Mathematical Comparison Between Marginal and Conditional Variances

504 Views Asked by At

Question: What I want to do is to mathematically prove the following two inequalities: $$Var(Y) \geq Var(Y | X)$$ $$Var(Y | X_1) \geq Var(Y | X_1, X_2)$$

Some hint to the proof will be appreciated. Thanks in advance!

Details Let $X_1, X_2$, and $Y$ be continuous random variables. We can also assume that they are not independent from each other.

1

There are 1 best solutions below

1
On BEST ANSWER

This is not true. (More precisely, this equality does not really make sense, at least not the way you most likely intended.)

By definition, $$\begin{align} \operatorname{Var}(Y\mid X) &= \mathbb{E}[(Y - \mathbb{E}[Y\mid X])^{2}\mid X] = \mathbb{E}[Y^2 - 2Y\mathbb{E}[Y\mid X]+\mathbb{E}[Y\mid X]^{2}\mid X]\\ &= \mathbb{E}[Y^2\mid X] - 2\mathbb{E}[Y\mathbb{E}[Y\mid X]\mid X]+\mathbb{E}[\mathbb{E}[Y\mid X]^{2}\mid X]\\ &= \mathbb{E}[Y^2\mid X] - 2\mathbb{E}[Y\mid X]^2+\mathbb{E}[Y\mid X]^{2}\\ &= \mathbb{E}[Y^2\mid X] - \mathbb{E}[Y\mid X]^2 \end{align}$$ expanding and using the fact that $\mathbb{E}[Y\mid X]$ is $\sigma(X)$-measurable.

Note that this is not a constant, but a random variable, so intuitively the inequality you seek is a "type error."

To see a concrete example, consider the simplest case, when $Y$ has an arbitrary distribution (to be picked later) and $X=Y^2$:, say for simplicity $Y=RN$ where is a Rademacher random variable (uniform on $\{-1,1\}$) independent of $N$ which is Gaussian$(0,1)$.

Then $\mathbb{E}[Y^2\mid X] = \mathbb{E}[X\mid X] = X$, $\mathbb{E}[Y\mid X]^2 = \mathbb{E}[Y\mid Y^2]^2 = \mathbb{E}[RN\mid N^2]^2= \mathbb{E}[R]\mathbb{E}[N\mid N^2]^2=0$ so that $\operatorname{Var}(Y\mid X) = X$.

But $\operatorname{Var}(Y)$ is just a nice number, and since $X$ is continuous and supported on $\mathbb{R}$ we have $\mathbb{P}\{X > \operatorname{Var}(Y) \} > 0$.

That is,

$$\mathbb{P}\{\operatorname{Var}(Y\mid X) > \operatorname{Var}(Y) \} > 0$$


By the law of total variance, however, you do have: $$ \operatorname{Var} Y = \mathbb{E}[\operatorname{Var}( Y \mid X)] + \underbrace{\operatorname{Var}( \mathbb{E}[Y\mid X])}_{\geq 0} \geq \mathbb{E}[\operatorname{Var}( Y \mid X)]. $$ Note that here both LHS and RHS are "numbers."