Scaling in the timing of extreme events

Abstract

Extreme events can come either from point processes, when the size or energy of the events is above a certain threshold, or from time series, when the intensity of a signal surpasses a threshold value. We are particularly concerned by the time between these extreme events, called respectively waiting time and quiet time. If the thresholds are high enough it is possible to justify the existence of scaling laws for the probability distribution of the times as a function of the threshold value, although the scaling functions are different in each case. For point processes, in addition to the trivial Poisson process, one can obtain double-power-law distributions with no finite mean value. This is justified in the context of renormalization-group transformations, where such distributions arise as limiting distributions after iterations of the transformation. Clear connections with the generalized central limit theorem are established from here. The non-existence of finite moments leads to a semi-parametric scaling law in terms of the sample mean waiting time, in which the (usually unkown) scale parameter is eliminated but not the exponents. In the case of time series, scaling can arise by considering random-walk-like signals with absorbing boundaries, resulting in distributions with a power-law "bulk" and a faster decay for long times. For large thresholds the moments of the quiet-time distribution show a power-law dependence with the scale parameter, and isolation of the latter and of the exponents leads to a non-parametric scaling law in terms only of the moments of the distribution. Conclusions about the projections of changes in the occurrence of natural hazards lead to the necessity of distinguishing the behavior of the mean of the distribution with the behavior of the extreme events.

Data-driven prediction of thresholded time series of rainfall and self-organized criticality models

“…which makes use of the fact that the scale parameter a is proportional to τ 2 / τ when −1 < γ < 0, as we have seen in [8,24]. Noting that a γ ∝ τ 2 / τ 2 , a non-parametric version of the scaling law for the ROC curve leads to the plot of specificity versus 1 − sensitivity τ 2 / τ 2 , which is also valid for when −1 < γ < 0 [24].…”

“…If 1 < β < 2, then the first two moments τ ∝ a 2−β and τ 2 ∝ a 3−β , from which it follows that a ∝ τ 2 / τ and a β ∝ τ 2 2 / τ 3 , thereby justifying the rescaling of the axes in Figs. 1(b) and (c) [8,24]. This rescaling is more appropriate than that used in…”

“…Noting that a γ ∝ τ 2 / τ 2 , a non-parametric version of the scaling law for the ROC curve leads to the plot of specificity versus 1 − sensitivity τ 2 / τ 2 , which is also valid for when −1 < γ < 0 [24]. For a better approximation, the finiteness of the time resolution, i.e., the fact that m < 0, should be taken into account in the calculation of the specificity, Eq.…”

Dependence of exponents on text length versus finite-size scaling for word-frequency distributions

“…where the moments of k are the original empirical ones (calculated for the discrete distribution). In a simple case (with no power laws involved [29,30]) we would have rescaled just by the mean k ; in this case the rescaling is a bit more involving [31,32]. Notice that this rescaling is totally equivalent to divide k by L, as shown in another section below; nevertheless, our choice is more general and makes the rescaling applicable when the data does not come from a text.…”

“…The case considered in the literature [32,38] assumes that g(z) has an intermediate power-law decay with exponent γ 1 followed by a much faster decay (exponential or so) for the largest k's. The pure power-law tail considered above is included in this framework when g(z) goes to zero abruptly, transforming the pure power law into a truncated power law.…”