math/big: Rat: add FloatPrec() (int, bool)

[mitar]

Update, Nov 13, 2023: Current API is at #50489 (comment). -gri

---

Currently if I parse string like 123.34 with SetString() and want to print it back to full precision, one has to manually pick the right precision for FloatString() function to get that. I propose that instead FloatString() should accept negative prec parameter. The full semantics of prec parameter would then be:

• For positive prec, it formats with prec digits after the radix point, padding with 0 on the right if necessary. The last digit is rounded to nearest, with halves rounded away from zero, if rounding is necessary.
• When prec is 0, there are no digits after the radix point.
• If prec is less than 0, then:
  • If there is a finite number of digits after the radix point to precisely represent the number, all these digits are appended after the radix point.
  • If the number of digits is infinite, then abs(prec) number of digits is used after the radix point, rounded the last digit to nearest, with halves rounded away from zero.

So FloatString(3) for the number above returns 123.340, but FloatString(-3) return 123.34.

big.NewRat(1, 3).FloatString(3) returns 0.333 and big.NewRat(1, 3).FloatString(-3) would return 0.333.

[robpike]

I wrote a very long and complicated answer to this, but perhaps it's better just to state that rather than change FloatString in such a peculiar and possibly incompatible way, a better answer is to add a function that tells you how many bits of precision you'd need. The problem with that is that there is no simple answer in general, as you may have a repeating decimal, but the function could tell you that:

func (r *Rat) FloatSize() (digitsLeftOfDecimal, digitsRightOfDecimal int, repeats bool)

That solves the problem but I am far from convinced the problem is significant enough to require such a messy fix, either mine or yours. If you want to be clever, you can compute (10 gcd denominator) and get some of the answer yourself, since infinite-length representation only happens when the factors of the denominator are not all factors of the printing base.

[mitar]

That is a good point and I like the idea of FloatSize.

I am far from convinced the problem is significant enough to require such a messy fix

With messy fix you mean the negative precision proposal or FloatSize?

[robpike]

Both. As I said, "mine or yours".

[mitar]

I see.

The problem I have is that when marshaling and unmarshaling JSON with large decimal numbers as strings it is useful to have a way to parse them and then marshal them back into the same precision as they came in. Because they are parsed from a fixed length strings I know that I can format them back to one as well (so that there will be no repeats or infinite digits). The issue is only that there is no easy way to do that back conversion. On the other hand I do not want to lose precision if it is not necessary.

So having FloatSize would address that for me.

[mitar]

I made an implementation with signature func RatPrecision(rat *big.Rat) (int, int), where the first int is the number of non-repeating digits after decimal dot, and the second the number of repeating (cycling) digits which follow. Seems to work pretty well, maybe it could be included into stdlib.

[rsc]

This seems like an awkward use of big.Rat. Why use big.Rat in this case at all? Why not use json.Number as the round-trippable representation?

[mitar]

Oh, sorry if that was unclear from the proposal. The roundtrip is the sanity check motivation, if I have a data structure and I read it from the JSON into numbers with big.Rat as fields, I would at least like to be able to generate back the original JSON with some ease. And currently that is hard. Of course, the main use is that after parsing JSON number info a struct with big.Rat field, one would take that and do further processing on the struct. Potentially later on try to save it to JSON, but at that point values might not be representable without losing some precision. That is fine, but I would at least like to be able to save it to JSON without losing precision when that is not necessary (e.g., when there is a finite number of digits). But currently it is not possible using stdlib to render the big.Rat to a number knowing that you are not losing precision (when that is possible).

[rsc]

Do we care about the repeat size or the digits on the left?
The repeat size of N/D in particular can be D and therefore requires a big.Int to represent.
And the digits on the left is irrelevant to using FloatPrec.

The finite digits on the right is guaranteed to be 10 log D, which will fit in an int whenever D fits in memory.
So perhaps we should have

func (r *Rat) FloatPrec() (digits int, ok bool)

that returns 0, false for rats with non-finite float representations?

[mitar]

Yes, I agree that digits on the left are not important. The implementation I made here has the following signature:

func (r *Rat) FloatPrec() (digits int, repeating int)

So repeating == 0 when your ok == true.

But maybe my implementation is not as elegant as it could be.

But I would be OK with func (r *Rat) FloatPrec() (digits int, ok bool) as well.

In practice, using the signature from my implementation, I generally do l+j number of digits, so all non-repeating digits followed by one cycle of repeating digits.

[rsc]

@mitar the problem is that there can be >2^64 repeating digits.

[rsc]

This proposal has been added to the active column of the proposals project
and will now be reviewed at the weekly proposal review meetings.
— rsc for the proposal review group

[mitar]

the problem is that there can be >2^64 repeating digits.

Sure, the caller of this function would then decide what limits to apply before passing digits further. Or are you saying that int return value type is not enough? I think yes, we could flag somehow if that would overflow or something.