Tensor Ring Decomposition

• Tensor Ring (TR) decomposition is a cyclic tensor factorization that represents an Nth-order tensor as interconnected third-order cores with circular invariance.
• It generalizes tensor-train decomposition by removing boundary constraints, thereby enhancing expressiveness and enabling linear parameter scaling.
• Efficient algorithms—including ALS variants and randomized methods—facilitate robust TR decomposition for large-scale and high-dimensional applications.

A tensor ring (TR) decomposition is a cyclic tensor network factorization representing an Nth-order tensor as a sequence of interconnected third-order core tensors, where the connections form a closed loop. Each entry of the tensor is computed by tracing the cyclic product of selected slices from the cores. The model generalizes and strictly contains the tensor-train (TT) decomposition by eliminating the open-boundary rank constraint and enabling full permutation symmetry under cyclic mode shifts. TR exhibits linear parameter scaling with respect to tensor order while maintaining high expressiveness and flexibility for approximating high-dimensional data.

1. Mathematical Formulation and Properties

Given a tensor $$\mathcal{X} \in \mathbb{R}^{I_1 \times \cdots \times I_N}$$, the TR decomposition posits third-order cores $$\{ \mathcal{G}^{(n)} \in \mathbb{R}^{R_n \times I_n \times R_{n+1}} \}_{n=1}^N$$ such that $R_{N+1} = R_1$. For each entry,

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

where $\mathcal{G}^{(n)}(:, i_n, :)$ denotes the $i_n$th lateral slice (an $R_n \times R_{n+1}$ matrix) of the $n$th core. The minimal set of integers $(R_1, \ldots, R_N)$ for which the above holds exactly is termed the TR-rank. The total number of parameters is $\sum_{n=1}^N R_n I_n R_{n+1}$, scaling linearly in $N$ for bounded $R_n, I_n$.

Salient structural properties include:

• Circular permutation invariance: Any cyclic reordering of core tensors yields the same reconstructed tensor, i.e., TR is invariant under circular shifts of modes (Zhao et al., 2016, Zhao et al., 2017).
• Generalization of TT: TR removes the boundary constraints of TT ($R_1=R_{N+1}=1$), allowing greater rank flexibility and expressivity.
• Gauge non-uniqueness: The representation is invariant under insertion of invertible $A_k$ ($R_k \times R_k$) matrices (gauge transformations), with $(U_1, \dots, U_N)$ identified to $(U_1 A_1, A_1^{-1}U_2A_2, \dots, A_N^{-1}U_NA_1)$ (Gao et al., 29 Jan 2026).

TR admits a trace-of-product algebraic structure, supporting efficient addition, inner/outer products, and mode contractions directly in the TR domain with $O(N I R^2 + N R^3)$ complexity (Zhao et al., 2016, Zhao et al., 2017).

2. Algorithmic Methods for TR Decomposition

2.1 Alternating Least Squares (ALS) and Variants

The canonical approach is block coordinate descent—optimizing each core with others fixed—yielding a least-squares subproblem per core. For core $n$: $$\min_{\mathcal{G}^{(n)}_{(2)}} \| X_{[n]} - G_{(2)}^{(n)} (G_{[2]}^{\neq n})^T \|_F^2$$ where $X_{[n]}$ is the mode-$n$ unfolding, $G_{(2)}^{(n)}$ the 2-unfolding of core $n$, and $G_{[2]}^{\neq n}$ the merged subchain of all other cores (Zhao et al., 2016, Yu et al., 2022).

Advanced ALS-based schemes include:

• Blockwise ALS with Adaptive Rank (TR-BALS): Merges and splits adjacent cores using SVD, adapting individual TR-ranks per core (Zhao et al., 2017, Zhao et al., 2016).
• ALS with Gram/QR Stabilization: Algebraic simplification of coefficient matrices and QR decomposition of the subchain tensors mitigate intermediate data explosion and improve numerical stability, leading to scalable implementations (TR-ALS-NE, TR-ALS-QR) (Yu et al., 2022).

