Colloquium de l'Institut

Seeing More Where It Matters: Spatially Adaptive Regularization of Inverse Problems

par Mirza Karamehmedovic

→ Europe/Paris
IMT

IMT

Description

Inverse problems arise whenever we attempt to recover hidden information
from indirect or imperfect measurements. A fundamental obstacle is
instability: recovering finer details typically requires amplifying
measurement errors. This creates the familiar trade-off between resolution
and stability, encountered in applications ranging from medical imaging to
signal processing.

But must the resolution of a reconstruction be the same everywhere? And can
we recover finer details in selected regions without making the overall
reconstruction less stable?

In this talk, I will discuss a spatially adaptive approach to deconvolution
based on variational mollification. Starting from classical Fourier-based
regularization, we introduce a reconstruction operator whose symbol depends
on both position and frequency, allowing the resolution to vary across the
domain. We establish consistency, convergence rates, and stability
estimates, including an explicit relationship between local resolution and
global stability.

The central message is that resolution can be redistributed: by accepting
lower resolution in less important regions, we can recover finer details
where they matter most, while maintaining a prescribed level of global
stability. Numerical examples in one and two dimensions illustrate this
principle.

The talk is based on joint work with Pierre Maréchal, Martin Sæbye Carøe,
and Lara Baalbaki.