cowan

127.10.177.1

WG1Ballot9Results

0

= Notes about Results = See [wiki:WG1BallotExplanation WG1BallotExplanation]. == WG1 - Core == === #460 Semantics of `eqv?` === The semantics of `eqv?` are highly contended, and as a result we are bringing this up for a third and final vote. The earlier votes on #125 (`eqv?` may not return true for procedures) and #229 (`eqv?` is unspecified for NaNs) were voted on and treated as orthogonal. There have been no objections to these, and so the results still stand. We're only focusing on the core `eqv?` semantics for inexact numbers. This is fundamentally a tale of three equivalences: mathematical, operational, and representational. {{{ Ultra-brief history: R2RS was purely mathematical, defining `eqv?` on numbers in terms of `=`. R3RS defined a simple operational equivalence by distinguishing exact and inexact. This was not a complete operational equivalence because R3RS already had multiple precisions and implicit NaNs and signed zeros. R[45]RS dropped the discussion of operational equivalence but kept the exactness separate. R6RS re-introduced the notion of operational equivalence, this time with a complete definition. The R7RS 7th draft introduced an incomplete notion of representational equivalence - two numbers are `eqv?` iff they have the same IEEE-style representation. }}} For this final vote there are three proposals under consideration, corresponding to the complete forms of each equivalence. Variations may be proposed and added if a suitable rationale is given. The `r5rs` proposal follows R5RS exactly, in the spirit of mathematical equivalence (assuming exacts and inexacts are different types to begin with). The advantage of this is it's very simple and it appeals to idealists - people who want to pretend that they are computing real mathematical values without such thing as NaNs or negative zeros. The disadvantage of this is, as with most things that appeal to idealists, it does not match reality. Our computers use crude hacks for efficiency, and even if someone manages to build an ideal Scheme, it will likely be impractical and most implementations will continue to use those hacks. Moreover, mathematical equivalence is already available via the `=` procedure. This is not a practical equivalence relation for a standard. The text for the true case of `eqv?` for inexact numbers under `r5rs` is: {{{ (3) obj1 and obj2 are both numbers, have the same exactness, and are numerically equal (see `=`). }}} The `r6rs` proposal follows R6RS exactly in the spirit of operational equivalence (with a small correction to avoid making everything unspecified via NaN transitivity). The advantages of this is that it's exactly what you want to distinguish if two values will always behave the same, for example for compiler optimizations or memoization. The disadvantage is that the definition is complicated and difficult to nail down - it doesn't account for non-standard extensions an implementation may provide which could distinguish certain new values. The `r6rs` text is: {{{ (3.1) obj1 and obj2 are both exact numbers and are numerically equal (see `=`) (3.2) obj1 and obj2 are both inexact numbers, are numerically equal (see `=`), and yield the same results (in the sense of `eqv?` and excluding `+nan.0`) when passed as arguments to any other procedure that can be defined as a finite composition of Scheme’s standard arithmetic procedures. }}} Finally, the `representational` proposal is based on the previous `same-bits` in the spirit of representational equivalence. Two numbers are `eqv?` if they are represented the same way. This is potentially finer grained than operational equivalence - it may in fact make useless distinctions, but it is generally safer to over-distinguish than to under-distinguish. The `representational` text is: {{{ (3.1) obj1 and obj2 are both exact numbers and are numerically equal (see `=`) (3.2) obj1 and obj2 are both numbers of the same composite numeric type, whose corresponding fields are all `eqv?` * numbers in the style of the IEEE 754-2008 floating point standard are considered composites of their radix, precision, maximum exponent, sign, exponent, and significand * non-real complex numbers are composites of their real and imaginary parts }}} * '''References:''' * [https://groups.google.com/d/msg/scheme-reports-wg1/BGvDFtD6A1M/5pHmfXHtvEIJ eqv? issues summarized] * [https://groups.google.com/d/msg/scheme-reports-wg1/2Nv6oIND8HI/Z2HXPQMNFooJ the history of eqv? on numbers] * [http://lists.scheme-reports.org/pipermail/scheme-reports/2012-November/002914.html Weaver's objection] * [http://www.math.purdue.edu/~lucier/r7rs-eqv-discuss Lucier on IEEE 754] * '''Options:''' r5rs, r6rs, representational * '''Default:''' r5rs * '''Voters:''' * [wiki:WG1BallotCowan Cowan]: r6rs, representational, r5rs * [wiki:WG1BallotGanz Ganz]: r6rs, representational, r5rs * [wiki:WG1BallotGleckler Gleckler]: representational, r6rs, r5rs * [wiki:WG1BallotHsu Hsu]: r6rs, representational, r5rs * [wiki:WG1BallotLucier Lucier]: representational, r6rs, r5rs * [wiki:WG1BallotMedernach Medernach]: r6rs, representational, r5rs * [wiki:WG1BallotSnellPym SnellPym]: r6rs, representational, r5rs * '''Results:''' r6rs, representational * '''Ratios:''' 4:3 * '''Rationales:''' `Cowan`:: I think the `r5rs` choice is Just Wrong as long as IEEE is the dominant representation of inexact numbers: there needs to be a way to distinguish between `0.0` and `-0.0` for memoization and other purposes, as they behave differently for some arithmetic procedures. For IEEE (always excepting !NaNs, which are implementation-dependent), `r6rs` and `representational` mean the same thing: with only one bit pattern per operationally distinct number, `representational` can't discriminate between them. In the non-IEEE space, I think the `r6rs` choice is more principled, but I can live with `representational`, though it may end up discriminating between non-IEEE numbers that aren't really different. `Gleckler`:: The representational definition clearly and concretely defines what an implementer should do to implement `eqv?` for inexact numbers, which makes it better than the r6rs option, and it takes into account issues like precision, which makes it better than the r5rs option. The r6rs option is better than the r5rs option for the same reason. Thus my ordering of preferences. `Hsu`:: I want operational equivalence in as implementation neutral a fashion that abstracts as much as possible away from implementational details. While r6rs may appear more subtle, the fact that it does not define itself in terms of implementation details directly is a plus for me. `Lucier`:: The representational definition clearly and concretely defines what an implementer should do to implement `eqv?` for inexact numbers, which makes it better than the r6rs option, and it takes into account issues like precision, which makes it better than the r5rs option. The r6rs option is better than the r5rs option for the same reason. Thus my ordering of preferences. `Medernach`:: IMHO the only sensible choice is operational equivalence, at the end of the day it is all about one implementation being consistent in its choice of equivalence classes. This left an open door for implementation and at the same time avoid inconvenience to users. '''representational''' is maybe too low-level, I mean the underlying concept here is about equivalence class and I would let implementors freedom/discretion of how they check for it. Again and of course de-composition is the way to go most of the time. I am sure the 2 proposals are not too far from each other. I suggest therefore that an editorial note may be added to the '''r6rs''' option saying something along the line of that: ''One possible implementation is to check subpart eqv?-eness if it is consistent with equivalence classes wished to be provided by the implementation.'' `SnellPym`:: I think that r6rs and representational will be equivalent in most implementations; certainly, I think an implementation should strive as best it can to provide operation equivalence of numbers, and in bizarre cases where it cannot (numbers represented as infinite lazy streams, perhaps? `(exact? pi)` might be fun), I'd rather that the implementation state that it's not completely R7RS conformant in this case (I'm sure this won't be the *only* numeric semantic surprise such an implementation has, and probably not even the biggest...) than be forced to go for some ugly implementation-exposing hack.

2013-07-07 01:23:48

1