Asymptotic analysis of higher-order perturbations of the Perona--Malik functional

1. Problem Statement

The paper investigates the asymptotic behavior (via $\Gamma$-convergence) of a specific class of singularly perturbed functionals related to the Perona–Malik equation, a model widely used in image processing for edge-preserving smoothing. The Perona–Malik functional is formally ill-posed because it involves a non-convex energy density that grows logarithmically.

The authors analyze the functional $F_\varepsilon$ defined on a bounded domain $\Omega \subset \mathbb{R}^d$ (with a focus on the 1D case for the rigorous proof) as:
$$F_\varepsilon(u) = \int_\Omega \frac{1}{\varepsilon|\log \varepsilon|} \log\left(1 + \varepsilon|\log \varepsilon||\nabla u|^2\right) dx + \varepsilon^{2k-1} \int_\Omega \|\nabla^{(k)} u\|^2 dx$$
where:

• The first term is a critically scaled logarithmic term approximating the Perona–Malik energy. It behaves like the Dirichlet integral for small gradients but allows for non-convexity (interface formation) for large gradients.
• The second term is a $k$-th order regularization (with $k \geq 2$) designed to prevent the concentration of energy on Cantor parts of the derivative and to balance bulk and interfacial effects.
• The scaling is chosen such that, in the limit $\varepsilon \to 0$, both a bulk energy (Dirichlet integral) and a surface energy (jump discontinuities) emerge simultaneously.

The goal is to identify the $\Gamma$-limit of $F_\varepsilon$ in the $L^1$ topology and to characterize the resulting energy functional, specifically the dependence of the surface energy on the jump amplitude.

2. Methodology

The authors employ the tools of Calculus of Variations, specifically $\Gamma$-convergence, combined with singular perturbation analysis. The methodology involves several key steps:

• Dimensional Reduction: The problem is reduced from $d$ dimensions to the one-dimensional case ($d=1$) using a slicing argument (Fubini's theorem) and density arguments. This allows the analysis of the energy density on intervals.
• Lower Bound via Truncated Potentials: A crucial step is establishing a lower bound for the logarithmic term. The authors show that the logarithmic term can be bounded from below by a truncated quadratic potential (a quadratic function that saturates at a threshold). This connects the problem to previous works on truncated quadratic functionals (e.g., by Solci).
• Interpolation Estimates: To handle the higher-order regularization ($k$-th derivative), the authors adapt local interpolation techniques from Solci [23]. They prove that if the gradient exceeds a certain threshold, the function must transition through a region where higher-order derivatives are controlled.
• Optimal Profile Problem: The surface energy density is derived by solving a one-dimensional optimal-profile problem. This involves minimizing the sum of the interval length and the $k$-th derivative energy over a transition layer, subject to specific boundary conditions.
• Boundary Condition Analysis: A novel aspect of this work is determining the correct boundary conditions for the optimal profile. The authors prove that for $k > 2$, the optimal transition requires homogeneous boundary conditions on all derivatives up to order $k-1$ (i.e., $v^{(\ell)}(0) = v^{(\ell)}(T) = 0$ for $\ell = 1, \dots, k-1$), not just the first derivative. This is achieved by showing that intervals with large gradients can be slightly extended to ensure derivatives vanish at the boundaries without significantly increasing the energy.
• Recovery Sequence Construction: For the upper bound, the authors construct a recovery sequence using the optimal profile scaled by the jump size and the parameter $\varepsilon$.

3. Key Contributions

1. Identification of the $\Gamma$-limit: The paper rigorously identifies the $\Gamma$-limit of the higher-order Perona–Malik functional as a free-discontinuity functional defined on the space of Special Functions of Bounded Variation ($SBV$).
2. Characterization of Surface Energy: The limit functional is shown to be:
   $$F(u) = \int_\Omega |\nabla u|^2 dx + m_k \int_{S_u \cap \Omega} |u^+ - u^-|^{1/k} d\mathcal{H}^{N-1}$$
   where $S_u$ is the jump set, $u^\pm$ are the traces, and the surface energy density scales with the jump amplitude to the power $1/k$.
3. Explicit Constant $m_k$: The constant $m_k$ is characterized as the solution to a specific variational problem:
   $$m_k = \inf_{T>0} \min \left\{ T + \int_0^T |v^{(k)}|^2 dt : v \in H^k(0,T), v(0)=0, v(T)=1, v^{(\ell)}(0)=v^{(\ell)}(T)=0 \forall \ell \in \{1, \dots, k-1\} \right\}$$
   This confirms a conjecture by Gobbino and Picenni regarding the $1/k$ power law.
4. Handling Low Coerciveness: The paper successfully adapts interpolation methods to a "low-coerciveness" framework where the logarithmic growth is weaker than the quadratic growth typically assumed in similar problems.
5. Staircasing Phenomenon Analysis: The results provide a theoretical basis for the "staircasing" phenomenon (the formation of piecewise constant regions) in image processing, linking it to the specific scaling of the regularization term.

4. Main Results

• Theorem 2.1 (Compactness and $\Gamma$-limit): For any sequence $u_\varepsilon$ with bounded energy, there exists a subsequence converging in measure to a function $u \in SBV(\Omega)$. The $\Gamma$-limit is the functional $F(u)$ described above.
• Theorem 5.1 (Alternative Scaling): The authors also analyze a slightly different scaling of the Perona–Malik term (without the $\varepsilon$ factor inside the log argument). In this case, the bulk Dirichlet term vanishes, and the limit functional consists only of the surface term:
  $$F(u) = m_k \int_{S_u \cap \Omega} |u^+ - u^-|^{1/k} d\mathcal{H}^{N-1}$$
  defined on functions where $\nabla u = 0$ almost everywhere. This regime is particularly relevant for explaining the staircasing effect where the image becomes piecewise constant.

5. Significance

• Mathematical Rigor: The paper provides a rigorous justification for the emergence of free-discontinuity problems from non-convex, ill-posed PDEs via singular perturbations.
• Higher-Order Effects: It extends previous results (which were mostly limited to $k=2$) to arbitrary order $k$, revealing how the order of regularization dictates the scaling of the jump energy ($|jump|^{1/k}$).
• Image Processing Theory: The results offer a theoretical explanation for the staircasing phenomenon in Perona–Malik denoising. The analysis shows that under specific scalings, the energy minimization favors piecewise constant solutions (jumps) over smooth transitions, with the cost of a jump depending on its magnitude in a non-linear way ($1/k$ power).
• Optimal Profiles: The derivation of the homogeneous boundary conditions for the optimal profiles is a significant technical advancement, showing that higher-order derivatives must vanish at the interface to minimize the transition energy efficiently.

In summary, this work bridges the gap between discrete approximations of image processing models and continuum free-discontinuity problems, providing a complete asymptotic picture of how higher-order regularization influences the formation of edges and singularities in variational models.