| 1 |
.. testsetup:: |
| 2 |
|
| 3 |
import math |
| 4 |
from fractions import Fraction |
| 5 |
|
| 6 |
.. _tut-fp-issues: |
| 7 |
|
| 8 |
************************************************** |
| 9 |
Floating-Point Arithmetic: Issues and Limitations |
| 10 |
************************************************** |
| 11 |
|
| 12 |
.. sectionauthor:: Tim Peters <[email protected]> |
| 13 |
.. sectionauthor:: Raymond Hettinger <python at rcn dot com> |
| 14 |
|
| 15 |
|
| 16 |
Floating-point numbers are represented in computer hardware as base 2 (binary) |
| 17 |
fractions. For example, the **decimal** fraction ``0.625`` |
| 18 |
has value 6/10 + 2/100 + 5/1000, and in the same way the **binary** fraction ``0.101`` |
| 19 |
has value 1/2 + 0/4 + 1/8. These two fractions have identical values, the only |
| 20 |
real difference being that the first is written in base 10 fractional notation, |
| 21 |
and the second in base 2. |
| 22 |
|
| 23 |
Unfortunately, most decimal fractions cannot be represented exactly as binary |
| 24 |
fractions. A consequence is that, in general, the decimal floating-point |
| 25 |
numbers you enter are only approximated by the binary floating-point numbers |
| 26 |
actually stored in the machine. |
| 27 |
|
| 28 |
The problem is easier to understand at first in base 10. Consider the fraction |
| 29 |
1/3. You can approximate that as a base 10 fraction:: |
| 30 |
|
| 31 |
0.3 |
| 32 |
|
| 33 |
or, better, :: |
| 34 |
|
| 35 |
0.33 |
| 36 |
|
| 37 |
or, better, :: |
| 38 |
|
| 39 |
0.333 |
| 40 |
|
| 41 |
and so on. No matter how many digits you're willing to write down, the result |
| 42 |
will never be exactly 1/3, but will be an increasingly better approximation of |
| 43 |
1/3. |
| 44 |
|
| 45 |
In the same way, no matter how many base 2 digits you're willing to use, the |
| 46 |
decimal value 0.1 cannot be represented exactly as a base 2 fraction. In base |
| 47 |
2, 1/10 is the infinitely repeating fraction :: |
| 48 |
|
| 49 |
0.0001100110011001100110011001100110011001100110011... |
| 50 |
|
| 51 |
Stop at any finite number of bits, and you get an approximation. On most |
| 52 |
machines today, floats are approximated using a binary fraction with |
| 53 |
the numerator using the first 53 bits starting with the most significant bit and |
| 54 |
with the denominator as a power of two. In the case of 1/10, the binary fraction |
| 55 |
is ``3602879701896397 / 2 ** 55`` which is close to but not exactly |
| 56 |
equal to the true value of 1/10. |
| 57 |
|
| 58 |
Many users are not aware of the approximation because of the way values are |
| 59 |
displayed. Python only prints a decimal approximation to the true decimal |
| 60 |
value of the binary approximation stored by the machine. On most machines, if |
| 61 |
Python were to print the true decimal value of the binary approximation stored |
| 62 |
for 0.1, it would have to display:: |
| 63 |
|
| 64 |
>>> 0.1 |
| 65 |
0.1000000000000000055511151231257827021181583404541015625 |
| 66 |
|
| 67 |
That is more digits than most people find useful, so Python keeps the number |
| 68 |
of digits manageable by displaying a rounded value instead: |
| 69 |
|
| 70 |
.. doctest:: |
| 71 |
|
| 72 |
>>> 1 / 10 |
| 73 |
0.1 |
| 74 |
|
| 75 |
Just remember, even though the printed result looks like the exact value |
| 76 |
of 1/10, the actual stored value is the nearest representable binary fraction. |
| 77 |
|
| 78 |
Interestingly, there are many different decimal numbers that share the same |
| 79 |
nearest approximate binary fraction. For example, the numbers ``0.1`` and |
| 80 |
``0.10000000000000001`` and |
| 81 |
``0.1000000000000000055511151231257827021181583404541015625`` are all |
| 82 |
approximated by ``3602879701896397 / 2 ** 55``. Since all of these decimal |
| 83 |
values share the same approximation, any one of them could be displayed |
| 84 |
while still preserving the invariant ``eval(repr(x)) == x``. |
| 85 |
|
| 86 |
Historically, the Python prompt and built-in :func:`repr` function would choose |
