Georgia Institute of Technology College of Sciences

We present new methods for segmentation of large datasets with graph based structure. The method combines ideas from classical nonlinear PDE-based image segmentation with fast and accessible linear algebra methods for computing information about the spectrum of the graph Laplacian. The goal of the algorithms is to solve semi-supervised and unsupervised graph cut optimization problems. I will present results for image processing applications such as image labeling and hyperspectral video segmentation, and results from machine learning and community detection in social networks, including modularity optimization posed as a graph total variation minimization problem.

PDEs (such as Navier-Stokes) are in principle infinite-dimensional dynamical systems. However, recent studies support conjecture that the turbulent solutions of spatially extended dissipative systems evolve within an `inertial' manifold spanned by a finite number of 'entangled' modes, dynamically isolated from the residual set of isolated, transient degrees of freedom. We provide numerical evidence that this finite-dimensional manifold on which the long-time dynamics of a chaotic dissipative dynamical system lives can be constructed solely from the knowledge of a set of unstable periodic orbits. In particular, we determine the dimension of the inertial manifold for Kuramoto-Sivashinsky system, and find it to be equal to the `'physical dimension' computed previously via the hyperbolicity properties of covariant Lyapunov vectors. (with Xiong Ding, H. Chate, E. Siminos and K. A. Takeuchi)

Accounting for Heterogenous Interactions in the Spread Infections, Failures, and Behaviors_ The scaled SIS (susceptible-infected-susceptible) network process that we introduced extends traditional birth-death process by accounting for heterogeneous interactions between individuals. An edge in the network represents contacts between two individuals, potentially leading to contagion of a susceptible by an infective. The inclusion of the network structure introduces combinatorial complexity, making such processes difficult to analyze. The scaled SIS process has a closed-form equilibrium distribution of the Gibbs form. The network structure and the infection and healing rates determine susceptibility to infection or failures. We study this at steady-state for three scales: 1) characterizing susceptibility of individuals, 2) characterizing susceptibility of communities, 3) characterizing susceptibility of the entire population. We show that the heterogeneity of the network structure results in some individuals being more likely to be infected than others, but not necessarily the individuals with the most number of interactions (i.e., degree). We also show that "densely connected" subgraphs are more vulnerable to infections and determine when network structures include these more vulnerable communities.

The commutator length of an element g in the commutator subgroup [G,G] of agroup G is the smallest k such that g is the product of k commutators. WhenG is the fundamental group of a topological space, then the commutatorlength of g is the smallest genus of a surface bounding a homologicallytrivial loop that represents g. Commutator lengths are notoriouslydifficult to compute in practice. Therefore, one can ask for asymptotics.This leads to the notion of stable commutator length (scl) which is thespeed of growth of the commutator length of powers of g. It is known thatfor n > 2, SL(n,Z) is uniformly perfect; that is, every element is aproduct of a bounded number of commutators, and hence scl is 0 on allelements. In contrast, most elements in SL(2,Z) have positive scl. This isrelated to the fact that SL(2,Z) acts naturally on a tree (its Bass-Serretree) and hence has lots of nontrivial quasimorphisms.In this talk, I will discuss a result on the stable commutator lengths inright-angled Artin groups. This is a broad family of groups that includesfree and free abelian groups. These groups are appealing to work withbecause of their geometry; in particular, each right-angled Artin groupadmits a natural action on a CAT(0) cube complex. Our main result is anexplicit uniform lower bound for scl of any nontrivial element in anyright-angled Artin group. This work is joint with Talia Fernos and MaxForester.