cpython/Doc/tutorial/floatingpoint.rst

293 lines
11 KiB
ReStructuredText
Raw Normal View History

2007-08-15 11:28:22 -03:00
.. _tut-fp-issues:
**************************************************
Floating Point Arithmetic: Issues and Limitations
**************************************************
.. sectionauthor:: Tim Peters <tim_one@users.sourceforge.net>
Floating-point numbers are represented in computer hardware as base 2 (binary)
fractions. For example, the decimal fraction ::
0.125
has value 1/10 + 2/100 + 5/1000, and in the same way the binary fraction ::
0.001
has value 0/2 + 0/4 + 1/8. These two fractions have identical values, the only
real difference being that the first is written in base 10 fractional notation,
and the second in base 2.
Unfortunately, most decimal fractions cannot be represented exactly as binary
fractions. A consequence is that, in general, the decimal floating-point
numbers you enter are only approximated by the binary floating-point numbers
actually stored in the machine.
The problem is easier to understand at first in base 10. Consider the fraction
1/3. You can approximate that as a base 10 fraction::
0.3
or, better, ::
0.33
or, better, ::
0.333
and so on. No matter how many digits you're willing to write down, the result
will never be exactly 1/3, but will be an increasingly better approximation of
1/3.
In the same way, no matter how many base 2 digits you're willing to use, the
decimal value 0.1 cannot be represented exactly as a base 2 fraction. In base
2, 1/10 is the infinitely repeating fraction ::
0.0001100110011001100110011001100110011001100110011...
2009-04-24 00:09:06 -03:00
Stop at any finite number of bits, and you get an approximation. On most
machines today, floats are approximated using a binary fraction with
the numerator using the first 53 bits following the most significant bit and
with the denominator as a power of two. In the case of 1/10, the binary fraction
is ``3602879701896397 / 2 ** 55`` which is close to but not exactly
equal to the true value of 1/10.
Many users are not aware of the approximation because of the way values are
displayed. Python only prints a decimal approximation to the true decimal
value of the binary approximation stored by the machine. On most machines, if
Python were to print the true decimal value of the binary approximation stored
for 0.1, it would have to display ::
2007-08-15 11:28:22 -03:00
>>> 0.1
2009-04-24 00:09:06 -03:00
0.1000000000000000055511151231257827021181583404541015625
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
That is more digits than most people find useful, so Python keeps the number
of digits manageable by displaying a rounded value instead ::
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
>>> 1 / 10
0.1
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
Just remember, even though the printed result looks like the exact value
of 1/10, the actual stored value is the nearest representable binary fraction.
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
Interestingly, there are many different decimal numbers that share the same
nearest approximate binary fraction. For example, the numbers ``0.1`` and
``0.10000000000000001`` and
``0.1000000000000000055511151231257827021181583404541015625`` are all
approximated by ``3602879701896397 / 2 ** 55``. Since all of these decimal
2009-04-24 16:06:29 -03:00
values share the same approximation, any one of them could be displayed
2009-04-24 00:09:06 -03:00
while still preserving the invariant ``eval(repr(x)) == x``.
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
Historically, the Python prompt and built-in :func:`repr` function would chose
the one with 17 significant digits, ``0.10000000000000001``, Starting with
Python 3.1, Python (on most systems) is now able to choose the shortest of
these and simply display ``0.1``.
2007-08-15 11:28:22 -03:00
Note that this is in the very nature of binary floating-point: this is not a bug
in Python, and it is not a bug in your code either. You'll see the same kind of
thing in all languages that support your hardware's floating-point arithmetic
(although some languages may not *display* the difference by default, or in all
output modes).
Python's built-in :func:`str` function produces only 12 significant digits, and
2007-08-15 11:28:22 -03:00
you may wish to use that instead. It's unusual for ``eval(str(x))`` to
reproduce *x*, but the output may be more pleasant to look at::
2009-04-24 00:09:06 -03:00
>>> str(math.pi)
'3.14159265359'
>>> repr(math.pi)
'3.141592653589793'
>>> format(math.pi, '.2f')
'3.14'
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
It's important to realize that this is, in a real sense, an illusion: you're
simply rounding the *display* of the true machine value.
