Oberseminar Discrete Optimization

Breaking the Barrier of Self-Concordant Barriers: Faster Interior Point Methods for M-MatricesOberseminar Discrete Optimization

by Adrian Vladu (CNRS & IRIF, Université Paris Cité)

Europe/Berlin
Arithmeum, Lennéstr., 2 - Seminarraum (Arithmeum / Research Institute for Discrete Mathematics)

Arithmeum, Lennéstr., 2 - Seminarraum

Arithmeum / Research Institute for Discrete Mathematics

100
Description

We study two fundamental problems in optimization: (1) scaling a symmetric positive definite matrix by a positive diagonal matrix so that the resulting matrix has row and column sums equal to 1, and (2) minimizing a quadratic function subject to hard non-negativity constraints. Both problems lend themselves to efficient algorithms based on interior point methods (IPMs). For general instances, standard self-concordance theory limits the iteration complexity of IPMs to $\widetilde{O}(n^{1/2})$, where $n$ is the matrix dimension. We show via an amortized analysis that, when the input matrix is an M-matrix, an IPM with adaptive step sizes solves both problems in only $\widetilde{O}(n^{1/3})$ iterations. As a corollary, using fast Laplacian solvers, we obtain an $\ell_{2}$ flow diffusion algorithm with depth $\widetilde{O}(n^{1/3})$ and work $\widetilde{O}(n^{1/3}\cdot \mathrm{nnz})$. This marks a significant instance where a standard log-barrier IPM permits provably fewer than $O(n^{1/2})$ iterations.

 

The Oberseminar takes place in the Seminarraum, 1st floor. Participants are invited to have coffee or tea in the lounge before.

Organized by

S. Held, S. Hougardy, L. Végh, J. Vygen