Functional Tensor Tucker Decomposition

• Functional Tensor Tucker Decomposition is a low‐rank model that represents continuous modes using function spaces rather than fixed factor matrices.
• It employs varied parameterizations—including sparse approximations, Chebyshev interpolation, and RKHS/GP-based approaches—to ensure smoothness and adaptability.
• The method delivers efficient storage, strong approximation guarantees, and robust performance in tasks like regression, classification, and scientific data compression.

Functional Tensor Tucker Decomposition denotes a class of Tucker-type low-rank models in which one or more mode factors are represented as functions of continuous variables rather than only as finite factor matrices. In the cited literature, this class includes deterministic functional approximations of multivariate functions, compression schemes for scientific data on unstructured meshes, RKHS-constrained decompositions for mixed discrete/continuous tensors, spline-regularized functional regression coefficients, Gaussian-process models for continuous-indexed tensors, and nonparametric infinite-feature Tucker constructions (Rai et al., 2019, Dolgov et al., 2020, Steidle et al., 26 Mar 2026, Li et al., 11 Jun 2025, Fang et al., 2023, Xu et al., 2011). Across these variants, the common structural element is a finite multilinear core together with mode-wise factors whose continuity, smoothness, interpolation behavior, or uncertainty model is made explicit.

1. Formal definition and scope

A standard functional Tucker representation replaces a full tensor-product coefficient array by a core tensor of low multilinear rank and a collection of univariate or mode-wise factor functions. In the formulation for a real-valued function $u(y_1,\dots,y_d)$, one begins with univariate spaces

$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$

so that a full expansion would require $\prod_k n_k$ coefficients. The functional Tucker approximation instead takes the form

$$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$

where $\alpha$ is a core tensor of size $r_1\times\cdots\times r_d$, each $w^{(k)}_{j_k}\in S^k_{n_k}$, and $\mathbf r=(r_1,\dots,r_d)$ are the multilinear ranks (Rai et al., 2019).

In the more general $N$-way tensor-valued setting, the same idea is written as

$$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$

with core tensor $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$0 and factor functions $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$1. When some modes are discrete, the corresponding factors reduce to ordinary factor matrices $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$2; when a mode is continuous, the factor functions may be constrained to belong to an RKHS $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$3 (Steidle et al., 26 Mar 2026).

The same multilinear structure also appears in continuous trivariate approximation. For $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$4, a functional Tucker approximation of rank $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$5 is

$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$6

where the minimal triple for exact representability is the multilinear rank of $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$7 (Dolgov et al., 2020).

A related formulation appears in regression, where the coefficient is a smoothly varying tensor-valued function $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$8. There the discretized coefficient tensor $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$9 is constrained to Tucker rank $\prod_k n_k$0, and the mode-$\prod_k n_k$1 factor $\prod_k n_k$2 induces $\prod_k n_k$3 smooth functional loadings $\prod_k n_k$4 (Li et al., 11 Jun 2025).

2. Representations of continuous factors

The principal distinction among functional Tucker models is how the continuous factors are parameterized.

In the functional sparse Tucker model for scientific compression, each singular vector extracted from a structured-grid surrogate is fitted in a prescribed dictionary, such as Legendre polynomials up to degree $\prod_k n_k$5 or multi-resolution wavelets of level $\prod_k n_k$6 and degree $\prod_k n_k$7. For each mode $\prod_k n_k$8 and component $\prod_k n_k$9, the coefficients are obtained from the sparse-regularized problem

$$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$0

with $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$1 selected by LARS + cross-validation. This produces factor expansions with $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$2 (Rai et al., 2019).

In Chebyshev-based functional Tucker approximation, the continuous factors are represented by univariate Chebyshev interpolants built from mode fibers sampled at Chebyshev nodes. The final approximation is assembled from interpolants $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$3, $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$4, and $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$5 derived from factor matrices on refined grids (Dolgov et al., 2020).

In the RKHS-based Functional Tucker Decomposition, each continuous factor $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$6 lies in $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$7, and by the Representer Theorem admits the expansion

$$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$8

At the sampled grid points, the factor evaluation matrix is $$\tilde u(y)=\sum_{j_1=1}^{r_1}\cdots\sum_{j_d=1}^{r_d}\alpha_{j_1\cdots j_d}\, w^{(1)}_{j_1}(y_1)\cdots w^{(d)}_{j_d}(y_d),$$9, where $\alpha$0 is the Gram matrix and $\alpha$1 contains the unknown coefficients. The paper states that this construction models continuous modes without requiring an a-priori basis (Steidle et al., 26 Mar 2026).

