Abstract
We collect some known results on the subdifferentials of a class of one-homogeneous functionals, which consist in anisotropic and nonhomogeneous variants of the total variation. It is known that the subdifferential at a point is the divergence of some “calibrating field”. We establish new relationships between Lebesgue points of a calibrating field and regular points of the level surfaces of the corresponding calibrated function.
Introduction
In the study of variational problems—particularly those arising in image processing, materials science, and plasticity—one-homogeneous functionals play a pivotal role. These functionals, which include the Total Variation (TV) and its anisotropic variants, are characterized by the property $F(\lambda u) = \lambda F(u)$ for $\lambda > 0$. A central object in the analysis of these functionals is the subdifferential, which characterizes the optimality conditions of the functional.
For a class of one-homogeneous functionals, the subdifferential at a point is represented by the divergence of a "calibrating field" $z$. While the relationship between this field and the geometry of the level sets of the function $u$ has been understood in a coarse sense, this paper establishes a rigorous pointwise link. Specifically, it investigates whether the analytical regularity of the field $z$ (expressed through Lebesgue points) corresponds to the geometric regularity of the level surfaces of $u$.
Theoretical Framework
The authors operate within the space of Functions of Bounded Variation (BV). Given a functional $F(u) = \int \phi(x, Du)$, where $\phi$ is one-homogeneous in its second argument, the subdifferential involves a vector field $z$ such that $z(x) \in \partial \phi(x, \nabla u(x))$.
To handle the product of the field $z$ and the gradient measure $Du$, the research utilizes Anzellotti pairings. This framework allows for a generalized integration by parts, which is essential when dealing with functions that may have jump discontinuities, such as those found in image edges.
Blow-up Analysis
The primary tool for establishing pointwise regularity is blow-up analysis. By "zooming in" on a point $\bar{x}$ through a sequence of rescalings, the authors examine the limit behavior of the function $u$ and the field $z$. If the rescaled level sets converge to a flat hyperplane, the point is considered geometrically regular. This analytical technique is coupled with density estimates to ensure that the level sets do not exhibit "porous" or "wispy" behavior near the point of interest.
Key Insights: The Calibration-Geometry Link
The research provides a definitive connection between the analytical properties of the calibrating field and the geometry of the underlying function.
From Lebesgue Points to Regularity
The first major result (Theorem 3.7) establishes that if a point $\bar{x}$ is a Lebesgue point of the calibrating field $z$, then $\bar{x}$ must be a regular point of the level set boundary (in the sense of the reduced boundary). At such a point, the value of the field $z(\bar{x})$ is uniquely determined by the normal vector to the level surface and the specific anisotropy of the functional. This confirms that the calibrating field effectively "picks out" the geometric normal where it is well-defined.
Dimensional Sensitivity and the Converse
While the implication from analytical regularity to geometric regularity is general, the converse—that a regular surface point implies a Lebesgue point for $z$—is more delicate. The authors prove this converse (Theorem 3.10) specifically for dimensions $d=2$ and $d=3$. In these lower dimensions, the stability of the blow-up limits is more manageable, preventing the formation of certain singularities that could decouple the field from the geometry.
The Critical Role of Divergence
A significant finding is the sensitivity of these results to the regularity of the source term $g = -\text{div } z$. The authors demonstrate that the relationship holds when $g \in L^d$. However, they provide a counterexample showing that if $g$ is only slightly less regular (e.g., $g \in L^{d-\epsilon}$), a point can fail to be a Lebesgue point even if the level set geometry is simple. This highlights $L^d$ as a sharp threshold for maintaining the pointwise subdifferential structure.
Applications and Extensions
Unlike classical studies focused solely on Euclidean Total Variation, this work extends to anisotropic (direction-dependent) and nonhomogeneous (position-dependent) functionals. This is particularly relevant for:
- Crystalline Modeling: Where energy depends on the orientation of the crystal lattice.
- Directional Image Processing: Where textures or edges have preferred orientations that vary across the image domain.
Limitations and Future Directions
The restriction of the converse result to $d \le 3$ remains a theoretical boundary. In higher dimensions ($d \ge 8$), the existence of complex singularities like the Simons cone suggests that the blow-up to a flat hyperplane is not guaranteed, potentially breaking the link between surface regularity and the calibrating field.
Furthermore, the current analysis assumes the functional is smooth and uniformly convex (elliptic). Future research may explore "crystalline" functionals, which possess facets and corners. Such functionals are non-elliptic and present additional challenges in defining the relationship between the subdifferential and the geometry of the level sets. Finally, relaxing the $L^d$ constraint on the divergence while preserving geometric links remains a high-priority open problem for the mathematical imaging community.
Summary updated