arxiv: 2605.04968 · v1 · submitted 2026-05-06 · 📊 stat.ME · math.ST· stat.TH

Recognition: unknown

Tests for white noise via asymptotically independent U-statistics in high-dimensions

Yuanya Xu

Authors on Pith no claims yet

Pith reviewed 2026-05-08 16:33 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.TH

keywords high-dimensional white noise testU-statisticsample autocovariancesasymptotic normalitymartingale differencecross-sectional dependencegraph structureserial correlation

0 comments

The pith

A U-statistic from autocovariances tests for white noise in high-dimensional series without assuming cross-sectional independence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a test for white noise in a vector time series with many components by constructing a U-statistic from the sample autocovariances at different lags. This statistic is designed to detect any form of serial correlation, both within individual series and across different series, without requiring the user to specify what the alternative looks like. Under the null hypothesis that the series is white noise, the test statistic is shown to be asymptotically normal when both the dimension p and the time span T grow large, with the proof relying on martingale difference theory. The analysis accommodates dependence across the p series by connecting it to an underlying graph structure and imposes only spectral conditions on the covariance matrix rather than independence. A reader would care because many applied datasets in finance, neuroscience, or sensor networks have both high dimension and possible cross-variable dependence, so a test that remains valid in this regime enables more reliable diagnostics than methods that force strong independence assumptions.

Core claim

We propose a high-dimensional white noise test that captures serial correlations within and across component series without specifying an alternative model. The test statistic is a U-statistic based on sample autocovariances. Under the null, asymptotic normality is established as p, T → ∞ jointly using martingale difference theory. Our approach imposes no cross-sectional independence assumption, requiring only spectral conditions on Σ₀. Theoretically, we link cross-sectional correlations to a graph structure, integrating algebraic and geometric analyses to facilitate the derivation.

What carries the argument

U-statistic constructed from sample autocovariances, whose asymptotic normality under the null is derived via martingale difference theory after linking cross-sectional correlations to a graph structure.

If this is right

The test maintains reliable size control in simulations for a range of (p, T) combinations.
The test exhibits satisfactory power against alternatives without any need to specify their form.
The procedure applies directly to data in which the component series are cross-sectionally dependent, provided the spectral conditions hold.
The graph structure supplies an algebraic-geometric route to control the dependence terms that appear in the variance calculation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same martingale framework could be reused to derive tests for white noise after fitting high-dimensional factor or graphical models to the data.
The graph representation of cross-sectional correlations might be turned into a diagnostic tool that flags which pairs of series contribute most to any detected dependence.
Because the proof avoids specifying alternatives, the statistic could serve as a general-purpose residual check in large multivariate autoregressive or volatility models.

Load-bearing premise

Spectral conditions on the covariance matrix Σ₀ plus regularity on the dependence structure suffice for the martingale difference argument and graph integration to deliver joint normality as p and T grow.

What would settle it

A concrete high-dimensional white noise process with p and T both large, cross-sectional correlations obeying the spectral conditions, yet whose normalized U-statistic fails to converge in distribution to standard normal.

read the original abstract

We propose a high-dimensional white noise test that captures serial correlations within and across component series without specifying an alternative model. The test statistic is a U-statistic based on sample autocovariances. Under the null, asymptotic normality is established as $p, T \to \infty$ jointly using martingale difference theory. Our approach imposes no cross-sectional independence assumption, requiring only spectral conditions on $\Sigma_0$. Theoretically, we link cross-sectional correlations to a graph structure, integrating algebraic and geometric analyses to facilitate the derivation. Simulations confirm reliable size control and satisfactory power across various $(p, T)$ settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper gives a workable high-dimensional white noise test via U-statistics on autocovariances that handles cross-sectional dependence through a graph model and martingale CLT without independence assumptions.

