Tensor Ring (TR) Decomposition

• Tensor Ring (TR) is a cyclic tensor network decomposition that factorizes high-order tensors into a ring of interconnected 3rd-order cores, providing circular invariance and flexible latent representation.
• It underpins various algorithms such as TR-SVD, TR-ALS, and randomized methods to efficiently perform tensor completion, compression, and robust recovery with linear storage scaling.
• TR's versatility has driven applications in hyperspectral imaging, deep learning dataset compression, and continuous functional representations, linking theoretical advances with practical solutions.

Searching arXiv for tensor ring decomposition papers to ground the article in the primary literature. Tensor Ring (TR) is a tensor network decomposition that represents a high-order tensor by a cyclic contraction of third-order core tensors. For an order-$N$ tensor $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$, the standard elementwise form is

$$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$

where $\mathbf{G}_n(i_n)$ is the $i_n$-th mode-2 slice of the core $$\mathcal{G}_n \in \mathbb{R}^{R_n \times I_n \times R_{n+1}}$$, with $R_{N+1}=R_1$. Introduced as a circular generalization of tensor train (TT), TR removes the TT boundary constraints, treats all latent cores equivalently, and achieves circular dimensional permutation invariance through the trace operator (Zhao et al., 2016). Subsequent work has developed TR into a broad framework for low-rank modeling, compression, completion, robust recovery, hyperspectral fusion, multiway representation learning, and more recently deterministic exact recovery and continuous functional representations (Zhao et al., 2017, Chen et al., 30 Nov 2025, Xu et al., 1 Mar 2026).

1. Formal definition and representation

The canonical TR model factorizes a $d$-way tensor into a ring of order-3 cores. In index form,

$$\mathcal{T}(i_1,\ldots,i_d)=\sum_{\alpha_1,\ldots,\alpha_d} Z_1(\alpha_1,i_1,\alpha_2)\cdots Z_d(\alpha_d,i_d,\alpha_1),$$

so the last latent index reconnects to the first and closes the network (Zhao et al., 2016). This cyclic closure distinguishes TR from chain-structured tensor networks and yields a graphical interpretation as a cyclic interconnection of $d$ third-order tensors.

The latent dimensions $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$0 are the TR-ranks. Under uniform mode size $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$1 and uniform rank $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$2, the storage complexity is $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$3, so the number of parameters grows linearly with tensor order rather than exponentially (Zhao et al., 2016). This linear scaling is one of the reasons TR is used for high-order tensors arising in images, videos, hyperspectral data, and tensorized machine-learning datasets.

An equivalent outer-product view expresses the tensor as a sum over mode-2 fibers of the cores,

$$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$4

which emphasizes that TR is a multilinear low-rank representation with cyclic latent coupling (Zhao et al., 2017). This form is frequently used in derivations of approximation algorithms and in relating TR to other tensor decompositions.

2. Structural properties and relations to other tensor models

A defining property of TR is circular dimensional permutation invariance. If the modes are circularly shifted, the TR representation is recovered by rotating the core sequence correspondingly, owing to the cyclic property of the trace (Zhao et al., 2016). A common misunderstanding is to interpret this as unrestricted invariance to arbitrary mode permutations. The property established in the foundational TR literature is circular or cyclic invariance, not an unrestricted invariance over all permutations.

TR is a strict generalization of TT. If the TT boundary condition $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$5 is imposed, TR reduces to TT; without that restriction, all TR cores remain third-order and the model has greater flexibility in allocating latent dimensions across modes (Zhao et al., 2017). The original TR paper further states that a general TR can be viewed as a linear combination of TT decompositions, which explains its stronger representation ability under comparable parameter budgets (Zhao et al., 2016).

Relations to other decompositions are also well documented. CP appears as a special case when all TR cores are diagonal, while Tucker can be absorbed into TR by representing the Tucker core itself in TR format and folding factor matrices into the multilinear contractions (Zhao et al., 2016). In applications, this positioning places TR between highly constrained rank-one expansions such as CP and core-expansive formats such as Tucker.

