A coordinate-free condition number for convex programming

TL;DR

This paper introduces a geometric, coordinate-free version of Renegar's condition number for convex feasibility problems, linking it to Grassmannian distances and matrix condition numbers, with implications for probabilistic analysis.

Contribution

It defines a new geometric condition number for convex programming, relating it to existing matrix condition numbers and differential geometry, extending prior work.

Findings

C_G(W) bounds R(A) and relates to matrix condition number

C_G(W) characterized via Riemannian distance on Grassmann manifold

Foundation for probabilistic analysis of condition numbers in convex programming

Abstract

We introduce and analyze a natural geometric version of Renegar's condition number R for the homogeneous convex feasibility problem associated with a regular cone C subseteq R^n. Let Gr_{n,m} denote the Grassmann manifold of m-dimensional linear subspaces of R^n and consider the projection distance d_p(W_1,W_2) := ||Pi_{W_1} - Pi_{W_2}|| (spectral norm) between W_1 and W_2 in Gr_{n,m}, where Pi_{W_i} denotes the orthogonal projection onto W_i. We call C_G(W) := max {d_p(W,W')^{-1} | W' \in Sigma_m} the Grassmann condition number of W in Gr_{n,m}, where the set of ill-posed instances Sigma_m subset Gr_{n,m} is defined as the set of linear subspaces touching C. We show that if W = im(A^T) for a matrix A in R^{m\times n}, then C_G(W) \le R(A) \le C_G(W) kappa(A), where kappa(A) =||A|| ||A^\dagger|| denotes the matrix condition number. This extends work by Belloni and Freund in Math. Program. 119:95-107 (2009). Furthermore, we show that C_G(W) can as well be characterized in terms of the Riemannian distance metric on Gr_{n,m}. This differential geometric characterization of C_G(W) is the starting point of the sequel [arXiv:1112.2603] to this paper, where the first probabilistic analysis of Renegar's condition number for an arbitrary regular cone C is achieved.