very cool - LGTM.
I am curious about actual speedup, do you happen to have any benchmark numbers comparing the two? In theory the O(log(n)) one is faster, but we don't know what the constant is...

Is there a reason to do this instead of using the identity 1 + ... + x^{n-1} = (1 - x^n) / (1 - x)?

It costs one inversion, but it's simple, and for the case that n is a power of 2 the x^n computation will just be repeated squarings.

Doesn't one inversion cost ≈200 multiplications? This implementation does (2..3)*lg n multiplications (2*lg(n) when n is power of two).

But i did not know about 1 + ... + x^{n-1} = (1 - x^n) / (1 - x), thanks for pointing it out!

Cool, I'm convinced this works but I'm not sure I understand why yet, so I'm working on some notes right now.

The power-of-two recurrence is really nicely explained and it makes sense, but I don't understand why the recurrence for general n works. Is it possible to prove it?

I just rebased this on main, for good hygiene, not with an expectation to merge this soon.

In fact, only RP code requires this helper, and in RP the input is always a power-of-two. For circuits, the delta is defined as an inner product of powers of a challenge with circuit-specific weights, so this whole function is not relevant.