Concentration inequalities

A: The right choice achieves a smaller exponent by combining variance and range, which matters when deviations aren't extreme; Hoeffding drops variance and pays a penalty, Markov ignores almost all structure and blows up the tail, and Chebyshev uses variance but decays too slowly to meet a guardband set this tight.

A: This works because bounded MGF growth is the defining property you're about to assume; a Q–Q plot is visual and fragile in the tails, Chebyshev sidesteps the question rather than verifying it, and kurtosis can look benign while tails still violate sub-Gaussian decay.

A: The goal is a tail bound under dependence with bounded steps, which Azuma-Hoeffding delivers; Chernoff relies on independence, Bernstein needs variance control that martingales may not give, and LLN doesn't quantify finite-sample tails.

A: This choice targets stability under sample replacement, which is exactly the bounded-difference condition; Hoeffding ignores the function-level sensitivity, Bennett needs variance and skew detail you didn't quantify, and a union bound loses the concentration structure.