Large and moderate deviations for random walks on nilpotent groups (Q1303912)

Let \((Z^j)_{j\in\mathbb N_0}\) be a sequence of i.i.d. random variables taking values in a simply connected, graded nilpotent Lie group \(G\), and consider the product \(T^n=Z^1\circ Z^2\circ\dots\circ Z^n\). One of the main goals of the paper is the derivation of a large-deviation principle (LDP) and a moderate-deviation principle (MDP) for the sequence \((T^n)_{n}\). Indeed, these two principles follow from the respective functional versions, which are the main results of the paper. Under the assumptions made in the paper, \(G\) may be identified (via the exponential mapping) with its Lie algebra \({\mathfrak g}\) which itself is identified with the Euclidean space \(\mathbb R^d\), equipped with a decomposition \(\mathbb R^d=V_1\oplus\dots\oplus V_l\) adapted to the structure of \({\mathfrak g}\). Define \(D_\alpha(g)=\sum_{i=1}^l\alpha^i g_i\), where \(g_i\) is the projection of \(g\in {\mathfrak g}\) on \(V_i\). Let \((c_n)_n\) be a sequence of positive numbers such that \(c_n=n\) in the LDP case and \(c_nn^{-1/2}\to\infty\) and \(c_n/n\to 0\) in the MDP case. Then a large-deviation principle is derived for \(D_{1/c_n}(T^n)\) at speed \(c_n^2/n\) with an explicit rate function in the LDP case resp. the MDP case. The main assumptions on the distribution of \(Z^1\) are that \(Z^1\) is centered and that the Laplace transform is finite in a neighborhood of the origin. Consider the random polygon \((S^n_t)_{t\in[0,1]}\) wich assumes the values \(Z^1+\dots+Z^k\) at times \(k/n\). Via two ODE's with random coefficients based on \(S^n_t\), one defines two random functions \((\widetilde Y^n_t)_{t\in[0,1]}\) and \((Y^n_t)_{t\in[0,1]}\) which interpolate the random walk \(D_{1/c_n}(Z^1\circ\dots\circ Z^k)\) at times \(k/n\), in the two cases \(c_n=n\) and \(c_n=o(n)\), respectively. A quite general functional version of the LDP and the MDP for \(\mathbb R^d\)-valued random walks has been proved by Mogul'skij (1976), with the rate functions explicitly given in terms of the Laplace transform respectively the covariance matrix of \(Z^1\). This result provides the basis for the proofs in the present paper. Essentially, the principles for \(\widetilde Y^n\) resp. for \(Y^n\) are derived via a transfer argument (reminiscent of the well-known projection lemma in large-deviation theory) from the above result. As a corollary, a law of the iterated logarithm is deduced from the MDP.