Functional Tucker Decomposition (FTD)

• Functional Tucker Decomposition (FTD) is a generalization of classic Tucker models that uses continuous latent functions to model multidimensional data with smoothness and uncertainty quantification.
• It integrates methodologies such as Gaussian processes, RKHS, and polynomial interpolation to achieve adaptive modeling and scalable computation for continuous evaluation.
• FTD has been successfully applied in domains like air pollution time series, geospatial climate analysis, and scientific simulations, yielding significant error reductions and high compression ratios.

Functional Tucker Decomposition (FTD) extends the classical Tucker tensor model to address the representation and analysis of multidimensional data indexed, in one or more modes, by continuous variables. By embedding mode-wise continuity constraints and equipping Tucker’s factor matrices with latent functions or elements in function spaces, FTD enables smooth, uncertainty-aware, and structure-preserving modeling of quasitensors arising from time series, spectral data, geospatial phenomena, and scientific simulations. FTD encompasses a spectrum of practical and theoretical frameworks: reproducing kernel Hilbert spaces (RKHS), Gaussian/process-based priors, polynomial interpolation, and sparse learning, all organized to exploit the multi-linear structure while supporting continuous evaluations, adaptive modeling, and scalable computation (Fang et al., 2023, Steidle et al., 26 Mar 2026, Dolgov et al., 2020, Rai et al., 2019).

1. Mathematical Framework and Variants

At its core, Functional Tucker Decomposition replaces discrete factor matrices with collections of mode-wise latent functions. Let $f(\mathbf{x})$ or $y(x_1,...,x_M)$ denote the continuous-indexed observation at index tuple $\mathbf{x} = (x_1,...,x_M)$. The general FTD form is

$$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$

where $\mathcal{G}$ is the core tensor and $f^{(m)}_{i_m}(\cdot)$ are latent functions for mode $m$ (Fang et al., 2023, Dolgov et al., 2020, Steidle et al., 26 Mar 2026). This basic structure admits several incarnations:

• Bayesian FTD: Mode functions are equipped with independent GP priors, yielding a functional Bayesian decomposition with uncertainty quantification (Fang et al., 2023).
• RKHS FTD: Mode functions are elements in RKHSs specified by positive-definite kernels; function evaluations and interpolations are intrinsic to the representation (Steidle et al., 26 Mar 2026).
• Chebyshev/Polynomial FTD: Factor functions are Chebyshev (or alternative basis) polynomial expansions, supporting exponential interpolation accuracy and spectral methods (Dolgov et al., 2020).
• Sparse Basis FTD: Factor functions are learned as sparse expansions in large orthonormal dictionaries (e.g., Legendre polynomials, wavelets), with $L_1$ regularization (Rai et al., 2019).

For quasitensors combining discrete and continuous modes, the FTD yields a fully functional generalization:

$$T(x_1, ..., x_{N-1}, x_N) \approx \mathcal{G} \times_1 A^{(1)} \cdots \times_{N-1} A^{(N-1)} \times_N U_N(x_N),$$

where $U_N(\cdot)$ maps the continuous mode $y(x_1,...,x_M)$0 into an $y(x_1,...,x_M)$1-dimensional latent subspace (Steidle et al., 26 Mar 2026).

2. Model Construction and Algorithmic Approaches

A. Gaussian Process and SDE Representation

Bayesian FTD frameworks deploy Gaussian process (GP) priors on mode functions. To address the $y(x_1,...,x_M)$2 cost of GP inference, for stationary Matérn kernels with half-integer smoothness, the GP prior is equivalently formulated as a linear time-invariant stochastic differential equation (SDE):

$y(x_1,...,x_M)$3

yielding a linear-chain state-space model supporting inference with $y(x_1,...,x_M)$4 complexity, where $y(x_1,...,x_M)$5 is white noise and $y(x_1,...,x_M)$6 contains $y(x_1,...,x_M)$7 and its derivatives (Fang et al., 2023).

B. RKHS and Adaptive Functional Subspaces

FTD can leverage RKHS modeling for functional modes. By the Representer Theorem, each mode function admits a data-adaptive expansion in terms of kernel functions centered at design points:

$y(x_1,...,x_M)$8

$y(x_1,...,x_M)$9

where $\mathbf{x} = (x_1,...,x_M)$0 is the kernel matrix, $\mathbf{x} = (x_1,...,x_M)$1 the expansion coefficients, and $\mathbf{x} = (x_1,...,x_M)$2 the grid of design points (Steidle et al., 26 Mar 2026). The optimization incorporates RKHS regularization, enabling smoothness control and support for new points $\mathbf{x} = (x_1,...,x_M)$3 outside training domains.

C. Practical Construction: Chebyshev/Polynomial and Sparse Functional Bases

FTD via spectral or polynomial interpolation (as in Chebfun3F) relies on tensorized Chebyshev interpolation and low-rank Tucker splitting of the coefficient tensor. The process includes:

1. Sampling univariate fibers on coarse Chebyshev grids (using Adaptive Cross Approximation).
2. Refining these fibers to high-accuracy 1D interpolants.
3. Extracting the core tensor via discrete empirical interpolation (DEIM) on selectively sampled cross points.

This yields a representation of the form $\mathbf{x} = (x_1,...,x_M)$4, where each $\mathbf{x} = (x_1,...,x_M)$5 etc. is a Chebyshev polynomial (Dolgov et al., 2020). Sparse functional Tucker methods further promote basis sparsity via $\mathbf{x} = (x_1,...,x_M)$6-penalized regression for each factor (Rai et al., 2019).

