## Math Calendar

Thursday, January 23, 2020

15:00-16:00

BBG.401

Master Seminar

Simon Veldkamp-Statistical postprocessing of wind spedd forecasts using convolutional neural networks

see description

I would like to invite all of you to the presentationof my master’s thesis, written under the supervision of Dr. Sjoerd Dirksen(UU), Dr. Maurice Schmeits (KNMI) and Kirien Whan (KNMI), with Prof. Dr.Jason Frank as the second reader.

Weatherforecasts provided by numerical weather prediction (NWP) models typically givea deterministic forecast. However, there is a certain amount of uncertainty inthese forecasts. The aim of statistical post-processing is to give aprobabilistic forecast instead. Current statistical post-processing methods forproviding a probabilistic forecast are not capable of using full spatialpatterns from the NWP model. Recent developments in deep learning (notablyconvolutional neural networks) have made it possible to use large gridded inputdata sets. This could potentially be useful in statistical postprocessing,since it allows us to use more spatial information.

In this research we consider wind speed forecasts for 48 hours ahead, asprovided by KNMI's Harmonie-Arome model. Convolutional neural networks, fullyconnected neural networks and quantile regression forests are used to obtainprobabilistic wind speed forecasts. Comparing these methods shows thatConvolutional neural networks are more skillful than the other methods,especially for medium to higher wind speeds.

Weatherforecasts provided by numerical weather prediction (NWP) models typically givea deterministic forecast. However, there is a certain amount of uncertainty inthese forecasts. The aim of statistical post-processing is to give aprobabilistic forecast instead. Current statistical post-processing methods forproviding a probabilistic forecast are not capable of using full spatialpatterns from the NWP model. Recent developments in deep learning (notablyconvolutional neural networks) have made it possible to use large gridded inputdata sets. This could potentially be useful in statistical postprocessing,since it allows us to use more spatial information.

In this research we consider wind speed forecasts for 48 hours ahead, asprovided by KNMI's Harmonie-Arome model. Convolutional neural networks, fullyconnected neural networks and quantile regression forests are used to obtainprobabilistic wind speed forecasts. Comparing these methods shows thatConvolutional neural networks are more skillful than the other methods,especially for medium to higher wind speeds.

16:00-17:00

Applied Mathematics Seminar

Rob Bisseling (UU) - Parallel Tomographic Reconstruction - Where Combinatorics Meets Geometry, HFG 409

see description

Title: Parallel Tomographic Reconstruction - Where Combinatorics Meets Geometry

Abstract: Today, high-resolution tomographic reconstruction of 3D objects is within reach, but the associated data sets are huge and calling for parallel computation. A typical 3D reconstruction with 4k resolution already produces an image of 256 Gbytes. Tomographic reconstruction is often done using iterative algorithms that involve repeated sparse matrix-vector multiplication (SpMV). The matrix, however, may be too large to store, requiring Tbytes of memory, and hence each matrix row is recomputed upon use.

In this talk, we present data partitioning methods for tomography matrices of increasing size. For small matrices, we can compute an optimal bipartitioning by an exact combinatorial method, as implemented in the packages MondriaanOpt and MP. This allows us to gauge the quality of medium-grain partitioning (default in the Mondriaan package), which is a heuristic combinatorial method that can handle larger problems. Medium-grain results in turn justified choosing column partitioning for the tomographic matrix-free SpMV. For this column partitioning, we developed a geometric recursive coordinate bisection algorithm with nearly the same output quality as combinatorial partitioning that can handle huge, matrix-free problems and is also faster.

We conclude with showing the scalability of an actual reconstruction that was written using Bulk, a modern C++ library for easy development of parallel programs in bulk-synchronous parallel style.

Abstract: Today, high-resolution tomographic reconstruction of 3D objects is within reach, but the associated data sets are huge and calling for parallel computation. A typical 3D reconstruction with 4k resolution already produces an image of 256 Gbytes. Tomographic reconstruction is often done using iterative algorithms that involve repeated sparse matrix-vector multiplication (SpMV). The matrix, however, may be too large to store, requiring Tbytes of memory, and hence each matrix row is recomputed upon use.

In this talk, we present data partitioning methods for tomography matrices of increasing size. For small matrices, we can compute an optimal bipartitioning by an exact combinatorial method, as implemented in the packages MondriaanOpt and MP. This allows us to gauge the quality of medium-grain partitioning (default in the Mondriaan package), which is a heuristic combinatorial method that can handle larger problems. Medium-grain results in turn justified choosing column partitioning for the tomographic matrix-free SpMV. For this column partitioning, we developed a geometric recursive coordinate bisection algorithm with nearly the same output quality as combinatorial partitioning that can handle huge, matrix-free problems and is also faster.

We conclude with showing the scalability of an actual reconstruction that was written using Bulk, a modern C++ library for easy development of parallel programs in bulk-synchronous parallel style.

Tuesday, January 28, 2020

16:00-17:00

MIN 014

Utrecht Geometry Center Seminar

Lior Bary-Soroker (Tel Aviv University) - Twin primes, a case study in number theory in function fields

see description

Abstract:

Number theory is famous for having many questions that are notoriously hard. Sometimes even to formulate a plausible conjecture is challenging. A classical theme in number theory is to study these question on ‘lab rats’, that is to say, on polynomials over finite fields. And indeed this approach produced extraordinary results such as the resolution of the Riemann Hypothesis for curves by Weil and the resolution of Weil's conjectures by Deligne, the equidistribution theorems of Katz, and many other results.

In this talk I will consider one problem that saw a great progress from 2006 til 2020. The problem, known as the twin prime conjecture, asks whether there are infinitely many primes p such that p+2 is also prime.

