org.apache.commons.math.dfp.package.html Maven / Gradle / Ivy
Show all versions of commons-math Show documentation
Decimal floating point library for Java
Another floating point class. This one is built using radix 10000
which is 104, so its almost decimal.
The design goals here are:
- Decimal math, or close to it
- Settable precision (but no mix between numbers using different settings)
- Portability. Code should be keep as portable as possible.
- Performance
- Accuracy - Results should always be +/- 1 ULP for basic
algebraic operation
- Comply with IEEE 854-1987 as much as possible.
(See IEEE 854-1987 notes below)
Trade offs:
- Memory foot print. I'm using more memory than necessary to
represent numbers to get better performance.
- Digits are bigger, so rounding is a greater loss. So, if you
really need 12 decimal digits, better use 4 base 10000 digits
there can be one partially filled.
Numbers are represented in the following form:
n = sign × mant × (radix)exp;
where sign is ±1, mantissa represents a fractional number between
zero and one. mant[0] is the least significant digit.
exp is in the range of -32767 to 32768
IEEE 854-1987 Notes and differences
IEEE 854 requires the radix to be either 2 or 10. The radix here is
10000, so that requirement is not met, but it is possible that a
subclassed can be made to make it behave as a radix 10
number. It is my opinion that if it looks and behaves as a radix
10 number then it is one and that requirement would be met.
The radix of 10000 was chosen because it should be faster to operate
on 4 decimal digits at once instead of one at a time. Radix 10 behavior
can be realized by add an additional rounding step to ensure that
the number of decimal digits represented is constant.
The IEEE standard specifically leaves out internal data encoding,
so it is reasonable to conclude that such a subclass of this radix
10000 system is merely an encoding of a radix 10 system.
IEEE 854 also specifies the existence of "sub-normal" numbers. This
class does not contain any such entities. The most significant radix
10000 digit is always non-zero. Instead, we support "gradual underflow"
by raising the underflow flag for numbers less with exponent less than
expMin, but don't flush to zero until the exponent reaches MIN_EXP-digits.
Thus the smallest number we can represent would be:
1E(-(MIN_EXP-digits-1)*4), eg, for digits=5, MIN_EXP=-32767, that would
be 1e-131092.
IEEE 854 defines that the implied radix point lies just to the right
of the most significant digit and to the left of the remaining digits.
This implementation puts the implied radix point to the left of all
digits including the most significant one. The most significant digit
here is the one just to the right of the radix point. This is a fine
detail and is really only a matter of definition. Any side effects of
this can be rendered invisible by a subclass.