In functional tensor regression, the continuous mode is represented through a basis $\alpha$2, typically spline-based, with

$\alpha$3

Smoothness is not left to the basis alone; it is enforced by the quadratic penalty

$\alpha$4

which encodes an $\alpha$5th-order derivative regularization on the functional mode (Li et al., 11 Jun 2025).

Bayesian continuous-indexed variants replace deterministic basis expansions by latent stochastic processes. FunBaT models each factor function with an independent Gaussian-process prior,

$\alpha$6

and converts each GP to an equivalent linear state-space model by constructing a corresponding stochastic differential equation. InfTucker likewise replaces finite factor entries by evaluations of latent GP functions $\alpha$7, then lets mode ranks tend to infinity to obtain a nonparametric Tucker model in an infinite feature space (Fang et al., 2023, Xu et al., 2011).

3. Construction and optimization procedures

The functional sparse Tucker compression procedure on unstructured meshes has three explicit stages. First, a small subset of the original unstructured samples, for example $\alpha$8, is interpolated onto a tensor-product grid. Second, sequentially truncated HOSVD is applied to the resulting structured tensor $\alpha$9: each mode unfolding $r_1\times\cdots\times r_d$0 is decomposed by SVD, truncated to the leading $r_1\times\cdots\times r_d$1 singular vectors $r_1\times\cdots\times r_d$2, and combined into a core

$r_1\times\cdots\times r_d$3

with guaranteed relative Frobenius error not exceeding $r_1\times\cdots\times r_d$4. Third, the singular vectors are fitted sparsely in functional dictionaries, after which the core coefficients are re-estimated against the original $r_1\times\cdots\times r_d$5 samples by randomized least squares using a sign-flip, a fast mixing transform such as FFT, and uniform row sampling (Rai et al., 2019).

Chebfun3F constructs a trivariate functional Tucker approximation by direct fiber sampling rather than slice-based selection. Its first phase identifies coarse factor matrices by Adaptive Cross Approximation applied successively to subtensors and mode matricizations. Its second phase refines the factors on fine Chebyshev grids by adaptive univariate interpolation. Its third phase QR-factorizes the factor matrices, chooses DEIM index sets, samples the function only at the selected triple indices, and assembles the core by oblique projection before converting factor columns to Chebyshev interpolants (Dolgov et al., 2020).

The RKHS-based FTD is optimized by an ALS-style loop. Discrete factors $r_1\times\cdots\times r_d$6 are updated by mode-wise least squares; continuous coefficients $r_1\times\cdots\times r_d$7 are updated by solving the linear system

$r_1\times\cdots\times r_d$8

with

$r_1\times\cdots\times r_d$9

and the core tensor $w^{(k)}_{j_k}\in S^k_{n_k}$0 is updated by least squares with the factor matrices fixed. Iteration continues until the relative reconstruction error

$w^{(k)}_{j_k}\in S^k_{n_k}$1

changes by less than a tolerance (Steidle et al., 26 Mar 2026).

Functional tensor regression uses a functional Riemannian Gauss--Newton algorithm on the manifold of tensors of fixed Tucker rank. Starting from an initialization such as T-HOSVD of $w^{(k)}_{j_k}\in S^k_{n_k}$2, each iteration computes a tangent-space basis, solves the Gauss--Newton linear system

$w^{(k)}_{j_k}\in S^k_{n_k}$3

lifts the tangent update back to ambient space, and retracts by T-HOSVD to rank $w^{(k)}_{j_k}\in S^k_{n_k}$4. The paper states that because the penalty acts only in mode $w^{(k)}_{j_k}\in S^k_{n_k}$5, the extra per-iteration cost enters only in small $w^{(k)}_{j_k}\in S^k_{n_k}$6 blocks (Li et al., 11 Jun 2025).

FunBaT performs approximate Bayesian inference by an Expectation-Propagation style message-passing scheme. Each likelihood factor is approximated by a product of Gaussian or Gamma messages over $w^{(k)}_{j_k}\in S^k_{n_k}$7, $w^{(k)}_{j_k}\in S^k_{n_k}$8, and the state variables, while each continuous mode is updated by Kalman filtering and Rauch--Tung--Striebel smoothing in the corresponding linear Gaussian state-space model. InfTucker uses a variational Bayes procedure on the latent tensor $w^{(k)}_{j_k}\in S^k_{n_k}$9, with Gaussian or probit likelihoods, a Kronecker-structured covariance $\mathbf r=(r_1,\dots,r_d)$0, and ELBO-based updates for latent variables, kernel hyperparameters, and latent inputs (Fang et al., 2023, Xu et al., 2011).

