Project

General

Profile

Actions

Bug #8883

closed

Rational canonicalization unexpectedly converts to Fixnum

Added by melquiades (Paul Cantrell) over 10 years ago. Updated over 10 years ago.

Status:
Closed
Assignee:
-
Target version:
-
ruby -v:
ruby 2.0.0p247 (2013-06-27 revision 41674) [x86_64-darwin12.3.0]
[ruby-core:57092]

Description

The documentation for Rational (http://www.ruby-doc.org/core-2.0.0/Rational.html) states that the result of creating or doing arithmetic on Rationals returns Rationals, as one would expect. Examples from the docs:

Rational(1)      #=> (1/1)
3.to_r           #=> (3/1)
Rational(-2, 9) * Rational(-9, 2)  #=> (1/1)

These all work as documented in 1.9. In 2.0, however, they all return Fixnum:

Rational(1)      #=> 1
3.to_r           #=> 3
Rational(-2, 9) * Rational(-9, 2)  #=> 1

This leads to unexpected behavior:

Rational(2) / Rational(3)  # => 0  ...but returns (2/3) in 1.9

That behavior is potentially dangerous. Math that may usually work, but suddenly start suffering from truncation errors depending on intermediate results. For example:

def should_always_return_one(a, b, c)
(Rational(a, c) + Rational(b, c)) / (a + b) * c
end

Under 1.9:

should_always_return_one(2, 3, 7) #=> (1/1)
should_always_return_one(2, 4, 7) #=> (1/1)
should_always_return_one(2, 5, 7) #=> (1/1)
should_always_return_one(2, 6, 7) #=> (1/1)

Under 2.0:

should_always_return_one(2, 3, 7) #=> 1
should_always_return_one(2, 4, 7) #=> 1
should_always_return_one(2, 5, 7) #=> 0   Oops!
should_always_return_one(2, 6, 7) #=> 1

Either the docs are wrong, or this is a bug. I vote bug. Whether arithmetic expressions truncate the result should not depend on whether intermediate values just happen to be integers! Such behavior renders Rational almost too dangerous to use in situations where exact results are required. (Yes, I realize that requiring 'mathn' fixes this, but even with such a workaround as an option, this is dangerously broken. See also #2121.) Note that floating point arithmetic does not exhibit this behavior.

Updated by phluid61 (Matthew Kerwin) over 10 years ago

=begin
I can't reproduce this in any version of Ruby that I have installed. What is your Ruby 2.0 patch level?

$ ruby2.0 -ve 'p Rational(2), Rational(3), Rational(2)/Rational(3)'
ruby 2.0.0p247 (2013-06-27 revision 41674) [x86_64-linux]
(2/1)
(3/1)
(2/3)
$ ruby2.0 -ve 'p (Rational(2,7)+Rational(5,7))/((2+5)/7)'
ruby 2.0.0p247 (2013-06-27 revision 41674) [x86_64-linux]
(1/1)
$ ruby2.1 -ve 'p Rational(2), Rational(3), Rational(2)/Rational(3)'
ruby 2.1.0dev (2013-08-27 trunk 42696) [x86_64-linux]
(2/1)
(3/1)
(2/3)
$ ruby2.1 -ve 'p (Rational(2,7)+Rational(5,7))/((2+5)/7)'
ruby 2.1.0dev (2013-08-27 trunk 42696) [x86_64-linux]
(1/1)

Edit: same behaviour on a fresh build of ruby 2.1.0dev (2013-09-10 trunk 42900) [x86_64-linux]
=end

Updated by nobu (Nobuyoshi Nakada) over 10 years ago

  • Category changed from core to lib

=begin
Rather, it seems caused by ((%mathn%)).

$ ~/ruby/1.9.3/bin/ruby -e 'p Rational(2)*Rational(1,2)'
(1/1)
$ ~/ruby/1.9.3/bin/ruby -rmathn -e 'p Rational(2)*Rational(1,2)'
1

$ ~/ruby/2.0.0/bin/ruby -e 'p Rational(2)*Rational(1,2)'
(1/1)
$ ~/ruby/2.0.0/bin/ruby -rmathn -e 'p Rational(2)*Rational(1,2)'
1
=end

Updated by nobu (Nobuyoshi Nakada) over 10 years ago

  • Status changed from Open to Rejected

Requiring only 'mathn/rational' causes this behavior.
It's a bug to use 'mathn/rational' solely.