| 87 |
the one with 17 significant digits, ``0.10000000000000001``. Starting with |
| 88 |
Python 3.1, Python (on most systems) is now able to choose the shortest of |
| 89 |
these and simply display ``0.1``. |
| 90 |
|
| 91 |
Note that this is in the very nature of binary floating point: this is not a bug |
| 92 |
in Python, and it is not a bug in your code either. You'll see the same kind of |
| 93 |
thing in all languages that support your hardware's floating-point arithmetic |
| 94 |
(although some languages may not *display* the difference by default, or in all |
| 95 |
output modes). |
| 96 |
|
| 97 |
For more pleasant output, you may wish to use string formatting to produce a |
| 98 |
limited number of significant digits: |
| 99 |
|
| 100 |
.. doctest:: |
| 101 |
|
| 102 |
>>> format(math.pi, '.12g') # give 12 significant digits |
| 103 |
'3.14159265359' |
| 104 |
|
| 105 |
>>> format(math.pi, '.2f') # give 2 digits after the point |
| 106 |
'3.14' |
| 107 |
|
| 108 |
>>> repr(math.pi) |
| 109 |
'3.141592653589793' |
| 110 |
|
| 111 |
It's important to realize that this is, in a real sense, an illusion: you're |
| 112 |
simply rounding the *display* of the true machine value. |
| 113 |
|
| 114 |
One illusion may beget another. For example, since 0.1 is not exactly 1/10, |
| 115 |
summing three values of 0.1 may not yield exactly 0.3, either: |
| 116 |
|
| 117 |
.. doctest:: |
| 118 |
|
| 119 |
>>> 0.1 + 0.1 + 0.1 == 0.3 |
| 120 |
False |
| 121 |
|
| 122 |
Also, since the 0.1 cannot get any closer to the exact value of 1/10 and |
| 123 |
0.3 cannot get any closer to the exact value of 3/10, then pre-rounding with |
| 124 |
:func:`round` function cannot help: |
| 125 |
|
| 126 |
.. doctest:: |
| 127 |
|
| 128 |
>>> round(0.1, 1) + round(0.1, 1) + round(0.1, 1) == round(0.3, 1) |
| 129 |
False |
| 130 |
|
| 131 |
Though the numbers cannot be made closer to their intended exact values, |
| 132 |
the :func:`math.isclose` function can be useful for comparing inexact values: |
| 133 |
|
| 134 |
.. doctest:: |
| 135 |
|
| 136 |
>>> math.isclose(0.1 + 0.1 + 0.1, 0.3) |
| 137 |
True |
| 138 |
|
| 139 |
Alternatively, the :func:`round` function can be used to compare rough |
| 140 |
approximations: |
| 141 |
|
| 142 |
.. doctest:: |
| 143 |
|
| 144 |
>>> round(math.pi, ndigits=2) == round(22 / 7, ndigits=2) |
| 145 |
True |
| 146 |
|
| 147 |
Binary floating-point arithmetic holds many surprises like this. The problem |
| 148 |
with "0.1" is explained in precise detail below, in the "Representation Error" |
| 149 |
section. See `Examples of Floating Point Problems |
| 150 |
<https://jvns.ca/blog/2023/01/13/examples-of-floating-point-problems/>`_ for |
| 151 |
a pleasant summary of how binary floating point works and the kinds of |
| 152 |
problems commonly encountered in practice. Also see |
| 153 |
`The Perils of Floating Point <http://www.indowsway.com/floatingpoint.htm>`_ |
| 154 |
for a more complete account of other common surprises. |
| 155 |
|
| 156 |
As that says near the end, "there are no easy answers." Still, don't be unduly |
| 157 |
wary of floating point! The errors in Python float operations are inherited |
| 158 |
from the floating-point hardware, and on most machines are on the order of no |
| 159 |
more than 1 part in 2\*\*53 per operation. That's more than adequate for most |
| 160 |
tasks, but you do need to keep in mind that it's not decimal arithmetic and |
| 161 |
that every float operation can suffer a new rounding error. |
| 162 |
|
| 163 |
While pathological cases do exist, for most casual use of floating-point |
| 164 |
arithmetic you'll see the result you expect in the end if you simply round the |
| 165 |
display of your final results to the number of decimal digits you expect. |
| 166 |
:func:`str` usually suffices, and for finer control see the :meth:`str.format` |
| 167 |
method's format specifiers in :ref:`formatstrings`. |
| 168 |
|
| 169 |
For use cases which require exact decimal representation, try using the |
