Events

Fragmented Bose-Einstein condensates are large systems of identical bosons displaying multiple macroscopic occupations of one-body states, in a suitable sense. The quest for an effective dynamics of the fragmented condensate at the leading order in the number of particles, in analogy to the much more controlled scenario for complete condensation in one single state, is deceptive both because characterising fragmentation solely in terms of reduced density matrices is unsatisfactory and ambiguous, and because as soon as the time evolution starts the rank of the reduced marginals generically passes from finite to infinite, which is a signature of a transfer of occupations on infinitely many more one-body states. In this work we review these difficulties, we refine previous characterisations of fragmented condensates in terms of marginals, and we provide a quantitative rate of convergence to the leading effective dynamics in the double limit of infinitely many particles and infinite energy gap. This is a joint work with Alessandro Michelangeli.

In this talk, we will present a framework for analyzing dynamics of stochastic optimization algorithms (e.g., stochastic gradient descent (SGD) and momentum (SGD+M)) when both the number of samples and dimensions are large. For the analysis, we will introduce a stochastic differential equation, called homogenized SGD. We show that homogenized SGD is the high-dimensional equivalent of SGD -- for any quadratic statistic (e.g., population risk with quadratic loss), the statistic under the iterates of SGD converges to the statistic under homogenized SGD when the number of samples n and number of features d are polynomially related. By analyzing homogenized SGD, we provide exact non-asymptotic high-dimensional expressions for the training dynamics and generalization performance of SGD in terms of a solution of a Volterra integral equation. The analysis is formulated for data matrices and target vectors that satisfy a family of resolvent conditions, which can roughly be viewed as a weak form of delocalization of sample-side singular vectors of the data. By analyzing these limiting dynamics, we can provide insights into learning rate, momentum parameter, and batch size selection. For instance, we identify a stability measurement, the implicit conditioning ratio (ICR), which regulates the ability of SGD+M to accelerate the algorithm. When the batch size exceeds this ICR, SGD+M converges linearly at a rate of $O(1/ \kappa)$, matching optimal full-batch momentum (in particular performing as well as a full-batch but with a fraction of the size). For batch sizes smaller than the ICR, in contrast, SGD+M has rates that scale like a multiple of the single batch SGD rate. We give explicit choices for the learning rate and momentum parameter in terms of the Hessian spectra that achieve this performance. Finally we show this model matches performances on real data sets.

Random matrices frequently appear in many different fields — physics, computer science, applied and pure mathematics. Oftentimes the random matrix of interest will have non-​trivial structure — entries that are dependent and have potentially different means and variances (e.g. sparse Wigner matrices, matrices corresponding to adjacencies of random graphs, sample covariance matrices). However, current understanding of such complex random matrices remains lacking. In this talk, I will discuss recent results concerning the spectrum of sums of independent random matrices with a.s. bounded operator norms. In particular, I will demonstrate that under some fairly general conditions, such sums will exhibit the following universality phenomenon — their spectrum will lie close to that of a Gaussian random matrix with the same mean and covariance. No prior background in random matrix theory is required — basic knowledge of probability and linear algebra are sufficient. (joint with Ramon van Handel) Pre-​print link: https://web.math.princeton.edu/~rvan/tuniv220113.pdf

In this talk, we will present a framework for analyzing dynamics of stochastic optimization algorithms (e.g., stochastic gradient descent (SGD) and momentum (SGD+M)) when both the number of samples and dimensions are large. For the analysis, we will introduce a stochastic differential equation, called homogenized SGD. We show that homogenized SGD is the high-dimensional equivalent of SGD -- for any quadratic statistic (e.g., population risk with quadratic loss), the statistic under the iterates of SGD converges to the statistic under homogenized SGD when the number of samples n and number of features d are polynomially related. By analyzing homogenized SGD, we provide exact non-asymptotic high-dimensional expressions for the training dynamics and generalization performance of SGD in terms of a solution of a Volterra integral equation. The analysis is formulated for data matrices and target vectors that satisfy a family of resolvent conditions, which can roughly be viewed as a weak form of delocalization of sample-side singular vectors of the data. By analyzing these limiting dynamics, we can provide insights into learning rate, momentum parameter, and batch size selection. For instance, we identify a stability measurement, the implicit conditioning ratio (ICR), which regulates the ability of SGD+M to accelerate the algorithm. When the batch size exceeds this ICR, SGD+M converges linearly at a rate of $O(1/ \kappa)$, matching optimal full-batch momentum (in particular performing as well as a full-batch but with a fraction of the size). For batch sizes smaller than the ICR, in contrast, SGD+M has rates that scale like a multiple of the single batch SGD rate. We give explicit choices for the learning rate and momentum parameter in terms of the Hessian spectra that achieve this performance. Finally we show this model matches performances on real data sets.

