Let X1,X2,… be independent identically distributed (i.i.d.) random variables with EXk=0, . Suppose that 蠁(t)鈮攍ogEetXk<∞ for all t>−蟽0 and some 蟽0>0. Let Sk=X1+鈰?Xk and S0=0. We are interested in the limiting distribution of the multiscale scan statistic
We prove that for an appropriate normalizing sequence an, the random variable converges to the Gumbel extreme-value law exp{−e−cx}. The behavior of depends strongly on the distribution of the Xk’s. We distinguish between four cases. In the superlogarithmic case we assume that 蠁(t)<t2/2 for every t>0. In this case, we show that the main contribution to comes from the intervals (i,j) having length l鈮攋−i of order a(logn)p, a>0, where aa1bbdb5badb954f925d7778" title="Click to view the MathML source">p=q/(q−2) and q∈{3,4,…} is the order of the first non-vanishing cumulant of X1 (not counting the variance). In the logarithmic case we assume that the function 蠄(t)鈮?蠁(t)/t2 attains its maximum m∗>1 at some unique point t=t∗∈(0,∞). In this case, we show that the main contribution to comes from the intervals (i,j) of length , a∈R, where d∗=1/蠁(t∗)>0. In the sublogarithmic case we assume that the tail of Xk is heavier than exp{−x2−蔚}, for some 蔚>0. In this case, the main contribution to comes from the intervals of length o(logn) and in fact, under regularity assumptions, from the intervals of length 1. In the remaining, fourth case, the Xk’s are Gaussian . This case has been studied earlier in the literature. The main contribution comes from intervals of length alogn, a>0. We argue that our results cover most interesting distributions with light tails. The proofs are based on the precise asymptotic estimates for large and moderate deviation probabilities for sums of i.i.d. random variables due to Cramér, Bahadur, Ranga Rao, Petrov and others, and a careful extreme value analysis of the random field of standardized increments by the double sum method.