Creating square matrix variable with specific attribute

Hi,

I have got a problem with creating a square matrix and

I would really appreciate if someone could have comment or suggest a way.

I am trying to create a complex square matrix variable.

This matrix should be a lower triangular matrix and also its diagonal elements should be real and positive.

To sum up, a complex square matrix which is lower triangular and its diagonal elements are real and positive values.

I tried to use

variable A complex lower triangular;

imag(diag(A)) == 0;

Somehow CVX does not accept ‘imag(diag(A)) == 0;’

(i.e. if I erase the constraint ‘imag(diag(A)) == 0;’), it works)

If someone has other ways to create this matrix variable,

It would be really thankful.

Thank you!!

CVX 2.2 accepts the following:

cvx_begin
variable A complex lower triangular
imag(diag(A))==0

Actually, though, you want lower_triangular (note the underscore), but both are accepted without error message. And of course to be non-trivial, A should not be a scalar. So
variable A(n,n) complex lower_triangular .

Are you applying imag(diag(A))==0 to a nonlinear expression A, as opposed to a variable A? That would be a disallowed equality constraint.

Please ensure you are using CVX 2.2, not 3.0beta (which is known to have bugs).

I am using version 2.2, but I got the error message that says

Error using *
Inner matrix dimensions must agree.

Error in cvxprob/eliminate (line 137)
P = P * cvx_invert_structure(
xR );

Error in cvxprob/solve (line 18)
[ At, cones, sgn, Q, P, dualized ] = eliminate(
prob, true, shim.dualize );

Error in cvx_end (line 88)
solve( prob );

Error in optimize1 (line 154)
cvx_end

Is it kind of famous bug??

That looks like a CVX bug (although perhaps not famous, until now). It would help if you can show a minimum reproducible example illustrating the error.

I’ve got better insight on this error

Thank you for your interest and suggestion again

I’ve got better insight on this error

As you said, CVX was able to accept


variable W_m(N_RU,N+F,M) complex;
variable Q_m(N_RU,N_RU,M) lower_triangular complex;
variable V_m(N_CP,N_CP,M) hermitian

for m=1:M

      real(diag(Q_m(:,:,m))) >= 0;

      imag(diag(Q_m(:,:,m))) == 0;             

end


However, I have also another constraint as below


for m=1:M

-sum(sum_square_abs(R(:,:,m)*W_m(:,:,m)))-sum(sum_square_abs(R(:,:,m)*Q_m(:,:,m)))+log_det(Z_m(:,:,m))+N_RU…
>= -(B_fso/((1-alpha_1-alpha_2)*f_s))c_fso…
-((alpha_1
B_rf)/((1-alpha_1-alpha_2)*f_s))*log_det(eye(N_RU)+ ((F_rf_m(:,:,m)*V_m(:,:,m)F_rf_m(:,:,m)’)/sigma2_rf))…
-2
log_det(Q_m(:,:,m));

end


where

for m=1:M
R(:,:,m)=chol(Z_m(:,:,m));
end

Those two constraints can not be compatible.
In other words, either one of them should be erased for the CVX to work properly.
What I mean by two constraints is

  1. imag(diag(Q_m(:,:,m))) == 0;

and

  1. -sum(sum_square_abs(R(:,:,m)*W_m(:,:,m)))-sum(sum_square_abs(R(:,:,m)*Q_m(:,:,m)))+log_det(Z_m(:,:,m))+N_RU…

    = -(B_fso/((1-alpha_1-alpha_2)*f_s))c_fso…
    -((alpha_1
    B_rf)/((1-alpha_1-alpha_2)*f_s))*log_det(eye(N_RU)+ ((F_rf_m(:,:,m)*V_m(:,:,m)F_rf_m(:,:,m)’)/sigma2_rf))…
    -2
    log_det(Q_m(:,:,m));

I do not know what the constraint 2 above has to do with constraint 1 (i.e. making diagonal entries real values)

Thank you for your interest and suggestion again

I do not know

I suggest you produce a minimal reproducible example, and submit a bug report as instructed in http://cvxr.com/cvx/support/ .

I have tried several things and I’ve got to find this was because of

2*log_det(Q_m(:,:,m));.

Somehow

2*log_det(Q_m(:,:,m));

and

imag(diag(Q_m(:,:,m))) == 0;

This two can not be written in same problem.

Do you think the fact that this two can not be compatible is kind of bug of CVX?

I have tried several things and I’ve got to find this was because of

2*log_det(Q_m(:,:,m));.

Somehow

2*log_det(Q_m(:,:,m));

and

imag(diag(Q_m(:,:,m))) == 0;

This two can not be written in same problem.

Do you think the fact that this two can not be compatible is kind of bug of CVX?