Pith. sign in

REVIEW 4 major objections 5 minor 71 references

This paper claims that multi-way tensor imputation can be cast as kernel regression with coefficient tensors on fixed-rank tensor-train manifolds and Hadamard overparameterization, yielding automated kernel hyperparameter selection and stat

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:06 UTC pith:5C73DKC3

load-bearing objection Solid kernel-TT imputation method whose empirical superiority claim is undercut by test-set hyperparameter selection and missing error bars. the 4 major comments →

arxiv 2607.17390 v2 pith:5C73DKC3 submitted 2026-07-19 stat.ML cs.LGeess.SP

Kernel Regression with Tensor Trains and Hadamard Overparameterization

Duc Thien Nguyen , Konstantinos Slavakis , Eleftherios Kofidis , Dimitris Pados This is my paper
classification stat.ML cs.LGeess.SP
keywords tensor completiontensor trainskernel regressionRKHSHadamard overparameterizationRiemannian optimizationfMRI imputationedge-flow imputation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Multi-way data—brain scans, traffic flows on networks—arrive with missing entries. This paper tries to show that those gaps can be filled by a nonparametric regression model: each entry is approximated by kernel functions, the regression coefficients are forced onto low-dimensional tensor-train manifolds, and the kernel bandwidth matrices are learned on the manifold of positive-definite matrices rather than tuned by cross-validation. The Hadamard product overparameterization of the coefficient tensors is claimed to induce sparsity and improve representation capacity. On two real tasks, 4D fMRI reconstruction and dynamic graph edge-flow imputation, the paper reports that this framework consistently outperforms state-of-the-art tensor, Bayesian, and neural-network baselines in normalized root-mean-squared error. If true, automated kernel hyperparameter selection becomes viable for kernel tensor regression.

Core claim

On the paper's own terms, the discovery is that an unknown data tensor Y can be modeled as a sum of terms (⊙_p U_{ν,p}) ×_1 K_ν ×_1 (⊙_q V_{ν,q}), where the U and V factors live on fixed-rank tensor-train manifolds and K_ν are Gaussian kernel matrices with learnable covariance matrices on the manifold of positive-definite matrices. Solving the resulting smooth inverse problem by Riemannian gradient descent—with retractions computed by TT-SVD and gradients orthogonally projected onto tangent spaces—the paper claims KReTTaH achieves lower NRMSE than leading tensor, Bayesian, and neural-network baselines across sampling ratios in both fMRI and edge-flow experiments, while also reducing paramete

What carries the argument

The central object is the Cartesian-product Riemannian manifold M = M_{r1}^{NK·P} × M_{r2}^{NK·Q} × (S_{D_l}^{++})^{NK}, with the model X = Σ_ν (⊙_p U_{ν,p}) ×_1 K_ν ×_1 (⊙_q V_{ν,q}). The machinery has four parts: (i) fixed-rank tensor-train manifolds constrain the coefficient tensors; (ii) Hadamard overparameterization U = ⊙_p U_p induces a quasi-norm regularizer that promotes sparsity; (iii) learnable covariance matrices C_ν on the SPD manifold replace manual kernel-bandwidth tuning; (iv) Riemannian gradient descent with Armijo line search, tangent-space projection, and TT-SVD retraction solves the smooth nonconvex problem, with every limit point a critical point of the loss.

Load-bearing premise

The model assumes every unknown entry can be represented in the span of a small set of landmark feature maps chosen from the observed data; if those landmarks do not cover the geometry of the missing entries (e.g., at very low sampling ratios), the approximation error cannot be reduced by optimization.

What would settle it

Take a tensor with known low-dimensional structure, sample entries only within a few well-separated spatial patches so that LLE landmarks cannot cover the unsampled patches, run KReTTaH at a very low sampling ratio (s=0.05), and check whether its NRMSE is no better than a simple k-NN imputation or a fixed-bandwidth kernel baseline; if so, the landmark-span assumption fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Automated kernel hyperparameter selection: learning covariance matrices on the SPD manifold removes the need for cross-validation over Gaussian bandwidths.
  • Sparsity with lower storage: Hadamard overparameterization produces sparse coefficient tensors; the (1,2) setting reported to cut parameter storage by more than 40 percent while improving accuracy.
  • Rank reduction: using multiple kernels allows tensor-train ranks three to eight times smaller than a single kernel, due to the additive property of tensor ranks.
  • Training-data-free and interpretable: the model performs regression in an RKHS and needs no external training labels, unlike neural-network baselines.
  • Joint Riemannian updates reduce sensitivity to initialization and open a path to stochastic and online learning from streaming multi-way data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The landmark/navigator construction is the real bottleneck: accuracy likely tracks how well LLE plus max-min landmark selection covers the manifold of observed slices; at very low sampling ratios the landmarks may miss the geometry of missing entries, so the empirical superiority claim may not generalize.
  • Because learned covariances reportedly match hand-tuned bandwidths, the SPD-manifold optimization is doing genuine work; a testable extension is to replace landmarks with a learned or adaptive landmark set and measure whether NRMSE improves further.
  • Since the model is nonparametric and fitted to the observed entries themselves, the same framework should transfer to other structured sparse-tensor problems (recommender systems, sensor networks) whenever navigator data can be defined.
  • The open question about the TT-SVD retraction domain makes step-size control via line search load-bearing; a direct comparison against retraction-free or projected-gradient variants would isolate that mechanism.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces KReTTaH, a kernel-regression framework for multi-way tensor imputation. Each entry of the data tensor is modeled as a sum of multi-kernel RKHS regressions, where the coefficient tensors are Hadamard products of fixed-rank tensor-train factors and the Gaussian kernel covariance matrices are learned on the SPD manifold. The inverse problem (8) is solved by Riemannian gradient descent (Algorithm 2), with partial Riemannian gradients given in Proposition 10. The method is evaluated on 4D fMRI imputation and dynamic graph edge-flow imputation against RTTC, VKBTR, NCP, STTC, S-VAR, PS, and HodgeNet, and the paper claims consistent superiority in NRMSE.

