When we want to understand a complex system—whether it is a climate simulation, a financial model, or a deep neural network—we naturally ask: which inputs actually drive the output?
To answer this, fields like Global Sensitivity Analysis (GSA) and machine learning interpretability have long relied on a comforting assumption: that the input variables are statistically independent. Under independence, the total variance of a system can be cleanly sliced into orthogonal pieces, attributing a distinct percentage of the output's behavior to each individual parameter.
But in the real world, variables are rarely independent. Temperature and pressure move together; genes in a biological pathway are tightly co-regulated; features in an image exhibit complex spatial correlations. When you change one variable, you implicitly drag others along with it. This fundamental coupling breaks the mathematical machinery of standard statistical inference, leading to misleading conclusions and fragile models.
A growing body of research focuses on relaxing this independence assumption, rebuilding our tools to accurately quantify variable influence within highly correlated systems.
Why Correlation Breaks Variance Decomposition
To understand why input dependence causes traditional methods to fail, consider the classical framework of Sobol sensitivity indices. As explained by Fabrice Gamboa et al., these indices rely on an ANOVA-like decomposition to split a model's output variance into contributions from individual inputs and their interactions. This decomposition fundamentally depends on the orthogonality of the input space.
When inputs are correlated, this orthogonality vanishes. Imagine a model where the output $Y$ is heavily influenced by two highly correlated inputs, $X_1$ and $X_2$. If we vary $X_1$ while ignoring the correlation, we are testing combinations of $(X_1, X_2)$ that are physically or statistically impossible in the real system. Conversely, if we try to isolate the effect of $X_1$ by averaging over the distribution of $X_2$, the shared variance gets confounded.
As a result, standard sensitivity analysis techniques produce highly misleading metrics. As demonstrated by Sébastien Da Veiga et al., ignoring these correlations can cause researchers to completely misidentify the primary drivers of variance, leading to flawed risk assessments and sub-optimal engineering designs. The influence becomes ambiguous: is $X_1$ driving the output because of its inherent impact, or simply because it acts as a proxy for $X_2$?
The Theoretical Domino Effect
The breakdown caused by input dependence extends far beyond variance decomposition; it shakes the foundations of probability theory and global signal detection.
Coordinate Influence and Concentration of Measure
In discrete harmonic analysis, the "influence" of a coordinate measures how much a function's output changes when that specific coordinate is toggled. Traditional proofs for bounding these influences—such as the KKL or Talagrand theorems—rely heavily on product measures, where every coordinate is independent.
When moving to non-product measures where variables exhibit complex dependencies, these standard analytical tools fail. As explored by Nathan Keller et al. and Frederic Koehler et al., quantifying coordinate influence in dependent systems requires entirely new geometric frameworks to track how a single perturbation propagates through a network of dependencies.
Global Signal Detection
In large-scale inference, frameworks like Higher Criticism are used to detect rare, weak signals buried in massive noise (the "needle in a haystack" problem). As originally formulated by Jiashun Jin, the theoretical guarantees of these tests assume independent test statistics. When the underlying data points are correlated, false-positive rates inflate rapidly. Mapping out the "impossibility regions" where signal detection is guaranteed to fail requires explicitly accounting for these structured dependencies, a challenge detailed by Alon Kipnis.
Reclaiming Sensitivity via Conditional Expectations
To fix sensitivity analysis for dependent variables, researchers must shift their perspective from marginal distributions to conditional distributions. Instead of evaluating how an input varies across its entire independent range, we must evaluate how it varies given the specific values its correlated partners must take.
Efficient Conditional Estimation
Computing these conditional expectations directly is notoriously difficult, traditionally requiring nested, multi-dimensional integration loops that are computationally intractable for complex models. To bypass this bottleneck, Sébastien Da Veiga et al. introduced a "black-box" approach using local polynomial estimation. By using local polynomials to approximate the conditional moments of the output based on existing data, the authors successfully sidestepped expensive multidimensional integration, making dependent sensitivity analysis computationally feasible.
Generalized Decompositions and Distributional Shifts
Rather than trying to force dependent inputs into an independent mold, Gaëlle Chastaing et al. developed a robust numerical framework for generalized Sobol indices. This framework explicitly incorporates the joint distribution's covariance structure, allowing variance to be decomposed accurately without losing the true physical meaning of the parameters.
However, variance itself can sometimes be an inadequate metric when dependencies distort the shape of the output distribution. To address this, Fabrice Gamboa et al. proposed a framework based on the Cramér-von Mises distance. Instead of measuring how an input shifts the variance of the output, this method quantifies how an input shifts the entire cumulative distribution function of the output, providing a more holistic and unambiguous measure of influence under dependence.
Independent Inputs (Orthogonal Decomposition)
[X1] ───> [ Clean, Isolated Variance Contribution ]
[X2] ───> [ Clean, Isolated Variance Contribution ]
Dependent Inputs (Confounded Variance)
[X1] ──┐
│ ├─> [ Shared / Confounded Variance ] ──> Misleading Sensitivity Indices
[X2] ──┘
Conditional / Distributional Frameworks
[X1 | X2] ──> [ Conditional Moments / Cramér-von Mises Distance ] ──> True Influence
Tracking Propagation with Wasserstein Matrices
For purely mathematical challenges like the concentration of measure—proving that a function of many variables stays close to its average value—how do we handle dependence without the clean factorization of product measures?
Aryeh Kontorovich solved this by combining the classical martingale method with Wasserstein matrices. The core intuition is elegant: if you alter the value of a single coordinate $X_1$, it doesn't just change the immediate output of the function; because of the underlying dependencies, it shifts the conditional probability distributions of $X_2, X_3, \dots, X_n$.
By constructing a Wasserstein matrix, the author quantified exactly how a perturbation in one coordinate ripples through the conditional distributions of all other coordinates. This matrix acts as a mathematical ledger, tracking the propagation of influence across the system and allowing researchers to prove sharp concentration inequalities without ever invoking the independence assumption.
Key Foundational Work
Higher criticism for detecting sparse heterogeneous mixtures (2004) — contributed 2 concept elements to this direction.
Learning Coordinate Covariances via Gradients (2006) — contributed 1 concept elements to this direction.