Regularity for convex viscosity solutions of Lagrangian mean curvature equation

We establish regularity for convex viscosity solutions of the Lagrangian mean curvature equation

under the assumption that $\psi:B_{1}(0)\to[0,n\pi/2)$ is in $C^{2,\alpha}(B_{1}(0))$ for some $\alpha\in(0,1)$. Here $\lambda_{i}$’s are the eigenvalues of the Hessian matrix $D^{2}u$ and then the phase $\psi$ becomes a potential for the mean curvature of the Lagrangian submanifold $L=(x,Du(x))\subset\mathbb{C}^{n}$. The induced Riemannian metric $g$ can be written as $g=I_{n}+(D^{2}u)^{2}$ or

In [HL82, (2.19)], the mean curvature vector $\vec{H}$ of this Lagrangian submanifold was shown to be

where $\nabla_{g}$ is the gradient operator for the metric $g$ and $J$ is the complex structure, or the $\frac{\pi}{2}$ rotation matrix in $\mathbb{C}^{n}$. Note that by our assumption on $\psi$, $|H|$ is bounded.

When the phase is constant, denoted by $c$, $u$ solves the special Lagrangian equation

or equivalently,

Equation (1.3) originates in the special Lagrangian geometry of Harvey- Lawson [HL82]. The Lagrangian graph $(x,Du(x))\subset\mathbb{C}^{n}$ is called special when the argument of the complex number $(1+i\lambda_{1})...(1+i\lambda_{n})$ or the phase $\psi$ is constant, and it is special if and only if $(x,Du(x))$ is a (volume minimizing) minimal surface in $(\mathbb{C}^{n},dx^{2}+dy^{2})$ [HL82].

A dual form of (1.3) is the Monge-Ampère equation

This is the potential equation for special Lagrangian submanifolds in $(\mathbb{C}^{n},dxdy)$ as introduced in [Hit97]. The gradient graph $(x,Du(x))$ is volume maximizing in this pseudo-Euclidean space, as shown by Warren [War10]. In the 1980s, Mealy [Mea89] showed that an equivalent algebraic form of the above equation is the potential equation for his volume maximizing special Lagrangian submanifolds in $(\mathbb{C}^{n},dx^{2}-dy^{2})$.

The regularity of solutions is a fundamental problem for these geometrically and analytically significant equations. Our main results are the following:

We use the above regularity to prove the following interior estimate.

We state the following application of our result.

The convexity of the special Lagrangian equation plays a dominant role in its regularity theory. The arctangent operator is concave if $u$ is convex, or if the Hessian has a lower bound $\lambda\geq 0$. Bao and Chen [BC03] showed regularity for convex $W^{2,p}$ strong solutions. For smooth convex solutions of (1.3), Hessian estimates have been obtained by Chen-Warren-Yuan [CWY09]. Recently in [CSY19], Chen-Shankar-Yuan confirmed that convex viscosity solutions of (1.3) are smooth. The semiconvex singular solutions constructed by Wang-Yuan [WY13] show that the convexity condition is necessary. Similarly, if in (1.3) we have critical phase $|c|=(n-2)\pi/2$ or supercritical phase $|c|>(n-2)\pi/2$, then $F(D^{2}u)$ has convex level sets, but it was shown by Yuan [Yua06] that this fails for subcritical phases $|c|<(n-2)\pi/2$. For critical and supercritical phases, Hessian estimates for (1.3) have been obtained by Warren-Yuan [WY09, WY10] and Wang-Yuan [WY14], see also [Li19]. For subcritical phases, $C^{1,\alpha}$ solutions of (1.3) were constructed by Nadirashvili-Vlăduţ [NV10] and Wang-Yuan [WY13]. Hessian lower bounds also play a role in the dual equation (1.4). The Monge-Ampére type equation $\det D^{2}u=h(x)$ has Pogorelov-type singular $C^{1-2/n}$ convex viscosity solutions whose graphs contain a line. However, under the necessary strict convexity assumption $\lambda>0$, interior regularity was obtained by Pogorelov [Pog78] for a smooth enough right hand side $h(x)$, by Urbas [Urb88] if $h(x)$ is Lipschitz and $u(x)\in C^{(1-2/n)^{+}}$, and by Caffarelli [Caf90] if $h(x)$ is merely Hölder.

The convexity of $u$ creates its own challenges in proving regularity, since it is unstable under smooth approximations of the boundary value. To use the method of a priori estimates, one would solve the Dirichlet problem for a modified concave equation $\tilde{F}(D^{2}u)=\psi(x)$, where $\tilde{f}(\lambda)=\arctan\lambda$ for $\lambda\geq 0$ and $=\lambda$ for $\lambda<0$. If we approximate the boundary data and find the smooth solution of this problem, it may no longer be convex, and therefore lacks a connection to the original problem. Although a priori estimates are available for (1.1) and (1.3), estimates for $\tilde{F}$ are unknown.

