Table of Contents
Fetching ...

Convergence rate of alternating projection method for the intersection of an affine subspace and the second-order cone

Hiroyuki Ochiai, Yoshiyuki Sekiguchi, Hayato Waki

Abstract

We study the convergence rate of the alternating projection method (APM) applied to the intersection of an affine subspace and the second-order cone. We show that when they intersect non-transversally, the convergence rate is $O(k^{-1/2})$, where $k$ is the number of iterations of the APM. In particular, when the intersection is not at the origin or forms a half-line with the origin as the endpoint, the obtained convergence rate can be exact because a lower bound of the convergence rate is evaluated. These results coincide with the worst-case convergence rate obtained from the error bound discussed in [Borwein et al., SIOPT, 2014] and [Drusvyatskiy et al., Math. Prog., 2017]. Moreover, we consider the convergence rate of the APM for the intersection of an affine subspace and the product of two second-order cones. We provide an example that the worst-case convergence rate of the APM is better than the rate expected from the error bound for the example.

Convergence rate of alternating projection method for the intersection of an affine subspace and the second-order cone

Abstract

We study the convergence rate of the alternating projection method (APM) applied to the intersection of an affine subspace and the second-order cone. We show that when they intersect non-transversally, the convergence rate is , where is the number of iterations of the APM. In particular, when the intersection is not at the origin or forms a half-line with the origin as the endpoint, the obtained convergence rate can be exact because a lower bound of the convergence rate is evaluated. These results coincide with the worst-case convergence rate obtained from the error bound discussed in [Borwein et al., SIOPT, 2014] and [Drusvyatskiy et al., Math. Prog., 2017]. Moreover, we consider the convergence rate of the APM for the intersection of an affine subspace and the product of two second-order cones. We provide an example that the worst-case convergence rate of the APM is better than the rate expected from the error bound for the example.
Paper Structure (26 sections, 18 theorems, 83 equations, 1 figure)

This paper contains 26 sections, 18 theorems, 83 equations, 1 figure.

Key Result

Lemma 2.1

Assume that $\bm{x}\in\mathop{\mathrm{bd}}\nolimits\mathcal{K}\setminus\{\bm{0}\}$. If $\bm{y}\in\mathcal{K}$ is orthogonal to $\bm{x}$, then $\bm{y}$ is formed as $\bm{y} = \alpha \hat{\bm{x}}$ for some $\alpha \ge 0$.

Figures (1)

  • Figure 1: Transition diagram of the recurrence formulas \ref{['Proj3']}, \ref{['Proj4']} and \ref{['Proj5']}: (15A) means the recurrence formula \ref{['Proj5']} with $Y_0 < Z_0$ and $Z_0>0$, and (15B) means \ref{['Proj5']} with $Y_0 > Z_0$ and $Z_0<0$. The solid line from \ref{['Proj4']} to (15B) means that the recurrence formula changes in one iteration. A dotted line means that the recurrence formula changes in a few iterations.

Theorems & Definitions (28)

  • Lemma 2.1
  • Theorem 2.1
  • Lemma 2.2
  • Lemma 2.3
  • Remark 2.1
  • Theorem 2.2
  • Proposition 2.1
  • Lemma 2.4
  • Theorem 3.1
  • Proposition 3.1
  • ...and 18 more