Core Concepts
The paper presents a new dependent rounding algorithm for bipartite graphs that achieves strong negative correlation properties. This algorithm is then used to obtain an improved approximation algorithm for the problem of scheduling on unrelated machines to minimize weighted completion time.
Abstract
The paper introduces a new dependent rounding algorithm for bipartite graphs that generates random variables with strong negative correlation properties. The algorithm takes as input a fractional assignment of values to the edges of the graph and outputs an integral solution where each right-node has at most one neighboring edge with value 1.
The key features of the algorithm are:
It generates negatively-correlated Exponential random variables and uses them for a rounding method inspired by a contention-resolution scheme.
It provides stronger and more flexible negative correlation properties compared to prior work.
The negative correlation guarantees scale with the size of each cluster or individual items, rather than being determined by worst-case bounds.
The algorithm has a parameter ρ that allows tuning the amount of anti-correlation, providing flexibility in the rounding.
The authors then apply this dependent rounding algorithm to the problem of scheduling on unrelated machines to minimize weighted completion time. They:
Solve an SDP relaxation to obtain a fractional assignment of jobs to machines.
Partition the jobs on each machine into clusters based on processing time.
Use the dependent rounding algorithm to convert the fractional assignment into an integral one, where each job is assigned to exactly one machine.
Schedule the jobs on each machine in non-increasing order of Smith ratio.
The authors show that this approach leads to a 1.398-approximation algorithm for the scheduling problem, improving upon the previous best 1.45-approximation.