Functional Maximum Correlation Analysis

• FMCA is a methodology that quantifies maximal statistical dependence between high-dimensional random objects, such as probability densities and planar shapes.
• It leverages tangent-space linearization and functional PCA to reduce infinite-dimensional data for practical multivariate analysis.
• FMCA extends to multimodal and network settings, achieving high cross-modal correlations and superior performance in applications like affective computing and biomedical imaging.

Functional Maximum Correlation Analysis (FMCA) is a methodology for quantifying and extracting maximal statistical dependence between high-dimensional or infinite-dimensional random objects, typically functional data such as probability densities or planar shapes, and for generalizing measures of nonlinear association to the multimodal setting. FMCA subsumes multivariate, nonlinear, and functional data dependencies by leveraging Hilbert space geometry, spectral theory, and information-theoretic objectives, forming the core of recent frameworks for multimodal, self-supervised, and manifold-aware correlation analysis.

1. Geometric and Functional Foundations

FMCA builds on the geometry of Hilbert spheres and functional manifolds, particularly for commonly encountered functional data types:

• Probability Densities: The space of 1D densities $$P = \{ p:[0,1]\rightarrow \mathbb{R}_+ : \int_0^1 p(t)\,\mathrm{d}t = 1 \}$$ is endowed with the Fisher–Rao metric, making $P$ a non-Euclidean manifold. The square-root transformation $\psi = \sqrt{p}$ maps $P$ onto the positive orthant of the unit Hilbert sphere $S^\infty$ in $L^2[0,1]$. The induced metric is the $L^2$-metric on the sphere with tangent space $$T_\psi S^\infty = \{\delta\psi: \langle \delta\psi, \psi \rangle = 0\}$$.
• Closed Planar Shapes: For planar closed curves, the square-root velocity function (SRVF) $q(t) = \dot{\beta}(t)/\sqrt{|\dot{\beta}(t)|}$—subject to unit-length—embeds the set of shapes as a quotient manifold under the $L^2$-sphere, accounting for rotations and reparameterizations. The quotient space $S = C/(SO(2)\times\Gamma)$ inherits this structure, supporting geodesic and variational analysis (Cho et al., 2021).

These constructions allow representations of densities and shapes as points on the unit Hilbert sphere, supporting both geometric and statistical operations. The geometry enables tractable analysis via local linearization and dimension reduction.

2. Local Tangent-Space Linearization and Functional PCA

Analysis of infinite-dimensional objects is simplified via tangent-space linearization. Given a reference mean $\mu$ (typically the Karcher mean of the sample), the logarithm map projects any point $x$ on the sphere to the tangent space at $\mu$: $$\log_\mu(x) = \frac{\theta}{\sin\theta} (x - \cos\theta\,\mu), \quad \theta = \arccos\langle x, \mu\rangle$$ This yields Euclidean structure locally, supporting standard linear multivariate techniques.

Functional Principal Component Analysis (FPCA) is then performed in the tangent space: for $n$ samples $x_i$, compute tangent vectors $v_i = \log_\mu(x_i)$, estimate the empirical covariance operator, and extract $r$ leading eigenfunctions. Each sample is projected to $r$-dimensional Euclidean coordinates via inner products with these eigenfunctions, effecting dimension reduction (Cho et al., 2021).

3. FMCA Objective: Nonlinear and Functional Canonical Correlation

For two or more modalities represented as Euclidean coordinates in their respective tangent spaces, FMCA seeks linear projections $(a, b)$ maximizing the canonical correlation: $\max_{a,b} \,\mathrm{Corr}(a^\top U,\, b^\top V)$ where $U, V \in \mathbb{R}^{n \times r}$ are the projected coordinates for each modality. The solution is given by the leading eigenpairs of the associated block matrix system. Generalizations include the use of nonlinear projection functions and kernel-based approaches for higher-order, multimodal settings (Zheng et al., 28 Dec 2025).

The conceptual extension to Hilbert–Schmidt operators enables nonparametric computation of dependence structure using the expansion of the density-ratio operator

$$\rho(x,y) = \frac{p(x,y)}{p(x)p(y)} = 1+\sum_k \sigma_k\,\varphi_k(x)\psi_k(y)$$

where $\{\sigma_k\}$ are canonical correlation coefficients, and the FMCA objective can be formulated in terms of maximizing the sum or log-determinant of $\sigma_k$ (Zheng et al., 28 Dec 2025).

