Scientific

Inverse problems arise in many imaging applications, such as image
reconstruction (e.g., computed tomography), image deblurring, and
digital super-resolution. These inverse problems are very difficult
to solve; in addition to being large scale, the underlying
mathematical model is often ill-posed, which means that noise and
other errors in the measured data can be highly magnified in computed
solutions. Regularization methods are often used to overcome this
difficulty. In this talk we describe hybrid Krylov subspace based
regularization approaches that combine matrix factorization methods
with iterative solvers. The methods are very efficient for large scale
imaging problems, and can also incorporate methods to automatically
estimate regularization parameters. We also show how the approaches
can be adapted to enforce sparsity and nonnegative constraints.

We will use many imaging examples that arise in medicine and astronomy
to illustrate the performance of the methods, and at the same time
demonstrate a new MATLAB software package that provides an easy to use
interface to their implementations.

This is joint work with Silvia Gazzola (University of Bath) and
Per Christian Hansen (Technical University of Denmark).

Phase retrieval is a problem that arises in a wide range of imaging
applications, including x-ray crystallography, x-ray diffraction imaging
and ptychography. The data in the phase retrieval problem are samples of
the modulus of the Fourier transform of an unknown function. To
reconstruct this function one must use auxiliary information to determine
the unmeasured Fourier transform phases. There are many algorithms to
accomplish task, but none work very well. In this talk we present an
analysis of the geometry that underlies these failures and points to new
approaches for solving this class of problems.

This talk will discuss results in a joint paper of mine with Steve Lalley from 1992.

Suppose a particle, starting at position 0, moves according to a diffusion process on the real line. Suppose also that this particle emits daughter particles according to a branching process whose instantaneous rate can depend on location, though not on time. The daughter particles move independently according to the same diffusion process, starting at their points of birth, and in turn emit their own daughters according to the same branching process. The simplest special case of this situation is standard branching Brownian motion, with the rate of reproduction not depending on location.

Let R_t be the position of the right-most particle at time t, and let m_t be the median of R_t. In 1937, Kolmogorov, Petrovskii, and Piskunov showed that, for standard branching Brownian motion, (m_t)/ t converges to SQRT(2) and that (R_t - m_t) converges in distribution to a nondegenerate limiting distribution.It turns out that results like those proved by Kolmogorov, et al, hold in great generality for one-dimensional branching diffusions. If the branching diffusion is "recurrent" (in the sense that the initial position is re-visited at arbitrarily large times by _some_ particle), and if space is rescaled so that m_t grows linearly, then (R_t - m_t) converges in distribution to a location-mixture of extreme value distributions. We also have the Andy Warhol Theorem, according to which every particle ever born has a descendant in the lead at some point in the future.

Classifying solutions is one of central themes in nonlinear partial differential equations. This is the content of various Liouville type theorems and more recently De Giorgi type conjectures. I will report recent progress towards De Giorgi's conjecture for Allen-Cahn equation and free boundary problems, and related issues in nonlinear Schroedinger equations.