Bounds in Wasserstein Distance for Locally Stationary Functional Time Series
Jan Nino G. Tinio, Mokhtar Z. Alaya, Salim Bouzebda · 2025
Read Bounds in Wasserstein Distance for Locally Stationary Functional Time Series by Jan Nino G. Tinio, Mokhtar Z. Alaya, Salim Bouzebda (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-2504-06453.txt.
Introduction
Jan Nino Tinio and co-authors study a 2025 problem in statistics: estimating how a curve-valued series behaves when its pattern drifts slowly over time. Functional time series record whole curves, not single numbers, and their dependence over time is hard to capture. Locally stationary models help by letting mean and variance shift in a smooth way.
The team examines a Nadaraya-Watson estimator for the conditional distribution of such series, where inputs live in a semi-metric space. Under set conditions on mixing and small ball probability, they prove convergence rates measured in Wasserstein distance. Wide numerical tests on simulated curves and real data show how the method behaves in finite samples. The work gives practitioners firmer ground for forecasting with slowly changing curve data.
Catalog record
| CREATOR | Jan Nino G. Tinio, Mokhtar Z. Alaya, Salim Bouzebda |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 131 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-06453
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Bounds in Wasserstein Distance for Locally Stationary Functional Time Series Jan Nino G. Tinio, Mokhtar Z. Alaya, Salim Bouzebda Abstract Functional time series (FTS) extend traditional methodologies to accommodate data observed as functions/curves. A significant challenge in FTS consists of accurately capturing the time-dependence structure, especially with the presence of time-varying covariates. When analyzing time series with time-varying statistical properties, locally stationary time series (LSTS) provide a robust framework that allows smooth changes in mean and variance over time. This work investigates Nadaraya-Watson (NW) estimation procedure for the conditional distribution of locally stationary functional time series (LSFTS), where the covariates reside in a semi-metric space endowed with a semi-metric. Under small ball probability and mixing condition, we establish convergence rates of NW estimator for LSFTS with respect to Wasserstein distance. The finite-sample performances of the model and the estimation method are illustrated through extensive numerical experiments both on functional simulated and real data. \DefineFNsymbolsTM myfnsymbols Keywords: Conditional distribution estimation; Locally stationary functional time series; Mixing condition; Nadaraya-Watson estimation; Wasserstein distance 1 Introduction In recent years, data collected across various fields increasingly exhibit functional or curve-like characteristics, commonly referred to as functional data whose values come from infinite-dimensional space. This advancement is driven by the proliferation of data collected on progressively refined temporal and spatial grids, for instance, in meteorology, medicine, satellite imagery, economics and finance, environmental science, and many others ( 65 , 16 , 18 ) . Specifically, [ 25 ] demonstrated intriguing applications in biometrics, [ 4 ] explored the relevance of functional data to environmental science, and [ 23 , 3 ] focused on applications in econometrics. The statistical modeling of such data, conceptualized as random functions/curves, has given rise to various intricate theoretical and numerical research inquiries. Functional data are investigated for multiple reasons: they assist in developing data representation strategies that highlight significant features, analyzing mean behaviors and deviations, and constructing models when temporal dependency is present in the data ( 62 , 41 ) . The statistical challenges in studying these data have drawn increasing attention over the last few decades from the statistical community ( 20 , 19 , 21 ) . Much of the existing research in FTS analysis is predicated on stationary models since the stationarity requirement is often adopted in time series modeling, leading to various models and methodologies. Theoretical foundations of function spaces concerning linear processes were explored in [ 11 ] , which are crucial for modeling and forecasting functional time series. Statistical analysis of autoregressive processes within Banach spaces was emphasized in [ 12 ] , a specialized class of functional spaces. Similarly, [ 34 ] examined asymptotic properties of the sample mean and variance for autoregressive processes of order one in a separable Banach space with independent innovations. Extending this, [ 2 ] incorporated wavelet analysis alongside nonparametric kernel regression for FTS forecasting. However, many FTS exhibit nonstationarity. Even with detrending and deseasonalizing, the stationarity assumption is not always beneficial for modeling functional