A quadratic programming problem is a mathematical programming problem, which consists of an objective function composed of linear and quadratic terms and a set of linear constraints. A quadratic programming problem can be solved by a modification of the simplex procedure.
The general quadratic programming problem is given by
Maximize f = cjxj + cjkxjxk
subject to
aijxj - bi ≤ 0 , i = 1, 2, ...., m
xj ≥ 0, j = 1, 2, ...., n
If the quadratic form is negative semidefinite and symmetric, then it is a concave function of the decision variables. Since the constraints are linear, the feasible region is convex. Thus, Kuhn-Tucker conditions in this case are both necessary and sufficient.
To derive these conditions, we use multipliers λi ( i = 1, 2, ...., m) corresponding to the constraints
aijxj - bi ≤ 0;
and μj ( j = 1, 2, ....,
n) corresponding to the non-negativity constraints.
df/dxj = cj + 2 cjkxk , j = 1, 2, ...., n
d ---- dxj |
aijxj - bi | = | aij ; i = 1, 2, ...., m; j = 1, 2, ...., n |
Let Si = bi - aijxj; i = 1, 2, ...., m .....(i)
Where Si is a slack variable.
Hence, Kuhn-Tucker conditions are given by
cj + 2 cjkxk - λiaij + μj = 0; j = 1, 2, ...., n
or - 2 cjkxk + λiaij - μj = cj; j = 1, 2, ...., n .....(ii)
Siλi = 0, xjμj = 0
Si, λi, μj,
xj ≥ 0, i = 1, 2, ...., m;
j = 1, 2, ...., n
If we ensure that the pairs ( λi, Si) and (μj, xj) are not into the basis simultaneously, then the conditions Siλi = 0 and xjμj = 0 will be automatically satisfied. Therefore, we use simplex procedure with a restricted entry rule.