4. Storage, complexity, and approximation guarantees

A central motivation for functional Tucker models is reduction of storage from the full tensor-product scale to a sum of core and factor costs. In the functional sparse Tucker formulation, storing the factors requires $\mathbf r=(r_1,\dots,r_d)$1 coefficients and storing the core requires $\mathbf r=(r_1,\dots,r_d)$2, so the total storage is

$\mathbf r=(r_1,\dots,r_d)$3

For $\mathbf r=(r_1,\dots,r_d)$4 original samples stored in double precision, the compressed representation requires

$\mathbf r=(r_1,\dots,r_d)$5

and the compression ratio is

$\mathbf r=(r_1,\dots,r_d)$6

The paper states that practical choices with $\mathbf r=(r_1,\dots,r_d)$7 and sparse expansions yield $\mathbf r=(r_1,\dots,r_d)$8 (Rai et al., 2019).

The randomized refinement step avoids the direct least-squares cost $\mathbf r=(r_1,\dots,r_d)$9 and storage $N$0 of the full design matrix $N$1. With sketch size $N$2, the reduced problem costs $N$3, and Johnson--Lindenstrauss type bounds imply that for $N$4 the sketched solution approximates the full least-squares solution within factor $N$5. The same paper reports rapid self-convergence once $N$6 (Rai et al., 2019).

For trivariate function approximation, Chebfun3F reduces function evaluations by selecting all mode fibers directly. The paper gives idealized function-evaluation costs $N$7 for Chebfun3 and $N$8 for Chebfun3F, and states that in practice Chebfun3F typically reduces evaluations by $N$9 and sometimes by over $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$0. The final storage in Tucker-Chebyshev form is $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$1 rather than $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$2. Error splitting is expressed as

$$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$3

and for analytic $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$4, the interpolation error decays exponentially as $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$5 (Dolgov et al., 2020).

Functional tensor regression provides explicit convergence and estimation guarantees. Under a functional-TRIP and smoothness assumptions on $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$6, the authors prove quadratic convergence of the Riemannian Gauss--Newton iterates,

$$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$7

derive an estimation error bound for $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$8, and state a matching minimax lower bound that is optimal in the tensor dimensions (Li et al., 11 Jun 2025).

Bayesian continuous-index models target scalability by exploiting structure in the latent processes. FunBaT states overall complexity $$X(t_1,\dots,t_N)\approx \sum_{i_1=1}^{r_1}\cdots\sum_{i_N=1}^{r_N} G_{i_1,\dots,i_N}\, f_1^{(i_1)}(t_1)\cdots f_N^{(i_N)}(t_N),$$9, where $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$00, because message passing and state-space smoothing are linear in the number of observations and modes. InfTucker reduces naïve $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$01 covariance operations by Kronecker eigendecompositions, giving costs of order $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$02, with further reduction under truncated SVD (Fang et al., 2023, Xu et al., 2011).

5. Empirical domains and reported results

In combustion direct numerical simulation, functional sparse Tucker compression was tested on two datasets derived from a turbulent premixed flame. For SP3D, a $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$03 grid with approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$04 points and size $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$05, TuckerMPI at precision $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$06 produced ranks $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$07 and core size $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$08; the final sparse functional Tucker representation used approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$09, achieved compression ratio approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$10, and reconstruction error approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$11. At $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$12, the reported ranks were $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$13, $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$14, storage approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$15, compression ratio $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$16, and error approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$17. For SP4D, the full $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$18 dataset of approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$19 points and size $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$20 yielded ranks $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$21, storage approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$22, compression ratio approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$23, and error approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$24 at $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$25. The same experiments also reported that higher-order singular vectors required richer bases, and that reconstruction was extremely fast because it only evaluated one-dimensional basis functions and assembled the Tucker sum (Rai et al., 2019).

The RKHS-based FTD was evaluated in domain-variant classification on three tasks, each represented as a $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$26-way tensor with one continuous mode. In a semi-synthetic digit experiment with continuous mode $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$27, training on every fourth sample and testing on a shifted domain, HOSVD accuracy dropped from approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$28 to approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$29, while FTD remained above approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$30, and Macro-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$31 showed the same gap. In multivariate time-series classification on NYC taxi trips, with hour as the continuous mode, HOSVD test accuracy fell to approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$32, while FTD maintained approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$33; Macro-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$34 was approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$35 under FTD versus approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$36 for HOSVD. In hyperspectral mango ripeness classification, with spectral bands as the continuous mode, HOSVD accuracy on the shifted domain was at or below $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$37, described as chance level, whereas FTD reached approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$38 accuracy and Macro-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$39 (Steidle et al., 26 Mar 2026).

