# Consistent handling of division by zero in numpy array

I want to populate a numpy array with values from the smooth bump function

f(x) = exp ( - 1 / (1 - x^2) )     if |x| < 1,  f(x) =  0 otherwise


Currently I have something that works (as in gives me the right numbers on my platform)

x = linspace(-1.1 , 1.1, 300)   #Sample 300 points between [-1.1,1.1]
bump = exp( 1 - 1 / (1 - clip(square(x), 0,1)) )


when the absolute value of an entry in x is at least 1, its square gets clipped to 1, and we have "1/(1-1) = 1/0 = +inf" as "expected" on my platform, which then gets set by "exp(1 - inf) = 0" which is exactly the behaviour I want.

My questions:

1. I suspect that the above is not the best practice. Am I correct in my suspicions?
2. Are there better ways of handling this division by zero? At the end of the day the array x may not be just simply a linear list of values. So I want something that can compute f(x) from x efficiently.

Why do you want to generate DivisionByZero exceptions?

import numpy as np
x= np.linspace(-1.1,1.1,300)

def f(x):
return np.exp(-1.0/(1.0-x**2))



If you want, you can do the masking in your function (by having boundary arguments)

• shakes fist at GertVdE for beating me to the answer. Gert has it, use array masks, don't rely on dangerous boundaries of the floating-point standard. For example, your code didn't work on my OS X.8 Intel Mac. – Aron Ahmadia Feb 6 '13 at 10:46
• "Why do you want to generate DivisionByZero exceptions?" Because I didn't know any better. Thanks for the answer. Let me try it out. – Willie Wong Feb 6 '13 at 11:14
• Okay, so after I do the arithmetic on the non-zero values, I can just just use the filled() function to set the masked values to 0. That works great. Thanks again. – Willie Wong Feb 6 '13 at 11:37

Another take:

_f = lambda x: np.exp(-1.0/(1.0-x*x))
f = lambda x: np.piecewise(x, [np.abs(x) < 1, np.abs(x) >= 1], [_f, 0.0])

x = np.linspace(-1.1,1.1,300)
bump = f(x)


my point here is that what you really need for efficiency is a ufunc, possibly implemented in C. In the absence of a true ufunc you can go with any numpy trick.