4. Multimodal and Network Extensions

The FMCA framework supports extension to higher-order and networked associations as in Multimodal Functional Maximum Correlation (MFMC) and Network Maximal Correlation (NMC):

• Dual Total Correlation (DTC): FMCA is adapted for multimodal dependency by maximizing the dual total correlation, defined by

$$\mathrm{DTC}(X_1,...,X_M) = H(X_{[M]}) - \sum_{i=1}^M H(X_i|X_{[M]\setminus\{i\}})$$

A tight sandwich bound relates DTC to cyclic joint mutual informations, allowing practical trace-based surrogate optimization in neural settings (Zheng et al., 28 Dec 2025).

• Network Maximal Correlation: For random variables $X_1,\dots,X_n$ on a graph $G=(V,E)$, NMC seeks transforms $\phi_i$ maximizing $$\sum_{(i,j)\in E}\mathbb{E}[\phi_i(X_i)\phi_j(X_j)]$$ under mean-zero, unit-variance constraints, solvable via Hilbert space expansion and, in Gaussian cases, reductions to Max-Cut optimization (Feizi et al., 2016).

These extensions permit FMCA to model not just pairwise, but also higher-order and network-structured dependencies among arbitrary modalities and data types.

5. Algorithmic Implementation

The standard FMCA workflow comprises the following steps:

1. Transformation: Functional data are mapped to Hilbert spheres using SRT (for densities) or SRVF (for shapes).
2. Karcher Mean and Tangent Mapping: Compute the mean, then project samples to the tangent space via the log map.
3. Dimension Reduction: Perform FPCA in the tangent space to extract principal components and obtain finite-dimensional representations.
4. Correlation Analysis: For two modalities, deploy CCA in $\mathbb{R}^r$; for more, maximize joint mutual information using cyclic trace objectives as surrogates.
5. Visualization and Back-Mapping: Canonical directions are reconstructed in the original function space by the exponential map; densities are squared, and shapes are recovered by integration (Cho et al., 2021, Zheng et al., 28 Dec 2025).
6. Neural Architectures for MFMC: For multimodal emotion recognition, deep encoders extract modality embeddings, which are fused and projected. Cyclic trace losses over joint covariance structures are minimized to capture higher-order dependencies (Zheng et al., 28 Dec 2025).

6. Empirical Results and Applications

• Densities and Shapes: Simulated and biomedical data (e.g., MRI for glioblastoma) demonstrate that FMCA precisely recovers true canonical correlations and directions, with cross-modal correlations up to $0.85$ and canonical variate regression yielding lower MSE and improved C-index versus PC regression (Cho et al., 2021).
• Affective Computing: MFMC with FMCA outperforms or matches supervised and baseline SSL methods in subject-independent protocols on emotion recognition benchmarks. Notably, it achieves state-of-the-art accuracy (e.g., 86.8% on CEAP-360VR, 44.2% on MAHNOB-HCI EEG), directly benefiting from higher-order dependence modeling (Zheng et al., 28 Dec 2025).
• Nonlinear Network Inference: NMC applied to gene expression uncovers nonlinear dependencies and modules undetectable by linear methods; in Gaussian settings, NMC acts as a nonparametric analogue of Max-Cut, supporting graphical model discovery (Feizi et al., 2016).

7. Conceptual Insights and Generalizations

FMCA provides a unifying, information-theoretic foundation for nonlinear, functional, and multimodal correlation analysis. By operating intrinsically on functional manifolds, leveraging local tangent linearization and Hilbert space theory, FMCA effectively decouples nonlinear structure for tractable multivariate analysis.

A key distinction with FMCA-relative approaches is their ability to capture higher-order synergistic dependence (DTC), surpassing the limitations of pairwise contrastive losses and avoiding redundancy double-counting intrinsic to total correlation. These properties enable FMCA and its multimodal extensions to robustly identify joint structure amidst subject-level variability and complex, coordinated multimodal signals.

The methodology generalizes to any functional manifold embeddable in a Hilbert sphere with analytically tractable exponential and logarithm maps, extending its applicability to a broad class of nonlinear, infinite-dimensional data contexts in modern statistical learning (Cho et al., 2021, Zheng et al., 28 Dec 2025, Feizi et al., 2016).