The elephant random walk with gradually increasing memory (Q2170249)

The elephant random walk is given by \(S_n=X_1+\cdots+X_n\), where \(X_1=1\) with probability \(p\) and \(X_1=-1\) with probability \(1-p\). Subsequent steps are then chosen such that, for \(n\geq0\), \(X_{n+1}=X_K\) with probability \(p\) and \(X_{n+1}=-X_K\) with probability \(1-p\), where \(K\) is uniformly distributed over some subset \(\mathfrak{M}\) of \(\{1,2,\dots,n\}\). In the present paper, the authors mainly study the case where \(\mathfrak{M}=\{1,2,\dots,m_n\}\), where \(m_n\to\infty\) as \(n\to\infty\) in such a way that \(m_n/n\to0\). The authors establish asymptotics of the mean and variance of \(S_n\), and limit theorems for (suitably scaled versions of) \(S_n\). These asymptotics and limiting distributions depend on how \(p\) compares to the transitional value \(p=3/4\). Asymptotics for moments and limit theorems are also derived for the case where the random walk has some positive probability of not moving at each step, and for the case where \(\mathfrak{M}=\{1,2,\dots,m_n,n\}\), i.e., the random walk also remembers the most recent step. Finally, asymptotics of the first two moments are also given for the case where \(m_n/n\to\alpha\in(0,1]\) as \(n\to\infty\). The question of limit theorems in this case is left open; in the previous cases the proofs of the limit theorems relied on certain conditional variances vanishing asymptotically, a property which no longer holds in this case.