In the MSE decomposition formula why does the following hold?

$ {\begin{aligned}{E} _{\theta }\left[2\left({\hat {\theta }}-\operatorname {E} _{\theta }[{\hat {\theta }}]\right)\left(\operatorname {E} _{\theta }[{\hat {\theta }}]-\theta \right)\right]+\operatorname {E} _{\theta }\left[\left(\operatorname {E} _{\theta }[{\hat {\theta }}]-\theta \right)^{2}\right] &=& \\2\left(\operatorname {E} _{\theta }[{\hat {\theta }}]-\theta \right)\operatorname {E} _{\theta }\left[{\hat {\theta }}-\operatorname {E} _{\theta }[{\hat {\theta }}]\right]+\left(\operatorname {E} _{\theta }[{\hat {\theta }}]-\theta \right)^{2}&&\\\\\end{aligned}}$

I know we already have this question here which clearly explains that $\mathbb{E}[\mathbb{E}[\hat{\theta}] - \hat{\theta}]$ is 0 since:

$\mathbb{E}[\mathbb{E}[\hat{\theta}] - \hat{\theta}] = \mathbb{E}[\mathbb{E}[\hat{\theta}]] + \mathbb{E}[\hat{\theta}] = \mathbb{E}[\hat{\theta}] - \mathbb{E}[\hat{\theta}] = 0$

But that alone doesn't seem to explain all the steps that actually took place. How does one derive the bottom equation from the first?

  • $\begingroup$ Thanks @SextusEmpiricus it's verbatim from the Wikipedia page (I just added a link). Maybe that's because $\theta$ is assumed to be a random variable? $\endgroup$
    – Josh
    Commented May 30, 2020 at 16:55

2 Answers 2


In that Wikipedia page people have been messing up the equations with many edits and eventually adding subscripts which makes it not more clear and too much cluttered.

That subscript started as $E_\theta$. Somebody changed it into $E_{\hat\theta}$. And now it is changed back into $E_\theta$ with the explanation

expectation given theta not hat{theta}

That is not the typical use of subscripts with an expectation operator.

My understanding of the subscript is as

$$E_Y(g(Y)) = \int_{\forall Y} g(y) f_Y(y) dy$$

(where $f_Y(y)$ is the probability density of $Y$) and then it is the expectation of the value of $g(Y)$ where $Y$ is the variable, and we integrate over the probabilities for all different $Y$.

(Or at least it is not what I am used to and not the expectation that I expected. But there seem to be multiple uses for the subscript. Anyway it is at least ambiguous and should be clarified in the text)

With the different interpretation, $E_{\theta}(X)$ as the expectation of $X$ given $\theta$ (ie. conditional on $\theta$), then it is like the previous question here on SE (the one the OP linked to), where in my answer I also commented that it is to be considered conditional on $\theta$ (a comment that was made because some people where confused about what was constant or not).


The struggle seems to be that in order to distribute $\mathbb{E}[a \cdot b]$ within the middle term shown in step 3-4 of the referenced question (or the first term above), $a$ and $b$ must be statistically independent. The joint probability must be equal to the product of individual probabilities so that $\mathbb{E}(a)*\mathbb{E}(b)$ is valid algebra. It is questionable that the distances $a$ and $b$ are independent, as would not an estimator be correlated with the truth.

The resolution consists in the problem framing. If we want to define a metric for the accuracy of a regression model, knowing or assuming a true distribution function $f(X)$, we generate a sampling distribution (over various subsets $D$ within $X$). In this setup, $Y$ is considered deterministic and we say $Y$ = $f(X)$ + $\epsilon$. Then, armed with this distribution, we would want to recast accuracy/error using its parameters. Thus, squared error distance is divided into segments with this mean (the first step in the derivation). And because we squared the error, we happen to end up with model variance.

It might otherwise be difficult to evaluate an estimator if $Y$ cannot be considered deterministic.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.