The Euclidean Distance Degree of Orthogonally Invariant Matrix Varieties

Dmitriy Drusvyatskiy,Hon-Leung Lee,Giorgio Ottaviani,Rekha R. Thomas

DOI: https://doi.org/10.48550/arXiv.1601.07210

2016-01-27

Abstract:We show that the Euclidean distance degree of a real orthogonally invariant matrix variety equals the Euclidean distance degree of its restriction to diagonal matrices. We illustrate how this result can greatly simplify calculations in concrete circumstances.

Optimization and Control

What problem does this paper attempt to address?

The problem that this paper attempts to solve is about the calculation of the Euclidean Distance Degree (ED degree) of orthogonal - invariant matrix varieties. Specifically, the authors studied a special class of matrix varieties - orthogonal - invariant matrix varieties \( M \subseteq \mathbb{R}^{n \times t} \), which remain invariant under the action of any real orthogonal matrices \( U \in O(n) \) and \( V \in O(t) \), that is, \( UMV^\top = M \).

The core problem of the paper is to prove an elegant formula:
\[ \text{ED degree}(M) = \text{ED degree}(S) \]
where \( S \) is the diagonal restriction of \( M \), that is, \( S = \{ x : \text{Diag}(x) \in M \} \). This formula shows that the Euclidean distance degree of an orthogonal - invariant matrix variety \( M \) is equal to the Euclidean distance degree of its diagonal restriction \( S \).

Through this result, the authors provide a method to simplify the calculation of the Euclidean distance degree of the orthogonal - invariant matrix variety \( M \), because the diagonal restriction \( S \) usually has a simpler geometric structure, making the calculation more intuitive and simple.

In addition, the paper also explores the application of this result on specific matrix varieties, such as the rank - \( r \) matrix variety \( R_{n \times t}^r \) and the essential matrix variety \( E \) in computer vision, and shows the specific significance of these applications.