jean-eric

Does this snippet is correct when using matplotlib Ellipse ? ```python import matplotlib.patches as mpatches minor,major,angle = props.axis_minor_length,props.axis_major_length,props.orientation*180./np.pi ellipse = mpatches.Ellipse(xy=(x, y),fill=False,width=major,height=minor,angle=angle,color="red") ```

Hi @mblondel you may be interested in the discussion [jax/jit](https://github.com/google/jax/discussions/10648) By the way, what is the difference with the [jax.scipy BFGS solver](https://jax.readthedocs.io/en/latest/jax.experimental.sparse.html?highlight=BFGS#example-sparse-logistic-regression), does it is the bounds?

I have coded the linear approx with the normalization of (a,b) mentioned in the last post, and the pytest is passed. So one can have a linear approx with a...

Hi, I'm quite interested by this new implementation. It would wonderful if the new code could be differentiated with the use of `jacrev` at least wrt to z argument.

Hi, I have updated [my Colab nb](https://colab.research.google.com/drive/1Cr7oDffnpc6q_hUqNYI423nzjo8MQvn3?usp=sharing) where I show an implementation of J0 and J1 that matches the sc.special.j0 & j1 and fulfils the use of jit/vmap/jacrev. This is...

Hi Notice also instabilities at high `x` which incline to use high n_iter but this is in conjunction with`nan` at low `x`: ```python N=300 x = jnp.linspace(0,N,50*N) jscs.bessel_jn(x, v=0, n_iter=200)...

Here may be a solution for J0(x) using discussion #14132 and Boost code; ```python def _evaluate_rational(z, num, denom): count = len(num) def true_fn_update(z): def body_true(val1): # decode val1 s1, s2,...

If you can simplify the code and improve it , I would be glad... Or may be it is already a better implementation. Moreover, Scipy may be not the most...

Well in Angpow I had a mechanism to generate ell list like ``` ell=0/1/2/3/4/5/6/7/8/9/10/11/12/13/14/16/18/20/22/25/28/32/36/41/47/54/62/71/81/93/106/121/139/159/182/209/240/275/315/355/395/435/475/515/555/595/635/675/715/755/795/835/875/915/955/995/999/ ``` with a linear part at low ell and a logscale a high ell. Then, from...

I have cooked a Google Collab nb to show what can be done: [have a look here ](https://colab.research.google.com/drive/1007W_evrn6q4UjCoeLsSp0vixegblTtI?usp=sharing) With ~ 50 ells sampling values in the range [2, 1000] for...