Probability flux as a method for detecting scaling

We introduce a new method for detecting scaling in time series. The method uses the properties of the probability flux for stochastic self-affine processes and is called the probability flux analysis (PFA). The advantages of this method are: 1) it is independent of the finiteness of the moments of the self-affine process; 2) it does not require a binning procedure for numerical evaluation of the the probability density function. These properties make the method particularly efficient for heavy tailed distributions in which the variance is not finite, for example, in Levy alpha-stable processes. This utility is established using a comparison with the diffusion entropy (DE) method.

💡 Research Summary

The paper introduces a novel technique called Probability Flux Analysis (PFA) for detecting scaling behavior in stochastic self‑affine time series. Traditional scaling detection methods, such as variance‑based approaches or the Diffusion Entropy (DE) method, rely on the existence of finite moments or require the construction of histograms to estimate probability density functions (PDFs). These requirements become problematic for processes with heavy‑tailed distributions, like Lévy α‑stable processes, where moments may diverge and binning introduces bias.

PFA circumvents these limitations by focusing on the probability flux J(x,t), defined as the time derivative of the PDF, J(x,t)=−∂P(x,t)/∂t. For a self‑affine process that obeys the scaling transformation x→λx and t→λ^α, the flux satisfies a corresponding scaling law J(λx,λ^α t)=λ^{−α−1}J(x,t). By taking the absolute value and averaging over space, one obtains ⟨|J|⟩∝t^{−(1+1/α)}. Consequently, a log‑log plot of ⟨|J|⟩ versus time yields a straight line whose slope m = −(1+1/α). The scaling exponent α can then be extracted directly as α = −1/(m+1).

The algorithm proceeds as follows: (1) sample the time series at uniform intervals to obtain positions x(t); (2) compute the flux numerically using a finite‑difference approximation of the time derivative of the empirical PDF; (3) average the absolute flux over the spatial domain for each time window; (4) perform linear regression on the log‑log data to determine the slope and thus α. Notably, the method does not require explicit binning of the PDF because the flux is derived from local changes rather than global density estimates.

The authors validate PFA through extensive simulations. For standard Brownian motion (α=2), both PFA and DE recover the correct exponent. For Lévy α‑stable processes with α=1.5 and α=1.2, where the variance is infinite, PFA accurately estimates α within a few percent, while DE shows substantial bias due to insufficient sampling of the heavy tails. Moreover, PFA remains robust when the data length is reduced to as few as 10^3 points, whereas DE’s estimates become highly unstable.

A discussion of strengths and limitations follows. Advantages include independence from moment finiteness, elimination of binning parameters, and applicability to short data segments because the flux captures instantaneous dynamics. The method also extends naturally to higher dimensions by defining a vector flux. The primary limitation is sensitivity to numerical differentiation noise; therefore, preprocessing steps such as smoothing filters or optimal time‑step selection are recommended. Additionally, processes with abrupt, discontinuous jumps may challenge the flux definition.

In conclusion, Probability Flux Analysis provides a theoretically sound and practically efficient alternative to existing scaling detection tools, especially for heavy‑tailed, non‑Gaussian processes. The paper suggests future work on incorporating drift terms, developing online versions for streaming data, and applying PFA to real‑world complex systems such as high‑frequency financial markets and seismic event catalogs.