Fast (Multi-)Evaluation of Linearly Recurrent Sequences: Improvements and Applications

Abstract

For a linearly recurrent vector sequence P[n+1] = A(n) * P[n], consider the problem of calculating either the n-th term P[n] or L<=n arbitrary terms P[n_1],...,P[n_L], both for the case of constant coefficients A(n)=A and for a matrix A(n) with entries polynomial in n. We improve and extend known algorithms for this problem and present new applications for it. Specifically it turns out that for instance * any family (p_n) of classical orthogonal polynomials admits evaluation at given x within O(n^{1/2} log n) operations INDEPENDENT of the family (p_n) under consideration. * For any L indices n_1,...,n_L <= n, the values p_{n_i}(x) can be calculated simultaneously using O(n^{1/2} log n + L log(n/L)) arithmetic operations; again this running time bound holds uniformly. * Every hypergeometric (or, more generally, holonomic) function admits approximate evaluation up to absolute error e>0 within O((log(1/e)^{1/2} loglog(1/e)) -- as opposed to O(log(1/e)) -- arithmetic steps. * Given m and a polynomial p of degree d over a field of characteristic zero, the coefficient of p^m to term X^n can be computed within O(d^2 M(n^{1/2})) steps where M(n) denotes the cost of multiplying two degree-n polynomials. * The same time bound holds for the joint calculation of any L<=n^{1/2} desired coefficients of p^m to terms X^{n_i}, n_1,...,n_L <= n.