ENH: Support for both close

[PaleNeutron]

Is your feature request related to a problem?

import pandas as pd
import piso
piso.register_accessors()
arr = pd.arrays.IntervalArray.from_tuples(
       [(1,5), (3,6), (2,4)], closed="both")
 
arr.piso.intersection()

Describe the solution you'd like

[[3, 4]]

In interval computation closed interval is a quite normal case, can you support that?

[venaturum]

Hi John, thanks for the interest.

Piso is intended to be a stop-gap until Pandas implements similar functionality. Under the hood, piso uses Staircase to perform a lot of operations, and this approach only facilitates half-open intervals. Adding the functionality for closed, or open intervals would require a complete rewrite to move away from Staircase and towards another approach - likely Interval Trees written in C code. As such it will remain out of scope for this project.

[PaleNeutron]

Thanks for reply! Does pandas already milestone similar functions?

[venaturum]

Not milestoned, but an open Issue.

I don't expect it will be actioned soon, I suspect it would require significant development to do properly.

[marade]

The docs for piso.bridge say that 'both' and 'neither' are accepted:

https://piso.readthedocs.io/en/latest/reference/api/piso.bridge.html

Can we get that updated please?

[venaturum]

Thanks for pointing that out @marade, will sort out by the end of the month when a new package is released.

[marade]

Thanks. I've found piso to be very useful. I wish pandas would adopt these features and others for Intervals.

[searchlie]

I too am one of those eagerly awaiting set operations for closed intervals.
At present, it seems there are no plans to implement this feature, which is somewhat disappointing.

So, how are you all currently handling set operations on closed intervals?
In my view, for union operations, closed intervals and half-open intervals produce almost identical results:

$$ (1,2] \cup (2,3] \rightarrow (1,3] $$ $$ [1,2] \cup [2,3] \rightarrow [1,3] $$

Therefore, if only union is being computed for closed intervals, I believe piso's current specification is sufficient.

On the other hand, for intersection operations, it clearly makes a difference.

$$ (1,2] \cap (2,3] \rightarrow \phi $$ $$ [1,2] \cap [2,3] \rightarrow 2 $$

What I'm currently considering is subtracting a small amount to the leftside of the half-open interval:

$$ (1 - \varepsilon, 2 ] \cap (2 - \varepsilon, 3] \rightarrow (2 - \varepsilon, 2] $$

Simply subtract $\varepsilon$ before using piso, and add $\varepsilon$ after using it.
Mathematically, the notation $(2,2]$ has no meaning, but since pd.Interval permits this expression, we can simply convert it to closed="both" later.

Does anyone have a better idea?

[PaleNeutron]

@searchlie I finally go to this project https://github.com/AlexandreDecan/portion

It is not array based, I think is much slower for thousands of rows but have full interval operations.

[searchlie]

@PaleNeutron
Thank you! I'll try it out straight away.
Come to think of it, I do feel that pandas compatibility isn't always necessary in many cases.