Abstract
This paper investigates the asymptotic behavior of geometric sums $S_M = X_1 + \dots + X_M$, where $M$ is a geometrically distributed random variable with parameter $p$. As $p \searrow 0$, Renyi's theorem provides a first-order approximation, suggesting that $pS_M$ converges to an exponential distribution. We extend this result by developing high-order asymptotic expansions. When $E[X_1] > 0$, the expansion proceeds in powers of $p$; when $E[X_1] = 0$, it proceeds in powers of $\sqrt{p}$. These results are derived through a new uniform renewal theorem, which remains robust even as the mean of the summands approaches zero. We demonstrate the utility of these expansions by deriving corrected diffusion approximations for the steady-state waiting time in M/G/1 queues.
Introduction
In stochastic modeling, we frequently encounter random variables defined as the sum of a random number of independent components. A primary example is the geometric sum $S_M = X_1 + \dots + X_M$, where the number of terms $M$ follows a geometric distribution with success probability $p$. These sums are foundational in fields ranging from insurance risk (modeling ruin probabilities) to telecommunications (modeling packet delays).
As the probability $p$ becomes small, the number of terms $M$ grows large, and the scaled sum $pS_M$ typically converges to an exponential distribution. This result, known as Renyi’s theorem, serves as a "limit law" analogous to the Central Limit Theorem. However, in high-stakes engineering or financial contexts, this first-order approximation often lacks the precision required to capture tail behavior or extreme risks. This paper addresses the need for higher-fidelity approximations by providing a unified framework for calculating correction terms.
The Challenge of Heavy Traffic
The primary mathematical hurdle in refining Renyi's theorem occurs in the "heavy traffic" regime, where the mean of the individual components $E[X_1]$ is very close to zero. Standard renewal theorems, which describe the behavior of sums as the number of terms tends toward infinity, often break down at this singularity.
Existing literature typically treats the cases of positive mean and zero mean separately. This paper bridges that gap by introducing a Uniform Renewal Theorem. This theorem is "uniform" because its error bounds remain valid even as the mean of the underlying distribution vanishes, allowing for a seamless transition between different scaling regimes.
Methodology
Uniform Renewal Theory
The authors utilize Fourier analysis, building upon the classical frameworks of Stone and Carlsson, to derive estimates for the renewal measure. By employing sophisticated Fourier transform bounds, they establish a renewal theorem that is robust against variations in the mean of the summands.
Expansion Regimes
The paper identifies two distinct scaling behaviors based on the expectation of the summands:
- Positive Mean ($E[X_1] > 0$): The distribution of $S_M$ is expanded in powers of $p$. This provides higher-order corrections to the exponential limit, accounting for the skewness and kurtosis of the underlying $X_i$ variables.
- Zero Mean ($E[X_1] = 0$): The expansion must proceed in powers of $\sqrt{p}$. This shift reflects the fundamental scaling of Brownian motion and random walks without drift.
Truncation for Heavy Tails
To ensure the theory applies to a wide range of distributions, including those with "heavy tails" (where standard moments may not exist), the authors implement a truncation strategy. By ignoring extreme outliers that fall beyond a specific threshold, they can apply their renewal machinery to the truncated variables and then bound the error introduced by the truncation.
Key Results and Applications
Corrected Diffusion Approximations (CDAs)
The theoretical heart of the paper is the derivation of explicit, computable formulas for the distribution of $S_M$. These are termed "Corrected Diffusion Approximations." Unlike the basic exponential limit, these CDAs incorporate the specific moments (mean, variance, skewness, etc.) of the service time or claim size distribution, leading to significantly higher accuracy.
M/G/1 Queueing Systems
The authors apply their results to the M/G/1 queue—a standard model for a single-server queue with Poisson arrivals and general service times. They derive explicit formulas for the steady-state waiting time distribution. These formulas allow engineers to predict queue behavior in heavy-traffic conditions with much greater precision than previously possible using only first-order limits.
Discussion and Limitations
While the uniform renewal theorem provides a powerful tool for asymptotic analysis, it requires the underlying random variables to be strongly nonlattice. This means the variables cannot be restricted to a discrete set of values (like integers). While many physical processes are continuous, applying these results to purely discrete systems requires additional "smoothing" or discretization adjustments.
Furthermore, while the paper provides a recursive algorithm to determine the coefficients of the expansion, the complexity of these coefficients increases rapidly. For example, the second-order correction term may require knowledge of the first five moments of the underlying distribution.
Conclusion
By developing a uniform renewal theory, this research provides a rigorous and unified approach to the study of random geometric sums. The resulting high-order expansions offer a significant improvement over Renyi's theorem, particularly in the heavy-traffic regimes critical to queueing theory and risk management. Future work may extend these uniform results to more complex Markovian environments or multi-server networks.
Summary updated