Date/Time : 24 Jun 2026 11:00
Venue : S17 #06-11 (Seminar Room 6)
Speaker : Xiuyuan Cheng, Duke University

Wasserstein regularization has become a common tool in learning problems over distributions, while its practical computation remains challenging, especially in high dimensions. This talk presents a framework for Wasserstein-regularized learning based on directly parameterizing transport maps with neural networks. Rather than working with dual potentials or entropic relaxations, we explicitly model the transport map, which enables principled optimization in Wasserstein space as well as scalable out-of-sample evaluation. One example is the implementation of Wasserstein-2 proximal steps via flow networks that realize the Jordan–Kinderlehrer–Otto (JKO) scheme, yielding a variational interpretation of flow-based generative models. The talk will also discuss a minimax regime motivated by distributionally robust optimization (DRO), where transport is chosen adversarially. In this setting, neural transport maps provide a practical framework for studying Wasserstein minimax problems and representing worst-case distributions. Together, these examples highlight the flexibility of transport-map formulations for incorporating Wasserstein geometry into learning problems over distributions.

Add to calendar :

Date/Time : 29 May 2026 16:00
Venue : S17 #04-05 (Seminar Room 2)
Speaker : Yufeng Wu, University of Connecticut

Single cell sequencing technologies are generating DNA sequences for individual cells of an organism. Such data enables novel study of single cell evolution. Cell lineage tree is an evolutionary model for single cells, where the leaves of the tree are the extant cells and the internal nodes are the progenitor cells. Reconstructing cell lineage tree from single cell data has been actively studied recently. The main computational challenges for cell lineage tree inference are the inherent noises in single cell data and also the size of the data. In this talk, I will first introduce perfect phylogeny, a simple evolutionary tree model that has been used extensively in cancer and single cell evolution. Then I will present my own work ScisTree, a probabilistic cell lineage tree inference method based on the perfect phylogeny model. Then, I will focus on presenting ScisTree2, a state-of-the-art cell lineage tree inference method that is orders of magnitude faster than ScisTree. I will show some results of ScisTree2 on simulated and biological data.

Add to calendar :

Date/Time : 17 Mar 2026 15:00
Venue : IMS Executive Seminar Room (S17 Level 3)
Speaker : Mihai Anitescu, Argonne National Laboratory and University of Chicago

Since 2018, with several co-authors, we have pursued a research program to explain a phenomenon first observed in long-horizon optimal control problems. That is when the data of the problem is perturbed at a certain temporal index, the effect of the perturbation on the solution at a different temporal index decays exponentially fast with respect to the distance between the two. We have shown, that, under uniform second order and linear independence constraint qualifications this phenomenon holds and we extended this proof for spatially-indexed, and more generally, graph-indexed problems. In turn, this allows proving that Schwarz-type decompositions with overlap used as either preconditioners or as solution schemes improve convergence exponentially fast with the size of the overlap. We discuss the consequences of this observation for design of efficient parallel algorithms and decentralized optimization schemes and present a few of their applications.

Add to calendar :

Date/Time : 10 Feb 2026 16:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Jean Lasserre, Laboratory for Analysis and Architecture of Systems, CNRS, France

In this talk we propose and describe a hierarchy of semidefinite relaxations that approximate as closely as desired the total variation distance between two given measures $\mu$ and $\nu$, from the sole knowledge of their respective moments. We only assume that both satisfy Carleman’s condition (hence with possibly non-compact support). We obtain a sequence of pseudo-moments that converges to moments of the Hahn-Jordan decomposition of $\mu-\nu$. We then apply this technique to the problem of computing mixtures of Gaussians that minimize the TV-distance to a given measure.

