In 1923 (Lectures on Cauchy's Problem in Linear PDEs (New York, 1953)), J Hadamard considered a particular example to illustrate the ill-posedness of the Cauchy problem for elliptic partial differential equations, which consists in recovering data on the whole boundary of the domain from partial but over-determined measures. He achieved explicit computations for the Laplace operator, due to the squared shape of the domain, to observe, in fine, that the solution does not depend continuously on the given boundary data. The primary subject of this contribution is to extend the result to general domains by proving that the Cauchy problem has a variational formulation that can be put under a (variational) pseudo-differential equation, set on the boundary where the data are missing, and defined by a compact Steklov–Poincaré-type operator. The construction of this operator is based on the Dirichlet-to-Neumann mapping, and its compactness is derived from the elliptic regularity theory. Next, using mathematical tools from the linear operator theory and the convex optimization, we provide a comprehensive analysis of the reduced problem which enables us to state that (i) the set of compatible data, for which existence and uniqueness are guaranteed, is dense in the admissible data space; (ii) when the existence fails, due to possible noisy data, the variational problem can be consistently approximated by the least-squares method, that is the incompatibility measure (the deviation indicator or the variational crime made on the Steklov–Poincaré equation) equals zero though all the minimizing sequences blow up.