Add more general linear operator constraint to sdp #350
Conversation
|
Thanks for opening your first pull request in this repository! Someone will review it when they have a chance. In the mean time, please be sure that you've handled the following things, to make the review process quicker and easier:
Thank you again for your contributions! 👍 |
Solely to avoid build errors.
There was a problem hiding this comment.
Hey @SuvarshaChennareddy, thanks for taking the time to look into this. I have some basic comments before you proceed further. I think that tests are definitely necessary both to ensure that the implementation is correct, and also that it might behave like a user expects.
If I were you, I would start by assembling a couple reasonable examples of how a user might want to express a constraint in this way, and then designing to make those cases easy. 👍
| @@ -109,6 +141,8 @@ inline void Smat(const MatAType& input, MatBType& output) | |||
| } | |||
| } | |||
|
|
|||
|
|
|||
|
|
|||
There was a problem hiding this comment.
No need to add this whitespace :)
| */ | ||
| template<typename ElemType, typename MatBType> | ||
| inline void convertToMatrix(std::function<ElemType(arma::Mat<ElemType>)> input, | ||
| MatBType& output) |
There was a problem hiding this comment.
It looks like the approach you are taking is to turn a std::function that the user gives into a matrix, and then use that as a constraint. But this is the opposite of what we are looking for in #72: take the given example of the constraint that the sum of elements is a constant. You can express this very tediously as the constraint trace(1 1^T X) = b, but the much easier way to represent this is sum(X) = b. So, instead of extracting the matrix 1 1^T, it would be better to just store the std::functions in the SDP objects and evaluate them directly to get the value.
There was a problem hiding this comment.
First of all, thank you for taking the time to look into this. I'm glad you asked this. I honestly didn't know if this would be alright.
However, (from what I understand) in the interior point method, the constraints need to be expressed as matrices (as matrix multiplication and transposing is involved).
Although, in the LRSDP implementation, I have used these std::functions wherever the matrix conversion isn't necessary (again like in the case of matrix multiplication or transposing).
I had given it some thought but using these std::functions directly would be extremely tedious as a lot of factors must be accounted for. For example, how would I calculate M=A(E^-1)F(A^T) where A is the matrix of vector forms of the constraints.
Nevertheless, I will think about this a bit more and try to come with a more appropriate solution.
|
Will do! Thank you! |
Co-authored-by: Ryan Curtin <[email protected]>
|
This issue has been automatically marked as stale because it has not had any recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions! 👍 |
Hello everyone,
This is related to issue #72. I've allowed the ability to store constraints as
std::functionobjects.I have not modified/added tests yet. However, if this PR looks alright, I'll be willing to do so.
Any comments would be greatly appreciated.