Quadratics as NLconstraints

[coreysharris]

(JuMP v18.5)

I am trying to solve a MIQCP with Bonmin. The problem has several quadratic constraints, but JuMP tells me that I can not use @constraint because

ERROR: LoadError: Solver does not support quadratic constraints

If I enter the quadratic expression "by hand", I can get it to work. Like so:

using JuMP, AmplNLWriter
m = Model(solver=AmplNLSolver("bonmin"))

@variable(m, a, Int)
@variable(m, b, Int)
@variable(m, s >= 0 )
@variable(m, t >= 0 )

@NLconstraint(m, 2*x+(a+b)*y >= 0)

But I actually I have an array of linear forms in a and b, and I'd like to create many constraints from them.

My first try

M = [2 a-b; a+b 3]
v = [s t]

for j = 1:2
    @NLconstraint(m, sum(M[i,j]*v[i] for i=1:2) == 0)
end

fails at the @NLconstraint: Unexpected object 2 in nonlinear expression. Ok, so I use the @NLexpression macro. This fails on the first line (same error).

M = @NLexpression(m, [i=1:2,j=1:2], [2 a-b; a+b 3][i,j])
v = @NLexpression(m, [i=1:2], [s t][i])

for j = 1:2
    @NLconstraint(m, sum(M[i,j]*v[i] for i=1:2) == 0)
end

What am I doing wrong?

[odow]

Reproducible on JuMP 0.19

model = Model()
@variable(model, x)
A = [x]
B = [2, x]
@NLconstraint(model, A[1] * B[1] == 0)

The problem is that B = [2, x] promotes 2 to an AffExpr, which are not allowed in @NL macros. Take a read of the syntax notes.

A quick fix is to type B as Any.

model = Model()
@variable(model, x)
A = [x]
B = Any[2, x]
@NLconstraint(model, A[1] * B[1] == 0)

This needs an improved error message.

[coreysharris]

Ah, I see. In my case,

M = Any[2 a-b; a+b 3]

still generated the error at a-b, but that's obviously an AffExpr. My solution (in case it's useful for others):

@variable(m, aux[1:2,1:2])
M = [2 a-b; a+b 3]
@constraint(m, aux .== M)

I created #1872 to address the error message.

[mlubin]

Closed by #1872. The underlying "quadratics not supported" issue is jump-dev/AmplNLWriter.jl#49.