Basic multilinear operations can be performed directly in TR format. The literature gives efficient constructions for addition via block-diagonal stacking of cores, Hadamard product via Kronecker products of slices, multilinear products, and inner-product or Frobenius-norm evaluation without materializing the full tensor (Zhao et al., 2017). These closure properties are important for optimization algorithms and for downstream tasks such as feature extraction and compression.

3. Algorithms for learning TR decompositions

The initial algorithmic toolkit for TR consisted of four families: TR-SVD, TR-ALS, TR-ALSAR, and TR-BALS (Zhao et al., 2016). TR-SVD is a sequential truncated-SVD procedure analogous to TT-SVD, with truncation thresholds chosen to control the global relative error. TR-ALS alternates least-squares updates of single cores but requires the TR-ranks to be specified in advance. TR-ALSAR augments ALS with adaptive rank growth, and TR-BALS merges two adjacent cores, solves a local least-squares problem, then splits the block by truncated SVD so that intermediate ranks are selected adaptively.

The later paper on ring-structure networks sharpened the practical role of TR-SVD and TR-BALS. TR-SVD was emphasized as a non-iterative rank-adaptive initializer, while TR-BALS was described as a cyclic blockwise ALS method that refines the representation and adjusts ranks through repeated low-rank separations of merged core pairs (Zhao et al., 2017). Both papers note that different starting modes may lead to different TR-rank configurations.

Large-scale settings motivated randomized and sketching-based methods. Randomized Tensor Ring Decomposition applies tensor random projection mode by mode: $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$6 then decomposes the smaller tensor $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$7 using ALS or SVD-based TR routines (Yuan et al., 2019). The paper reports that the proposed algorithms are $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$8 times faster than traditional algorithms without loss of accuracy, and reports superior performance in deep learning dataset compression and hyperspectral image reconstruction experiments compared to other randomized algorithms (Yuan et al., 2019).

A second randomized line replaces dense projections by sketching inside ALS subproblems. Practical Sketching-Based Randomized Tensor Ring Decomposition introduces KSRFT- and TensorSketch-based algorithms that use new tensor products to avoid explicit construction of huge coefficient matrices, and provides theoretical sketch-size and complexity analyses (Yu et al., 2022). Block-Randomized Stochastic Methods for Tensor Ring Decomposition moves further toward lightweight updates by combining randomized block coordinate descent with mini-batch stochastic gradient descent; its scaled variant acts as an adaptive preconditioned or diagonally scaled method for ill-conditioned problems (Yu et al., 2023). For streaming data, Tracking Tensor Ring Decompositions of Streaming Tensors develops deterministic and randomized update rules that reuse TR structure and auxiliary quantities so that updates are much faster than repeated batch recomputation while maintaining similar accuracy (Yu et al., 2023).

4. Completion, low-rank regularization, and recovery theory

Tensor completion has been one of the most active TR application areas. TR-WOPT formulates completion as

$$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$9

where the binary tensor $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$0 masks observed entries and missing values are inferred from the low-TR-rank reconstruction (Yuan et al., 2018). The paper states that TR-WOPT performs well in various high-dimension tensors and that, especially when the missing rate of the test images is high, such as over $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$1, its image completion performance is significantly better than the compared algorithms (Yuan et al., 2018).

A separate response to the TR-rank selection problem is to regularize the latent space rather than the data space. Tensor Ring Decomposition with Rank Minimization on Latent Space imposes nuclear norms on the unfoldings of the TR cores, performs the SVD steps on much smaller matrices, and uses ADMM to infer low-rank latent factors and the completed tensor simultaneously (Yuan et al., 2018). The stated motivation is robustness to model selection, since the initial TR-rank acts only as an upper bound.

