copysign/signum/whatever to differentiate ±0.0?

[XeCycle]

Description

Currently we have sign() which is the same as numpy.sign, in that it converts 0.0 and -0.0 to 0.0. In the absence of copysign I failed to find any tools to differentiate 0.0 and -0.0, unless I take that series and call numpy.copysign on it. Can we have expr.copysign(expr) so that I can polars.lit(1).copysign(polars.col("c"))?

[orlp]

Polars doesn't respect the difference between 0.0 and -0.0 in general in its data model: https://docs.pola.rs/user-guide/concepts/data-types-and-structures/#floating-point-numbers.

Why do you want this?

[XeCycle]

These values may be parsed from feather/parquet/whatever; in an IPC application using arrow streaming format, if this sign bit can be used to mark two different cases, I can drop a boolean column and thus reduce memory usage and/or IO transfers. I agree that's not a good design, and it isn't usually useful but IMO it doesn't hurt to include it?

[orlp]

I don't think that's a good enough reason, sorry. And as mentioned before, we don't actually respect the sign bit for zeroes/NaNs in Polars and may drop it arbitrarily, so your idea wouldn't work robustly in Polars either.