---
title: Functional Tensor Tucker Decomposition
url: https://www.emergentmind.com/topics/functional-tensor-tucker-decomposition
type: topic
---

# Functional Tensor Tucker Decomposition

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 [1907.05884] [2007.16126] [2603.25530] [2506.09358] [2311.04829] [1108.6296]. 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 [1907.05884].

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 \(G\in\mathbb R^{r_1\times\cdots\times r_N}\) and factor functions \(f_n^{(i)}:I_n\to\mathbb R\). When some modes are discrete, the corresponding factors reduce to ordinary factor matrices \(A^{(n)}\in\mathbb R^{I_n\times r_n}\); when a mode is continuous, the factor functions may be constrained to belong to an RKHS \(\mathcal H_{K_n}\) [2603.25530].

The same multilinear structure also appears in continuous trivariate approximation. For \(f:[-1,1]^3\to\mathbb R\), a functional Tucker approximation of rank \((r_1,r_2,r_3)\) is
\[
f(x,y,z)\approx \sum_{i=1}^{r_1}\sum_{j=1}^{r_2}\sum_{k=1}^{r_3}
G_{ijk}\,\phi_i(x)\,\psi_j(y)\,\theta_k(z),
\]
where the minimal triple for exact representability is the multilinear rank of \(f\) [2007.16126].

A related formulation appears in regression, where the coefficient is a smoothly varying tensor-valued function \(B:\mathbb R\to\mathbb R^{p_1\times\cdots\times p_D}\). There the discretized coefficient tensor \(\Theta\in\mathbb R^{p_0\times p_1\times\cdots\times p_D}\) is constrained to Tucker rank \((r_0,\dots,r_D)\), and the mode-\(0\) factor \(U_0\) induces \(r_0\) smooth functional loadings \(u_0(s)\) [2506.09358].

## 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 \(p\) or multi-resolution wavelets of level \(s\) and degree \(p\). For each mode \(k\) and component \(j_k\), the coefficients are obtained from the sparse-regularized problem
\[
\min_{v\in\mathbb R^{n_k}}
\bigl\|W^{(k)}_{:,j_k}-\Phi^{(k)}v\bigr\|_2^2+\lambda\|v\|_1,
\]
with \(\lambda\) selected by LARS + cross-validation. This produces factor expansions with \(\|v\|_0\ll n_k\) [1907.05884].

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 \(u_i(x)\), \(v_j(y)\), and \(w_k(z)\) derived from factor matrices on refined grids [2007.16126].

In the RKHS-based Functional Tucker Decomposition, each continuous factor \(f_n^{(i)}\) lies in \(\mathcal H_{K_n}\), and by the Representer Theorem admits the expansion
\[
f_n^{(i)}(t)=\sum_{j=1}^{p_n} w^{(n)}_{j,i}\,K_n(t,x_{n,j}).
\]
At the sampled grid points, the factor evaluation matrix is \(C^{(n)}=K_nW^{(n)}\), where \(K_n\) is the Gram matrix and \(W^{(n)}\in\mathbb R^{p_n\times r_n}\) contains the unknown coefficients. The paper states that this construction models continuous modes without requiring an a-priori basis [2603.25530].

In functional tensor regression, the continuous mode is represented through a basis \(\psi(s)=(\psi_1(s),\dots,\psi_{p_0}(s))^\top\), typically spline-based, with
\[
B(s)=\Theta\times_0\{\psi(s)^\top(\Psi^\top\Psi)^{-1}\Psi^\top\},
\qquad
\Theta=\mathcal G\times_0 U_0\times_1 U_1\times\cdots\times_D U_D.
\]
Smoothness is not left to the basis alone; it is enforced by the quadratic penalty
\[
J(\Theta)=\langle \Theta\times_0 A_m,\Theta\rangle,
\]
which encodes an \(m\)th-order derivative regularization on the functional mode [2506.09358].

Bayesian continuous-indexed variants replace deterministic basis expansions by latent stochastic processes. FunBaT models each factor function with an independent Gaussian-process prior,
\[
f_{m,i}\sim \mathcal{GP}(0,k_m(\mathbf r,\mathbf r')),
\]
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 \(f_r^{(n)}(u_{i_n}^{(n)})\), then lets mode ranks tend to infinity to obtain a nonparametric Tucker model in an infinite feature space [2311.04829] [1108.6296].

## 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 \(10\%\), is interpolated onto a tensor-product grid. Second, sequentially truncated HOSVD is applied to the resulting structured tensor \(U\): each mode unfolding \(U_{(k)}\) is decomposed by SVD, truncated to the leading \(r_k\) singular vectors \(W^{(k)}\), and combined into a core
\[
G=U\times_1 W^{(1)T}\times_2\cdots\times_d W^{(d)T},
\]
with guaranteed relative Frobenius error not exceeding \(\epsilon\). Third, the singular vectors are fitted sparsely in functional dictionaries, after which the core coefficients are re-estimated against the original \(Q\) samples by randomized least squares using a sign-flip, a fast mixing transform such as FFT, and uniform row sampling [1907.05884].