In showing regularity for convex solutions of (1.3), the authors of [CSY19] managed to avoid a priori estimates altogether. The basic idea was to change variables using the Lewy-Yuan rotation of the gradient graph, $\bar{x}+iD\bar{u}(\bar{x})=e^{-i\pi/4}(x+iDu(x))$, such that the Hessian bounds decreased from $0\leq\lambda\leq+\infty$ to $-1\leq\bar{\lambda}\leq 1$. Since minimal graphical tangent cones with such bounds are flat, they were able to deduce regularity in new coordinates using ideas from [Yua01]. One problem, however, is that defining such a rotation is unclear if convex $u$ is not $C^{1}$. To adapt to the lower regularity setting, it was shown in [CSY19] that in the smooth case, $\bar{u}(\bar{x})$ can be constructed using the Legendre transform, which is still well defined in the Lipschitz case.

Extending the constant phase (1.3) results to variable phases (1.1) is subtle, and not always possible. Interior regularity for solutions of (1.1) with a $C^{1,1}$ supercritical and critical phase were recently obtained in [Bha21, Bha22, BMS22], but $C^{1,\alpha}$ singular solutions are known if $\psi$ has Hölder regularity [Bha21, Remark 1.3]. In fact, these singular solutions are convex, so Theorem 1.1 is not valid unless $\psi(x)$ is sufficiently smooth. Moreover, in our companion paper [BS20], we exhibit convex $C^{1,\alpha}$ singular solutions which solve an equation with $\psi=f(x,Du)$ depending also on the gradient. In such cases, $f$ is smooth in both arguments. Convex singular solutions do not exist for analogous uniformly elliptic PDEs $F(D^{2}u)=f(x,Du)$. The non-uniform ellipticity of the arctangent operator in (1.3) makes the PDE highly sensitive to the structure of the variable phase. Nevertheless, the results of this paper show that no conditions on $\psi(x)$ are needed for regularity of convex viscosity solutions, apart from some smoothness such as $C^{2,\alpha}$.

In proving Theorem 1.1, the treatment of the variable phase is delicate, and the technique does not always work. The key step to deduce $C^{1,1}$ regularity, or $\lambda_{max}(D^{2}u)<\infty$, is to show that in rotated coordinates, $\bar{\lambda}_{max}$ never saturates its upper bound of $1$. For constant phases, this is done using the strong maximum principle for $\Delta_{\bar{g}}$-subharmonic $\bar{\lambda}_{max}$. For variable phases, subharmonicity only holds up to extra terms, such as $D^{2}\bar{\psi}(e_{max},e_{max})$. If $\psi$ is convex, then subharmonicity is restored, but Theorem 1.1 does not assume any conditions on $\psi$ other than $C^{2,\alpha}$ smoothness. To handle these terms, we carefully account for the coordinate change, and relate $D^{2}\bar{\psi}$ to $D^{2}\psi$, for example. It turns out that the resulting expression vanishes when $\bar{\lambda}_{max}$ saturates its upper bound, restoring the maximum principle. To understand how this could happen, observe that for a convex singular solution, the forward map $\bar{x}=x\cos\alpha+Du(x)\sin\alpha$ has large Jacobian matrix in the $\lambda_{max}$ directions. It follows that the inverse map has small Jacobian in those directions, so the rotated phase $\bar{\psi}(\bar{x})=\psi\circ x(\bar{x})$ will inherit this flatness.

This paper is divided into the following sections. In section 2, we formulate the Lewy-Yuan rotation for the Lipschitz potential $u(x)$. In section 3, we establish regularity of the rotated potential $\bar{u}(\bar{x})$. In section 4, we deduce regularity of the original potential $u(x)$, thereby proving Theorem 1.1. In section 5, we prove the Hessian estimate (1.5).

Acknowledgments. The authors are grateful to Yu Yuan for his guidance, support, and helpful discussions. The authors thank D. H. Phong for helpful comments and suggestions. AB acknowledges the support of the AMS-Simons Travel Grant. RS was partially supported by the NSF Graduate Research Fellowship Program under grant No. DGE-1762114. The authors thank the anonymous referee for the referee’s thorough feedback.

We formulate Lewy-Yuan rotation for the convex function $u$ solving (1.1), using the idea introduced in [CSY19]; we refer to sections 2.1 and 2.2 in that paper for various details in this section. If $u(x)$ is smooth on $\Omega$, then the gradient graph $z=(x,Du(x))$ is a Lagrangian submanifold of $\mathbb{C}^{n}$, and if $u$ is convex, then the downward rotation $\bar{z}=e^{-i\pi/4}z$ yields another Lagrangian submanifold $(\bar{x},D\bar{u}(\bar{x}))$ with smooth potential $\bar{u}(\bar{x})$ on $\bar{x}(\Omega)$. Because the canonical angles decrease by an angle of $\pi/4$

it follows that the rotation sends solutions of (1.1) to solutions $\bar{u}(\bar{x})$ of another Lagrangian mean curvature equation, which now is uniformly elliptic by the Hessian bounds

This makes regularity theory tractable in new coordinates using the ideas developed in [Yua01].

