Python vs Julia autocorrelation

Ross Mariano :

I am trying to do autocorrelation using Julia and compare it to Python's result. How come they give different results?

Julia code

using StatsBase

t = range(0, stop=10, length=10)
test_data = sin.(exp.(t.^2))

acf = StatsBase.autocor(test_data)

gives

10-element Array{Float64,1}:
  1.0                   
  0.13254954979179642   
 -0.2030283419321465    
  0.00029587850872956104
 -0.06629381497277881   
  0.031309038331589614  
 -0.16633393452504994   
 -0.08482388975165675   
  0.0006905628640697538 
 -0.1443650483145533

Python code

from statsmodels.tsa.stattools import acf
import numpy as np

t = np.linspace(0,10,10)
test_data = np.sin(np.exp(t**2))

acf_result = acf(test_data)

gives

array([ 1.        ,  0.14589844, -0.10412699,  0.07817509, -0.12916543,
       -0.03469143, -0.129255  , -0.15982435, -0.02067688, -0.14633346])

Jakob Nissen :

This is because your test_data is different:

Python:

array([ 0.84147098, -0.29102733,  0.96323736,  0.75441021, -0.37291918,
        0.85600145,  0.89676529, -0.34006519, -0.75811102, -0.99910501])

Julia:

[0.8414709848078965, -0.2910273263243299, 0.963237364649543, 0.7544102058854344,
 -0.3729191776326039, 0.8560014512776061, 0.9841238290665676, 0.1665709194875013,
 -0.7581110212957692, -0.9991050130774393]

This happens because you are taking sin of enormous numbers. For example, with the last number in t being 10, exp(10^2) is ~2.7*10^43. At this scale, floating point inaccuracies are about 3*10^9. So if even the least significant bit is different for Python and Julia, the sin value will be way off.

In fact, we can inspect the underlying binary values of the initial array t. For example, they differ in the third last value:

Julia:

julia> reinterpret(Int, range(0, stop=10, length=10)[end-2])
4620443017702830535

Python:

>>> import struct
>>> s = struct.pack('>d', np.linspace(0,10,10)[-3])
>>> struct.unpack('>q', s)[0]
4620443017702830536

We can indeed see that they disagree by exactly one machine epsilon. And if we use Julia take sin of the value obtained by Python:

julia> sin(exp(reinterpret(Float64, 4620443017702830536)^2))
-0.3400651855865199

We get the same value Python does.