read the letter

The main point is that this paper gives a U-statistic test for white noise in high dimensions. It uses sample autocovariances and derives asymptotic normality as p and T grow together through martingale difference theory. The test works without assuming cross-sectional independence by imposing spectral conditions on the covariance and modeling the cross-correlations with a graph structure that combines algebraic and geometric arguments. The paper does well in avoiding a common restrictive assumption. Many existing high-dimensional tests require independence across series, but this one links the correlations to a graph and controls the variance accordingly. The simulations back this up with reliable size and satisfactory power in different p and T settings. The argument looks internally consistent. The spectral conditions are explicit, and the graph analysis feeds directly into the Lindeberg and variance steps without gaps. I see no load-bearing circularity or unstated dependencies. One soft spot is that the power results come from post-hoc simulation choices rather than a broad range of alternatives, though this is typical and does not affect the null distribution result. The conditions on the graph might also need checking in applications, but the paper keeps the focus on theory. This is useful for statisticians working on multivariate time series testing where cross-dependence cannot be ignored. A reader interested in high-dimensional asymptotics will get a clear extension of U-statistic methods. It deserves a serious referee. The math checks out and the contribution is targeted and usable.

Referee Report

0 major / 3 minor

Summary. The manuscript proposes a high-dimensional white noise test for multivariate time series that detects serial correlations both within and across component series without specifying an alternative. The test statistic is constructed as a U-statistic from sample autocovariances. Under the null of white noise, asymptotic normality is derived as p and T tend to infinity jointly, using martingale difference central limit theory. The method requires only spectral conditions on the covariance matrix Σ₀ and handles cross-sectional dependence by associating it with an algebraic-geometric graph structure whose properties are integrated into the variance calculations and Lindeberg condition.

Significance. If the central derivations hold, the paper offers a technically coherent extension of white-noise testing to the joint high-dimensional regime without cross-sectional independence assumptions. The explicit use of martingale theory for the U-statistic and the graph-theoretic control of dependence are strengths that could make the test applicable to settings such as high-dimensional financial or neuroimaging series. The absence of free parameters in the limiting null distribution and the provision of simulation evidence for size and power are additional positive features.

minor comments (3)

[§3.2] §3.2: the precise statement of the spectral condition on Σ₀ (eigenvalue bounds or decay rate) should be restated immediately before the variance formula to make the dependence on the graph geometry transparent.
[Table 1] Table 1: the reported empirical sizes for p=100, T=200 under the AR(1) cross-sectional case appear slightly conservative; a brief remark on whether this is due to the graph-diameter term would be helpful.
[Appendix] The proof of the Lindeberg condition in the appendix relies on the maximum degree of the dependence graph; a short sentence in the main text linking this quantity to the spectral radius of Σ₀ would improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and accurate summary of our manuscript, as well as for recommending minor revision. The referee correctly identifies the key technical elements, including the U-statistic construction, the use of martingale difference CLT for joint high-dimensional asymptotics, and the graph-theoretic handling of cross-sectional dependence without requiring independence. Since the report lists no specific major comments, we have no points requiring rebuttal or revision at this stage. We remain available to incorporate any minor suggestions or clarifications in a revised version.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines the test statistic directly as a U-statistic constructed from sample autocovariances of the observed series. Asymptotic normality under the null is obtained via martingale-difference central limit theory under the joint (p,T)→∞ regime, with variance terms controlled by explicitly stated spectral conditions on Σ₀. The algebraic-geometric graph analysis is introduced as an auxiliary device to handle cross-sectional dependence within the Lindeberg and variance-stabilization arguments; it does not presuppose the target result. No parameter is fitted to a data subset and then relabeled as a prediction, no self-citation supplies a uniqueness theorem or ansatz that the present derivation relies upon, and no renaming of a known empirical pattern occurs. The central claim therefore remains independent of its own inputs and is self-contained against standard U-statistic and martingale CLT machinery.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents full audit; no explicit free parameters, axioms, or invented entities are stated beyond standard martingale and spectral assumptions typical in time series.

pith-pipeline@v0.9.0 · 5390 in / 1078 out tokens · 52480 ms · 2026-05-08T16:33:01.395406+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

54 extracted references

[1]

Anderson, T. W. , title =. 1984 , isbn =

1984
[2]

Statistica Sinica , volume =

Bai, Zhidong and Saranadasa, Hewa , title =. Statistica Sinica , volume =. 1996 , issn =

1996
[3]

, title =

Bai, Zhidong and Silverstein, Jack W. , title =. The Annals of Probability , volume =. 2004 , issn =

2004
[4]

2012 , isbn =

Billingsley, Patrick , title =. 2012 , isbn =

2012
[5]

