Nonlocal Tensor Factorization

• Nonlocal Tensor Factorization is a paradigm that captures long-range dependencies by leveraging patch grouping, optimal-transport losses, and kernel-based regularization.
• These models employ techniques such as Tensor Train and CP decompositions to enhance tasks like tensor completion, image restoration, compressive sensing, and imputation.
• Efficient solvers like ADMM, block coordinate descent, and ALS are used to optimize the models, ensuring robust performance even with noisy or incomplete data.

to=arxiv_search 天天中彩票大奖json code 大发快三怎么看{"query":"\"Nonlocal Tensor Factorization\" tensor factorization nonlocal self-similarity tensor completion Wasserstein factorization arXiv", "max_results": 10} to=arxiv 天天好彩票 code : {"query":"nonlocal tensor factorization arXiv (Ding et al., 2020, Afshar et al., 2020, Zhang et al., 2018, Lei et al., 2024)","top_k":5} Nonlocal tensor factorization denotes a family of tensor models in which latent structure is not inferred solely from entrywise proximity or modewise smoothness, but from dependencies that connect distant patches, indices, or graph nodes. In the cited literature, nonlocality is realized in at least three distinct ways: by grouping similar patches or cubes into higher-order tensors and imposing low-rank structure; by replacing elementwise losses with optimal-transport losses defined through modewise ground metrics; and by regularizing tensor factors with covariance kernels that encode long-range correlations while reserving a separate residual process for local variation (Ding et al., 2020, Afshar et al., 2020, Zhang et al., 2018, Lei et al., 2024). The resulting models are used for tensor completion, image restoration, compressive sensing, sparse nonnegative tensor factorization, and multidimensional imputation.

1. Conceptual scope and meanings of nonlocality

In the patch-grouping literature, nonlocality refers to nonlocal self-similarity (NSS): small patches or cubes that recur across an image, video, or hyperspectral scene even when they are spatially distant. Grouping these similar patches produces a tensor whose additional “group” mode is content-adaptive rather than spatially local, and low-rank structure is then imposed on that tensor (Ding et al., 2020, Zhang et al., 2018).

In the optimal-transport formulation of SWIFT, nonlocality does not arise from explicit patch grouping. Instead, it is induced by modewise ground cost matrices $C_n$ that allow mass to move between similar but distinct indices. This makes the loss nonlocal because discrepancies are propagated along the geometry encoded by $C_n$, rather than being penalized only entrywise (Afshar et al., 2020).

In GLSKF, nonlocality is encoded by covariance kernels $K_d^u$ on factor matrices. Large length scales or graph-diffusion kernels create non-negligible covariance between distant indices along each mode, so the low-rank component captures long-range, low-frequency structure, while a separate residual process models short-scale variation (Lei et al., 2024).

Formulation	Tensor model	Mechanism of nonlocality
NSS-grouped completion	TT or CP on grouped patches/cubes	Stacking similar patches/cubes into a group tensor
Wasserstein factorization	Nonnegative CP	Transport across indices via $C_n$
Kernelized completion	CP plus residual process	Kernel/covariance coupling between distant indices

A common misconception is that nonlocal tensor factorization is synonymous with patch-based image denoising. The cited work suggests a broader interpretation: nonlocality may be patch recurrence, transport over mode geometry, or kernelized long-range covariance, depending on the modeling assumptions.

2. Patch-grouped low-rank tensor models

For visual tensor completion, the TT-based model of “Tensor train rank minimization with nonlocal self-similarity for tensor completion” forms higher-order group tensors by stacking similar cubes. Given a $d$th-order tensor $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$, its TT representation is

$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$

with TT cores $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$, $r_0=r_d=1$, and TT ranks $r_k=\operatorname{rank}(X_{[k]})$. The convex surrogate used for grouped completion is

$C_n$0

For a color image, overlapping cubes of size $C_n$1 are extracted, similar cubes are found by BM3D-style block matching with Euclidean distance, and the $C_n$2 most similar cubes are stacked into $C_n$3. After groupwise completion, overlapping estimates are mapped back and averaged pixel-by-pixel. The paper explicitly contrasts this with ket augmentation (KA), reporting that canonical matricizations of KA-augmented tensors have an average ratio of singular values larger than $C_n$4 of the maximum of approximately $C_n$5, whereas NSS-grouped tensors show approximately $C_n$6, which is presented as evidence of stronger TT low-rankness and reduced block artifacts (Ding et al., 2020).

The CP-based NLR-TFA model constructs a 3D tensor from grouped image patches rather than cubes. Similar $C_n$7 patches are found by $C_n$8-nearest-neighbor search in Euclidean distance and stacked as

$C_n$9

The tensor is then factorized by a CP model,

$K_d^u$0

and only the $K_d^u$1 most significant components are retained,

$K_d^u$2

The paper emphasizes that this avoids vectorizing patches into a 2D matrix and therefore preserves within-patch spatial structure. Typical settings are $K_d^u$3, $K_d^u$4, and $K_d^u$5 (Zhang et al., 2018).

