Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let us consider a Markov process in discrete time $X = (X_n)_{n\geq 0}$ which values belong to $\mathbb{R}$ and $X_0 = x$. For each $x\in\mathbb{R}$ there is a probability measure $\mathsf P_x$. When one would like to make $X$ be a martingale, the change of measure can help - namely the process $Z_n \in \mathcal F_n$ such that $$ \frac {dQ}{dP}|_{\mathcal F_n} = Z_n $$ and $X$ is a $Q$-martingale. I am a bit confused since for a Markov process there is family of measures $(\mathsf P_x)_{x\in\mathbb{R}}$ rather than a unique one. How one defines a change of measure for a Markov process?

I guess that here $Z_n = f(X_n)$ - but I am looking for the formal definition of change of measure.

share|cite|improve this question

One asks that properties (1), (2) and (3) below hold:

(1) The process $(X_n)$ is a $Q$-martingale, that is, $E_Q(X_{n+1}|F_n)=X_n$.

(2) There exists a $P$-martingale $(Z_n)$ such that $E_Q(U)=E(UZ_n)$ for every $F_n$-measurable random variable $U$.

(3) The process $(Z_n)$ is nonnegative.

Condition (1) means that $E_Q(X_{n+1}U)=E_Q(X_nU)$ for every $F_n$-measurable random variable $U$. Condition (2) implies that $$ E_Q(X_{n+1}U)=E(X_{n+1}UZ_{n+1}),\qquad E_Q(X_nU)=E(X_nUZ_n). $$ The two must coincide for every $F_n$-measurable random variable $U$ hence one asks that $$ E(X_{n+1}Z_{n+1}|F_n)=X_nZ_n. $$ Finally, the goal is to find a nonnegative process $(Z_n)$ such that both $(Z_n)$ and $(X_nZ_n)$ are $(F_n)$-martingales. Such a process is called a martingale density for $(X_n)$, see this paper by Martin Schweizer for an introduction in the continuous time setting.

Let us now try to exhibit a martingale density for $(X_n)$. For every nonnegative $n$, write $E_n$ for the conditional expectation conditionally on $F_n$ and introduce the processes $(Y_n)$ and $(R_n)$ defined by $$ Y_{n+1}=X_{n+1}-E_nX_{n+1},\qquad R_{n+1}=\frac{T_nY_{n+1}}{E_n(X_{n+1}Y_{n+1})}. $$ for an $F_n$-measurable random variable $T_n$, to be specified later on.

Then $E_nY_{n+1}=0$ and both $T_n$ and the denominator of $R_{n+1}$ are $F_n$-measurable hence $E_nR_{n+1}=0$. Furthermore, $E_n(R_{n+1}X_{n+1})=T_n$.

Finally, a solution to conditions (1)-(2) is given by $Z_0=1$ and, for every positive $n$, $$ Z_n=\prod_{k=1}^n(1-R_k)\quad\text{with}\ T_n=E_nX_{n+1}-X_n. $$ An equivalent formula for $R_n$ is $$ R_{n}=\frac{(X_{n}-E_{n-1}X_{n})(E_{n-1}X_{n}-X_{n-1})}{E_{n-1}(X_{n}^2)-(E_{n-1}X_{n})^2}. $$ The denominator of $R_n$ is always nonnegative, and one can probably assume it is almost surely positive and that $R_n$ is well defined. But it is not clear (to me) that $R_n\le1$ almost surely, as it should if $Z_n$ is to be a Radon-Nykodym density... In other words I do not know if condition (3) holds for the process $(Z_n)$ defined above (but there might be a simple argument showing either that (3) always holds or, that it may happen that (3) does not hold and that in these cases there is no solution to your problem).

share|cite|improve this answer
Thank you for the answer. Usually when I am looking for the density process $Z_t$ in a continuous time setting, I solve appropriate SDE to find a Doléans-Dade exponential which certainly admits (2) and (3) and the only question is to admit (1). Is there a similar approach in the discrete time? – Ilya Jul 6 '11 at 14:45

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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