2007-08-15 11:28:22 -03:00
One illusion may beget another. For example, since 0.1 is not exactly 1/10,
summing three values of 0.1 may not yield exactly 0.3, either::
>>> .1 + .1 + .1 == .3
False
Also, since the 0.1 cannot get any closer to the exact value of 1/10 and
0.3 cannot get any closer to the exact value of 3/10, then pre-rounding with
:func:`round` function cannot help::
>>> round(.1, 1) + round(.1, 1) + round(.1, 1) == round(.3, 1)
False
Though the numbers cannot be made closer to their intended exact values,
the :func:`round` function can be useful for post-rounding so that results
have inexact values that are comparable to one another::
>>> round(.1 + .1 + .1, 1) == round(.3, 1)
True
2007-08-15 11:28:22 -03:00
Binary floating-point arithmetic holds many surprises like this. The problem
with "0.1" is explained in precise detail below, in the "Representation Error"
section. See `The Perils of Floating Point <http://www.lahey.com/float.htm>`_
for a more complete account of other common surprises.
As that says near the end, "there are no easy answers." Still, don't be unduly
wary of floating-point! The errors in Python float operations are inherited
from the floating-point hardware, and on most machines are on the order of no
more than 1 part in 2\*\*53 per operation. That's more than adequate for most
tasks, but you do need to keep in mind that it's not decimal arithmetic, and
that every float operation can suffer a new rounding error.
While pathological cases do exist, for most casual use of floating-point
arithmetic you'll see the result you expect in the end if you simply round the
display of your final results to the number of decimal digits you expect.
:func:`str` usually suffices, and for finer control see the :meth:`str.format`
method's format specifiers in :ref:`formatstrings`.
2007-08-15 11:28:22 -03:00
For use cases which require exact decimal representation, try using the
:mod:`decimal` module which implements decimal arithmetic suitable for
accounting applications and high-precision applications.
Another form of exact arithmetic is supported by the :mod:`fractions` module
which implements arithmetic based on rational numbers (so the numbers like
1/3 can be represented exactly).
If you are a heavy user of floating point operations you should take a look
at the Numerical Python package and many other packages for mathematical and
statistical operations supplied by the SciPy project. See <http://scipy.org>.
Python provides tools that may help on those rare occasions when you really
*do* want to know the exact value of a float. The
:meth:`float.as_integer_ratio` method expresses the value of a float as a
fraction::
>>> x = 3.14159
>>> x.as_integer_ratio()
(3537115888337719L, 1125899906842624L)
Since the ratio is exact, it can be used to losslessly recreate the
original value::
>>> x == 3537115888337719 / 1125899906842624
True
The :meth:`float.hex` method expresses a float in hexadecimal (base
16), again giving the exact value stored by your computer::
>>> x.hex()
'0x1.921f9f01b866ep+1'
This precise hexadecimal representation can be used to reconstruct
the float value exactly::
>>> x == float.fromhex('0x1.921f9f01b866ep+1')
True
Since the representation is exact, it is useful for reliably porting values
across different versions of Python (platform independence) and exchanging
data with other languages that support the same format (such as Java and C99).
2009-04-26 19:01:46 -03:00
Another helpful tool is the :func:`math.fsum` function which helps mitigate
loss-of-precision during summation. It tracks "lost digits" as values are
added onto a running total. That can make a difference in overall accuracy
so that the errors do not accumulate to the point where they affect the
final total:
>>> sum([0.1] * 10) == 1.0
False
>>> math.fsum([0.1] * 10) == 1.0
True
2007-08-15 11:28:22 -03:00
.. _tut-fp-error:
Representation Error
====================
This section explains the "0.1" example in detail, and shows how you can perform
an exact analysis of cases like this yourself. Basic familiarity with binary
floating-point representation is assumed.
