← Visual companionIn-app Blueprint viewLean 4 source indexPDFDependency graph

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

4 Lemma 3 and Constant-Step Convergence — Theorem 4

Definition 26 Step-size threshold \(\bar{\tau }\) and constant-step iterates
✓

Set \(G := \sup _{x \in K_{\beta _0}} \| \nabla F(x)\| _\infty \ge 0\) and

\[ \bar{\tau } := \min \! \left\{ \frac{1}{(a + \xi ) G},\; \frac{1}{L_{\beta _1}}\right\} {\gt} 0 \]

(taking \(\bar{\tau } := 1 / L_{\beta _1}\) if \(G = 0\)).

Theorem 27 Single-step well-definedness, descent (6), and relative error (7) — Lemma 3
✓

Let \(x^0 \in [0, a]^n\) with \(\Psi (x^0) {\lt} +\infty \). For every \(x \in L_0(x^0)\) and every \(\tau \in (0, \bar{\tau })\):

  1. \(1 + \tau (x_j + \xi ) \partial _j F(x) {\gt} 0\) for every coordinate \(j \in \{ 1, \dots , n\} \);

  2. \(T_\tau (x) \in K_{\beta _1} \cap L_0(x^0)\);

  3. Bregman sufficient decrease (6) holds:

    \[ \Psi (x) - \Psi (T_\tau (x)) \ge \left(\frac{1}{\tau } - L_{\beta _1}\right) D_{h_\xi }(T_\tau (x), x); \]
  4. setting \(\omega := \nabla F(T_\tau (x)) - \nabla F(x) + \frac{1}{\tau }(\nabla h_\xi (x) - \nabla h_\xi (T_\tau (x)))\), we have \(\omega \in \partial \Psi (T_\tau (x))\) and the relative error bound (7):

    \[ \| \omega \| \le \left(M_{\beta _1} + \frac{1}{\tau \xi ^2}\right) \| T_\tau (x) - x\| . \]
Proof ▶

Using the continuous map \(t \mapsto T_t(x)\) on \([0, \tau ]\), define \(U := \{ t \in [0, \tau ] \mid T_t(x) \in K_{\beta _0}\} \) (closed, containing \(0\)) and \(V := \{ t \in \mathbb {R}\mid T_t(x) \in U_+ \text{ and } \mathrm{KL}(y, A T_t(x)) {\lt} \beta _1\} \) (open). Any \(t \in [0, \tau ] \cap V\) has \(T_t(x) \in K_{\beta _1}\) and \(1 + t (x_j + \xi ) \partial _j F(x) {\gt} 0\), so Theorem 25 and Theorem 18 give \(\Psi (T_t(x)) \le \Psi (x) \le \Psi (x^0)\), putting \(t \in U\). By connectedness of \([0, \tau ]\), \(\tau \in U\), yielding \(T_\tau (x) \in K_{\beta _0} \subseteq K_{\beta _1}\), (6), and \(T_\tau (x) \in L_0(x^0)\). Combining Theorem 18(1) with Lemma 4 and Lipschitz bounds for \(\nabla F\) and \(\nabla h_\xi \) gives (7).

Definition 28 Constant-step Mirror Descent sequence
✓
#

For fixed \(\tau {\gt} 0\) and \(x^0 \in [0, a]^n\), define \(x^0_{\mathrm{MD}} := x^0\) and \(x^{k+1} := T_\tau (x^k)\) for \(k \in \mathbb {N}\).

Theorem 29 Global convergence for constant step size — Theorem 4
✓

Let \(x^0 \in [0, a]^n\) with \(\Psi (x^0) {\lt} +\infty \), let \(\tau \in (0, \bar{\tau })\), and assume \(\Psi \) satisfies the KŁ property (\(\mathrm{HasKL}(\Psi )\)). Then the sequence \((x^k)_{k \in \mathbb {N}}\) satisfies

\[ \sum _{k=0}^\infty \| x^{k+1} - x^k\| {\lt} +\infty \]

and converges to a critical point \(x^\infty \in \mathrm{crit}(\Psi ) \cap L_0(x^0)\) which is a fixed point of \(T_\tau \), i.e., \(x^\infty = T_\tau (x^\infty )\).

Proof ▶

By induction via Theorem 27, every \(x^k\) stays in the bounded closed set \(L_0(x^0) \subseteq K_{\beta _0}\). Combining (6) with Lemma 16 verifies (H1) with \(c_1 = \frac{1/\tau - L_{\beta _1}}{2(a+\xi )^2} {\gt} 0\), while (7) gives (H2) with \(b = M_{\beta _1} + \frac{1}{\tau \xi ^2} {\gt} 0\), and continuity of \(F\) on \(U_+ \cap U_R \supseteq K_{\beta _0}\) verifies (H3). Theorem 11 yields summability and convergence \(x^k \to x^\infty \in \mathrm{crit}(\Psi ) \cap L_0(x^0)\), and continuity of \(T_\tau \) on \([0, a]^n \cap U_+\) passes to the limit \(x^\infty = T_\tau (x^\infty )\).