Optimal Change Point Detection and Inference in the Spectral Density of General Time Series Models

Sepideh Mosaferi, Department of Mathematics, Statistics · 2025

Read Optimal Change Point Detection and Inference in the Spectral Density of General Time Series Models by Sepideh Mosaferi, Department of Mathematics, Statistics (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-23211.txt.

Introduction

Sepideh Mosaferi, Abolfazl Safikhani, and Peiliang Bai detect change points in the spectral density of time series in two thousand twenty-five. Their drive comes from EEG reads of seizure patients. Seizures break coherence and links across brain zones. Sharp timing of that break is a clinical need.

Their method rests on the Wold form, which writes a series as an endless lag process. That process is cut and fit near each candidate change point. A first scan hunts across time in a set order. A refined step hits the best rate and gives trust bands. Tests on EEG and video beat past methods.

Catalog record

CREATORSepideh Mosaferi, Department of Mathematics, Statistics
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE148 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-23211

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Optimal Change Point Detection and Inference in the Spectral Density of General Time Series Models Sepideh Mosaferi, Department of Mathematics, Statistics Abstract This paper addresses the problem of detecting change points in the spectral density of time series, motivated by EEG analysis of seizure patients. Seizures disrupt coherence and functional connectivity, necessitating precise detection. Departing from traditional parametric approaches, we utilize the Wold decomposition, representing general time series as autoregressive processes with infinite lags, which are truncated and estimated around the change point. Our detection procedure employs an initial estimator that systematically searches across time points. We examine the localization error and its dependence on time series properties and sample size. To enhance accuracy, we introduce an optimal rate method with an asymptotic distribution, facilitating the construction of confidence intervals. The proposed method effectively identifies seizure onset in EEG data and extends to event detection in video data. Comprehensive numerical experiments demonstrate its superior performance compared to existing techniques. Keywords: Confidence interval; spectral density functions; optimal detection; Wold representation; Yule-Walker equation. Introduction Change point detection is a critical field in time series analysis, focused on identifying points in data where statistical properties, such as mean, variance, or distribution, change abruptly ( Basseville and Nikiforov, 1993 ) . These change points often indicate significant shifts in the underlying processes, which are important for various applications in fields. For example, in neuroscience, change point detection helps identify transitions in brain activity, such as shifts in neural dynamics during cognitive tasks or disease progression ( Safikhani and Shojaie, 2022 ) . In finance, detecting shifts in market trends can be essential for identifying regime changes and risk assessment ( Frisén, 2008 ) . Also, in signal processing, it is used to detect abrupt changes in sensor data, such as anomalies in speech or sound signals ( Foggia et al., 2015 ) . Finally, in manufacturing, it can help identify faults or quality control problems by detecting changes in production processes ( Qiu, 2013 ) . Detection algorithms can be broadly categorized into online/sequential and offline approaches. In the online/sequential setting, change points must be detected in real-time as data are observed sequentially ( Wang and Mei, 2015 ; Chan et al., 2021 ) . In contrast, the offline approach involves a retrospective analysis of the entire data set to identify all change points ( Yu, 2020 ) . In this paper, we focus on the offline detection problem, which is particularly relevant for applications requiring a comprehensive post hoc analysis of structural changes in time series data. There are various methods in the literature (CUSUM-based, fused lasso, dynamic programming, PCA-based, rolling window-based, etc.) to perform offline change point detection in univariate or high-dimensional data across different statistical models. Many of these methods are mainly focused on detecting mean changes. Examples include pruned exact linear time (PELT) method ( Killick et al., 2012 ) which uses optimal partitioning approach of Jackson et al. (2005) and a pruning step within the dynamic program to detect the structural breaks, Fryzlewicz (2018) that apply a tail-greedy Haar transformation to consistently estimate the number and locations of multiple change points in the univariate piecewise-constant model, Inspect method ( Wang and Samworth, 2018 ) which proposes a high-dimensional change-point detection method that uses a sparse projection to project the high-dimensional data into a univariate case while Aue et al. (2018) develop a method based on the (scaled) functional cumulative sum (CUSUM) statistic for detecting shifts in the mean of a functional data model ( Wang et al., 2020 ; Fryzlewicz, 2014 ) . While some other methods accommodate a broader range of change types (e.g. changes in regression coefficients), they often assume a parametric structure, which can be limiting in certain contexts ( Pein et al., 2017 ; Bai and Safikhani, 2023 ; Kolar and Xing, 2012 ; Aue et al., 2009 ; Avanesov and Buzun, 2018 ) . Additionally, more recent methods have emerged that are capable of handling changes in temporal dependence, which manifest as shifts in spectral density ( Davis et al., 2006 ; Safikhani and Shojaie, 2022 ; Cho and Fryzlewicz, 2015 ; Cho, 2016 ; Wang et al., 2019 ; Bai et al., 2023 ) . These methods offer a more flexible modeling framework while focusing on detecting changes. However, they typically do not extend to providing inference on the location of the change points, which is a critical aspect for a comprehensive analysis of temporal dynamics. Providing inference for change point locations has become a major focus in the recent literature. For example, Xu et al. (2024) performs estimation and inference for change point locations and allows the covariate and noise sequences to be temporally dependent in the functional dependence framework, but the changes are only in the regression coefficients, and it is unclear whether their method can handle changes in the temporal dependence. Kaul et al. (2021) and Kaul and Michailidis (2025) provide inference for multiple change points in mean shift models, and thus cannot handle changes in spectral density ( Wang and Shao, 2023 ) . Further, Madrid Padilla et al. (2021) provides inference for change points in multivariate data, but focuses on independent observations. More recently, Madrid Padilla et al. (2023) and Padilla et al. (2021) provide detection and inference in multivariate non-parametric models. However, their detection method focuses on estimating marginal distributions of data, and thus, it is unclear whether they are suitable for change point inference due to changes in temporal dependence/spectral density ( Chen et al., 2022 ; Fryzlewicz, 2024 ) . To the best of our knowledge, there are no methods in the literature designed for detection and inference for the spectral density of general time series models. Thus, there remains a gap in the literature that our paper aims to address by developing a method that not only accommodates changes in the temporal dependence through different spectral densities but also offers inference on the location of the change point. In this article, we investigate the problem of change point detection in the spectral density of time series models, motivated by the analysis of EEG data from patients undergoing seizures. Seizure events are known to alter the coherence and functional connectivity, making precise detection critical for understanding brain dynamics ( Safikhani and Shojaie, 2022 ) . While existing methods in the literature primarily assume specific parametric forms before and after the change point, we take a nonparametric approach by leveraging the well-known Wold decomposition. This allows us to represent the pre- and post-change time series models as autoregressive processes with infinite lags ( Wang and Politis, 2021 ) . The core idea is to appropriately truncate these infinite-lag autoregressive models and estimate model parameters before and after the change point. Our detection procedure begins with an initial estimation of the change point location by systematically searching over all possible time points, while excluding boundary regions as per standard practice in change point analysis. We theoretically establish the localization error for this estimator, linking its non-asymptotic rate to key characteristics of the time series before and after the change point, as well as the sample size (Theorem 3.1 ). To further enhance detection accuracy, we develop an

Read the complete text