In this talk, we will present a framework for analyzing dynamics of stochastic optimization algorithms (e.g., stochastic gradient descent (SGD) and momentum (SGD+M)) when both the number of samples and dimensions are large. For the analysis, we will introduce a stochastic differential equation, called homogenized SGD. We show that homogenized SGD is the high-​dimensional equivalent of SGD -- for any quadratic statistic (e.g., population risk with quadratic loss), the statistic under the iterates of SGD converges to the statistic under homogenized SGD when the number of samples n and number of features d are polynomially related. By analyzing homogenized SGD, we provide exact non-​asymptotic high-​dimensional expressions for the training dynamics and generalization performance of SGD in terms of a solution of a Volterra integral equation. The analysis is formulated for data matrices and target vectors that satisfy a family of resolvent conditions, which can roughly be viewed as a weak form of delocalization of sample-​side singular vectors of the data. By analyzing these limiting dynamics, we can provide insights into learning rate, momentum parameter, and batch size selection. For instance, we identify a stability measurement, the implicit conditioning ratio (ICR), which regulates the ability of SGD+M to accelerate the algorithm. When the batch size exceeds this ICR, SGD+M converges linearly at a rate of $O(1/ \kappa)$, matching optimal full-​batch momentum (in particular performing as well as a full-​batch but with a fraction of the size). For batch sizes smaller than the ICR, in contrast, SGD+M has rates that scale like a multiple of the single batch SGD rate. We give explicit choices for the learning rate and momentum parameter in terms of the Hessian spectra that achieve this performance. Finally we show this model matches performances on real data sets.

In this talk we will discuss some extremal problems related to embeddings between weighted Paley–Wiener spaces. We will present some asymptotic results for sharp constants in terms of the parameters involved, deduce existence results for extremal functions as well as radial symmetry of those. For certain cases, these extremal problems can be reformulated in terms of sharp Poincaré inequalities, and for those cases we will present a characterisation of extremizers and sharp constants that recover several classical results.

I will present new results on the asymptotic growth rate of the Euler characteristic of Kontsevich's commutative graph complex. By a work of Chan, Galatius and Payne, these results imply the same asymptotic growth rate for the top-weight Euler characteristic of M_g, the moduli space of curves, and establish the existence of large amounts of unexplained top-weight cohomology in this space. I will illustrate the ingredients for the proof and comment on related recent work with Karen Vogtmann on Out(F_n) and the moduli space of graphs.

The Fyodorov-Hiary-Keating conjecture has two parts, one in random matrix theory and one about the Riemann zeta function. In the random matrix part, it gives the precise distributional limit for the maximum of a characteristic polynomial of a Haar Unitary matrix. Using the replica method and a physically motivated `freezing’ ansatz, they derived one of the most precise log-correlated field predictions todate, and they did it for a process which was not even Gaussian. While existing work shows that Haar Unitary matrices had many log correlated field connections, techniques for showing convergence of the maximum typically rely on either the Gaussianity of the underlying process or precise branching structures built into the problem; the characteristic polynomial has neither. We will describe the problem and the current state of the art, in which we (the speaker and Ofer Zeitouni) show the convergence in law of the maximum of a Circular-beta ensemble random matrix to a convolution of a gumbel and the total mass of a (non-Gaussian) critical multiplicative chaos.

Hyperbolic groups were introduced in the 80's by M.Gromov as a generalization of fundamental groups of Riemannian manifold with non-positive curvature. These groups form a large class of finitely generated groups with lots of good combinatorial properties. Property (T) was introduced in the 60's by D.Kahzdan to prove that a large class of lattices are finitely generated. This is a rigidity property, this means that if a group has this property, actions on some particular metric spaces have to be trivial. For example, actions on trees have a fixed point. I will discuss these two notions and links between them. Time permitting, I will talk about a strengthening of Property (T) called Strong Property (T).

It was known since 1960s (Novikov, Mischenko) that the logarithm of the formal group in complex cobordisms can be written explicitly in terms of the complex projective spaces, but the algebro-geometric nature of the coefficients of the corresponding exponential was not clear until recently. In the talk I will explain that the answer can be given by the smooth theta divisors of principally polarised abelian varieties. It will be shown that the topological characteristics of the theta divisors and their intersections can be expressed in terms of the combinatorics of permutohedra. We reveal also interesting relations between the theta divisors, permutohedral varieties and Tomei manifolds from the Toda lattice theory. The talk is based on joint work with V.M. Buchstaber.

We consider a company who models the dependence of its surplus - a Brownian motion with drift - on business cycles by a Markov chain. Depending on the chosen target functional and the impact of the control on the surplus, the value function and the optimal strategy can be found explicitly, recursively or not at all. The constant "once and forever" strategies, optimal for the case of only one state, turn out to be suboptimal in most multi-state problems. In this talk, we consider some "explicit"- and "recursive"-solution cases. As a pudding, we look at the "beyond" case, where the model features a stochastic continuous discounting rate - an infinite-regime setting.

By a classical theorem of Denjoy, any sufficiently regular piece-wise smooth circle homeomorphism with finitely many branches (often called a circle homeomorphism with breaks) and irrational rotation number is topologically conjugated to an irrational circle rotation. In particular, it admits a unique invariant probability measure. We will discuss dimensional properties of this measure and show that, generically, this unique invariant probability measure has zero Hausdorff dimension. To encode this generic condition, we consider piece-wise smooth homeomorphisms as generalized interval exchange transformations of the interval and rely on the notion of combinatorial rotation number.