Functional tensor regression was studied by simulation and in ADHD fMRI analysis. In the simulation with $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$40, $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$41, true rank $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$42, and $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$43, the paper reports quadratic decay of $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$44 over iterations, generalized cross-validation selecting $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$45 with lowest relative integrated squared error, and penalized functional estimation outperforming the tabular variant with $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$46. In the ADHD fMRI example, with $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$47, $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$48 time points, and spatial grid $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$49, the model used rank $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$50, identified cortical surface and white-matter regions as having the largest smooth time-varying effects, and improved $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$51-fold cross-validation MSE from approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$52 to approximately $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$53, with in-sample $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$54 (Li et al., 11 Jun 2025).

FunBaT was tested on synthetic data and climate-related datasets. On a synthetic rank-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$55 two-mode surface sampled at $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$56 random points in $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$57 with Gaussian noise, FunBaT with a Matern-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$58 kernel accurately recovered both mode functions with credible bands and reconstructed the full surface at unobserved points. On BeijingAir datasets for PM$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$59, PM$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$60, and SO$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$61, and on US-TEMP with modes latitude, longitude, and time, FunBaT and FunBaT-CP were reported to achieve the lowest RMSE/MAE among the listed baselines, with the example $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$62 versus $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$63 on PM$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$64, while also revealing geographic and temporal patterns such as northern versus southern temperature gradients and historic warming trends (Fang et al., 2023).

InfTucker was evaluated on chemometrics and social network datasets, where the paper states that the proposed models achieved significantly higher prediction accuracy than the most state-of-art tensor decomposition methods. The reported motivation was to handle complex interactions, multiple data types including binary data, and noisy observations within a probabilistic nonlinear tensor framework (Xu et al., 2011).

6. Relation to classical Tucker decomposition and recurring distinctions

A fundamental reference point in this literature is the ordinary discrete Tucker decomposition. The RKHS-based FTD states explicitly that if all modes are discrete and every mode factor is unconstrained with $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$65, the model reduces exactly to the ordinary rank-$$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$66 Tucker decomposition (Steidle et al., 26 Mar 2026). The regression formulation makes the same connection operationally by placing Tucker rank constraints on a discretized coefficient tensor and then regularizing only the functional mode (Li et al., 11 Jun 2025).

A common misconception is that “functional Tucker decomposition” refers to a single implementation. Taken together, the cited works show instead a family of constructions with different assumptions about continuity and different computational objectives. Some formulations are deterministic approximation schemes based on prescribed bases or interpolation grids, such as sparse polynomial or wavelet fitting and Chebyshev interpolation (Rai et al., 2019, Dolgov et al., 2020). Others are basis-free at the modeling stage, as in the RKHS formulation that “does not require an a-priori basis” (Steidle et al., 26 Mar 2026). Bayesian variants place GP priors on the mode functions, either with finite cores and scalable state-space inference or with an infinite-feature construction coupled to Kronecker variational inference (Fang et al., 2023, Xu et al., 2011).

Another recurring distinction concerns the role of continuity. In some settings, continuity is introduced primarily to recover or interpolate data on unstructured or shifted sampling schemes. The functional sparse Tucker method is motivated by the limitation that tensor compression “can only consider scientific data represented on structured grids,” and therefore treats unstructured-mesh data as realizations of a function (Rai et al., 2019). The RKHS-based FTD emphasizes interpolation to new $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$67-values and domain transfer on novel grids (Steidle et al., 26 Mar 2026). FunBaT further couples continuity with Bayesian uncertainty quantification through posterior distributions and credible bands (Fang et al., 2023).

Reported limitations are likewise model-specific. FunBaT notes that a dense core $$S^k_{n_k}=\operatorname{span}\{\phi^{(k)}_1,\dots,\phi^{(k)}_{n_k}\},$$68 may overfit when data are very sparse, and states that in such cases FunBaT-CP is preferable (Fang et al., 2023). Chebfun3F is specialized to trivariate functions on tensor-product domains and is framed as an alternative to slice-Tucker construction rather than as a general-purpose continuous-index tensor model (Dolgov et al., 2020). Functional sparse Tucker compression depends on interpolation to a structured grid before low-rank decomposition, whereas RKHS and GP-based formulations encode continuity directly in the factors (Rai et al., 2019, Steidle et al., 26 Mar 2026).

The aggregate picture is that functional Tucker methodology preserves the multilinear subspace structure of Tucker decomposition while replacing discrete tabulation in selected modes by continuous representations. Depending on the formulation, the resulting benefits may include compression on unstructured meshes, adaptive subspace modeling under domain shift, smooth coefficient estimation in regression, probabilistic continuous-index prediction, or nonlinear infinite-feature generalization (Rai et al., 2019, Steidle et al., 26 Mar 2026, Li et al., 11 Jun 2025, Fang et al., 2023, Xu et al., 2011).