Global Sensitivity Analysis via Variance Decomposition

54 concepts48 papers1 proposal

When dealing with complex, non-linear systems—whether they are biophysical simulations of a beating heart, climate models, or deep neural networks—we often face a fundamental question: which inputs actually drive the behavior of the output?

The simplest way to answer this is to calculate a local derivative: change an input slightly, observe the change in the output, and call that the sensitivity. However, this local approach fails in high-dimensional spaces where inputs vary across wide ranges and interact with one another. If turning knob $A$ only matters when knob $B$ is set to a high value, a local derivative will completely miss this interaction.

To capture these dynamics, researchers turn to Variance-Based Global Sensitivity Analysis (GSA). Instead of looking at a single point, GSA explores the entire parameter space simultaneously. It treats the model as a black box and asks a elegant question: If we let all our inputs vary across their entire ranges, what fraction of the total variance in the output can be attributed to each individual input or their combinations?


The Intuition of Variance Decomposition

Imagine a model output $Y$ driven by several uncertain inputs $X_1, X_2, \dots, X_n$. Because the inputs are uncertain, the output $Y$ has a total variance, $V(Y)$, representing the overall unpredictability of the system.

Variance-based global sensitivity analysis decomposes this total variance into a sum of contributions from individual parameters and their interactions:

$$V(Y) = \sum_{i} V_i + \sum_{i < j} V_{ij} + \dots + V_{12\dots n}$$

From this decomposition, we derive Sobol' indices, which serve as the standard metrics for ranking parameter importance:

  • First-Order Index ($S_i = V_i / V(Y)$): This measures the main effect of an input $X_i$ acting alone. It answers: How much would the output variance drop, on average, if we could fix the exact value of $X_i$?
  • Total-Effect Index ($S_{Ti}$): This measures the total contribution of $X_i$, including its first-order effect and all its higher-order interactions with other variables.

By comparing these indices, we gain deep structural insights. If a parameter has a low first-order index but a high total index, it means the parameter is influential, but only through its interactions with other variables.

This framework is widely used to identify critical drivers in complex physical systems. For example, Parikh et al. applied global sensitivity analysis to ventricular myocyte models to determine which specific ion channel conductances drive proarrhythmic risk metrics. In engineering, Prorok et al. used Sobol indices to study autonomous vehicle traffic, discovering that the target velocity of individual vehicle controllers is the single most critical parameter affecting overall traffic throughput when different manufacturer styles interact. Similarly, Li et al. used variance-based analysis to demonstrate that the geometry of buried bedrock ridges is a far more influential driver of earth fissures than the actual volume of groundwater pumped.


Bypassing the Computational Bottleneck

The main challenge of variance-based GSA is its computational cost. Estimating Sobol indices traditionally requires evaluating the model thousands or millions of times using sampling strategies like Saltelli sampling (as seen in work by Munoz et al.). If a single simulation takes minutes or hours, computing these indices directly becomes intractable.

To bypass this bottleneck, researchers build surrogate models (also called emulators). A surrogate model is a fast mathematical approximation trained on a small, carefully selected set of expensive simulation runs. Once trained, the surrogate can evaluate millions of points in seconds, making the calculation of Sobol indices trivial.

[Expensive Physics Model] ---> Run ~100 times ---> [Train Fast Surrogate]
                                                           |
                                                   Evaluate 1,000,000x
                                                           |
                                                           v
                                                [Exact Sobol Indices]

Different fields leverage distinct surrogate architectures to achieve this:

  • Gaussian Processes / Kriging: Lazarus et al. deployed Gaussian Process emulators to perform sensitivity analysis on left ventricular passive mechanics, revealing which tissue properties are identifiable under normal versus high mechanical loads.
  • Polynomial Chaos Expansion (PCE): In power systems engineering, Wang et al. introduced a Data-Driven Sparse PCE method to rapidly quantify how continuous weather fluctuations and discrete line outages propagate variance through renewable energy grids.

From Physics to Deep Learning Explainability

A compelling modern application of variance-based sensitivity analysis is in Explainable AI (XAI). Deep neural networks are notoriously complex black boxes. Traditional explainability methods rely on local gradients (such as saliency maps) to show which pixels influenced a classification. However, just like local sensitivity analysis in physics, gradients fail to capture non-linear feature interactions and are highly sensitive to adversarial noise.

To solve this, researchers have adapted Sobol indices to explain model decisions based on high-level semantic concepts rather than raw pixels or tokens.

[Input Image] ---> [Extract Latent Concepts] ---> [Perturb Concept Space] ---> [Measure Output Variance] ---> [Total Sobol Index]

In the CRAFT framework, Fel et al. and Fel used recursive matrix factorization to isolate latent concepts inside a network, and then applied total Sobol indices to measure how much the variance of the model's output is explained by perturbations in those concept representations. This provides a highly robust measure of feature importance that accounts for global model behavior.

Similarly, in natural language processing, Jourdan et al. developed the COCKATIEL framework, which leverages total Sobol indices to rank the influence of continuous textual concepts on NLP classifiers, capturing both the direct impact of a concept and its complex interactions with other features without requiring model retraining.


Beyond Variance: Distribution-Based Analysis

While variance decomposition is incredibly powerful, variance itself only captures the first two moments of a distribution (mean and spread). If an input parameter drastically alters the skewness or the heavy tails of an output distribution without changing its overall variance, Sobol indices might label that parameter as unimportant.

To address this limitation, Gamboa et al. proposed a distribution-based sensitivity analysis framework using the Cramér-von Mises distance. Instead of measuring changes in output variance, this method measures the statistical distance between the entire unconditional cumulative distribution function (CDF) of the output and the conditional CDF obtained when a specific input is fixed. This ensures that changes in the entire shape of the output distribution—including risk-critical tails—are fully captured.

Key Foundational Work

SBML-SAT: a systems biology markup language (SBML) based sensitivity analysis tool (2008) — contributed 1 concept elements to this direction.
Global sensitivity analysis of stochastic computer models with joint metamodels (2012) — 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_variance_based_sensitivity_analysis_f4a01ae8,
  title = {variance-based sensitivity analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/variance-based-sensitivity-analysis/rdr_f4a01ae8d3bc483fa061f12308e75839},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_variance_based_sensitivity_analysis_f4a01ae8_v1,
  title = {variance-based sensitivity analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/variance-based-sensitivity-analysis/rdr_f4a01ae8d3bc483fa061f12308e75839/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake