Seminars and Colloquia by Series

Finite-time Convergence Guarantees of Contractive Stochastic Approximation: Mean-Square and Tail Bounds

Series
Stochastics Seminar
Time
Thursday, October 5, 2023 - 15:30 for 1 hour (actually 50 minutes)
Location
Skile 006
Speaker
Siva Theja MaguluriGeorgia Tech

Abstract: Motivated by applications in Reinforcement Learning (RL), this talk focuses on the Stochastic Appproximation (SA) method to find fixed points of a contractive operator. First proposed by Robins and Monro, SA is a popular approach for solving fixed point equations when the information is corrupted by noise. We consider the SA algorithm for operators that are contractive under arbitrary norms (especially the l-infinity norm). We present finite sample bounds on the mean square error, which are established using a Lyapunov framework based on infimal convolution and generalized Moreau envelope. We then present our more recent result on concentration of the tail error, even when the iterates are not bounded by a constant. These tail bounds are obtained using exponential supermartingales in conjunction with the Moreau envelop and a novel bootstrapping approach. Our results immediately imply the state-of-the-art sample complexity results for a large class of RL algorithms.

Bio: Siva Theja Maguluri is Fouts Family Early Career Professor and Associate Professor in the H. Milton Stewart School of Industrial and Systems Engineering at Georgia Tech. He obtained his Ph.D. and MS in ECE as well as MS in Applied Math from UIUC, and B.Tech in Electrical Engineering from IIT Madras. His research interests span the areas of Control, Optimization, Algorithms and Applied Probability and include Reinforcement Learning theory and Stochastic Networks. His research and teaching are recognized through several awards including the  Best Publication in Applied Probability award, NSF CAREER award, second place award at INFORMS JFIG best paper competition, Student best paper award at IFIP Performance, CTL/BP Junior Faculty Teaching Excellence Award, and Student Recognition of Excellence in Teaching: Class of 1934 CIOS Award.

Convergence of Frame Series: from Hilbert Space to Modulation Space

Series
Analysis Seminar
Time
Wednesday, October 4, 2023 - 14:00 for 1 hour (actually 50 minutes)
Location
Skiles 005
Speaker
Pu-Ting YuGeorgia Tech

It is known that if $\{x_n\}$ is a frame for a separable Hilbert space, then there exist some sequences $\{y_n\}$ such that $x= \sum x_n$, and this sum converges in the norm of H. This equation is called the reconstruction formula of x. In this talk, we will talk about the existence of frames that admit absolutely convergent and unconditionally convergent reconstruction formula. Some characterizations of such frames will also be presented. Finally, we will present an extension of this problem about the unconditional convergence of Gabor expansion in Modulation spaces.

Enumerating Patterns in Social Networks - A Distribution-Free Model

Series
Graph Theory Seminar
Time
Tuesday, October 3, 2023 - 15:30 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Fan WeiDuke University

Fox et al introduced the model of c-closed graphs, a distribution-free model motivated by one of the most universal signatures of social networks, triadic closure. Even though enumerating maximal cliques in an arbitrary network can run in exponential time, it is known that for c-closed graph, enumerating maximal cliques and maximal complete bipartite graphs is always fast, i.e., with complexity being polynomial in the number of vertices in the network. In this work, we investigate further by enumerating maximal blow-ups of an arbitrary graph H in c-closed graphs. We prove that given any finite graph H, the number of maximal blow-ups of H in any c-closed graph on n vertices is always at most polynomial in n. When considering maximal induced blow-ups of a finite graph H, we provide a characterization of graphs H when the bound is always polynomial in n. A similar general theorem is also proved when H is infinite.

Balanced truncation for Bayesian inference

Series
Applied and Computational Mathematics Seminar
Time
Monday, October 2, 2023 - 14:00 for 1 hour (actually 50 minutes)
Location
Clough Commons 125 and https://gatech.zoom.us/j/98355006347
Speaker
Elizabeth QianSchool of Aerospace Engineering and School of Computational Science and Engineering at Georgia Tech

We consider the Bayesian approach to the linear Gaussian inference problem of inferring the initial condition of a linear dynamical system from noisy output measurements taken after the initial time. In practical applications, the large dimension of the dynamical system state poses a computational obstacle to computing the exact posterior distribution. Model reduction offers a variety of computational tools that seek to reduce this computational burden. In particular, balanced truncation is a control-theoretic approach to model reduction which obtains an efficient reduced-dimension dynamical system by projecting the system operators onto state directions which trade off the reachability and observability of state directions.  We define an analogous balanced truncation procedure for the Bayesian inference setting based on the trade off between prior uncertainty and data information. The resulting reduced model inherits desirable theoretical properties for both the control and inference settings: numerical demonstrations on two benchmark problems show that our method can yield near-optimal posterior covariance approximations with order-of-magnitude state dimension reduction.

