How to avoid floating point errors?

temporary_user_name picture temporary_user_name · Oct 20, 2013 · Viewed 31.7k times · Source

I was trying to write a function to approximate square roots (I know there's the math module...I want to do it myself), and I was getting screwed over by the floating point arithmetic. How can you avoid that?

def sqrt(num):
    root = 0.0
    while root * root < num:
        root += 0.01
    return root

Using this has these results:

>>> sqrt(4)
2.0000000000000013
>>> sqrt(9)
3.00999999999998

I realize I could just use round(), but I want to be able to make this really accurate. I want to be able to calculate out to 6 or 7 digits. That wouldn't be possible if I'm rounding. I want to understand how to properly handle floating point calculations in Python.

Answer

Tim Peters picture Tim Peters · Oct 20, 2013

This really has nothing to do with Python - you'd see the same behavior in any language using your hardware's binary floating-point arithmetic. First read the docs.

After you read that, you'll better understand that you're not adding one one-hundredth in your code. This is exactly what you're adding:

>>> from decimal import Decimal
>>> Decimal(.01)
Decimal('0.01000000000000000020816681711721685132943093776702880859375')

That string shows the exact decimal value of the binary floating ("double precision" in C) approximation to the exact decimal value 0.01. The thing you're really adding is a little bigger than 1/100.

Controlling floating-point numeric errors is the field called "numerical analysis", and is a very large and complex topic. So long as you're startled by the fact that floats are just approximations to decimal values, use the decimal module. That will take away a world of "shallow" problems for you. For example, given this small modification to your function:

from decimal import Decimal as D

def sqrt(num):
    root = D(0)
    while root * root < num:
        root += D("0.01")
    return root

then:

>>> sqrt(4)
Decimal('2.00')
>>> sqrt(9)
Decimal('3.00')

It's not really more accurate, but may be less surprising in simple examples because now it's adding exactly one one-hundredth.

An alternative is to stick to floats and add something that is exactly representable as a binary float: values of the form I/2**J. For example, instead of adding 0.01, add 0.125 (1/8) or 0.0625 (1/16).

Then look up "Newton's method" for computing square roots ;-)