MixedIntegerLinearProgram.add_constraint: Allow False, True

[mkoeppe]

follow up from #34878

Also p.linear_constraints_parent()(False)

CC: @mantepse

Component: linear programming

Issue created by migration from https://trac.sagemath.org/ticket/34882

[mantepse]

comment:2

Meanwhile I understand that it woul be hard to do something sensible with p.linear_constraints_parent()(False), because it should probably be the same as p.linear_constraints_parent()(ZZ(0)), and this is used as the building block for chained (in)equalities. More precisely, it does not make sense (to me) to attach a truth value to p.linear_constraints_parent()(False).

[mantepse]

comment:3

There is actually a doctest introduced in #13646 that checks that True/False input fail, but I think it is outdated.

[mantepse]

comment:4

Matthias, you wrote in #34878 comment:22:

It would be OK with me for

• p.add_constraint(True) to add the trivial constraint "0 <= 0"
• p.add_constraint(False) to add the infeasible constraint "0 <= -1"

Wouldn't it be better to ignore p.add_constraint(True) entirely?

I am surprised how tricky this little modification is. In particular, it also affects _is_redundant_constraint.

[mantepse]

comment:5

To give some background, in #33238 I have, repeatedly, things like

                c = AZ_matrix[z][w] - B_matrix[z][w]
                if c.is_zero():
                    continue
                if c in ZZ:
                    raise MIPSolverException
                self.milp.add_constraint(c == 0, name="statistics")

Since the code is already hard to read, just having

                self.milp.add_constraint(AZ_matrix[z][w] == B_matrix[z][w], name="statistics")

seems nicer. Moreover, the check c in ZZ seems brittle. I guess I could use

               if c.coefficient(-1):

instead, but this also looks quite obscure.

Do you have an estimate for when #34251 will be ready? How does CVXPY deal with trivially feasible / infeasible constraints?

[mantepse]

comment:6

Oops, it would have to be if c.coefficient(-1) and (c - c.coefficient(-1)).is_zero():.

[mantepse]

comment:7

I am now convinced that this is not a good idea. It is much better to make the user convert the input to a linear constraint themselves.

[mantepse]

comment:8

In my case, if I simply make sure that AZ_matrix[z][w] - B_matrix[z][w] is a LinearFunction, everything works just fine.

[mkoeppe]

comment:9

Yes, working within the linear function parent is, of course, the correct solution.

[mkoeppe]

comment:10

Replying to Martin Rubey:

How does CVXPY deal with trivially feasible / infeasible constraints?

Usually modeling frameworks have a specialized function for forming fast sums, which also guarantees that the empty sum has the correct type. (Also Sage has MixedIntegerLinearProgram.sum for the same purpose.)