We'll cover this in CDS 110, on Wednesday. The basic idea is that you think of the convolution integral as a mapping from a function u(t) to a vector of n real numbers. The mapping is full rank if the set of all states that can be obtained by some input is all of R^n.