What is the probability that a random symmetric tensor is close to rank-one?

Alberto Cazzaniga, Antonio Lerario, Andrea Rosana

Abstract

We address the general problem of estimating the probability that a real symmetric tensor is close to rank-one tensors. Using Weyl's tube formula, we turn this question into a differential geometric one involving the study of metric invariants of the real Veronese variety. More precisely, we give an explicit formula for its reach and curvature coefficients with respect to the Bombieri-Weyl metric. These results are obtained using techniques from Random Matrix theory and an explicit description of the second fundamental form of the Veronese variety in terms of GOE matrices. Our findings give a complete solution to the original problem. In the case of rational normal curves, it leads to a simple formula describing explicitly exponential decay with respect to the degree of the tensor.

Introduction

In modern computational fields—ranging from machine learning and signal processing to quantum information theory—tensors serve as the primary language for representing multi-dimensional data. However, the inherent complexity and memory requirements of high-order tensors often necessitate approximation. The most fundamental approximation is the rank-one tensor, formed by the outer product of vectors. In quantum mechanics, these represent "separable" states; the distance of a state from the set of rank-one tensors provides a geometric measure of its entanglement.

This paper investigates a fundamental statistical question: given a real symmetric tensor drawn from a natural Gaussian distribution, what is the exact probability that it lies within a distance $\delta$ of the set of rank-one tensors?

Geometric Translation and Weyl’s Tube Formula

The authors approach this algebraic problem by translating it into the language of differential geometry. They map the space of symmetric tensors to the space of homogeneous polynomials, equipped with the Bombieri-Weyl (or Kostlan) probability distribution. In this geometric setting, the set of all rank-one tensors corresponds to the spherical Veronese variety.

Determining if a tensor is within a distance $\delta$ of a rank-one tensor is equivalent to calculating the volume of a "tubular neighborhood" of radius $\epsilon$ (where $\epsilon = \arcsin(\delta)$) around this variety on a unit sphere. To compute this volume, the authors employ Weyl’s tube formula, which expresses the volume of a tube around a manifold in terms of its intrinsic and extrinsic curvature invariants.

The Random Matrix Connection

The primary challenge in applying Weyl’s formula lies in evaluating the "curvature coefficients" of the Veronese variety. The authors provide a significant breakthrough by linking algebraic geometry with Random Matrix Theory (RMT). They prove that the second fundamental form of the Veronese variety—the operator that encodes how the variety curves within its ambient space—can be explicitly described using Gaussian Orthogonal Ensemble (GOE) matrices.

By demonstrating that the Weingarten operator behaves like a GOE matrix, the authors reduce complex geometric integrals into the more manageable task of computing the expected values of random matrix determinants. This connection allows for the derivation of exact formulas that were previously considered intractable.

Key Results

1. The Reach of the Veronese Variety

A critical metric invariant identified is the reach ($\rho$) of the variety—the maximum radius at which a tubular neighborhood remains "well-behaved" without self-intersecting. The authors calculated the reach of the spherical Veronese variety as:

  • $\rho = \pi/4$ for tensor degrees $2 \le d \le 5$.
  • $\rho = \frac{1}{\sqrt{2}}\sqrt{1 + \frac{1}{d-1}}$ for degrees $d \ge 6$.

From a practical standpoint, the reach provides a guarantee: any tensor within this distance of the Veronese variety is mathematically guaranteed to have a unique best rank-one approximation.

2. Exact Probability Formulas

The paper provides a comprehensive closed-form integral formula (Theorem E) for the probability of a tensor being $\delta$-close to rank-one. While the general formula is complex, it simplifies significantly in specific cases:

  • Rational Normal Curves: For these specific varieties, the probability simplifies to the elegant expression $\sqrt{d} \delta^{d-1}$, showing an exponential decay relative to the tensor's degree.
  • The Matrix Case ($d=2$): For standard symmetric matrices, the authors derive a novel result for the probability that the Frobenius norm of a matrix is bounded by its smallest singular value.

Discussion and Limitations

While the results provide a complete theoretical solution, the authors note certain limitations. The general probability formula (Theorem E) involves complex combinations of Gamma functions and integrals, making it computationally intensive. Furthermore, the validity of Weyl’s tube formula is strictly limited to distances smaller than the reach of the manifold; it cannot describe the distribution of tensors that are very far from the rank-one set.

Future research may extend this framework to higher-rank approximations. Geometrically, this would involve studying the volumes of tubular neighborhoods around secant varieties to the Veronese variety, which represent tensors of rank 2, 3, and beyond. This work establishes a robust foundation for using differential geometry and RMT to understand the landscape of tensor approximation.

Summary updated

Abstract

We address the general problem of estimating the probability that a real symmetric tensor is close to rank-one tensors. Using Weyl's tube formula, we turn this question into a differential geometric one involving the study of metric invariants of the real Veronese variety. More precisely, we give an explicit formula for its reach and curvature coefficients with respect to the Bombieri-Weyl metric. These results are obtained using techniques from Random Matrix theory and an explicit description of the second fundamental form of the Veronese variety in terms of GOE matrices. Our findings give a complete solution to the original problem. In the case of rational normal curves it leads to a simple formula describing explicitly exponential decay with respect to the degree of the tensor.
BibTeX
@article{cazzaniga2024what,
  author = {Alberto Cazzaniga and Antonio Lerario and Andrea Rosana},
  title = {{What is the probability that a random symmetric tensor is close to rank-one?}},
  journal = {SIAM J. Appl. Algebra Geom.},
  year = {2024},
  doi = {10.1137/23m1551675},
}
Figure