IPOPT: disable "Problem has too few degrees of freedom"

[cvanaret]

Dear all,

I'm writing a paper in which I compare my own NLP solver against state-of-the-art solvers on a subset of CUTEst problems.
On two problems (argauss and lewispol), IPOPT terminates with the error message "Problem has too few degrees of freedom", because there are more constraints than variables. Fair enough, either the problem is infeasible or some constraints are redundant.
However it's important that NLP methods be able to quickly detect (local) infeasibility. Therefore I'd like to run IPOPT anyway to monitor the number of iterations/evaluations until termination. Is it possible to disable this somehow?
Best,

Charlie

[svigerske]

As far as I know, there is no option to just "run Ipopt anyway". My general understanding is that the Ipopt datastructures are not setup in a way that they can handle overdetermined systems of equations.

If there are sufficiently many fixed variables, then Ipopt will already try automatically whether it can "activate" more degrees of freedom by relaxing the bounds on the fixed vars:

Ipopt/src/Interfaces/IpTNLPAdapter.cpp

Lines 619 to 629 in 5e3f0bc

	if( fixed_variable_treatment_ == RELAX_BOUNDS || n_x_fixed_ == 0 || n_x_var >= n_c || n_x_var == 0 )
	{
	done = true;
	}
	else
	{
	fixed_variable_treatment_ = RELAX_BOUNDS;
	jnlst_->Printf(J_WARNING, J_INITIALIZATION,
	"Too few degrees of freedom (n_x = %" IPOPT_INDEX_FORMAT ", n_c = %" IPOPT_INDEX_FORMAT ").\n Trying fixed_variable_treatment = RELAX_BOUNDS\n\n",
	n_x_var, n_c);
	}

If that isn't possible, then the only other option I am aware of is to try some experimental feature where Ipopt tries to detect redundant constraints and removes them: https://coin-or.github.io/Ipopt/OPTIONS.html#OPT_dependency_detector

[cvanaret]

Thank you @svigerske , I'll look into it!