Lattice approximation and linear discrepency of totally unimodular matrices

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

This paper shows that the lattice approximation problem for totally unimodular matrices A □ Rm×n can be solved efficiently and optimally via a linear programming approach. The complexity of our algorithm is &Ogr;(log m) times the complexity of finding an extremal point of a polytope in Rn described by 2(m + n) linear constraints. We also consider the worst-case approximability. This quantity is usually called linear discrepancy lindisc(/4). For any totally unimodular m × n matrix A we show lindisc(4) ≤ min{1 - 1/n+1, 1 - 1/m}. This bound is sharp. It proves Spencer's conjecture lindisc(A) ≤ (1 - 1/n+1) herdisc(4) for totally unimodular matrices. This seems to be the first time that linear programming is successfully used for a discrepancy problem.

Original languageEnglish
Title of host publicationProceedings of the 12th Annual ACM-SIAM Symposium on Discrete Algorithms
Pages119-125
Number of pages7
Publication statusPublished - 1 Dec 2001
Externally publishedYes
Event2001 Operating Section Proceedings, American Gas Association - Dallas, TX, United States
Duration: 30 Apr 20011 May 2001

Publication series

NameProceedings of the Annual ACM-SIAM Symposium on Discrete Algorithms

Conference

Conference2001 Operating Section Proceedings, American Gas Association
Country/TerritoryUnited States
CityDallas, TX
Period30/04/011/05/01

Keywords

  • Algorithms
  • Theory
  • Verification

Fingerprint

Dive into the research topics of 'Lattice approximation and linear discrepency of totally unimodular matrices'. Together they form a unique fingerprint.

Cite this