Functional ANOVA: Decomposing Functional Data

• Functional ANOVA is a framework that decomposes functional data into additive main and interaction effects using an orthogonal decomposition approach.
• It employs recursive inclusion–exclusion and permutation tests to partition variance and assess sensitivity in high-dimensional and complex data settings.
• The method is applicable in areas like high-dimensional regression, environmental monitoring, and explainable machine learning through robust, nonparametric techniques.

Functional ANOVA (fANOVA) is a principled framework for decomposing variation in functional data or multivariate mappings into additive components corresponding to low- and higher-order effects, with applications spanning inference, sensitivity analysis, explainable machine learning, and high-dimensional regression. The fundamental goal is to extend classical ANOVA concepts to data and models where responses are functions—curves, surfaces, or general Hilbert-space elements—rather than scalars or vectors.

1. Mathematical Foundation of Functional ANOVA

Let $f:X \to \mathbb{R}$ be a square-integrable function defined on a product space $X = X_1 \times X_2 \times \cdots \times X_p$, equipped with a product measure $\mu = \mu_1 \otimes \cdots \otimes \mu_p$. Then $f$ admits an orthogonal decomposition,

$$f(x) = f_\emptyset + \sum_{i=1}^p f_{\{i\}}(x_i) + \sum_{1\leq i<j\leq p} f_{\{i,j\}}(x_i,x_j) + \cdots + f_{[p]}(x_1,\dots,x_p),$$

where each $f_S$ is the unique effect of variables in $S\subseteq[p]$, hierarchically orthogonal: $$\int f_S(x_S)\, d\mu_j(x_j) = 0 \quad \forall j\in S \text{ for all } x_{S\setminus\{j\}}.$$ The decomposition can be constructed recursively using the inclusion–exclusion principle: $$f_S(x_S) = \sum_{T\subseteq S} (-1)^{|S|-|T|} \int f(x) \, d\mu_{-T}(x_{-T}),$$ admitting a Möbius inversion form. The term $f_\emptyset$ is the grand mean. For functional data (e.g., group-indexed curves $X = X_1 \times X_2 \times \cdots \times X_p$0), this logic applies to the mean curves over time $X = X_1 \times X_2 \times \cdots \times X_p$1 as well.

The sum-of-squares decomposition leads to a variance partition: $X = X_1 \times X_2 \times \cdots \times X_p$2 This enables Sobol’ sensitivity indices and effective dimension analysis (Fumagalli et al., 2024, Herren et al., 2022, Borgonovo et al., 2018, Ginsbourger et al., 2014).

2. Testing Group Mean Differences: Permutation Global Envelope Approach

In functional ANOVA for two or more groups of functional data, the principal inferential task is to test

$X = X_1 \times X_2 \times \cdots \times X_p$3

against the general alternative. Given discretized observations $X = X_1 \times X_2 \times \cdots \times X_p$4, the test statistics are built from the vectors:

• $X = X_1 \times X_2 \times \cdots \times X_p$5: Concatenation of group mean functions,
• $X = X_1 \times X_2 \times \cdots \times X_p$6: Pairwise mean differences,
• $X = X_1 \times X_2 \times \cdots \times X_p$7: Pointwise $X = X_1 \times X_2 \times \cdots \times X_p$8-statistics.

A permutation procedure is used under the null of exchangeability: functional observations are relabeled across groups, statistics recomputed, and a global envelope constructed by the "extreme-rank-length" procedure. Let $X = X_1 \times X_2 \times \cdots \times X_p$9 be the number of permutations; for each, compute extreme ranks and a lexicographically ordered depth statistic. The $\mu = \mu_1 \otimes \cdots \otimes \mu_p$0 global envelope $\mu = \mu_1 \otimes \cdots \otimes \mu_p$1 for each coordinate $\mu = \mu_1 \otimes \cdots \otimes \mu_p$2 is determined from the set of permutations whose depths exceed the critical threshold.

If any observed mean function or contrast departs from its envelope at some $\mu = \mu_1 \otimes \cdots \otimes \mu_p$3, $\mu = \mu_1 \otimes \cdots \otimes \mu_p$4 is rejected, with simultaneous family-wise error rate control at $\mu = \mu_1 \otimes \cdots \otimes \mu_p$5 (Mrkvicka et al., 2016). Two variants are supported:

• Original-space envelope (visualizes group mean functions vs. their envelopes);
• Pairwise difference envelope (locates $\mu = \mu_1 \otimes \cdots \otimes \mu_p$6 where specific group pairs differ).

This approach is computationally straightforward and localizes differences in the domain, with robust finite-sample calibration via permutations.

3. Multivariate and Repeated Measures Functional ANOVA

Functional ANOVA extends to:

• Multivariate and vector-valued functional observations: Each curve is discretized; group- and condition-wise coefficients are stacked and subjected to MANOVA on basis coefficients. Null hypotheses about group or treatment effects translate to contrasts on these coefficients. Sphericity corrections or permutation tests ensure validity under dependence or non-Gaussianity (Acal et al., 2024, Acal et al., 2024).
• Repeated-measures designs: A basis-expansion reduces the two-way repeated-measures functional ANOVA to classical MANOVA. Models allow for main group effects, within-subject (repeated) factors, and their interactions. Scores on principal components or basis vectors can serve as high-dimensional responses in the MANOVA or nonparametric test.