2.2 Randomized, Streaming, and Sketching Techniques

Scalability is achieved through randomized projections, leverage-score sampling, and sketching:

• Randomized TR via Tensor Projections: All modes are projected to lower dimensions via random matrices or KSRFT, and TR decomposition is performed on the reduced tensor, reducing per-iteration cost from $O(N I^N R^2)$ to $O(N K^N R^2)$ (Yuan et al., 2019).
• Sampling-Based ALS: Sketches the subchain and data using leverage-score or uniform sampling, providing (with high probability) $(1+\epsilon)$-approximate updates at sublinear cost in tensor size (Malik et al., 2020, Yu et al., 2022).
• Block-Randomized SGD: Combines randomized block coordinate descent with stochastic gradients computed on minibatches of fibers, enabling lightweight, memory-efficient updates and robust behavior under poor conditioning (Yu et al., 2023).
• Streaming and Online Extensions: Efficiently update the TR decomposition in real time as new slices arrive, avoiding recomputation from scratch and supporting arbitrary sketching types (Yu et al., 2023).

Empirical studies demonstrate that these strategies can speed up TR fitting by 10–100$\times$ with negligible loss in accuracy for large-scale and streaming tensors.

3. Theoretical Guarantees and Intrinsic Geometry

TR decomposition admits substantial theoretical analysis:

• Quotient Geometry: With full-rank (injectivity) conditions on each core's 2-unfolding, and modding out by the projective general linear group action, the space of TR factorizations can be described as a smooth quotient manifold $M_r^*$ of dimension $\sum_{k=1}^d (r_k n_k r_{k+1} - r_k^2) + 1$ (Gao et al., 29 Jan 2026). This geometry underpins Riemannian optimization for recovery and completion.
• Identifiability and Non-Uniqueness: Apart from cyclic gauge freedom, TR cores are essentially unique under mild genericity conditions (Gao et al., 29 Jan 2026, Zhao et al., 2016).
• Finite-Step Algebraic Recovery: There exist deterministic, finite-step algorithms (BLOSTR) that recover generic TR cores exactly (up to gauge) from only $O(r^2 \sum_k n_k)$ sampled entries, matching the parameter count. This applies to general and symmetric TR tensors, and extends to robust, noise-tolerant variants coupled with ALS refinement (Chen et al., 30 Nov 2025).
• Failure of Best Low-Rank Approximation: TR lacks a guaranteed minimal-rank truncation after operations such as contraction and Hadamard product. No stable best low-rank TR approximation exists in general, and “ring rounding” can fail to reduce ranks or recover minimal representations, sharply contrasting with the TT setting (Batselier, 2018).

4. Extensions: Constraints and Regularization

The flexibility of the TR ansatz enables principled incorporation of structure and regularization:

• Nonnegative TR (NTR) and Graph-Regularized NTR (GNTR): Imposing nonnegativity on cores leads to interpretable, parts-based basis extraction and improved clustering/classification accuracy. Graph Laplacian regularization (GNTR) further aligns representations with data manifold geometry. Both are optimized efficiently with accelerated proximal gradient (APG) (Yu et al., 2020).
• Latent-Space Rank Minimization: Imposing nuclear norms on core unfoldings rather than the full tensor relieves model selection burden and mitigates overfitting. ADMM-based schemes allow simultaneous learning of core ranks and reconstruction, with all SVDs performed on small matrices (Yuan et al., 2018).
• Robust and Scalable TR: Methods based on correntropy-induced loss, fast Gram matrix computation (FGMC), and randomized subtensor sketching (RStS) provide robust TR decomposition for tensors with missing data and heavy-tailed noise/outliers. This dramatically reduces computational cost while preserving statistical efficiency (He et al., 2023).

