Analysis of the autocorrelation function for time series with higher-order temporal correlations: An exponential case

Min-ho Yu, Hang-Hyun Jo · 2025

Read Analysis of the autocorrelation function for time series with higher-order temporal correlations: An exponential case by Min-ho Yu, Hang-Hyun Jo (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-2502-11043.txt.

Introduction

Min-ho Yu and Hang-Hyun Jo study how events that arrive in bursts leave fingerprints on a standard statistics curve. Many systems, from messages to trades, fire in clusters separated by quiet gaps, and the sizes of back-to-back bursts often correlate. The autocorrelation function, which measures how a series resembles itself over time, is the usual lens for such patterns. The authors build a model series with correlated burst sizes using a coupling technique from statistics.

They then solve exactly for that model's autocorrelation under any distributions of gaps and burst sizes. For a concrete case they plug in exponential distributions for both. The worked example shows how correlation between one burst and the next changes the way the curve decays. The result gives analysts a cleaner way to read burst structure out of recorded time series.

Catalog record

CREATORMin-ho Yu, Hang-Hyun Jo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-11043

Browse the full library · About the Tower of Babel

Full text

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

Analysis of the autocorrelation function for time series with higher-order temporal correlations: An exponential case Min-ho Yu, Hang-Hyun Jo Abstract Temporal correlations in the time series observed in various systems have been characterized by the autocorrelation function. Such correlations can be explained by heavy-tailed interevent time distributions as well as by correlations between interevent times. The latter is called higher-order temporal correlations, and they have been captured by the notion of bursts; a burst indicates a set of consecutive events that rapidly occur within a short time period and are separated from other bursts by long time intervals. The number of events in the burst is called a burst size. Some empirical analyses have shown that consecutive burst sizes are correlated with each other. To study the impact of such correlations on the autocorrelation function, we devise a model generating a time series with higher-order temporal correlations by employing the copula method. We successfully derive the analytical solution of the autocorrelation function of the model time series for arbitrary distributions of interevent times and burst sizes when consecutive burst sizes are correlated. For the demonstration of our analysis, we adopt exponential distributions of interevent times and burst sizes to understand how the correlations between consecutive burst sizes affect the decaying behavior of the autocorrelation function. I Introduction Time series observed in various natural and social systems have shown complex temporal patterns with long-term temporal correlations [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 ] . One of the most commonly used methods characterizing temporal correlations is the autocorrelation function (ACF) [ 11 , 6 ] . The ACF for the time series x(t) is defined as \displaystyle A(t_{\rm d})\equiv\frac{\langle x(t)x(t+t_{\rm d})\rangle-\langle x(t)\rangle^{2}}{\langle x(t)^{2}\rangle-\langle x(t)\rangle^{2}}, (1) where \langle\cdot\rangle is the time average taken over the entire period in the time series, and t_{\rm d} is the time lag. For the time series with long-term correlations, one often finds an algebraically decaying behavior such that A(t_{\rm d})\propto t_{\rm d}^{-\gamma} , where \gamma is the decay exponent. The time series that is given as a sequence of event timings is called an event sequence. The empirical event sequences have been characterized in terms of the interevent time (IET) which is the time interval between two consecutive events and denoted by \tau [ 12 , 10 , 13 ] . In particular, heavy-tailed IET distributions, P(\tau) , imply long-term memory effects in the time series as the memoryless, Poisson processes lead to exponential IET distributions. Recently, the notion of bursty trains was introduced to capture correlations between IETs that are missing in the IET distribution [ 14 ] . For a given timescale \Delta t , a bursty train or a burst is defined as a set of consecutive events where each pair of consecutive events in the burst are separated by an IET less than or equal to \Delta t , while events in different bursts are separated by IETs larger than \Delta t . The number of events in the burst is called a burst size, which is denoted by b . Empirical analyses have shown that the burst size distributions, Q_{\Delta t}(b) , are heavy-tailed for a wide range of \Delta t , implying the presence of correlations between arbitrarily many consecutive events [ 14 , 15 , 16 , 17 , 18 , 19 , 10 , 20 ] . In addition, consecutive burst sizes turn out to be positively correlated with each other for a wide range of \Delta t in several datasets from diverse backgrounds [ 20 ] . It means that bigger (smaller) bursts tend to follow bigger (smaller) ones, which can be related to the bursty-get-burstier mechanism [ 21 ] that has been proposed to explain heavy-tailed burst size distributions. We call such empirical findings higher-order temporal correlations as they are correlations beyond the IET distributions [ 22 ] . Temporal correlations in event sequences, characterized by the ACF, can be explained by the heavy-tailed IET distribution as well as by correlations between IETs. The effects of the latter, higher-order temporal correlations on the ACF have been largely unexplored. For the simplest, Poisson process, it is obvious that A(t_{\rm d})=\delta(t_{\rm d}) , where \delta(\cdot) is a Dirac delta function. For the renewal process in which IETs are fully uncorrelated with each other, the power-law IET distribution with the exponent \alpha has been analytically shown to lead to the decay exponent of the ACF as \gamma=|\alpha-2| for 1<\alpha<3 [ 23 ] . The cases with correlated IETs were also studied; when the IETs are power-law distributed with exponent \alpha and each IET depends only on the previous IET in the time series, the scaling relation between \alpha and \gamma remains the same, but the ACF is elevated for the range of the small t_{\rm d} due to the positive correlations between two consecutive IETs [ 24 ] . More recently, for the case where the correlations between IETs were implemented by a power-law burst size distribution with the exponent \beta , the relation between \alpha , \beta , and \gamma has been analytically obtained by devising a discrete-time model considering only one timescale \Delta t=1 for defining bursts [ 25 ] . However, the effects of correlations between two consecutive burst sizes on the ACF have not been studied, which is the aim of our current work. In this work, we extend the time series model proposed in Ref. [ 25 ] to incorporate the correlations between two consecutive burst sizes. In particular, the correlations between two consecutive burst sizes are implemented using the Farlie-Gumbel-Morgenstern (FGM) copula method [ 26 , 27 , 28 , 29 , 30 , 31 ] . The copula method enables us to write a joint probability distribution function of correlated variables in a tractable form [ 27 ] . Note that the FGM copula was employed in the previous works to implement correlations between two consecutive IETs [ 24 , 32 , 33 ] . Then we derive an analytical solution of the ACF as a function of correlations between two consecutive burst sizes for arbitrary IET and burst size distributions. To demonstrate the validity of our analysis, we consider exponential distributions of both IETs and burst sizes but with their lower bounds. We find that the positive correlation between two consecutive burst sizes elevates the ACF curve. Finally, we perform numerical simulations to confirm our analytical results. II Model Following Ref. [ 25 ] , we consider a discrete-time model that generates a time series, \left\{x(1),x(2),\ldots,x(T)\right\} , where T is the number of discrete times. We set x(t)=1 if an event occurs at a time t , x(t)=0 otherwise. By definition, the minimum interevent time (IET) is \tau_{\rm min}=1 . As for the timescale defining bursts, we consider only \Delta t=1 , implying that any maximal set of consecutively occurring events is detected as a burst. An isolated event can also make its own burst. The number of events in a burst is called a burst size, which is denoted by b . Considering the fact that each burst of size b generates exactly b-1 IETs of \tau=1

Read the complete text