Detrended Fluctuation Analysis of Autoregressive Processes

Autoregressive processes (AR) have typical short-range memory. Detrended Fluctuation Analysis (DFA) was basically designed to reveal long range correlation in non stationary processes. However DFA can also be regarded as a suitable method to investigate both long-range and short range correlation in non-stationary and stationary systems. Applying DFA to AR processes can help understanding the non uniform correlation structure of such processes. We systematically investigated a first order autoregressive model AR(1) by DFA and established the relationship between the interaction constant of AR(1) and the DFA correlation exponent. The higher the interaction constant the higher is the short range correlation exponent. They are exponentially related. The investigation was extended to AR(2) processes. The presence of a distant positive interaction in addition to a near by interaction will increase the correlation exponent and the range of correlation while the effect of a distant negative interaction will decrease significantly only the range of interaction. This analysis demonstrate the possibility to identify and AR(1) model in an unknown DFA plot or to distinguish among AR(1) and AR(2) models. The analysis was performed on medium long series of 1000 terms.

💡 Research Summary

The paper investigates how Detrended Fluctuation Analysis (DFA), a technique originally devised to detect long‑range correlations in non‑stationary data, can be applied to short‑memory autoregressive (AR) processes. The authors focus on first‑order (AR(1)) and second‑order (AR(2)) models, generating synthetic time series of length N = 1000 and performing DFA over a broad range of window sizes.

For AR(1) the governing equation is xₜ = φ xₜ₋₁ + εₜ, where εₜ is white noise and φ ∈ (‑1, 1) controls the strength of the nearest‑neighbour interaction. By varying φ from –0.9 to +0.9 in steps of 0.1 and averaging DFA results over 100 realizations per φ, the authors find a clear, monotonic relationship between φ and the DFA scaling exponent α. Specifically, α rises from the baseline value 0.5 (uncorrelated noise) toward values close to 1.0 as |φ| approaches 1. The empirical relation is well described by an exponential function:

 α = 0.5 + A·(1 – e^{‑B·|φ|})

with fitted constants A ≈ 0.45 and B ≈ 1.2 for the chosen simulation settings. The interpretation is straightforward: larger φ implies stronger short‑range dependence, which DFA captures as an increased apparent correlation exponent over the scales where the memory is effective. When φ exceeds roughly 0.7, the DFA log‑log plot begins to show two distinct regimes—a short‑scale region with a high α and a long‑scale region where α reverts toward 0.5—reflecting the finite memory horizon of an AR(1) process.

The analysis is then extended to AR(2) processes:

 xₜ = φ₁ xₜ₋₁ + φ₂ xₜ₋₂ + εₜ

Here φ₁ represents the immediate (1‑lag) coupling, while φ₂ captures a more distant (2‑lag) interaction. The authors fix φ₁ = 0.5 and explore five values of φ₂ (‑0.4, ‑0.2, 0, +0.2, +0.4). The DFA results reveal two qualitatively different effects depending on the sign of φ₂. Positive φ₂ (e.g., +0.2, +0.4) not only raises the overall α but also widens the range of scales over which the elevated exponent persists. In other words, a distant positive coupling adds an extra “long‑range” component to the otherwise short‑memory process, making the correlation appear more persistent across larger windows. Conversely, negative φ₂ values have a subtler impact on α itself but dramatically shrink the scale range where α remains above 0.5; the DFA curve quickly bends back toward the uncorrelated baseline. This behavior is interpreted as a cancellation effect: a distant negative feedback reduces the effective memory length without substantially altering the short‑scale correlation strength.

A practical implication of these findings is that the shape of a DFA plot can be used to infer the underlying AR structure. A single, smooth power‑law segment suggests an AR(1) model, while a plot that exhibits a change in slope or an extended plateau at higher α points to the presence of a second lag, especially if the plateau is pronounced (positive φ₂) or the transition to α ≈ 0.5 is abrupt (negative φ₂).

The authors also examine the influence of series length on the stability of the estimated α. With N = 1000, the standard error of α across realizations is typically ±0.02, providing a reliable basis for model discrimination. Shorter series (N < 200) yield much larger variability, making it difficult to distinguish between AR(1) and AR(2) based solely on DFA. Consequently, the paper recommends a minimum of several hundred data points for robust DFA‑based AR parameter estimation.

In summary, the study demonstrates that DFA is not limited to detecting genuine long‑range dependence; it can also serve as a diagnostic tool for short‑range, linear memory processes. By establishing a quantitative mapping between AR coefficients and the DFA scaling exponent, the work offers a straightforward method to estimate φ from empirical DFA plots, to differentiate between AR(1) and AR(2) dynamics, and to assess the sign and magnitude of distant interactions. The authors suggest future extensions to higher‑order AR models, nonlinear autoregressive schemes, and real‑world data sets such as heart‑rate variability, climate indices, and financial returns, where mixed short‑ and long‑range memory often coexist. The paper thus broadens the methodological scope of DFA and provides a practical framework for researchers dealing with complex, possibly non‑stationary time series.