Margaret Duff

Hi @epapoutsellis - I am going through this in detail and have a few questions: 1) Comparing the code to the paper, I have lost this term:  Can you...

> 2. How do we want to set default step sizes: > The paper suggests convergence if >  > where β=1/f.L and λ=δγ and thus I suggest > ...

> For you question 1 on AA^T having gone missing. Maybe I misunderstand, but I read the implemented code as based on equation 4 (also the code comment states this),...

> Eq. (4a) contains the gradient of l^*. I don't see that in the code? The paper considers optimising  where we are considering an objective $f(x)+g(x)+h(Ax)$. I don't know...

> For reference the link to the published version: https://link.springer.com/article/10.1007/s10915-018-0680-3 > > In the lines just below eq (4) it says l^* is considered equal to zero. So perhaps this...

> Re. question 3, whether a warning. I think it makes sense to allow the algorithm to run with ZeroFunction being passed. I think also consistent with allowing this in...

> * For the step-sizes, better to have `0.99*2` which close to theory. Thanks Vaggelis, I think i have done this. > In theory, you can use any (0,∞) parameter...

@epapoutsellis @jakobsj - please can I revive this PR. Do you have any more comments?

Jenkins is happy 

More clearly explained in algorithm 12 of https://arxiv.org/pdf/2406.06342