1
$\begingroup$

Let $x_0 \sim N(\mu_0,\Sigma_0)$, where $\mu_0 \in \mathbb R^n$ and $\Sigma_0 \in \mathbb R^{n \times n}$. Then, for $k = 0,\dots,N-1$, let $$x_{k+1} = Ax_k$$ for $A \in \mathbb R^{n \times n}$. Next, define $$E_{x_0,x_1,\dots,x_{N}}[x_0] = \int_{\mathbb R^n} \int_{\mathbb R^n} \cdots \int_{\mathbb R^n} x_0 p(x_0,x_1,\dots,x_N) \text{d}x_N \dots \text{d}x_1 \text{d}x_0$$ and $$E_{x_0}[x_0] = \int_{\mathbb R^n} x_0 p(x_0) \text{d}x_0$$ Is it true that $$E_{x_0,x_1,\dots,x_{N}}[x_0] = E_{x_0}[x_0]$$ It would seem to me that the answer is yes. First, the sequence $x_0 \to x_1 \to \cdots \to x_N$ forms a Markov chain, such that the joint distribution $p(x_0,x_1,\dots,x_N)$ simplifies to $$p(x_0,x_1,\dots,x_N) = p(x_N \mid x_{N-1})p(x_{N-1} \mid x_{N-2})\cdots p(x_1 \mid x_0) p(x_0)$$ Therefore, we can write \begin{align} E_{x_0,x_1,\dots,x_{N-1}}[x_0] &= \int_{\mathbb R^n} \int_{\mathbb R^n} \cdots \int_{\mathbb R^n} x_0 p(x_0,x_1,\dots,x_N) \text{d}x_N \dots \text{d}x_1 \text{d}x_0 \\ &= \int_{\mathbb R^n} \int_{\mathbb R^n} \cdots \int_{\mathbb R^n} x_0 p(x_N \mid x_{N-1})p(x_{N-1} \mid x_{N-2})\cdots p(x_1 \mid x_0) p(x_0) \text{d}x_N \dots \text{d}x_1 \text{d}x_0 \\ &= \int_{\mathbb R^n} \int_{\mathbb R^n} \cdots \int_{\mathbb R^n} p(x_N \mid x_{N-1}) \text{d}x_N \dots p(x_1 \mid x_0) \text{d}x_1 x_0 p(x_0) \text{d}x_0 \end{align} Then, the integrals $$\int_{\mathbb R^n} p(x_N \mid x_{N-1}) \text{d}x_N$$ and $$\int_{\mathbb R^n} p(x_{N-1} \mid x_{N-2}) \text{d}x_{N-1}$$ and so on evaluate to $1$, and so we are left with \begin{align} E_{x_0,x_1,\dots,x_{N-1}}[x_0] &= \int_{\mathbb R^n} \int_{\mathbb R^n} \cdots \int_{\mathbb R^n} p(x_N \mid x_{N-1}) \text{d}x_N \dots p(x_1 \mid x_0) \text{d}x_1 x_0 p(x_0) \text{d}x_0 \\ &= \int_{\mathbb R^n} x_0 p(x_0) \text{d}x_0 \\ &= E_{x_0}[x_0] \end{align} Is my reasoning correct?

$\endgroup$
6
  • $\begingroup$ Why do you think it forms a Markov chain instead of being deterministic? Imagine $A$ is a square matrix of all ones; then at each step, the first element of $x_{i+1}$ is just the sum of the elements of $x_i$, etc. It's a matrix multiplication, not a transition matrix. $\endgroup$
    – jbowman
    Feb 2 at 0:57
  • $\begingroup$ As written, $x_k = A^kx_0$, so $\mathbb{E}x_k = A^k\mathbb{E}x_0$ $\endgroup$
    – jbowman
    Feb 2 at 1:04
  • $\begingroup$ @jbowman What I mean by "the sequence $x_0 \to x_1 \to \cdots \to x_N$ forms a Markov chain" is that $x_{k}$ is conditionally independent of $x_{k-2}$ given $x_{k-1}$ for $k = 2,\dots,N-1$. Is there something wrong with saying so? $\endgroup$
    – mhdadk
    Feb 2 at 14:56
  • $\begingroup$ It's deterministic. One could say that $x_{i+t} = x_i + 1$ is a Markov chain, but why? There's no randomness in the transitions; describing multiplication and addition as a Markov chain adds nothing useful to the discussion. $\endgroup$
    – jbowman
    Feb 2 at 16:05
  • 1
    $\begingroup$ Ah, I see... that's actually kind of clever, if roundabout! $\endgroup$
    – jbowman
    Feb 2 at 17:55

1 Answer 1

5
$\begingroup$

You don't need the Markov property. For any two jointly defined random variables $X$ and $Y$, $$E_X[X]=E_{X,Y}[X]$$ The marginal distribution of $X$ is just the distribution of $X$. Now just take $X$ to be $X_0$ and $Y$ to be everything else.

$\endgroup$

Your Answer

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

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