diff --git a/source/ch-Expected-Value.ptx b/source/ch-Expected-Value.ptx index 4aefa7d..7a869e2 100644 --- a/source/ch-Expected-Value.ptx +++ b/source/ch-Expected-Value.ptx @@ -6,7 +6,7 @@

- Testing in codespace + Testing locally

@@ -871,10 +871,133 @@
Covariance + + +

+ Let X and Y be random variables. + The covariance of X and Y is: + + \Cov(X, Y) = \E\left[\left(X - \E(X)\right) \left(Y - \E(Y)\right)\right]. + +

+
+
+

- Text of section. + Let's write \mu_X = \E(X) and \mu_Y = \E(Y) to make this is a bit cleaner to look at: + + \Cov(X, Y) = \E\left[\left(X - \mu_X \right) \left(Y - \mu_Y\right)\right]. + + This looks like a pretty natural generalization of the variance. + In fact: + + \Cov(X, X) \amp = \E\left[ (X - \mu_X)(X - \mu_X)\right] = \E\left[\left(X - \mu_X\right)^2\right] = \Var(X). + + One difference worth noting is that \Var(X), as the average of squared values, must always be nonnegative. + The expression (X - \mu_X)(X - \mu_X) can be negative for combinations of X and Y values where one is greater than its average and the other is smaller. + If this is the average behavior, then \Cov(X, Y) can be negative overall. + This is useful information: a negative \Cov(X, Y) would imply that (on average, not necessarily always) when X is larger than typical, Y is smaller than typical, and vice versa.

+

+ Also as with variance, the covariance definition formula gives us some useful insight, but it can be unwieldy to work with. + There is an alternative formula, analogous to : +

+ + + +

+ Let X and Y be random variables. + Then: + + \Cov(X, Y) = \E\left(XY\right) - \E(X)\E(Y). + +

+
+
+ +

+ To make use of this formula, it's helpful to have both a joint distribution and marginal distributions. + The marginal distributions let you find \E(X) and \E(Y). + The joint distribution lets you find \E(XY). +

+ + + +

+ Let X and Y be indicator random variables with joint and marginal distributions below. +

+ + + + + + + + + X = 0 + X = 1 + + + + Y = 0 + 0.2 + 0.1 + + + + Y = 1 + 0.2 + 0.5 + + +
+ + + + + x\Pr(X = x) + + + 00.4 + + + 10.6 + + +
+ + + + + y\Pr(Y = y) + + + 00.3 + + + 10.7 + + +
+
+

+ Since they're indicator random variables, we can immediately see that \E(X) = 0.6 and \E(Y) = 0.7 (see ). + For \E(XY), we'll traverse cell by cell through the joint distribution table, constructing products of the form X-value times Y-value times probability: + + \E(XY) =\, \amp (0)(0)(0.2) + (1)(0)(0.1) + \amp (0)(1)(0.2) + (1)(1)(0.5) + =\, \amp 0.5. + + Finally: + + \Cov(X, Y) = \E(XY) - \E(X)\E(Y) = 0.5 - (0.6)(0.7) = 0.08. + + +

+
+
+ diff --git a/source/docinfo.ptx b/source/docinfo.ptx index ccf7d92..1a35ef9 100644 --- a/source/docinfo.ptx +++ b/source/docinfo.ptx @@ -35,6 +35,7 @@ \DeclareMathOperator{\E}{E} \DeclareMathOperator{\Var}{Var} + \DeclareMathOperator{\Cov}{Cov}