📊 statistics

Adaptive Ridge-Regularized Hotelling Change-Point Tests for Functional Data

This paper proposes a unified, adaptive ridge-regularized Hotelling framework that effectively detects and locates mean changes in functional time series by combining spectral regularization with Cauchy-transform aggregation and wild binary segmentation to ensure robust performance under weak dependence, non-Gaussianity, and multiple change points.

Original authors: Ping Zhao, Long Feng

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ping Zhao, Long Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a river flow. Usually, the water moves in a smooth, predictable rhythm. But sometimes, something changes: a dam opens upstream, a new tributary joins, or the weather shifts, causing the current to suddenly speed up, slow down, or change direction. In the world of data science, this "river" is a stream of information that changes over time, like daily temperature curves, stock market trends, or brain wave patterns. Scientists call this functional time series because the data isn't just a single number; it's a whole shape or curve that evolves.

The big challenge is spotting the exact moment the river changes course. This is called finding a change point. If you look at a single day's temperature, it's just a line. But if you look at a year's worth of daily temperatures, you get a smooth, wiggly curve. When you have hundreds of these curves, one after another, you want to know: "Did the climate shift in 1972?" or "Did the market crash in 1982?" The problem is tricky because these curves are messy. They are often connected to each other (today's weather depends on yesterday's), and they have hidden layers of detail. Some parts of the curve might be very loud and obvious, while others are quiet whispers. If you try to listen to the whole river at once, the loud parts might drown out the quiet, important signals, or the noise might make you think a change happened when it didn't.

This paper introduces a clever new way to listen to that river. The authors, Ping Zhao and Long Feng, propose a method that acts like a high-tech, adjustable hearing aid. Instead of trying to hear every single drop of water (which is too much data), they break the curves down into a set of musical notes (mathematical building blocks). Then, they use a special "volume knob" called ridge regularization. Think of this knob as a filter that lets you turn down the loud, dominant notes so you can hear the quiet, subtle ones that might actually be the signal you're looking for.

The paper's main finding is that by combining many different settings of this volume knob, they can detect changes in the data much more reliably than previous methods, especially when the data is noisy or when the changes are subtle. They proved mathematically that their method works under a wide range of conditions, including when the data points are connected to each other (dependent) and when the data doesn't follow a perfect bell-curve shape. They also showed how to use this method to find multiple changes in a row, not just one.

To test their idea, they ran computer simulations where they knew exactly where the changes were hidden. Their new method, which they call RHT (Ridge-regularized Hotelling Test), found the changes faster and more accurately than older techniques, even when the "loud" parts of the data tried to hide the "quiet" changes. They also applied their method to real-world data: daily minimum temperatures in Sydney and sea surface temperatures in the Pacific Ocean. In both cases, their method pinpointed specific years when the climate patterns shifted, matching what we know about historical warming trends.

The authors are very confident in their results because they didn't just guess; they built a solid mathematical proof showing why their method works. They also showed that their method is flexible: it doesn't require the data to be perfectly clean or independent. However, they are careful to note that while their method is excellent for finding where a change happened, it doesn't explain why it happened. That's a job for climate scientists and economists, not just statisticians.

In short, this paper gives scientists a sharper, more adaptable tool for spotting when the "river" of data changes its flow. Whether it's tracking climate change, monitoring industrial machines, or understanding brain activity, knowing exactly when things shift is crucial. This new method ensures we don't miss the quiet shifts that older tools might have ignored.

Technical Summary: Adaptive Ridge-Regularized Hotelling Change-Point Tests for Functional Data

Problem Statement
The paper addresses the detection and localization of mean changes in weakly dependent functional time series. Unlike standard vector time series, functional data often exhibit intrinsic spectral decay, where the covariance operator is trace-class and eigenvalues diminish rapidly. This structure creates a specific challenge: a few low-frequency directions may carry a non-negligible proportion of total variation, while high-frequency directions have vanishing variances. Standard high-dimensional change-point methods, which often assume bounded spectra or trace-dispersion conditions where no single direction dominates, are ill-suited for this regime. Direct inversion of the covariance matrix is unstable, and unregularized quadratic statistics can be dominated by a small number of directions. Furthermore, the presence of serial dependence necessitates the estimation of a long-run covariance operator rather than a simple lag-zero covariance, which is complicated by the potential contamination of global mean estimators when structural breaks exist.

