arXiv research

A modified SPA preconditioner enhances noise robustness in separable NMFs.

Let $X=G/K$ be a symmetric space of noncompact type and rank $k\ge 2$. We prove that horospheres in $X$ are Lipschitz $(k-2)$--connected if their centers are not contained in a proper join factor of the spherical building of $X$ at infinity. As a consequence, the distortion dimension of an irreducible $\mathbb{Q}$--ran…

Study financial contagion in networks using low-rank approximations and graphons.

We study streaming principal component analysis (PCA), that is to find, in $O(dk)$ space, the top $k$ eigenvectors of a $d\times d$ hidden matrix $\bf Σ$ with online vectors drawn from covariance matrix $\bf Σ$. We provide $\textit{global}$ convergence for Oja's algorithm which is popularly used in practice but lacks t…

Efficient knockoffs for large-scale feature selection.

We consider the pseudo-Anosov elements of the mapping class group of a surface of genus g that fix a rank k subgroup of the first homology of the surface. We show that the smallest entropy among these is comparable to (k+1)/g. This interpolates between results of Penner and of Farb and the second and third authors, who…

We develop an efficient alternating framework for learning a generalized version of Factorization Machine (gFM) on steaming data with provable guarantees. When the instances are sampled from $d$ dimensional random Gaussian vectors and the target second order coefficient matrix in gFM is of rank $k$, our algorithm conve…

In this paper, we study entire translating solutions $u(x)$ to a mean curvature flow equation in Minkowski space. We show that if $Σ=\{(x, u(x))| x\in\mathbb{R}^n\}$ is a strictly spacelike hypersurface, then $Σ$ reduces to a strictly convex rank k soliton in $\mathbb{R}^{k, 1}$ (after splitting off trivial factors) wh…

Truncated Singular Value Decomposition (SVD) calculates the closest rank-$k$ approximation of a given input matrix. Selecting the appropriate rank $k$ defines a critical model order choice in most applications of SVD. To obtain a principled cut-off criterion for the spectrum, we convert the underlying optimization prob…

RPCholesky approximates kernel matrices with few evaluations.

Solves Kobayashi's conjecture on homogeneous spaces.

We show that for k at least 3, given any matrix in GL(k,Z), there is a hyperbolic fully irreducible automorphism of the free group of rank k whose induced action on Z^k is the given matrix.

A martingale \int H.dZ is defined as having Dimension k if H has rank k almost surely, almost all t. Dimension can be used as a geometric invariant to classify and study martingales. We also define general Brownian motions in higher dimensions.

Optimizes FM model complexity for better feature interaction learning.

Improved Frank-Wolfe algorithm solves convex trace-norm ball problems.

The paper tackles robust submodular maximization under matroid constraints, providing approximation algorithms for summary extraction.

Given a free group $F_k$ of rank $k\ge 2$ with a fixed set of free generators we associate to any homomorphism $φ$ from $F_k$ to a group $G$ with a left-invariant semi-norm a generic stretching factor, $λ(φ)$, which is a non-commutative generalization of the translation number. We concentrate on the situation when $φ:F…

Hyperbolic space outperforms Euclidean in learning hierarchical data.

We show that every automorphism $α$ of a free group $F_k$ of finite rank $k$ has {\it asymptotically periodic} dynamics on $F_k$ and its boundary $\partial F_k$: there exists a positive power $α^q$ such that every element of the compactum $F_k \cup \partial F_k$ converges to a fixed point under iteration of $α^q$.

The paper develops algorithms to find a robust summary of data under deletion, achieving good approximation guarantees.

New framework for consistent submodular maximization with insertions and deletions.

Maximizes determinant of vector sums under matroid constraints.

New algorithms minimize non-zero entries in low-rank approximations.

Regular integer lattices are characterized by k unit vectors that build up their generator matrices. These have rank k for D-lattices, and are rank-deficient for A-lattices, for E_6 and E_7. We count lattice points inside hypercubes centered at the origin for all three types, as if classified by maximum infinity norm i…

The moduli space of solutions to Nahm's equations of rank (k,k+j) on the circle, and hence, of SU(2) calorons of charge (k,j), is shown to be equivalent to the moduli of holomorphic rank 2 bundles on P^1xP^1 trivialized at infinity with c_2=k and equipped with a flag of degree j along P^1x{0}. An explicit matrix descri…

Four conditions ensure low-rank projection costs using random matrix tricks.

