Steveot comments on [missing post]

[Steveot]

Thanks, I finally got it. What I just now fully understood is that the final inequality holds with high ${\pi }_0^n$ probability (i.e., as you say, ${\pi }_0$ is the data), while the learning bound or loss reduction is given for $\pi$.