Chester Holtz

Thank you so much for the resources! I really appreciate it. I guess that second paper is the most general, but doesn't guarantee optimality. I am happy to make a...

Hi yes - sorry for the delayed response. Still potentially interested. I ended up doing iteratively reweighted least squares to solve the l1 problem for my application since I did...

Hey - just started working on a prototype. I implemented a preliminary version of PGD for the smoothened l1 problem (huber) from the Trendafilov & Watson paper in Jax [here](https://colab.research.google.com/drive/1enh3AzhL2-G3LJczZdyMgX2enQr40b6p?usp=sharing)....

Sounds good. Yeah- those plots don't look right to me. Here is the result from my run:  Likely there is some issue with the implementation of the descent direction...

Hi, sorry for the late response. Quick follow-up. Took another look at my implementation. I think it may be possible that the implementation is okay - I reproduced the concrete...

Not going to make a new issue, but it's seems natural if GPs are included in the library to maybe also have some api for HMMs.

FYI: here is a minimal example: https://colab.research.google.com/drive/1Z802HuUZ_TCTRxeQNRvC6XJppEiGAiyQ?zusp=sharing For nearly singular matrices (e.g. conditioned graph Laplacians), cg returns different (sometimes valid, sometimes nan) solutions on GPU between different runs when x0...

Yes- so that's one thing. It's mainly an issue for poorly conditioned matrices. I guess typical and correct use cases wont suffer from this problem. But the main thing is...

I don't really have anything useful to add, but I found that cg and bicg were superior wrt to stability & much better wrt runtime on large sparse and large...

I'm also interested in collaborative best arm identification! Some simple meta-aggregate algorithms like [1, 2] might be a good place to start. https://arxiv.org/abs/1311.0800 https://arxiv.org/abs/1811.07763