Calendar

We study the supersymmetric partition function of a 2d linear sigma-model whose target space is a torus with a complex structure that varies along one worldsheet direction and a Kähler modulus that varies along the other. This setup is inspired by the dimensional reduction of a Janus configuration of 4d N=4 U(1) Super-Yang-Mills theory compactified on a mapping torus (T^2 fibered over S^1) times a circle with an SL(2,Z) duality wall inserted on S^1, but our setup has minimal supersymmetry. The partition function depends on two independent elements of SL(2,Z), one describing the duality twist, and the other describing the geometry of the mapping torus. It is topological and can be written as a multivariate quadratic Gauss sum. By calculating the partition function in two different ways, we obtain identities relating different quadratic Gauss sums, generalizing the Landsberg-Schaar relation. These identities are a subset of a collection of identities discovered by F. Deloup. Each identity contains a phase which is an eighth root of unity, and we show how it arises as a Berry phase in the supersymmetric Janus-like configuration. Supersymmetry requires the complex structure to vary along a semicircle in the upper half-plane, as shown by Gaiotto and Witten in a related context, and that semicircle plays an important role in reproducing the correct Berry phase.

Zoom: https://harvard.zoom.us/j/91780604388?pwd=d3BqazFwbDZLQng0cEREclFqWkN4UT09

Zoom: https://harvard.zoom.us/j/779283357?pwd=MitXVm1pYUlJVzZqT3lwV2pCT1ZUQT09

Zoom: https://harvard.zoom.us/j/779283357?pwd=MitXVm1pYUlJVzZqT3lwV2pCT1ZUQT09

Least-mean squares (LMS) solvers such as Linear / Ridge / Lasso-Regression, SVD and Elastic-Net not only solve fundamental machine learning problems, but are also the building blocks in a variety of other methods, such as decision trees and matrix factorizations.

We suggest an algorithm that gets a finite set of $n$ $d$-dimensional real vectors and returns a weighted subset of $d + 1$ vectors whose sum is exactly the same. The proof in Caratheodory’s Theorem (1907) computes such a subset in $O(n^2 d^2 )$ time and thus not used in practice. Our algorithm computes this subset in $O(nd)$ time, using $O(logn)$ calls to Caratheodory’s construction on small but “smart” subsets. This is based on a novel paradigm of fusion between different data summarization techniques, known as sketches and coresets.

As an example application, we show how it can be used to boost the performance of existing LMS solvers, such as those in scikit-learn library, up to $x100$. Generalization for streaming and distributed (big) data is trivial. Extensive experimental results and complete open source code are also provided.

Zoom: https://harvard.zoom.us/j/98231541450

Determining the computational complexity of matrix multiplication has been one of the central open problems in theoretical computer science ever since in 1969 Strassen presented an algorithm for multiplication of n by n matrices requiring only O(n^2.81) arithmetic operations. I will briefly discuss this problem and its reduction to deciding on which secant variety to the Segre embedding of a product of three projective spaces the matrix multiplication tensor lies. I will explain a recent technique to rule out membership of a fixed tensor in such secant varieties, border apolarity. Border apolarity establishes the existence of certain multigraded ideals implied by membership in a particular secant variety. These ideals may be assumed to be fixed under a Borel subgroup of the group of symmetries of the tensor, and in the simplest case, can consequently be tractably shown not to exist. When ideals exist satisfying the easily checkable properties, one must decide if they are limits of ideals of distinct points on the Segre. This talk discusses joint work with JM Landsberg, Alicia Harper, and Amy Huang.

Zoom: https://harvard.zoom.us/j/91794282895?pwd=VFZxRWdDQ0VNT0hsVTllR0JCQytoZz09