| 170 |
:mod:`decimal` module which implements decimal arithmetic suitable for |
| 171 |
accounting applications and high-precision applications. |
| 172 |
|
| 173 |
Another form of exact arithmetic is supported by the :mod:`fractions` module |
| 174 |
which implements arithmetic based on rational numbers (so the numbers like |
| 175 |
1/3 can be represented exactly). |
| 176 |
|
| 177 |
If you are a heavy user of floating-point operations you should take a look |
| 178 |
at the NumPy package and many other packages for mathematical and |
| 179 |
statistical operations supplied by the SciPy project. See <https://scipy.org>. |
| 180 |
|
| 181 |
Python provides tools that may help on those rare occasions when you really |
| 182 |
*do* want to know the exact value of a float. The |
| 183 |
:meth:`float.as_integer_ratio` method expresses the value of a float as a |
| 184 |
fraction: |
| 185 |
|
| 186 |
.. doctest:: |
| 187 |
|
| 188 |
>>> x = 3.14159 |
| 189 |
>>> x.as_integer_ratio() |
| 190 |
(3537115888337719, 1125899906842624) |
| 191 |
|
| 192 |
Since the ratio is exact, it can be used to losslessly recreate the |
| 193 |
original value: |
| 194 |
|
| 195 |
.. doctest:: |
| 196 |
|
| 197 |
>>> x == 3537115888337719 / 1125899906842624 |
| 198 |
True |
| 199 |
|
| 200 |
The :meth:`float.hex` method expresses a float in hexadecimal (base |
| 201 |
16), again giving the exact value stored by your computer: |
| 202 |
|
| 203 |
.. doctest:: |
| 204 |
|
| 205 |
>>> x.hex() |
| 206 |
'0x1.921f9f01b866ep+1' |
| 207 |
|
| 208 |
This precise hexadecimal representation can be used to reconstruct |
| 209 |
the float value exactly: |
| 210 |
|
| 211 |
.. doctest:: |
| 212 |
|
| 213 |
>>> x == float.fromhex('0x1.921f9f01b866ep+1') |
| 214 |
True |
| 215 |
|
| 216 |
Since the representation is exact, it is useful for reliably porting values |
| 217 |
across different versions of Python (platform independence) and exchanging |
| 218 |
data with other languages that support the same format (such as Java and C99). |
| 219 |
|
| 220 |
Another helpful tool is the :func:`sum` function which helps mitigate |
| 221 |
loss-of-precision during summation. It uses extended precision for |
| 222 |
intermediate rounding steps as values are added onto a running total. |
| 223 |
That can make a difference in overall accuracy so that the errors do not |
| 224 |
accumulate to the point where they affect the final total: |
| 225 |
|
| 226 |
.. doctest:: |
| 227 |
|
| 228 |
>>> 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 + 0.1 == 1.0 |
| 229 |
False |
| 230 |
>>> sum([0.1] * 10) == 1.0 |
| 231 |
True |
| 232 |
|
| 233 |
The :func:`math.fsum` goes further and tracks all of the "lost digits" |
| 234 |
as values are added onto a running total so that the result has only a |
| 235 |
single rounding. This is slower than :func:`sum` but will be more |
| 236 |
accurate in uncommon cases where large magnitude inputs mostly cancel |
| 237 |
each other out leaving a final sum near zero: |
| 238 |
|
| 239 |
.. doctest:: |
| 240 |
|
| 241 |
>>> arr = [-0.10430216751806065, -266310978.67179024, 143401161448607.16, |
| 242 |
... -143401161400469.7, 266262841.31058735, -0.003244936839808227] |
| 243 |
>>> float(sum(map(Fraction, arr))) # Exact summation with single rounding |
| 244 |
8.042173697819788e-13 |
| 245 |
>>> math.fsum(arr) # Single rounding |
| 246 |
8.042173697819788e-13 |
| 247 |
>>> sum(arr) # Multiple roundings in extended precision |
| 248 |
8.042178034628478e-13 |
| 249 |
>>> total = 0.0 |
| 250 |
>>> for x in arr: |
| 251 |
... total += x # Multiple roundings in standard precision |
| 252 |
... |
| 253 |
>>> total # Straight addition has no correct digits! |
| 254 |
-0.0051575902860057365 |
| 255 |
|
| 256 |
|
| 257 |
.. _tut-fp-error: |
| 258 |
|
| 259 |
Representation Error |
| 260 |
==================== |
| 261 |
|
| 262 |
This section explains the "0.1" example in detail, and shows how you can perform |
| 263 |
an exact analysis of cases like this yourself. Basic familiarity with binary |
