Abstract. In this paper, the problem of estimating Jacobian and Hessian matrices arising in the finite difference approximation of partial differential equations is considered. Using the notion of computational molecule or stencil, schemes are developed that require the minimum number of differences to estimate these matrices. A procedure applicable to more complicated structures is also given.1. Introduction. In the past few years, there has been a growing interest in studying efficient ways to obtain good estimates of sparse Jacobian or Hessian matrices. Two basic approaches have been followed. In the first approach, it is assumed that one has an estimate of the Jacobian (Hessian) which one wishes to revise given "arbitrary" changes in the variables and the corresponding changes in the functions (gradients). This has led to the development of several sparse quasi-Newton methods (see [11], [14], [12], for example). In the second approach, it is assumed that one can specify the changes in the variables. One then forms an estimate of the Jacobian (Hessian) matrix from the observed changes in the functions (gradients). When the dimension of the matrix is large and the evaluation of each function (gradient) vector is expensive, it is necessary to take the sparsity structure of the matrix into account, in order to make this approach efficient.