5. Applications and Empirical Performance

The TR format's compressibility, symmetry, and computational tractability make it suitable for a range of high-dimensional contexts:

• Tensor Completion: Weighted optimization (TR-WOPT) recovers missing entries via gradient-based optimization on the cores, consistently outperforming TT and CP alternatives, especially for high missing data rates and higher-order tensors (Yuan et al., 2018, Yuan et al., 2018).
• Representation Learning: NTR and GNTR yield interpretable factors for object and face recognition, hyperspectral image analysis, and clustering, with significant accuracy gains over unconstrained and TT-based approaches (Yu et al., 2020).
• Dynamic Networks: TR, coupled with single latent factor-dependent multiplicative updates and bias modeling, effectively captures evolving high-order dependencies in large-scale dynamic networks, surpassing CP-based factorizations in predictive accuracy (Wang, 2023).
• Quantum Tomography: Finite-step algebraic recovery supports provable matrix product state (MPS) tomography from local marginals, bridging TR and quantum information (Chen et al., 30 Nov 2025).
• Large-Scale/Streaming Tensors: Randomized, sketching-based, and streaming TR algorithms afford real-time tracking and rapid batch updates for high-dimensional time-series and video data (Yu et al., 2023, Yuan et al., 2019, Yu et al., 2022).

Across synthetic and real-world datasets, randomized and structure-exploiting TR algorithms deliver order-of-magnitude speedups as compared to standard ALS, while robust/regularized schemes improve performance under noise, outliers, or model misspecification (Chen et al., 30 Nov 2025, He et al., 2023, Malik et al., 2020).

6. Limitations and Open Problems

Despite its advantages, TR decomposition presents notable limitations and challenges:

• Absence of Stable Best Approximation: Unlike TT, there is no guarantee of a minimal-rank TR after common operations or a best low-rank approximation to arbitrary tensors. Truncation and rounding may be ineffective, with ranks growing and no singular value gap for thresholding (Batselier, 2018).
• Non-uniqueness (Gauge Freedom): Cyclic gauge redundancy complicates statistical identifiability, geometric analysis, and optimization, although recent work establishes appropriate quotient characterizations (Gao et al., 29 Jan 2026).
• Hyperparameter and Rank Selection: Model selection is a challenge: the possible TR-rank configurations grow exponentially with order. Latent regularization and adaptive rank selection ameliorate, but do not eliminate, this issue (Yuan et al., 2018).
• Computational Scaling: While randomized and blockwise algorithms mitigate the curse of dimensionality, operations on extremely large tensors remain bottlenecked without careful exploitation of data sparsity and structure (Yuan et al., 2019, Yu et al., 2022).

7. Future Directions

Current research trends seek to further extend the applicability and understanding of TR decomposition:

• Riemannian and Non-Euclidean Algorithms: Quotient manifold structures enable principled Riemannian optimization, with dimension reduction and improved convergence properties (Gao et al., 29 Jan 2026).
• Provable Algorithms for Structure Recovery: Finite-step, sample-optimal algorithms for exact and noisy recovery open new directions for provable tensor network learning in statistics and quantum physics (Chen et al., 30 Nov 2025).
• Scalable, Robust, and Interpretable TR: Ongoing work on automated rank selection, robust statistics, graph regularization, and domain-specific constraints will increase practical impact in scientific computing, machine learning, and network analysis (He et al., 2023, Yu et al., 2020).
• Extension to Quantized, Tree-structured, and Higher-Order Models: Tree-TT, quantized-TR, and hybrid tensor networks represent potential avenues for addressing ultra-high-dimensional and structured data (Zhao et al., 2016).

The TR decomposition thus constitutes a flexible, cyclic generalization of popular low-rank tensor factorizations, supporting both theoretical advances in tensor geometry and a range of scalable, robust algorithms for high-dimensional inference and completion. Its further development is likely to hinge on advances in both mathematical theory and large-scale randomized algorithms.