrate substantial sample efficiency gains for DGSM estimation, especially under stringent evaluation budgets.
5. Extensions to Functional Outputs and High-dimensional Inputs
For models with high-dimensional parameter spaces and vector- or function-valued outputs, DGSM compute through integration or weighting over functional domains. Karhunen–Loève (KL) expansions reduce both input and output dimensionality, enabling efficient computation of sensitivities with adjoint-based gradients (Cleaves et al., 2019). The adjoint approach allows all partial derivatives for a given sample with low additional computational burden, scalable to large numbers of input parameters.
Functional DGSM bounds guide reduced-order modeling: selecting only those input modes or basis functions with significant DGSM suffices to achieve accurate surrogates whose predicted uncertainty and variable ranking closely match the full model (Cleaves et al., 2019).
6. Advanced Frameworks: Shapley and Entropy-based Indices
The derivative-based Shapley value (DerSHAP) merges the axiomatic fairness of Shapley values with the efficiency of DGSM. Given
2
DerSHAP robustly evaluates interaction effects and is computationally tractable even in high dimensions, directly accommodating dependent inputs via the joint covariance structure (Duan et al., 2023).
Entropic DGSM-based proxies exploit inequalities from transformation entropy: 3 Exponentiation enforces positivity and defines tractable screening proxies that remain accurate in non-Gaussian and multi-modal settings, complementing variance-based proxies in such contexts (Yang, 2023).
7. Practical Considerations, Limitations, and Applications
DGSM are especially valuable for high-dimensional screening, efficient surrogate construction, and guiding reduced-order modeling. Their efficacy is maximized when gradient information is readily available (e.g., algorithmic or adjoint differentiation), and they are particularly well-suited for models where Sobol' MC estimation is prohibitive due to dimensionality (Kucherenko et al., 2016, Kucherenko et al., 2014).
Limitations include:
- Sensitivity estimates may be loose in cases of strong nonlinearity, synergy, or when distributions exhibit heavy tails or significant skewness.
- Poincaré and Cheeger constants must be chosen appropriately for tightness of the Sobol’ bound; otherwise, bounds may be overly conservative.
- DGSM-based screening relies on the assumption that small averaged derivatives imply a variable's unimportance, which can be violated if strong interactions exist without corresponding large derivatives (Kucherenko et al., 2014, Lamboni et al., 2012).
Applications range from engineering uncertainty quantification (e.g., environmental and flood models, biotransport, circuit and mechanical systems) to machine learning explainability tools where gradient information enables efficient, scalable sensitivity ranking (Lüthen et al., 2021, Constantine et al., 2015, Roustant et al., 3 Oct 2025, Duan et al., 2023).
In summary, derivative-based global sensitivity indices provide a mathematically rigorous, computationally efficient, and broadly generalizable foundation for high-dimensional global sensitivity analysis. They underpin state-of-the-art advances in dimension reduction, surrogate modeling, and experiment design in uncertainty quantification, and they serve as indispensable tools for both preliminary screening and quantitative ranking of influential variables in complex systems.