standard errors

I have to solve the following optimization problem: maxf(S), (f is a multivariate log likelihood) subject to: S>0 (since S is a matrix S>0 means that S is positive definite) and Sum(|sij|)<m (m is a positive scalar) The theory suggests solving the dual problem, which is a quadratic problem with linear constraints. The method for the dual is Block Coordinate Descent (BCD). Everything is ok, so far. Now I want to calculate the standard errors of the parameters. For that, I think I just have to calculate the gradient and hessian of the primal problem, at the optimum. However, since the constraints are not differenciable in zero in am not sure how to calculate the gradient and hessian of the problem, even though I can compute the objective function's gradient and hessian with both numerical or analytical methods. Any hints or suggestions? Thanks in advance!