Journal of Multivariate Analysis , volume =

Bose, Arup and Hachem, Walid , title =. Journal of Multivariate Analysis , volume =. 2020 , issn =

2020
[6]

Box, G. E. P. and Pierce, David A. , title =. Journal of the American Statistical Association , volume =. 1970 , issn =

1970
[7]

Tony and Jiang, Tiefeng , title =

Cai, T. Tony and Jiang, Tiefeng , title =. The Annals of Statistics , volume =. 2011 , issn =

2011
[8]

Tony and Ma, Zongming , title =

Cai, T. Tony and Ma, Zongming , title =. Bernoulli , volume =. 2013 , issn =

2013
[9]

Biometrika , volume =

Chang, Jinyuan and Yao, Qiwei and Zhou, Wen , title =. Biometrika , volume =. 2017 , issn =

2017
[10]

Journal of the American Statistical Association , volume =

Chen, Songxi and Zhang, Lixin and Zhong, Pingshou , title =. Journal of the American Statistical Association , volume =. 2010 , issn =

2010
[11]

Journal of the American Statistical Association , volume =

Chen, Likai and Wu, Wei Biao , title =. Journal of the American Statistical Association , volume =. 2019 , issn =

2019
[12]

Statistica Sinica , volume =

Chen, Dachuan and Song, Fengyi and Feng, Long , title =. Statistica Sinica , volume =. 2025 , issn =

2025
[13]

, title =

Eaton, Morris L. , title =. 1983 , isbn =

1983
[14]

The Annals of Statistics , volume =

Feng, Long and Jiang, Tiefeng and Liu, Binghui and Xiong, Wei , title =. The Annals of Statistics , volume =. 2022 , issn =

2022
[15]

2022 , note =

Feng, Long and Liu, Binghui and Ma, Yanyuan , title =. 2022 , note =

2022
[16]

Journal of the Royal Statistical Society: Series B , volume =

Gao, Jiti and Han, Xiao and Pan, Guangming and Yang, Yanrong , title =. Journal of the Royal Statistical Society: Series B , volume =. 2017 , issn =

2017
[17]

and Heyde, C

Hall, P. and Heyde, C. C. , title =
[18]

, title =

Halmos, Paul R. , title =. The Annals of Mathematical Statistics , volume =. 1946 , issn =

1946
[19]

Journal of Multivariate Analysis , volume =

He, Daojiang and Liu, Huanyu and Xu, Kai and Cao, Mingxiang , title =. Journal of Multivariate Analysis , volume =. 2021 , issn =

2021
[20]

The Annals of Statistics , volume =

He, Yinqiu and Xu, Gongjun and Wu, Chong and Pan, Wei , title =. The Annals of Statistics , volume =. 2021 , issn =

2021
[21]

Annals of Mathematical Statistics , volume =

Hoeffding, Wassily , title =. Annals of Mathematical Statistics , volume =. 1948 , issn =

1948
[22]

Hosking, J. R. M. , title =. Journal of the American Statistical Association , volume =. 1980 , issn =

1980
[23]

The Annals of Applied Probability , volume =

Jiang, Tiefeng , title =. The Annals of Applied Probability , volume =. 2004 , issn =

2004
[24]

The Annals of Statistics , volume =

Jiang, Tiefeng and Yang, Fan , title =. The Annals of Statistics , volume =. 2013 , issn =

2013
[25]

Scandinavian Journal of Statistics , volume =

Jiang, Tiefeng and Qi, Yongcheng , title =. Scandinavian Journal of Statistics , volume =. 2015 , issn =

2015
[26]

The Annals of Applied Probability , volume =

Jiang, Tiefeng , title =. The Annals of Applied Probability , volume =. 2019 , issn =

2019
[27]

Journal of Business & Economic Statistics , volume =

Lan, Wei and Luo, Ronghua and Tsai, Chihling and Wang, Hansheng and Yang, Yunhong , title =. Journal of Business & Economic Statistics , volume =. 2015 , issn =

2015
[28]

Li, W. K. and McLeod, A. I. , title =. Journal of the Royal Statistical Society: Series B , volume =. 1981 , issn =

1981
[29]

Li, W. K. and McLeod, A. I. , title =. Biometrika , volume =. 1986 , issn =

1986
[30]

