Scalable Function-on-Scalar Quantile Regression for Densely Sampled Functional Data

Yusha Liu, Meng Li, Jeffrey S. MorrisarXiv 2020 · DOI · arXiv

Abstract

Functional quantile regression (FQR) is a useful alternative to mean regression for functional data as it provides a comprehensive understanding of how scalar predictors influence the conditional distribution of functional responses. In this article, we study the FQR model for densely sampled, high-dimensional functional data without relying on parametric error or independent stochastic process assumptions, with the focus on statistical inference under this challenging regime along with scalable implementation. This is achieved by a simple but powerful distributed strategy, in which we first perform separate quantile regression to compute $M$-estimators at each sampling location, and then carry out estimation and inference for the entire coefficient functions by properly exploiting the uncertainty quantification and dependence structure of $M$-estimators. We derive a uniform Bahadur representation and a strong Gaussian approximation result for the $M$-estimators on the discrete sampling grid, leading to dimension reduction and serving as the basis for inference. An interpolation-based estimator with minimax optimality is proposed, and large sample properties for point and simultaneous interval estimators are established. The obtained minimax optimal rate under the FQR model shows an interesting phase transition phenomenon that has been previously observed in functional mean regression. The proposed methods are illustrated via simulations and an application to a mass spectrometry proteomics dataset.

Introduction

In many scientific domains, such as proteomics, genomics, and neuroimaging, data is increasingly collected as functions or curves rather than scalar points. Functional Data Analysis (FDA) typically focuses on functional mean regression, which models the relationship between predictors and the average response curve. However, the "average" often fails to capture critical biological or physical signals that reside in the tails of the distribution. For instance, in mass spectrometry, certain biomarkers may only be significantly elevated in a small subset of a clinical cohort, making them invisible to mean-based analysis but prominent in high-quantile regression.

Functional Quantile Regression (FQR) offers a more granular view, but its application to high-dimensional, densely sampled data—where thousands of measurements are taken per curve—presents significant computational and theoretical hurdles. Existing methods often rely on restrictive assumptions, such as independent errors across time points or specific parametric likelihoods. This paper introduces a scalable framework that overcomes these limitations, providing both computational efficiency and rigorous statistical inference.

Methodology: A Distributed Strategy

The authors propose a "divide and conquer" approach to handle the high dimensionality of dense functional data. The framework consists of three primary stages:

1. Pointwise Estimation

Rather than attempting to fit a complex, high-dimensional model simultaneously across the entire functional domain, the method performs separate quantile regressions at each discrete sampling location $t$. This generates a collection of $M$-estimators. Because these regressions are independent of one another at this stage, the process is highly parallelizable and computationally efficient.

2. Interpolation and Smoothing

To reconstruct the continuous coefficient functions from the discrete pointwise estimates, the authors employ an interpolation-based estimator. By using linear or spline interpolation to "connect" the $M$-estimators, they produce a continuous function that represents the predictor's effect across the entire domain.

3. Statistical Inference via Gaussian Approximation

The theoretical core of the paper involves proving that these pointwise estimates, when viewed collectively, behave like a Gaussian process. The authors derive a uniform Bahadur representation and a strong Gaussian approximation for the $M$-estimators on the sampling grid. This allows for the construction of simultaneous confidence bands using a weighted bootstrap method. These bands are essential for researchers to determine which regions of the functional response are statistically significant.

Theoretical Insights: The Phase Transition

A key contribution of this work is the identification of a "phase transition" phenomenon in FQR. The accuracy of the functional estimator is governed by the relationship between the number of subjects ($n$) and the number of sampling points ($T$).

  • Dense Regime: When $T \gg n^{1/2r}$ (where $r$ relates to the smoothness of the function), the estimation error is dominated by the number of subjects. In this regime, sampling more points per curve does not significantly improve accuracy.
  • Sparse Regime: When $T$ is small, the sampling frequency becomes the bottleneck for convergence.

This finding aligns FQR theory with previously observed patterns in functional mean regression, suggesting a fundamental structural similarity in how information is recovered from functional data.

Empirical Results and Applications

Computational Scalability

The proposed method demonstrates superior scalability compared to existing Bayesian FQR approaches. In simulations with $n=8,000$ subjects and $T=512$ sampling points, the distributed framework completed the analysis in approximately 4 hours. In contrast, a leading Bayesian competitor was projected to require over 36 hours. The linear scaling with respect to $T$ makes this method uniquely suited for modern, high-resolution sensors.

Mass Spectrometry Proteomics

The utility of the method was demonstrated using a proteomics dataset consisting of blood serum samples from 139 pancreatic cancer patients and 117 healthy controls. Each sample contained $T = 3,279$ spectral locations.

The FQR analysis revealed that the differences between the cancer and control groups were significantly more pronounced at the 90th percentile ($\tau = 0.9$) than at the median. This suggests that certain proteins are "aberrantly expressed" only in a subset of the cancer cohort. By identifying these high-quantile differences, the model flagged candidate biomarkers that would have been overlooked by standard mean regression.

Conclusion

This research provides a robust and scalable framework for Functional Quantile Regression on dense data. By bridging the gap between simple pointwise estimation and rigorous functional inference through Gaussian approximation, the authors enable the discovery of signals in the tails of functional distributions. While interpolation is theoretically optimal, future work may explore more adaptive smoothing techniques, such as Bayesian priors or wavelets, to better handle spiky functional data.

Summary updated

Abstract

Functional quantile regression (FQR) is a useful alternative to mean regression for functional data as it provides a comprehensive understanding of how scalar predictors influence the conditional distribution of functional responses. In this article, we study the FQR model for densely sampled, high-dimensional functional data without relying on parametric error or independent stochastic process assumptions, with the focus on statistical inference under this challenging regime along with scalable implementation. This is achieved by a simple but powerful distributed strategy, in which we first perform separate quantile regression to compute $M$-estimators at each sampling location, and then carry out estimation and inference for the entire coefficient functions by properly exploiting the uncertainty quantification and dependence structure of $M$-estimators. We derive a uniform Bahadur representation and a strong Gaussian approximation result for the $M$-estimators on the discrete sampling grid, leading to dimension reduction and serving as the basis for inference. An interpolation-based estimator with minimax optimality is proposed, and large sample properties for point and simultaneous interval estimators are established. The obtained minimax optimal rate under the FQR model shows an interesting phase transition phenomenon that has been previously observed in functional mean regression. The proposed methods are illustrated via simulations and an application to a mass spectrometry proteomics dataset.
BibTeX
@misc{liu2020scalable,
  author = {Yusha Liu and Meng Li and Jeffrey S. Morris},
  title = {{Scalable Function-on-Scalar Quantile Regression for Densely Sampled Functional Data}},
  year = {2020},
  eprint = {2002.03355},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.03355},
}

Loading related content...

Figure