Methodology
The proposed framework, termed RHT (Ridge-Regularized Hotelling Test), integrates three core components:

Growing Basis Expansion: Functional observations are projected onto a deterministic basis (specifically a real Fourier basis) to generate high-dimensional score vectors. The dimension of this expansion, $p$ , grows with the sample size $n$ , preserving the full functional information without pre-selecting a fixed number of principal components.
Edge-Corrected Difference-Based Estimation: To handle serial dependence and avoid bias from mean shifts, the long-run covariance operator is estimated using a difference-based procedure. This method utilizes lagged products of differenced scores. Crucially, the authors introduce an "edge correction" by using the actual number of available lagged products in the denominator for each lag, removing a deterministic finite-sample attenuation that would otherwise propagate through the spectral estimation and ridge inversion. The estimator is further stabilized by retaining only its positive spectral part before inversion to ensure positive semi-definiteness.
Adaptive Ridge Regularization and Cauchy Aggregation: A family of ridge-regularized Hotelling CUSUM statistics is constructed over a deterministic grid of ridge parameters. Theoretical analysis reveals that the power-maximizing ridge parameter depends on the unknown spectral orientation of the mean change; no single fixed ridge is uniformly optimal. To address this, the method aggregates the marginal $p$ -values from the ridge grid using a Cauchy transform. This aggregation preserves cross-ridge dependence by utilizing a common set of weighted Brownian bridge paths for calibration, avoiding the need for resampling the original curves.

For multiple change-point detection, the framework embeds these local maximum-ridge statistics into a Wild Binary Segmentation (WBS) procedure. Random intervals are drawn to isolate individual changes, followed by recursive segmentation and local refinement. The global long-run covariance estimate is reused throughout the recursion to maintain efficiency.

Key Theoretical Results
Under mild conditions involving weak dependence (modeled via a Bernoulli-shift approximation) and finite moments, the paper establishes:

Null Limit Distribution: The limiting null distribution of the ridge-regularized CUSUM process is not Gaussian. Instead, it converges to a centered and standardized weighted sum of independent squared Brownian bridges. This non-Gaussian limit arises from the trace-class nature of the covariance operator and the persistence of leading directions.
Direct Spectral Calibration: The authors derive explicit critical values from the limiting weighted-bridge process. This allows for direct calibration without bootstrapping or permuting the original functional curves.
Local Power and Oracle Ridge: An explicit local power formula is derived, showing that the optimal ridge depends on the spectral coordinates of the change. The Cauchy aggregation is justified theoretically as a method to achieve robust power across unknown change directions without requiring prior knowledge of the signal structure.
Consistency and Localization: For single changes, the tests are shown to be consistent, and the change-point location is estimated with uniform convergence rates. For multiple changes, the WBS procedure is proven to consistently recover the number of changes and uniformly estimate their locations, provided the changes are well-separated.

Empirical Performance
Simulation studies compare the proposed RHT method against established competitors: the fully functional test (ARS), the functional principal component test (BGHK), and the self-normalized functional principal component test (SN).

Size Control: RHT maintains satisfactory empirical size across various covariance spectra and dimensions, including cases with functional autoregressive (FAR) errors. Competitors like BGHK exhibit liberal behavior under dependence, particularly with fast spectral decay.
Power: RHT demonstrates superior power, especially under finite-rank and fast-decay spectral settings where changes may load on low-variance directions poorly captured by principal component truncation.
Localization: The proposed method provides more accurate and stable change-point localization compared to competitors, particularly when the change direction is not aligned with the leading eigenfunctions.

Empirical Applications
The framework is applied to two temperature datasets:

Sydney Annual Minimum Temperatures: The method identifies a structural shift around 1971/1972, corresponding to a broad upward displacement in the annual temperature curve. Residual analysis confirms the presence of serial dependence in the annual level, justifying the use of long-run covariance estimation.
Niño 1+2 Sea-Surface Temperatures: The method detects a change around 1981/1982, characterized by a vertical displacement in the monthly temperature curves.

Significance and Claims
The paper claims to provide a unified framework that resolves the tension between dimension reduction, spectral estimation, and the utilization of full functional information in change-point analysis. Its primary contributions are:

Developing a basis-expansion formulation that accommodates intrinsic spectral decay via trace-scaled ridge regularization.
Establishing consistency rates for an edge-corrected difference-based estimator of the functional long-run covariance under weak dependence without requiring Gaussianity or specific linear process representations.
Deriving a non-Gaussian weighted Brownian-bridge-square null limit and providing direct spectral critical values, eliminating the need for computationally intensive resampling of functional curves.
Justifying the Cauchy aggregation over a ridge grid as a theoretically sound method to handle the unknown spectral orientation of changes.
Demonstrating that the proposed multiple-change procedure consistently recovers the number and locations of changes under the same finite-moment weak-dependence framework used for single changes.

The authors note that the method accommodates weak dependence and non-Gaussian errors, and the simulation results suggest favorable finite-sample performance. They acknowledge that the multiple-change theory currently assumes a fixed number of proportionally separated changes, leaving the extension to diverging numbers of changes as a direction for future work.

More like this