Abstract
We investigate the sub-Gaussian property for almost surely bounded random variables. If sub-Gaussianity per se is de facto ensured by the bounded support of said random variables, then exciting research avenues remain open. Among these questions is how to characterize the optimal sub-Gaussian proxy variance? Another question is how to characterize strict sub-Gaussianity, defined by a proxy variance equal to the (standard) variance? We address the questions in proposing conditions based on the study of functions variations. A particular focus is given to the relationship between strict sub-Gaussianity and symmetry of the distribution. In particular, we demonstrate that symmetry is neither sufficient nor necessary for strict sub-Gaussianity. In contrast, simple necessary conditions on the one hand, and simple sufficient conditions on the other hand, for strict sub-Gaussianity are provided. These results are illustrated via various applications to a number of bounded random variables, including Bernoulli, beta, binomial, Kumaraswamy, triangular, and uniform distributions.
Introduction
In probability and statistics, sub-Gaussian bounds are essential tools for proving that a random variable does not deviate excessively from its mean. While sub-Gaussianity is guaranteed for variables with bounded support (e.g., Bernoulli trials or variables restricted to a finite interval), the "proxy variance" ($\sigma^2$) used in these bounds is frequently much larger than the actual variance. This discrepancy leads to loose and overly pessimistic concentration inequalities.
This research addresses two fundamental questions:
- Optimization: How can we determine the smallest possible proxy variance ($\sigma^2_{opt}$) for a given bounded distribution?
- Strictness: Under what conditions is the optimal proxy variance exactly equal to the actual variance? This property, termed strict sub-Gaussianity, represents the tightest possible Gaussian-like bound.
A central focus of this work is the role of symmetry. While common intuition suggests that symmetric distributions are strictly sub-Gaussian, this paper rigorously explores the boundaries of this relationship.
Methodology
The authors move beyond case-by-case analysis to develop a generalized analytical framework:
- Function Variation Analysis: They define a function $h(\lambda)$ derived from the cumulant generating function (CGF). The problem of finding the optimal proxy variance is reformulated as finding the global maximum of $h(\lambda)$.
- ODE Characterization: The research demonstrates that $h(\lambda)$ satisfies specific ordinary differential equations (ODEs). This approach provides an elegant way to prove the uniqueness of global maxima for standard distributions like the Bernoulli and Uniform distributions.
- Moment-Based Conditions: Necessary conditions for strict sub-Gaussianity are derived using 3rd and 4th cumulants (kurtosis), while sufficient conditions are established based on even moments for symmetric distributions.
- Constructive Counter-examples: The authors use specific mixtures of Dirac masses to demonstrate that symmetry is neither necessary nor sufficient for strict sub-Gaussianity.
Key Insights: The Role of Symmetry and Kurtosis
The most significant contribution of this paper is the debunking of the "symmetry implies strictness" myth.
- Symmetry is not a Silver Bullet: As illustrated in the figure above, a distribution can be perfectly symmetric yet fail to be strictly sub-Gaussian (Panels A & C), or it can be asymmetric and pass the test (Panels B & D).
- The Kurtosis Connection: A necessary condition for strict sub-Gaussianity is that the distribution must be "platykurtic," meaning its kurtosis is less than or equal to 3 (the kurtosis of a standard normal distribution). If a distribution has "fat tails" (even if bounded), it cannot be strictly sub-Gaussian.
- The "Standard" Exception: Interestingly, for common "named" distribution families—including Bernoulli, Beta, Binomial, Kumaraswamy, and Triangular—symmetry is equivalent to strict sub-Gaussianity. For these specific families, if it is symmetric, it is strict.
Results for Common Distributions
The framework was applied to several standard benchmarks to provide closed-form solutions and proof of strictness:
- Bernoulli and Binomial: These are strictly sub-Gaussian if and only if $p = 1/2$. The authors also derived a closed-form expression for the optimal proxy variance when $p \neq 1/2$.
- Uniform Distribution: Confirmed to be always strictly sub-Gaussian, with $\sigma^2_{opt} = \text{Var}(X) = \frac{1}{12}(b-a)^2$.
- Triangular and Beta/Kumaraswamy: In these families, strict sub-Gaussianity holds if and only if the distribution is symmetric.
Conclusion and Future Work
This paper provides a rigorous foundation for characterizing optimal sub-Gaussian bounds. By proving that symmetry is neither necessary nor sufficient for strict sub-Gaussianity in the general case, the authors settle a long-standing intuition in the field.
Future research directions include extending these tight characterizations to unbounded sub-Gaussian variables and seeking a "grand unified" necessary and sufficient condition that applies universally across all distribution types without relying on numerical optimization for asymmetric cases.
Summary updated