Abstract:In this work we derive higher order error estimates for inverse problems distorted by non-additive noise, in terms of Bregman distances. The results are obtained by means of a novel source condition, inspired by the dual problem. Specifically, we focus on variational regularization having the Kullback-Leibler divergence as data-fidelity, and a convex penalty term. In this framework, we provide an interpretation of the new source condition, and present error estimates also when a variational formulation of the source condition is employed. We show that this approach can be extended to variational regularization that incorporates more general convex data fidelities.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: in inverse problems under the influence of non - additive noise, derive higher - order error estimates. Specifically, the author focuses on the performance of variational regularization methods when dealing with data distorted by non - additive noise, especially when using the Kullback - Leibler divergence as the data fidelity term. To achieve this goal, the author introduces a new source condition based on the dual problem and measures the error by the Bregman distance.
### Problem Background
Inverse problems are a classic problem in mathematics and applied sciences, usually represented as an operator equation \(Au = f\), where \(A\) is a bounded linear operator from Banach space \(X\) to Banach space \(Y\), \(u\) is the unknown solution, and \(f\) is the observed data. However, in practical applications, the observed data \(f\) may be contaminated by noise, resulting in the equation becoming \(Au = f^{\delta}\), where \(\|f^{\delta}-f\|\leq\delta\). If the problem is ill - posed, that is, the solution does not exist, is not unique, or is unstable, then regularization methods need to be used to obtain a stable approximate solution.
### Research Focus
The research focuses of this paper are as follows:
1. **Non - additive noise**: Different from the traditional additive Gaussian noise, this paper considers more complex and realistic non - additive noises, such as Poisson noise.
2. **Higher - order error estimates**: In addition to the existing low - order error estimates, the author aims to derive higher - order error estimates to better understand the performance of regularization methods under different noise conditions.
3. **New source condition**: In order to obtain these higher - order error estimates, the author introduces a new source condition, which is not only applicable to the dual problem but also can explain the relationship between the regularization parameter and the solution.
### Formula Summary
The key formulas involved in the article include:
- **Bregman distance**:
\[
D_{\xi}^R(u, v)=R(u)-R(v)-\langle\xi, u - v\rangle
\]
If \(\partial R(v)\) is a singleton, it is simplified to \(D_R(u, v)\).
- **Symmetric Bregman distance**:
\[
D_s^R(u, v)=D_{\xi}^R(u, v)+D_{\zeta}^R(v, u)
\]
- **New source condition**:
\[
\exists p^{\dagger}\in L^{\infty}(\Gamma), v^{\dagger}\in L^1(\Omega)\text{ such that }A^{*}p^{\dagger}\in\partial R(u^{\dagger})\text{ and }f p^{\dagger}=A v^{\dagger}
\]
- **Error estimate**:
\[
D_R(u_{\alpha}, u^{\dagger})\leq D_R(u^{\dagger}-s(\alpha)v^{\dagger}, u^{\dagger})+O(\alpha^3)
\]
Through these formulas and the newly introduced source condition, the author successfully derives higher - order error estimates under non - additive noise conditions, providing theoretical support for the numerical solution of inverse problems.