Journal of Multivariate Analysis , volume =

Li, Deli and Qi, Yongcheng and Rosalsky, Andrew , title =. Journal of Multivariate Analysis , volume =. 2012 , issn =

2012
[31]

Scandinavian Journal of Statistics , volume =

Li, Weiming and Li, Zeng and Yao, Jianfeng , title =. Scandinavian Journal of Statistics , volume =. 2018 , issn =

2018
[32]

The Annals of Statistics , volume =

Li, Zeng and Lam, Clifford and Yao, Jianfeng and Yao, Qiwei , title =. The Annals of Statistics , volume =. 2019 , issn =

2019
[33]

and Yang, Yaxing , title =

Ling, Shiqing and Tsay, Ruey S. and Yang, Yaxing , title =. Journal of Business & Economic Statistics , volume =. 2021 , issn =

2021
[34]

The Annals of Applied Probability , volume =

Liu, Weidong and Lin, Zhengyan and Shao, Qiman , title =. The Annals of Applied Probability , volume =. 2008 , issn =

2008
[35]

Ljung, G. M. and Box, G. E. P. , title =. Biometrika , volume =. 1978 , issn =

1978
[36]

, title =

Muirhead, Robb J. , title =. 1982 , isbn =

1982
[37]

, title =

Petrov, Valentin V. , title =. 1975 , isbn =

1975
[38]

, title =

Schott, James R. , title =. Biometrika , volume =. 2005 , issn =

2005
[39]

2003 , isbn =

Shao, Jun , title =. 2003 , isbn =

2003
[40]

, title =

Srivastava, Muni S. , title =. Journal of the Japan Statistical Society , volume =. 2005 , issn =

2005
[41]

and Kollo, T

Srivastava, Muni S. and Kollo, T. Some Tests for the Covariance Matrix with Fewer Observations than the Dimension under Non-Normality , journal =. 2011 , issn =

2011
[42]

Scandinavian Actuarial Journal , volume =

Stuart, Alan , title =. Scandinavian Actuarial Journal , volume =. 1954 , issn =

1954
[43]

, title =

Tsay, Ruey S. , title =. Journal of Econometrics , volume =. 2020 , issn =

2020
[44]

The Annals of Statistics , volume =

Wang, Runmin and Shao, Xiaofeng , title =. The Annals of Statistics , volume =. 2020 , issn =

2020
[45]

The Annals of Statistics , volume =

Wang, Runmin and Zhu, Changbo and Volgushev, Stanislav and Shao, Xiaofeng , title =. The Annals of Statistics , volume =. 2022 , issn =

2022
[46]

Statistica Sinica , volume =

Wang, Xiaoyi and Xu, Gongjun and Zheng, Shurong , title =. Statistica Sinica , volume =. 2023 , issn =

2023
[47]

Journal of Business & Economic Statistics , volume =

Wang, Lengyang and Kong, Efang and Xia, Yingcun , title =. Journal of Business & Economic Statistics , volume =. 2023 , issn =

2023
[48]

The Annals of Probability , volume =

Wu, Wei Biao and Woodroofe, Michael , title =. The Annals of Probability , volume =. 2004 , issn =

2004
[49]

Biometrika , volume =

Xu, Gongjun and Lin, Lifeng and Wei, Peng and Pan, Wei , title =. Biometrika , volume =. 2016 , issn =

2016
[50]

Journal of Multivariate Analysis , volume =

Xu, Yuanya and Li, Weiming , title =. Journal of Multivariate Analysis , volume =. 2026 , issn =

2026
[51]

Journal of Time Series Analysis , pages =

Zhang, Bingbing and Liu, Mengya , title =. Journal of Time Series Analysis , pages =. 2025 , issn =

2025
[52]

Statistics and Its Interface , volume =

Zhang, Bingbing and Liu, Mengya and Yan, Ting and Zhu, Ke , title =. Statistics and Its Interface , volume =. 2025 , issn =

2025
[53]

Journal of the American Statistical Association , pages =

Zhang, Yangfan and Wang, Runmin and Shao, Xiaofeng , title =. Journal of the American Statistical Association , pages =. 2025 , issn =

2025
[54]

Transactions of the American Mathematical Society , volume =

Zhou, Wang , title =. Transactions of the American Mathematical Society , volume =. 2007 , issn =

2007