On Tue, 15 Apr 2014 09:52:08 0700, Phil Steitz wrote:
> On 4/15/14, 5:19 AM, Gilles wrote:
>> On Tue, 15 Apr 2014 08:03:51 +0900, Cyrille Artho wrote:
>>> Dear all,
>>> Java's default contract (in Object) states that two objects with
>>> equal data should return the same hashCode. In other words, if
>>> a.equals(b), their return value of hashCode() must be the same.
>>>
>>> Unfortunately, with Double values, new Double(0.0d).equals(new
>>> Double(0.0d)) is false. This is because their internal
>>> representation
>>> differs. Java's builtin Double thus returns a different hashCode
>>> for
>>> 0.0d and 0.0d.
>>>
>>> However, Apache's math library uses a mathematical comparison for
>>> Complex/Dfp (perhaps also others), where 0.0d == 0.0d. This breaks
>>> the contract, and thus causes problems when Complex or Dfp
>>> instances
>>> are used in containers such as HashMap, HashSet, etc. See the bug
>>> report for more details and test cases:
>>>
>>> https://issues.apache.org/jira/browse/MATH1118
>>>
>>> It is not quite clear how this should be fixed. Gilles posted a
>>> patch
>>> that uses Java's equals to perform the comparison. This fixes the
>>> behavior w.r.t. the contract but may surprise people who expect a
>>> mathematical comparison.
>>>
>>> Another possible fix (without a tentative patch at this time) would
>>> be to change hashCode:
>>>
>>> public int hashCode() {
>>> if (imaginary == 0.0d && real == 0.0d) {
>>> return ZERO.hashCode();
>>> }
>>> ...
>>>
>>> and similar for Dfp.
>>>
>>> This fixes the issue for zero values. However, there may be issues
>>> with normalized vs. unnormalized floating point values with the
>>> same
>>> (mathematical) value but different internal representations, where
>>> this kind of fix cannot be used. I'm not familiar enough with
>>> various
>>> floating point implementations to know if Java always normalizes
>>> the
>>> values before using them in a statement like hashCode().
>>>
>>> Therefore, the suggestion above is mathematically nice for +/
>>> 0.0d,
>>> but otherwise much less safe than Gilles' patch.
>>>
>>> Any comments? Probably best post them on JIRA:
>>>
>>> https://issues.apache.org/jira/browse/MATH1118
>>
>> IIUC, fixing the "hashCode" method as above would not solve the
>> problem
>> since it deals only with "0" whereas the unit test fails for "i".
>
> Thanks. I get that now.
>>
>> Do we agree that we first have to decide whether "Complex" should
>> behave
>> as "Double"?
>>
>> If the answer is "no" (as is the case now), shouldn't we document
>> that
>> "hash tables will not operate properly" (cf. Javadoc for "Double")?
>> And is this behaviour acceptable?
>
> I don't think having equal instances hashing to different values is
> acceptable. We have to do something to fix that.
>>
>> One possibility is to have the "equals(Complex)" behave according to
>> the JDK semantics, and add a static method "equals(Complex,
>> Complex)"
>> that would implement mathematical comparison (where "0" == "0").
>
> I like the idea of imitating the way Double works for equals.
>
I've uploaded[1] a tentative patch that
1. changes the behaviour of "equals(Object)", so that it has it has
the
same semantics as "Double",
2. adds static methods to test floatingpoint equality of instances
(by
delegating to the corresponding methods in class "Precision").
Regards,
Gilles
[1] https://issues.apache.org/jira/browse/MATH1118

To unsubscribe, email: devunsubscribe@commons.apache.org
For additional commands, email: devhelp@commons.apache.org
