Projected Gradient Descent
@zhou told me that you work on projected gradient descent algorithms for sparse tensor recovery.
I implemented a basic version of it using HOSVD as projection in https://git-ce.rwth-aachen.de/lamBOO/SparseTensorMoments.jl/-/blob/cf6ab3c97b1e2667e24d52f769ba2fd0ebf1eab8/examples/reconstruct/test-projected-gd.jl.
It seems to work when I use the optimal step size. Do you have experience with that or know some tricks to further speed it up?
Best Lambert