Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Why OrderedFloat is not IEEE conformant? #36

Open
kryptan opened this issue Oct 29, 2017 · 19 comments
Open

Why OrderedFloat is not IEEE conformant? #36

kryptan opened this issue Oct 29, 2017 · 19 comments

Comments

@kryptan
Copy link

kryptan commented Oct 29, 2017

For OrderedFloat:

NaN is sorted as greater than all other values and equal to itself, in contradiction with the IEEE standard.

IEEE standard requires different NaNs to be not equal, positive NaN to be greater than +∞ and negative NaN to be less than -∞. It also requires -0 to be less than +0. Why non-conformant behavior was chosen? Wouldn't it make more sense to implement total order as defined in standard instead?

This Stack Overflow answer is presumably providing a fast way to implement IEEE total order comparison.

@1011X
Copy link

1011X commented Jan 20, 2019

So according to what I could find about the definition of total order provided by IEEE 754, part of it would require that negative zero and positive zero compare differently from the default implementation. In other words, -0.0 < +0.0 returns false by default, but for total order it would have to be true. It also describes how different representations of the same floating point number would be ordered, and how NaNs with different payloads would compare with each other.

This looks like a big problem because Rust's documentation for Ord and PartialOrd explicitly say:

Implementations of PartialEq, PartialOrd, and Ord must agree with each other.

If IEEE-conformant total order is implemented using Ord, then it would disagree with the implementation of PartialOrd for f32 and f64, likely violating assumptions made by the standard library. I'm not sure what the exact consequences of that would be, but they don't sound good.

The only way I can think of to avoid this would be to have a separate trait specifically for floating-point total order.

(Btw, this is all based on what I could find from the section I linked above. If I misunderstood something or got something wrong, feel free to correct me.)

@kryptan
Copy link
Author

kryptan commented Jan 21, 2019

If IEEE-conformant total order is implemented using Ord, then it would disagree with the implementation of PartialOrd for f32 and f64

If we would implement Ord for f32 and f64 in standard library then yes, this would be a problem and PartialOrd should be compatible with Ord. But this library instead provides wrappers for f32 and f64, implementation of Ord and PartialOrd for these wrappers can agree with each other regardless of what standard library does for f32 and f64.

This is definition of total order from the standard:

5.10 Details of totalOrder predicate

For each supported arithmetic format, an implementation shall provide the following predicate that defines an ordering among all operands in a particular format:

boolean totalOrder(source, source)

totalOrder(x, y) imposes a total ordering on canonical members of the format of x and y:

a) If x<y, totalOrder(x, y) is true.
b) If x>y, totalOrder(x, y) is false.
c) If x=y:

  1. totalOrder(−0, +0) is true.
  2. totalOrder(+0, −0) is false.
  3. If x and y represent the same floating-point datum:
    i) If x and y have negative sign, totalOrder(x, y) is true if and only if the exponent of x≥ the exponent of y
    ii) otherwise totalOrder(x, y) is true if and only if the exponent of x≤ the exponent of y.

d) If x and y are unordered numerically because x or y is NaN:

  1. totalOrder(−NaN, y) is true where -NaN represents a NaN with negative sign bit and y is a floating-point number.
  2. totalOrder(x, +NaN) is true where +NaN represents a NaN with positive sign bit and x is a floating-point number.
  3. If x and y are both NaNs, then totalOrder reflects a total ordering based on:
    i) negative sign orders below positive sign
    ii) signaling orders below quiet for +NaN, reverse for -NaN
    iii) lesser payload, when regarded as an integer, orders below greater payload for +NaN,reverse for -NaN.

Neither signaling NaNs nor quiet NaNs signal an exception. For canonical x and y, totalOrder(x, y) and totalOrder(y, x) are both true if x and y are bitwise identical.

NOTE—totalOrder does not impose a total ordering on all encodings in a format. In particular, it does not distinguish among different encodings of the same floating-point representation, as when one or both encodings are non-canonical.

@Diggsey
Copy link

Diggsey commented Aug 7, 2019

@kryptan because people (myself included) still expect these wrappers to behave like a number. The aim is to be as close as possible to how ordering works for the native float types, whilst still upholding the guarantees of Ord. If negative zero compared differently to positive zero it would be useless for me and many other people.

It may make sense to have another type that implements the IEEE-defined ordering, but I suspect it will see less usage...