Taken together, these two patch-group formulations show that “nonlocal” does not specify a unique factorization family. One line uses TT rank on higher-order grouped tensors; another uses CP factor truncation on 3D grouped tensors. The common principle is that similarity-driven grouping creates a tensor in which multiway low-rank structure is more explicit than in the original observation domain.

3. Distribution-agnostic and kernelized formulations

SWIFT replaces distribution-specific reconstruction losses with a tensor Wasserstein loss defined over CP reconstructions of sparse nonnegative tensors. Its nonnegative CP model is

$K_d^u$6

with $K_d^u$7. The modewise Wasserstein construction begins with entropy-regularized vector transport,

$K_d^u$8

extends it to matrices, and then defines the tensor distance

$K_d^u$9

The final objective combines OT terms with KL soft constraints:

$C_n$0

This formulation is explicitly distribution-agnostic, assumes nonnegative input, and uses $C_n$1 to exploit cross-dimensional correlation in a way that is not restricted to Gaussian side-information models (Afshar et al., 2020).

GLSKF places nonlocality in the factor priors rather than in the loss. The global component is a CP tensor

$C_n$2

while the completed tensor is modeled through a global low-rank term plus a local residual process. The core objective is

$C_n$3

Here the covariance norm

$C_n$4

is used both for factor smoothness and for the residual process, with a separable residual covariance

$C_n$5

The paper uses Matérn $C_n$6 kernels for temporal and pixel modes, graph-regularized Laplacian kernels for graph modes, Bohman tapering for locality in residual kernels, and alternatively sparse GMRF precisions (Lei et al., 2024).

These two formulations expand the meaning of nonlocal tensor factorization beyond grouped images. SWIFT is nonlocal because its discrepancy measure transports mass along modewise geometry; GLSKF is nonlocal because distant indices are linked by kernel covariance in the factor space.

4. Optimization architectures

The TT-NSS completion model is solved by ADMM with auxiliary tensors $C_n$7. The $C_n$8 updates apply singular value thresholding to each canonical TT unfolding,

$C_n$9

while the $d$0 update is a projection onto the observation set,

$d$1

Stopping uses the relative change criterion $d$2 or $d$3. The dominant cost is the SVD of canonical unfoldings, giving an approximate per-iteration cost of

$d$4

for $d$5 groups (Ding et al., 2020).

NLR-TFA also uses ADMM, but for a compressive sensing objective with auxiliary variable $d$6:

$d$7

Its $d$8-update is a diagonal quadratic solve because $d$9 counts patch overlaps, and the expensive $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$0-update is handled either by a Fourier-domain closed form under partial Fourier sensing or by a learned convolutional module $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$1 that approximates $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$2 in the Sherman–Morrison–Woodbury reduction. The paper states that the learned module is not a plug-and-play denoiser nor an unrolled ADMM net, but a reusable inverse-approximation block for the $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$3-update (Zhang et al., 2018).

SWIFT is optimized by block coordinate descent. For transport updates it uses Sinkhorn-like fixed-point iterations without explicitly forming transport matrices:

$X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$4

with

$X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$5

where $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$6. Factor matrices are then updated multiplicatively. The paper emphasizes sparsity exploitation through the number of nonzero columns $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$7, symbolic use of $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$8, and parallel OT subproblems (Afshar et al., 2020).

GLSKF uses ALS with closed-form linear systems for each factor block and CG or PCG for scalable solution. For mode $X \in \mathbb{R}^{n_1 \times \cdots \times n_d}$9, the update solves

$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$0

and the residual is updated either in precision form,

$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$1

or in covariance form with the “imaginary” observation construction that preserves Kronecker structure (Lei et al., 2024).

Across these models, the algorithmic pattern is consistent: nonlocal structure is first encoded in the model class, and efficient solvers are then designed to exploit the resulting algebraic structure—TT unfoldings, small grouped tensors, sparse OT plans, or Kronecker products.

5. Theoretical properties and interpretation

The TT-NSS paper develops explicit perturbation analysis for grouped tensors. Its Theorem 1 gives a rank-$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$2 approximation from the average column fiber with error bound $$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$3; Theorem 2 gives a rank-$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$4 approximation from the average row-cube with $$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$5; Theorem 3 gives a rank-$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$6 approximation by retaining the first $$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$7 cubes and duplicating one cube for the remainder; and Theorem 4 states TT nuclear-norm stability,

$$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$8

The stated interpretation is that grouped tensors can be well approximated by low-TT-rank tensors and that TT nuclear norm is robust to small perturbations (Ding et al., 2020).

SWIFT provides two theoretical statements of a different kind. First, it is described as a Block Coordinate Descent method that guarantees convergence to a stationary point. Second, the paper states that the tensor Wasserstein distance $$X(i_1,\dots,i_d)=G_1(:,i_1,:)G_2(:,i_2:)\cdots G_d(:,i_d,:),$$9 is a valid metric under standard assumptions on the per-mode cost matrices, with proof in the appendix (Afshar et al., 2020).

