Nonzero elements in the Hessian that should be zero

[abelsiqueira]

The code

using JuMP, MathProgBase
model = Model()
@variable(model, x[1:2])
@NLobjective(model, Min, 0.0)
@NLconstraint(model, x[1] * x[2] == 2)
d = JuMP.NLPEvaluator(model)
MathProgBase.initialize(d, [:Hess])
hx = zeros(3)
MathProgBase.eval_hesslag(d, hx, [0.659; 0.702], 1.0, [1.0])

corresponds to problem

min 0   s.t.  x_1 * x_2 == 2

which has Hessian of Lagrangian

\nabla^2 L(x,y) = [0 y_1; y_1 0]

i.e., only one nonzero is stored. However, a small error appears on the diagonal. Varies with the inputs.

Related to JuliaSmoothOptimizers/NLPModelsJuMP.jl/issues/13.

[mlubin]

This is an interesting bug. I can reproduce under JuMP 0.19. The call to @NLobjective isn't needed.

model = Model()
@variable(model, x[1:2])
@NLconstraint(model, x[1] * x[2] == 2)
d = NLPEvaluator(model)
MOI.initialize(d, [:Hess])
hx = zeros(3)
MOI.eval_hessian_lagrangian(d, hx, [0.659, 0.702], 0.0, [1.0])