Impossibility of almost extension

Abstract

Let $({\mathbf X},\|\cdot\|_{\mathbf X}), ({\mathbf Y},\|\cdot\|_{\mathbf Y})$ be normed spaces with ${\mathrm{dim}}({\mathbf X})=n$. Bourgain's almost extension theorem asserts that for any ${\varepsilon}>0$, if ${\mathcal{N}}$ is an ${\varepsilon}$-net of the unit sphere of ${\mathbf X}$ and $f:{\mathcal{N}}\to {\mathbf Y}$ is $1$-Lipschitz, then there exists an $O(1)$-Lipschitz $F:{\mathbf X}\to {\mathbf Y}$ such that $\|F(a)-f(a)\|_{\mathbf Y}\lesssim n{\varepsilon}$ for all $a\in \mathcal{N}$. We prove that this is optimal up to lower order factors, i.e., sometimes $\max_{a\in {\mathcal{N}}} \|F(a)-f(a)\|_{\mathbf Y}\gtrsim n^{1-o(1)}{\varepsilon}$ for every $O(1)$-Lipschitz $F:{\mathbf X}\to {\mathbf Y}$. This improves Bourgain's lower bound of $\max_{a\in {\mathcal{N}}} \|F(a)-f(a)\|_{\mathbf Y}\gtrsim n^{c}{\varepsilon}$ for some $0<c<\frac12$. If ${\mathbf X}=\ell_2^n$, then the approximation in the almost extension theorem can be improved to $\max_{a\in {\mathcal{N}}} \|F(a)-f(a)\|_{\mathbf Y}\lesssim \sqrt{n}{\varepsilon}$. We prove that this is sharp, i.e., sometimes $\max_{a\in {\mathcal{N}}} \|F(a)-f(a)\|_{\mathbf Y}\gtrsim \sqrt{n}{\varepsilon}$ for every $O(1)$-Lipschitz $F:\ell_2^n\to {\mathbf Y}$.