Significance. If the empirical claims can be substantiated, the contribution is significant: the paper offers a principled way to couple RKHS regression with fixed-rank TT geometry and Hadamard overparameterization, and it automates Gaussian-kernel bandwidth selection via Riemannian optimization on the SPD manifold, avoiding cross-validation. The gradient derivation in Proposition 10 appears internally consistent, and the multi-kernel extension over the authors' preliminary work is a genuine step forward. The Hadamard sparsity mechanism is also well motivated by the cited literature. However, the experimental validation as reported does not support the central 'consistently outperforms' claim because the hyperparameter-selection protocol and the lack of uncertainty quantification undermine the comparison.

major comments (4)
  1. [§4.1–4.2 (evaluation protocol)] The central empirical claim rests on the statement that 'All methods are finely tuned to reach their lowest NRMSE' and on grid-search selection of N_l, r1, r2, P, Q, λ, and baseline settings. The reported NRMSE is computed on the held-out missing entries, and no validation split is described. If the same test entries are used to select hyperparameters, the comparison is optimistically biased and does not measure predictive performance. The paper needs a nested validation protocol (or at least a separate validation set used identically for all methods) before the superiority claim can be accepted. The absence of code and data currently prevents an independent check of this tuning protocol.
  2. [§4.1–4.2, Figures 3, 7, 12] All results are reported as averages over 10 runs, without error bars, standard deviations, or significance tests. Some reported differences are small (e.g., Figure 7, BF network at s=0.1; Figure 12, KReTTaH vs. KReTTaHσ). Without uncertainty quantification, the claim of 'consistently outperforms' is not statistically supported, even under a corrected evaluation protocol. Per-run results and significance tests (or at least error bars) should be reported for every table/figure that supports a comparison claim.
  3. [§3.3 (landmark span assumption)] The model assumes both the regression functions f_{i1,...,im} and the feature vectors φ(µ_{im+1,...,iN}) lie in the span of the N_l landmark feature maps {φ(l_j)}. This is not the standard representer theorem, because the location parameters µ are also optimized and the landmark set is chosen by LLE/max-min selection from navigator data. If the landmarks do not cover the low-dimensional geometry of the missing entries — for example at very low sampling ratios — the representational error cannot be reduced by better optimization. The paper provides no sensitivity analysis with respect to N_l, landmark selection strategy, or sampling ratio, so the empirical generalization claim is not yet grounded.
  4. [§3.5 (convergence guarantees)] The paper states that by [32, Thm. 4.3.1] every limit point of Algorithm 2 is a critical point, while immediately acknowledging that M_r is not closed and limit points may leave the manifold. Since Term2 is not included in the implemented algorithm, the hypotheses of the cited convergence theorems are not verified. Similarly, the claim that 'convergence rates can also be obtained' via [32, Thm. 4.5.6] is not justified for the specific retraction and line search used here. The section should distinguish rigorously established results from empirical observations, and should state explicitly which conditions are assumed rather than proved.
minor comments (5)
  1. [Equation (9h)–(9i)] The contraction notation in the definition of ˜K_ν and ˜C_ν is very dense and hard to parse. A short worked scalar example or a verbal description of the contraction modes would substantially improve readability.
  2. [Figure 12 and §4.1] KReTTaHσ is introduced in the text only after it appears in Figure 12; define it clearly before first use, and state that for KReTTaHσ the σ_z values are tuned on the same criterion as the other hyperparameters.
  3. [Figure 11] The sparsity measure is defined with an arbitrary threshold 10^-3 after normalization. This is fine, but the threshold should be stated in the main text and ideally accompanied by the actual fraction of exact zeros or a sensitivity check.
  4. [§4.1] The λ selection is described via an L-curve, but the surrounding text says hyperparameters are selected by grid search. Clarify whether λ is part of the grid search or selected separately by the L-curve criterion, and by which criterion the other hyperparameters are selected.
  5. [Appendix/Notation] Minor typos and notation inconsistencies remain, e.g., 'crank TT' used in Figure 1 and Definition 5, and the unspecified use of 'n' in the line-search description of Algorithm 2. These should be cleaned up.

Circularity Check

0 steps flagged

No significant circularity: the model, gradients, and optimization are derived in-paper, and the empirical claim is checked against external baselines.

full rationale

The paper's derivation chain is self-contained. The data model (Eq. 6) is an explicit modeling ansatz; the inverse problem (Eq. 8) and Riemannian gradient formulas in Proposition 10 are derived in the paper (Appendix C) from the model and objective, not imported from the authors' prior work by fiat. The learned covariance matrices C_nu are genuine optimization variables, and the comparison against fixed-bandwidth KReTTaH_sigma is a proper ablation rather than a fitted quantity relabeled as a prediction. Self-citations to [19-21] and [28] are used for positioning and for the claimed extension to multi-way data, but they are not load-bearing for the gradient computation, convergence argument, or experimental comparison; the empirical claims are evaluated against external methods (RTTC, VKBTR, NCP, HodgeNet, S-VAR, PS, STTC) on public datasets. The concern raised by the skeptic about test-set hyperparameter selection and absent error bars is a correctness/evaluation-protocol issue, not a circularity of the derivation: no step reduces by definition to its own input. Thus the paper does not exhibit self-definitional, fitted-input-called-prediction, or self-citation-driven circularity.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The central claim rests on a small set of modeling choices: landmark feature-space span (Section 3.3), TT-SVD retraction validity (Remark 18), convergence of R-GD on nonclosed manifolds (Section 3.5), and the dataset/measurement protocol (uniform random missingness). The method also carries many hand-selected hyperparameters (ranks, P, Q, N_l, D_l, N_K, m, λ, graph-prior weights); the learned C_ν are parameters fit to the imputation loss, not free external constants.

