Let $(S,\mathcal{S},P)$ be a probability space. Suppose that $\mathcal{F}$ is a countable set of real measurable functions on $S$ such that $\int fdP=0$ and $\sup_{f\in\mathcal{F}}\|f\|_\infty\leq 1$. Let $X_1,\dots,X_n$ be random variables taking values in $S$ and define $$Z = \sup_{f\in\mathcal{F}}\sum_{i=1}^n f(X_i),\qquad Z_k = \sup_{f\in\mathcal{F}}\sum_{i\ne k}f(X_i),\qquad k=1,\dots,n.$$

We need to prove that $\sum_{k=1}^n(Z - Z_k)\leq Z$. The prof goes as follows. Assume that $\mathcal{F}$ is finite, then there exists $f\in\mathcal{F}$ such that $$(n-1)Z = (n-1)\sum_{i=1}^nf(X_i) = \sum_{k=1}^n\sum_{i\ne k}f(X_i)\leq \sum_{k=1}^nZ_k,$$ which gives the result. Now the paper says that the case when $\mathcal{F}$ is not countable and so the supremum is not achievable, can be established by considering a sequence of finite sets and passing to the limit.

I'm looking for justification for this, why can we claim that there exists a sequence of sets such that $$\lim_{m\to\infty}\sup_{f\in\mathcal{F}_m}\sum_{k=1}^nf(X_k) = \sup_{f\in\mathcal{F}}\sum_{k=1}^nf(X_k)?$$

This relates to another my question Is countable pointwise supremum of bounded functions achievable?, namely the comment given by Brian M. Scott.

This is a part of the proof of the Bousquet inequality appeared in Bousquet, "Concentration Inequalities for Sub-Additive Functions Using the Entropy Method", 2003.


Given $\varepsilon>0$, there exists $f$ such that $Z<\sum_{i=1}^n f(X_i)+\varepsilon$. Hence, $$ \begin{split} (n-1)Z &= (n-1)\left(\sum_{k=1}^nf(X_k)+\varepsilon\right) \\ &= \sum_{k=1}^n\sum_{i\ne k}f(X_i)+(n-1)\varepsilon\leq \sum_{k=1}^nZ_k+(n-1)\varepsilon. \end{split} $$ In other words, there is no need to consider several cases.


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.