GLSKF supplies an additional interpretive bridge by showing that covariance-norm regularization subsumes familiar smoothness penalties: choosing $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$0 recovers QV regularization as a special case. It also introduces the imaginary-observation trick, in which missing entries are augmented with infinitesimal-precision observations so that GLS updates preserve Kronecker structure (Lei et al., 2024).

These results clarify that nonlocal tensor factorization is not defined by a single proof technique or rank notion. TT perturbation bounds, Wasserstein metric properties, and kernel-norm interpretations all serve the same broader purpose: justifying why information from distant but correlated configurations can stabilize inference under incompleteness or corruption.

6. Empirical behavior, applications, and limitations

The cited models are evaluated on different tasks and therefore emphasize different performance criteria. In grouped TT completion, experiments cover color images, multispectral images, and videos with PSNR and SSIM. Representative results at image random sampling with $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$1 include Lena at $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$2 dB and SSIM $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$3 for NL-TT, compared with $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$4 dB and $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$5 for TMac-TT, and $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$6 dB and $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$7 for SiLRTC-TT. Under structured missingness, House with curves reaches $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$8 dB and $G_k \in \mathbb{R}^{r_{k-1}\times n_k \times r_k}$9, and Peppers with random blocks reaches $r_0=r_d=1$0 dB and $r_0=r_d=1$1. On the CAVE multispectral dataset, the paper states that across 32 MSIs, NL-TT held the best PSNR consistently; for videos, per-frame PSNR/SSIM curves are reported to be highest across frames (Ding et al., 2020).

NLR-TFA is evaluated for compressive sensing restoration on eight standard images, under pseudo-radial Fourier sampling and random Gaussian sensing, using PSNR and SSIM. At $r_0=r_d=1$2, the average over eight images is $r_0=r_d=1$3 dB and $r_0=r_d=1$4 SSIM for NLR-TFA, compared with $r_0=r_d=1$5 dB and $r_0=r_d=1$6 for BM3D-CS, $r_0=r_d=1$7 dB and $r_0=r_d=1$8 for TVAL3, $r_0=r_d=1$9 dB and $r_k=\operatorname{rank}(X_{[k]})$0 for NLR-CS, and $r_k=\operatorname{rank}(X_{[k]})$1 dB and $r_k=\operatorname{rank}(X_{[k]})$2 for D-AMP. The paper further states that average PSNR drops only about $r_k=\operatorname{rank}(X_{[k]})$3 dB as $r_k=\operatorname{rank}(X_{[k]})$4 decreases from $r_k=\operatorname{rank}(X_{[k]})$5 to $r_k=\operatorname{rank}(X_{[k]})$6, whereas NLR-CS and D-AMP drop about $r_k=\operatorname{rank}(X_{[k]})$7 dB and $r_k=\operatorname{rank}(X_{[k]})$8 dB, respectively (Zhang et al., 2018).

SWIFT is assessed on sparse nonnegative tensors for downstream prediction rather than image fidelity. On BBC News, downstream classification accuracy improves by $r_k=\operatorname{rank}(X_{[k]})$9 to $C_n$00 relative over baselines; on Sutter EHR, PR-AUC improves by $C_n$01 to $C_n$02. Under noisy conditions, the reported gains reach up to $C_n$03 relative improvement on BBC and up to $C_n$04 on Sutter. The paper also reports speedups over a direct dense Wasserstein tensor factorization implementation of up to $C_n$05 for OT computation, up to $C_n$06 overall, and $C_n$07 in appendix experiments (Afshar et al., 2020).

GLSKF is evaluated on traffic speed imputation, color image inpainting, color video completion, and MRI completion. Highlights include PeMS at $C_n$08 with MAE $C_n$09 and RMSE $C_n$10, Airplane at $C_n$11 with PSNR $C_n$12, Hall video at $C_n$13 with PSNR $C_n$14, and MRI at $C_n$15 with PSNR $C_n$16. The ablations LSTF, LSKF, and GLSlocal are used to argue that global nonlocal regularization and local residual modeling play complementary roles (Lei et al., 2024).

The limitations reported across the literature are similarly diverse. TT-NSS requires sufficiently strong NSS, can be degraded by heavy noise in block matching, and may incur significant runtime and memory costs for large $C_n$17 or large data volumes (Ding et al., 2020). NLR-TFA depends on accurate patch grouping, is sensitive to retained rank $C_n$18, and requires retraining of the learned inverse when $C_n$19 or $C_n$20 changes (Zhang et al., 2018). SWIFT assumes nonnegative tensors and depends on the quality of the ground metric $C_n$21 (Afshar et al., 2020). GLSKF depends on kernel choice, remains nonconvex despite monotone ALS descent, and uses empirically set hyperparameters such as $C_n$22, $C_n$23, $C_n$24, and taper ranges (Lei et al., 2024).

A plausible implication is that the term “nonlocal tensor factorization” should be understood as a modeling principle rather than a single algorithmic template. The common claim across these works is that explicit mechanisms for long-range dependence—patch recurrence, Wasserstein transport, or kernel covariance—can materially improve tensor reconstruction or representation when entrywise or purely local models are inadequate.