Christian Lorentzen

If we want, we could go on and optimize with multi-threaded Cython code like ```Cython def sqrt_D_Lt_matmul_cython( double[::1, :] x, const double[::1, :] p, const double[::1, :] q_inv, int n_threads=1,...

Another benchmark, this time with the 20 newsgroup dataset (vectorized): `n_classes=20`, `n_features=130_107`, `n_samples=18_846` and sparse `X`.  ```python import warnings from pathlib import Path import numpy as np from scipy...

After an important improvement in commit e7368e775c5fb5ea34f18022c0141b218d0212a7 that sets a better inner stopping criterion with a forcing sequence like in Newton CG, the benchmarks are looking much better, even impressive!...

> This looks very impressive. However it seems that there are missing results and code snippet for the 20 newsgroups case in the previous comment On purpose, note the TODO....

> so that we can get a finer insight on a robust strategy to set the lsmr stopping conditions. Honestly, we don’t have a good infrastructure for that and I...

> For GLMs with non-canonical link functions (e.g. Gamma and Tweedie with p > 1, the log link is not the canonical link) the Hessian is not equal to the...

Commit https://github.com/scikit-learn/scikit-learn/pull/25462/commits/83ce34fc7a528f096718bfff03f07d94253d3b0b passes `SKLEARN_TESTS_GLOBAL_RANDOM_SEED="all" pytest -n auto -rfE sklearn/linear_model/_glm/tests/test_glm.py` locally on my laptop.

I opened #26707 for investigating the inner solver stopping criterion and run a log of benchmarks. There is no clear winner. I have to leave it as is: Either it...

The first (non-zero) value is `x = 1.452e+5` and `alpha=-1`. Now, `wright_bessel(1, 0, 1e5) = 2.356e+275` and `np.finfo(np.float64).max = 1.798e+308`. So the simple answer is that `wright_bessel(1, 0, 1.4e5)` is...

That's a question for @josef-pkt to answer and it also depends on you expected timeline. I will only pursue the "proper" solution with scipy. Let's see how they respond.