Non-improvability of sharp endpoint estimates

For an integer $n$ and the parameter $\gamma\in(0,n)$, the Riesz potential $I_\gamma$ is known to take boundedly $L^1(\mathbb{R}^n)$ into $L^{\frac{n}{n-\gamma},\infty}(\mathbb{R}^n)$, and also that the target space is the smallest possible among all rearrangement-invariant Banach function spaces. We study the natural question whether the target space can be improved when the domain space is replaced with a (smaller) Lorentz space $L^{1,q}(\mathbb{R}^n)$ with $q\in(0,1)$. The classical methods cannot be used because the spaces $L^{1,q}(\mathbb{R}^n)$ are not equivalently normable. We develop two new abstract methods, establishing rather general results, a particular consequence of each (albeit achieved through completely different means) being the negative answer to this question. The methods are based on special functional properties of endpoint spaces. The results can be applied to a wide field of operators satisfying certain minimal requirements.