Asymptotics of Yule’s nonsense correlation for Ornstein-Uhlenbeck paths: The correlated case.
Soukaina Douissi, Philip Ernst, Frederi Viens · 2025
Read Asymptotics of Yule’s nonsense correlation for Ornstein-Uhlenbeck paths: The correlated case. by Soukaina Douissi, Philip Ernst, Frederi Viens (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-2504-17175.txt.
Introduction
Soukaina Douissi, Philip Ernst, and Frederi Viens study a classic trap in statistics: two wandering series can look correlated when they are not. Yule named this nonsense correlation a century ago. The authors take its continuous-time form for two Ornstein-Uhlenbeck paths, which may themselves be truly correlated, and study the sample correlation between the paths over a long window.
Using sharp tools from Wiener chaos analysis, they prove the fluctuations of the sample value around its long-run limit follow a normal law. They measure the closeness in Kolmogorov distance, which yields error speeds for two simple independence tests, one built on the correlation and one on its numerator. An application tests independence between two views of the stochastic heat equation, using only a few Fourier modes. The tests there show excellent power.
Catalog record
| CREATOR | Soukaina Douissi, Philip Ernst, Frederi Viens |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 119 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-17175
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Asymptotics of Yule’s nonsense correlation for Ornstein-Uhlenbeck paths: The correlated case. Soukaina Douissi, Philip Ernst, Frederi Viens Abstract We study the continuous-time version of the empirical correlation coefficient between the paths of two possibly correlated Ornstein-Uhlenbeck processes, known as Yule’s nonsense correlation for these paths. Using sharp tools from the analysis on Wiener chaos, we establish the asymptotic normality of the fluctuations of this correlation coefficient around its long-time limit, which is the mathematical correlation coefficient between the two processes. This asymptotic normality is quantified in Kolmogorov distance, which allows us to establish speeds of convergence in the Type-II error for two simple tests of independence of the paths, based on the empirical correlation, and based on its numerator. An application to independence of two observations of solutions to the stochastic heat equation is given, with excellent asymptotic power properties using merely a small number of the solutions’ Fourier modes. 1 Introduction and Setup The first purpose of this paper is to provide a detailed asymptotic study of the empirical correlation coefficient \rho\left(T\right) between two standard Ornstein-Uhlenbeck (OU) processes X_{1} and X_{2} on a time interval [0,T] , as the time horizon T increases to infinity, where \rho\left(T\right) is defined below in ( 1 ). The OU paths X_{1} and X_{2} may or may not be correlated. This paper’s second purpose is to use those asymptotics to evaluate the power of independence tests based on \rho\left(T\right) itself as a test statistic, or on its constituent components as test statistics. It is important to note from the outset that the data available to compute these test statistics are the single pair of paths (X_{1},X_{2}) , not on repeated measurements of X_{1} and/or X_{2} . This is why we chose to investigate increasing-horizon (large time) asymptotics. This framework is well adapted to longitudinal obervational studies with high-frequency observations, as can occur commonly in environmental data, financial data, and many other areas where it is inconvenient or impossible to work with highly repeatable designed experiments. The notion of empirical correlation coefficient \rho\left(T\right) for any pair of paths of continuous stochastic processes (X_{1},X_{2}) defined on [0,T] can be defined by analogy with the standard Pearson correlation coefficient for these same paths observed in discrete time, e.g. at regular time intervals. Because the paths are continuous, it is a trivial application of standard Riemann integration that the standard Pearson correlation coefficient for the discrete-time observations of (X_{1},X_{2}) converges, as the time step converges to 0, to the following continuous-time statistic \rho(T):=\frac{Y_{12}(T)}{\sqrt{Y_{11}(T)}\sqrt{Y_{22}(T)}}, (1) where the random variables Y_{ij}(T) , i,j=1,2 are given via the following Riemann integrals Y_{ij}(T):=\int_{0}^{T}X_{i}(u)X_{j}(u)du-T\bar{X}_{i}\left(T\right)\bar{X_{j}}\left(T\right),\text{ \ }\text{ \ }\bar{X}_{i}\left(T\right):=\frac{1}{T}\int_{0}^{T}X_{i}(u)du. (2) This classical analysis statement holds almost surely, as soon as the paths (X_{1},X_{2}) are