← Visual companionIn-app Blueprint viewLean 4 source indexPDFDependency graph

Erratum: Deep Equilibrium Models for Poisson Imaging Inverse Problems via Mirror Descent

Formalization Blueprint (Lean 4)

This blueprint accompanies a complete, sorry-free Lean 4 formalization (using pure Mathlib v4.27.0) of the erratum to Daniele, Villa, Vaiter, and Calatroni (SIAM J. Imaging Sci., 2026, arXiv:2507.11461) concerning Deep Equilibrium Models for Poisson imaging inverse problems via Mirror Descent. We formalize:
  1. the variational calculus primitives and a full proof of the abstract Kurdyka–Łojasiewicz (KŁ) convergence theorem (Theorem A.8 of Attouch, Bolte, and Svaiter);

  2. the Poisson/Kullback–Leibler variational setup with zero counts permitted (\(y_i \ge 0\)), the counterexample (Example 1) showing that the classical Burg entropy gradient fails to be Lipschitz continuous on sublevel sets \(L_0\), and the shifted Burg geometry (\(h_\xi \), \(D_{h_\xi }\)) yielding the variational equivalence \((1) \iff (4) \iff (5)\);

  3. sublevel separation from the boundary singular set and relative smoothness \(L_\beta h_\xi - F\) convex on \(K_\beta \) (Lemma 2);

  4. single-step positivity of denominators, sublevel invariance, Bregman sufficient decrease (6), and relative error bound (7) (Lemma 3), followed by constant-step sequence convergence (Theorem 4, passing the definable KŁ property \(\mathrm{HasKL}(\Psi )\) as an explicit hypothesis);

  5. monotone backtracking step-size stabilization and global convergence (Algorithm 1).