Minimax Nonparametric Two-sample Test under Smoothing

Xin Xing, Zuofeng Shang, Pang Du, Ping Ma, Wenxuan Zhong, Jun S. LiuarXiv 2019 · DOI · arXiv

Abstract

We consider the problem of comparing probability densities between two groups. A new probabilistic tensor product smoothing spline framework is developed to model the joint density of two variables. Under such a framework, the probability density comparison is equivalent to testing the presence/absence of interactions. We propose a penalized likelihood ratio test for such interaction testing and show that the test statistic is asymptotically chi-square distributed under the null hypothesis. Furthermore, we derive a sharp minimax testing rate based on the Bernstein width for nonparametric two-sample tests and show that our proposed test statistics is minimax optimal. In addition, a data-adaptive tuning criterion is developed to choose the penalty parameter. Simulations and real applications demonstrate that the proposed test outperforms the conventional approaches under various scenarios.

Introduction

The two-sample test—determining whether two sets of data originate from the same underlying distribution—is a fundamental task in statistical inference, with applications ranging from clinical trials to genomic analysis. While classical parametric tools like the $t$-test are effective for Gaussian distributions, modern datasets often exhibit complex characteristics such as multimodality and long tails.

Existing nonparametric alternatives, such as the Kolmogorov-Smirnov (KS) test, often lack sensitivity in the tails of distributions. More recent kernel-based methods, such as Maximum Mean Discrepancy (MMD), offer improvements but are highly sensitive to the choice of tuning parameters (e.g., kernel bandwidth). This paper introduces a new penalized likelihood ratio (PLR) test within a smoothing spline framework that achieves minimax optimality and provides a robust, data-adaptive solution to the two-sample problem.

Methodology

The Interaction Framework

The core innovation of this work is the reframing of the two-sample problem as an interaction test. Rather than comparing two densities directly, the authors model the joint density $f(x, z)$ of the data $X$ and the group label $Z$. Using a Functional ANOVA Decomposition, the log-transformed joint density is decomposed as:

$$\log f(x, z) = \eta(x, z) = \mu + \alpha(x) + \beta(z) + \gamma(x, z)$$

In this decomposition, $\gamma(x, z)$ represents the interaction effect. If the two distributions are identical, the interaction term $\gamma(x, z)$ is zero. Thus, testing for equality of distributions becomes equivalent to testing for the absence of interaction within a probabilistic tensor product smoothing spline framework.

Penalized Likelihood Ratio (PLR) Test

The authors propose a test statistic based on the ratio of the likelihood under the null hypothesis (no interaction) versus the alternative. To ensure the estimated densities are smooth and to prevent overfitting, a penalty term is incorporated to control the "roughness" of the estimates.

Theoretical Guarantees

  1. Wilks’ Phenomenon: The authors demonstrate that the PLR test statistic follows an asymptotic chi-square distribution. This allows for the calculation of p-values without the need for computationally expensive resampling or bootstrapping methods.
  2. Minimax Optimality: Using the concept of "Bernstein width," the paper proves that the PLR test achieves the sharpest possible detection rate. This means the test is theoretically guaranteed to detect the smallest possible differences between distributions that any statistical test could feasibly identify.
  3. Data-Adaptive Tuning: A specific criterion is developed to automatically select the penalty parameter $\lambda$, addressing a significant pain point in nonparametric testing where manual parameter selection often leads to loss of power.

Experimental Results

Simulation Studies

The PLR test was evaluated against KS, MMD, and Anderson-Darling tests across various scenarios, including Gaussian mixtures, bimodal distributions, and asymmetric Beta distributions. In almost every case, the PLR test reached a statistical power of 1.0 with significantly smaller sample sizes than its competitors, particularly when dealing with non-Gaussian shapes.

Real-World Biological Applications

The practical utility of the PLR test was demonstrated on two high-dimensional biological datasets:

  • Metagenomic Analysis (Type II Diabetes): In a study of 2,450 microbial species, the PLR test identified 101 significant species, whereas the KS test found only 4 and the MMD test found 13. The species identified by PLR often exhibited bimodal distributions that traditional tests failed to capture.
  • Cancer Genomics (CLL): Analyzing 18,863 genes in Chronic Lymphocytic Leukaemia patients, the PLR test identified 1,071 differentially expressed genes, vastly outperforming KS (275) and MMD (412). Notably, the PLR test flagged known biological targets like TGFB2 and DTX1 that were overlooked by other methods.

Discussion and Conclusion

By transforming the two-sample comparison into an interaction test within a smoothing spline ANOVA framework, this research provides a powerful new tool for nonparametric statistics. The PLR test's ability to automatically adapt to the shape of the density makes it uniquely effective for modern, complex data.

While the method currently relies on data splitting for parameter tuning—which can be a limitation for very small sample sizes—its theoretical minimax optimality and superior performance in high-dimensional biological applications mark a significant advancement. Future work will explore extending this framework to $k$-sample tests and conditional independence testing.

Summary updated

Abstract

We consider the problem of comparing probability densities between two groups. A new probabilistic tensor product smoothing spline framework is developed to model the joint density of two variables. Under such a framework, the probability density comparison is equivalent to testing the presence/absence of interactions. We propose a penalized likelihood ratio test for such interaction testing and show that the test statistic is asymptotically chi-square distributed under the null hypothesis. Furthermore, we derive a sharp minimax testing rate based on the Bernstein width for nonparametric two-sample tests and show that our proposed test statistics is minimax optimal. In addition, a data-adaptive tuning criterion is developed to choose the penalty parameter. Simulations and real applications demonstrate that the proposed test outperforms the conventional approaches under various scenarios.
BibTeX
@misc{xing2019minimax,
  author = {Xin Xing and Zuofeng Shang and Pang Du and Ping Ma and Wenxuan Zhong and Jun S. Liu},
  title = {{Minimax Nonparametric Two-sample Test under Smoothing}},
  year = {2019},
  eprint = {1911.02171},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1911.02171},
}
Figure