Add to calendar :

Date/Time : 04 Feb 2026 13:30
Venue : S16 #03-09
Speaker : Minwoo Chae, Pohang University of Science and Technology

In recent years, diffusion-based deep generative models have achieved remarkable success in various applications. In this talk, we present statistical theories for diffusion models within the framework of nonparametric structured density estimation. To address the curse of dimensionality in nonparametric density estimation, we assume that the underlying density function factorizes into several low-dimensional components. Such factorizable densities are common in important examples, such as Bayesian networks and Markov random fields. We prove that an implicit density estimator constructed from diffusion models achieves the minimax optimal convergence rate with respect to total variation. Technically, we design a novel network architecture, which includes convolutional neural networks as a special case, to construct a minimax optimal estimator.

Add to calendar :

Date/Time : 30 Jan 2026 16:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Soon Hoe Lim, KTH

Flow matching has emerged as a powerful paradigm for generative modeling. From the viewpoint of dynamical measure transport, it constructs continuous-time ODE samplers by prescribing probability paths that connect a base and a target distribution. In this talk, we discuss both the potential and pitfalls of flow matching for probabilistic forecasting of dynamical systems.

We observe that forecasting performance is highly sensitive to the choice of probability path, motivating principled constructions that lead to improved computational efficiency and predictive performance on spatio-temporal dynamical system benchmarks. We then revisit flow matching in the empirical setting and analyze the structure of the induced velocity fields, uncovering connections between flow matching, memory effects, and nonparametric dynamical systems. This perspective leads to new sampling strategies and raises questions about the role of parameterization in learning dynamical systems from data.

Add to calendar :

Date/Time : 27 Jan 2026 16:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Ivan Au Yeung , Nvidia

This talk explores solving nonlinear p-Laplacian problems using NVIDIA PhysicsNeMo, a GPU-accelerated framework for building and training physics AI models at scale. The p-Laplacian is motivated through applications such as nonlinear diffusion and variational problems, with emphasis on the numerical difficulties that arise for large values of p and in complex 2D and 3D geometries. The core of the talk is a PhysicsNeMo-based workflow that recasts the p-Laplacian into Physics-Informed Neural Network and operator-learning formulations, enabling end-to-end training directly from PDE specifications and boundary conditions while leveraging multi-GPU hardware. Numerical experiments compare these neural solvers with traditional mesh-based methods in terms of accuracy, robustness for high p. The talk concludes with practical guidelines for setting up p-Laplacian problems and tuning architectures and loss terms within PhysicsNeMo.

Add to calendar :

Date/Time : 27 Jan 2026 15:00
Venue : Conference room (S17 Level 4)
Speaker : Deborah Hendrych, Zuse Institute Berlin

Mixed Integer Nonlinear Problem (MINLP) solvers have traditionally relied on Outer Approximation schemes. Boscia, a recent framework, takes a different approach by leveraging the Frank-Wolfe algorithm—a first-order method for convex optimization—within a branch-and-bound structure. This approach eliminates the need for additional cutting planes while maintaining the oracle-based flexibility of Frank-Wolfe and capitalizing on recent advances in mixed-integer programming solvers. This talk present the intuition behind the framework. highlights the main advantages and showcase the usage and modelling through two examples: portfolio optimization and D-optimal experiment design.

Add to calendar :

Date/Time : 26 Jan 2026 15:00
Venue : S17 #05-11 (Seminar Room 5)
Speaker : Ernest Tan, National University of Singapore

Research in quantum information theory, particularly for quantum cryptography or quantum computing, often makes use of numerical methods and software for linear algebra. In earlier literature, most work was based around well-established methods such as semidefinite programming. However, more recent developments have highlighted the need to handle broader classes of convex optimization problems, involving a family of generalizations of the Kullback-Leibler divergence or relative entropy, for which many questions remain open. Furthermore, results from convex analysis have also proven useful in deriving powerful techniques for statistical estimation in quantum information, connected to large-deviation theory. I will present an introductory overview of some active areas of research and open problems in the field, in particular the areas that could benefit from further insights in convex optimization.

Add to calendar :