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

Suppose $f,g$ are continuous functions from $\mathbb R$ to $\mathbb R$, with the property that $$f(x)+f(y)=g(x+y)$$ for all $x,y$. Taking $x=y=z/2$ implies that $g(x)=2f(x/2)$ so that the above condition becomes $$f(x)+f(y)=2f((x+y)/2).$$ This is known as Jensen's functional equation, and it implies that $f$ is linear.

There's also a generalization of Jensen's equation (I've seen it in work of Rassias, but it could be earlier): if $|f(x)+f(y)-2f((x+y)/2)|\leq\epsilon$ (and assuming WLOG that $f(0)=0$), then there is a linear function $L$ such that $|f(x)-L(x)\|\leq \epsilon$.

What I am interested in a generalization of all this: Suppose there are independent random variables $X,Y$ such that $$E[(f(X)+f(Y)-g(X+Y))^2]\leq\epsilon.$$ Is it possible to say anything about $f$ being (appropriately) approximately linear?

share|cite|improve this question
I've removed the soft-question tag, as this doesn't really fit the bill. – j.c. Jul 8 '10 at 22:23
Is this true for some class of $X$ and $Y$ or just two particular random variables? If it's two particular random variables then the condition certain doesn't imply anything without further knowledge of the distributions: X and Y could just be particular constants, and the inequality would tell you very little about f globally. – Noah Stein Jul 9 '10 at 16:50

Comment : I think you could give more precision in the formulation of your question because you already have most of the answer.

In particular:

1-Is the generalization of Jensen equation you give valid on Hilbert Space ? (seems to be true, see

2-What assumption do you make on the random variables? do they have the same distribution ? are their distribution "equivakent" (i.e mutually absolutly continuous)

Answer: If (the answer to my question 1 is yes and if) $X$ and $Y$ have the same distribution $P$, because $L_2(P)$ is a hilbert space then you are done. Here, approximate linearity will mean that there exists a linear form $L$ such that $E_P[(f(X)-L(X))^2]<\epsilon$.

If $X$ and $Y$ have different distribution (say $P_X$ and $P_Y$) I see an easy case: when $E_{P_Y}[( dP_X/dP_Y )^2] ) < c$ ($\chi^2$ divergence between distributions bounded). Indeed, in this case you can work in $L_2(P_Y)$ using cauchy swartz inequality. Otherwise the question needs to be clarifyed (i.e what do you mean by "approximatly linear"? : in what banach space do you choose to work ? ).

Hope this helps !

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.