Convex surrogates based on circular unfoldings form another major line. Tensor-Ring Nuclear Norm Minimization defines a tensor circular unfolding $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$2 and establishes

$$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$3

thereby connecting matrix ranks of circular unfoldings to TR-ranks (Yu et al., 2019). Completion is then posed as minimizing a weighted sum of nuclear norms of these unfoldings subject to observed-entry constraints, producing a convex alternative to non-convex factor updates (Yu et al., 2019). Tensor completion via nonconvex tensor ring rank minimization with guaranteed convergence replaces the nuclear norm by a logdet-based smooth nonconvex surrogate that penalizes small singular values more aggressively and proves convergence of the resulting ADMM scheme to a stationary point under mild assumptions (Ding et al., 2020).

Theoretical recovery guarantees for TR completion were established in Provable Tensor Ring Completion. Under a strong TR incoherence condition analogous to matrix incoherence, a $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$4-order tensor of dimensional size $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$5 and TR rank $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$6 can be exactly recovered with high probability by solving a convex optimization program, given $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$7 samples (Huang et al., 2019). The convex program minimizes a weighted sum of nuclear norms of balanced TR unfoldings, and the paper identifies balanced unfoldings as critical for reducing sample complexity (Huang et al., 2019). Coupled Tensor Completion extends the factor-sharing idea to multiple tensors, enforces equality constraints on shared TR factors, and derives an excess risk bound showing theoretical performance enhancement compared with coupled nuclear norm based methods (Huang et al., 2020).

5. Bayesian and robust formulations

A limitation of many deterministic TR methods is the need to preassign ranks. Bayesian Low Rank Tensor Ring Model for Image Completion addresses this by placing hierarchical Gaussian-Gamma priors on TR core entries so that the induced marginal is Student-T and redundant components shrink automatically (Long et al., 2020). Variational Bayesian inference updates posterior distributions over cores, precision variables, and noise precision, and the effective TR-ranks are obtained by pruning components whose precisions become large (Long et al., 2020). The paper emphasizes that the method is free of parameter-tuning and that the TR-ranks can be obtained by Bayesian inference.

Bayesian Robust Tensor Ring Model for Incomplete Multiway Data extends this idea to robust tensor completion with sparse corruptions and dense Gaussian noise. The observed tensor is modeled as

$$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$8

where $$\mathcal{X}(i_1,\ldots,i_N)=\operatorname{Tr}\left\{\prod_{n=1}^{N}\mathbf{G}_n(i_n)\right\},$$9 is low-rank and represented by TR decomposition, $\mathbf{G}_n(i_n)$0 is sparse corruption, and $\mathbf{G}_n(i_n)$1 is dense Gaussian noise (Huang et al., 2022). Slice-wise automatic relevance determination priors on the core tensors again drive unnecessary TR components to zero, and variational Bayes jointly infers low-rank structure, sparse corruption, and noise levels without manual selection of TR-rank or penalty parameters (Huang et al., 2022).

Robustness to outliers has also been addressed from an optimization rather than Bayesian perspective. Scalable and Robust Tensor Ring Decomposition for Large-scale Data proposes an auto-weighted correntropy objective and converts it by half-quadratic analysis into an auto-weighted least-squares problem over TR cores and weights (He et al., 2023). The same paper introduces fast Gram matrix computation, reducing the Gram-matrix bottleneck from $\mathbf{G}_n(i_n)$2 time and $\mathbf{G}_n(i_n)$3 storage to $\mathbf{G}_n(i_n)$4 time and $\mathbf{G}_n(i_n)$5 storage, and combines it with randomized subtensor sketching so that large tensors with missing entries and gross corruptions can be handled efficiently (He et al., 2023).

6. Applications and specialized variants

TR has been used in a wide range of domain-specific recovery and representation tasks. In hyperspectral image fusion, Hyperspectral Image Fusion via Logarithmic Low-rank Tensor Ring Decomposition studies the low-rankness of TR factors from the tensor nuclear norm perspective and regularizes each TR factor by a mode-2 logarithmic tensor nuclear norm together with weighted total variation (Zhang et al., 2023). The resulting model is optimized by proximal alternating minimization with ADMM-based subproblem solvers, and the paper reports improvements over CSTF, CTRF, and FSTRD on Pavia University and CAVE benchmarks across PSNR, SSIM, ERGAS, SAM, and UIQI (Zhang et al., 2023).