:dfn:`Representation error` refers to the fact that some (most, actually)
decimal fractions cannot be represented exactly as binary (base 2) fractions.
This is the chief reason why Python (or Perl, C, C++, Java, Fortran, and many
2009-04-24 00:09:06 -03:00
others) often won't display the exact decimal number you expect.
2007-08-15 11:28:22 -03:00
Why is that? 1/10 is not exactly representable as a binary fraction. Almost all
machines today (November 2000) use IEEE-754 floating point arithmetic, and
almost all platforms map Python floats to IEEE-754 "double precision". 754
doubles contain 53 bits of precision, so on input the computer strives to
Merged revisions 69129-69131,69139-69140,69143,69154-69159,69169,69288-69289,69293,69297-69301,69348 via svnmerge from svn+ssh://pythondev@svn.python.org/python/trunk ........ r69129 | benjamin.peterson | 2009-01-30 19:42:55 -0600 (Fri, 30 Jan 2009) | 1 line check the errno in bad fd cases ........ r69130 | andrew.kuchling | 2009-01-30 20:50:09 -0600 (Fri, 30 Jan 2009) | 1 line Add a section ........ r69131 | andrew.kuchling | 2009-01-30 21:26:02 -0600 (Fri, 30 Jan 2009) | 1 line Text edits and markup fixes ........ r69139 | mark.dickinson | 2009-01-31 10:44:04 -0600 (Sat, 31 Jan 2009) | 2 lines Add an extra test for long <-> float hash equivalence. ........ r69140 | benjamin.peterson | 2009-01-31 10:52:03 -0600 (Sat, 31 Jan 2009) | 1 line PyErr_BadInternalCall() raises a SystemError, not TypeError #5112 ........ r69143 | benjamin.peterson | 2009-01-31 15:00:10 -0600 (Sat, 31 Jan 2009) | 1 line I believe the intention here was to avoid a global lookup ........ r69154 | benjamin.peterson | 2009-01-31 16:33:02 -0600 (Sat, 31 Jan 2009) | 1 line fix indentation in comment ........ r69155 | david.goodger | 2009-01-31 16:53:46 -0600 (Sat, 31 Jan 2009) | 1 line markup fix ........ r69156 | gregory.p.smith | 2009-01-31 16:57:30 -0600 (Sat, 31 Jan 2009) | 4 lines - Issue #5104: The socket module now raises OverflowError when 16-bit port and protocol numbers are supplied outside the allowed 0-65536 range on bind() and getservbyport(). ........ r69157 | benjamin.peterson | 2009-01-31 17:43:25 -0600 (Sat, 31 Jan 2009) | 1 line add explanatory comment ........ r69158 | benjamin.peterson | 2009-01-31 17:54:38 -0600 (Sat, 31 Jan 2009) | 1 line more flags which only work for function blocks ........ r69159 | gregory.p.smith | 2009-01-31 18:16:01 -0600 (Sat, 31 Jan 2009) | 2 lines Update doc wording as suggested in issue4903. ........ r69169 | guilherme.polo | 2009-01-31 20:56:16 -0600 (Sat, 31 Jan 2009) | 3 lines Restore Tkinter.Tk._loadtk so this test doesn't fail for problems related to ttk. ........ r69288 | georg.brandl | 2009-02-05 04:30:57 -0600 (Thu, 05 Feb 2009) | 1 line #5153: fix typo in example. ........ r69289 | georg.brandl | 2009-02-05 04:37:07 -0600 (Thu, 05 Feb 2009) | 1 line #5144: document that PySys_SetArgv prepends the script directory (or the empty string) to sys.path. ........ r69293 | georg.brandl | 2009-02-05 04:59:28 -0600 (Thu, 05 Feb 2009) | 1 line #5059: fix example. ........ r69297 | georg.brandl | 2009-02-05 05:32:18 -0600 (Thu, 05 Feb 2009) | 1 line #5015: document PythonHome API functions. ........ r69298 | georg.brandl | 2009-02-05 05:33:21 -0600 (Thu, 05 Feb 2009) | 1 line #4827: fix callback example. ........ r69299 | georg.brandl | 2009-02-05 05:35:28 -0600 (Thu, 05 Feb 2009) | 1 line #4820: use correct module for ctypes.util. ........ r69300 | georg.brandl | 2009-02-05 05:38:23 -0600 (Thu, 05 Feb 2009) | 1 line #4563: disable alpha and roman lists, fixes wrong formatting of contributor list. ........ r69301 | georg.brandl | 2009-02-05 05:40:35 -0600 (Thu, 05 Feb 2009) | 1 line #5031: fix Thread.daemon property docs. ........ r69348 | benjamin.peterson | 2009-02-05 19:47:31 -0600 (Thu, 05 Feb 2009) | 1 line fix download link ........
