Stochastic dominance and the bijective ratio of online algorithms

TL;DR

This paper advances the analysis of online algorithms by introducing the bijective ratio, a generalization of bijective analysis and stochastic dominance, enabling more comprehensive performance evaluation across various problems and metrics.

Contribution

It proposes sufficient conditions for bijective optimality, introduces the bijective ratio as a new performance measure, and demonstrates its application to the continuous k-server problem.

Findings

Greedy algorithm has bijective ratios of O(k) on line, circle, and star metrics.

The bijective ratio generalizes stochastic dominance and applies to all online problems.

Results align with practical efficiency of algorithms beyond competitive analysis.

Abstract

Stochastic dominance is a technique for evaluating the performance of online algorithms that provides an intuitive, yet powerful stochastic order between the compared algorithms. Accordingly this holds for bijective analysis, which can be interpreted as stochastic dominance assuming the uniform distribution over requests. These techniques have been applied to some online problems, and have provided a clear separation between algorithms whose performance varies significantly in practice. However, there are situations in which they are not readily applicable due to the fact that they stipulate a stringent relation between the compared algorithms. In this paper, we propose remedies for these shortcomings. First, we establish sufficient conditions that allow us to prove the bijective optimality of a certain class of algorithms for a wide range of problems; we demonstrate this approach in the context of some well-studied online problems. Second, to account for situations in which two algorithms are incomparable or there is no clear optimum, we introduce the bijective ratio as a natural extension of (exact) bijective analysis. Our definition readily generalizes to stochastic dominance. This renders the concept of bijective analysis (and that of stochastic dominance) applicable to all online problems, and allows for the incorporation of other useful techniques such as amortized analysis. We demonstrate the applicability of the bijective ratio to one of the fundamental online problems, namely the continuous $k$-server problem on metrics such as the line, the circle, and the star. Among other results, we show that the greedy algorithm attains bijective ratios of $O(k)$ consistently across these metrics. These results confirm extensive previous studies that gave evidence of the efficiency of this algorithm on said metrics in practice, which, however, is not reflected in competitive analysis.