the formula for layout optimization is slightly different from that in UMAP paper?

Nobody has claimed this yet.

Hi,
Just trying to understand the layout optimization part of code, I find that the formula for positive edges here:
` if dist_squared > 0.0:
grad_coeff = -2.0 * a * b * pow(dist_squared, b - 1.0)
grad_coeff /= a * pow(dist_squared, b) + 1.0
else:
grad_coeff = 0.0`
is different from what explained in https://arxiv.org/pdf/1802.03426.pdf chapter 3.2:
I understand the W is missing because we are doing importance sampling, but the denominator is also slightly different.

Also for negative edges, we do something like this:
` if dist_squared > 0.0:
grad_coeff = 2.0 * gamma * b
grad_coeff /= (0.001 + dist_squared) * (
a * pow(dist_squared, b) + 1
)
elif j == k:
continue
else:
grad_coeff = 0.0`

but in the paper, it is:

Am I missing something? Or perhaps the version in the code works better?

Thanks a lot in advance

Start by locating the positive- and negative-edge layout optimization code shown in the issue, then compare both gradient formulas with chapter 3.2 of the UMAP paper. Determine whether the denominator differences are intentional consequences of importance sampling or an implementation error; done means the discrepancy is explained or corrected with supporting validation.

Written by the indexing model from the issue text.