Radial and Non-Radial Solution Structures for Quasilinear Hamilton--Jacobi--Bellman Equations in Bounded Settings

This paper establishes the existence, uniqueness, and global regularity of positive classical solutions to quasilinear Hamilton–Jacobi–Bellman equations on bounded convex domains via a constructive weighted monotone iteration scheme, while providing a probabilistic derivation from controlled Itô diffusions and demonstrating applications in stochastic production planning and image restoration.

The Gist

Imagine you are the captain of a ship navigating through a foggy, stormy sea (the domain). Your goal is to reach the shore (the boundary) with the least amount of fuel and effort possible, but the sea is unpredictable, with random waves pushing you off course (the random noise or diffusion).

This paper is about finding the perfect "navigation map" (the solution) that tells you exactly how to steer your ship at every single point to minimize your total cost. This map is governed by a complex mathematical rule called the Hamilton–Jacobi–Bellman (HJB) equation.

Here is a breakdown of what the author, Dragos-Patru Covei, achieved, using simple analogies:

1. The Core Problem: A Tough Navigation Puzzle

The equation in the paper is a bit like a recipe for a very tricky cake. It has three main ingredients:

• The Smoothness (Diffusion): The random waves trying to mess up your path.
• The Steering Cost (Gradient): How hard you have to turn the wheel. The paper looks at a specific type of "steering cost" that gets very expensive if you turn too sharply (a "supra-quadratic" penalty).
• The Terrain (Source Term): The landscape you are sailing over, which might have hills and valleys (represented by the function $h$).

The Challenge: For a long time, mathematicians could only solve this puzzle if the "steering cost" was simple (like a standard quadratic curve) or if the map was perfectly round (a circle). This paper solves the puzzle for weirdly shaped islands (convex domains) and strange, expensive steering costs (sub-quadratic growth).

2. The Solution: Building a "Safety Net"

The author didn't just guess the answer; he built a constructive proof. Think of it like building a bridge across a canyon:

• Step 1: The Safety Net (Sub- and Super-solutions):
  First, the author builds two imaginary bridges:

  • A lower bridge (Sub-solution) that is definitely too low to be the real answer.
  • An upper bridge (Super-solution) that is definitely too high.
  • He knows the real answer must be sandwiched somewhere between these two. He uses a "torsion function" (imagine a rubber sheet stretched over the island) to build the upper bridge.

• Step 2: The Iterative Ladder (Monotone Iteration):
  Now, he starts at the top bridge and takes a step down. Then he takes another step down.

  • He has a special rule (a "weighted linear monotone iteration") that ensures every step he takes is closer to the truth and never jumps back up.
  • It's like descending a staircase where every step is guaranteed to be lower than the last, but you are guaranteed not to fall through the floor. Eventually, you land exactly on the perfect path.

• Step 3: The Result:
  This process proves that a perfect, smooth map exists, that there is only one such map (uniqueness), and that the map is smooth enough to be useful (regularity).

3. Two Special Cases: The Circle and The Blob

The paper handles two scenarios:

• The Radial Case (The Perfect Circle): If your island is a perfect circle and the terrain is the same in all directions, the solution is also a perfect circle. The complex 2D math simplifies into a simple 1D line (an Ordinary Differential Equation), like peeling an onion layer by layer.
• The Non-Radial Case (The Irregular Blob): If your island is an ellipse or a weird shape, the solution isn't a perfect circle. The author's method works here too, proving that even on a lopsided island, a perfect navigation map exists.

4. Where Does This Come From? (The Stochastic Story)

The paper also explains why this equation exists. It comes from Stochastic Optimal Control.

• Imagine you are playing a video game where you control a character. You want to reach the exit, but the game has random glitches (noise).
• The "Value Function" ($V$) is the game's "cheat sheet" that tells you the minimum score you can get from any position.
• The author shows that the complex math equation is just the formal way of describing this "cheat sheet" for a character moving randomly but trying to be smart.

5. Real-World Applications: Why Should We Care?

The author doesn't just leave this in the math world; he shows how to use it in two very different fields:

• Factory Planning (Production Planning):
  Imagine a factory manager trying to decide how much to produce. If they produce too much, they pay storage costs. If they produce too little, they miss sales. The "random waves" are unpredictable customer demand.

