Tensor Star (TS) Decomposition

• Tensor Star (TS) is a tensor network decomposition that represents an Nth-order tensor using order-3 factor tensors and ring-connected order-4 core tensors arranged in a star structure.
• It captures direct correlations between arbitrary modes by separating physical modes from latent interactions, offering a balance between connectivity and storage efficiency.
• Optimization via ALS and PAM methods enables effective tensor compression and completion, as demonstrated by notable reconstruction quality even under high missing data conditions.

Searching arXiv for the primary paper and closely related tensor-network/tubal references. Tensor Star (TS) is a tensor network decomposition in which an order-$N$ tensor

$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$

is represented by $N$ order-$3$ factor tensors and $N$ order-$4$ core tensors arranged in a star shape, with the core tensors themselves connected in a ring. The model was introduced to combine two objectives that are often in tension: direct correlation characterization between arbitrary modes and storage growth that remains linear in tensor order rather than exponential. In this construction, the factor tensors carry the physical modes, the core tensors act as skip connections between factors, and the resulting network is positioned between sequential low-storage models such as Tensor Train and Tensor Ring, and higher-connectivity but more expensive models such as Tucker, Tensor Wheel, and Fully-Connected Tensor Network (Zhou et al., 2024).

1. Conceptual origin and modeling objective

TS decomposition is motivated by two recurring bottlenecks in higher-order representation learning. The first is the curse of dimensionality: direct storage of an order-$N$ tensor scales as $\mathcal{O}(I^N)$ when the mode sizes are comparable. The second is the curse of ranks: richer tensor networks often require rank structures or cores whose size grows rapidly with order. TS is designed to alleviate the first while controlling the second by using low-order local components and a structured latent topology (Zhou et al., 2024).

The central design choice is to separate physical modes from latent interaction structure. Each physical mode $I_k$ is attached to its own order-$3$ factor tensor, while latent dependence is mediated by order-$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$0 core tensors connected cyclically. Unlike Tensor Train (TT) and Tensor Ring (TR), consecutive factor tensors are not directly connected. Unlike Tucker or HOSVD, there is no single order-$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$1 core. The intended effect is that arbitrary-mode interactions can still be represented directly, but through the aggregate ring of cores rather than through one exponentially large central tensor (Zhou et al., 2024).

This places TS in a specific structural niche. Compared with TT and TR, it aims to move beyond primarily consecutive-dimension interactions. Compared with Tensor Wheel (TW) and Fully-Connected Tensor Network (FCTN), it sacrifices some direct global connectivity in exchange for substantially lower storage. A plausible implication is that TS should be read not merely as another low-rank factorization, but as a topology choice about how latent correlations are routed through the network.

2. Network topology and formal mathematical definition

TS decomposes $$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$2 into factor tensors

$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$3

and core tensors

$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$4

for $$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$5, with cyclic conditions

$$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$6

The $$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$7 tensors are the star “spokes,” each carrying one observed mode $$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$8, while the $$\mathcal{X}\in \mathbb{R}^{I_1\times I_2\times \cdots \times I_N}$$9 tensors form the ring-shaped latent backbone (Zhou et al., 2024).

Using the generalized tensor contraction notation of the paper, the decomposition is written as

$N$0

The paper denotes this compactly as

$N$1

Its contraction operator is explicitly non-commutative and non-associative, so ordering is part of the model specification rather than a mere notational convenience (Zhou et al., 2024).

The entrywise reconstruction formula makes the latent wiring explicit: $N$2 This formula shows why the model is described as a star with a ring core: physical indices appear only in the $N$3, whereas the $N$4 indices close a cyclic latent loop through the $N$5 (Zhou et al., 2024).

3. Rank structure, direct-correlation property, and storage complexity

TS introduces two rank families. The factor ranks

$N$6

control the two latent modes of each factor tensor, and the ring ranks

$N$7

control the cyclic core backbone. This already distinguishes TS from TT and TR, where consecutive factors share a single common bond dimension. In TS, consecutive factor tensors may have different latent ranks because their coupling is mediated by an order-$N$8 core rather than a single shared edge (Zhou et al., 2024).

A key theorem in the paper formalizes the claim of direct correlation characterization between any two arbitrary dimensions. After circular permutation of modes, all core tensors can be contracted into an aggregate tensor

$N$9

and then two selected factor tensors $3$0 and $3$1 become directly connected through that aggregate core. This is the paper’s formal justification for the claim that TS is not restricted to consecutive-dimension interactions (Zhou et al., 2024).

The rank theory reflects this topology. The mode-$3$2 matricization satisfies

$3$3

while the generalized unfolding rank obeys

$3$4

The paper emphasizes that the second bound involves only four latent terms, rather than a number of factors that grows with the number of unfolded modes. This suggests a more controlled rank geometry than in higher-connectivity networks such as FCTN and TW (Zhou et al., 2024).

Under the uniform assumptions $3$5, $3$6, and $3$7, the storage complexities given in the paper are as follows.