free parameters (9)
  • λ (sparsity regularization weight) = 0.1 (L-curve corner, Fig. 14)
    Controls the balance between data fidelity and Hadamard sparsity in (8a); selected per dataset via the L-curve criterion.
  • Kernel covariance matrices C_ν = Learned on S^{D_l}_{++}; no single reported value
    The Gaussian kernel bandwidths are fit to the imputation loss in (8) rather than fixed a priori; this is the paper's 'automated kernel hyperparameter selection' mechanism.
  • TT ranks r1, r2 = Grid search over (1, 8r, ..., 8r, 1), r ∈ {1,...,12}; best ranks per dataset
    Rank selection is a free model choice; the paper grid-searches ranks to minimize NRMSE on the evaluation data.
  • Hadamard factor counts P, Q = P, Q ∈ {1,2,3}; (1,2) consistently best
    Number of Hadamard factors selected by grid search; directly controls sparsity and overparameterization.
  • Number of landmark points N_l = N_l ∈ {50, 100, 150} (10 l, l = 5,...,15)
    Landmark sample size chosen by grid search; affects kernel matrix size and model capacity.
  • LLE embedding dimension D_l = 5 (used in Fig. 13 and experiments)
    Dimensionality of navigator data after LLE; chosen by the user, kept small to make SPD covariance optimization tractable.
  • Number of kernels N_K = 1 or 6 in experiments
    Number of Gaussian kernels; N_K = 6 improves graph-flow accuracy in Fig. 12 and is selected as a model hyperparameter.
  • Splitting index m = m ∈ {1,2,3} for fMRI; m = 1 for edge flows
    The tensor mode split in model (5)–(6) is user-chosen; all results depend on this choice.
  • Graph-prior weights λ_l, λ_u = Set to 0 in the reported best results
    Divergence/curl regularization weights in Section 4.2 were tuned and found best at zero; this is part of the method's parameter selection.
axioms (5)
  • ad hoc to paper Regression functions f_{i1,...,im} and feature vectors φ(µ) lie in the span of the N_l landmark feature maps {φ(l_j)}.
    Invoked in Section 3.3 before equation (6); not guaranteed by the representer theorem because the location parameters µ are also optimized rather than fixed training points.
  • domain assumption TT-SVD (Algorithm 1) acts as a retraction and line-search step sizes remain small enough that iterates stay on the fixed-rank manifold M_r.
    Remark 18 explicitly states the retraction domain is not fully characterized; the algorithm's practical convergence relies on this assumption.
  • domain assumption Riemannian gradient descent on the nonclosed fixed-rank product manifold converges to a useful critical point rather than a spurious local minimum.
    Section 3.5 cites [32, Thm. 4.3.1] although M_r is not closed; the paper discusses regularization fixes but does not implement Term 2, and no landscape analysis for (8) is provided.
  • domain assumption Navigator data, LLE compression, and max-min landmark selection faithfully represent the low-dimensional geometry of the missing entries.
    Section 3.1; if the observed entries' geometry is not captured by LLE on P_Ω(Y), the learned kernels cannot encode the missing entries.
  • domain assumption Observed entries are missing uniformly at random and are representative of the full tensor.
    Experiments sample Ω uniformly; real fMRI and traffic data may have structured missingness, so the benchmark claim is conditional on this assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 25614 in / 15154 out tokens · 152421 ms · 2026-08-01T18:06:19.483131+00:00 · methodology

0 comments
read the original abstract

Kernel regression with tensor trains and Hadamard overparameterization (KReTTaH) is introduced as a training-data-free, interpretable, and nonparametric framework for multi-way data imputation. The imputation problem is reformulated as regression in reproducing kernel Hilbert spaces (RKHS), where the tensor regression coefficients are explicitly constrained to lie on fixed-rank tensor-train (TT) manifolds and structured via Hadamard overparameterization to promote sparsity and high representational efficiency. Rather than relying on costly cross-validation, KReTTaH jointly optimizes the TT coefficient tensors and the kernel covariance matrices within a Riemannian product-manifold framework -- the former on fixed-rank TT manifolds, the latter on the manifold of positive-definite matrices -- thereby enabling automated kernel-hyperparameter selection. Numerical tests on two challenging applications -- imputation of high-dimensional functional magnetic resonance imaging (fMRI) data and recovery of missing edge flows in dynamic graphs -- demonstrate that KReTTaH consistently outperforms state-of-the-art tensor-, Bayesian-, and neural-network-based baselines in terms of modeling accuracy.

Figures

Figures reproduced from arXiv: 2607.17390 by Dimitris Pados, Duc Thien Nguyen, Eleftherios Kofidis, Konstantinos Slavakis.