  • The Result: The math gives the factory a perfect "production policy" that tells them exactly how much to make at any moment to minimize costs, even when demand is chaotic.

• Image Restoration (Making Photos Pop):
  Imagine you have a blurry, dull photo. You want to make the edges sharp and the contrast high without making it look fake.

  • The Result: The author uses this equation as a "smart filter." By tweaking a single knob (the parameter $\alpha$), the math can either gently smooth the image or aggressively sharpen the edges.
  • The Magic: When $\alpha$ is close to 1, the math acts like a super-powerful edge detector, making the image look incredibly crisp (better than standard methods like "Histogram Equalization").

Summary

In short, this paper is a masterclass in taming chaos.

1. It proves that a perfect solution exists for a very difficult math problem involving random noise and complex costs.
2. It provides a step-by-step recipe (algorithm) to find that solution.
3. It shows that this recipe works for both factory managers trying to save money and photographers trying to fix blurry pictures.

The author essentially built a universal navigation tool that works on any shape, for any level of difficulty, bridging the gap between abstract probability theory and practical, real-world engineering.

Technical Summary

Here is a detailed technical summary of the paper "Radial and Non-Radial Solution Structures for Quasilinear Hamilton–Jacobi–Bellman Equations in Bounded Settings" by Dragos-Patru Covei.

1. Problem Statement

The paper investigates the existence, uniqueness, and global $C^{1,\beta}$ regularity of positive classical solutions to a specific class of quasilinear Hamilton–Jacobi–Bellman (HJB) equations on bounded convex domains. The governing equation is:

$$-\frac{\sigma^2}{2} \Delta V(y) + C_\alpha |\nabla V(y)|^p - h(y) = 0 \quad \text{in } \Omega,$$
$$V = g \quad \text{on } \partial\Omega,$$

Key Parameters and Conditions:

• Domain: $\Omega \subset \mathbb{R}^N$ is a bounded $C^2$ convex domain.
• Nonlinearity: The gradient term has a power $p = \frac{\alpha}{\alpha-1}$, where $\alpha \in (1, 2]$. This implies $p \in [2, \infty)$, covering supra-quadratic gradient penalties ($p > 2$).
• Coefficients: $\sigma > 0$ is the diffusion coefficient; $C_\alpha = \frac{\alpha-1}{\alpha} (\frac{\alpha}{\alpha-1})^\alpha$ is a constant derived from the Legendre transform of the control cost.
• Source Term: $h(y)$ is a continuous, non-negative function with sub-quadratic growth (specifically, $\lim_{|x|\to\infty} h(x)/|x|^p \in (0, \infty)$).
• Boundary Condition: $g \geq 0$ is a constant.

The paper addresses a gap in the literature where previous works often focused on quadratic costs ($\alpha=2$) or radially symmetric settings, leaving the general non-radial case with sub-quadratic forcing and supra-quadratic gradient penalties less explored.

2. Methodology

The author employs a constructive existence proof based on a weighted linear monotone iteration scheme, avoiding reliance on the vanishing viscosity method or purely probabilistic arguments for the existence proof.

A. Analytical Framework

1. Sub- and Super-Solutions:
   • Sub-solution ($V_-$): Constructed as a constant $g$.
   • Super-solution ($V_+$): Constructed using the torsion function $\phi$ (solution to $-\Delta \phi = 1, \phi|_{\partial\Omega}=0$). Specifically, $V_+ = g + B\phi$, where $B$ is chosen large enough to dominate the source term $h$.
2. Comparison Principle:
   • A rigorous comparison principle is established using a perturbation argument involving the torsion function to handle the nonlinearity of the gradient term. This ensures that if $u \leq v$ on the boundary, then $u \leq v$ in the domain.
3. Weighted Monotone Iteration:
   • A sequence $\{V^{(k)}\}$ is defined starting from the super-solution $V^{(0)} = V_+$.
   • Each iteration solves a linear elliptic equation:
     $$-\frac{\sigma^2}{2} \Delta V^{(k+1)} + \Lambda_0 V^{(k+1)} = h(x) + \Lambda_0 V^{(k)} - C_\alpha |\nabla V^{(k)}|^p$$
   • Weight Selection: The constant $\Lambda_0$ is chosen based on the maximum gradient of the barriers ($M$) and a geometric constant of the domain ($C_\Omega$) to ensure the operator is monotone non-decreasing.
   • Convergence: The sequence is proven to be monotone decreasing and bounded below by $V_-$. Using uniform $W^{2,q}$ estimates and Morrey embedding, the sequence converges uniformly in $C^1(\Omega)$ to a unique classical solution $V \in C^{2,\beta}(\Omega) \cap C^{1,\beta}(\Omega)$.

