# What is the correlation of [1,2,3] and [1,5,7] to 8 decimal digits?

This stackoverflow post describes computing a Pearson correlation of [1,2,3] and [1,5,7] in several different ways in Python. The most straightforward implementation from the definition in Wikipedia comes up with

0.973328526785


while Excel, R, NumPy, an online calculator, and a different Python implementation (involving what looks like a more numerically unstable calculation to me) come up with

0.981980506


I am just curious to know what you think.

-
The exact correlation is $\frac{3}{2}\sqrt{\frac{3}{7}} = \sqrt{\frac{27}{28}} \approx$ 0.981 980 506 061 965 716. The value you obtain for the Wikipedia definition looks like a mistake in your implementation: the discrepancy is far too large to be due to numerical instability. (Besides, the Wikipedia formula is the more stable of the two algorithms!) Your result is almost exactly $\sqrt{\frac{18}{19}}$. –  whuber Oct 30 '11 at 19:20
I put a comment in the SO post: Beware of the type of the variables! You have encountered an int/float problem. In sum(x) / len(x) you divide ints, not floats. So sum([1,5,7]) / len([1,5,7]) = 13 / 3 = 4, according to integer division (whereas you want 13. / 3. = 4.33...). To fix it rewrite this line as float(sum(x)) / float(len(x)) (one float suffices, as Python converts it automatically). –  Piotr Migdal Oct 30 '11 at 19:47
Nice catch, @Piotr! –  whuber Oct 30 '11 at 19:48
Nice! If you submit this as an answer, I'll accept it. –  dfrankow Oct 30 '11 at 20:00
You have encountered an int/float problem. In sum(x) / len(x) you divide ints, not floats. So sum([1,5,7]) / len([1,5,7]) = 13 / 3 = 4, according to the integer division rules (whereas you want 13. / 3. = 4.33...). To fix it rewrite this line as float(sum(x)) / float(len(x)) (one float suffices, as Python converts it automatically).