ProximalOperators.jl

ProximalOperators.jl copied to clipboard

Hi Lorenzo,

I wanted to follow up on this comment: https://github.com/JuliaFirstOrder/ProximalOperators.jl/issues/124#issuecomment-894072191

I think these weighted sums of norms can be useful for Group/Fused Lasso type problems. For the Group Lasso, one would need to think about how the group indices are specified and then you can simply execute the prox of the sum on each subvector.

I could try it out if you guys are interested in it.

@fabian-sp I think it would be useful to have them in the package. There could be multiple solutions for indexing the groups: I guess the simplest (and most general?) thing is to specify a vector of ranges, or a vector of vectors of indices, and then rely on views internally.

Another option, for matrix variables, would be to have each row or column as groups. But this could rely on the general solution above.

Do you have any ideas?

We had something like that for grouping in SlicedSeparableSum