Consider a small illustrative example. Suppose you want to minimize a two-variable quadratic function
on the nonnegative quadrant, subject to two constraints:
![\[ \begin{array}{rccccccccl} \min & 2x_1 & + & 3x_2 & + & x_1^2 & + & 10x_2^2 & + & 2.5 x_1 x_2 \\ \mbox{subject to} & x_1 & - & x_2 & \le & 1 & & & & \\ & x_1 & + & 2x_2 & \ge & 100 & & & & \\ & x_1 & & & \ge & 0 & & & & \\ & & & x_2 & \ge & 0 & & & & \end{array} \]](images/ormpug_optqp0021.png)
The linear objective function coefficients, vector of right-hand sides, and lower and upper bounds are identified immediately as
![\[ \mathbf{c} = \left[\begin{array}{c} 2 \\ 3 \end{array}\right],\quad \mathbf{b} = \left[\begin{array}{c} 1 \\ 100 \end{array}\right],\quad \mathbf{l} = \left[\begin{array}{c} 0 \\ 0 \end{array}\right],\quad \mathbf{u} = \left[\begin{array}{c} +\infty \\ +\infty \end{array}\right] \]](images/ormpug_optqp0022.png)
Carefully construct the quadratic matrix
. Observe that you can use symmetry to separate the main-diagonal and off-diagonal elements:
![\[ \frac{1}{2} \mathbf{x}^\textrm {T} \mathbf{Qx} \equiv \frac{1}{2} \sum _{i, j = 1}^ n\; x_ i\, q_{ij}\, x_ j = \frac{1}{2} \sum _{i = 1}^ n\; q_{ii}\, x_ i^2 + \sum _{i > j}\; x_ i\, q_{ij}\, x_ j \]](images/ormpug_optqp0023.png)
The first expression
![\[ \frac{1}{2} \sum _{i = 1}^{n}\; q_{ii}\, x_ i^2 \]](images/ormpug_optqp0024.png)
sums the main-diagonal elements. Thus, in this case you have
![\[ q_{11} = 2,\quad q_{22} = 20 \]](images/ormpug_optqp0025.png)
Notice that the main-diagonal values are doubled in order to accommodate the 1/2 factor. Now the second term
![\[ \sum _{i > j} x_ i\, q_{ij}\, x_ j \]](images/ormpug_optqp0026.png)
sums the off-diagonal elements in the strict lower triangular part of the matrix. The only off-diagonal (
) term in the objective function is
, so you have
![\[ q_{21} = 2.5 \]](images/ormpug_optqp0029.png)
Notice that you do not need to specify the upper triangular part of the quadratic matrix.
Finally, the matrix of constraints is as follows:
![\[ \mathbf{A} = \left[\begin{array}{cr} 1 & -1 \\ 1 & 2 \end{array}\right] \]](images/ormpug_optqp0030.png)
The SAS input data set with a quadratic programming system (QPS) format for the preceding problem can be expressed in the following manner:
data gsdata; input field1 $ field2 $ field3 $ field4 field5 $ field6 @; datalines; NAME . EXAMPLE . . . ROWS . . . . . N OBJ . . . . L R1 . . . . G R2 . . . . COLUMNS . . . . . . X1 R1 1.0 R2 1.0 . X1 OBJ 2.0 . . . X2 R1 -1.0 R2 2.0 . X2 OBJ 3.0 . . RHS . . . . . . RHS R1 1.0 . . . RHS R2 100 . . RANGES . . . . . BOUNDS . . . . . QUADOBJ . . . . . . X1 X1 2.0 . . . X1 X2 2.5 . . . X2 X2 20 . . ENDATA . . . . . ;
For more details about the QPS-format data set, see Chapter 17: The MPS-Format SAS Data Set.
Alternatively, if you have a QPS-format flat file named gs.qps, then the following call to the SAS macro %MPS2SASD translates
that file into a SAS data set, named gsdata:
%mps2sasd(mpsfile =gs.qps, outdata = gsdata);
Note: The SAS macro %MPS2SASD is provided in SAS/OR software. See Converting an MPS/QPS-Format File: %MPS2SASD for details.
You can use the following call to PROC OPTQP:
proc optqp data=gsdata primalout = gspout dualout = gsdout; run;
The procedure output is displayed in Figure 14.2.
Figure 14.2: Procedure Output
| Problem Summary | |
|---|---|
| Problem Name | EXAMPLE |
| Objective Sense | Minimization |
| Objective Function | OBJ |
| RHS | RHS |
| Number of Variables | 2 |
| Bounded Above | 0 |
| Bounded Below | 2 |
| Bounded Above and Below | 0 |
| Free | 0 |
| Fixed | 0 |
| Number of Constraints | 2 |
| LE (<=) | 1 |
| EQ (=) | 0 |
| GE (>=) | 1 |
| Range | 0 |
| Constraint Coefficients | 4 |
| Hessian Diagonal Elements | 2 |
| Hessian Elements Above the Diagonal | 1 |
The optimal primal solution is displayed in Figure 14.3.
Figure 14.3: Optimal Solution
The SAS log shown in Figure 14.4 provides information about the problem, convergence information after each iteration, and the optimal objective value.
Figure 14.4: Iteration Log
| NOTE: The problem EXAMPLE has 2 variables (0 free, 0 fixed). |
| NOTE: The problem has 2 constraints (1 LE, 0 EQ, 1 GE, 0 range). |
| NOTE: The problem has 4 constraint coefficients. |
| NOTE: The objective function has 2 Hessian diagonal elements and 1 Hessian |
| elements above the diagonal. |
| NOTE: The QP presolver value AUTOMATIC is applied. |
| NOTE: The QP presolver removed 0 variables and 0 constraints. |
| NOTE: The QP presolver removed 0 constraint coefficients. |
| NOTE: The presolved problem has 2 variables, 2 constraints, and 4 constraint |
| coefficients. |
| NOTE: The QP solver is called. |
| NOTE: The Interior Point algorithm is used. |
| NOTE: The deterministic parallel mode is enabled. |
| NOTE: The Interior Point algorithm is using up to 4 threads. |
| Primal Bound Dual |
| Iter Complement Duality Gap Infeas Infeas Infeas Time |
| 0 3.5863E+03 4.8823E+00 1.0251E+00 1.0354E+02 2.3142E-15 0 |
| 1 1.9345E+03 9.6222E-01 4.4158E-01 4.4602E+01 5.7855E-16 0 |
| 2 2.2140E+03 1.2297E-01 4.4158E-03 4.4602E-01 5.5926E-15 0 |
| 3 5.0020E+01 3.2272E-03 4.4158E-05 4.4602E-03 9.6502E-15 0 |
| 4 4.9973E-01 3.2332E-05 4.4158E-07 4.4602E-05 2.5148E-14 0 |
| 5 4.9972E-03 3.2332E-07 4.4158E-09 4.4602E-07 3.0701E-14 0 |
| 6 0.0000E+00 3.6334E-16 1.5730E-16 0.0000E+00 1.2342E-14 0 |
| NOTE: Optimal. |
| NOTE: Objective = 15018. |
| NOTE: The Interior Point solve time is 0.00 seconds. |
| NOTE: The data set WORK.GSPOUT has 2 observations and 9 variables. |
| NOTE: The data set WORK.GSDOUT has 2 observations and 10 variables. |
See the section Interior Point Algorithm: Overview and the section Iteration Log for the OPTQP Procedure for more details about convergence information given by the iteration log.