Efficient implementation of D x / D.T @ v

[svaiter]

In several solvers, sparse matrix is used to implement the finite difference operator as

len_y = len(self.y)
data = np.array([np.ones(len_y), -np.ones(len_y)])
diags = np.array([0, 1])
D = spdiags(data, diags, len_y-1, len_y)

I believe it could be implemented with a forward operator np.diff and an adjoint -np.diff(x, append=0, preprend=0)
A quick profiling give me x2 - x3 speedup, but most importantly it generalizes better if one wants to do 2D.

[EnLAI111]

Thanks! I will change them.

todo for me:

• CondatVu: K = (D A) for huber
• CondatVu: spnorm(D)
• ADMM: D.T @ D