A fully nonlinear flow for two-convex hypersurfaces

We consider a one-parameter family of closed, embedded hypersurfaces moving with normal velocity $G_\kappa = \big ( \sum_{i < j} \frac{1}{\lambda_i+\lambda_j-2\kappa} \big )^{-1}$, where $\lambda_1 \leq \hdots \leq \lambda_n$ denote the curvature eigenvalues and $\kappa$ is a nonnegative constant. This defines a fully nonlinear parabolic equation, provided that $\lambda_1+\lambda_2>2\kappa$. In contrast to mean curvature flow, this flow preserves the condition $\lambda_1+\lambda_2>2\kappa$ in a general ambient manifold. Our main goal in this paper is to extend the surgery algorithm of Huisken-Sinestrari to this fully nonlinear flow. This is the first construction of this kind for a fully nonlinear flow. As a corollary, we show that a compact Riemannian manifold satisfying $\overline{R}_{1313}+\overline{R}_{2323} \geq -2\kappa^2$ with non-empty boundary satisfying $\lambda_1+\lambda_2 > 2\kappa$ is diffeomorphic to a $1$-handlebody. The main technical advance is the pointwise curvature derivative estimate. The proof of this estimate requires a new argument, as the existing techniques for mean curvature flow due to Huisken-Sinestrari, Haslhofer-Kleiner, and Brian White cannot be generalized to the fully nonlinear setting. To establish this estimate, we employ an induction-on-scales argument; this relies on a combination of several ingredients, including the almost convexity estimate, the inscribed radius estimate, as well as a regularity result for radial graphs. We expect that this technique will be useful in other situations as well.