Given three vectors $a$, $b$, and $c$, is it possible that correlations between $a$ and $b$, $a$ and $c$, and $b$ and $c$ are all negative? I.e. is this possible?

\begin{align} \text{corr}(a,b) < 0\\ \text{corr}(a,c) < 0 \\ \text{corr}(b,c) < 0\\ \end{align}

  • 3
    $\begingroup$ Negative correlations mean, geometrically, that the centered vectors mutually make obtuse angles. You should have no problem drawing a configuration of three vectors in the plane that have this property. $\endgroup$ – whuber Jan 24 '18 at 15:52
  • $\begingroup$ They cannot be completely negatively correlated ($\rho=-1$), but in general there can be some negative correlation, again bounds set by the other correlations. $\endgroup$ – karakfa Jan 24 '18 at 20:27
  • 2
    $\begingroup$ @whuber Your comment seems to contradict Heikki Pulkkinen's answer, which claims it's impossible for vectors in a plane. If you stand by it, you should turn your comment into an answer. $\endgroup$ – R.M. Jan 24 '18 at 20:58
  • 2
    $\begingroup$ @R.M. There is no contradiction between whuber and Heikki. This question asks about data matrix $X$ of $n\times 3$ size. Normally we would talk about $n$ data points in 3 dimensions, but this Q is talking about three "vectors" in $n$ dimensions. Heikki says that all negative correlations cannot happen if $n=2$ (indeed, two points after centering are always perfectly correlated, so correlations must be $\pm 1$ and cannot be all $-1$). Whuber says that 3 vectors in $n$ dimensions can effectively lie in a 2-dimensional subspace (i.e. $X$ is rank 2) and suggests to imagine a Mercedes logo. $\endgroup$ – amoeba Jan 24 '18 at 21:18
  • 1
    $\begingroup$ Related: Bound for the correlation of three random variables. (cc, @amoeba) $\endgroup$ – gung - Reinstate Monica Jan 24 '18 at 21:19

It is possible if the size of the vector is 3 or larger. For example

\begin{align} a &= (-1, 1, 1)\\ b &= (1, -9, -3)\\ c &= (2, 3, -1)\\ \end{align}

The correlations are \begin{equation} \text{cor}(a,b) = -0.80...\\ \text{cor}(a,c) = -0.27...\\ \text{cor}(b,c) = -0.34... \end{equation}

We can prove that for vectors of size 2 this is not possible: \begin{align} \text{cor}(a,b) &< 0\\[5pt] 2\Big(\sum_i a_i b_i\Big) - \Big(\sum_i a_i\Big)\Big(\sum_i b_i\Big) &< 0\\[5pt] 2(a_1 b_1 + a_2 b_2) - (a_1 + a_2)(b_1 b_2) &< 0\\[5pt] 2(a_1 b_1 + a_2 b_2) - (a_1 + a_2)(b_1 b_2) &< 0\\[5pt] 2(a_1 b_1 + a_2 b_2) - a_1 b_1 + a_1 b_2 + a_2 b_1 + a_2 b_2 &< 0\\[5pt] a_1 b_1 + a_2 b_2 - a_1 b_2 + a_2 b_1 &< 0\\[5pt] a_1 (b_1-b_2) + a_2 (b_2-b_1) &< 0\\[5pt] (a_1-a_2)(b_1-b_2) &< 0 \end{align}

The formula makes sense: if $a_1$ is larger than $a_2$, $b_1$ has to be larger than $b_1$ to make the correlation negative.

Similarly for correlations between (a,c) and (b,c) we get

\begin{equation} (a_1-a_2)(c_1-c_2) < 0\\ (b_1-b_2)(c_1-c_2) < 0\\ \end{equation}

Clearly, all of these three formulas can not hold in the same time.

  • 3
    $\begingroup$ Another example of something unexpected that only happens in dimension three or higher. $\endgroup$ – nth Jan 24 '18 at 19:21
  • 1
    $\begingroup$ With vectors of size $2$, correlations are usually $\pm1$ (straight line through two points), and you cannot have three correlations of $-1$ with three vectors of any size $\endgroup$ – Henry Jan 25 '18 at 15:14

Yes, they can.

Suppose you have a multivariate normal distribution $X\in R^3, X\sim N(0,\Sigma)$. The only restriction on $\Sigma$ is that it has to be positive semi-definite.

So take the following example $\Sigma = \begin{pmatrix} 1 & -0.2 & -0.2 \\ -0.2 & 1 & -0.2 \\ -0.2 & -0.2 & 1 \end{pmatrix} $

Its eigenvalues are all positive (1.2, 1.2, 0.6), and you can create vectors with negative correlation.


let's start with a correlation matrix for 3 variables

$\Sigma = \begin{pmatrix} 1 & p & q \\ p & 1 & r \\ q & r & 1 \end{pmatrix} $

non-negative definiteness creates constraints for pairwise correlations $p,q,r$ which can be written as

$$ pqr \ge \frac{p^2+q^2+r^2-1}2 $$

For example, if $p=q=-1$, the values of $r$ is restricted by $2r \ge r^2+1$, which forces $r=1$. On the other hand if $p=q=-\frac12$, $r$ can be within $\frac{2 \pm \sqrt{3}}4$ range.

Answering the interesting follow up question by @amoeba: "what is the lowest possible correlation that all three pairs can simultaneously have?"

Let $p=q=r=x < 0$, Find the smallest root of $2x^3-3x^2+1$, which will give you $-\frac12$. Perhaps not surprising for some.

A stronger argument can be made if one of the correlations, say $r=-1$. From the same equation $-2pq \ge p^2+q^2$, we can deduce that $p=-q$. Therefore if two correlations are $-1$, third one should be $1$.


A simple R function to explore this:

f <- function(n,trials = 10000){
  count <- 0
  for(i in 1:trials){
    a <- runif(n)
    b <- runif(n)
    c <- runif(n)
    if(cor(a,b) < 0 & cor(a,c) < 0 & cor(b,c) < 0){
      count <- count + 1

As a function of n, f(n) starts at 0, becomes nonzero at n = 3 (with typical values around 0.06), then increases to around 0.11 by n = 15, after which it seems to stabilize:

enter image description here So, not only is it possible to have all three correlations negative, it doesn't seem to be terribly uncommon (at least for uniform distributions).


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.