| 264 |
floating-point representation is assumed. |
| 265 |
|
| 266 |
:dfn:`Representation error` refers to the fact that some (most, actually) |
| 267 |
decimal fractions cannot be represented exactly as binary (base 2) fractions. |
| 268 |
This is the chief reason why Python (or Perl, C, C++, Java, Fortran, and many |
| 269 |
others) often won't display the exact decimal number you expect. |
| 270 |
|
| 271 |
Why is that? 1/10 is not exactly representable as a binary fraction. Since at |
| 272 |
least 2000, almost all machines use IEEE 754 binary floating-point arithmetic, |
| 273 |
and almost all platforms map Python floats to IEEE 754 binary64 "double |
| 274 |
precision" values. IEEE 754 binary64 values contain 53 bits of precision, so |
| 275 |
on input the computer strives to convert 0.1 to the closest fraction it can of |
| 276 |
the form *J*/2**\ *N* where *J* is an integer containing exactly 53 bits. |
| 277 |
Rewriting |
| 278 |
:: |
| 279 |
|
| 280 |
1 / 10 ~= J / (2**N) |
| 281 |
|
| 282 |
as :: |
| 283 |
|
| 284 |
J ~= 2**N / 10 |
| 285 |
|
| 286 |
and recalling that *J* has exactly 53 bits (is ``>= 2**52`` but ``< 2**53``), |
| 287 |
the best value for *N* is 56: |
| 288 |
|
| 289 |
.. doctest:: |
| 290 |
|
| 291 |
>>> 2**52 <= 2**56 // 10 < 2**53 |
| 292 |
True |
| 293 |
|
| 294 |
That is, 56 is the only value for *N* that leaves *J* with exactly 53 bits. The |
| 295 |
best possible value for *J* is then that quotient rounded: |
| 296 |
|
| 297 |
.. doctest:: |
| 298 |
|
| 299 |
>>> q, r = divmod(2**56, 10) |
| 300 |
>>> r |
| 301 |
6 |
| 302 |
|
| 303 |
Since the remainder is more than half of 10, the best approximation is obtained |
| 304 |
by rounding up: |
| 305 |
|
| 306 |
.. doctest:: |
| 307 |
|
| 308 |
|
| 309 |
|
| 310 |
>>> q+1 |
| 311 |
7205759403792794 |
| 312 |
|
| 313 |
Therefore the best possible approximation to 1/10 in IEEE 754 double precision |
| 314 |
is:: |
| 315 |
|
| 316 |
7205759403792794 / 2 ** 56 |
| 317 |
|
| 318 |
Dividing both the numerator and denominator by two reduces the fraction to:: |
| 319 |
|
| 320 |
3602879701896397 / 2 ** 55 |
| 321 |
|
| 322 |
Note that since we rounded up, this is actually a little bit larger than 1/10; |
| 323 |
if we had not rounded up, the quotient would have been a little bit smaller than |
| 324 |
1/10. But in no case can it be *exactly* 1/10! |
| 325 |
|
| 326 |
So the computer never "sees" 1/10: what it sees is the exact fraction given |
| 327 |
above, the best IEEE 754 double approximation it can get: |
| 328 |
|
| 329 |
.. doctest:: |
| 330 |
|
| 331 |
>>> 0.1 * 2 ** 55 |
| 332 |
3602879701896397.0 |
| 333 |
|
| 334 |
If we multiply that fraction by 10\*\*55, we can see the value out to |
| 335 |
55 decimal digits: |
| 336 |
|
| 337 |
.. doctest:: |
| 338 |
|
| 339 |
>>> 3602879701896397 * 10 ** 55 // 2 ** 55 |
| 340 |
1000000000000000055511151231257827021181583404541015625 |
| 341 |
|
| 342 |
meaning that the exact number stored in the computer is equal to |
| 343 |
the decimal value 0.1000000000000000055511151231257827021181583404541015625. |
| 344 |
Instead of displaying the full decimal value, many languages (including |
| 345 |
older versions of Python), round the result to 17 significant digits: |
| 346 |
|
| 347 |
.. doctest:: |
| 348 |
|
| 349 |
>>> format(0.1, '.17f') |
| 350 |
'0.10000000000000001' |
| 351 |
|
| 352 |
The :mod:`fractions` and :mod:`decimal` modules make these calculations |
| 353 |
easy: |
| 354 |
|
| 355 |
.. doctest:: |
| 356 |
|
| 357 |
>>> from decimal import Decimal |
| 358 |
>>> from fractions import Fraction |
| 359 |
|
| 360 |
>>> Fraction.from_float(0.1) |
| 361 |
Fraction(3602879701896397, 36028797018963968) |
| 362 |
|
| 363 |
>>> (0.1).as_integer_ratio() |
| 364 |
(3602879701896397, 36028797018963968) |
| 365 |
|
| 366 |
>>> Decimal.from_float(0.1) |
| 367 |
Decimal('0.1000000000000000055511151231257827021181583404541015625') |
| 368 |
|
| 369 |
>>> format(Decimal.from_float(0.1), '.17') |
| 370 |
'0.10000000000000001' |