Georgia Institute of Technology College of Sciences

In this presentation the analytical background of nonlinear observers based on minimal energy estimation is discussed. Originally, such strategies were proposed for the reconstruction of the state of finite dimensional dynamical systems by means of a measured output where both the dynamics and the output are subject to white noise. Our work aims at lifting this concept to a class of partial differential equations featuring deterministic perturbations using the example of a wave equation with a cubic defocusing term in three space dimensions. In particular, we discuss local regularity of the corresponding value function and consider operator Riccati equations to characterize its second spatial derivative.

Designing effective positional encodings for graphs is key to building powerful graph transformers and enhancing message-passing graph neural networks’ expressive power. However, since there lacks a canonical order of nodes in the graph-structure data, the choice of positional encodings for graphs is often tricky. For example, Laplacian eigenmap is used as positional encodings in many works. However, it faces two fundamental challenges: (1) Non-uniqueness: there are many different eigen-decompositions of the same Laplacian, and (2) Instability: small perturbations to the Laplacian could result in completely different eigenvectors, leading to unpredictable changes in positional encoding. This is governed by the Davis-Kahan theorem, which further negatively impacts the model generalization. In this talk, we are to introduce some ideas on building stable positional encoding and show their benefits in model out-of-distribution generalization. The idea can be extended to some other types of node positional encodings. Finally, we evaluate the effectiveness of our method on molecular property prediction, link prediction, and out-of-distribution generalization tasks, finding improved generalization compared to existing positional encoding methods.

I will mainly talk about three papers:

1. Distance Encoding: Design Provably More Powerful Neural Networks for Graph Representation Learning,NeurIPS20, Pan Li, Yanbang Wang, Hongwei Wang, Jure Leskovec

2. Equivariant and Stable Positional Encoding for More Powerful Graph Neural Networks, ICLR22 Haorui Wang, Haoteng Yin, Muhan Zhang, Pan Li

3. On the Stability of Expressive Positional Encodings for Graphs, ICLR24 Yinan Huang, William Lu, Joshua Robinson, Yu Yang, Muhan Zhang, Stefanie Jegelka, Pan Li

When studying symplectic 4-manifolds, it is useful to consider Lefschetz fibrations over the 2-sphere due to their one-to-one correspondence uncovered by Freedman and Gompf. Lefschetz fibrations of genera 0 and 1 are well understood, but for genera greater than or equal to 2, much less is known. However, some Lefschetz fibrations with monodromies that respect the hyperelliptic involution of a genus-g surface have stronger properties which make their invariants easier to compute. In this talk, we will explore Terry Fuller's results from the late 90's which explore two-fold branched covers of hyperelliptic genus-g Lefschetz fibrations. We will look at his proof of why a Lefschetz fibration with only nonseparating vanishing cycles is a two-fold cover of $S^2 \times S^2$ branched over an embedded surface. The talk will include definitions, constructions, and Kirby pictures of branched covers in 4 dimensions. If time, we will discuss his results on hyperelliptic genus-g Lefschetz fibration which contain at least one separating vanishing cycles.