Numpy argsort strange indices

Question

The first output is correct. But somehow argsort gives wrong output for the second array:

>>> np.argsort(np.array([ 0.62678927,  0.36816272,  0.31044763,  0.44873312,  0.3101446 ]))
array([4, 2, 1, 3, 0])

>>> np.argsort(np.array([ 0.36816272,  0.62678927,  0.13509969,  0.54590815,  0.13493432]))
array([4, 2, 0, 3, 1])

After spending 2 hours on it, I'm convinced that this problem is either too trivial or too technical. I'm using Anaconda virtual environments and have tested it with numpy 1.11.3 and 1.10.4

What is the issue exactly? The output is correct as 0.13493432 < 0.13509969 < 0.36816272 < 0.54590815 < 0.62678927. — fuglede, Jan 12 '17 at 21:10
Thanks for your responses. But I was expecting the second command to return `[2, 4, 1, 3, 0]`. As 0.6 is the largest it should have a number `4` associated with it. What am I missing? — MediocreMyna, Jan 12 '17 at 21:18
But that would be the indices after sorting the array but `argsort` returns the "_indices that would sort the array_" (Source: [argsort documentation](https://docs.scipy.org/doc/numpy/reference/generated/numpy.argsort.html#numpy-argsort)) — MSeifert, Jan 12 '17 at 21:20
Note also that your expected result is exactly the inverse of the output of `np.argsort` as a permutation; that is, you could get your expected result through `s = np.argsort(a)`, `s_inv = np.empty(5)`, `s_inv[s] = np.arange(5)`. It is thus rather unfortunate that in your first case, `s[s]` happens to be `np.arange(5)`. — fuglede, Jan 12 '17 at 21:34

score 2 · Answer 1 · answered Jan 12 '17 at 21:13

It looks like you're under the impression that np.argsort(x) returns an array such that np.argsort(x)[i] is where x[i] goes in sorted order. That's not how it works.

np.argsort(x) returns an array such that x[np.argsort(x)[i]] goes in position i in sorted order. This produces the effect that

x[np.argsort(x)]

is equivalent to

np.sort(x)

This happens to produce the same result as the behavior you expected for the first array, but not for the second.

Numpy argsort strange indices

1 Answers1