Covariance

The covariance generalizes the concept of variance to multiple random variables. Instead of measuring the fluctuation of a single random variable, the covariance measures the fluctuation of two variables with each other.

Recall that the variance is the mean squared deviation from the mean for a single random variable $X$: $\text{Var}(X) = E[\left(X - E[X]\right)^2].$ The covariance adopts an analogous functional form.

The covariance $\text{Cov}(X, Y)$ of random variables $X$ and $Y$ is defined as $\text{Cov}(X, Y) = E\left[(X - E[X])(Y - E[Y])\right].$

It is generally simpler to find the covariance by taking $\begin{aligned} \text{Cov}(X, Y) &= E[XY - E[X] Y - X E[Y] + E[X] E[Y]] \\ &= \boxed{E[XY] - E[X] E[Y].} \end{aligned}$ In other words, to compute the covariance, one can equivalently find $E[XY]$ (in addition to the means of $X$ and $Y$).

Similarly, one can find an expression in terms of variances: $\begin{aligned} \text{Var}(X + Y) &= E\left[(X + Y - E[X] - E[Y])^2\right] \\ &= E[\left(X - E[X]\right)^2] + E[\left(Y - E[Y]\right)^2] + 2 E\left[(X - E[X])(Y - E[Y])\right] \\ &= \boxed{\text{Var}(X) + \text{Var}(Y) + 2 \text{Cov}(X, Y).} \end{aligned}$

A generalized statement of this result is as follows.

Variance of a sum. Given random variables $X_i$, each with finite variance,

$\text{Var}\left( \sum_i X_i \right) = \sum_i \ \text{Var}(X_i) + 2 \sum_{i<j} \text{Cov}(X_i, X_j).$

The covariance inherits many of the same properties as the inner product from linear algebra. The proof involves straightforward algebra and is left as an exercise for the reader.

Given a constant $a$ and random variables $X$, $Y$, and $Z$, the following properties hold:

• $\text{Cov}(X, X) = \text{Var}(X) \geq 0$
• $\text{Cov}(X, Y) = \text{Cov}(Y, X)$
• $\text{Cov}(aX, Y) = a \text{Cov}(X, Y)$
• $\text{Cov}(X, a) = 0$
• $\text{Cov}(X + Y, Z) = \text{Cov}(X, Z) + \text{Cov}(Y, Z)$.

Let $X$ and $Y$ be random variables such that $\text{Var}(X) = \sigma^2$ and $Y = aX$, where $\sigma$ and $a$ are constants. Determine $\text{Cov}(X, Y)$.

---

The inner product properties yield

$\text{Cov}(X, Y) = \text{Cov}(X, aX) = \text{Cov}(aX, X) = a\text{Cov}(X, X) = a \sigma^2.$

As a result, the Cauchy-Schwarz inequality holds for covariances.

Cauchy-Schwarz inequality. Given random variables $X$ and $Y$,

$\left[ \text{Cov}(X ,Y) \right]^2 \leq \text{Var}(X) \text{Var}(Y).$

One of the key properties of the covariance is the fact that independent random variables have zero covariance.

Covariance of independent variables. If $X$ and $Y$ are independent random variables, then $\text{Cov}(X, Y) = 0.$

If $X$ and $Y$ are independent, then $E[XY] = E[X] E[Y]$ and therefore $\text{Cov}(X, Y) = 0$. (Recall that $E[XY] = E[X] E[Y]$ is a simple consequence of the fact that $P(X | Y) = P(X)$.)

Dependent variables with zero covariance. However, the converse is not in general true. As a simple example, suppose that $X$ is a standard normal random variable and that $Y = X^2$. Notice that knowledge of $X$ completely determines $Y$, in which case $X$ and $Y$ are very clearly dependent. However, by symmetry it holds that $\text{Cov}(X, Y) = E[XY] - E[X] E[Y] = 0.$

A simple corollary is as follows.

Variance of the sum of independent variables. Given independent random variables $X_i$, each with finite variance,

$\text{Var}\left( \sum_i X_i \right) = \sum_i \ \text{Var}(X_i).$

Since the $X_i$ are independent, it must be the case that $\text{Cov}(X_i, X_j) = 0$ for all $i \neq j$, and the result follows directly from the variance of a sum theorem.

When dealing with a large number of random variables $X_i$, it makes sense to consider a covariance matrix whose $m,n$th entry is $\text{Cov}(X_m, X_n)$.

Since $\text{Cov}(X, Y) = \text{Cov}(Y, X)$, the covariance matrix is symmetric.

[1] DeGroot, Morris H. Probability and Statistics. Second edition. Addison-Wesley, 1985.