These expressions give 1 - 1 as an alternative, instead of 0

[acoustoelectrically]

https://herbie.uwplse.org/demo/ddf7fb9e4a8c8bd9fa5bdbff8e69e2debb344ab1.1a83c9f5186db102b678a2d5f869f99992469071/graph.html

https://herbie.uwplse.org/demo/dd307cef8a2c9370cbe9d542578c6c1a1a6d965d.1a83c9f5186db102b678a2d5f869f99992469071/graph.html

We may also want to see if other similar expressions give any similar looking alternatives

[acoustoelectrically]

it reminds me of the absolute value bug

[pavpanchekha]

This is the same cause as the absolute value bug—Herbie can't constant eval through an approx. One option: fix that. But for the abs bug our conclusion was that actually we should just let Herbie series expand abs. The reason that doesn't work here is that the series expansion would be 0, which Herbie doesn't allow for series expansion. Maybe we should just allow it?

[pavpanchekha]

@acoustoelectrically copying what I said on Slack so you can find it more easily.

Why don't you try to modify Herbie, so that when doing a series expansion it also tries 0? That would probably fix this bug. The code that handles series expansions is in taylor-alt in src/core/patch.rkt. There's a nested for loop in that function; the innermost for loop (the one that goes up to (*taylor-order-limit*)) is the one that handles series expansions of different lengths. (So, for example, for exp(x) it would produces the series 1, 1 + x, 1 + x + x^2/2, and so on; I think it does up to four.)

What you're going to want to do, I think, is copy and paste the body of the loop to just before the loop; you want to go from:

(for (...)
  body ...)

to

body ...
(for (...)
  body ...)

Then in the new copy of the body outside the loop, there's a call to (genexpr), replace that with 0. (genexpr is what generates the series expansion; if we replace it with 0 we're just generating the series expansion 0.)

I think everything else in the new copy of the body can stay the same. Try that out and see if it fixes this bug. If you run into trouble let me know.