Arnold diffusion in Hamiltonian systems with small dissipation

Series
CDSNS Colloquium
Time
Monday, October 2, 2023 - 14:00 for 1 hour (actually 50 minutes)
Location
In-person in Skiles 005
Speaker
Marian GideaYeshiva University

We consider a mechanical system consisting of a rotator and a pendulum, subject to a small, conformally symplectic perturbation. The resulting system has energy dissipation. We provide explicit conditions on the dissipation parameter, so that the resulting system exhibits Arnold diffusion. More precisely, we show that there are diffusing orbits along which the energy of the rotator grows by an amount independent of the smallness parameter. The fact that Arnold diffusion may play a role  in  systems with small dissipation was conjectured by Chirikov. Our system can be viewed as a simplified  model for an energy harvesting device, in which context the energy growth translates into generation of electricity.
Joint work with S.W. Akingbade and T-M. Seara.

The L^p metrics on Teichmüller space by Hannah Hoganson

Series
Geometry Topology Seminar
Time
Monday, October 2, 2023 - 14:00 for 1 hour (actually 50 minutes)
Location
Speaker
Hannah HogansonUMD

We will start by introducing the Teichmüller space of a surface, which parametrizes the possible conformal structures it supports. By defining this space analytically, we can equip it with the Lp metrics, of which the Teichmüller and Weil-Petersson metrics are special cases. We will discuss the incompleteness of the Lp metrics on Teichmüller space and what we know about their completions.

A stronger Torelli theorem for graphs

Series
Algebra Seminar
Time
Monday, October 2, 2023 - 13:00 for 1 hour (actually 50 minutes)
Location
Skiles 006
Speaker
Sarah GriffithBrown University

Please Note: There will be a pre-seminar (aimed toward grad students and postdocs) from 11:00 am-11:30 am in Skiles 006.

Recent research trends have explored curious analogies between the theory of graphs and Riemann surfaces. To each graph we can associate a real metric torus, known as its Jacobian. It was previously known that isomorphisms of graph Jacobians yield isomorphisms of the associated graphic matroids, partially mirroring a classical algebraic geometry result known as the Torelli theorem. However, the result on graphs is not as strong as a direct analogue of the Riemann surface result would be, nor does it use as much data. I will discuss how the graph Torelli theorem can be refined to incorporate additional data as with Riemann surfaces, in which case it produces isomorphisms of graphs. If time permits, I will describe further recent work in this direction.

Topological dynamics of knotted and tangled matter

Series
CDSNS Colloquium
Time
Friday, September 29, 2023 - 15:30 for 1 hour (actually 50 minutes)
Location
Skiles 249
Speaker
Vishal PatilStanford
Knots and tangles play a fundamental role in the dynamics of biological and physical systems, from DNA and root networks to surgical sutures and shoelaces. Despite having been studied for centuries, the subtle interplay between topology and mechanics in tangled elastic filaments remains poorly understood. Here we investigate the dynamical rules governing the behavior of knotted and tangled matter. We first study the human-designed knots used to tie ropes together. By developing an analogy with long-range ferromagnetic spin systems, we identify simple topological counting rules to predict the relative mechanical stability of commonly used climbing and sailing knots. Secondly, we examine the complex tangling dynamics exhibited by California blackworms, which form living tangled structures in minutes but can rapidly untangle in milliseconds. Using ultrasound imaging datasets, we construct a minimal model that explains how the kinematics of individual active filaments determines their emergent collective topological dynamics. By identifying generic dynamical principles of topological transformations, our results can provide guidance for designing classes of self-adaptive topological metamaterials.

 

 

 

 

Maximising copies of H in K_{r+1}-free graphs

Series
Combinatorics Seminar
Time
Friday, September 29, 2023 - 15:15 for 1 hour (actually 50 minutes)
Location
Skiles 308
Speaker
Natasha MorrisonUniversity of Victoria

Let H be a graph. We show that if r is large enough as a function of H,
then the r-partite Turán graph maximizes the number of copies of H among
all Kr+1-free graphs on a given number of vertices. This confirms a
conjecture of Gerbner and Palmer.

State Space Variance Ratio (SSVR) Test for Sequential Change Point Detection

Series
Mathematical Biology Seminar
Time
Friday, September 29, 2023 - 15:00 for 1 hour (actually 50 minutes)
Location
Skiles 005
Speaker
Vanja DukicUniversity of Colorado - Boulder

This talk will present a new online algorithm for sequential detection of change points in state-space models. The algorithm is computationally fast, and sensitive to changes in model parameters (including observation and evolution variances), as well as model structure. We consider change point detection in a sequential way, when observations are received one by one, or in batches, with a (possibly soft) restart after each detected change point. We provide the theoretical foundation of the algorithm, and study its performance in different state space models used to model the growth of epidemics over time, using simulated data and the recent COVID-19 dataset.  This work is joint work with Ruyu Tan.

This seminar is in a Hybrid format.  The in-person version is on campus at Georgia Tech in Skiles 005.  The virtual version will be at: https://gatech.zoom.us/j/92952024862

Pages