Figure 1
Figure 1. Figure 1: A visualization of an order-3 TT tensor (see Definition 5). The (3, 3, 1)th entry of a TT-tensor A ∈ R 4×4×4 is computed by the matrix multiplication of the (lateral) slices of the core tensors, i.e., A(3, 3, 1) = A1(:, 3, :) A2(:, 3, :) A3(:, 1, :). The core tensors A1 ∈ R 1×4×4 , A2 ∈ R 4×4×3 , A3 ∈ R 3×4×1 yield the TT compression rank crankTT(A) = (1, 4, 3, 1). 2.2 Background on tensor trains Definitio… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Illustration of a smooth manifold M. The tangent space TΘM at Θ comprises all tangent vectors ξ of M at Θ. The retraction map RΘ(ξ) maps a tangent vector ξ ∈ TΘM to a point in M; it serves as an extension and a tractable alternative of the exponential map; see Definition 15 in Appendix B. (b) Illustration of one Riemannian-gradient descent step in Algorithm 2. The Euclidean gradient of a smooth loss L(… view at source ↗
Figure 3
Figure 3. Figure 3: Mean NRMSE value curves (↓) vs. sampling ratios. RTTC [22]: , VKBTR [17]: , NCP [12]: , KReTTaH (NK = 1): . NCP RTTC VKBTR (1, 1) (1, 2) (2, 1) 0 20 40 60 80 100 120 140 Methods Time (minutes) COBRE BOLD [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average run-time (in minutes) across all sampling ratios (with hyperparameters achieving the lowest NRMSE) in the 4D-fMRI application. The horizontal axis labels in brackets denote KReTTaH at different (P, Q). Time is measured until a stopping criterion, similar to the one in Line 9 of Algorithm 2 with ϵ = 10−4 , is satisfied. NCP RTTC VKBTR (1, 1) (1, 2) (2, 1) 5 6 7 8 9 Methods Time (hours) COBRE BOLD [… view at source ↗
Figure 5
Figure 5. Figure 5: Average run-time (in hours) across all sampling ratios in the 4D-fMRI application. The horizontal axis labels in brackets denote KReTTaH at different (P, Q). Measured time includes the hyperparameter-tuning step. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: An example of a graph, G = (V, E), and its set of triangles, T . nodes, N1 = |E| the number of edges, and N2 = |T | the number of triangles of G [47–50]. The incidence matrix B1 ∈ R N0×N1 captures the node-to-edge adjacencies, while B2 ∈ R N1×N2 encodes the edge-to-triangle ones. These incidence matrices depend on an arbitrary choice of orientation: the e-th edge is oriented as an ordered pair of nodes (i,… view at source ↗
Figure 7
Figure 7. Figure 7: Mean NRMSE value curves (↓) vs. sampling ratios. HodgeNet [63]: , S-VAR [52, 53]: , PS [62]: , RTTC [22]: , STTC [64]: , NCP [12]: , VKBTR [17]: , KReTTaH (NK = 1): , KReTTaH (NK = 6): . d·e is the ceiling function. This sampling pattern suggests that the number of observations is consistent over time. Navigator data and landmark points are also built in the same manner as in Section 4.1. Hyperparameters a… view at source ↗
Figure 8
Figure 8. Figure 8: The impact of different (P, Q) values in (6) (NK = 1) on performance. NRMSE values are averaged over all sampling ratios. HodgeNet RTTC STTC NCP VKBTR (1, 1) (1, 2) (1, 3) (2, 1) (2, 2) (2, 3) (3, 1) (3, 2) (3, 3) 0 5 10 15 20 25 30 Methods Time (minutes) EMA BF [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Average run-time (in minutes) across all sampling ratios (with hyperparameters achieving the lowest NRMSE) in the edge-flow estimation application. The horizontal axis labels in brackets are KReTTaH at different (P, Q). Time is measured until a stopping criterion, similar to the one in Line 9 of Algorithm 2 with ϵ = 10−4 , is satisfied. The notation (P, Q) refers to KReTTaH with P, Q corresponding to the H… view at source ↗
Figure 10
Figure 10. Figure 10: Average run-time (in minutes) across all sampling ratios in the edge flow imputation task. The horizontal axis labels in brackets denote KReTTaH at different (P, Q). Measured time includes the hyperparameter-tuning step. (1, 1) (1, 2) (1, 3) (2, 1) (2, 2) (2, 3) (3, 1) (3, 2) (3, 3) 10 20 30 40 50 60 70 80 90 100 (P, Q) Sparsity (%) ˆ (∗) ˆ(∗) [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Average sparsity of the U (∗) ν and V (∗) ν tensors depending on the number of HP factors P and Q in the EMA dataset, when NK = 1. U (∗) ν and V (∗) ν are calculated from the outputs of Algorithm 2. The sparsity is measured as the percentage of the tensor entries with absolute value less than 10−3 after having normalized all entries to a maximum of unity. 0.1 0.2 0.3 0.4 0.5 0.05 0.07 0.11 0.12 0.15 Sampl… view at source ↗
Figure 12
Figure 12. Figure 12: Mean NRMSE value curves (↓) vs. sampling ratios. KReTTaHσ (NK = 1): , KReTTaH (NK = 1): , KReTTaHσ (NK = 6): , KReTTaH (NK = 6): . KReTTaHσ (dashed line) denote the case of KReTTaH where, for all ν ∈ J1, NKK, Cν = σzIDl is fixed and σz need to be tuned to achieve to lowest NRMSE. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Effect of ranks r1 and r2. Both datasets are set at a sampling ratio s = 0.5. P = Q = 1 , Nl = 100, Dl = 5. P = Q = 1, Nl = 100, Dl = 5. Rank selection is a crucial and challenging aspect of TD methods. While this paper proposes KReTTaH that uses grid-search to locate the ranks, it acknowledges that there are techniques to improve, e.g., the sparsifying effect in Bayesian TD [14, 17] or the classic MDL [6… view at source ↗
Figure 14
Figure 14. Figure 14: L-curve for the selection of the regularization hyperparameter λ in Eq. (8a). The plot displays the data-fidelity loss term F versus the regularization term H in logarithmic scale at different λ (Eq. (8a)). According to [46], a good value of λ is located at the corner of the L-shaped curve (λ = 0.1). [2] Q. Song, H. Ge, J. Caverlee, and X. Hu, “Tensor completion algorithms in big data analytics,” ACM Tran… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 3 linked inside Pith

  1. [1]

    Tensordecompositionforsignalprocessing andmachinelearning,

    N.D.Sidiropoulos,L.DeLathauwer,X.Fu,K.Huang,E.E.Papalexakis,andC.Faloutsos,“Tensordecompositionforsignalprocessing andmachinelearning,” IEEE Trans. Signal Process.,vol.65,no.13,pp.3551–3582,2017. 19 102 Data-fidelity loss term F (log scale) 102 3 × 101 4 × 101 6 × 101 Regularization term H (log scale) = 1e-03 = 1e-02 = 5e-02 = 1e-01 = 5e-01 = 1e+00 = 1e+0...

