Queue Stability and Probability 1 Convergence via Lyapunov Optimization

TL;DR

This paper extends Lyapunov optimization theory to show that under certain conditions, queue stability and performance bounds hold with probability 1, not just in expectation, using martingale convergence techniques.

Contribution

It demonstrates that quadratic Lyapunov functions and mild moment conditions ensure all major stability forms, and that bounds hold almost surely, not only on average.

Findings

All major stability forms are implied by the basic drift condition with mild moment assumptions.

Probability 1 bounds for queue backlog and penalties are established under drift-plus-penalty.

The analysis combines Lyapunov drift theory with martingale convergence laws.

Abstract

Lyapunov drift and Lyapunov optimization are powerful techniques for optimizing time averages in stochastic queueing networks subject to stability. However, there are various definitions of queue stability in the literature, and the most convenient Lyapunov drift conditions often provide stability and performance bounds only in terms of a time average expectation, rather than a pure time average. We extend the theory to show that for quadratic Lyapunov functions, the basic drift condition, together with a mild bounded fourth moment condition, implies all major forms of stability. Further, we show that the basic drift-plus-penalty condition implies that the same bounds for queue backlog and penalty expenditure that are known to hold for time average expectations also hold for pure time averages with probability 1. Our analysis combines Lyapunov drift theory with the Kolmogorov law of large numbers for martingale differences with finite variance.