Probably because the SQP algorithm must extract subsets of rows of the constraint matrice quite often. This can be much faster for full matrices than sparse (see below). Also, even if your constraints are sparse, there is no reason to expect that the quadratic programming sub-problems of the algorithm will be. So it would be of questionable benefit to have constraints alone in sparse form.
A=sprand(4000,4000,4000/4e5)+speye(4000);
Af=full(A);
tic
for i=1:size(A,1)
a=A(i,:);
end
toc
tic
for i=1:size(A,1)
a=Af(i,:);
end
toc
Finally, even if everything in your problem is sparse, any matrix inversions that must be done will often be faster in full form than sparse for small scale problems, which sqp problems are expected to be. Example,
b=rand(4000,1);
tic
A\b;
toc
tic
Af\b;
toc
