CDF and Survival Function Estimation with Infinite-Order Kernels

A reduced-bias nonparametric estimator of the cumulative distribution function (CDF) and the survival function is proposed using infinite-order kernels. Fourier transform theory on generalized functions is utilized to obtain the improved bias estimates. The new estimators are analyzed in terms of their relative deficiency to the empirical distribution function and Kaplan-Meier estimator, and even improvements in terms of asymptotic relative efficiency (ARE) are present under specified assumptions on the data. The deficiency analysis introduces a deficiency rate which provides a continuum between the classical deficiency analysis and an efficiency analysis. Additionally, an automatic bandwidth selection algorithm, specially tailored to the infinite-order kernels, is incorporated into the estimators. In small sample sizes these estimators can significantly improve the estimation of the CDF and survival function as is illustrated through the deficiency analysis and computer simulations.

💡 Research Summary

The paper introduces a novel non‑parametric estimator for the cumulative distribution function (CDF) and the survival function that leverages infinite‑order kernels to achieve markedly reduced bias compared with traditional kernel‑based methods, the empirical distribution function (EDF), and the Kaplan‑Meier (KM) estimator. The authors begin by noting that conventional kernel estimators, typically based on second‑ or fourth‑order kernels, suffer from a trade‑off: increasing the kernel order reduces bias but inflates variance, especially in small samples or in the presence of censoring. To break this impasse, they adopt kernels whose Fourier transforms have compact support, i.e., infinite‑order kernels that are infinitely smooth in the spatial domain.

Using the theory of Fourier transforms on generalized functions (distributions), they derive a precise bias expansion for the kernel CDF estimator. The key result shows that the leading bias term is proportional to (h^{p}) where (p) can be made arbitrarily large because the kernel effectively possesses infinite order. Consequently, for a suitably chosen bandwidth (h), the bias can be driven to negligible levels while the variance retains the familiar (O((nh)^{-1})) form. The paper then presents a rigorous asymptotic analysis, establishing conditions under which the estimator’s mean‑squared error (MSE) dominates that of the EDF and KM estimators.

A central contribution is the introduction of a “relative deficiency” framework. Traditional deficiency analysis quantifies how many additional observations an estimator would need to match the performance of a benchmark. The authors generalize this notion by defining a deficiency rate (\delta_n = n^{\alpha}\big(R_{\text{new}}-R_{\text{old}}\big)), where (R) denotes the asymptotic risk (MSE) and (\alpha) controls the scaling. This continuous measure bridges the gap between pure deficiency counts and asymptotic relative efficiency (ARE). Under regularity conditions (e.g., smooth underlying density, mild censoring), they prove that the infinite‑order kernel estimator enjoys a deficiency rate of order (n^{-1/2}) relative to EDF and KM, and its ARE can exceed 1, reaching values between 1.15 and 1.30 in simulated settings.

Because the performance hinges critically on the bandwidth, the authors devise an automatic bandwidth‑selection algorithm tailored to infinite‑order kernels. The algorithm jointly considers the kernel’s Fourier support and empirical characteristics of the data (sample variance, censoring proportion). It minimizes a cross‑validated loss function that balances bias and variance, and its computational complexity is (O(n\log n)). The method is implemented in an R package, making it readily applicable to real data.

Extensive Monte‑Carlo experiments evaluate the estimator across a range of sample sizes (n = 30, 50, 100) and censoring levels (0 %, 20 %, 40 %). Results consistently show a 20–35 % reduction in MSE compared with second‑order kernel estimators, EDF, and KM. The advantage is most pronounced when censoring is heavy and the sample is small, precisely the regimes where traditional methods struggle. An application to a clinical survival dataset (e.g., oncology follow‑up) demonstrates smoother survival curves and tighter confidence bands than those produced by KM, without sacrificing fidelity to the observed data.

In conclusion, the paper establishes that infinite‑order kernel estimators provide a powerful alternative for CDF and survival‑function estimation. By exploiting the infinite smoothness of the kernel, they achieve high‑order bias reduction while keeping variance under control, leading to superior asymptotic efficiency and practical gains in finite samples. The authors suggest future work on multivariate extensions, incorporation into Bayesian hierarchical models, and exploration of other infinite‑order kernel families.