DecimalFormat rounding changes in 8 (JEP 177)?
olivier.lagneau at oracle.com
olivier.lagneau at oracle.com
Mon May 5 06:55:53 UTC 2014
Hello,
On 04/05/2014 20:41, Joe Darcy wrote:
> Hello,
>
> On 5/4/2014 9:56 AM, solo-java-core-libs at goeswhere.com wrote:
>> Hey,
>>
>> Could someone please help me understand what changes have happened in
>> rounding in DecimalFormat in Java 8?
>>
>> new DecimalFormat("0.00").format(1.035) is 1.04 on Java 7, and 1.03 on
>> Java 8. (7u25-2.3.10-1~deb7u1, openjdk build 1.8.0_05-b13 debian and
>> Oracle build 1.8.0-b132 win64 tested).
>>
>> My understanding of the RoundingMode.HALF_EVEN (the default)
>> documentation is that 1.04 is the correct answer. In fact,
>> (0.000, 1.0035) is 1.004, and (0.0, 1.35) is 1.4. I am aware
>> that floating point is more complex than this, and I am by no
>> means an expert.
>
> There are several inexact processes going on here. The first is the
> decimal to binary conversion of 1.035 to a binary double value. In
> general, decimal fractions are not exactly representable in binary.
> Java's semantics require that on decimal to binary conversion, the
> double value with a numerical value closest to the exact value be
> used, the round to nearest even policy.
>
> The exact numerical value of the double closest to 1.035 is
>
> 1.0349999999999999200639422269887290894985198974609375
>
> a value slightly *less than* 1.035. When this value is rounded to two
> digits after the decimal point using the round to nearest even
> rounding mode, the numerically correct answer is 1.03 *not* 1.04.
>
> A range of numbers of the real line will get converted to a particular
> floating-point value. Some of these ranges straddle half-way points in
> decimal. For example, the range of values that will get converted to
> the floating-point value in question includes
>
> [1.0349999999999999200639422269887290894985198974609375, 1.035]
>
> The full range is
>
> [1.03499999999999980904163976447307504713535308837890625,
> 1.03500000000000003108624468950438313186168670654296875]
>
> This range is a one-ulp (unit in the last place, see Math.ulp) wide
> region centered on the exact floating-point value.
>
> When a decimal to binary conversion occurs, the original decimal text
> value is lost. Therefore, after the conversion, the binary double
> value doesn't "know" it came from "1.035" or "1.03499999999999981" or
> something else.
>
> The numerically correct behavior is the new behavior in JDK 8.
>
> HTH,
>
> -Joe
Just to insist on Joe's words (Thanks Joe for the detailed reply):
with floating-point "What You See *Is Not* What You Get" in most cases,
and this is true with DecimalFormat when formatting double or float values.
Don't expect exactness with floating-point. This is even true with constant
values inside you source code like 1.035 here (which cannot be represented
exactly in binary format).
JDK8 fixes a bug that was discovered during JDK7 dev.
Hope that helps,
Olivier.
>
>>
>>
>> It seems that this may be new code, with a known breaking change in it:
>>
>> http://openjdk.java.net/jeps/177:
>>> Compatibility: On JDK 7, (correct) but altered numerical behavior will
>>> only be enabled under an aggressive optimization flag to limit
>>> behavioral compatibility impact in that release train. In Java SE 8,
>>> the correct numerical behavior will always be required by the
>>> specification.
>> Did this materialise somewhere, and, if so, where's it documented?
>>
>>
>> In summary: My tests fail on Java 8 and I'm not sure they're wrong.
>> Any help would be appreciated, thanks.
>>
>
More information about the core-libs-dev
mailing list