Using exponentiation **0.5 less efficient than math.sqrt?

Question

A quote from "Python Programming: An Introduction to Computer Science"

We could have taken the square root using exponentiation **. Using math.sqrt is somewhat more efficient.

"Somewhat", but to what extent, and how?

Answer 1

Theoretically, hammar's answer and duffymo's answer are good guesses. But in practice, on my machine, it's not more efficient:

>>> import timeit
>>> timeit.timeit(stmt='[n ** 0.5 for n in range(100)]', setup='import math', number=10000)
0.15518403053283691
>>> timeit.timeit(stmt='[math.sqrt(n) for n in range(100)]', setup='import math', number=10000)
0.17707490921020508

Part of the problem is the . operation. If you import sqrt directly into the namespace, you get a slight improvement.

>>> timeit.timeit(stmt='[sqrt(n) for n in range(100)]', setup='from math import sqrt', number=10000)
0.15312695503234863

Key word there: slight.

Further testing indicates that as the number gets larger, the benefit you get from using sqrt increases. But still not by a lot!

>>> timeit.timeit(stmt='[n ** 0.5 for n in range(1000000)]', setup='import math', number=1)
0.18888211250305176
>>> timeit.timeit(stmt='[math.sqrt(n) for n in range(1000000)]', setup='import math', number=1)
0.18425297737121582
>>> timeit.timeit(stmt='[sqrt(n) for n in range(1000000)]', setup='from math import sqrt', number=1)
0.1571958065032959

Answer 2

No need to guess the implementation, we can read the code!

math.sqrt is a thin wrapper about sqrt from the standard C library: see mathmodule.c, line 956

The ** operator has multiple implementations depending on the types of the arguments, but in the case of a floating-point exponent, it eventually dispatches to pow from the standard C library (see floatobject.c line 783).

Modern CPUs often have special square root instructions which general exponentiation routines don't use (compare and contrast the implementations of pow and sqrt in glibc for x86-64, for example). But once all the interpreter overhead is added (byte codes, type checking, method dispatch etc), the difference in raw speed doesn't matter all that much, and can be dominated by issues like whether you call sqrt directly or look it up via the math module (as shown by the timings in other answers).

Answer 3

** has to support any exponent while math.sqrt knows it's always 0.5. math.sqrt can therefore use a more specialized (and therefore probably more efficient) algorithm.