    2017

  2. [2]

    Tensorcompletionalgorithmsinbigdataanalytics,

    Q.Song,H.Ge,J.Caverlee,andX.Hu,“Tensorcompletionalgorithmsinbigdataanalytics,” ACM Trans. Knowl. Disc. Data,vol.13, no.1,pp.1–48,Jan.2019

    2019

  3. [3]

    Tensor-traindecomposition,

    I.V.Oseledets,“Tensor-traindecomposition,” SIAM J. Sci. Comput.,vol.33,no.5,pp.2295–2317,2011

    2011

  4. [4]

    Tensorringdecomposition,

    Q.Zhao,G.Zhou,S.Xie,L.Zhang,andA.Cichocki,“Tensorringdecomposition,”arXiv:1606.05535[cs.NA],Jun.2016

    Pith/arXiv arXiv 2016

  5. [5]

    Tensor-train decomposition for image classification problems,

    D. Brandoni, “Tensor-train decomposition for image classification problems,” Ph.D. dissertation, University of Bologna, 2022. [Online].Available: https://amsdottorato.unibo.it/id/eprint/10121/3/phd_thesis_DomitillaBrandoni_final.pdf

    2022

  6. [6]

    The Tucker and Tensor Train decompositions,

    C. F. Van Loan, “The Tucker and Tensor Train decompositions,” CIME-EMS Summer School, Cetraro, Italy, Jun. 2015. [Online]. Available:https://www.dm.unibo.it/~simoncin/CIME/vanloan3.pdf

    2015

  7. [7]

    On manifolds of tensors of fixed TT-rank,

    S. Holtz, T. Rohwedder, and R. Schneider, “On manifolds of tensors of fixed TT-rank,”Numer. Math., vol. 120, no. 4, pp. 701–731, 2012

    2012

  8. [8]

    Berlin: Springer,2022

    J.W.RobbinandD.A.Salamon, Introduction to Differential Geometry. Berlin: Springer,2022

    2022

  9. [9]

    Theoryofreproducingkernels,

    N.Aronszajn,“Theoryofreproducingkernels,” Trans. Amer. Math. Soc.,vol.68,no.3,pp.337–404,1950

    1950

  10. [10]

    Cambridge,MA:MITPress,2002

    B.SchölkopfandA.J.Smola, Learning with Kernels. Cambridge,MA:MITPress,2002

    2002

  11. [11]

    Some results on Tchebycheffian spline functions,

    G. Kimeldorf and G. Wahba, “Some results on Tchebycheffian spline functions,”J. Math. Anal. Appl., vol. 33, no. 1, pp. 82–95, Jan. 1971

    1971

  12. [12]

    Nonparametric low-rank tensor imputation,

    J. A. Bazerque, G. Mateos, and G. B. Giannakis, “Nonparametric low-rank tensor imputation,” inProc. SSP, Ann Arbor, MI, Aug. 2012

    2012

  13. [13]

    Tensor decomposition meets RKHS: Efficient algorithms for smooth andmisaligneddata,

    B. W. Larsen, T. G. Kolda, A. R. Zhang, and A. H. Williams, “Tensor decomposition meets RKHS: Efficient algorithms for smooth andmisaligneddata,” arXiv preprint arXiv:2408.05677,2024

    Pith/arXiv arXiv 2024

  14. [14]

    When bayesian tensor completion meets multioutput gaussian processes: Functional universality and rank learning,

    S. Li, S. Fang, L. Cheng, F. Yin, Y.-C. Wu, P. Gerstoft, and S. Theodoridis, “When bayesian tensor completion meets multioutput gaussian processes: Functional universality and rank learning,”IEEE Transactions on Signal Processing , vol. 73, pp. 5319–5335, 2025

    2025

  15. [15]

    Kernel-based tensor partial least squares for reconstruction of limb move- ments,

    Q. Zhao, G. Zhou, T. Adalı, L. Zhang, and A. Cichocki, “Kernel-based tensor partial least squares for reconstruction of limb move- ments,”in Proc. IEEE ICASSP,Vancouver,Canada,May2013

  16. [16]

    Gradient-based optimization for regression in the functional tensor-train format,

    A. A. Gorodetsky and J. D. Jakeman, “Gradient-based optimization for regression in the functional tensor-train format,”J. Comput. Phys.,vol.374,pp.1219–1238,2018

    2018

  17. [17]

    Kernel Bayesian tensor ring decomposition for multiway data recovery,

    Z. Huang, G. Zhou, Y. Qiu, X. Chen, and Q. Zhao, “Kernel Bayesian tensor ring decomposition for multiway data recovery,”Neural Netw.,p.107500,2025

    2025

  18. [18]

    Györfi, M

    L. Györfi, M. Kohler, A. Krzyżak, and H. Walk,A Distribution-Free Theory of Nonparametric Regression . New York: Springer, 2010

    2010

  19. [19]

    Multilinear kernel regression and imputation via manifold learning,

    D. T. Nguyen and K. Slavakis, “Multilinear kernel regression and imputation via manifold learning,”IEEE Open J. Signal Process. , vol.5,pp.1073–1088,Nov.2024

