A Bound on the Rate of Convergence in the Central Limit Theorem for Renewal Processes under Second Moment Conditions

Gesine Reinert, Ce Yang

Abstract

A famous result in renewal theory is the Central Limit Theorem for renewal processes. As in applications usually only observations from a finite time interval are available, a bound on the Kolmogorov distance to the normal distribution is desirable. Here we provide an explicit non-uniform bound for the Renewal Central Limit Theorem based on Stein's method and track the explicit values of the constants. For this bound the inter-arrival time distribution is required to have only a second moment. As an intermediate result of independent interest we obtain explicit bounds in a non-central Berry-Essén theorem under second moment conditions.

Introduction

In probability theory, a renewal process counts the number of occurrences of an event over a specific time interval $t$. While the Central Limit Theorem (CLT) guarantees that as $t$ approaches infinity, the distribution of these events $X_t$ will resemble a normal distribution, real-world applications are inherently finite. Engineers and scientists require a "speed limit"—a Berry-Esséen bound—to determine how closely their finite-time data approximates a true bell curve.

Historically, these bounds required the existence of a third moment, assuming the data was not too skewed or heavy-tailed. This paper by Gesine Reinert and Ce Yang fills a critical gap by providing explicit, non-uniform bounds for cases where only the second moment (variance) is known. This is particularly vital for analyzing processes with heavier tails, such as those following a Pareto distribution.

Methodology: Stein’s Method

The researchers utilize Stein’s Method, a functional approach that bounds the distance between two probability distributions by solving a specific differential equation. This departs from traditional Fourier-based proofs and involves three key technical steps:

  1. Decomposition: The error is split into the error inherent in the sum of inter-arrival times and the error caused by the "smoothness" of the normal distribution's cumulative distribution function (c.d.f.).
  2. Non-Uniform Bounding: Rather than a single global error value, the authors derive bounds that adjust based on the position on the distribution curve.
  3. Tracking Constants: The authors explicitly calculate the numerical values of constants throughout the proofs, moving beyond mere existence proofs to provide concrete mathematical baselines.

Key Insights and Theoretical Contributions

Second Moment Sufficiency

The most significant finding is that a finite third moment is not required to guarantee a convergence rate of $O(t^{-1/2})$. This demonstrates that the normal approximation is more robust for "messier" data than previously proven, provided the variance is finite.

Intermediate Berry-Esséen Result

To solve the renewal process problem, the authors first proved a new version of the non-central Berry-Esséen theorem for general sums of independent variables under second-moment conditions. This result is a significant contribution to probability theory in its own right.

Explicit Constants

The paper provides actual numerical values for the constants in the bounds. For example, in Theorem 3.3, they show that a specific constant $C_2$ is $\le 37$ when the variance is small and the value is near the mean, though it can jump to $25,431$ in the tails. While these numbers are currently large, they provide the first concrete baseline for this specific problem.

Results and Limitations

The paper confirms that the rate of convergence remains $t^{-1/2}$, matching the rate seen in cases where the third moment is known. The "main event" of the paper, Theorem 3.1, provides the complex formula for the explicit upper bound on the difference between the renewal process and the normal distribution.

However, the authors acknowledge certain limitations:

  • Large Constants: Mathematically, a bound of 50,000 on a probability (which must be between 0 and 1) is true but not yet practically useful for small datasets.
  • Tail Deterioration: The bound becomes less informative when the time $t$ is very close to the expected value $n\mu$.

Conclusion

By using Stein's method to establish these bounds under second-moment conditions, Reinert and Yang have cleared a theoretical path. While the constants are currently high, this work provides the foundation for future research to "tighten" these numbers, eventually making the bounds practical for real-world statistical engineering.

Summary updated

Abstract

A famous result in renewal theory is the Central Limit Theorem for renewal processes. As in applications usually only observations from a finite time interval are available, a bound on the Kolmogorov distance to the normal distribution is desirable. Here we provide an explicit non-uniform bound for the Renewal Central Limit Theorem based on Stein's method and track the explicit values of the constants. For this bound the inter-arrival time distribution is required to have only a second moment. As an intermediate result of independent interest we obtain explicit bounds in a non-central Berry-Essén theorem under second moment conditions.
BibTeX
@article{reinert2020bound,
  author = {Gesine Reinert and Ce Yang},
  title = {{A Bound on the Rate of Convergence in the Central Limit Theorem for Renewal Processes under Second Moment Conditions}},
  journal = {J. Appl. Probab.},
  year = {2020},
  doi = {10.1017/jpr.2019.101},
}
Figure