different results from c++ to Matlab for quadratic programming with osqp

[Mahmoud Omidvar]

Hi ,

I'm trying to use your OSQP library and we face now some problems.

On one hand, we use the matlab code to compute a solution to our quadratic programming problem. It provides a solution with

-----------------------------------------------------------------

           OSQP v0.2.1  -  Operator Splitting QP Solver

              (c) Bartolomeo Stellato,  Goran Banjac

        University of Oxford  -  Stanford University 2017

-----------------------------------------------------------------

problem:  variables n = 10, constraints m = 39

          nnz(P) + nnz(A) = 120

settings: linear system solver = suitesparse ldl,

          eps_abs = 1.0e-12, eps_rel = 1.0e-12,

          eps_prim_inf = 1.0e-12, eps_dual_inf = 1.0e-12,

          rho = 1.00e-01 (adaptive)

          sigma = 1.00e-06, alpha = 1.60, max_iter = 100000

          check_termination: on (interval 25)

          scaling: on, scaled_termination: off

          warm start: on, polish: off

iter   objective    pri res    dua res    rho        time

   1  -1.6006e+02   1.07e+01   1.22e+09   1.00e-01   1.20e-04s

 150   5.8662e+07   6.47e-14   1.44e-06   7.94e-01   2.54e-04s

status:               solved

number of iterations: 150

optimal objective:    58661573.9305

run time:             2.70e-04s

optimal rho estimate: 1.42e-02

On th other hand, we used C++ code to find the solution for the same problem. Here are the results. We are surpised that for the same operating conditions, we get different iteration number and a different result for nnz(P) + nnz(A).

This seems to come from the triangulation function of matrix P (csc_to_triu).

We start with a dense 10 by 10 P matrix , and  surpirisingly, by triangularizing, we obtain 110 non zero entries. This seems impossible.

 -----------------------------------------------------------------

           OSQP v0.2.1  -  Operator Splitting QP Solver

              (c) Bartolomeo Stellato,  Goran Banjac

        University of Oxford  -  Stanford University 2017

-----------------------------------------------------------------

problem:  variables n = 10, constraints m = 39

          nnz(P) + nnz(A) = 130

settings: linear system solver = suitesparse ldl,

          eps_abs = 1.0e-12, eps_rel = 1.0e-12,

          eps_prim_inf = 1.0e-12, eps_dual_inf = 1.0e-12,

          rho = 1.00e-01 (adaptive)

          sigma = 1.00e-06, alpha = 1.60, max_iter = 100000

          check_termination: on (interval 25)

          scaling: on, scaled_termination: off

          warm start: on, polish: off

iter   objective    pri res    dua res    rho        time

   1   0.0000e+00   1.07e+01   5.38e+09   1.00e-01   8.80e-05s

 125   5.8837e+07   3.73e-15   1.05e-06   1.14e+00   2.36e-04s

status:               solved

number of iterations: 125

optimal objective:    58836945.3983

run time:             2.65e-04s

optimal rho estimate: 4.43e-03

Did we do something wrong or is it a mistake?

Thank you in advance for your answer

Best Regards

Mahmoud

[Paul Goulart]

Could you post a minimal working example showing how you have called the solver in both the matlab and C++ versions?    It is difficult to debug otherwise.