Extreme deviations for sums of random variables

Stretched exponential probability density functions (pdf), having the form of the exponential of minus a fractional power of the argument, are commonly found in turbulence and other areas. They can arise because of an underlying random multiplicative process. For this, a theory of extreme deviations is developed, devoted to the far tail of the pdf of the sum X of a finite number n of independent random variables with a common pdf e −f (x). The function f (x) is chosen (i) such that the pdf is normalized and (ii) with a strong convexity condition that f ′′ (x) > 0 and that x 2 f ′′ (x) → +∞ for |x| → ∞. Additional technical conditions ensure the control of the variations of f ′′ (x). The tail behavior of the sum comes then mostly from individual variables in the sum all close to X/n and the tail of the pdf is ∼ e −nf (X/n). This theory is then applied to products of independent random variables, such that their logarithms are in the above class, yielding usually stretched exponential tails. An application to fragmentation is developed and compared to data from fault gouges. The pdf by mass is obtained as a weighted superposition of stretched expo-nentials, reflecting the coexistence of different fragmentation generations. For sizes near and above the peak size, the pdf is approximately log-normal, while it is a power law for the smaller fragments, with an exponent which is a decreasing function of the peak fragment size. The anomalous relaxation of glasses can also be rationalized using our result together with a simple multiplicative model of local atom configurations. Finally, we indicate the possible relevance to the distribution of small-scale velocity increments in turbulent flow.