Model	Storage complexity
Original tensor	$3$8
CP	$3$9
Tucker	$N$0
TT	$N$1
TR	$N$2
FCTN	$N$3
TW	$N$4
TS	$N$5

For TS, the factor contribution is $N$6 and the core contribution is $N$7. Hence the model is linear in $N$8 but fourth-order in the latent rank. This is precisely the sense in which the paper says TS alleviates the curse of dimensionality while controlling, rather than eliminating, the curse of ranks (Zhou et al., 2024).

4. Optimization procedures: ALS decomposition and PAM completion

The basic approximation problem is

$N$9

The paper proposes an ALS for TS decomposition and a PAM for tensor completion. Both methods isolate one tensor block at a time, construct an environment tensor by contracting everything else, unfold the corresponding subproblem into matrix form, and solve a least-squares or regularized least-squares update (Zhou et al., 2024).

For factor updates, the paper introduces an environment tensor $4$0 such that

$4$1

In the PAM formulation, the proximal update becomes

$4$2

Core updates are derived analogously through a second environment tensor $4$3, leading to another closed-form proximal least-squares step (Zhou et al., 2024).

For completion, the objective is

$4$4

where $4$5 enforces equality on observed entries. The $4$6-update has the simple entrywise form

$4$7

Thus observed entries are clamped, while missing entries are updated by a proximal average of the current TS reconstruction and the previous iterate (Zhou et al., 2024).

Algorithm 1 in the paper specifies the practical initialization used for completion: $4$8 on $4$9 and $N$0 otherwise; $N$1 and $N$2 sampled from $N$3; proximal parameter $N$4; maximum iteration $N$5; and convergence threshold

$N$6

The stopping criterion is

$N$7

The paper cites general PAM convergence literature for motivation, but it does not prove a TS-specific convergence theorem (Zhou et al., 2024).

5. Compression, completion, and empirical evidence

The paper presents TS primarily as a framework for higher-order tensor decomposition/compression and tensor completion, and also suggests relevance for neural network compression because the core tensors act as skip connections and may shorten implicit gradient paths relative to sequential TT/TR structures (Zhou et al., 2024).

The explicit experiment described is a proof-of-concept completion study on the Toy MSI dataset from the Columbia CAVE multispectral image database, with tensor size

$N$8

The task uses 90\% missing entries. The observed data have

$N$9

and TSD-PAM, using

$\mathcal{O}(I^N)$0

achieves

$\mathcal{O}(I^N)$1

The paper also reports that the recovered 20th spectral slice visually reconstructs the original tensor well despite the high missingness and relatively small latent ranks, with largest rank $\mathcal{O}(I^N)$2 (Zhou et al., 2024).

The empirical description is intentionally limited. No direct baseline comparison table is reported for this experiment, no ablation study over rank configurations or structural variants is given, and no case in which TS underperforms is discussed. The experiment is therefore best read as a proof-of-concept establishing feasibility rather than as a comprehensive benchmark campaign. This suggests that the principal contributions of the paper are architectural and algebraic, with empirical validation still at an early stage.

6. Relation to adjacent tensor frameworks and terminological scope

TS decomposition should be distinguished from the tubal tensor algebra literature in which “tensor star” can informally refer to a tensor–tensor product induced by a tube algebra, especially the classical t-product and its transform-generalization $\mathcal{O}(I^N)$3. In that setting, third-order tensors are treated as matrices of tubes, and the product is

$\mathcal{O}(I^N)$4

with

$\mathcal{O}(I^N)$5

That framework is an algebra of tensor multiplication; TS decomposition is instead a specific tensor-network topology built from factor and core tensors (Avron et al., 3 Jun 2025).

There is nevertheless a clear family resemblance to broader matrix-of-structured-entries approaches. In generalized visual information analysis via tensorial algebra, matrices with entries that are fixed-size arrays—t-scalars—are manipulated via convolution-based multiplication, leading to TSVD, THOSVD, TPCA, T2DPCA, and TGCA. The shared principle is to preserve local multiway structure while retaining matrix-like or network-like operations, but the object class and decomposition topology differ from TS decomposition proper (Liao et al., 2020).

A later development, the Tubal Tensor Train (TTT) decomposition, makes the connection to network topology especially explicit: it keeps one distinguished tube mode, replaces scalar contractions by t-products, and organizes the remaining modes in a train. The paper states that the same tubal-network principle can be combined with more general tensor networks. This suggests that a tubal analogue of Tensor Star could be obtained by applying the same recipe to a star topology, although such a construction is not worked out there (Ahmadi-Asl et al., 11 Mar 2026).

The acronym TS is also overloaded across fields. In randomized numerical linear algebra it denotes Tensor Sketch (Cao et al., 2021); in neural networks it denotes Tensor Switching (Tsai et al., 2016); in STAR-RIS communications it denotes time switching (Mu et al., 2021); and in graph theory it can refer to the star chromatic number of a tensor product of graphs (Choudhary et al., 2023). Within tensor decomposition, however, Tensor Star specifically denotes the star-shaped tensor network introduced in 2024, characterized by $\mathcal{O}(I^N)$6 order-$\mathcal{O}(I^N)$7 factor tensors, $\mathcal{O}(I^N)$8 order-$\mathcal{O}(I^N)$9 ring-connected core tensors, direct arbitrary-mode correlation characterization, and storage complexity $I_k$0 under uniform-rank assumptions (Zhou et al., 2024).