Unable to read complex numeric literals with NaN as one of the components

[Oxyd]

> 0+nan.0i
ERROR on line 23: invalid numeric syntax: #\n
> ERROR: undefined variable: an.0i
> +nan.0i
ERROR: undefined variable: |+nan.0i|
> +nan.0+1i
ERROR: undefined variable: |+nan.0+1i|
> +nan.0+i
ERROR: undefined variable: |+nan.0+i|
> +nan.0+nan.0i
ERROR: undefined variable: |+nan.0+nan.0i|

Even though the analogous syntax with infinites works:

> 0+inf.0i
0+inf.0i
> +inf.0i
0+inf.0i
> +inf.0+1i
+inf.0+i
> +inf.0+i
+inf.0+i
> +inf.0+inf.0i
+inf.0+inf.0i

[Oxyd]

Additional observation: Chbi can't write complex numbers with NaN's in them either:

> (make-rectangular 0 +nan.0)
0++nan.0i
> (make-rectangular 0 +inf.0)
0+inf.0i

Should be just a single + for the first output, like it is for the second number.

And, for completeness, Chibi can't read its own write output in this case:

> 0++nan.0i
ERROR on line 10: invalid numeric syntax: #\n
> ERROR: undefined variable: an.0i

[ashinn]

The standard is not strict on this - most of the numeric tower and any NaN support at all is optional.

From a mathematical perspective I am opposed to the notion of any number that is somehow "partially" not a number. From a practical perspective, everyone else does it, and it makes certain things easier.

I have to decide whether to make those make-rectangular calls return a single real +nan.0, or break down and fix read/write.

[Oxyd]

Well the standard says that implementations must “implement a coherent subset consistent with both the purposes of the implementation and the spirit of the Scheme language” – I'm not sure if supporting NaN's but not supporting the read syntax for complex NaN's counts as a coherent subset.

Even if Chibi were to collapse complex numbers with any component being NaN to just a single real NaN, I'd expect it to be able to read such complex numbers. I.e. 1+nan.0i should evaluate to +nan.0 instead of giving a read error.

From a practical perspective I don't see any gains from doing such collapsing. If anything, not collapsing complex numbers in that way should be slightly easier to implement. From a mathematical perspective, I find numbers that are not numbers already objectionable, so if one accepts those, one might as well accept them as components of complex numbers.

[ashinn]

If we take the view that there are no partial NaN's, the logic would be to collapse intermediate computations resulting this to a single real NaN. Explicitly trying to create such invalid values with make-rectangular or read syntax would then logically signal an error.

However, I realize the ship has sailed with SRFI 160. Even if I try to patch up the homogeneous complex vectors to respect this behavior, other libraries such the usage of BLAS in (chibi math linalg) exposes these values.