Adaptive nonparametric estimation of a component density in a two-class mixture model

Gaëlle Chagny, Antoine Channarond, Van Hà Hoang, Angelina Roche

This paper, published in early 2021, presents a robust statistical framework for disentangling a specific signal from a mixture of data. It focuses on a "two-class mixture model," a common scenario in fields like genomics and signal processing where observations come from two sources: a known "null" distribution (often representing noise or inactive states) and an unknown "alternative" distribution (the signal of interest).

TL;DR

The authors introduce a new "randomly weighted kernel estimator" that can reconstruct an unknown probability density even when it is contaminated by a known distribution. Their method is fully "adaptive," meaning it automatically tunes itself to the smoothness of the data without prior knowledge, achieving optimal convergence rates through a data-driven bandwidth selection process.

What Problem Does This Paper Address?

Imagine you are looking at p-values from thousands of genetic tests. Most of these tests represent "null" results (no effect), which follow a uniform distribution. A small fraction represents true discoveries, following an unknown distribution $f$. Mathematically, your observed data follows the density: $$g(x) = \theta + (1 - \theta)f(x)$$ where $\theta$ is the proportion of null results.

The challenge is two-fold:

  1. Indirect Observation: You never see samples directly from $f$; you only see the mixture $g$.
  2. Unknown Smoothness: To estimate a density accurately using kernels (a common smoothing technique), you need to pick a "bandwidth" (window size). If the bandwidth is too small, the estimate is noisy; if it's too large, the estimate is blurry. Usually, picking the perfect bandwidth requires knowing how "smooth" the target function $f$ is—information we rarely have.

Methods

The authors propose a Randomly Weighted Kernel Estimator (AWKE). The core innovation lies in how they weight the observations.

  • Weighting Trick: They derive a weight function based on a "change of density" principle. By weighting each observation $X_i$ by a factor that accounts for the local density of the mixture versus the known null, they can create an estimator whose expected value is exactly the target density $f$.
  • Plug-in Estimators: Since the mixing proportion $\theta$ and the total mixture density $g$ are unknown, the authors use preliminary estimates ($\tilde{\theta}_n$ and $\hat{g}$) and "plug" them into the weight formula.
  • Goldenshluger-Lepski (GL) Adaptation: This is the "brain" of the method. Instead of the researcher picking a bandwidth $h$, the GL method compares estimators at different bandwidths and selects the one that best balances bias and variance based purely on the observed data.
  • Identifiability: To ensure the model can actually be solved, they assume $f$ vanishes near the edge of the domain (specifically near $x=1$). This allows them to estimate $\theta$ by looking at the "flat" part of the mixture distribution.

Datasets

The paper relies on Simulated Data to validate the theory, as is standard for foundational statistical methodology papers. They tested three distinct shapes for the unknown density $f$:

  1. A polynomial decay ($f_1$).
  2. A truncated power distribution ($f_2$).
  3. A truncated exponential distribution ($f_3$).

They varied the mixing proportion $\theta$ (from 0.35 to 0.65) and the sample size $n$ (500, 1000, and 2000) to see how the estimator scales.

Key Insights and Ideas

  • The "Oracle" Property: The authors prove an "oracle inequality." This means their data-driven bandwidth selection performs almost as well as if an "oracle" had told them the mathematically perfect bandwidth in advance.
  • Robustness to Estimation Error: A major contribution is proving that the errors introduced by using estimated weights (instead of true weights) do not ruin the final convergence rate. The estimator still hits the optimal minimax rate $(\frac{\log n}{n})^{2\beta/(2\beta+1)}$ for a function with $\beta$ levels of smoothness.
  • Symmetry for Boundary Correction: To estimate the proportion $\theta$ near the boundary $x=1$ without falling into the "boundary effect" trap (where kernel methods often fail), they use a "reflection method" to artificially symmetrize the data.

Key Results

  • Superior Accuracy: In numerical simulations, the proposed AWKE method consistently achieved lower Mean Squared Error (MSE) than existing state-of-the-art methods (specifically the Nguyen and Matias estimator).
  • Automatic Smoothing: The method successfully adapted to different shapes—it stayed sharp for the exponential distribution and smooth for the polynomial one without any manual tuning.
  • Boundary Performance: While all kernel methods struggle at the very edges of a dataset, the AWKE showed significantly better stability at $x=0.1$ and $x=0.9$ compared to previous weighted kernel approaches.

Limitations and Future Work

  • Identifiability Constraints: The method requires the signal $f$ to be zero in some neighborhood (e.g., near $x=1$). If the signal is spread across the entire domain, $\theta$ becomes much harder to estimate accurately.
  • Computational Complexity: The Goldenshluger-Lepski method requires computing a grid of estimators and comparing them pairwise, which is more computationally expensive than simple cross-validation.
  • Extension to General Known Components: While the paper focuses on a uniform null distribution (common in p-value analysis), the authors suggest the framework can be extended to any known density $\phi(x)$, though they leave the full theoretical proof of that extension for future work.

Summary updated

Abstract

A two-class mixture model, where the density of one of the components is known, is considered. We address the issue of the nonparametric adaptive estimation of the unknown probability density of the second component. We propose a randomly weighted kernel estimator with a fully data-driven bandwidth selection method, in the spirit of the Goldenshluger and Lepski method. An oracle-type inequality for the pointwise quadratic risk is derived as well as convergence rates over Holder smoothness classes. The theoretical results are illustrated by numerical simulations.
BibTeX
@article{chagny2021adaptive,
  author = {Gaëlle Chagny and Antoine Channarond and Van Hà Hoang and Angelina Roche},
  title = {{Adaptive nonparametric estimation of a component density in a two-class mixture model}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2021},
  doi = {10.1016/j.jspi.2021.05.004},
}

Loading related content...

Figure