Date/Time : 26 Aug 2025 10:00
Venue : S17 #04-06 (Seminar Room 1)
Speaker : Marcel Nutz, Columbia University

Entropic optimal transport — the optimal transport problem regularized by KL divergence — is highly successful in statistical applications. Thanks to the smoothness of the entropic coupling, its sample complexity avoids the curse of dimensionality suffered by unregularized optimal transport. The flip side of smoothness is overspreading: the entropic coupling always has full support, whereas the unregularized coupling that it approximates is usually sparse, even given by a map. Regularizing optimal transport by less-smooth divergences such as L^p regularization is known to allow for sparse approximations, but is often thought to suffer from the curse of dimensionality as the couplings have limited differentiability and the dual is not strongly concave. We refute this conventional wisdom and show, for a broad family of divergences, that the key empirical quantities converge at the parametric rate, independently of the dimension. More precisely, we provide central limit theorems for the optimal cost, the optimal coupling, and the dual potentials induced by i.i.d. samples from the marginals. (Joint work with Alberto Gonzalez-Sanz and Stephan Eckstein.)

Add to calendar :

Date/Time : 21 Aug 2025 TBA
Venue : TBA
Speaker : Liu Sheng, Stanford University

While modern machine learning models and systems, including most advanced large language models, have achieved impressive results, their success is still predicated on massive collections of human-curated, carefully filtered data. This reliance exposes a fundamental vulnerability: when deployed in the real world, these models encounter data that is noisy, incomplete, or ambiguous, leading to unpredictable and unreliable behavior. This challenge is more urgent in high-stakes domains such as medicine, where robustness and trustworthiness are paramount.
In this talk, I will present how we can move beyond the limitations of curated data by developing machine learning algorithms and theory that enable models to learn and adapt directly from the complexity of real-world environments. I will share new insights into the underlying mechanisms of failure for current models under imperfect data, and introduce principled methods that address these vulnerabilities. Finally, I will illustrate how these advances empower reliable AI systems for critical medical applications, including early Alzheimer’s disease detection and automated radiotherapy treatment planning.

Add to calendar :

Date/Time : 13 Aug 2025 15:00
Venue : S17 #04-06 (SR1)
Speaker : Bjoern Bringmann, Princeton University

Gibbs measures were first introduced in statistical mechanics, where they have been used to describe statistical equilibria of physical systems. Over the last decades, Gibbs measures have also become central objects in probability theory, partial differential equations, and mathematical physics. In this talk, we are interested in the heat, wave, and Schrödinger dynamics corresponding to Gibbs measures. In the first part of the talk, we discuss the invariance of the $\Phi^4_3$-measure under the three-dimensional cubic nonlinear wave equation. In the second part, we then discuss the global well-posedness of the stochastic Abelian-Higgs model, which is the Langevin equation for the Abelian-Higgs measure.

Add to calendar :

Date/Time : 12 Aug 2025 14:00
Venue : S17 #06-11 (Seminar Room 6)
Speaker : Bjoern Bringmann, Princeton University

There has been much recent progress on the local solution theory for geometric singular SPDEs. However, the global theory is still largely open. In this talk, we discuss the global well-posedness of the stochastic Abelian-Higgs model in two dimension, which is a geometric singular SPDE arising from gauge theory. The proof is based on a new covariant approach, which consists of two parts: First, we introduce covariant stochastic objects, which are controlled using covariant heat kernel estimates. Second, we control nonlinear remainders using a covariant monotonicity formula, which is inspired by earlier work of Hamilton.

Add to calendar :

Date/Time : 16 Jul 2025 15:00
Venue : S17 #04-04 (SR3)
Speaker : Naihuan Jing, North Carolina State University

I will discuss some polynomials parameterized by Young diagrams, whose coefficients are central elements of the quantized enveloping algebra of the general linear algebra. These polynomials are q-immanants and are shown to provide the higher quantum Capelli identities and generalize Drinfeld-Reshetikhin’s construction of central elements in Z(U_q(g)) as well as q-analogues of Okounkov’s quantum immanants for U(gl(n)). The Harish-Chandra image of the these central elements are certain factorial Schur polynomials. The main results are joint works with Ming Liu and Alex Molev.

Add to calendar :

Date/Time : 10 Jul 2025 15:00
Venue : S17 #04-04 (Seminar Room 3)
Speaker : Si-Jong Kwak, Korea Advanced Institute of Science & Technology

The gonality conjecture of a smooth projective curve raised by Green and Lazarsfeld. has been one of the leading problems in geometry and syzygy. The conjecture says that the gonality of a curve can be encoded in the Betti table if a given curve is embedded by the line bundle of sufficiently large degree. It was proved by Ein-Lazarsfeld in 2016.

In this talk, we generalize this conjecture to the secant varieties of a curve. The gonality sequence d_1, d_2, . . . , d_g are encoded in syzygies of the secant varieties of a curve. This is a joint work with J. Choe and J. Park.

Add to calendar :

Date/Time : 10 Jul 2025 10:00
Venue : S17 #04-04 (Seminar Room 3)
Speaker : Nimish Shah, The Ohio State University

For the space of unimodular lattices in a Euclidean space, we give necessary and sufficient conditions for equidistribution of expanding translates of any real-analytic submanifold under a diagonal flow. This extends the earlier result of Shah in the case of non-degenerate submanifolds.

We apply the above dynamical result to show that if the affine span of a real-analytic submanifold in a Euclidean space satisfies certain Diophantine and arithmetic conditions, then almost every point on the manifold is not Dirichlet-improvable.

This is a joint work with Pengyu Yang

Add to calendar :

Date/Time : 08 Jul 2025 13:30
Venue : S17 Level 3 (IMS Executive Seminar Room)
Speaker : Lou van den Dries, The University of Illinois Urbana-Champaign

This will be a survey of our work on these topics. No familiarity with transseries or Hardy fields is required, but I mention here that both are natural settings for solving ordinary differential equations and studying their asymptotics (and for much more). Transseries are formal objects, and Hardy fields consist of actual functions (or rather germs of such). Around 2015 we showed that the differential field of transseries has a good model theory. We went on to prove that sufficiently rich Hardy fields have the same (complete) model theory, with interesting consequences even on the level of linear differential equations. More recently we extended this even to analytic Hardy fields. Some of the ideas involved go back to Sturm and Liouville (1830’s).

This is joint work with Aschenbrenner and Van der Hoeven.

Add to calendar :

Date/Time : 04 Jul 2025 16:00
Venue : S17 #04-04 (Seminar Room 3)
Speaker : Ye Jieping, Alibaba

The rapid advancement of large language models (LLMs) showcases unprecedented capabilities, yet their opaque internal decision-making remains a critical challenge. Mechanistic interpretability has emerged as a critical frontier for enhancing the safety, reliability, and controllability of LLMs. In this talk, we synthesize cutting-edge research that reverse-engineers LLM internals to establish foundational principles and deployable applications. Through mechanistic analysis of diverse capabilities—including arithmetic computation, multilingual processing, and factual recall—we reveal how LLMs represent and process information as discrete circuits. These insights enable novel applications: a) Precise Control: Token-, activation-, neuron- and module-level intervention to adjust LLM’s behaviour; b) Precise Tuning: Head- and neuron-level tuning to enhance LLM’s performance c) Precise Quantization: Interpretability-guided quantization strategies achieving 3.4-bit parity with 4-bit baselines.

Add to calendar :