For nonnegative multiway representation learning, Graph Regularized Nonnegative Tensor Ring Decomposition introduces NTR and GNTR. NTR imposes nonnegativity on all cores to obtain parts-based basis elements, while GNTR adds a graph Laplacian penalty on the last core to preserve manifold geometry (Yu et al., 2020). Both are optimized by accelerated proximal gradient updates and were reported to yield better performance than state-of-the-art tensor based methods in clustering and classification tasks, while the basis visualizations showed richer colors and rich lines (Yu et al., 2020).

TR has also been extended beyond real-valued tensors. Quaternion tensor left ring decomposition generalizes TR to quaternion-valued cores for color image processing, replacing ordinary matrix multiplication by left quaternion matrix multiplication and redefining the cyclic property to respect quaternion non-commutativity (Miao et al., 2023). The associated low-rank quaternion tensor completion model uses weighted nuclear norms of quaternion circular unfoldings and ADMM-type optimization for color image inpainting (Miao et al., 2023).

Several application papers also use TR as a compression primitive. Randomized Tensor Ring Decomposition reports strong performance on deep learning dataset compression and hyperspectral image reconstruction (Yuan et al., 2019). The original and early TR papers reported compact feature extraction for COIL-100 object images and KTH action videos, with higher compression or better classification accuracy than TT or CP at comparable reconstruction errors (Zhao et al., 2016, Zhao et al., 2017).

7. Recent theory and emerging extensions

A major recent theoretical development is A Provably Efficient Method for Tensor Ring Decomposition and Its Applications, which presents the first deterministic, finite-step algorithm for exact TR decomposition under appropriate identifiability conditions (Chen et al., 30 Nov 2025). The method, BLOSTR, uses blockwise simultaneous diagonalization to recover TR cores from a limited number of tensor observations. The paper states that only $\mathbf{G}_n(i_n)$6 entries are required in the general setting and $\mathbf{G}_n(i_n)$7 entries in the symmetric TR setting, matching the degrees of freedom up to constants (Chen et al., 30 Nov 2025). For noisy observations, the paper combines its spectral initialization with ALS and reports faster convergence and improved accuracy compared to classic methods. It further identifies applications in matrix product state tomography and provable learning of pushforward distributions (Chen et al., 30 Nov 2025).

Another recent direction moves TR beyond discrete meshgrids. Reparameterized Tensor Ring Functional Decomposition for Multi-Dimensional Data Recovery replaces discrete TR factors by implicit neural representations, so the factor functions can be evaluated on meshgrid and non-meshgrid coordinates (Xu et al., 1 Mar 2026). The paper analyzes the spectral structure of TR factors, proves that factor bandwidth limits the output bandwidth, and argues that standard INR-based TR functional decomposition has limited high-frequency modeling capacity (Xu et al., 1 Mar 2026). Its reparameterized model expresses each factor as a structured combination of a learnable latent tensor and a fixed basis,

$\mathbf{G}_n(i_n)$8

and derives a variance-preserving initialization together with a global Lipschitz continuity result for the resulting coordinate-to-tensor map (Xu et al., 1 Mar 2026). The reported experiments cover image inpainting, denoising, super-resolution, and point cloud recovery, indicating that TR methodology is moving from discrete tensor compression toward continuous data representations.

These developments suggest two parallel trajectories for TR research. One is algebraic and algorithmic, centered on identifiability, exact recovery, and efficient solvers; the other is modeling-oriented, where TR is combined with probabilistic priors, graph structure, quaternion algebra, and neural function parameterizations. A plausible implication is that TR now functions less as a single decomposition algorithm than as a general cyclic low-rank prior whose concrete realization depends on the application regime.