A Note on Secure Computation of the Moore-Penrose Pseudoinverse and Its Application to Secure Linear Algebra

This work deals with the communication complexity of secure multi-party protocols for linear algebra problems. In our model, complexity is measured in terms of the number of secure multiplications required and protocols terminate within a constant number of rounds of communication. Previous work by Cramer and Damgard proposes secure protocols for solving systems Ax = b of m linear equations in n variables over a finite field, with m ≤ n. The complexity of those protocols is n5. We show a new upper bound of m4 + n2m secure multiplications for this problem, which is clearly asymptotically smaller. Our main point, however, is that the advantage can be substantial in case m is much smaller than n. Indeed, if m = √n, for example, the complexity goes down from n5 to n2.5. Our secure protocols rely on some recent advances concerning the computation of the Moore-Penrose pseudo-inverse of matrices over fields of positive characteristic. These computations are based on the evaluation of a certain characteristic polynomial, in combination with variations on a well-known technique due to Mulmuley that helps to control the effects of non-zero characteristic. We also introduce a new method for secure polynomial evaluation that exploits properties of Chebychev polynomials, as well as a new secure protocol for computing the characteristic polynomial of a matrix based on Leverrier's lemma that exploits this new method.