@DDtKey
Copy link

DDtKey commented Jul 7, 2024

Hi there!

I wanted to revive this topic. As I understand it, the main argument here is that Ord and PartialOrd may disagree in the case of following IEEE.

But I see a more important disagreement here between Eq and Div:

    // they are equal, but why?
    assert_eq!(OrderedFloat(f64::NAN), OrderedFloat(f64::NAN));
    // but division returns NaN (and that's correct!), but if they were truly equal it should result in 1.0
    assert_eq!(
        OrderedFloat(f64::NAN) / OrderedFloat(f64::NAN),
        OrderedFloat(f64::NAN)
    );

I don't think it's really a good idea to deviate from the standard here 🤔

Moreover, in hash sets they will be represented as single identical element, and that’s wrong. Such expectations may lead to many hidden issues

@ssokolow
Copy link

ssokolow commented Jul 7, 2024

I don't think it's really a good idea to deviate from the standard here

Right now, it looks like the question is "Is it possible to comply with one standard (IEEE) without breaking compliance with another standard (the invariants the Rust standard library and ecosystem were promised) and, if not, which standard has worse consequences for breakage?"

@DDtKey
Copy link

DDtKey commented Jul 7, 2024

But current implementation breaks mathematical rules and contradicts rust standard library (Div impl, as I mentioned above) 🤔

@ssokolow
Copy link

ssokolow commented Jul 7, 2024

Then it's a question of whether the two invariants in different parts of the standard library can be reconciled and, if not, which invariant is worse to break.

@DDtKey
Copy link

DDtKey commented Jul 7, 2024

In general, I think it can be interpreted in different ways:

Implementations of PartialEq, PartialOrd, and Ord must agree with each other.

rust-lang/rust#57157 (see follow-up PR there)

That is, a.cmp(b) == Ordering::Equal if and only if a == b and Some(a.cmp(b)) == a.partial_cmp(b) for all a and b.

However it looks like these mentions have been removed from documentation of Ord.

UPD:
Ok, current documentation of PartialOrd is pretty clear actually (updated here):

The methods of this trait must be consistent with each other and with those of PartialEq. The following conditions must hold:

  1. a == b if and only if partial_cmp(a, b) == Some(Equal).
  2. a < b if and only if partial_cmp(a, b) == Some(Less)
  3. a > b if and only if partial_cmp(a, b) == Some(Greater)
  4. a <= b if and only if a < b || a == b
  5. a >= b if and only if a > b || a == b
  6. a != b if and only if !(a == b).
    Conditions 2–5 above are ensured by the default implementation. Condition 6 is already ensured by PartialEq.

@mbrubeck
Copy link
Collaborator

mbrubeck commented Jul 8, 2024

but division returns NaN (and that's correct!), but if they were truly equal it should result in 1.0

There is no such invariant in floating point arithmetic or in Rust’s Div trait. For example, IEEE 754 requires that INFINITY is equal to itself, but also INFINITY / INFINITY is not 1.0.

Also, 0 == 0, but 0 / 0 is not 1.

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

For example, IEEE 754 requires that INFINITY is equal to itself, but also INFINITY / INFINITY is not 1.0.
Also, 0 == 0, but 0 / 0 is not 1.

Hm, right, that's correct 🤔

But these ones follows the standard at least (and division results generally based on math, there is an explanation for that). While NaN == NaN is very confusing part and breaks the standard and their nature in general.

In fact, NaNs are introduced by IEEE 754 (at least their wide usage).
And from math perspective it's undefined (which can't be equal to undefined). However, according to this argument, infinity should not be equal to infinity either, but it is in IEEE 😄

@mbrubeck
Copy link
Collaborator

mbrubeck commented Jul 8, 2024

But these ones follows the standard at least (and division results generally based on math, there is an explanation for that). While NaN == NaN is very confusing part and breaks the standard and their nature in general.

The standard IEEE 754 totalOrder predicate (which this issue is about) also includes cases where two NaN values have equal ordering. Note that the standard totalOrder predicate is not the same as the standard comparison predicates. That is necessary both for this crate and for IEEE, because the standard comparison predicates do not form a total order. I’m not sure exactly what you are proposing to solve this problem, since neither this crate’s ordering nor the IEEE standard ordering fulfill your requirements.