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 [2007.16126].

The RKHS-based FTD is optimized by an ALS-style loop. Discrete factors \(A^{(m)}\) are updated by mode-wise least squares; continuous coefficients \(W^{(n)}\) are updated by solving the linear system
\[
L\,\mathrm{vec}(W^{(n)})=R,
\]
with
\[
L=(G_{(n)}G_{(n)}^\top)\otimes K_n+\lambda_n I_{r_np_n};
\]
and the core tensor \(G\) is updated by least squares with the factor matrices fixed. Iteration continues until the relative reconstruction error
\[
\epsilon_t=\|T-[\![G;\dots]\!]\|_F/\|T\|_F
\]
changes by less than a tolerance [2603.25530].

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 \(\mathcal Z^*y\), each iteration computes a tangent-space basis, solves the Gauss--Newton linear system
\[
[R^*\mathcal Z^*\mathcal ZR+n\rho R^*AR]\Delta=R^*\mathcal Z^*y,
\]
lifts the tangent update back to ambient space, and retracts by T-HOSVD to rank \(r\). The paper states that because the penalty acts only in mode \(0\), the extra per-iteration cost enters only in small \(r_0\times r_0\) blocks [2506.09358].

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 \(\tau\), \(\mathcal G\), 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 \(\mathcal M\), with Gaussian or probit likelihoods, a Kronecker-structured covariance \(\mathbf K=K^{(1)}\otimes\cdots\otimes K^{(N)}\), and ELBO-based updates for latent variables, kernel hyperparameters, and latent inputs [2311.04829] [1108.6296].

## 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 \(\sum_{k=1}^d n_k r_k\) coefficients and storing the core requires \(\prod_{k=1}^d r_k\), so the total storage is
\[
\prod_{k=1}^d r_k+\sum_{k=1}^d n_k r_k.
\]
For \(Q\) original samples stored in double precision, the compressed representation requires
\[
8\Bigl(R+\sum_k n_k r_k\Bigr)\ \text{bytes},
\qquad
R=r_1\cdots r_d,
\]
and the compression ratio is
\[
\mathrm{CR}=\frac{Q}{R+\sum_k n_k r_k}.
\]
The paper states that practical choices with \(r_k\ll I_k\) and sparse expansions yield \(\mathrm{CR}\sim 10^3\text{--}10^5\) [1907.05884].

The randomized refinement step avoids the direct least-squares cost \(O(QR^2)\) and storage \(O(QR)\) of the full design matrix \(\Phi_{\rm core}\). With sketch size \(S\), the reduced problem costs \(O(SR^2)\), and Johnson--Lindenstrauss type bounds imply that for \(S=O(R\log R/\varepsilon^2)\) the sketched solution approximates the full least-squares solution within factor \((1\pm\varepsilon)\). The same paper reports rapid self-convergence once \(S\gtrsim 2.5R\) [1907.05884].

For trivariate function approximation, Chebfun3F reduces function evaluations by selecting all mode fibers directly. The paper gives idealized function-evaluation costs \(O(r^3+nr^2)\) for Chebfun3 and \(O(r^3+nr)\) for Chebfun3F, and states that in practice Chebfun3F typically reduces evaluations by \(75\%\) and sometimes by over \(98\%\). The final storage in Tucker-Chebyshev form is \(O(r_1r_2r_3+n_1r_1+n_2r_2+n_3r_3)\) rather than \(O(n_1n_2n_3)\). Error splitting is expressed as
\[
\|f-\hat f\|_\infty
\le
\|f-\tilde f\|_\infty
+
\Lambda_{n_1-1}\Lambda_{n_2-1}\Lambda_{n_3-1}\,\|T-\hat T\|_\infty,
\]
and for analytic \(f\), the interpolation error decays exponentially as \(C\rho_{\min}^{-n}\) [2007.16126].

Functional tensor regression provides explicit convergence and estimation guarantees. Under a functional-TRIP and smoothness assumptions on \(B(\cdot)\), the authors prove quadratic convergence of the Riemannian Gauss--Newton iterates,
\[
\|\Theta^{k+1}-\Theta^*\|_F
\le
C_1\|\Theta^k-\Theta^*\|_F^2+C_2\eta,
\]
derive an estimation error bound for \(\|\hat\Theta-\Theta^*\|_F^2\), and state a matching minimax lower bound that is optimal in the tensor dimensions [2506.09358].

Bayesian continuous-index models target scalability by exploiting structure in the latent processes. FunBaT states overall complexity \(\mathcal O(NMR)\), where \(R=\max_m R_m\), because message passing and state-space smoothing are linear in the number of observations and modes. InfTucker reduces naïve \(O(n^3)\) covariance operations by Kronecker eigendecompositions, giving costs of order \(O(\sum_n I_n^3+n\sum_n I_n)\), with further reduction under truncated SVD [2311.04829] [1108.6296].

