Abstract. An analysis of moving least squares (m.l.s.) methods for smoothing and interpolating scattered data is presented. In particular, theorems are proved concerning the smoothness of interpolants and the description of m.l.s. processes as projection methods. Some properties of compositions of the m.l.s. projector, with projectors associated with finiteelement schemes, are also considered. The analysis is accompanied by examples of univariate and bivariate problems.1. Introduction. While the theory and practice of interpolation and approximation of functions of a single variable on the basis of a finite amount of information is well developed, the same is not true for functions of several variables. When the information consists of function values at the points of a rectangular grid, tensor product and blended interpolants based on univariate schemes can be employed. For irregularly distributed function-value data, the situation is much worse. Finiteelement techniques are of value, although, in order to produce C1 interpolants, more than just function-value information is required. As well, if the distribution of data is irregular, only triangular elements seem feasible. Their use has become more attractive in view of the recent development of fast and efficient triangulation algorithms. Nevertheless, the additional nodal information consisting of derivative data must somehow be concocted.Least squares approximation by polynomials is in widespread use and is, of course, not expected to produce an interpolant. However, as has been shown by McLain [7]