Towards a Geometric Understanding of Tensor Learning via the t-Product

Andong Wang, Yuning Qiu, Haonan Huang, Zhong Jin, Guoxu Zhou, Qibin ZhaoNeurIPS 2025 · OpenReview

Introduction

Tensor-based modeling is a staple in machine learning for handling structured, multi-dimensional data like 3D videos or multispectral images. Specifically, the t-product framework (and the associated t-SVD) has been highly successful because it treats "tubes" of data as single algebraic units, preserving spatial and spectral correlations.

However, a glaring theoretical gap remains: we lack a geometric language to describe these models. Classical differential geometry is built on real-valued functions and Euclidean spaces, meaning it cannot "see" the algebraic structure induced by the transforms used in tensor operations. Without a geometric theory, we cannot formally define concepts like smoothness, gradients, or geodesics in a way that respects the tensor's internal structure. This paper fills that gap by introducing a rigorous differential geometric framework for tensors based on the t-product algebra, moving the field from "flat" linear tensor models to "curved" tensor manifolds.

The t-Manifold Framework

The authors build their geometric framework from the ground up, establishing a bridge between discrete and continuous tensor settings within a unified algebraic-geometric perspective.

  1. Algebraic Foundation: They first define t-scalars (the "numbers" of this system, which are actually tubes or fibers) and t-modules (the "vector spaces" where these scalars live).
  2. t-Manifolds: They introduce the t-manifold, a topological space that looks locally like a t-module. Unlike a standard manifold that is locally Euclidean ($\mathbb{R}^n$), a t-manifold is locally modeled on a ring of commutative t-scalars.
Conceptual overview of the t-manifold framework, illustrating the construction from t-scalars to t-manifolds and their decomposition in the transform domain.
  1. Sheaf Theory for Smoothness: To handle "smoothness" across different frequencies in the transform domain, the authors employ a sheaf-theoretic approach. This allows them to define transform-consistent geometric objects, including Riemannian metrics, Laplacians, and geodesics.

Key Insights and Theoretical Guarantees

A core insight of this work is Frequency-Wise Decomposition. In the transform domain (e.g., after a Fourier or Cosine transform), a t-manifold decomposes into a collection of independent classical manifolds—one for each frequency. This decomposition makes complex tensor geometry mathematically tractable.

Furthermore, the "intrinsic dimension" $p$ is defined relative to the t-scalar ring. A dataset might appear incredibly high-dimensional in a standard Euclidean sense, but possess a very low "t-dimension," making learning much more efficient.

With the geometry established, the authors develop a robust learning theory:

  • Hypothesis Testing & Manifold Fitting: They provide a statistical procedure to test if tensor data lies near a low-dimensional t-manifold, proving that their manifold estimator converges to the true manifold at a non-asymptotic rate of $O(\sigma^2 \log(1/\sigma))$, where $\sigma$ is the noise level.
  • Tensor Neural Networks (TNNs): They establish generalization bounds for TNNs learning smooth functions over these manifolds. The generalization error depends on the effective dimension $p_{eff}$, which scales with the intrinsic complexity of the t-manifold rather than the raw number of pixels, effectively avoiding the curse of dimensionality.

Empirical Validation

The theoretical framework also yields practical insights. The authors show that enforcing constraints in two directions (e.g., row-wise and column-wise t-products) creates a "geometric twist" or nonlinearity, explaining why bidirectional tensor models often generalize better than simpler ones.

Inspired by this geometric view, they propose a Bidirectional Tensor Representation (BTR). BTR was evaluated on clustering and denoising tasks across several benchmark datasets (e.g., FRDUE, Olivetti, PIE-10, USPS1000, and YUV video sequences). It consistently outperformed state-of-the-art methods like R-TPCA and TLRR. For example, on the PIE-10 dataset, BTR achieved a clustering accuracy of 0.6110, significantly higher than the 0.4276 achieved by standard R-TPCA.

Limitations and Future Work

While foundational, the framework has limitations to address in future work:

  • Frequency Independence: The current theory assumes that different frequency components in the transform domain are independent. In reality, many datasets exhibit "cross-frequency coupling."
  • Unit Element Assumption: The mathematics currently requires a "unit element" in the t-scalar ring, which might not hold for all transform types without extending into distribution theory.
  • Computational Scalability: Developing numerically stable and scalable algorithms to perform optimization directly on these t-manifolds remains a major next step for practical deployment.

Summary updated

Abstract

Despite the growing success of transform-based tensor models such as the t-product, their underlying geometric principles remain poorly understood. Classical differential geometry, built on real-valued function spaces, is not well suited to capture the algebraic and spectral structure induced by transform-based tensor operations. In this work, we take an initial step toward a geometric framework for tensors equipped with tube-wise multiplication via orthogonal transforms. We introduce the notion of smooth t-manifolds, defined as topological spaces locally modeled on structured tensor modules over a commutative t-scalar ring. This formulation enables transform-consistent definitions of geometric objects, including metrics, gradients, Laplacians, and geodesics, thereby bridging discrete and continuous tensor settings within a unified algebraic-geometric perspective. On this basis, we develop a statistical procedure for testing whether tensor data lie near a low-dimensional t-manifold, and provide nonasymptotic guarantees for manifold fitting under noise. We further establish approximation bounds for tensor neural networks that learn smooth functions over t-manifolds, with generalization rates determined by intrinsic geometric complexity. This framework offers a theoretical foundation for geometry-aware learning in structured tensor spaces and supports the development of models that align with transform-based tensor representations.
BibTeX
@inproceedings{wang2025towards,
  author = {Andong Wang and Yuning Qiu and Haonan Huang and Zhong Jin and Guoxu Zhou and Qibin Zhao},
  title = {{Towards a Geometric Understanding of Tensor Learning via the t-Product}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2025},
  url = {https://openreview.net/forum?id=0uo2GPd94f},
}
Figure