2009-02-05 22:40:07 -04:00
convert 0.1 to the closest fraction it can of the form *J*/2**\ *N* where *J* is
2007-08-15 11:28:22 -03:00
an integer containing exactly 53 bits. Rewriting ::
1 / 10 ~= J / (2**N)
as ::
J ~= 2**N / 10
and recalling that *J* has exactly 53 bits (is ``>= 2**52`` but ``< 2**53``),
the best value for *N* is 56::
>>> 2**52
2008-08-10 09:16:45 -03:00
4503599627370496
2007-08-15 11:28:22 -03:00
>>> 2**53
2008-08-10 09:16:45 -03:00
9007199254740992
2007-08-15 11:28:22 -03:00
>>> 2**56/10
2008-08-10 09:16:45 -03:00
7205759403792794.0
2007-08-15 11:28:22 -03:00
That is, 56 is the only value for *N* that leaves *J* with exactly 53 bits. The
best possible value for *J* is then that quotient rounded::
>>> q, r = divmod(2**56, 10)
>>> r
2008-08-10 09:16:45 -03:00
6
2007-08-15 11:28:22 -03:00
Since the remainder is more than half of 10, the best approximation is obtained
by rounding up::
>>> q+1
2008-08-10 09:16:45 -03:00
7205759403792794
2007-08-15 11:28:22 -03:00
Therefore the best possible approximation to 1/10 in 754 double precision is
that over 2\*\*56, or ::
7205759403792794 / 72057594037927936
2009-04-24 00:09:06 -03:00
Dividing both the numerator and denominator by two reduces the fraction to::
3602879701896397 / 36028797018963968
2007-08-15 11:28:22 -03:00
Note that since we rounded up, this is actually a little bit larger than 1/10;
if we had not rounded up, the quotient would have been a little bit smaller than
1/10. But in no case can it be *exactly* 1/10!
So the computer never "sees" 1/10: what it sees is the exact fraction given
above, the best 754 double approximation it can get::
2009-04-24 00:09:06 -03:00
>>> 0.1 * 2 ** 55
3602879701896397.0
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
If we multiply that fraction by 10\*\*60, we can see the value of out to
60 decimal digits::
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
>>> 3602879701896397 * 10 ** 60 // 2 ** 55
1000000000000000055511151231257827021181583404541015625
2007-08-15 11:28:22 -03:00
meaning that the exact number stored in the computer is approximately equal to
the decimal value 0.100000000000000005551115123125. Rounding that to 17
significant digits gives the 0.10000000000000001 that Python displays (well,
will display on any 754-conforming platform that does best-possible input and
output conversions in its C library --- yours may not!).
2009-04-24 00:09:06 -03:00
The :mod:`fractions` and :mod:`decimal` modules make these calculations
easy::
2007-08-15 11:28:22 -03:00
2009-04-24 00:09:06 -03:00
>>> from decimal import Decimal
>>> from fractions import Fraction
>>> print(Fraction.from_float(0.1))
3602879701896397/36028797018963968
>>> print(Decimal.from_float(0.1))
0.1000000000000000055511151231257827021181583404541015625