Updated by david_macmahon (David MacMahon) over 10 years ago

But your previous example required just mathn:

$ ruby -rmathn -e 'p Rational(2,1)'
2

It seems like a mathn bug to me.

Dave

On Sep 9, 2013, at 9:56 PM, nobu (Nobuyoshi Nakada) wrote:

Issue #8883 has been updated by nobu (Nobuyoshi Nakada).

Status changed from Open to Rejected

Requiring only 'mathn/rational' causes this behavior.
It's a bug to use 'mathn/rational' solely.

Bug #8883: Rational canonicalization unexpectedly converts to Fixnum
https://bugs.ruby-lang.org/issues/8883#change-41708

Author: melquiades (Paul Cantrell)
Status: Rejected
Priority: Normal
Assignee:
Category: lib
Target version:
ruby -v: ruby 2.0.0p247 (2013-06-27 revision 41674) [x86_64-darwin12.3.0]
Backport: 1.9.3: UNKNOWN, 2.0.0: UNKNOWN

The documentation for Rational (http://www.ruby-doc.org/core-2.0.0/Rational.html) states that the result of creating or doing arithmetic on Rationals returns Rationals, as one would expect. Examples from the docs:

Rational(1) #=> (1/1)
3.to_r #=> (3/1)
Rational(-2, 9) * Rational(-9, 2) #=> (1/1)

These all work as documented in 1.9. In 2.0, however, they all return Fixnum:

Rational(1) #=> 1
3.to_r #=> 3
Rational(-2, 9) * Rational(-9, 2) #=> 1

This leads to unexpected behavior:

Rational(2) / Rational(3) # => 0 ...but returns (2/3) in 1.9

That behavior is potentially dangerous. Math that may usually work, but suddenly start suffering from truncation errors depending on intermediate results. For example:

def should_always_return_one(a, b, c)
(Rational(a, c) + Rational(b, c)) / (a + b) * c
end

Under 1.9:

should_always_return_one(2, 3, 7) #=> (1/1)
should_always_return_one(2, 4, 7) #=> (1/1)
should_always_return_one(2, 5, 7) #=> (1/1)
should_always_return_one(2, 6, 7) #=> (1/1)

Under 2.0:

should_always_return_one(2, 3, 7) #=> 1
should_always_return_one(2, 4, 7) #=> 1
should_always_return_one(2, 5, 7) #=> 0 Oops!
should_always_return_one(2, 6, 7) #=> 1

Either the docs are wrong, or this is a bug. I vote bug. Whether arithmetic expressions truncate the result should not depend on whether intermediate values just happen to be integers! Such behavior renders Rational almost too dangerous to use in situations where exact results are required. (Yes, I realize that requiring 'mathn' fixes this, but even with such a workaround as an option, this is dangerously broken. See also #2121.) Note that floating point arithmetic does not exhibit this behavior.

--
http://bugs.ruby-lang.org/

Updated by marcandre (Marc-Andre Lafortune) over 10 years ago

  • Status changed from Rejected to Open

david_macmahon (David MacMahon) wrote:

But your previous example required just mathn:

$ ruby -rmathn -e 'p Rational(2,1)'
2

It seems like a mathn bug to me.

Agreed.

Updated by marcandre (Marc-Andre Lafortune) over 10 years ago

  • Status changed from Open to Rejected

Mmm, sorry, misread.

I think the idea is that the buggy part (Rational(2) / Rational(3) # => 0) won't happen if you require 'mathn'

Updated by david_macmahon (David MacMahon) over 10 years ago

OK. I agree that requiring mathn avoids that buggy part. Thanks for clarifying. I guess I'm just a little uncomfortable with Rationals and Fixnums being promoted/demoted as needed, but maybe it's all OK and I'm just being paranoid.

While playing around with this, I see that integer Floats also have some special handling:

Without mathn...

$ ruby -e 'p [1/2.0, 1/2.5]'
[0.5, 0.4]

With mathn...

