2 Poisson Problem Setup, Counterexample, and Shifted Burg Geometry
Let \(a {\gt} 0\), \(\xi {\gt} 0\), \(y \in \mathbb {R}^m_{\ge 0}\) (zero counts allowed), \(A \in \mathbb {R}^{m \times n}_{\ge 0}\) with every row nonzero, and \(R_\theta : \mathcal{E}_n \to \mathbb {R}\) of class \(C^2\) on an open neighborhood \(U_R \supseteq [0, a]^n\). Set \(C := [0, a]^n\),
with \(0 \log (0/u) := 0\) (even for \(u = 0\)), and \(\Psi (x) := \mathrm{KL}(y, Ax) + R_\theta (x) + \iota _{[0, a]^n}(x)\). Define the open positive-observation set \(U_+ := \{ x \in \mathcal{E}_n \mid \forall i,\; y_i {\gt} 0 \implies (Ax)_i {\gt} 0\} \) and the boundary singular set
We have \(\mathrm{dom}(\Psi ) = [0, a]^n \cap U_+\), \(\overline{\mathrm{dom}(\Psi )} = [0, a]^n\), and \(\overline{\mathrm{dom}(\Psi )} \setminus \mathrm{dom}(\Psi ) = \mathcal{S}\).
Membership \((Ax)_i {\gt} 0\) for every \(i\) with \(y_i {\gt} 0\) characterizes finiteness of \(\mathrm{KL}(y, Ax)\), while \(\iota _{[0,a]^n}(x) {\lt} +\infty \) characterizes \(x \in [0, a]^n\). Approximating any \(x \in [0, a]^n\) along convex combinations \((1 - t) x + t (a, \dots , a)\) stays in \([0, a]^n \cap U_+\) for \(t \in (0, 1]\), proving \(\overline{\mathrm{dom}(\Psi )} = [0, a]^n\) and \(\overline{\mathrm{dom}(\Psi )} \setminus \mathrm{dom}(\Psi ) = \mathcal{S}\).
The classical Burg entropy on \(\mathcal{E}_n\) is \(h(x) := -\sum _{j=1}^n \log x_j\) on \(\mathbb {R}^n_{{\gt}0}\) (and \(+\infty \) otherwise), with \((\nabla h(x))_j = -1/x_j\). For shift \(\xi {\gt} 0\), the shifted Burg entropy and Bregman divergence are
With \(A = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}\), \(R_\theta \equiv 0\), \(a \ge 1\), \(y \in \mathbb {R}^2_{{\gt}0}\), and \(x^0 = (0, 1)\), we have \(\Psi (x^0) {\lt} +\infty \) and every \(x_t := (t, 1)\) for \(t \in (0, 1]\) lies in \(L_0(x^0) \cap (0, a]^n\) with \(t \mapsto \Psi (x_t)\) bounded above on \((0, 1]\). Yet \(\| \nabla h(x_t) - \nabla h(x_1)\| = |1 - 1/t| \to +\infty \) while \(\| x_t - x_1\| = 1 - t \le 1\), so \(\nabla h\) is not Lipschitz continuous on \(\mathrm{dom}(h) \cap L_0(x^0)\).
Direct evaluation gives \(A x_t = (t + 1, 1)\) componentwise \(\ge 1\), keeping \(\mathrm{KL}(y, A x_t)\) continuous on \([0, 1]\), hence bounded by its maximum. Meanwhile \(\nabla h(x_t)_0 - \nabla h(x_1)_0 = 1 - 1/t\) grows arbitrarily large as \(t \downarrow 0\) whereas \(\| x_t - x_1\| \le 1\).
For all \(u, x \in [0, a]^n\),
By the 1D integral/mean-value identity for \(s \mapsto -\log (s + \xi )\) on \([0, a]\), each coordinate Bregman divergence satisfies \(-\log (\frac{u_j + \xi }{x_j + \xi }) + \frac{u_j - x_j}{x_j + \xi } \ge \frac{(u_j - x_j)^2}{2(a + \xi )^2}\), and \(|-(u_j + \xi )^{-1} + (x_j + \xi )^{-1}| \le \frac{|u_j - x_j|}{\xi ^2}\).
For step size \(\tau {\gt} 0\) and \(x \in [0, a]^n\), the shifted-Burg Mirror Descent subproblem objective is
We define the operator \(T_\tau \) coordinatewise by
For any \(\tau {\gt} 0\) and \(x \in [0, a]^n\) satisfying \(1 + \tau (x_j + \xi ) \partial _j F(x) {\gt} 0\) for all \(j\):
\(T_\tau (x)\) satisfies the optimality inclusion (5):
\[ -\nabla F(x) - \frac{1}{\tau }\bigl(\nabla h_\xi (T_\tau (x)) - \nabla h_\xi (x)\bigr) \in \partial \iota _{[0, a]^n}(T_\tau (x)) = N_{[0, a]^n}(T_\tau (x)); \]\(T_\tau (x)\) is the unique global minimizer (1) of \(\Phi _{x, \tau }\) over \(\mathcal{E}_n\), and conversely any global minimizer of \(\Phi _{x, \tau }\) coincides with \(T_\tau (x)\) (4).
Furthermore, for every \(\tau \in \mathbb {R}\), \(T_\tau \) is continuous on \([0, a]^n \cap U_+\).
The coordinate scalar inclusion for \(N_{[0, a]}(y_j)\) follows by inspecting the three cases \((x_j + \xi )/d_j - \xi \le 0\), \(\ge a\), and \(\in (0, a)\). The three-point Bregman identity together with Lemma 16 makes \(\Phi _{x, \tau }(z) - \Phi _{x, \tau }(T_\tau (x)) {\gt} 0\) for all \(z \ne T_\tau (x)\). Clamping the denominator below by \(\delta = \frac{\xi }{2(a + \xi )}\) agrees identically with \(T_\tau \) on \([0, a]^n \cap U_+\), establishing continuity even when denominators approach or cross zero.