Assume that we have two normalized $n\times 1$ vectors $\bf x$ and $\bf y$. In terms of Pearson correlation, these two signals are uncorrelated if ${\bf x}^T {\bf y} = 0$. Now, assume that ${\bf x}^T {\bf y} \neq 0$. If we perform the following projection: $$ \widehat{\bf x} = {\bf x} - ({\bf x}^T {\bf y}) {\bf y}, $$ we get $\widehat{\bf x}^T {\bf y} = 0$.

In all the above computations, we use Pearson correlation.

My questions is if we can perform such projection for Kendall Tau correlation. In other words, if we have two normalized $n\times 1$ vectors $\bf x$ and $\bf y$ such that their Kendall Tau correlation is not zero, can we perform a projection of $\bf x$ to get $\widehat{\bf x}$ such that the Kendall Tau correlation of $\widehat{\bf x}$ and $\bf y$ is zero?


Covariance is linear, so a linear projection can be used to zero it out.

Concordance is not linear, so a linear projection won't (in general) work to zero it out.

However, one can still come up with vectors which have zero Kendall correlation.

Specifically, if $\hat{\beta}^K$ is the slope estimate for the Theil-Sen regression of $y$ on $x$, then the Kendall correlation of $x$ and $r=y-\hat{\beta}^K x$ will be 0.

  • $\begingroup$ This is what I was looking for, thanks. Is there an extension of this on higher-dimensional data. Say, I want to make y uncorrelated with two vectors x1 and x2. $\endgroup$
    – MMM
    Mar 17 '15 at 13:02
  • 1
    $\begingroup$ That one might be best asked as a new question. $\endgroup$
    – Glen_b
    Mar 17 '15 at 13:10

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.