I’ve arXived the paper “On the condition number of a randomly perturbed matrix”, authored with Van Vu, which will appear at STOC 2007. (This paper was actually finished and submitted several months ago, but for some reason I neglected to upload it to the arXiv until now.) Here we show that given an arbitrary n \times n matrix M with polynomially bounded entries (in particular, M could be highly singular), a random discrete perturbation M+N, where N is (say) a random Bernoulli \pm 1 matrix, will have polynomially bounded condition number with high probability (specifically, given any A there exists a B such that the condition number is O(n^B) with probability 1-O(n^{-A})). This is the discrete analogue of an analogous result by Spielman and Teng for continuous random perturbations. The proof uses the inverse Littlewood-Offord technology as well as a discretisation lemma for generalized arithmetic progressions, both from our previous paper (which essentially addressed the M=0 case). It turns out that the effect of M is essentially just to add a few deterministic rows to a random matrix which one needs to obtain lower bounds for the least singular value.