data. Many time series models, commonly observed in various physical phenomena and economic data, are non-stationary ( 48 , 65 , 16 ) . Conventional approaches become inappropriate when the (weak) stationarity assumption is violated. Specifically, global climate changes in meteorology affect the distribution of a region’s daily temperature, precipitation, and cloud cover when viewed interconnectedly. In finance, an option’s implied volatility evolves over time depending on its moneyness ( 68 ) . To address such nonstationarity, the locally stationary time series (LSTS) framework provides a practical modeling approach. The parameters of LSTS exhibit temporal dependence. These nonstationary processes can be approximated by locally stationary counterparts that remain stationary within smaller time windows. As a result, asymptotic theories can be developed to estimate time-dependent characteristics ( 26 , 29 , 28 ) . Many works have addressed the intuitive concept of local stationarity. The seminal works in [ 26 , 27 , 29 ] provide a strong foundation for the inference of LSTS. Parametric ( 29 , 42 , 28 ) and nonparametric ( 71 , 48 , 17 ) frameworks were developed to analyze LSTS. For instance, [ 26 ] introduced a parametric framework by leveraging local periodograms to minimize the generalized Whittle function tailored to locally stationary models. Expanding the parametric framework beyond time-invariant assumptions, [ 47 ] proposed a semiparametric model. This approach combined the strengths of parametric and nonparametric methods, allowing for greater flexibility in modeling complex time-varying structures without imposing stringent parametric constraints. Parallel to the parametric advancements, nonparametric approaches have garnered significant attention from researchers aiming to model conditional mean and variance functions without relying on predefined parametric forms. In particular, [ 48 ] developed an estimation theory for nonparametric regression problems involving LSFTS. Central to nonparametric approaches in the literature is Nadaraya-Watson (NW) ( 55 , 72 ) estimation procedure, which has been extensively used to estimate conditional mean functions. Despite the proven efficacy of NW procedure in estimating conditional mean functions, its usefulness extends to conditional distribution estimation (CDE). Estimating accurately conditional distributions is essential in prediction and forecasting, as it comprehensively describes the conditional law for any given random variable. Due to its critical role in predictive modeling and inference, numerous studies have focused on developing robust estimation theories for CDE within FTS framework ( 39 , 45 , 14 ) . Two distinct approaches for estimating the conditional distribution of a target variable within a prediction set, given a functional covariate, were introduced in [ 45 ] . The first method relies on the empirical distribution of the estimated model residuals, while the second involves fitting functional parametric models to the residuals. These approaches provide flexible frameworks for handling complex dependencies between functional covariates and target variables, particularly in high-dimensional settings where traditional parametric models may be inadequate. Additionally, [ 14 ] developed a local polynomial estimator for conditional CDF of a scalar target variable given a functional covariate in the context of stationary strongly mixing processes. They established asymptotic normality property. In dealing with CDE, optimal transport (OT) theory has emerged as a powerful mathematical framework for quantifying difference between probability distributions. OT measures the minimal cost required to transport one distribution to another, providing a meaningful metric for distributional comparison ( 59 ) . This approach addresses shortest-path problems by enabling the concurrent transportation of multiple items along geodesic curves or straight paths. Among various metrics derived from OT, Wasserstein distance stands out for its robustness and versatility in comparing probability distributions ( 70 , 50 , 66 ) . The landscape of LSTS analysis is enriched by parametric and nonparametric approaches, each contributing unique strengths to the modeling of time-evolving data. Parametric approaches offer structured and interpretable models for capturing dynamic behaviors through frameworks like those proposed in [ 26 , 47 ] . Concurrently, nonparametric techniques, particularly those leveraging NW estimator and OT theory, provide flexible and robust tools for estimating conditional means, variances, and distributions. The integration of these procedures, underpinned by rigorous theoretical advancements, continues to enhance the capacity of statisticians and data scientists to model, predict, and infer from complex, non-stationary time series data. Contributions. This paper investigates CDE for LSFTS characterized by weakly dependent sequences