Duality and stability for functional inequalities (Q2012076)

Suppose we have a functional inequality on a real normed linear space \(X\), \[ \mathcal{F}(x) \leq \mathcal{E}(x), \,x \in X. \] Such an inequality is sharp if, for any \(\lambda < 1\), there is an \(x \in X\), \(\mathcal{E}(x) > \lambda \mathcal{F}(x)\). If the inequality is sharp, \(X_0 = \{ x \in X: \mathcal{E}(x) = \mathcal{F}(x),\, \mathcal{F}(x) < \infty \}\) is the set of optimizers of the inequality. If \(X_0 \neq \emptyset\), the inequality is optimal. An optimal inequality is sharp, but the converse is not true. The question of stability concerns whether, for an optimal inequality, if a sequence \(\{x_n\} \in X\) approaches equality, does it approach an optimizer and at what rate, i.e., if \[ \lim_{n \to \infty} (\mathcal{F}(x_n) - \mathcal{E}(x_n)) = 0, \] do we have \[ \lim_{n \to \infty} d(x_n, X_0) = 0, \] where \(d\) is some metric on \(X\) which need not be the norm on \(X\)? The measure of this is a rate function, which is a strictly convex function, \(\Phi:[0, \infty) \mapsto [0, \infty)\), \(\Phi(0) = 0\). An optimal inequality \(\mathcal{E} \leq \mathcal{F}\) is \((d, \Phi)\) stable if \[ \mathcal{F}(x) - \mathcal{E}(x) \geq \Phi(d(x, X_0)), \] and the inequality it \(\kappa\)-stable if this occurs with \(d(x,y) = || x - y ||_X\), \(\Phi(t) = \kappa t^2\). When an optimal inequality is \((d, \Phi)\) stable, the rate of convergence of \(\mathcal{F}(x_n) - \mathcal{E}(x_n) \to 0\) controls the rate of convergence of \(d(x_n, X_0) \to 0\). For example, Sobolev's inequality in its optimal form says that, if \(1/p = 1/2 - \alpha/n\), \[ \left( \int_{\mathbb R^n} |f(x)|^p \, dx)^{1/p} \right) \leq \mathcal{S}_{n, \alpha} \int_{\mathbb R^n} | (\Delta)^{-\alpha/2} f(\eta)|^2 d^n\eta, \] where \(\mathcal{S}_{n, \alpha}\) is the left hand side of the inequality divided by the right hand side for the particular function \(h(\eta) = (1 + |\eta|^2)^{(n - 2 \alpha)/2}\), which therefore must be an optimizer. It is the case that all optimizers of this inequality are multiples of dilates and translates of this function. This fits in the above setting \(X\) to be the completion of the compactly supported infinitely differentiable functions in the norm \[ || f ||_X = || (\Delta)^{-\alpha/2} f ||_2, \quad \mathcal{E}(f) = || f ||_p^2, \quad \mathcal{F}(x) = \mathcal{S}_{n, \alpha} || f ||_X^2. \] It was recently proved by \textit{S.-B. Chen} et al. [Indiana Univ. Math. J. 62, No. 4, 1381--1397 (2013; Zbl 1296.46032)] that, for all \(\alpha \in (0, n/2)\), there is a constant \(\kappa_{BE}>0\), depending on \(n, \alpha\), such that \[ \mathcal{F}(f) - \mathcal{E}(f) \geq \kappa_{BE} \inf_{z \in X_0} || f - z ||_X^2, \] or, in the language of this paper, the Sobolev inequality is \(\kappa_{BE}\)-stable. The object of this paper is to prove stability theorems for functional inequalities using duality and flow arguments. The duality is via the Legendre transform of an inequality. Let \(Y\) be the dual space of the space of measurable \(L_p\) functions considered in the Sobolev inequality, and \[ \mathcal{E}^*(g) = \sup_{f \in X} (\langle f, g \rangle - \mathcal{E}(f) ), \] where \(\langle f, g \rangle\) is an appropriately normalized form of the duality integral between \(Y\) and the \(L_p\) space, and make a similar definition for \(\mathcal{F}^*\). Duality is used to show that the inequality \(\mathcal{E} \leq \mathcal{F}\) implies that \(\mathcal{F}^* \leq \mathcal{E}^*\). The Legendre transform of the optimal Sobolev inequality is the optimal Hardy-Littlewood-Sobolev inequality first found by \textit{E. H. Lieb} [Ann. Math. (2) 118, 349--374 (1983; Zbl 0527.42011)]. The author's main result is that when \(\mathcal{E} \leq \mathcal{F}\) is not only optimal, but is also stable, then under suitable quantitative convexity conditions on \(\mathcal{E}\), there is an optimality and stability result for the dual inequality \(\mathcal{F}^* \leq \mathcal{E}^*\). The convexity condition is met for the Sobolev inequality and, as a result, from the result of Chen et al. [loc. cit.], the author obtains the new result that the Hardy-Littlewood-Sobolev inequality is \(\kappa_{BE}^*\)-stable for some \(\kappa_{BE}^* > 0\) which can be explicitly computed in terms of the constant \(\kappa_{BE}\) above. A direct approach to this result looks difficult, but becomes unnecessary to attempt. Finally, the author introduces the concept of monotone flows for a functional inequality. He uses such flows with duality methods to prove stability bounds for the Sobolev and the Hardy-Littlewood-Sobolev inequalities with computable constants.