If the matrix is not sparse then the times are the same, O(n^2)
If the matrices are sparse, then the time of sum(A) is potentially O(Nz) where Nz is the number of non-zero values in the matrix. Then full(sum(A)) is going to require filling out the sparse row vector, and the time for that is going to be at least n (number of columns) because each column in the output would have to be written to; in practice it would be more like (n + p) where p is the number of non-sparse columns, which is a value you cannot directly compute from knowing the number of non-sparse elements of A (but you can apply the pigeon-hole principle to put a lower bound on it as ceiling(Nz/p)) . Assuming a best-case implementation rather than a practical implementation, call it n, so the overall order would be n+Nz where n is number of rows and Nz is number of non-zero entries (which is going to be between 0 and n^2). Best case is O(n), worst case is O(n^2).
