AN earlier paper by one of us (Farrell, 1957) developed, inter alia, a method for estimating efficient production functions from observations of the inputs and outputs of a number of individual production units. The method consisted essentially in plotting these observations as points in a space of a suitable number of dimensions, forming the convex closure of this set of points and taking the appropriate part of the surface of this convex closure as the estimate of the efficient production function. In the case where n inputs are used to produce a single output under conditions of constant returns to scale, if output is represented by X and the inputs by xl, x2, ..., Xw then each firm can be represented as a point in n-dimensional space with coordinates (xl/X, x2/X, ..., x./X). This set of points is then augmented by the points at infinity on each of the n axes, and the convex closure of the augmented set is formed. The surface of this convex closure, omitting the facet at infinity, is our estimate of the efficient production function. The earlier paper was primarily concerned with this simple version of the method. Not only was it used for the basic exposition but a large part of the paper was devoted to applying it to observations on American agricultural production. The paper did, however, discuss at a formal level the relaxation of the assumptions of single product and constant returns to scale, and in Section 2.3 showed how the method could be generalized to the case of m outputs X1, X2, ..., Xm by treating each observation as a point in n + m-dimensional space with coordinates (X,, X2, ..., Xm, x1, x2, ..., xn). A broadly similar procedure which would accommodate the possibility of decreasing returns to scale was indicated in Section 2.4. Underlying the use of the method in all these cases is the assumption that the efficient production function to be estimated is convex. In no case does this assumption hold necessarily, but it is one frequently made in economics, and will probably hold in most actual situations. Certainly there is nothing in the nature of multiple outputs or decreasing returns to scale that is inconsistent with the convexity assumption. This happy state of affairs ceases to obtain as soon as we consider a situation with increasing returns to scale, because this essentially involves a non-convex production function; indeed they are in practice probably the most important example of nonconvex production functions. The estimation of efficient production functions under