3. Inference, Optimization, and Computational Aspects

Message Passing and Scalable Posterior Approximation

In Bayesian FTD, the posterior over core $\mathbf{x} = (x_1,...,x_M)$7, noise precision $\mathbf{x} = (x_1,...,x_M)$8, and state variables is approximated by site-factored Expectation Propagation (EP), updated using Conditional EP and collapsed with Kalman–Rauch–Tung–Striebel (RTS) smoothing over the state-space chains. Each update scales as $\mathbf{x} = (x_1,...,x_M)$9 for tensor rank $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$0 and $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$1 per observation, leading to an overall $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$2 run-time (Fang et al., 2023).

Alternating Minimization for RKHS/Composite FTD

Optimization in RKHS-based FTD proceeds via alternating minimization on the discrete factors, continuous factor coefficients $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$3, and the core tensor. The $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$4-step solves a Kronecker-structured ridge regression of size $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$5, while core and discrete steps follow the standard HOOI or ALS pipeline (Steidle et al., 26 Mar 2026).

Randomization and Compression

Sparse randomized FTD exploits sketching and randomized least-squares to re-estimate the core, reducing computational complexity from $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$6 to $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$7, where $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$8 is the data size, $$f(x_1,\dots,x_M) \approx \sum_{i_1=1}^{R_1} \cdots \sum_{i_M=1}^{R_M} \mathcal{G}_{i_1\cdots i_M} \prod_{m=1}^M f^{(m)}_{i_m}(x_m),$$9 the parameter count, and $\mathcal{G}$0 (Rai et al., 2019). High-dimensional data (e.g., scientific simulations) benefit from dramatic storage reductions, with compression ratios up to $\mathcal{G}$1.

4. Error Analysis and Theoretical Properties

• Approximation Guarantees: For Chebyshev FTD, the total error splits into interpolation and low-rank errors, with exponential decay of interpolation error for analytic functions, and quasi-optimal Frobenius error for judiciously selected fiber sets (Dolgov et al., 2020).
• RKHS Theory: If the target tensor is smooth in the functional mode, RKHS projection error decays as $\mathcal{G}$2, where $\mathcal{G}$3 is the fill distance of the design points (Steidle et al., 26 Mar 2026).
• Continuity and Identifiability: FTD with RKHS or GP priors guarantees smooth factor functionals, enabling out-of-domain evaluation and domain transfer. Core and factor orthogonality promote identifiability up to rotation and scaling, as in standard Tucker decomposition.
• Compression Bounds: Sparse FTD retains only as many nonzero coefficients per mode as necessary, with error controlled by Lasso regularization (Rai et al., 2019).

5. Applications and Empirical Comparisons

Functional Tucker Decomposition has been validated in several domains:

Application	Task & Setting	FTD Outcome
Air Pollution Time Series (Beijing PM$\mathcal{G}$4)	3-way (pressure, temp, time)	FTD RMSE $\mathcal{G}$5 0.29 vs $\mathcal{G}$6 0.80 for top discrete Tucker; full trajectory/uncertainty recovery (Fang et al., 2023)
US Temperature Data (Lat, Long, Year)	Spatio-temporal analysis	Lat/Long functions show physical gradients; time-mode detects post-1950 warming, historical climate dips (Fang et al., 2023)
Domain-variant Tensor Classification	Synthetic digits, time series, hyperspectral	Accurate domain-transfer classification: FTD maintains performance under distribution shift, HOSVD degrades by 30pp or drops to chance (Steidle et al., 26 Mar 2026)
Scientific Data Compression (Combustion)	3D/4D simulation, unstructured	Compression ratios $\mathcal{G}$7–$\mathcal{G}$8, storage reduction by orders of magnitude, relative error $\mathcal{G}$9–$f^{(m)}_{i_m}(\cdot)$0 (Rai et al., 2019)
Function Approximation (Chebfun3F)	Black-box trivariate $f^{(m)}_{i_m}(\cdot)$1	75–98% reduction in function evaluations vs slice-based Chebfun3, near-optimal accuracy (Dolgov et al., 2020)

FTD consistently outperforms discrete Tucker when the data-generating process is fundamentally continuous or when sampling schemes vary, thanks to its structural interpolation, smoothness, and adaptability (Fang et al., 2023, Steidle et al., 26 Mar 2026, Dolgov et al., 2020).

6. Limitations and Extensions

Current FTD implementations typically handle one or a few continuous modes; computations for multi-functional-mode cases become dominated by large Kronecker or tensor contractions (Steidle et al., 26 Mar 2026). Storage and computational costs in the RKHS setting scale with the number of design points $f^{(m)}_{i_m}(\cdot)$2; thus, selection of $f^{(m)}_{i_m}(\cdot)$3 and kernel is operationally significant. Arbitrary missing or unaligned data requires model extensions or more general factorization frameworks.

Chebfun3F and similar procedures can be generalized to higher dimensions, but suffer the curse of dimensionality due to exponential growth of the core. Tensor-train or hierarchical Tucker variants, as well as fiber-cross and DEIM strategies, offer scalability avenues (Dolgov et al., 2020). Transfer of these advances to generalized tensor-based models is a topic of ongoing research.

7. Summary and Outlook

Functional Tucker Decomposition provides a unified approach for representing, compressing, and analyzing multiway data indexed on continuous domains. By merging the multilinear core structure of Tucker models with functional, RKHS, or Bayesian priors on mode-wise factors, FTD addresses both the “continuity gap” of classical decomposition and the need for statistical adaptivity and uncertainty quantification. Applications in time series, geospatial and spectral data analysis, scientific simulation, and adaptive subspace modeling have demonstrated substantial practical and theoretical benefits, particularly under domain shift, irregular sampling, and large-scale regimes. Methodological developments continue toward incorporating multi-mode functionality, active design selection, and robustification against missing or asynchronous data (Fang et al., 2023, Steidle et al., 26 Mar 2026, Dolgov et al., 2020, Rai et al., 2019).