A quantitative Sobolev inequality in BV (Q2499246)

The Sobolev inequality in \(W^{1,1}(\mathbb R^n)\) (or \(\text{BV}(\mathbb R^n)\)) states that \(n \omega_n^{1/n} \| u\| _{L^{n'}} \leq \| \nabla u\| _{L^1}\). It holds with equality if and only if \(u(x) = \lambda \chi_B(x)\) for some ball \(B\) (\(\chi_B\) denotes the characteristic function of \(B\)). The idea of this paper is the following: if we know that \(u\) is not close to a function of the type \(\lambda \chi_B\), then it should be possible to improve the inequality. More concretely, the following measure of discrepancy is introduced: \[ d_{n'}(u) = \inf_{\lambda, B} \frac{\| u - \lambda \chi_B \| _{L^{n'}}}{\| u\| _{L^{n'}}} \] if \(u\not \equiv 0\), and \(d_{n'}(0)=0\). The main theorem reads: Theorem~1.1 Let \(n\geq 2\). Then there exists a constant \(c(n)\) so that \[ n \omega_n^{1/n} \big( 1 + c(n) d_{n'}(u)^{4n' + 1}\big)\| u\| _{L^{n'}} \leq \| D u\| , \] for every \(u\in \text{BV}(\mathbb R^n)\). It is not known whether the power \(4n' + 1\) in the theorem is optimal. Theorem~1.2 gives a slightly stronger version of this result using the Lorentz space scale.