Alternating Minimizations Converge to Second-Order Optimal Solutions

Authors: Qiuwei Li, Zhihui Zhu, Gongguo Tang

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical This work studies the second-order convergence for both standard alternating minimization and proximal alternating minimization. We show that under mild assumptions on the (nonconvex) objective function, both algorithms avoid strict saddles almost surely from random initialization. Together with known first-order convergence results, this implies that both algorithms converge to a second-order stationary point.
Researcher Affiliation Academia 1Department of Electrical Engineering, Colorado School of Mines 2Mathematical Institute for Data Science, Johns Hopkins University. Correspondence to: Gongguo Tang <gtang@mines.edu>.
Pseudocode Yes Algorithm 1 Alternating Minimization 1: Initialization: x0. 2: For k = 1, 2, . . . , recursively generate (xk, yk) by yk = arg min y Rm f(xk 1, y), xk = arg min x Rn f(x, yk). (2) Algorithm 2 Proximal Alternating Minimization 1: Input: β > Lf. 2: Initialization: (x0, y0). 3: For k = 1, 2, . . . , recursively generate (xk, yk) by yk = arg min y Rm f(xk 1, y) + β 2 y yk 1 2 2, xk = arg min x Rn f(x, yk) + β 2 x xk 1 2 2. (3)
Open Source Code No The paper does not contain an explicit statement about the release of open-source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets No The paper is purely theoretical, focusing on mathematical convergence proofs and 'stylized applications' which are theoretical problem formulations, not empirical studies using datasets. Therefore, no information about publicly available or open datasets is provided.
Dataset Splits No The paper is theoretical and does not involve empirical experiments with data splits for training, validation, or testing.
Hardware Specification No The paper is theoretical and does not describe empirical experiments, so no hardware specifications are mentioned.
Software Dependencies No The paper is theoretical and does not describe empirical experiments or software implementations with specific version numbers.
Experiment Setup No The paper is theoretical and does not describe empirical experiments, so no experimental setup details such as hyperparameters or training settings are provided.