Let $\Delta$ be an open triangle in $\mathbb{R}^2$ and $u\in C^0(\overline{\Delta})\cap C^\infty(\Delta)$ be the convex function satisfying $$ \det D^2u=1,\quad u|_{\partial\Delta}=0. $$ Classical results on Monge-Ampère equations imply that there exists a unique such $u$. I need informations on the boundary regularity of $u$, but could not find any literature on regularity of such Monge-Ampère equations near a boundary line segment.

Question 1. What can be said about the boundary regularity of $u$? In particular, is $\|\nabla u(x)\|$ bounded as $x\in\Delta$ approaches a point on $\partial \Delta$ which is not a vertex?


  1. The second question is equivalent to asking whether $u$ has infinite slope at a non-vertex point $p\in\partial\Delta$, or more precisely, whether the limit $$ \lim_{t\rightarrow 0^+}\frac{u((1-t)p+tq)}{t}\in [-\infty,0) $$ equals $-\infty$ for some $q\in\Delta$ (which implies the same limit for every $q\in \Delta$).

  2. It can be shown $u$ has finite slope at a vertex (namely, the above limit is finite if $p\in\partial\Delta$ is a vertex): assuming $p=0$ w.l.o.g., by Comparison Principle, $u$ is minorized by a convex function $v$ of the form $$ v(x)=c|x|^\alpha-L(x), $$ where the constant $\alpha\in (1,2)$ is arbitrary, $c>0$ is a constant to make $\det D^2v=c^2\alpha^2(\alpha-1)|x|^{2\alpha-4}$ bigger than $1$ on $\Delta$, and $L$ is a linear function on $\mathbb{R}^2$ to make $v|_{\partial\Delta}\leq 0$; whereas $v$ has value $0$ and finite slope at $x=0$.

  3. Attempting to apply Comparison Principle near non-vertex boundary points, one is lead to the following problem:

Question 2. Let $D=\{x\in\mathbb{R}^2\mid |x|<1,\, x_1>0\}$ be the half-disk. Is there a nonnegative convex function $u\in C^0(\overline{D})$ which vanishes on the boundary line segment $\{x_1=0,\, |x_2|\leq 1\}$, such that $\det D^2u\geq \lambda>0$?

I tried to construct such a $u$ of the form $u(x_1,x_2)=f(x_1)g(x_2)$ (with $f$ and $g$ nonnegative, convex and $C^2$) but was lead to the conclusion that such $u$ can never satisfy the hypotheses.


1 Answer 1


The answer to both questions is no.

For Question $2$: After subtracting a function of the form $cx_1$ we may assume that $u \geq 0$ and $u(t,0) = o(t)$. (Take $c$ to be the slope of the tangent line to $u(t,0)$ at $t = 0$; then $u - cx_1 \geq 0$ on the $x_1$ axis and vanishes on the $x_2$ axis, so by convexity is nonnegative in $D$). Then $\{u < \epsilon\}$ contains the triangle $T_{\epsilon}$ with vertices $\pm e_2$ and $4l_{\epsilon}e_1$ where $\epsilon^{-1}l_{\epsilon} \rightarrow \infty$ as $\epsilon \rightarrow 0$. It is straightforward to check that the paraboloid $P_{\epsilon} = 2\epsilon(l_{\epsilon}^{-2}(x_1 - l_{\epsilon})^2 + 4x_2^2)$ satisfies $P_{\epsilon} > \epsilon \geq u$ on $\partial T_{\epsilon}$, vanishes at $(l_{\epsilon},0)$, and $\det D^2P_{\epsilon} \rightarrow 0$ as $\epsilon \rightarrow 0$. For $\epsilon$ small this violates the comparison principle.

Question $1$ can be reduced to Question $2$ as follows: if there exists a sequence of points approaching the flat boundary where the graph of $u$ has supporting planes of uniformly bounded slope, then in the limit we get a supporting plane to the graph of $u$ with bounded slope that vanishes on one of the edges. Then subtract the plane, translate, rotate and rescale.

However, near the flat boundary I suspect there is some expansion of $u$ in powers of "singular building blocks," maybe of the form $x_1(-\log x_1)^{1/2}$ and $x_2$. (This is motivated by the observation that $w := x_1(-\log x_1)^{1/2}(x_2^2-1)$ satisfies $0 < \lambda < \det D^2w < \Lambda$ in the domain $\{|x_2| < c(-\log (x_1))^{-1/2}\}$ near the origin, which has boundary that separates from the $x_2$ axis slower than any polynomial).

  • $\begingroup$ That's really a great argument that I should have known earlier! By the way, I need to further consider the equation with right-hand side $|x|^\alpha$ instead of $1$ (still with zero boundary value, on a triangle $\Delta$ with $0\in\mathbb{R}^2$ as a vertex) and determine whether the derivative of the solution $u$ at $0$ (along a vector pointing inside $\Delta$) is finite. I can show that it is finite if $\alpha>-2$ and infinite if $\alpha<-2$, but have trouble with the $\alpha=-2$ case. Could you maybe shed some light on this? $\endgroup$
    – Xin Nie
    Commented Nov 9, 2018 at 23:04
  • 1
    $\begingroup$ The gradient will blow up at the vertex in the case $\alpha = -2$. To see this, consider the barrier $w = -y(\log(1/y))^{1/2} + x^2(\log(1/y))^{1/2}/y$ in the quarter space $\{y > |x|\}$. A computation gives $\det D^2w = y^{-2}(1-x^2/y^2 + O(1/|\log y|))$, so $w$ is a super solution with gradient that blows up logarithmically at $0$. $\endgroup$ Commented Nov 15, 2018 at 17:57
  • $\begingroup$ My computations show that $\Omega:=\{\det D^2w>0\}$ is the raindrop-shaped region $\left\{0<y<e^{1/2},\ x^2<\frac{y^2\left(\frac{1}{2}-\log y\right)}{\frac{3}{2}-\log y}\right\}\subset\{y>|x|\}$. Do you mean using $\tilde{w}=w+cy$ (with $c>0$ large enough to ensure $\tilde{w}\geq 0$ on $\partial\Omega$) as a barrier on $\Omega$? $\endgroup$
    – Xin Nie
    Commented Nov 16, 2018 at 11:50
  • 1
    $\begingroup$ That is right, thanks for doing the computation carefully- in any corner of angle $< \pi$ we can take an affine rescaling of $w$, multiply by a small constant, and add a large multiple of $y$ to obtain a super-solution that is nonnegative on the boundary of a region where $u \leq 0$ and thus show blowup of the gradient at the vertex. $\endgroup$ Commented Nov 16, 2018 at 15:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.