Sketchy reduces memory and compute requirements for adaptive regularization in deep learning.

Paper proposes a new method to improve classification performance over PCA.

We study the column subset selection problem with respect to the entrywise $\ell_1$-norm loss. It is known that in the worst case, to obtain a good rank-$k$ approximation to a matrix, one needs an arbitrarily large $n^{Ω(1)}$ number of columns to obtain a $(1+ε)$-approximation to the best entrywise $\ell_1$-norm low ra…

In this paper we study regular irreducible algebraic monoids over $\fldc$ equipped with the euclidean topology. It is shown that, in such monoids, the Green classes and the spaces of idempotents in the Green classes all have natural manifold structures. The interactions of these manifold structures and the semigroup st…

For any complex vector bundle $E^k$ of rank $k$ over a manifold $M^m$ with Chern classes $c_i \in H^{2i}(M^m,\Z)$ and any non-negative integers $l_1, >..., l_k$ we show the existence of a positive number $N(k,m)$ and the existence of a complex vector bundle $\hat E^k$ over $M^m$ whose Chern classes are $ N(k,m) \cdot l…

Improved guarantees and multiple-descent curve for data approximations.

Study of correlated Wigner matrices with BBP transitions.

Most of machine learning deals with vector parameters. Ideally we would like to take higher order information into account and make use of matrix or even tensor parameters. However the resulting algorithms are usually inefficient. Here we address on-line learning with matrix parameters. It is often easy to obtain onlin…

Study the spectral flow of Dirac operators on spinor bundles.

Study growth of systoles in arithmetic manifolds, focusing on $k$-dimensional cases.

Spatial Adapter adds structured spatial representation to frozen predictors.

Motivated by Bonahon's result for hyperbolic surfaces, we construct an analogue of the Patterson-Sullivan-Bowen-Margulis map from the Culler-Vogtmann outer space $CV(F_k)$ into the space of projectivized geodesic currents on a free group. We prove that this map is a topological embedding. We also prove that for every $…

Algorithm learns a better sketch matrix for low-rank approximations.

For certain manifolds, nonnegative Ricci curvature limits dimension and forces almost abelian fundamental group.

When data is sampled from an unknown subspace, principal component analysis (PCA) provides an effective way to estimate the subspace and hence reduce the dimension of the data. At the heart of PCA is the Eckart-Young-Mirsky theorem, which characterizes the best rank k approximation of a matrix. In this paper, we prove …

The study finds stably free modules and distinct 2-complexes for large ranks.

We prove a rigidity theorem for the geometry of the unit ball in random subspaces of the scl norm in B_1^H of a free group. In a free group F of rank k, a random word w of length n (conditioned to lie in [F,F]) has scl(w)=log(2k-1)n/6log(n) + o(n/log(n)) with high probability, and the unit ball in a subspace spanned by…

Matroid bundles, introduced by MacPherson, are combinatorial analogues of real vector bundles. This paper sets up the foundations of matroid bundles, and defines a natural transformation from isomorphism classes of real vector bundles to isomorphism classes of matroid bundles, as well as a transformation from matroid b…

Consider a flat bundle over a complex curve. We prove a conjecture of Fei Yu that the sum of the top k Lyapunov exponents of the flat bundle is always greater or equal to the degree of any rank k holomorphic subbundle. We generalize the original context from Teichmueller curves to any local system over a curve with non…

We develop an efficient algorithm for low-rank approximation with improved approximation guarantees.

Generalizes randomized SVD for better matrix approximations using Gaussian vectors.

The spectral $k$-support norm enjoys good estimation properties in low rank matrix learning problems, empirically outperforming the trace norm. Its unit ball is the convex hull of rank $k$ matrices with unit Frobenius norm. In this paper we generalize the norm to the spectral $(k,p)$-support norm, whose additional para…