Uniqueness of conditional distributions

48 Views Asked by At

Let $X$ denote a real-valued bounded random variable and $Y$ denote a discrete-valued random variable with $K$ states. Given

$E[X^m]= \sum_{k} c_{km} P(Y=k)$ with constants $c_{km} \in\mathbb{R}$ and $k=1,...,K$ and $m\in\mathbb{R}\cup\{0\}$.

Can we infer from $E[X^m]= \sum_k E[X^m|Y=k] P(Y=k)$ that $c_{km}=E[X^m|Y=k]$?

What confuses me is that the equality $\sum_{k} c_{km} P(Y=k)=\sum_k E[X^m|Y=k] P(Y=k)$ holds only for the sum. Does it also hold for each element in the sum $c_{km}=E[X^m|Y=k]$ and, if yes, why?

Thanks for your comments / suggestions / counter examples?

1

There are 1 best solutions below

0
On BEST ANSWER

Either the conjecture is obviously false, or I am missing something obvious...

Each $c_{km}$ only appears in one equation, that of $E[X^m]$, right? So, for every $m$, you have $K$ unknowns (the $K$ coefficients) and one equation. How can the coefficients be determined?

E.g. lets say $P(Y=1) > 0$. One possible set of coefficients would be:

$$c_{1m} = {E[X^m] \over P(Y=1)}; ~~~~~~~~~~~~~~~ c_{km} = 0 ~~ \forall k \neq 1$$

Does that work? Or am I missing something really obvious...