! XNLP example for dual multipliers for non-linear problems
!
! This example solves a series of problem, some of which do not have valid dual
! multipliers (Lagrange multipliers) in the absense of one of the regularity
! conditions (like Slater's condition)
!
! This example demonstrates a particular non-linear optimization concept as related
! to Xpress NonLinear.
! The version of the example is for Xpress 7.5.
!
!  (c) 2013-2024 Fair Isaac Corporation
!       author: Zsolt Csizmadia
model mmxnlp_nlp_duals
uses "mmxnlp"; 

declarations
 x,y: mpvar
end-declarations

! Simple model minimizing a linear objective over a ball
Ball1 := x^2 + y^2 <= 1
x is_free
y is_free

! Solve with a local solver to be able to obtain duals
setparam("XPRS_NLPSOLVER",1)

! Optimize with defaults. Note that y solution of y is not exact
writeln("Step one: default options")
minimize(y)
writeln("Optimal solution: x=",getsol(x)," y=",getsol(y))
writeln("-----------------------------------------------")
writeln

! Set a strict complementarity target. Bargaptarget is a "wish", it is safe to
! set it small, if the solver cannot satisfy it, it will stop early
! (in constrast to bargapstop)
! This is a solver specific setting, use the xprs_ pretag.
setparam("xprs_bargaptarget",1e-20)
minimize(y)
writeln("Optimal solution: x=",getsol(x)," y=",getsol(y))
! Look at the dual values
! Observe that the multipliers provide a proof of local optimality
writeln("Optimal solution: pi(Ball1)=",getdual(Ball1))
writeln("Linearization of Ball1 at the optimal solution: (", 2*getsol(x),
        ",",2*getsol(y),")")
writeln("Multiplied with its dual: (", getdual(Ball1)*2*getsol(x),
        ",",getdual(Ball1)*2*getsol(y),")")
writeln("Linearization of the objective at the optimal solution: (0,1)")
writeln("-----------------------------------------------")
writeln

! Now introduce a new nonlinear constraint that only intersects the first
! constraint in a single point
Ball2 := (x-2)^2 + y^2 <= 1
minimize(y)
writeln("Optimal solution: x=",getsol(x)," y=",getsol(y))
! Look at the dual values, and note the very large numbers
! Note that the solver kept pertubing y even though the very strict
! complementarity target
writeln("Optimal solution: pi(Ball1)=",getdual(Ball1), " pi(Ball2)=",getdual(Ball2))
writeln("-----------------------------------------------")
writeln

! The solver has artifically pertuebed (relaxed) the problem.
! Let we do the relaxation instead by the means of redefining the second constraints
Ball2 := (x-2)^2 + y^2 <= 1.001
minimize(y)
writeln("Optimal solution: x=",getsol(x)," y=",getsol(y))
! Even though the relaxation is small, 0.1%, the duals are now meaningful
writeln("Optimal solution: pi(Ball1)=",getdual(Ball1), " pi(Ball2)=",getdual(Ball2))
writeln("-----------------------------------------------")
writeln

! The case of x^2 <= 0 is interseting, as there is simply no constraint gradient
! The effect is the same: the solver perturbs
sethidden(Ball1, true)
sethidden(Ball2, true)
Ball := x^2 + y^2 <= 0
minimize(y)
writeln("Optimal solution: x=",getsol(x)," y=",getsol(y))
! Even though the relaxation is small, 0.1%, the duals are now meaningful
writeln("Optimal solution: pi(Ball)=",getdual(Ball))
writeln("-----------------------------------------------")
writeln



end-model