## 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 \(500\times 500\times 300\) grid with approximately \(7.5\times 10^7\) points and size \(0.6\,\mathrm{GB}\), TuckerMPI at precision \(\epsilon=10^{-2}\) produced ranks \((25,24,8)\) and core size \(R=4800\); the final sparse functional Tucker representation used approximately \(155\,\mathrm{KB}\), achieved compression ratio approximately \(3.9\times 10^3\), and reconstruction error approximately \(1.0\times 10^{-2}\). At \(\epsilon=10^{-4}\), the reported ranks were \((57,50,17)\), \(R=48\,450\), storage approximately \(640\,\mathrm{KB}\), compression ratio \(936\), and error approximately \(1.9\times 10^{-4}\). For SP4D, the full \(500^3\times 300\) dataset of approximately \(3.75\times 10^{10}\) points and size \(300\,\mathrm{GB}\) yielded ranks \((30,38,5,11)\), storage approximately \(673\,\mathrm{KB}\), compression ratio approximately \(4.5\times 10^5\), and error approximately \(1.1\times 10^{-2}\) at \(\epsilon=10^{-2}\). 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 [1907.05884].

The RKHS-based FTD was evaluated in domain-variant classification on three tasks, each represented as a \(4\)-way tensor with one continuous mode. In a semi-synthetic digit experiment with continuous mode \(p=50\), training on every fourth sample and testing on a shifted domain, HOSVD accuracy dropped from approximately \(90\%\) to approximately \(60\%\), while FTD remained above approximately \(80\%\), and Macro-\(F_1\) 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 \(30\%\), while FTD maintained approximately \(70\%\); Macro-\(F_1\) was approximately \(0.7\) under FTD versus approximately \(0.3\) for HOSVD. In hyperspectral mango ripeness classification, with spectral bands as the continuous mode, HOSVD accuracy on the shifted domain was at or below \(40\%\), described as chance level, whereas FTD reached approximately \(60\%\) accuracy and Macro-\(F_1\approx 0.6\) [2603.25530].

Functional tensor regression was studied by simulation and in ADHD fMRI analysis. In the simulation with \(p_0=12\), \(p_1=p_2=8\), true rank \((2,3,3)\), and \(n=500\), the paper reports quadratic decay of \(\|\Theta^k-\Theta^*\|\) over iterations, generalized cross-validation selecting \(\rho\) with lowest relative integrated squared error, and penalized functional estimation outperforming the tabular variant with \(\rho=0\). In the ADHD fMRI example, with \(n\approx 50\), \(p_0=50\) time points, and spatial grid \(8\times 8\times 4\), the model used rank \((2,2,2,2)\), identified cortical surface and white-matter regions as having the largest smooth time-varying effects, and improved \(10\)-fold cross-validation MSE from approximately \(265\) to approximately \(122\), with in-sample \(R^2\approx 0.99\) [2506.09358].

FunBaT was tested on synthetic data and climate-related datasets. On a synthetic rank-\(1\) two-mode surface sampled at \(650\) random points in \([0,1]^2\) with Gaussian noise, FunBaT with a Matern-\(3/2\) kernel accurately recovered both mode functions with credible bands and reconstructed the full surface at unobserved points. On BeijingAir datasets for PM\(_{2.5}\), PM\(_{10}\), and SO\(_2\), 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 \(0.29\) versus \(1.0\) on PM\(_{2.5}\), while also revealing geographic and temporal patterns such as northern versus southern temperature gradients and historic warming trends [2311.04829].

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 [1108.6296].

## 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 \(\lambda_n=0\), the model reduces exactly to the ordinary rank-\((r_1,\dots,r_N)\) Tucker decomposition [2603.25530]. 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 [2506.09358].

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 [1907.05884] [2007.16126]. Others are basis-free at the modeling stage, as in the RKHS formulation that “does not require an a-priori basis” [2603.25530]. 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 [2311.04829] [1108.6296].

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 [1907.05884]. The RKHS-based FTD emphasizes interpolation to new \(t\)-values and domain transfer on novel grids [2603.25530]. FunBaT further couples continuity with Bayesian uncertainty quantification through posterior distributions and credible bands [2311.04829].

Reported limitations are likewise model-specific. FunBaT notes that a dense core \(\mathcal G\) may overfit when data are very sparse, and states that in such cases FunBaT-CP is preferable [2311.04829]. 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 [2007.16126]. 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 [1907.05884] [2603.25530].

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 [1907.05884] [2603.25530] [2506.09358] [2311.04829] [1108.6296].

Source: https://www.emergentmind.com/topics/functional-tensor-tucker-decomposition