4. Sensitivity Analysis, Model Explainability, and Machine Learning

The functional ANOVA decomposition underpins global sensitivity analysis (GSA) and explainability in complex models:

• Sobol’ indices, truncation and superposition dimensions, and the breakdown of variance by effect order are built directly on the orthogonality of the $\mu = \mu_1 \otimes \cdots \otimes \mu_p$7 (Fumagalli et al., 2024, Herren et al., 2022, Ferrere et al., 3 Mar 2026).
• Explainable ML methods such as SHAP (Shapley Additive Explanations) derive attribution values for each feature $\mu = \mu_1 \otimes \cdots \otimes \mu_p$8 by distributing the fANOVA terms among variables:

$\mu = \mu_1 \otimes \cdots \otimes \mu_p$9

(Herren et al., 2022, Fumagalli et al., 2024). Interventional and observational SHAP correspond to different choices of the underlying distribution (e.g., independent marginals, empirical, or full joint).

Advanced machine learning architectures now directly incorporate the fANOVA structure. Additive and low-order (pairwise) interaction models, such as GAMI-Net, EBM, and GAMI-Lin-T, fit separate neural or tree submodels for each $f$0 or $f$1, with identifiability constraints (zero-mean with respect to marginals) enforced for orthogonality (Hu et al., 2023, Hu et al., 2022, Park et al., 21 Feb 2025, Park et al., 1 Oct 2025). In categorical input spaces, explicit discrete orthogonal decompositions are computed via extensions of Fourier analysis (Ferrere et al., 3 Mar 2026).

5. Robust and Nonparametric Functional ANOVA

Robust fANOVA methods address the vulnerability of classical $f$2 statistics to outliers or non-Gaussian errors:

• RoFANOVA uses robust $f$3-estimators for functional means and dispersion, combined with permutation tests for global inference, replacing sums of squares by sums of robust loss functions (e.g., bisquare, Hampel) (Centofanti et al., 2021).
• Heavy-tailed error processes, such as t-processes, are modeled within a functional regression, yielding bounded-influence, information-consistent estimators robust to anomalous curves (Zhang et al., 2018).
• Nonparametric approaches employ wavelet-based estimation and thresholding for irregular or locally heterogeneous functional deviations; adaptive or Bayesian wavelet shrinkage can be performed under dependent errors with explicit optimality guarantees (Kist et al., 2015, Ma et al., 2016).

6. Extensions and Computational Considerations

Functional ANOVA generalizes to multidimensional domains (e.g., spatial fields), multiple input distributions, and correlated errors:

• The effect of the spatial domain and boundary conditions is managed via Hilbert space projections and eigenfunction expansions. When errors follow autoregressive Hilbertian processes (ARH(1)), estimation relies on generalized least squares in basis coordinates with domain-adapted eigenfunctions (Álvarez-Liébana et al., 2017).
• When the input distribution is uncertain or a mixture, fANOVA decompositions are only unique up to "cores" of measures. Under prior mixing, decomposition remains valid but orthogonality and variance decomposition require modification (Borgonovo et al., 2018).
• In high-dimensional settings, efficient model-based algorithms with explicit regularization, sparsity, and basis expansion deliver computational tractability and interpreted component-wise models (Park et al., 21 Feb 2025, Park et al., 1 Oct 2025).

7. Empirical Illustration and Applications

Applications span a variety of domains:

• Fiscal policy: fANOVA envelope tests identify temporal regions (e.g., policy changes in the EU in 2006) of significant difference in government expenditure decentralization not captured by integral-based $f$4 tests (Mrkvicka et al., 2016).
• Environmental science: Multivariate-fANOVA uncovers period and site-location effects for air pollutant curves, crucial for environmental monitoring (Acal et al., 2024).
• Human behavior: Functional group pattern analysis and kernel-based group classification extract temporally localized differences in facial action units for emotion recognition (Ji et al., 2022).
• Engineering: Orthogonal decompositions extract actuator contributions and interaction significance in control of shape in engineering structures, with nearly all variance often captured by main effects (Tan et al., 15 Jun 2025).

---

In summary, functional ANOVA organizes the decomposition of functional or multivariate outcomes into interpretable main and interaction effects, with mathematically rigorous and computationally scalable frameworks for hypothesis testing, variance decomposition, sensitivity analysis, robust inference, and explainable artificial intelligence. Contemporary advances increasingly unify classical statistical principles, modern machine learning, and efficient computational methods (Mrkvicka et al., 2016, Fumagalli et al., 2024, Hu et al., 2023, Herren et al., 2022, Park et al., 21 Feb 2025, Ferrere et al., 3 Mar 2026).