Rectangular Jacobian matrices support

[tversteeg]

Thanks for the highly underrated library, it's very nice to use.

It would be nice to be able to return a non-square Jacobian matrix from the LeastSquaresProblem::jacobian() function. Whenever I try it I get a termination: WrongDimension("jacobian") as part of the report.

At least a mention of this limitation in the documentation would be nice.

Also, for my understanding: is this a mathematical limitation, a missing implementation, or am I doing something wrong?

[vadixidav]

Thank you for reporting your issue. You should be able to use LeastSquaresProblem with non-square matrices, which is something we use it for in Rust CV, so if it is having an issue with non-square matrices right now then I will definitely look at that.

Before I do that, I would like you to double check something. Notably, you need to make sure that your residuals have M rows, your params have N rows, and your Jacobian is MxN (rows x colums). Our example in the documentation likely isn't clear on this since it uses a square matrix, so it could be more clear by using a non-square matrix in the documentation.

If you check the above and that all seems correct, let me know, and I can investigate further. If you have a code sample that fails, I would appreciate that as well, but simply verifying for me that those things are true should be helpful. Thanks!

[tversteeg]

Thanks a lot, the problem was that the residuals I returned was the same size as the parameters. So I had 2 residuals, 2 parameters but a 2x1 Jacobian.

Maybe a message in the form of expected $x residuals but got $y would be better instead of the simple jacobian string that's returned currently?

[vadixidav]

Opened #11 to improve this error message.