InexactError: Float64 when problem gets big

[jluoaz]

I'm running an optimization problem on some simulated data:

using LinearAlgebra, JuMP, Ipopt

d = 10
n = 1000
A = I + 0.5*(rand(d,d)-0.5*ones(d,d))
A = Symmetric(A)
J = sqrt(A)
eps = randn(d,n)
z = J*eps

m = Model(with_optimizer(Ipopt.Optimizer))
@variable(m, theta >= 0, start = 1.0)
@variable(m, beta[1:d-1], start = 0.0)
@NLobjective(m, Min, theta/n*sum( (z[1,j] + sum(beta[i-1] * z[i,j] for i=2:d))^2 for j=1:n) - log(theta))

optimize!(m)

This code runs great when d=10, but change to d=100, and I get the following:

InexactError: Float64(0.3362800773416026 + 0.03631914074869952im)

Stacktrace:
 [1] Type at .\complex.jl:37 [inlined]
 [2] convert at .\number.jl:7 [inlined]
 [3] push!(::Array{Float64,1}, ::Complex{Float64}) at .\array.jl:853
 [4] _parse_NL_expr_runtime(::Model, ::Complex{Float64}, ::Array{JuMP._Derivatives.NodeData,1}, ::Int64, ::Array{Float64,1}) at C:\Users\<Username>\.julia\packages\JuMP\tyMag\src\parse_nlp.jl:171
 [5] top-level scope at C:\Users\<Username>\.julia\packages\JuMP\tyMag\src\parse_expr.jl:358

Not sure what is going on here.

[odow]

J can become a Array{Complex{Float64}, 2}, and JuMP cannot handle complex numbers.

This needs a better error message.

[jluoaz]

Ah, I see, I should have checked that first. Thank you!

[odow]

I'm going to re-open until this has a nicer error message.

[jluoaz]

Yeah, actually, I had originally thought that log(theta) had become complex. If I knew z was complex, I would have figured it out easily.