[
https://issues.apache.org/jira/browse/NUMBERS-120?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16877281#comment-16877281
]
Heinrich Bohne commented on NUMBERS-120:
The fix in PR #63 should make the methods {{doubleValue()}} and
{{floatValue()}} _always_ calculate that {{float}}/{{double}} value that is
closer to the actual value of the fraction than any other {{float}}/{{double}}
value, rounding the result according to IEEE 754's round-to-nearest,
ties-to-even rounding mode, which is the same rounding mode used by primitive
floating-point arithmetic operations. The code is, unfortunately, a bit
convoluted, but it's the best way I could come up with to ensure maximum
precision even in "mad" corner cases like 2^54^ / (2^53^ + 1).
Also, I am not entirely convinced that I did not, at some point, re-invent the
wheel, for example with the helper method {{roundAndRightShift}}. But I don't
think the class {{BigDecimal}} would have been of much help here, because it is
designed for decimal numbers and not binary numbers, and other than that, I
don't know what other already existing code I could have used.
> Major loss of precision in BigFraction.doubleValue() and
> BigFraction.floatValue()
> -
>
> Key: NUMBERS-120
> URL: https://issues.apache.org/jira/browse/NUMBERS-120
> Project: Commons Numbers
> Issue Type: Bug
> Components: fraction
>Affects Versions: 1.0
>Reporter: Heinrich Bohne
>Priority: Minor
> Time Spent: 10m
> Remaining Estimate: 0h
>
> The method {{BigFraction.doubleValue()}} calculates the double value of
> fractions with numerators or denominators that, when converted to a
> {{double}}, round up to {{Double.POSITIVE_INFINITY}}, by right-shifting both
> the numerator and denominator synchronously until both numbers fit into 1023
> bits. Apart from the fact that the maximum number of bits an integer
> representable as a finite {{double}} can have is 1024 (an unbiased exponent
> of 1023, which is the largest possible unbiased exponent of a {{double}}
> number, means 1. ⋅ 2^1023^, which amounts to 1024 bits), this way of
> converting the fraction to a {{double}} is incredibly wasteful with precision
> if the numerator and denominator have a different bit length, because the
> smaller of the two numbers will be truncated beyond what is necessary to
> represent it as a finite {{double}}. Here is an extreme example:
> The smallest integer that rounds up to {{Double.POSITIVE_INFINITY}} when
> converted to a {{double}} is 2^1024^ - 2^970^. This is because
> {{Double.MAX_VALUE}} as an integer is a 1024-bit number with the most
> significant 53 bits set to 1 and all other 971 bits set to 0. If the 970
> least significant bits are changed in any way, the number will still round
> down to {{Double.MAX_VALUE}} as long as the 971st bit remains 0, but as soon
> as the 971st bit is set to 1, the number will round up to
> {{Double.POSITIVE_INFINITY}}.
> The smallest possible denominator greater than 1 where a single right-shift
> will cause a loss of precision is 3. 2^1024^ - 2^970^ is divisible by 3, so
> in order to create an irreducible fraction, let's add 1 to it:
> (2^1024^ - 2^970^ + 1) / 3 ≈ 5.992310449541053 ⋅ 10^307^ (which can be
> verified with {{BigDecimal}}, or, more easily, with [this online
> tool|https://www.wolframalpha.com/input/?i=(2%5E1024+-+2%5E970+%2B+1)+%2F+3].
> However, the current implementation of BigFraction.doubleValue() returns
> 8.98846567431158 ⋅ 10^307^, which differs from the correct result by a
> relative error of 50%! The same problem applies to the method
> {{BigFraction.floatValue()}}.
> This can be prevented by truncating the numerator and denominator separately,
> so that for each of the two numbers, the maximum possible precision is
> retained.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)