$ ruby -r mathn -e 'p [1/2.0, 1/2.5]'
[(1/2), 0.4]

Oddly though, this can result in non-reduced Rationals:

$ ruby -r mathn -e 'p [2/2.0, 2/2.5]'
[(2/2), 0.8]

Weird.

Also, why do integer Floats not get changed to Fixnums like Rational and Complex do?

$ ruby -r mathn -e 'p Rational(1).class'
Fixnum

$ ruby -r mathn -e 'p Complex(1).class'
Fixnum

$ ruby -r mathn -e 'p Float(1).class'
Float

Thanks,
Dave

On Sep 10, 2013, at 3:28 PM, marcandre (Marc-Andre Lafortune) wrote:

Issue #8883 has been updated by marcandre (Marc-Andre Lafortune).

Status changed from Open to Rejected

Mmm, sorry, misread.

I think the idea is that the buggy part (Rational(2) / Rational(3) # => 0) won't happen if you require 'mathn'


Bug #8883: Rational canonicalization unexpectedly converts to Fixnum
https://bugs.ruby-lang.org/issues/8883#change-41727

Author: melquiades (Paul Cantrell)
Status: Rejected
Priority: Normal
Assignee:
Category: lib
Target version:
ruby -v: ruby 2.0.0p247 (2013-06-27 revision 41674) [x86_64-darwin12.3.0]
Backport: 1.9.3: UNKNOWN, 2.0.0: UNKNOWN

The documentation for Rational (http://www.ruby-doc.org/core-2.0.0/Rational.html) states that the result of creating or doing arithmetic on Rationals returns Rationals, as one would expect. Examples from the docs:

Rational(1) #=> (1/1)
3.to_r #=> (3/1)
Rational(-2, 9) * Rational(-9, 2) #=> (1/1)

These all work as documented in 1.9. In 2.0, however, they all return Fixnum:

Rational(1) #=> 1
3.to_r #=> 3
Rational(-2, 9) * Rational(-9, 2) #=> 1

This leads to unexpected behavior:

Rational(2) / Rational(3) # => 0 ...but returns (2/3) in 1.9

That behavior is potentially dangerous. Math that may usually work, but suddenly start suffering from truncation errors depending on intermediate results. For example:

def should_always_return_one(a, b, c)
(Rational(a, c) + Rational(b, c)) / (a + b) * c
end

Under 1.9:

should_always_return_one(2, 3, 7) #=> (1/1)
should_always_return_one(2, 4, 7) #=> (1/1)
should_always_return_one(2, 5, 7) #=> (1/1)
should_always_return_one(2, 6, 7) #=> (1/1)

Under 2.0:

should_always_return_one(2, 3, 7) #=> 1
should_always_return_one(2, 4, 7) #=> 1
should_always_return_one(2, 5, 7) #=> 0 Oops!
should_always_return_one(2, 6, 7) #=> 1

Either the docs are wrong, or this is a bug. I vote bug. Whether arithmetic expressions truncate the result should not depend on whether intermediate values just happen to be integers! Such behavior renders Rational almost too dangerous to use in situations where exact results are required. (Yes, I realize that requiring 'mathn' fixes this, but even with such a workaround as an option, this is dangerously broken. See also #2121.) Note that floating point arithmetic does not exhibit this behavior.

--
http://bugs.ruby-lang.org/

Updated by marcandre (Marc-Andre Lafortune) over 10 years ago

david_macmahon (David MacMahon) wrote:

While playing around with this, I see that integer Floats also have some special handling:

Right. Floats are inexact while Integers & Rational are exact, and so are Complex with exact components. Rational(1/1) and 1 should yield the same mathematical result, but with floats that can be tricky. For example there are infinitely many different bigdecimals that will map to 1.0 (say 1.000....1 and 1.000...2 with enough zeros), but they don't behave exactly the same way, for example if you substract 1), so we can't freely map them.

