Long-range dependence

Long-range dependence (LRD), also called long memory or long-term persistence, is a property of stochastic processes in which correlations between observations decay more slowly than exponentially — typically as a power law — so that events separated by arbitrarily long time lags remain correlated. Unlike Markov processes, where the future depends only on the present, a process with long-range dependence retains memory of its entire history. The rate of decay is quantified by the Hurst exponent H: for H \u003e 0.5, the process exhibits persistence (trends continue); for H \u003c 0.5, it exhibits anti-persistence (trends reverse).

Long-range dependence appears across domains that seem unrelated. In hydrology, Harold Hurst's study of Nile River flood levels revealed that wet years tend to cluster with wet years and dry years with dry years — a pattern that defies the independence assumption of standard stochastic models. In finance, price fluctuations exhibit persistent correlations on intermediate timescales, violating the efficient market hypothesis's assumption of independent returns. In network traffic, packet arrivals show bursty, self-similar patterns that persist across many orders of magnitude in time.

The mathematical signature of LRD is that the spectral density diverges at zero frequency and the autocorrelation function is not summable. This has profound consequences for statistical inference: standard estimators of variance, confidence intervals, and hypothesis tests assume either independence or short-range dependence, and they fail catastrophically when applied to LRD data. The apparent variance of a long-range dependent process grows without bound as the sample size increases, a property that invalidates many classical statistical procedures.

Long-range dependence is not a statistical nuisance to be removed by detrending or differencing. It is a diagnostic of systemic memory — evidence that the process is generated by mechanisms operating across multiple timescales, from fast local dynamics to slow global forcing. The systems thinker does not ask how to eliminate long memory but what structure produces it: feedback loops, aggregation of heterogeneous processes, or hierarchical coupling across scales. The Hurst exponent is merely the symptom; the architecture of memory is the disease — or the design.