OverflowError: long int too large to convert to float in python

Question

I tried to calculate poisson distribution in python as below:

p = math.pow(3,idx)
depart = math.exp(-3) * p 
depart = depart / math.factorial(idx)

idx ranges from 0

But I got OverflowError: long int too large to convert to float

I tried to convert depart to float but no results.

Solution 1

Factorials get large real fast:

>>> math.factorial(170)
7257415615307998967396728211129263114716991681296451376543577798900561843401706157852350749242617459511490991237838520776666022565442753025328900773207510902400430280058295603966612599658257104398558294257568966313439612262571094946806711205568880457193340212661452800000000000000000000000000000000000000000L

Note the L; the factorial of 170 is still convertable to a float:

>>> float(math.factorial(170))
7.257415615307999e+306

but the next factorial is too large:

>>> float(math.factorial(171))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
OverflowError: long int too large to convert to float

You could use the decimal module; calculations will be slower, but the Decimal() class can handle factorials this size:

>>> from decimal import Decimal
>>> Decimal(math.factorial(171))
Decimal('1241018070217667823424840524103103992616605577501693185388951803611996075221691752992751978120487585576464959501670387052809889858690710767331242032218484364310473577889968548278290754541561964852153468318044293239598173696899657235903947616152278558180061176365108428800000000000000000000000000000000000000000')

You'll have to use Decimal() values throughout:

from decimal import *

with localcontext() as ctx:
    ctx.prec = 32  # desired precision
    p = ctx.power(3, idx)
    depart = ctx.exp(-3) * p 
    depart /= math.factorial(idx)

Solution 2

When idx gets large either the math.pow and/or the math.factorial will become insanely large and be unable to convert to a floating value (idx=1000 triggers the error on my 64 bit machine). You'll want to not use the math.pow function as it overflows earlier than the built in ** operator because it tries to keep higher precision by float converting earlier. Additionally, you can wrap each function call in a Decimal object for higher precision.

Another approach when dealing with very large numbers is to work in the log scale. Take the log of every value (or calculate the log version of each value) and perform all required operations before taking the exponentiation of the results. This allows for your values to temporary leave the floating domain space while still accurately computing a final answer that lies within floating domain.

3 ** idx  =>  math.log(3) * idx
math.exp(-3) * p  =>  -3 + math.log(p)
math.factorial(idx)  =>  sum(math.log(ii) for ii in range(1, idx + 1))
...
math.exp(result)

This stays in the log domain until the very end so your numbers can get very, very large before you'll hit overflow problems.

Solution 3

Try using the decimal library. It claims to support arbitrary precision.
from decimal import Decimal

Also, you don't need to use math.pow. pow is in-built.

Solution 4

The scipy module could help you.

scipy.misc.factorial is a factorial function that can use the gamma function approximation to calculate the factorial, and returns the result using floating points.

import numpy
from scipy.misc import factorial

i = numpy.arange(10)
print(numpy.exp(-3) * 3**i / factorial(i))

Gives:

[ 0.04978707  0.14936121  0.22404181  0.22404181  0.16803136  0.10081881
  0.05040941  0.02160403  0.00810151  0.0027005 ]

There is also a module to calculate Poisson distributions. For example:

import numpy
from scipy.stats import poisson

i = numpy.arange(10)
p = poisson(3)
print(p.pmf(i))

Gives:

[ 0.04978707  0.14936121  0.22404181  0.22404181  0.16803136  0.10081881
  0.05040941  0.02160403  0.00810151  0.0027005 ]

View more solutions

Comments

[Hunter McMillen]

What is the large number class in Python called? because 170! is well above the limits of a 64-bit integer. I assume the type is coerced to that? If so, why is the L still appended to the end?

[Martijn Pieters]

@HunterMcMillen: Python long integers are only limited by OS memory allocation. The L signifies we moved beyond the maximum C integer size for the platform (64-bit in my case) to show long integers are being used; the transition is automatic. See docs.python.org/2/library/…

[Hunter McMillen]

Ahh, I see. Most other languages have BigInteger or something similar. Thanks.

[DSM]

+1 for scipy.stats, but the factorial function scipy offers will still inf out at large argument.

[user2312186]

Thanks. I am going to generate Poisson numbers as generating packet time in the network. How can i use this function for generating a Poisson number each continually?

[user2312186]

Hi guys I tried to use decimal or scipy but i got ImportError: No module named scipy.stats the same error for decimal

[xylon97]

what version are you running? decimal library is supposed to be standard for all. perhaps you have deleted it. re-install python and it will fix it. as for scipy, it is a third-party package, so you'll have to download it from scipy.org

[user2312186]

I am going to use poisson like this: depart = depart + Poisson(3) so that inter-arrival times are selected by Poisson. but it does not work.

[Shubham Saini]

I think 3 ** idx => math.log(3) * math.log(idx) should be 3 ** idx => math.log(3) * idx