Gradient estimates for the Perona-Malik equation (Q877161)

This paper concerns the initial boundary value problem \[ \begin{aligned} & u_t=(\varphi^\prime(u_x))_x\quad \text{in }\;(-1,1) \times [0,T), \\ &u_x(-1,t) = u_x(1,t) = 0\quad \text{for }\;t \in[0,T), \\ & u(x,0) = u_0(x)\quad \text{for }\;x \in (-1,1), \end{aligned} \] which is the formal gradient flow of the integral functional \[ \int_{-1}^1\varphi(u_x)\,dx. \] As for \(\varphi\), the authors assume that \(\varphi \in C^2(\mathbb R)\) and \(\varphi^\prime(0)=0\). Therefore, the equation is forward-parabolic where \(\varphi^{\prime\prime}(u_x) > 0\), and backward-parabolic where \(\varphi^{\prime\prime}(u_x) < 0\). The interesting case is when the initial data \(u_0\) is transcritical, that is, such that \(\varphi^{\prime\prime}(u_{0x})\) changes its sign in \([-1,1]\). It is known that this problem has no global solution if \(u_0\) is transcritical, and existence of local solutions, even for special classes of transcritical initial data, is still open problem. The \(n\)-dimensional generalization of this problem is \[ \begin{aligned} & u_t=\operatorname{div}\left( \varphi^\prime(| \nabla u| )\frac{\nabla u}{| \nabla u| }\right)\quad \text{in }\;\Omega \times [0,T), \\ &\frac {\partial u}{\partial n}= 0\quad \text{on }\;\partial\Omega\times[0,T), \\ & u(x,0) = u_0(x)\quad \text{for }\;x \in \Omega, \end{aligned} \] which is the gradient flow of the integral functional \(\int_\Omega\varphi(| \nabla u| )\,dx\), where \(\Omega\) is a bounded open set with piecewise \(C^1\) boundary. The typical example is the so called Perona-Malik equation corresponding to \(\varphi(\sigma) = \frac 12\log(1+\sigma^2)\) which is introduced by \textit{P. Perona} and \textit{J. Malik} [IEEE Trans. Pattern Anal. Mach. Intell. 12, 629--639 (1990)] in the context of image processing [see \textit{B. Kawohl}, Math. Methods Appl. Sci. 27, 1803--1814 (2004; Zbl 1060.35054)]. In this paper, in the one dimensional problem, it is shown that if \(u_0\) is transcritical, then the \(L^1\) norm of \(u_x\) in the space variable is non-increasing in time and the \(L^\infty\) norm of \(u_x\) in the space variable is non-decreasing in time. Then, the authors consider the semi-discrete scheme obtained by replacing the space derivatives by finite differences. Extending the previous estimates to the discrete setting they prove a compactness result for the scheme and characterize the possible limits in some cases. In the \(n\)-dimensional problem \((n > 1)\), examples show that the corresponding estimates on \(\nabla u\) are in general false.