VC theory provides an answer to Problem 1 specified below. I am wondering what is known about a similar issue, Problem 2.

$$ ~ $$

**Problem 1**

Let $X$ be a set, let $\mathcal{D}$ be a distribution over $X$, let $F$ be a set of functions from $X$ to $\{0,1\}$. Let $\varepsilon,\delta \in (0,1)$, and let $x_1,x_2,\dots,x_m$ be a set of i.i.d. samples from $\mathcal{D}$. Find the minimal $m$ such that with probability at least $1-\delta$ over the samples,

$$ \sup_{f \in F}\Big|\mathbb{E}_{x \sim \mathcal{D}}[f(x)] - \frac{1}{m}\sum_{i=1}^m f(x_i)\Big| \leq \varepsilon. $$

It is well known that the answer is $m = \Theta(\mathsf{VC}(F))$, where $\mathsf{VC}(F)$ is the VC dimension of $F$.

$$ ~ $$

**Problem 2**

Under the same conditions, find the minimal $m$ such that with probability at least $1-\delta$ over the samples,

$$ \Big|\sup_{f \in F}\mathbb{E}_{x \sim \mathcal{D}}[f(x)] - \sup_{f \in F}\frac{1}{m}\sum_{i=1}^m f(x_i)\Big| \leq \varepsilon. $$

Clearly, taking $m = \Theta(\mathsf{VC}(F))$ as before is sufficient to achieve this, but in some cases $m$ can be much smaller. I am wondering if there are any better bounds on $m$ that are known? Can someone refer me to any papers on this issue?

I am aware of some related papers in the context of passive testing in computer science (e.g. here and here), but I am wondering if there is a more direct or comprehensive treatment from a statistics perspective?