Abstract. This paper studies the approximation of the solution of nonlinear ordinary differential equations by (discontinuous) piecewise polynomials of degree K and traces at the nodes of discretization. A mesh-dependent variational framework underlying this discontinuous approximation is derived. Several families of one-step, hybrid and multistep schemes are obtained. It is shown that the convergence rate in the ¿2-norm is K + 1. The nodal-convergence rate can go up to 2 K + 2, depending on the particular scheme under consideration. The mesh-dependent variational framework introduced here is of special interest in the approximation of the solution of optimal control problems governed by differential equations.1. Introduction. The object of this paper is the study of (discontinuous) piecewise polynomial approximations to the solution of systems of nonlinear ordinary differential equations defined on a fixed interval [0, T], T > 0. The type of approximation we shall use can be briefly described in the following way. The interval [0, T] is first partitioned into TV intervals by specifying a sequence {t"}^0, 0 = t0 < tx < ■ ■ ■ < tN = T, of real numbers. On each interval /" = [t"_x, t"], n = 1,..., TV, we construct a polynomial u" in PK(In), the space of polynomials of degree at most K > 0 defined on the interval /". At each node t", we specify a trace (or a point) U", n = 0,..., TV. So the approximation problem consists in finding the TV polynomials {un}"=x and the TV + 1 traces (or points) {[7"}^=0. We shall see that this kind of approximation leads to a global //-convergence rate of K + 1 and a nodal-convergence rate (for the traces {t/"}'s) of 2K + 2.In this paper we adopt a more general formulation, of which the above-described approximation is a special case. On each interval /" we allow L, 0 < L < K + 1, additional conditions on each polynomial un in PK(In). For instance, when L = 2 and «"('"_i)= U"_x and u"(t") = U", we obtain the continuous piecewise-polynomial approximation of B. L. Hulme [22], [23]. For that method, the global L2-convergence rate is K + 1 and the convergence rate at the nodes is 2K. This framework encompasses the a-method of Delfour, Hager and Trochu [14] for L = 1, and a"un(t") + (1 -an)u"+x(t") = X", n = 0,...,TV, where {a"}N=0 is an a priori