Technical Summary: A Morse-Bott Framework for Blind Inverse Problems

Problem Statement

The paper addresses blind inverse problems, specifically the joint recovery of an image $\bar{x}$ and a forward operator parameter $\bar{\theta}$ (e.g., a blur kernel) from an indirect observation $y = A(\bar{\theta})\bar{x} + b$. These problems are severely ill-posed, as infinitely many pairs $(x, \theta)$ can explain the same observation.

The authors focus on the Maximum A Posteriori (MAP) estimation framework, where the solution is sought by minimizing the negative log-posterior:
$$L(x, \theta; y) = \frac{1}{2\sigma^2} \|A(\theta)x - y\|^2 - \log p_x(x) - \log p_\theta(\theta)$$
While handcrafted priors (e.g., sparsity) have historically dominated, recent advances utilize learned priors (specifically diffusion models) to capture the complexity of natural images. However, a paradox persists: while local optimization with careful initialization often succeeds, global MAP optimization consistently fails, converging to degenerate "no-blur" solutions (the "blurry trap") where the operator collapses to the identity and the image remains blurry. The paper aims to rigorously explain this phenomenon and establish conditions under which local recovery is theoretically guaranteed.

Methodology

1. The Morse-Bott Modeling Hypothesis

The authors propose modeling the image potential $q(x) = -\log p_x(x)$ as a Morse-Bott function. Unlike traditional convex assumptions, this framework acknowledges that natural images reside on low-dimensional manifolds.

• Definition: A potential $q$ is Morse-Bott at a critical point $\bar{x}$ if the set of critical points forms a smooth submanifold $\mathcal{M}$ locally, and the Hessian $\nabla^2 q(\bar{x})$ is non-degenerate in directions normal to $\mathcal{M}$.
• Implication: The potential is flat along the tangent space of the image manifold ($T_{\bar{x}}\mathcal{M}$) but strictly convex in the normal directions.

2. Empirical Validation of the Hypothesis

The authors validate this hypothesis using state-of-the-art diffusion models (FFHQ, ImageNet, AFHQ) by:

• Locating Critical Points: Running gradient descent from natural images, unrelated datasets, and noise. They observe convergence to critical points that act as "denoised" versions of the input.
• Landscape Curvature: Demonstrating that the potential increases sharply when moving away from the manifold (transverse directions) but remains relatively flat along it.
• Hessian Spectral Analysis: Computing the Hessian spectrum at critical points. They find a clear spectral gap: a small number of near-zero eigenvalues (corresponding to the tangent space of the manifold) and a vast majority of large, positive eigenvalues (normal directions). This confirms the Morse-Bott structure for learned diffusion potentials.

3. Theoretical Analysis of Stability

Using the Morse-Bott assumption, the authors derive local recovery guarantees for both non-blind and blind inverse problems.

• Non-blind Case: They prove that if the ground-truth image $\bar{x}$ is a second-order critical point of the potential and the operator $A$ is injective on the tangent space of the manifold ($T_{\bar{x}}\mathcal{M} \cap \ker A = \{0\}$), then $\bar{x}$ is a stable local minimizer.
• Blind Case: They extend this to joint recovery $(\bar{x}, \bar{\theta})$. Stability requires three geometric conditions:
  1. Identifiability on the manifold: $T_{\bar{x}}\mathcal{M} \cap \ker A(\bar{\theta}) = \{0\}$.
  2. Identifiability of parameters: $\ker J_{\bar{x}}(\bar{\theta}) = \{0\}$ (variations in $\theta$ produce measurable changes).
  3. Manifold-operator decoupling: $(A(\bar{\theta})T_{\bar{x}}\mathcal{M}) \cap \text{Im } J_{\bar{x}}(\bar{\theta}) = \{0\}$ (image variations cannot be confused with operator variations).
• Result: Under these conditions, the ground truth is a stable local minimizer. The solution depends smoothly on data perturbations, and gradient descent initialized nearby converges to this minimizer.

4. Analysis of Global Failure (The "Blurry Trap")

The authors investigate why global MAP fails despite local stability.

• Empirical Observation: They demonstrate that for diffusion-based potentials trained on natural images, the potential $q(x)$ strictly decreases as the image becomes blurrier. This is a global property of the landscape, not just a local artifact.
• Theoretical Proof: They prove that if the potential satisfies a "blurry preference" condition ($q(h * x) \leq q(x)$ for all blur kernels $h$), the global minimizer of the MAP objective is always a degenerate "no-blur" solution, regardless of the generative quality of the prior.
• Conclusion: The failure of global MAP is an intrinsic characteristic of the landscape's topology, not a limitation of the prior's quality.

Key Results

1. Local Stability: The paper establishes that for a broad class of blind inverse problems, the ground truth is a stable local minimizer provided the operator and image manifold satisfy specific geometric transversality conditions. This justifies the empirical success of local optimization methods when properly initialized.
2. Global Instability: The "blurry trap" is confirmed to persist even with state-of-the-art diffusion priors. The global landscape is biased toward smooth, blurry structures because natural image statistics favor low-frequency content.
3. Numerical Verification:
   • Experiments on various blur kernels (Gaussian, motion, diffraction, simplex) show that the Morse-Bott valley (local basin) exists but is often overshadowed by the global attractor.
   • Identifiability varies by kernel: Motion blur and defocus blur exhibit sharper local minima and better conditioning than Gaussian or Airy blurs, satisfying the decoupling condition more effectively.
   • Noise Sensitivity: As noise levels increase, the local recovery basin collapses, and the global "blurry trap" dominates, rendering the MAP estimator inconsistent beyond a critical noise threshold.
4. Initialization Strategy: The authors propose a heuristic alternating optimization scheme (Algorithm 6.1) to navigate the landscape:
   • Basin-aware initialization: Initialize with a "large" (maximal support) kernel rather than a small one. This forces the image component to contain high-frequency content, placing the optimization in the transverse region where the manifold-operator decoupling holds.
   • Asymmetric updates: Perform multiple image updates for every kernel update to keep the image close to the manifold floor.
   • Periodic reset: Reset the image to the observation $y$ periodically to escape small local pits and maintain data consistency.
   • Validation: This strategy successfully recovers ground truths in synthetic experiments and on real-world data (Köhler dataset), whereas small-kernel initialization fails.

Significance and Claims

The paper claims to provide a rigorous theoretical framework (Morse-Bott) that explains the dichotomy between the success of local optimization and the failure of global optimization in blind inverse problems with learned priors.

• Theoretical Contribution: It formalizes the geometry of learned potentials, showing they are Morse-Bott functions. It provides explicit stability constants and conditions (identifiability, decoupling) for local recovery, extending compressed sensing theories to non-convex, learned settings.
• Practical Insight: It demonstrates that the "blurry trap" is a fundamental topological feature of natural image statistics, not a flaw in specific model architectures. Therefore, improving the generative quality of the prior alone cannot solve the global MAP failure.
• Strategic Shift: The authors argue that the focus of blind inverse problem research should shift from seeking global maximizers to characterizing and targeting local recovery basins. Success depends on "topologically informed optimization," specifically strategic initialization that avoids the global attractor.

The paper concludes that while local recovery is theoretically guaranteed under geometric conditions, the global landscape is inherently biased. Thus, practical algorithms must rely on heuristics (like large-kernel initialization) to navigate this complex terrain, rather than relying on the global optimality of the MAP estimator.