Oddly though, this can result in non-reduced Rationals:

$ ruby -r mathn -e 'p [2/2.0, 2/2.5]'
[(2/2), 0.8]

Oh oh, that's a bug. It's not even related to 'mathn'. I opened a new issue about this: https://bugs.ruby-lang.org/issues/8894

Updated by david_macmahon (David MacMahon) over 10 years ago

On Sep 10, 2013, at 9:09 PM, marcandre (Marc-Andre Lafortune) wrote:

Issue #8883 has been updated by marcandre (Marc-Andre Lafortune).

david_macmahon (David MacMahon) wrote:

While playing around with this, I see that integer Floats also have some special handling:

Right. Floats are inexact while Integers & Rational are exact, and so are Complex with exact components. Rational(1/1) and 1 should yield the same mathematical result, but with floats that can be tricky. For example there are infinitely many different bigdecimals that will map to 1.0 (say 1.000....1 and 1.000...2 with enough zeros), but they don't behave exactly the same way, for example if you substract 1), so we can't freely map them.

That's all fine from a numerical/mathematical point of view, but it still seems like there is something missing from the duck typing:

$ ruby -e 'p (1/1.0).nan?'
false

$ ruby -r mathn -e 'p (1/1.1).nan?'
false

$ ruby -r mathn -e 'p (1/1.0).nan?'
-e:1:in <main>': undefined method nan?' for (1/1):Rational (NoMethodError)

Though admittedly this is getting a bit far from the OP.

Oddly though, this can result in non-reduced Rationals:

$ ruby -r mathn -e 'p [2/2.0, 2/2.5]'
[(2/2), 0.8]

Oh oh, that's a bug. It's not even related to 'mathn'. I opened a new issue about this: https://bugs.ruby-lang.org/issues/8894

Thanks!

Dave

Updated by melquiades (Paul Cantrell) over 10 years ago

Somewhere in all the discussion, the actual bug got lost. This issue shouldn't be closed.

To clarify:

(1) The bug occurs when you do not include mathn, and has nothing to do with mathn.

(2) The bug occurs when you include nothing at all:

$ ~/.rvm/rubies/ruby-2.0.0-p247/bin/ruby -e 'p Rational(2) / Rational(3)'
0

That is clearly wrong. The should_always_return_one example demonstrates why this behavior is terribly dangerous, and is probably causing mathematically incorrect results in production code right now for poor unsuspecting souls out there in the world.

(3) The bug does not occur in 1.9.3:

$  ~/.rvm/rubies/ruby-1.9.3-p448/bin/ruby -e 'p Rational(2) / Rational(3)'
(2/3)

Bottom line: promoting the results of Rational calculations to Fixnum is never safe without mathn, not ever, and Ruby should never do it.

Updated by marcandre (Marc-Andre Lafortune) over 10 years ago

Hi

melquiades (Paul Cantrell) wrote:

(2) The bug occurs when you include nothing at all:

$ ~/.rvm/rubies/ruby-2.0.0-p247/bin/ruby -e 'p Rational(2) / Rational(3)'
0

I can't reproduce this, with ruby 2.0.0p247, p195 nor trunk.

Updated by nagachika (Tomoyuki Chikanaga) over 10 years ago

Hello, melquiades

Don't you build your binary with --with-static-linked-ext ?
A similar issue is reported when extension library mathn/rational is statically linked.
See #8879

If so, require "mathn" explicitly ease the problem.

Updated by nobu (Nobuyoshi Nakada) over 10 years ago

  • Status changed from Rejected to Closed

Updated by nagachika (Tomoyuki Chikanaga) over 10 years ago

  • Backport changed from 1.9.3: UNKNOWN, 2.0.0: UNKNOWN to 1.9.3: UNKNOWN, 2.0.0: DONE

r43449, r43514 and r43525 are backported to ruby_2_0_0 at r43656.

Updated by melquiades (Paul Cantrell) over 10 years ago

@nagachika (Tomoyuki Chikanaga): Yes, your guess is correct. I am using rvm, which passes --with-static-linked-ext.

I verified that patch 43656 does indeed fix the issue:

$ rvm install 2.0.0-patch43656 --patch changeset_r43656.diff
...
 ~/.rvm/rubies/ruby-2.0.0-p247-patch43656/bin/ruby -e 'p Rational(2) / Rational(3)'
(2/3)

Hooray!

(Apologies for my slow responses. Apparently I'm not receiving email notifications on this thread, despite having watched it.)

Actions

Also available in: Atom PDF

Like0
Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0Like0