If we do not assume that $u$ is in $C^{1}$, then the subdifferential $(x,\partial u(x))$, i.e., the slopes of tangent planes touching $u$ from below at $x$, is not a graph over $\Omega$ (for example, if $u(x)=|x|$, then $\partial u(0)=\bar{B}_{1}(0)$). Nevertheless, if we rotate the subdifferential $(x,\partial u(x))$ downwards by $\pi/4$, then we still obtain a gradient graph $(\bar{x},D\bar{u}(\bar{x}))$, where $\bar{u}\in C^{1,1}$ and is given explicitly by

where $c=\cos\pi/4$, $s=\sin\pi/4$, and

is the Legendre transform of convex $f(x)$; see [CSY19, Proposition 2.1]. The image domain $\bar{\Omega}=\bar{x}(\Omega)=\partial\tilde{u}(\Omega)$ is open and connected by [CSY19, Lemma 2.1], which comes from the fact that

is strictly convex.

The transform (2.2) is order preserving, $f\leq g\to\bar{f}\leq\bar{g}$, hence, it preserves uniform limits. It follows that $\bar{u}(\bar{x})$ satisfies the same Hessian bounds (2.1) as in the smooth case. We could then use interpolation to see that $D\bar{u}_{n}\to D\bar{u}$ locally uniformly provided $u_{n}\to u$ uniformly, but this also follows from strict convexity, as in [CSY19, Proposition 2.3], i.e. if $\bar{x}\in\partial\tilde{u}(a)\cap\partial\tilde{v}(b)$ for $u,v$ convex, then

The smallness of $|D\bar{u}-D\bar{v}|$ would then follow from reverse rotation $a=-s\bar{x}+cD\bar{u}$.

Following [CSY19, Propositions 2.2,2.3], we show that viscosity solutions are preserved under this rotation.

In the remainder, we abbreviate the rotated phase by

In particular, since $\bar{u}\in C^{1,1}$, we see that $\bar{\psi}$ is Lipschitz, so far.

We first show that $\bar{u}\in VMO(B_{1/2})$ followed by $\bar{u}\in C^{2,\alpha}(B_{1/2})$. We recall the notion of VMO functions.

By translation invariance of the VMO seminorms, $D^{2}\bar{u}(\bar{x}+h)$ will also be in $VMO(B_{1/2})$ if $h$ is small. This means we can conclude the following but for the sake of completeness, we provide a proof similar to Proposition 3.1.

Note that the $VMO(B_{1/2})$ seminorms of $K(D^{2}\bar{u},D^{2}\bar{u}^{h})$ are independent of $h$ if $h$ is small. The point of $K$ is we need to take a difference quotient of (2.5).

Now from Schauder theory, the rotated potential $\bar{u}(\bar{x})$ is as regular as the equation allows.

Indeed, since $\bar{u}\in C^{1,1}$, we have $D\bar{u}\in C^{0,1}$ and $\bar{\psi}\in C^{0,1}$, and Proposition 3.2 applies, so $D\bar{u}\in C^{1,\alpha}$. If we also know $\psi\in C^{2,\alpha}$, then this means $\bar{\psi}\in C^{1,\alpha}$. Recalling equation (3.1) for the difference quotient of $\bar{u}$, it follows from Schauder estimates that $D\bar{u}\in C^{2,\alpha}$, so $\bar{\psi}\in C^{2,\alpha}$. Taking two difference quotients in a similar way, we deduce $D^{2}\bar{u}\in C^{2,\alpha}$.

The final step is to show that $\bar{\lambda}_{max}<1$ on $\bar{x}(B_{1})$. Indeed, once we prove this, we can compare with quadratics to convert to original variables. Let $\bar{Q}(\bar{x})$ touch $\bar{u}$ from above near $\bar{x}$ and suppose, after lowering, that $D^{2}\bar{Q}<I_{n}$. By the order preservation of rotations, we see that $Q(x)$ touches $u$ from above near $x$. Moreover, $D^{2}Q<\infty$ by the transformation law

Thus $\lambda_{max}<\infty$ on $B_{1}$, i.e. $u\in C^{1,1}(B_{1})$. This means $\bar{x}(x)=cx+sDu(x)$ is Lipschitz, so the above formula with $Q$ replaced by $u$ implies $u\in C^{2,1}(B_{1})$. This means $\bar{x}\in C^{1,1}$, and as before, we deduce $u\in C^{3,1}(B_{1})$. Since $D^{2}\bar{u}\in C^{2,\alpha}(B_{1})$ and $\bar{x}\in C^{2,1}$, we conclude that $u\in C^{4,\alpha}(B_{1})$. We now establish the desired inequality.