    2024

  20. [20]

    Imputation of time-varying edge flows in graphs by multilinear kernel regression and manifoldlearning,

    D. T. Nguyen, K. Slavakis, and D. Pados, “Imputation of time-varying edge flows in graphs by multilinear kernel regression and manifoldlearning,” Signal Process.,vol.237,Dec.2025

    2025

  21. [21]

    Estimating dynamic graph flows with kernel models and Hadamard-structured Riemannian constraints,

    ——, “Estimating dynamic graph flows with kernel models and Hadamard-structured Riemannian constraints,” in2025 Asia Pacific 20 Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) ,2025,pp.1538–1543

    2025

  22. [22]

    Riemannianoptimizationforhigh-dimensionaltensorcompletion,

    M.Steinlechner,“Riemannianoptimizationforhigh-dimensionaltensorcompletion,” SIAM J. Sci. Comput.,vol.38,no.5,2016

    2016

  23. [23]

    Taking the 4D nature of fMRI data into account promises significant gains in data completion,

    I. Belyaeva, S. Bhinge, Q. Long, and T. Adalı, “Taking the 4D nature of fMRI data into account promises significant gains in data completion,” IEEE Access,vol.9,pp.145334–145362,2021

    2021

  24. [24]

    LASSO, fractional norm and structured sparse estimation using a Hadamard product parametrization,

    P. D. Hoff, “LASSO, fractional norm and structured sparse estimation using a Hadamard product parametrization,”Comput. Stat. & Data Anal.,vol.115,pp.186–198,Nov.2017

    2017

  25. [25]

    The tail-Hadamard product parametrization algorithm for compressed sensing,

    G. Li, S. Li, D. Li, and C. Ma, “The tail-Hadamard product parametrization algorithm for compressed sensing,”Signal Process., vol. 205,p.108853,2023

    2023

  26. [26]

    SPRED:solving L1 penaltywithSGD,

    L.ZiyinandZ.Wang,“SPRED:solving L1 penaltywithSGD,”in Proc. ICML. Honolulu,HI:PMLR,Jul.2023

    2023

  27. [27]

    Smoothing the edges: Smooth optimization for sparse regularization using Hadamardoverparametrization,

    C. Kolb, C. L. Müller, B. Bischl, and D. Rügamer, “Smoothing the edges: Smooth optimization for sparse regularization using Hadamardoverparametrization,”arXiv:2307.03571v3[cs.LG],Apr.2024

    Pith/arXiv arXiv 2024

  28. [28]

    Kernel regression via tensor trains with Hadamard overparametrization and imputationofdynamicgraphedgeflows,

    D. T. Nguyen, K. Slavakis, E. Kofidis, and D. Pados, “Kernel regression via tensor trains with Hadamard overparametrization and imputationofdynamicgraphedgeflows,”in Proc. IEEE ICASSP,Barcelona,Spain,May2026

  29. [29]

    LearninggeneralGaussiankernelhyperparametersofSVMsusingoptimization onsymmetricpositive-definitematricesmanifold,

    H.Laanaya,F.Abdallah,H.Snoussi,andC.Richard,“LearninggeneralGaussiankernelhyperparametersofSVMsusingoptimization onsymmetricpositive-definitematricesmanifold,” Pattern Recogn. Lett.,vol.32,no.13,pp.1511–1515,2011

    2011

  30. [30]

    Tensordecompositionsandapplications,

    T.G.KoldaandB.W.Bader,“Tensordecompositionsandapplications,” SIAM Rev.,vol.51,no.3,pp.455–500,2009

    2009

  31. [31]

    Tensor networks for dimensionality reduction and large- scale optimization—Part 1: Low-rank tensor decompositions,

    A. Cichocki, N. Lee, I. Oseledets, A.-H. Phan, Q. Zhao, and D. P. Mandic, “Tensor networks for dimensionality reduction and large- scale optimization—Part 1: Low-rank tensor decompositions,”Foundations and Trends in Machine Learning , vol. 9, no. 4–5, pp. 249–429,2016

    2016

  32. [32]

    PrincetonUniversityPress,2008

    P.-A.Absil,R.Mahony,andR.Sepulchre, Optimization Algorithms on Matrix Manifolds. PrincetonUniversityPress,2008

    2008

  33. [33]

    Geometric methods on low-rank matrix and tensor manifolds,

    A. Uschmajew and B. Vandereycken, “Geometric methods on low-rank matrix and tensor manifolds,” inHandbook of Variational Methods for Nonlinear Geometric Data. Springer,2020,pp.261–313

    2020

  34. [34]

    Nonlinear dimensionality reduction by locally linear embedding,

    S. T. Roweis and L. K. Saul, “Nonlinear dimensionality reduction by locally linear embedding,”Science, vol. 290, no. 5500, pp. 2323–2326,2000

    2000

  35. [35]

    Sparsemanifoldclusteringandembedding,

    E.ElhamifarandR.Vidal,“Sparsemanifoldclusteringandembedding,” Advances in neural information processing systems,vol.24, 2011

    2011

  36. [36]

    Using the Nyström method to speed up kernel machines,

    C. Williams and M. Seeger, “Using the Nyström method to speed up kernel machines,”Advances in neural information processing systems,vol.13,2000

    2000

  37. [37]

    Sparse multidimensional scaling using landmark points,

    V. De Silva and J. B. Tenenbaum, “Sparse multidimensional scaling using landmark points,” Stanford University, Tech. Rep., Jun. 2004.[Online].Available: https://graphics.stanford.edu/courses/cs468-05-winter/Papers/Landmarks/Silva_landmarks5.pdf

    2004

  38. [38]

