Asymmetric Huber Periodogram
Tianbo Chen · 2025
Read Asymmetric Huber Periodogram by Tianbo Chen (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-2510-25316.txt.
Introduction
Tianbo Chen and Xiaojun Song introduce an asymmetric Huber periodogram. Their 2025 statistics paper generalizes several existing periodograms. A periodogram helps find hidden cycles in time series. The ordinary version uses squared error loss. Quantile and Huber versions offer other trade-offs. Their new method aims for a fuller picture.
They build it through asymmetric Huber regression with trigonometric terms. A special loss replaces the usual squared loss. Scanning an asymmetry parameter shows more data structure. The method can be cheaper than quantile versions. It also resists outliers better than ordinary versions. Chen and Song establish its theoretical properties. They link it to an asymmetric Huber spectrum. They develop confidence intervals and a Fisher-type test. Simulations and two applications show its value. Uses include detecting periodicities and clustering time series.
Catalog record
| CREATOR | Tianbo Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-25316
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Asymmetric Huber Periodogram Tianbo Chen Abstract This paper introduces a novel spectral M-estimator, called the asymmetric Huber periodogram (AHP), for periodicity detection in time series. The AHP is constructed from trigonometric asymmetric Huber regression, where a specially designed check function is used to substitute the squared ℓ 2 \ell_{2} norm that defines the ordinary periodogram (PG). The AHP is statistically more efficient than the quantile periodogram (QP), while offering a more comprehensive picture than the Huber periodogram (HP) by examining the data across the entire range of the asymmetric parameter. We prove the theoretical properties of the AHP and investigate the relationship between the AHP and the so-called asymmetric Huber spectrum (AHS). Finally, simulations and three real-world data examples demonstrate that the AHP’s capability in detecting periodicity and its robustness against outliers. Keywords : Asymmetric Huber regression; Periodicity detection; Periodogram; Robustness; Spectral analysis. 1 Introduction The PG, a raw non-parametric estimator of the power spectrum, plays a crucial role in spectral analysis and is widely applied across various fields ( Caiado et al., 2006 ; Polat and Güneş, 2007 ; Baud et al., 2018 ; Euán et al., 2018 ; Maadooliat et al., 2018 ; Martínez-Murcia et al., 2019 ; Chen et al., 2021a ) . The PG is constructed by trigonometric ordinary least squares (OLS) regression. However, OLS exhibits limitations in terms of robustness and efficiency, particularly in handling data with asymmetric or heavy-tailed distributions. Moreover, OLS regression focuses on the conditional mean, lacking the ability to provide a complete picture of the distribution of the data, which subsequently affects the performance of the PG ( Bloomfield, 2004 ) . As a result, alternative regression techniques have been proposed to overcome the issues of OLS. The pioneering work of Koenker and Bassett Jr, 1978 introduced the concept of quantile regression (QR). By substituting the \ell_{2} loss with asymmetric \ell_{1} loss, QR provides a more complete picture of the relationship between the response variable and the covariates across different quantile levels, while also demonstrating strong robustness against outliers and heavy-tailed distributions. For a detailed and systematic introduction to quantile regression and its extensions, we refer to Kouretas et al., 2005 ; Cai and Xu, 2008 ; Cai and Xiao, 2012 ; Koenker, 2017 . However, QR is not free of limitations: the non-smooth check (loss) function causes both theoretical and computational burdens, and it suffers efficiency losses under Gaussian or homoscedastic errors. Based on this, asymmetric least squares (ALS) regression, also known as expectile regression (ER), was proposed by Newey and Powell, 1987 . The quadratic form of the loss function partly addresses the limitations inherent in QR, but this improvement comes at the expense of a certain degree of robustness. A comprehensive comparative analysis of quantiles and expectiles is presented in Waltrup et al., 2015 , wherein the relationships between these two approaches are thoroughly examined. Further theoretical analysis between QR and ER can be found in Efron, 1991 ; Jones, 1994 ; Yao and Tong, 1996 . Because of their advantages, QR and ER, as well as their extensions, have been widely used across various scientific fields ( Garcia et al., 2001 ; Machado and Mata, 2005 ; Gu and Zou, 2016 ; Ziegel, 2016 ; Bellini and Di Bernardino, 2017 ; Jiang et al., 2017 ; Daouia et al., 2018 ; Xu et al., 2020 ; Alvarado et al., 2021 ; Sharif et al., 2021 ; Xu et al., 2021 ) . As a generalization of both QR and ER, Jiang et al., 2021 introduced the k -th power expectile regression (KER) with 1<k\leq 2 . With these regression techniques, researches have developed the corresponding periodograms. One innovative development is the QP ( Li, 2012b ) constructed by trigonometric QR, which demonstrates its ability to detect hidden periodicities in time series. Similar to ordinary spectral analysis, large-sample analysis reveals that the QP is associated with the so called quantile spectrum, which is a scaled version of the ordinary power spectrum of the level-crossing process. The Laplace periodogram (LP) is a special case of the QP by setting the quantile to 0.5 . Related works on the QP include Li, 2012a ; Hagemann, 2013 ; Li, 2014 ; Dette et al., 2015 ; Kley, 2016 ; Birr et al., 2017 ; Meziani et al., 2020 ; Chen et al., 2021b ; Li, 2023 . Chen, 2024 further proposed the EP, and comprehensively compared it with the QP. For more robust periodograms, Wen et al., 2021 proposed the HP and Thieler et al., 2016 developed the R package RobPer , which allows users to combine different regression techniques ( Rousseeuw and Yohai, 1984 ; Oh et al., 2004 ; Ahdesmäki et al., 2007 ; Thieler, 2014 ) to calculate robust periodograms. Besides KER, the asymmetric Huber regression (AHR) ( Allende et al., 2006 ) also strikes a balance between robustness and efficiency using piecewise linear and quadratic losses, with a threshold parameter to control the ratio of the two losses. In addition to the original Huber loss ( Huber, 1964 ) , the AHR uses an asymmetric parameter that similar to quantile and expectile, to comprehensively analyze the distribution of the data. In this paper, we apply asymmetric Huber trigonometric regression and propose a spectral estimator called AHP. We demonstrate that the AHP not only shares similar properties with the PG, LP, and HP as a frequency-domain representation of serial dependence in time series, but also provides a more comprehensive analysis by exploring the data across different asymmetric parameters. The remainder of the paper is organized as follows. We first introduce the AHP in Section 2 . The large sample theory revealing the asymptotic distribution of the AHP, which is presented in Section 3 . In Section 4 , we present comparative studies evaluating the performance of different types of periodograms by simulation studies. In Section 5 , we apply the AHP to three real data examples, demonstrating its ability to detect hidden periodicities and its robustness against outliers. Finally, we conclude the paper and list the limitations and possible future work in Section 6 . 2 Asymmetric Huber Periodogram Suppose that \{y_{t}\} is a stationary process with distribution function F(\cdot) . For any fixed asymmetric parameter \alpha\in(0,1) and threshold parameter \psi>0 , define the asymmetric Huber loss 1 1 1 Note that some literature, such as Gupta et al., 2020 and Hazarika et al., 2023 , uses a different version, in which the loss function is symmetric when |u|\leq\psi . as \rho_{\alpha,\psi}(u)=\begin{cases}|\alpha-I(u<0)|\cdot\frac{1}{2}u^{2},&|u|\leq\psi,\\ |\alpha-I(u<0)|\cdot\psi(|u|-\frac{1}{2}\psi),&|u|>\psi.\end{cases} In the special case of \psi\to 0 and \psi\to\infty , the loss function reduces to the check loss function of quantile regression and the asymmetric quadratic loss function of expectile regression, respectively. The \alpha -th asymmetric Huber quantile (AHQ) of y_{t} is defined as \mu(\alpha,\psi):=\arg\min_{\mu\in\mathbb{R}}\mathrm{E}\left\{\rho_{\alpha,\psi}(y_{t}-\mu)\right\}, (1) with the normal equation \mathrm{E}\left\{\dot{\rho}_{\alpha,\psi}(y_{t}-\mu(\alpha,\psi))\right\}=0. (2) Given a time series \{y_{t}:t=1,\ldots,n\} and regressors \mathbf{x}_{nt}(\omega_{\nu}):=[\cos(\omega_{\nu}t),\sin(\omega_{\nu}t)]^{\top} , the AHR solves the optimization problem: \widehat{\bm{\beta}}_{n}(\alpha,\psi):=\arg\min_{\bm{\beta}\in\mathbb{R}^{2}}\sum_{t=1}^{n}\rho_{\alpha,\psi}(y_{t}-\mu-\mathbf{x}_{nt}^{\top}\bm{\beta}), (3) where \omega_{\nu} denotes the Fourier frequencies in (0,\pi) , \widehat{\bm{\beta}}_{n}(\omega_{\nu},\alpha,\psi):=[\widehat{\beta}_{1}(\omega_{\nu},\alpha,\psi),\widehat{\beta}_{2}(\omega_{\nu},\alpha,\psi)]^{\top}, and \mu is a suitable constant, typically the sample \alpha -th AHQ of y_{t} . Then, the Asymmetric Huber discrete Fourier transform (AHDFT) of \{y_{t}:t=1,\ldots,n\} at level \alpha is defined as z_{n}(\omega_{\nu},\alpha,\psi):=\frac{n}{2}\{\hat{\beta}_{1}(\omega_{\nu},\alpha,\psi)-i\hat{\beta}_{2}(\omega_{\nu},\alpha,\psi)\}, where i:=\sqrt{-1} . The