Hi, thanks for the prompt response,
X is indeed desired to have the same sparsity pattern as the different R_s (which all have the same sparsity pattern). This is why I was trying to reduce the problem size by minimizing the sum of squares over the sparsity pattern only.
Unfortunately, I don't fully understand what you mean by
"you are then destroying it by adding the full vectorized constraint vec(X)>>0 through the SVECPSD cone."
My idea was to have only the non zero coefficients of the lower triangle of X in the decision variable, and to create the full vectorized lower triangle of X through Fx+g, which results in a lot of true zeros in this conic vector. I feel like I must be understanding something wrong then? Could you elaborate on how you see this?
Thank you for the example you provide. It is not exactly what I am after as it does not enforce the sparsity pattern of R on X, but it is still interesting to me. In particular I have the following question over the formulation choice and expected solving time/complexity.
Some context first: clearly your example runs faster than what I came up with. Looking at the PTF, I see that CVX formulates the Frobenius norm (over the sparsity pattern) by taking, for each non zero coefficient, the inner product of X with a matrix that is non-zero only at that coefficient, and then proceeding "as expected" from there. This results in as many symmetric matrices as there are non-zeros. Actually, this is the way I first implemented the Frobenius norm on my side as well (I don't think there is another way if X is a barvar given that the only available operations are inner products), which resulted in poor performance on my side. I then read documentation and conversations in this group, which led me to the LMI formulation I presented in my first email.
So my question is: did I get it wrong that this large number of symmetric matrices and inner products would result in poor performance? Is the LMI choice predictably slower in this case? How come the formulation by CVX is "so fast" even though it relies on large matrices instead of smaller vectors?
On a slightly unrelated note, I attach a simplified version of the initial demo that does not use randsample. R has a slightly different sparsity pattern as a result and is created in a naive way, but it does not really matter to this conversation I believe.
Thanks again for your time and looking forward to your explanations,
Kind Regards,
Nathan