Start with optimal solution as initial point

Large NLP's are difficult, not in the least because the chance increases that numerics become more unstable. Here is an example of such a problem with a square system (having 224k equations and variables):

solve m minimizing dummy using nlp;
x.fx(ivar,t) = x.l(ivar,t);
solve m minimizing dummy using nlp;

We fix the variables to their optimal values and solve again. This gives for the second solve:

               S O L V E      S U M M A R Y
 
     MODEL   m                   OBJECTIVE  dummy
     TYPE    NLP                 DIRECTION  MINIMIZE
     SOLVER  CONOPT              FROM LINE  554557
 
**** SOLVER STATUS     1 NORMAL COMPLETION         
**** MODEL STATUS      4 INFEASIBLE                
**** OBJECTIVE VALUE                0.0000
 

We get some good advice how to work around this by loosening some tolerance:

 ** An equation is inconsistent with other equations in the
    pre-triangular part of the model.
  
    Residual=           3.78931873E-08
    Tolerance (Rtnwtr)= 2.00000000E-08
  
    The pre-triangular feasibility tolerance may be relaxed with
    option:
  
    Rtnwtr=x.xx
  
    E_CRIMSONSTEOQEN(2017): Inconsistency in pre-triangular part of model.
 
    The solution order of the critical equations and
    variables is:
 
    All variables in equation E_CRIMSONSTEOQEN(2017) are now fixed
    and the equation is infeasible. Residual =-3.7893187255E-08