    Unconstrainedparametrizationsforvariance-covariancematrices,

    J.C.PinheiroandD.M.Bates,“Unconstrainedparametrizationsforvariance-covariancematrices,” Statist. Comput.,vol.6,no.3,pp. 289–296,1996

    1996

  39. [39]

    Springer,2006

    J.NocedalandS.J.Wright, Numerical optimization. Springer,2006

    2006

  40. [40]

    Low-rank matrix completion by Riemannian optimization,

    B. Vandereycken, “Low-rank matrix completion by Riemannian optimization,”SIAM Journal on Optimization , vol. 23, no. 2, pp. 1214–1236,2013

    2013

  41. [41]

    Tensorringdecomposition: Optimizationlandscapeandone-loopconvergenceofalternatingleastsquares,

    Z.Chen,Y.Li,andJ.Lu,“Tensorringdecomposition: Optimizationlandscapeandone-loopconvergenceofalternatingleastsquares,” SIAM J. Matrix Anal. Appl.,vol.41,no.3,pp.1416–1442,2020

    2020

  42. [42]

    Riemannian preconditioned algorithms for tensor completion via tensor ring decomposition,

    B. Gao, R. Peng, and Y.-X. Yuan, “Riemannian preconditioned algorithms for tensor completion via tensor ring decomposition,” Comput. Optim. Appl.,vol.88,no.2,pp.443–468,2024

    2024

  43. [43]

    Impact oftheresolutionofbrainparcelsonconnectome-wideassociationstudiesinfMRI,

    P.Bellec,Y.Benhajali,F.Carbonell,C.Dansereau,G.Albouy,M.Pelland,C.Craddock,O.Collignon,J.Doyon,E.Stip et al.,“Impact oftheresolutionofbrainparcelsonconnectome-wideassociationstudiesinfMRI,” Neuroimage,vol.123,pp.212–228,2015

    2015

  44. [44]

    BOLD5000, a public fMRI dataset while viewing 5000 visualimages,

    N. Chang, J. A. Pyles, A. Marcus, A. Gupta, M. J. Tarr, and E. M. Aminoff, “BOLD5000, a public fMRI dataset while viewing 5000 visualimages,” Sci. Data,vol.6,no.1,p.49,2019

    2019

  45. [45]

    fMRIPrep: arobustpreprocessingpipelineforfunctionalMRI,

    O. Esteban, C. J. Markiewicz, R. W. Blair, C. A. Moodie, A. I. Isik, A. Erramuzpe, J. D. Kent, M. Goncalves, E. DuPre, M. Snyder et al.,“fMRIPrep: arobustpreprocessingpipelineforfunctionalMRI,” Nature Meth.,vol.16,no.1,pp.111–116,2019

    2019

  46. [46]

    Data adaptive RKHS Tikhonov regularization for learning kernels in operators,

    F. Lu, Q. Lang, and Q. An, “Data adaptive RKHS Tikhonov regularization for learning kernels in operators,” inMathematical and Scientific Machine Learning. PMLR,2022,pp.158–172

    2022

  47. [47]

    CambridgeUniversityPress,2010

    P.J.Giblin, Graphs, Surfaces, and Homology,3rded. CambridgeUniversityPress,2010

    2010

  48. [48]

    HodgeLaplaciansongraphs,

    L.-H.Lim,“HodgeLaplaciansongraphs,” SIAM Rev.,vol.62,no.3,pp.685–715,2020

    2020

  49. [49]

    Signal processing on higher-order networks: Livin’ on the edge...andbeyond,

    M. T. Schaub, Y. Zhu, J.-B. Seby, T. M. Roddenberry, and S. Segarra, “Signal processing on higher-order networks: Livin’ on the edge...andbeyond,” Signal Process.,vol.187,Oct.2021

    2021

  50. [50]

    Topological signal processing over simplicial complexes,

    S. Barbarossa and S. Sardellitti, “Topological signal processing over simplicial complexes,”IEEE Trans. Signal Process., vol. 68, pp. 2992–3007,Mar.2020

    2020

  51. [51]

    Networksbeyondpairwiseinteractions: Structureanddynamics,

    F.Battiston,G.Cencetti,I.Iacopini,V.Latora,M.Lucas,A.Patania,J.-G.Young,andG.Petri,“Networksbeyondpairwiseinteractions: Structureanddynamics,” Phys. Rep.,vol.874,pp.1–92,2020

    2020

  52. [52]

    Simplicial vector autoregressive models,

    J. Krishnan, R. Money, B. Beferull-Lozano, and E. Isufi, “Simplicial vector autoregressive models,”IEEE Trans. Signal Process. , vol.72,pp.5454–5469,Nov.2024

    2024

  53. [53]

    Evolutionbackcastingofedgeflowsfrompartialobservationsusingsimplicial vectorautoregressivemodels,

    R.Money,J.Krishnan,B.Beferull-Lozano,andE.Isufi,“Evolutionbackcastingofedgeflowsfrompartialobservationsusingsimplicial vectorautoregressivemodels,”in Proc. IEEE ICASSP,Seoul,Korea,Apr.2024

    2024

  54. [54]

    Hodge-compositional edge Gaussian processes,

    M. Yang, V. Borovitskiy, and E. Isufi, “Hodge-compositional edge Gaussian processes,” inInternational Conference on Artificial Intelligence and Statistics. PMLR,2024,pp.3754–3762. 21

    2024

  55. [55]

    Simplicial neural networks,

    S. Ebli, M. Defferrard, and G. Spreemann, “Simplicial neural networks,” inTopological Data Analysis and Beyond workshop at NeurIPS,2020

    2020

  56. [56]

