interp in NumPy/SciPy

[oyamad]

This is a minor issue.
In the whole library, interp is imported as from scipy, and one of the lectures states "Linear interpolation is performed by SciPy’s interp function", but it is actually a function from NumPy (numpy.interp).

Importing from scipy is fine because the __init__.py in SciPy has from numpy import *. The problem is that if the user looks into SciPy for the documentation (as I did), it is not found there, but in NumPy.

[mmcky]

Thanks @oyamad - you are correct - this should be fixed. The alternative would be to use interp1d which is imported from scipy.

[lbui01]

Hi @oyamad - I have opened a PR to change the scipy.interp to be directly used from numpy. See PR jstac/QuantEcon.lectures#90

This also needs to be updated in the solutions_py.ipynb. See issue jstac/QuantEcon.lectures#92.

We should be able to close this issue soon.

Also note:

I ran a speed test for np.interp and scipy.interpolate.interp1d and found that np.interp was significantly faster. 8.31ms for interp and 38.4ms for interp1d (including initializing the function). If anyone is interested, see this gist: https://gist.github.com/lbui01/1e7f2bffeae8296a31e49727bb710ef4

[jstac]

@lbui01 Thanks, nice work.

I guess the main difference is that interp returns an array whereas interp1d returns a callable object.

Having a callable object (e.g., a function) is often more convenient. In fact you can see in this line

w_func = lambda x: interp(x, grid, w)

from the optimal growth lecture that I'm creating a callable using lambda syntax. If I'd known I might have just used interp1d1.

I looked at the relative performance as well, and, if you do want to make a callable, then interp is only slightly faster. See here. That kind of difference is negligible.

[jstac]

In summary, if we want to make a callable I suggest we use interp1d imported from SciPy, whereas if we only wish to make an array (i.e., evaluate once off) then we use interp, imported from NumPy.

[lbui01]

Thanks @jstac, I see why interp1d is useful.