B. Probabilistic Derivation

The paper bridges stochastic control theory and PDE analysis by deriving the equation from a controlled Itô diffusion problem:

• Control Problem: Minimize the expected cost $J(x; v) = \mathbb{E}[\int_0^\tau (h(X_t) + |v_t|^\alpha) dt + g]$, where the state evolves as $dX_t = v_t dt + \sigma dW_t$.
• Dynamic Programming: Applying Itô's formula and the dynamic programming principle leads to the HJB equation.
• Legendre Transform: The infimum over the control $v$ is computed explicitly using the Legendre–Fenchel transform, recovering the specific quasilinear term $C_\alpha |\nabla V|^p$.

C. Radial Symmetry

For the specific case where $\Omega$ is a ball and $h$ is radial, the paper proves that the unique solution inherits radial symmetry, reducing the PDE to an Ordinary Differential Equation (ODE) in the radial coordinate $r$.

3. Key Contributions

1. Unified Existence Theory: Provides a complete, constructive proof for the existence and uniqueness of classical solutions for $\alpha \in (1, 2]$ on general bounded convex domains, extending beyond the classical quadratic case ($\alpha=2$).
2. Regularity Results: Establishes global $C^{1,\beta}$ regularity for solutions with sub-quadratic source terms, a result previously missing in the literature connecting 1980s theory to modern applications.
3. Algorithmic Construction: The proof is not just existential; it yields a numerically stable iteration scheme that preserves physical properties (positivity, concavity) at every step.
4. Probabilistic Bridge: Formally derives the specific quasilinear PDE from a stochastic optimal control framework with exit-time costs, validating the equation's physical relevance.
5. Radial vs. Non-Radial: Explicitly characterizes the solution structure in both general non-radial domains and radially symmetric settings.

4. Results

• Theorem 1.1 (Non-radial): Proves the existence of a unique positive classical solution $V \in C^{2}(\Omega) \cap C^{1,\beta}(\Omega)$ for general convex domains. If $g>0$ and $h>0$, then $V>0$ in $\Omega$.
• Theorem 1.2 (Radial): Proves that for a ball domain with radial data, the solution is radial and satisfies a specific ODE.
• Numerical Convergence: The weighted monotone iteration scheme converges rapidly (e.g., 9 iterations for production planning) to a tolerance of $10^{-6}$.
• Applications:
  • Stochastic Production Planning: The value function represents minimal expected production costs. The method successfully models inventory dynamics with cubic gradient penalties ($\alpha=1.5, p=3$).
  • Image Restoration: The HJB equation is used for nonlinear contrast enhancement. The parameter $\alpha$ acts as a "contrast knob."
    • Findings: Small $\alpha$ (close to 1, e.g., 1.02–1.10) yields strong geometric enhancement (high EME and Tenengrad scores), significantly outperforming histogram equalization and Total Variation (TV) methods.
    • Mechanism: The term $|\nabla V|^p$ acts as a geometric amplifier of local intensity variations.

5. Significance

This work is significant for both theoretical mathematics and applied sciences:

• Theoretical: It resolves the existence and regularity of solutions for a class of quasilinear HJB equations with supra-quadratic gradient penalties in non-radial settings, filling a critical gap between classical elliptic theory and modern stochastic control.
• Computational: The constructive nature of the proof provides a robust, monotone numerical algorithm that is stable and efficient, avoiding the instabilities often associated with nonlinear PDE solvers.
• Applied: It demonstrates the versatility of HJB equations in diverse fields:
  • Operations Research: Optimizing production planning under uncertainty.
  • Computer Vision: Providing a principled, tunable framework for image contrast enhancement that outperforms classical methods, particularly in preserving edges and enhancing geometric features without excessive smoothing.

In summary, the paper successfully unifies the theoretical analysis of quasilinear HJB equations with practical numerical implementation, offering a powerful tool for solving complex stochastic control problems and image processing tasks.