Since various requirements (standard comparisons, standard ordering, cmp::Eq invariants, etc.) are contradictory, it is impossible to offer a single type that meets all of them. We need multiple types, so users can choose an appropriate one for each use case:

  • Users who need standard IEEE 754 comparison behavior should use the primitive f64 and f32 types directly.
  • Users who need IEEE 754 totalOrder behavior can use (for example) the total_cmp methods from libcore.
  • Users who want a total ordering that is as consistent as possible with IEEE 754 comparison predicates can use the OrderedFloat type offered by this crate.
  • Users who want to avoid comparisons with NaN entirely can use the NotNan type offered by this crate.

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

First of all thanks for detailed answers!

The standard IEEE 754 totalOrder predicate (which this issue is about) also includes cases where two NaN values have equal ordering.

Could you point me to this part, please?

Note that the standard totalOrder predicate is not the same as the standard comparison predicates. That is necessary both for this crate and for IEEE, because the standard comparison predicates do not form a total order. I’m not sure exactly what you are proposing to solve this problem, since neither this crate’s ordering nor the IEEE standard ordering fulfill your requirements.

Actually, I'm referring to the comparison behavior. Because as I mentioned earlier, Eq impl of this crate returns true for two NaNs.

And this is not something expected to me. At least, Comparison with NaN claims:

But when at least one operand of a comparison is NaN, this trichotomy does not apply, and a fourth relation is needed: unordered. In particular, two NaN values compare as unordered, not as equal.

And:

The IEEE floating-point standard requires that NaN ≠ NaN hold.

Thus, I believe IEEE fulfill my expectations.

But these are good points actually:

  • Users who need standard IEEE 754 comparison behavior should use the primitive f64 and f32 types directly.
  • Users who need IEEE 754 totalOrder behavior can use (for example) the total_cmp methods from libcore.

And due to the contradictions, this is probably the only way:

it is impossible to offer a single type that meets all of them. We need multiple types

@mbrubeck
Copy link
Collaborator

mbrubeck commented Jul 8, 2024

The standard IEEE 754 totalOrder predicate (which this issue is about) also includes cases where two NaN values have equal ordering.

Could you point me to this part, please?

IEEE 754-2019 §5.10 states:

For canonical x and y, totalOrder(x, y) and totalOrder(y, x) are both true if x and y are bitwise identical.

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

In fact, OrderedFloat documentation claims it clearly:

NaN is sorted as greater than all other values and equal to itself, in contradiction with the IEEE standard.

So, I think there is no any issue here. As it's expected and explicitly stated.

It was just not entirely clear why this was so. But these contradictions explain the chosen behavior.

Thank you for quick and detailed replies!

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

Btw, I'm not sure there would be a contradiction if OrderedFloat(f64::NAN) == OrderedFloat(f65::NAN) would return false, but Ord would be Equal.

Based on Rust's documentation (see comment above), there is no mention about transitivity. It's seems logical to apply, but might not be necessary.

a == b if and only if partial_cmp(a, b) == Some(Equal).

It doesn't mean that a == b must be true for all partial_cmp(a, b) == Some(Equal). It tells us that a & b can be equal only if their ordering is Equal

WDYT? 🤔

@Diggsey
Copy link

Diggsey commented Jul 8, 2024

It doesn't mean that a == b must be true for all partial_cmp(a, b) == Some(Equal).

That's exactly what "if and only if" means.

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

That's exactly what "if and only if" means.

Oh, right, it's biconditional statement

Thanks, for some reason I totally ignored "iff"

@Diggsey
Copy link

Diggsey commented Jul 8, 2024

But there is no transitivity claimed. Note, it's about PartialEq & PartialOrd. Thus, I see it like:

Btw, I think you're talking about symmetry here rather than transitivity. Symmetry is the property that a <op> b implies b <op> a, whereas transitivity is the property that a <op> b && b <op> c implies a <op> c, so whilst iff is symmetric, transitive and reflexive (a <op> a == true) the relevant property here is symmetry.

@DDtKey
Copy link

DDtKey commented Jul 8, 2024

Yes, you're absolutely right!
Accurately noted, here, of course, symmetry was implied and “iff” fully corresponds to this.

In general, I have no more questions.
For now, I consider the current behavior as a feature that is explicitly stated, and I was just trying to understand it better

Thanks everybody for answers!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

6 participants