continuous almost surely and are not constant over time. That the denominators in ( 1 ) are non-zero comes an application of Jensen’s inequality where equality does not hold because the paths are not constant. All other details, including the definition of \rho in discrete time, are omitted, since many references including several cited below, such as [ 11 , 2 ] , cover this topic. The topic of independence testing for continuous-time modeled stochastic processes, as a mathematical framework approximating time series observed over long horizons or in high frequency, using \rho\left(T\right) as above, has gained renewed attention since Ernst, Shepp, and Wyner solved a long-standing mathematical conjecture regarding the exact quantitative behavior of the continuous-time version of Pearson’s correlation coefficient for independent random walks. Their paper [ 11 ] discusses the history of how \rho\left(T\right) relates to the discrete-time classical Pearson correlation coefficient. This paper can be consulted, along with its references, for why \rho\left(T\right) is the correct object of study, as we claim above. In the case of random walks, their paper explains, as was known since the 1960s, that the pair of processes (X_{1},X_{2}) should be Brownian motions (Wiener processes), and their paper proves that when these paths are independent, neither \rho\left(T\right) nor its discrete version converge to 0 , as one would expect for standard interpretations of Pearson correlation coefficients. Rather, \rho\left(T\right) is constant in distribution, with a variance which they compute explicitly. This is the so-called phenomenon of “Yule’s nonsense correlation”, and indeed this appellation is a label for \rho\left(T\right) itself. It was named after G. Udny Yule who discovered this phenomenon empirically in 1926 in [ 30 ] , and who had conjectured that the variance of \rho\left(T\right) should be computable. The fact that, for random walks, as for other self-similar processes such as fractional Brownian motions, \rho\left(T\right) has a stationary distribution as T increases, was presumably well known by the scholars who had studied Yule’s nonsense correlation since the 1960s, and most likely by Yule himself for the case of standard random walks. This fact was not recorded in the literature until it was pointed out in the introduction of [ 2 ] when discussing the distinction in the behavior of \rho\left(T\right) between highly non-stationary paths like random walks and Brownian motion on one hand, and i.i.d. data and stationary time series and processes like OU on the other. This brings us to the core of this paper’s topic. It has been known for some time that, when a pair of paths of times series is sufficiently stationary (with some limits on how long their memory (auto-correlation) is), the phenomenon of Yule’s nonsense correlation does not hold: the Pearson correlation of the pair of time series paths typically converges to their underlying mathematical correlation, just as one would expect for i.i.d data. This was established in continuous time for OU processes in the paper [ 9 ] : if \left(X_{1},X_{2}\right) are two OU paths with correlation r , then \rho\left(T\right)\rightarrow r almost surely i.e. as T\rightarrow\infty , and the fluctuations in this convergence are Gaussian, i.e. a Central Limit Theorem (CLT) holds for \sqrt{T}(\rho\left(T\right)-r) as T\rightarrow\infty . The paper [ 9 ] was published in 2025, but an arXiv version with this result in the case r=0 was posted in 2022 as [ 10 ] , which predates the publication of the paper [ 2 ] . In the paper [ 2 ] , the case of r=0 was studied in detail, and a speed of convergence in this CLT was established, at the so-called Berry-Esséen rate 1/\sqrt{T} , using tools from the so-called Wiener chaos analysis. That paper also studied the discrete high-frequency version of \rho\left(T\right) and established a rate of convergence of its normal fluctuations which depends on T and on the rate of observations. A study of moderate deviations for \rho\left(T\right) is given in the 2025 paper [ 31 ] , where the OU processes are observed at discrete time and in high frequency, similarly to the dicrete observation restrictions placed on the OU processes in the earlier contribution [ 2 ] . This leaves the case of correlated paths ( r\neq 0 ) in continuous time open. That question was taken up in fully dicrete time in the preprint [ 7 ] in the context of AR(1) processes, i.e. without simultaneous restrictions on high-frequency observations and increasing horizon. They establish an exact distribution theory, and they study asymptotics of the discrete-time version of the empirical correlation, quantitatively,