# How can this counterintiutive result with the Mahalanobis distance be explained?

I encountered a strange issue when performing Mahalanobis distance matching. Let's say I have one treated unit with the following values on two variables: $$T:(17, 4)$$. I have two control units with values $$A:(16, 3)$$ and $$B:(15, 3)$$. Intuitively, it seems like $$T$$ should be closer to $$A$$ than to $$B$$, and that would be true regardless of any affine transformation of the variables. The covariance matrix for my variables (which is computed in the full sample that includes many more observations) is $$\Sigma = \begin{bmatrix}25 & 9\\9 & 5\end{bmatrix}$$

When I compute the Mahalanobis distance $$d(.,.)$$, I find that $$d(T,A) = 0.522$$ and $$d(T,B)=0.452$$; that is, $$B$$ is closer to $$T$$ than $$A$$ is on the Mahalanobis distance. This doesn't make much sense to me; intuitively, what is going on here?

Some R code to play around with:

T <- c(17, 4)
A <- c(16, 3)
B <- c(15, 3)

S <- matrix(c(25, 9,
9, 5), nrow = 2)

mahalanobis(T, A, S) |> sqrt()
## [1] 0.522233
mahalanobis(T, B, S) |> sqrt()
## [1] 0.452267

• The correlation between the two dimensions is very high, creating a "ridge" that runs more or less diagonally. $(16,3)$ is off to one side of the ridge, so it is less probable (Gaussian) or farther away (Mahalanobis) than $(17,4)$, which is closer to the "ridgeline". Commented Jun 28, 2023 at 0:31
• @jbowman The question is why $(15,3)$ is closer to $(17,4)$ on the Mahalanobis distance than $(16,3)$ is. Surely $(15,3)$ is even less probable and farther away from $(17,4)$, no?
– Noah
Commented Jun 28, 2023 at 1:25
• Why not draw a picture? More specifically, draw the points $T,A,$ and $B$ on a pair of axes. Then, with $T$ at the center, start drawing concentric ovals that represent the level sets of the Mahalanobis distance with respect to $T$. You should then find that $B$ lies on an oval that is inside the oval that $A$ lies on. Commented Jun 28, 2023 at 1:54
• I know what you're asking, and the accepted answer illustrates in a picture that is worth 1000+ words, what I was trying to say :) Commented Jun 28, 2023 at 2:32
• I'm just remaking jbowman's point a little differently, but notice that for a jointly Gaussian $(X,Y)$ with covariance $\Sigma$ and mean $(17,4)$, the conditional mean of $X$ given $Y = 3$ is $17 + (9/5)(3-4) = 15 + 1/5.$ Of course, $15$ is much closer to this conditional mean than $16$, so given $Y = 3,$ $X = 15$ is much more likely than $X = 16.$ The ordering of Mahalanobis distances reflects this fact. Commented Jun 30, 2023 at 8:18

Here are contours of the Mahalanobis distance/Gaussian likelihood centred at T (17, 4) (open circle), and two points A: (16,3) and B(15,3). You can see the point at (15,3) is closer than that at (16,3) in this metric.

library(ellipse)
plot(ellipse(S,centre=T,level=0.6),type="n")
points(c(16,15,17),c(3,3,4),pch=c(19,19,1))
polygon(ellipse(S,centre=T,level=0.6),col="#AA00AA20",border=NA)
polygon(ellipse(S,centre=T,level=0.5),col="#AA00AA20",border=NA)
polygon(ellipse(S,centre=T,level=0.4),col="#AA00AA20",border=NA)
polygon(ellipse(S,centre=T,level=0.3),col="#AA00AA20",border=NA)
polygon(ellipse(S,centre=T,level=0.2),col="#AA00AA20",border=NA)
polygon(ellipse(S,centre=T,level=0.1),col="#AA00AA20",border=NA)

• This is extremely helpful and just what I needed to understand the comments on the original post. Thanks to you and @mhdadk.
– Noah
Commented Jun 28, 2023 at 2:10
• Nice one! Looks really good. Commented Jun 28, 2023 at 12:13

Very qualitative but visual explanation: since your covariance matrix is not diagonal, the distribution of the samples is going to be tilted with respect to the main axes. Both $$(17,4)$$ and $$(15,3)$$ are close to its main axis, and are thus close to each other under the Mahalanobis distance. On the opposite, $$(16,3)$$ is closer to the "outside" of the samples cloud.

The other answers provided excellent graphical explanations. In this answer, I'll provide a numerical one.

Given a positive-definite covariance matrix $$\Sigma$$, the Mahalanobis distance between two points $$x$$ and $$y$$ in $$\mathbb R^n$$ is $$\sqrt{(x-y)^T \Sigma^{-1}(x-y)}$$ Because every covariance matrix is positive semi-definite, then its square root always exists and is symmetric. Let's call this square root $$\Sigma^{1/2}$$, such that $$\Sigma = \Sigma^{1/2}\Sigma^{1/2}$$. Let's also assume that $$\Sigma$$ is positive-definite, such that $$\Sigma^{-1}$$ exists and is equal to $$\Sigma^{-1/2}\Sigma^{-1/2}$$, where $$\Sigma^{-1/2}$$ is the inverse of $$\Sigma^{1/2}$$. Then, the Mahalanobis distance becomes \begin{align} \sqrt{(x-y)^T \Sigma^{-1}(x-y)} &= \sqrt{(x-y)^T\Sigma^{-1/2}\Sigma^{-1/2}(x-y)} \\ &= \sqrt{\left(\Sigma^{-1/2}(x-y)\right)^T\left(\Sigma^{-1/2}(x-y)\right)} \\ &= \sqrt{z^T z} \\ &= \|z\|_2 \end{align} where $$z = \Sigma^{-1/2}(x-y)$$. We can therefore see that the Mahalanobis distance is nothing but the Euclidean norm of the "whitened" version of $$x-y$$, such that if $$x-y$$ has a covariance of $$\Sigma$$, then $$z = \Sigma^{-1/2}(x-y)$$ has a covariance of $$I$$.

Going back to our case, let $$\Sigma = \begin{bmatrix}25 & 9\\9 & 5\end{bmatrix}$$ Then (computed using MATLAB's sqrtm function), $$\Sigma^{1/2} \approx \begin{bmatrix}4.8091 & 1.3683\\1.3683 & 1.7686\end{bmatrix}$$ and $$\Sigma^{-1/2} \approx \begin{bmatrix}0.2666 & -0.2063\\-0.2063 & 0.7250\end{bmatrix}$$ We can then compute $$z_A = \Sigma^{-1/2}(T - A)$$ and $$z_B = \Sigma^{-1/2}(T - B)$$ as \begin{align} T - A &= \begin{bmatrix}17 \\ 4\end{bmatrix} - \begin{bmatrix}16 \\ 3\end{bmatrix} = \begin{bmatrix}1 \\ 1\end{bmatrix} \\ z_A &= \Sigma^{-1/2}(T - A) \approx \begin{bmatrix}0.0604 \\ 0.5187\end{bmatrix} \\ T - B &= \begin{bmatrix}17 \\ 4\end{bmatrix} - \begin{bmatrix}15 \\ 3\end{bmatrix} = \begin{bmatrix}2 \\ 1\end{bmatrix} \\ z_B &= \Sigma^{-1/2}(T - B) \approx \begin{bmatrix}0.3270 \\ 0.3125\end{bmatrix} \end{align} By computing $$\|z_A\|$$ and $$\|z_B\|$$, we can see that $$\|z_A\| > \|z_B\|$$.