Abstract
In this paper, we obtain some new exponential inequalities for partial sums and their finite maximum of acceptable random variables by the results of Sung et al. (J. Korean Stat. Soc., 40, 109-114, 2011) and in different ways from theirs. The inequalities we obtained improve the existing corresponding results and, in some sense, are optimal. In addition, we introduce some concepts and examples of widely acceptable random variables to extend our results mentioned above.
Introduction
In probability theory and statistics, exponential inequalities—such as the Hoeffding or Chernoff bounds—are fundamental tools for estimating the probability that a sum of random variables deviates significantly from its mean. These "tail bounds" are well-established for independent and identically distributed (i.i.d.) variables. However, real-world stochastic processes often involve complex dependencies that standard models fail to capture.
This research focuses on Acceptable Random Variables (ARVs). ARVs are a class of dependent variables where the expected value of the exponent of their sum is bounded by the product of the expected values of the exponents of the individual variables. While previous literature provided functional bounds for ARVs, they often lacked optimality or failed to address the maximum of partial sums—a critical metric in risk insurance and financial modeling. This paper addresses these gaps by introducing the Petrov-exponent to sharpen these mathematical tools.
Methodology
The study is a theoretical derivation within probability theory. The authors utilize several key mathematical frameworks to establish tighter bounds:
- The Petrov-Exponent ($\lambda_{k0}$): Rather than employing a generic parameter $\lambda$ in the exponential bound, the authors prove the existence of a unique, optimal $\lambda$ that minimizes the upper bound for a specific distribution. This is achieved by solving what is termed the "Petrov equation."
- Analytical Tools: Cauchy’s Inequality and the properties of strictly increasing functions are used to demonstrate the uniqueness and optimality of these exponents.
- Copula Theory: To extend the scope of their findings, the authors use Farlie-Gumbel-Morgenstern (FGM) copulas. This allows for the construction of "widely acceptable" variables that exhibit dependency structures not captured by traditional ARV definitions, including certain types of positive dependence.
Key Insights and Results
The core contribution of this work lies in the refinement of existing inequalities and the expansion of the "acceptable" framework to more general dependency structures.
The Petrov-Exponent Advantage
The authors demonstrate that by identifying the unique Petrov-exponent, one can find a value for $\lambda$ that makes the resulting exponential inequality optimal. This approach removes restrictive conditions (such as $\epsilon \leq K\delta$) that were necessary in previous theorems, such as those published by Sung et al. just months prior to this work.
Tighter Bounds for Partial Sums
Theorem 2.1 establishes a new exponential inequality for the partial sums of identically distributed ARVs. Proposition 2.2 mathematically proves that this new bound is strictly smaller—and therefore more accurate—than previous results. Specifically, the authors show that their bound: $$\exp\{-n(\lambda_{n0}\epsilon - \log E \exp\{\lambda_{n0}(X-EX)\})\}$$ is strictly less than the previous standard of $\exp\{-n\epsilon^2 / 4K\}$.
Addressing Finite Maximums
A significant practical advancement is presented in Theorem 2.2, which provides an exponential inequality for the finite maximum of partial sums: $$\max_{1 \leq k \leq n} \sum_{i=1}^k (X_i - EX_i)$$ In fields like risk management and insurance, understanding the maximum deviation over a period is often more critical than understanding the final sum at the end of that period.
Generalization to Widely Acceptable (WA) Variables
The paper introduces Widely Acceptable (WA) variables. While traditional ARVs often model negative dependence (where one variable being high makes others likely to be low), WA variables can accommodate certain types of positive dependence. By adding a scaling factor $g(n)$ to the bounds, the authors extend their inequalities to these more complex dependency structures, significantly broadening the applicability of the theory to various stochastic processes.
Discussion and Limitations
While the paper provides significant theoretical improvements, certain limitations remain for practical application:
- Identical Distribution Assumption: The primary theorems assume that the random variables are identically distributed. In many empirical scenarios, variables may follow different distributions over time or across different sources.
- Parameter Calculation: Finding the optimal $\lambda$ requires knowledge of the moment-generating function of the underlying distribution. In practice, if the distribution is unknown or empirical data is sparse, calculating this optimal parameter becomes a non-trivial task.
Future research may focus on extending these results to non-identically distributed variables or deriving other limit theorems, such as the Law of Large Numbers, within the "widely acceptable" framework.
Summary updated