MathOverflow is a question and answer site for professional mathematicians. 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 $X, Y$ be two $L^1$ random variables on the probablity space $(\Omega, \mathcal{F}, P)$. Let $\mathcal{G} \subset \mathcal{F}$ be a sub-$\sigma$-algebra. Consider the conditional expectation operator $E(\cdot|\mathcal{G}) \colon L^1(\mathcal{F}) \to L^1(\mathcal{G})$. When is $E(X \ast Y|\mathcal{G}) = E(X|\mathcal{G}) \ast E(Y|\mathcal{G})$? Here $\ast$ is the convolution product on $L^1$ (which makes $(L^1, \ast)$ a Banach algebra, so I'm asking when is $E(\cdot|\mathcal{G})$ an algebra homomorphism?).

share|cite|improve this question
I'm not quite sure how you're defining convolution - normally one can only define convolution of (equivalence classes of) L^1-functions which are defined on some semigroup that carries a suitable measure. Could you please write out your definition? – Yemon Choi Dec 10 '09 at 3:13
@davidk01 I think that this is indeed the usual notion of conditional expectation wrt a sub $\sigma$-algebra; so I don't think that bit needs changing - or at least, it's not so important – Yemon Choi Dec 10 '09 at 6:33
@Yemon Choi: You are right. I completely ignored the question of what kind of structure Omega has. – davidk01 Dec 10 '09 at 8:12
-1 since it is still (at time of writing) not clear what semigroup structure the author is assuming on the measure space, and hence I still don't follows what the "convolution product is supposed to be... – Yemon Choi Dec 11 '09 at 5:34

As long as $\mathcal{G}$ is invariant under whatever operation you use in the convolution ("+", say), $\mathcal{G}$-measurable functions will convolve to $\mathcal G$-measurable functions and the equalities of integrals that define the conditional expectation will be automatic: for any $H \in \mathcal G$ we have $$ \int_{\Omega} E(X|\mathcal{G}) * E(Y|\mathcal{G}) (t) \chi_H(t) dP(t) = $$ $$ = \int_{\Omega\times\Omega} E(X|\mathcal{G})(x)E(Y|\mathcal{G})(y) \chi_{H}(x+y) dP(x)dP(y) = $$ $$ = \int_{\Omega} E(X|\mathcal{G})(x) \int_{\Omega} E(Y|\mathcal{G})(y) \chi_{H}(x+y) dP(y)dP(x) = $$ $$ = \int_{\Omega} E(X|\mathcal{G})(x) \int_{\Omega} Y(y) \chi_{H}(x+y) dP(y)dP(x) = $$ $$ = \int_{\Omega} Y(y) \int_{\Omega} E(X|\mathcal{G})(x) \chi_{H}(x+y) dP(x)dP(y) = $$ $$ = \int_{\Omega} Y(y) \int_{\Omega} X(x) \chi_{H}(x+y) dP(x)dP(y) = $$ $$ = \int_{\Omega} X*Y(t) \chi_H(t) dP(t) = \int_{\Omega} E(X*Y|\mathcal{G})(t) \chi_H(t) dP(t) $$ If $\mathcal{G}$ is not invariant under the operation, though, I see no reason for the convolution to be $\mathcal{G}$-measurable. Is an example for that what you're requesting?

share|cite|improve this answer

Assuming $\Omega$ has the structure for defining convolutions I don't think it is ever an algebra homomorphism. Take $X$ to be supported on $\mathcal{G}^c$, i.e. take some set of non-zero measure in $\mathcal{G}^c$ and let $X$ be a function whose support lies in that set, then $E(X\ast Y|\mathcal{G})\neq 0$ but $E(X|\mathcal{G})=0$ so $E(X|\mathcal{G})\ast E(Y|\mathcal{G})=0$.

Edit: Scratch what I said. I was confusing sub-$\sigma$-algebra with sub-algebra of random variables and even in the finite case my statement is completely incorrect. In almost every instance $E(X|\mathcal{G})$ will not be zero as Jonas points out in the comments.

share|cite|improve this answer
What do you mean when you say that X is supported on the complement of a sigma algebra of subsets of \Omega? – Jonas Meyer Dec 10 '09 at 10:48
I don't understand. There may be an element of F\G on which X is supported, but that doesn't imply that E(X|G) = 0. Of course the support of X lies in $\Omega\in\mathcal{G}$. – Jonas Meyer Dec 11 '09 at 2:26
As I mentioned, G will always contain one particular set which contains the support of X, namely $\Omega$, and G will typically contain many more sets on which X is not identically zero, even if the support itself is an element of F\G. Consider for example Omega = the unit circle, F = the set of Lebesgue measurable sets, G = the set of sets that are countable or have countable complement, X = a function that is identically 1 on a semicircle and 0 otherwise. Then the support of X is not an element of G, but E(X|G) is not zero because its integral over sets with countable complement is > 0. – Jonas Meyer Dec 11 '09 at 3:11
I'm glad that's cleared up, but I think it can be confusing when you delete your comments that appeared as part of an exchange. – Jonas Meyer Dec 11 '09 at 17:34
Ya, you have a point but hopefully the edit makes it clear. – davidk01 Dec 12 '09 at 0:51

I hadn't really thought through the question of which structure I want $\Omega$ to have specifically. I guess what is needed is some convolution semigroup structure which accurately resembles the convolution product (for $\Omega = \mathbb{R}^n$ and the Lebesgue measure) and $P$ being some kind of probability Haar measure which makes $(L^1(\Omega), \ast)$ a Banach algebra. But I consider Thorny's post a sufficient answer of my question.. Thanks for your help.

share|cite|improve this answer

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.