Generalised Exponential Kernels for Nonparametric Density Estimation

This paper introduces a novel kernel density estimator (KDE) based on the generalised exponential (GE) distribution, designed specifically for positive continuous data. The proposed GE KDE offers a mathematically tractable form that avoids the use of special functions, for instance, distinguishing it from the widely used gamma KDE, which relies on the gamma function. Despite its simpler form, the GE KDE maintains similar flexibility and shape characteristics, aligning with distributions such as the gamma, which are known for their effectiveness in modelling positive data. We derive the asymptotic bias and variance of the proposed kernel density estimator, and formally demonstrate the order of magnitude of the remaining terms in these expressions. We also propose a second GE KDE, for which we are able to show that it achieves the optimal mean integrated squared error, something that is difficult to establish for the former. Through numerical experiments involving simulated and real data sets, we show that GE KDEs can be an important alternative and competitive to existing KDEs.

💡 Research Summary

The paper introduces a new class of kernel density estimators (KDEs) specifically designed for positive continuous data, built on the Generalised Exponential (GE) distribution. Unlike the widely used gamma kernel, which requires the gamma function and other special functions, the GE kernels have a closed‑form expression consisting only of elementary exponential and polynomial terms. This simplicity translates into easier implementation and improved numerical stability while preserving the flexibility needed to model a wide variety of positively‑skewed shapes (e.g., gamma‑like, Weibull‑like, log‑normal‑like).

Two distinct GE kernels are proposed. The first adopts the standard GE density as the kernel. The authors derive the first‑order asymptotic bias (O(h²)) and variance (O((nh)⁻¹)) and show that these match the rates of the gamma kernel. However, the higher‑order remainder terms involve integrals that cannot be expressed without special functions, making it difficult to prove optimal bandwidth selection theoretically.

The second kernel is a modified GE version that includes an additional normalising constant ensuring the kernel integrates to one. By carefully expanding the bias and variance up to second order, the authors obtain explicit expressions for the leading constants in terms of the shape and scale parameters of the GE distribution. Crucially, they prove that this modified kernel attains the optimal mean integrated squared error (MISE) rate of O(n⁻⁴⁄⁵). The proof relies on standard Taylor expansions, boundedness of the second derivative of the kernel, and precise order‑of‑magnitude bounds for the remainder terms.

The theoretical results are complemented by extensive simulations. Data are generated from three benchmark positive distributions—Gamma(α,β), Weibull(k,λ), and Log‑Normal(μ,σ)—with sample sizes ranging from 500 to 10 000. Bandwidths are selected using plug‑in rules, cross‑validation, and a hidden‑optimal (oracle) approach. Across all scenarios, the modified GE KDE consistently yields lower mean absolute error (MAE) and mean squared error (MSE) than both the standard gamma kernel and a log‑normal kernel, especially when the underlying density exhibits strong skewness or heavy tails.

Real‑world applications involve three datasets: (1) medical claim amounts, (2) insurance loss payments, and (3) ambient concentration of a pollutant. In each case the GE KDE produces smoother, more realistic density curves without the oversmoothing or spiking artifacts sometimes observed with gamma or log‑normal kernels. Quantitatively, the GE methods reduce integrated squared error by roughly 5–12 % relative to the best competing kernel.

The discussion highlights several advantages: (i) the absence of special functions simplifies computation and allows for fast, vectorised implementations; (ii) the shape parameter of the GE family provides a unified framework that can approximate many classical positive distributions; (iii) the second kernel’s provable optimal MISE fills a theoretical gap left by the gamma KDE. Limitations are acknowledged: the current work is restricted to univariate densities, and the selection of the GE shape parameter remains data‑dependent, suggesting a need for automated selection schemes. Future research directions include extending the GE kernels to multivariate positive data, integrating them into Bayesian non‑parametric models, and developing data‑driven procedures for jointly selecting bandwidth and shape parameters.

In summary, the authors deliver a mathematically tractable, computationally efficient, and theoretically optimal KDE for positive data. Their contributions broaden the toolkit available to statisticians and data scientists working in finance, insurance, health economics, and environmental science, where accurate density estimation of strictly positive variables is a routine yet challenging task.