0
$\begingroup$

I want to show that $ (e^{\alpha X_t} \int^ t_ 0 e ^{-\alpha X_u}du, t \geq 0) $ is a Markov process whereas $(\int^ t_ 0 e ^{-\alpha X_u}du, t\geq 0)$ is not. Here $X_t$ is a Levy Process and nothing has been said about $\alpha$.

I could express $$Y_t = e^{\alpha X_t} \int^ t_ 0 e ^{-\alpha X_u}du = e^{\alpha (X_t-X_s)}Y_s+ e^{\alpha (X_t-X_s)}\int^ t_ s e ^{-\alpha (X_u-X_s)}du $$ Then how does one use the independence to show the markov process and for showing that other one is not a markov process.

$\endgroup$
  • 1
    $\begingroup$ For every $t<s$, $$Y_{s}=A_{t,s}Y_t+B_{t,s}$$ where $A_{t,s}$ and $B_{t,s}$ are measurable with respect to $\sigma(X_u-X_t;u>t)$, since $$A_{t,s}=e^{\alpha(X_s-X_t)}\qquad B_{t,s}=A_{t,s}\int_t^se^{-\alpha(X_u-X_t)}du$$ Thus, $A_{t,s}$ and $B_{t,s}$ are are independent of $\sigma(X_u;u\leqslant t)$. $\endgroup$ – Did Mar 1 '18 at 19:58
1
$\begingroup$

I think there is a typo in the identity which you mentioned in your question; it should read

$$Y_t = e^{\alpha(X_t-X_s)} Y_s + e^{\alpha(X_t-X_s)} \int_s^t e^{-\alpha(X_u-X_s)} \, du.\tag{1}$$

In order to show that $(Y_t)_{t \geq 0}$ is a Markov process, we can use the following statement

Let $\mathcal{F}$ be a $\sigma$-algebra and let $X$, $Y$ be two random variables such that $X$ is independent from $\mathcal{F}$ and $Y$ is $\mathcal{F}$-measurable. Then $$\mathbb{E}(f(X,Y) \mid \mathcal{F}) = g(Y)$$ for any bounded measurable function $f$ where $$g(y) := \mathbb{E}(f(X,y)).$$

Denote by $\mathcal{F}^X$ the canonical filtration of the Lévy process $(X_t)_{t \geq 0}$. Using the above lemma and $(1)$, we find from the independence of the increments of $(X_t)_{t \geq 0}$ that for $s \leq t$

$$\mathbb{E}(f(Y_t) \mid \mathcal{F}_s^X) = g(Y_s) \tag{2} $$

for $$g(y) = \mathbb{E}\left[ f \left( e^{\alpha (X_t-X_s)} y + e^{\alpha(X_t-X_s)} \int_s^t e^{-\alpha (X_u-X_s)} \, ds \right) \right] .$$

Since the canonical filtration $\mathcal{F}^Y$ of $(Y_t)_{t \geq 0}$ satisfies $\mathcal{F}_t^Y \subseteq \mathcal{F}_t^X$ for all $t$, the tower property gives

$$\mathbb{E}(f(Y_t) \mid \mathcal{F}_s^Y) = g(Y_s).$$

In particular,

$$\mathbb{E}(f(Y_t) \mid \mathcal{F}_s^Y) = \mathbb{E}(f(Y_t) \mid Y_s),$$

and this shows that $(Y_t)_{t \geq 0}$ is a Markov process.


In order to show that

$$Z_t := \int_0^t e^{-\alpha X_u} \, du$$

is not a Markov process, you actually have to assume that $(X_t)_{t \geq 0}$ is a non-trivial Lévy process. Using a very similar reasoning as above, we find

$$\mathbb{E}(f(Z_t) \mid \mathcal{F}_s^X) = Z_s + c e^{-\alpha X_s} $$

where

$$c := \mathbb{E} \left( \int_s^t e^{-\alpha (X_u-X_s)} \, du \right).$$

Since $e^{-\alpha X_s}$ is $\mathcal{F}_s^Z$-measurable, the tower property gives

$$\mathbb{E}(f(Z_t) \mid \mathcal{F}_s^Z) =Z_s +c e^{-\alpha X_s}. \tag{3}$$

It is intuitively clear that there cannot exist a function $h$ such that

$$e^{-\alpha X_s}= h(Z_s), \tag{4}$$

and therefore $(3)$ shows that $(Z_t)_{t \geq 0}$ is not Markovian. (Note that $(4)$ would mean that for $F(s) := e^{-\alpha X_s}$ you can determine the value $F(s)$ if you know the value of the integral $\int_0^s F(u) \, du$... this is pretty absurd.)

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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