Convergence Criteria

Convergence of the DSDP solver may be defined using several options. The precision of the solution can be set by using the subroutine

DSDPSetGapTolerance(DSDP dsdp, double rgaptol);

The solver will terminate if there is a sufficiently feasible solution such that the difference between the objective values in (DD) and (PP), divided by the sum of their absolute values, is less than the prescribed number. A tolerance of 0.001 provides roughly three digits of accuracy, whereas a tolerance of 1.0e-5 provides roughly five digits of accuracy. The subroutine

DSDPSetMaxIts(DSDP dsdp, int maxits)

specifies the maximum number of iterations. The subroutine DSDPSetRTolerance(DSDP, double) specifies how small the constant $r$ representing the infeasibility in (D) must be to be an approximate solution, and the subroutine DSDPSetDualBound( DSDP, double) specifies an upper bound on the objective value in (D). The algorithm will terminate when it finds a point when the variable $r$ in (DD) is less than the prescribed tolerance and the objective value in (DD) is greater than this number.