Moderate deviations on different scales: no relations (Q476497)

For a sum \(S_n = \sum_{i=1}^n X_i\) of i.i.d. bounded variables \(\{X_i\}\), the classical central limit theorem as well as large deviation principles hold. For a large threshold not in the large deviations area, one can show so-called moderate deviation results, which in some sense extend the normal approximation to larger values. In this work, the \(\{X_i\}\) form a stationary process. Examples are constructed where a moderate deviations principle holds in a certain given region, and is violated outside the region. In particular, let \(G_0 \subset (0,{1\over2})\) be a finite union of open intervals bounded away from zero and \(G_1\) be the interior of the complement \((0,{1\over 2})\setminus G_0\). Then, there exists a stationary process \(\{X_i\}\) such that, for any \(\gamma \in G_0 \cup G_1\) and \(c \in (0,\infty)\), \[ \lim_{n \to \infty} {1\over n^{2\gamma}} \log {\mathbb P}[S_n > c n^{\gamma+1/2}] = \begin{cases} -{1\over 2}c^2 & \text{if \(\gamma \in G_1\)};\\ 0 & \text{if \(\gamma \in G_0\)}.\end{cases} \] The construction is based on a Markov chain on \(\{1,2,\ldots\}^2 \cup \{(0,0)\}\). From \((0,0)\) the process jumps to \((1,n-1)\) with probability \(p_n\) and stays in \((0,0)\) with probability \(p_1\). From \((k,\ell)\) with \(\ell \neq 1\) the process jumps to \((k+1,\ell-1)\), and from \((k,1)\) it jumps to \((0,0)\). The times where the origin is visited form a renewal process. If the Markov chain is denoted by \((A_t,B_t)\), then \(A_t\) is the time since the last visit of the origin and \(B_t\) is the time to the next visit of the origin. The process is now constructed as \(X_t = I_{t-A_t} \varphi(A_t,B_t)\) for some function \(\varphi\) and i.i.d. Bernoulli variables \(\{I_k\}\). In a first step, it is shown how to choose the parameters such that the result holds for \(G_0 = (u,v)\) with \(0 < u < v \leq {1\over2}\). Adding a finite number of such processes and normalising the sum will prove the result for arbitrary \(G_0\).