    Principledsimplicialneuralnetworksfortrajectoryprediction,

    T.M.Roddenberry,N.Glaze,andS.Segarra,“Principledsimplicialneuralnetworksfortrajectoryprediction,”in Proc. ICML. PMLR, 2021,pp.9020–9029

    2021

  57. [57]

    Simplicialconvolutionalneuralnetworks,

    M.Yang,E.Isufi,andG.Leus,“Simplicialconvolutionalneuralnetworks,”in Proc. IEEE ICASSP,2022,pp.8847–8851

    2022

  58. [58]

    Simplicial complex neural networks,

    H. Wu, A. Yip, J. Long, J. Zhang, and M. K. Ng, “Simplicial complex neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence,vol.46,no.1,pp.561–575,2024

    2024

  59. [59]

    Position: Topologicaldeeplearningisthenewfrontierforrelationallearning,

    T.Papamarkou,T.Birdal,M.M.Bronstein,G.E.Carlsson,J.Curry,Y.Gao,M.Hajij,R.Kwitt,P.Lio,P.DiLorenzo et al.,“Position: Topologicaldeeplearningisthenewfrontierforrelationallearning,”in Proc. ICML,Vienna,Austria,Jul.2024

    2024

  60. [60]

    Transportation networks for research,

    Transportation Networks for Research Core Team, “Transportation networks for research,” accessed: Nov. 23, 2024. [Online]. Available:https://github.com/bstabler/TransportationNetworks

    2024

  61. [61]

    UXsim: LightweightmesoscopictrafficflowsimulatorinpurePython,

    T.Seo,“UXsim: LightweightmesoscopictrafficflowsimulatorinpurePython,” J. Open Source Softw.,vol.10,no.106,Feb.2025

    2025

  62. [62]

    Signal processing on product spaces,

    T. M. Roddenberry, V. P. Grande, F. Frantzen, M. T. Schaub, and S. Segarra, “Signal processing on product spaces,” inProc. IEEE ICASSP,RhodesIsland,Greece,Jun.2023

    2023

  63. [63]

    HodgeNet: Graphneuralnetworksforedgedata,

    T.M.RoddenberryandS.Segarra,“HodgeNet: Graphneuralnetworksforedgedata,”in Proc. ACSSC,PacificGrove,CA,Nov.2019

    2019

  64. [64]

    Robusttensorringdecompositionforurbantrafficdataimputation,

    L.Yu,C.Guan,H.Wang,Y.He,W.Cao,andC.-S.Leung,“Robusttensorringdecompositionforurbantrafficdataimputation,” IEEE Trans. Intell. Transp. Syst.,vol.26,no.6,pp.8707–8719,Jun.2025

    2025

  65. [65]

    Tensordenoisingusinglow-ranktensortraindecomposition,

    X.Gong,W.Chen,J.Chen,andB.Ai,“Tensordenoisingusinglow-ranktensortraindecomposition,” IEEE Signal Processing Letters, vol.27,pp.1685–1689,2020

    2020

  66. [66]

    MosttensorproblemsareNP-hard,

    C.J.HillarandL.-H.Lim,“MosttensorproblemsareNP-hard,” Journal of the ACM (JACM),vol.60,no.6,pp.1–39,2013

    2013

  67. [67]

    Projection-likeretractionsonmatrixmanifolds,

    P.-A.AbsilandJ.Malick,“Projection-likeretractionsonmatrixmanifolds,” SIAM Journal on Optimization,vol.22,no.1,pp.135–158, 2012

    2012

  68. [68]

    Time integration of tensor trains,

    C. Lubich, I. V. Oseledets, and B. Vandereycken, “Time integration of tensor trains,”SIAM Journal on Numerical Analysis , vol. 53, no.2,pp.917–941,2015

    2015

  69. [69]

    quasi-optimal

    N.Boumal, An introduction to optimization on smooth manifolds. CambridgeUniversityPress,2023. A Background on Tensor Trains Fact 11([3,Thm.2.1]) . Given r := (1, r1, . . . , rN , 1)2 NN +1 ∗ ,any X2 RI1×···×IN hasaTTD A∗ X ,r. Computing A∗ X ,r is an NP-hard problem in general [66]. The TT singular value decomposition (TT-SVD) [3]—see Algo- rithm 1—comput...

    2023

  70. [70]

    Cν isderivedbythechainruleofdifferentiationasfollows ∂L ∂X (i1,

    TheEuclideangradientw.r.t. Cν isderivedbythechainruleofdifferentiationasfollows ∂L ∂X (i1, . . . , iN ) = PΩ(X Y)(i1, . . . , iN ) + ∂R ∂X (i1, . . . , iN ) = ∆(i1, . . . , iN ) , (15a) ∂X (i1, . . . , iN ) ∂Kν(j, j′) = Uν(i1, . . . , im, j)Vν(j′, im+1, . . . , iN ) , (15b) ∂L ∂Kν(j, j′) = X i1,...,iN ∂L ∂X (i1, . . . , iN ) ∂X (i1, . . . , iN ) ∂Kν(j, j′...

  71. [71]

    Therefore,theEuclideangradientw.r.t

    Notice that the mapUν,p7! (Uν,p Uν,̸=p) 1 Kν 1 Vν is a linear map w.r.t.Uν,p when Uν,̸=p, Vν, and Kν areconsideredfixed. Therefore,theEuclideangradientw.r.t. Uν,p is rUν,pL(Θ) =rUν,p(F +R)(Uν,p) + λUν,p (15m) = rX (F +R)(X ) m+1,...,N m+1,...,N Vν 1 Kν Uν,̸=p + λUν,p (15n) = ∆ m+1,...,N m+1,...,N Vν 1 Kν Uν,̸=p + λUν,p . (15o) Bysimilararguments,theEuclid...