Free variables occur often in practice. Handling free variables in interior point algorithms is a pending issue (see for example [Andersen2002], [Anjos2007], and [Meszaros1998]). Frequently users convert a problem with free variables into one with restricted variables by representing the free variables as a difference of two nonnegative variables. This approach increases the problem size and introduces ill-posedness, which may lead to numerical difficulties.
For an example we consider the test problem nb_L1 from the DIMACS test library [Pataki2002]. The problem originates from side lobe minimization in antenna engineering. This is a second order cone programming problem with 915 equality constraints, 793 SOCP blocks each of size 3, and 797 nonnegative variables. Moreover, the problem has two free variables that are described as the difference of four nonnegative variables. This problem can be loaded from the examples directory of VSDP. As the computation is more expensive, only the results are reported here:
SDPT3 solves the problem without warnings, although it is ill-posed according to Renegar’s definition [Renegar1994].
Now we try to get rigorous error bounds using the approximation of SDPT3.
These results reflect that the interior of the dual feasible solution set is
empty. An ill-posed problem has the property that the distance to primal or
dual infeasibility is zero. If as above the distance to dual infeasibility
is zero, then there are sequences of dual infeasible problems with input data
converging to the input data of the original problem. Each problem of the
sequence is dual infeasible and thus has the dual optimal solution $-\infty$.
Hence, the result $-\infty$ of
rigorous_lower_bound is exactly the limit of
the optimal values of the dual infeasible problems and reflects the fact that
the distance to dual infeasibility is zero. This demonstrates that the
infinite bound computed by VSDP is sharp, when viewed as the limit of a
sequence of infeasible problems. We have a similar situation if the distance
to primal infeasibility is zero.
If the free variables are not converted into restricted ones then the problem is well-posed and a rigorous finite lower bound can be computed.
By using the computed approximations we obtain the following rigorous bounds:
Therefore, without splitting the free variables, we get rigorous finite lower and upper bounds of the exact optimal value with an accuracy of about eight decimal digits. Moreover, verified interior solutions are computed for both the primal and the dual problem, proving strong duality.
In Table benchmark_dimacs_free_2012_12_12.html we display rigorous bounds for the optimal value of eight problems contained in the DIMACS test library that have free variables (see [Anjos2007] and [Kobayashi2007]). These problems have been modified by reversing the substitution of the free variables. We have listed the results for the problems with free variables and for the same problems when representing the free variables as the difference of two nonnegative variables. The table contains the rigorous upper bounds $fU$, the rigorous lower bounds $fL$, and the computing times measured in seconds for the approximate solution $t_s$, the lower bound $t_u$, and the upper bound $t_l$, respectively. The table demonstrates the drastic improvement if free variables are not split.
Independent of the transformation of the free variables the primal problems of the nql instances are ill-posed. The weak error bound of the optimal constraints. A solution for the qssp180 instance is due to the large number of equality system with 130141 equality constraints and 261365 variables has to be solved rigorously. In the next version of VSDP the accuracy for such large problems will be improved.
Published with GNU Octave 4.4.1