Thursday, January 30, 2020

15:30-16:30

MIN 201

The Utrecht Mathematical Colloquium

Lior Bary-Soroker (Tel Aviv) - Virtually all polynomials are irreducible

see description

Abstract: It has been known for almost a hundred years that most polynomials with large integral coefficients are irreducible and have a big Galois group. For a few dozen years, people have been interested in whether the same holds when one considers families of polynomials with small coefficients—notably, polynomials with plus-minus 1 coefficients. In particular, “some guy on the street” conjectures that the probability for a random plus-minus 1 coefficient polynomial to be irreducible tends to 1 as the degree tends to infinity (a much earlier conjecture of Odlyzko-Poonen is about the 0-1 coefficients model). In the talk, I will discuss these types of problems, and approach to attack them using combination of combinatorics, analytic number theory, random permutations, and number theory in function fields.

Tuesday, February 4, 2020

16:00-17:00

MIN 014

Utrecht Geometry Center Seminar

Job Kuit (Universität Paderborn) - Discrete series representations for real spherical spaces

see description

Abstract:

Let $Z = G/H$ be a homogeneous space attached to a real reductive group $G$ and a closed subgroup $H$. A principal objective in the harmonic analysis of $Z$ is the understanding of the $G$-equivariant spectral decomposition of the space $L^{2}(Z)$ of square integrable half-densities. The irreducible components of $L^{2}(Z)$ are of particular interest, they comprise the discrete series for $Z$. In this talk I will focus on real spherical homogeneous spaces $Z$ and present some recent results.

Let $Z = G/H$ be a homogeneous space attached to a real reductive group $G$ and a closed subgroup $H$. A principal objective in the harmonic analysis of $Z$ is the understanding of the $G$-equivariant spectral decomposition of the space $L^{2}(Z)$ of square integrable half-densities. The irreducible components of $L^{2}(Z)$ are of particular interest, they comprise the discrete series for $Z$. In this talk I will focus on real spherical homogeneous spaces $Z$ and present some recent results.

Thursday, February 6, 2020

16:00-17:00

Applied Mathematics Seminar

Jeremy Budd (TU Delft) - Allen-Cahn and MBO on graphs, HFG 611

see description

Title: Allen-Cahn and MBO on graphs

Abstract: An emerging technique in clustering, segmentation and classification problems is to consider the dynamics of flows defined on finite graphs. In particular Bertozzi and co-authors considered dynamics related to Allen-Cahn flow (Bertozzi, Flenner, 2012) and the MBO algorithm (Merkurjev, Kostic, Bertozzi, 2013) for this purpose.

This talk will exhibit our recent work showing rigorous links between these two flows, explaining why MBO can be used as an alternative to Allen-Cahn.

Abstract: An emerging technique in clustering, segmentation and classification problems is to consider the dynamics of flows defined on finite graphs. In particular Bertozzi and co-authors considered dynamics related to Allen-Cahn flow (Bertozzi, Flenner, 2012) and the MBO algorithm (Merkurjev, Kostic, Bertozzi, 2013) for this purpose.

This talk will exhibit our recent work showing rigorous links between these two flows, explaining why MBO can be used as an alternative to Allen-Cahn.

Tuesday, February 11, 2020

Thursday, February 13, 2020

16:00-17:00

Applied Mathematics Seminar

Guus Regts (UvA) - On the location of zeros of the independence for bounded degree graphs, HFG 611

see description

Title: On the location of zeros of the independence for bounded degree graphs

Abstract: In this talk I will introduce the independence polynomial (a.k.a. the partition function of the hard-core model in statistical physics) and motivate the study of the location of its zeros by applications to statistical physics and theoretical computer science.

Along the way I will indicate how the theory of complex dynamical systems has been used to settle a conjecture of Alan Sokal on the location of the zeros of the independence polynomial for bounded degree graphs. I will end with some open problems.

Based on joint works with Viresh Patel and Han Peters.

Abstract: In this talk I will introduce the independence polynomial (a.k.a. the partition function of the hard-core model in statistical physics) and motivate the study of the location of its zeros by applications to statistical physics and theoretical computer science.

Along the way I will indicate how the theory of complex dynamical systems has been used to settle a conjecture of Alan Sokal on the location of the zeros of the independence polynomial for bounded degree graphs. I will end with some open problems.

Based on joint works with Viresh Patel and Han Peters.

Tuesday, February 18, 2020

16:00-17:00

MIN 014

Thursday, February 20, 2020

16:00-17:00

Tuesday, February 25, 2020

Thursday, February 27, 2020

16:00-17:00

Tuesday, March 3, 2020

16:00-17:00

MIN 014

Thursday, March 5, 2020

Tuesday, March 10, 2020

16:00-17:00

MIN 014

Tuesday, March 17, 2020

16:00-17:00

MIN 014

Thursday, March 19, 2020

Tuesday, March 24, 2020

16:00-17:00

MIN 014

Thursday, March 26, 2020

Tuesday, March 31, 2020

Tuesday, April 7, 2020

Tuesday, April 14, 2020

Thursday, April 16, 2020

16:00-17:00

Tuesday, April 21, 2020

Thursday, April 23, 2020

16:00-17:00

Tuesday, April 28, 2020

Tuesday, May 5, 2020

Tuesday, May 12, 2020

Tuesday, May 19, 2020

Tuesday, May 26, 2020

Tuesday, June 2, 2020

Tuesday, June 9, 2020

Tuesday, June 16, 2020

Tuesday, June 23, 2020

Tuesday, June 30, 2020

Tuesday, July 7, 2020

Tuesday, July 14, 2020

Tuesday, July 21, 2020