We consider a family of regression problems in a semi-supervised setting. Given real-valued labels on a small subset of data the task is to recover the function on the whole data set while taking advantage of the (geometric) structure provided by the large number of unlabeled data points. We consider a random geometric graph to represent the geometry of the data set. We study objective functions which reward the regularity of the estimator function and impose or reward the agreement with the training data. In particular we consider discrete p-Laplacian and fractional Laplacian regularizations.
We investigate asymptotic behavior in the limit where the number of unlabeled points increases while the number of training points remains fixed. We uncover a delicate interplay between the regularizing nature of the functionals considered and the nonlocality inherent to the graph constructions. We rigorously obtain almost optimal ranges on the scaling of the graph connectivity radius for the asymptotic consistency to hold. The talk is based on joint works with Matthew Dunlop, Andrew Stuart, and Matthew Thorpe.
Search
Upcoming Seminars
Mon Oct 16, 2017
We describe a number of related questions at the interface of set theory and homology theory, centering on (1) the additivity of strong homology, and (2) the cohomology of the ordinals. In the first, the question is, at heart: To how general a category of topological spaces may classical homology theory be continuously extended? And in the tension between various potential senses of continuity lie a number of delicate set-theoretic questions. These questions led to the consideration of the Cech cohomology of the ordinals; the surprise was that this is a meaningful thing to consider at all. It very much is, describing or suggesting at once (i) distinctive combinatorial principles associated to the nth infinite cardinal, for each n, holding in ZFC, (ii) rich connections between cofinality and dimension, and (iii) higher-dimensional extensions of the method of minimal walks.
Tue Oct 17, 2017
Modern financial portfolio construction uses mean-variance optimisation that requiers the knowledge of a very large covariance matrix. Replacing the unknown covariance matrix by the sample covariance matrix (SCM) leads to disastrous out-of-sample results that can be explained by properties of large SCM understood since Marcenko and Pastur. A better estimate of the true covariance can be built by studying the eigenvectors of SCM via the average matrix resolvent. This object can be computed using a matrix generalisation of Voiculescu’s addition and multiplication of free matrices. The original result of Ledoit and Peche on SCM can be generalise to estimate any rotationally invariant matrix corrupted by additive or multiplicative noise. Note that the level of rigor of the seminar will be that of statistical physics.
This is a joint probability/applied math seminar.
Modern financial portfolio construction uses mean-variance optimisation that requiers the knowledge of a very large covariance matrix. Replacing the unknown covariance matrix by the sample covariance matrix (SCM) leads to disastrous out-of-sample results that can be explained by properties of large SCM understood since Marcenko and Pastur. A better estimate of the true covariance can be built by studying the eigenvectors of SCM via the average matrix resolvent. This object can be computed using a matrix generalisation of Voiculescu’s addition and multiplication of free matrices. The original result of Ledoit and Peche on SCM can be generalise to estimate any rotationally invariant matrix corrupted by additive or multiplicative noise. Note that the level of rigor of the seminar will be that of statistical physics.
This is a joint applied math/probability seminar.
The general average distance problem, introduced by Buttazzo, Oudet, and Stepanov, asks to find a good way to approximate a high-dimensional object, represented as a measure, by a one-dimensional object. We will discuss two variants of the problem: one where the one-dimensional object is a measure with connected one-dimensional support and one where it is an embedded curve. We will present examples that show that even if the data measure is smooth the nonlocality of the functional can cause the minimizers to have corners. Nevertheless the curvature of the minimizer can be considered as a measure. We will discuss a priori estimates on the total curvature and ways to obtain information on topological complexity of the minimizers. We will furthermore discuss functionals that take the transport along the network into account and model best ways to design transportation networks. (Based on joint works with Xin Yang Lu and Slav Kirov.)
In 1934, Wilhelm Blaschke’s attention focused on a recent construction in metric geometry proposed by Dan Barbilian as a generalization of various models of hyperbolic geometry. It was the year when S.-S. Chern started his doctoral program under Blaschke’s supervision in Hamburg and when in several academic centers in Europe scholars were interested in generalizations of Riemannian geometry. Introduced originally in 1934, Barbilian’s metrization procedure induces a distance on a planar domain through a metric formula given by the so-called logarithmic oscillation. In 1959, Barbilian generalized this process to more general domains. In our discussion we plan to show that these spaces are naturally related to Gromov hyperbolic spaces. In several works written with W.G. Boskoff, we explore this connection. We conclude our talk by stating several open problems related to this content.
Wed Oct 18, 2017
Thu Oct 19, 2017
On a hyperelliptic curve over the rationals, there are infinitely many points defined over quadratic fields: just pull back rational points of the projective line through the degree two map. But for a positive proportion of genus g odd hyperelliptic curves, we show there can be at most 24 quadratic points not arising in this way. The proof uses tropical geometry work of Park, as well as results of Bhargava and Gross on average ranks of hyperelliptic Jacobians. This is joint work with Jackson Morrow.
Fri Oct 20, 2017
In this talk, we first discuss existence and uniqueness of weak solutions to general time fractional equations and give their probabilistic representation. We then talk about sharp two- sided estimates for fundamental solutions of general time fractional equations in metric measure spaces. This is a joint work with Zhen-Qing Chen(University of Washington, USA), Takashi Kumagai (RIMS, Kyoto University, Japan) and Jian Wang (Fujian Normal University, China).
In this talk, I will present the Kingman’s subadditive ergodic theorem, which is an extension of Birkhoff’s theorem.