arXiv subjects

Toni Volkmer

Publications and source records attributed to Toni Volkmer.

10 recordsLinked to original sources

arXiv2021

Worst-case recovery guarantees for least squares approximation using random samples

We construct a least squares approximation method for the recovery of complex-valued functions from a reproducing kernel Hilbert space on $D \subset \mathbb{R}^d$. The nodes are drawn at random for the whole class of functions and the error is measured in $L_2(D,\varrho_D)$. We prove worst-case recovery guarantees by explicitly controlling all the involved constants. This leads to new preasymptotic recovery bounds with high probability for the error of Hyperbolic Fourier Regression on multivariate data. In addition, we further investigate its counterpart Hyperbolic Wavelet Regression also based on least-squares to recover non-periodic functions from random samples. Finally, we reconsider the analysis of a cubature method based on plain random points with optimal weights and reveal near-optimal worst-case error bounds with high probability. It turns out that this simple method can compete with the quasi-Monte Carlo methods in the literature which are based on lattices and digital nets.

math.NA↗

arXiv2021

Semi-supervised Learning for Aggregated Multilayer Graphs Using Diffuse Interface Methods and Fast Matrix Vector Products

We generalize a graph-based multiclass semi-supervised classification technique based on diffuse interface methods to multilayer graphs. Besides the treatment of various applications with an inherent multilayer structure, we present a very flexible approach that interprets high-dimensional data in a low-dimensional multilayer graph representation. Highly efficient numerical methods involving the spectral decomposition of the corresponding differential graph operators as well as fast matrix-vector products based on the nonequispaced fast Fourier transform (NFFT) enable the rapid treatment of large and high-dimensional data sets. We perform various numerical tests putting a special focus on image segmentation. In particular, we test the performance of our method on data sets with up to 10 million nodes per layer as well as up to 104 dimensions resulting in graphs with up to 52 layers. While all presented numerical experiments can be run on an average laptop computer, the linear dependence per iteration step of the runtime on the network size in all stages of our algorithm makes it scalable to even larger and higher-dimensional problems.

cs.LG↗

arXiv2020

Sparse Fourier Transforms on Rank-1 Lattices for the Rapid and Low-Memory Approximation of Functions of Many Variables

We consider fast, provably accurate algorithms for approximating functions on the $d$-dimensional torus, $f: \mathbb{ T }^d \rightarrow \mathbb{C}$, that are sparse (or compressible) in the Fourier basis. In particular, suppose that the Fourier coefficients of $f$, $\{c_{\bf k} (f) \}_{{\bf k} \in \mathbb{Z}^d}$, are concentrated in a finite set $I \subset \mathbb{Z}^d$ so that $$\min_{Ω\subset I s.t. |Ω| =s } \left\| f - \sum_{{\bf k} \in Ω} c_{\bf k} (f) e^{ -2 πi {\bf k} \cdot \circ} \right\|_2 < ε\|f \|_2$$ holds for $s \ll |I|$ and $ε\in (0,1)$. We aim to identify a near-minimizing subset $Ω\subset I$ and accurately approximate the associated Fourier coefficients $\{ c_{\bf k} (f) \}_{{\bf k} \in Ω}$ as rapidly as possible. We present both deterministic as well as randomized algorithms using $O(s^2 d \log^c (|I|))$-time/memory and $O(s d \log^c (|I|))$-time/memory, respectively. Most crucially, all of the methods proposed herein achieve these runtimes while satisfying theoretical best $s$-term approximation guarantees which guarantee their numerical accuracy and robustness to noise for general functions. These are achieved by modifying several one-dimensional Sparse Fourier Transform (SFT) methods to subsample a function along a reconstructing rank-1 lattice for the given frequency set $I$ to rapidly identify a near-minimizing subset $Ω\subset I$ without using anything about the lattice beyond its generating vector. This requires new fast and low-memory frequency identification techniques capable of rapidly recovering vector-valued frequencies in $\mathbb{Z}^d$ as opposed to simple integer frequencies in the univariate setting. Two different strategies are proposed and analyzed, each with different accuracy versus computational speed and memory tradeoffs.

math.NA↗

arXiv2020

A sample efficient sparse FFT for arbitrary frequency candidate sets in high dimensions

In this paper a sublinear time algorithm is presented for the reconstruction of functions that can be represented by just few out of a potentially large candidate set of Fourier basis functions in high spatial dimensions, a so-called high-dimensional sparse fast Fourier transform. In contrast to many other such algorithms, our method works for arbitrary candidate sets and does not make additional structural assumptions on the candidate set. Our transform significantly improves upon the other approaches available for such a general framework in terms of the scaling of the sample complexity. Our algorithm is based on sampling the function along multiple rank-1 lattices with random generators. Combined with a dimension-incremental approach, our method yields a sparse Fourier transform whose computational complexity only grows mildly in the dimension and can hence be efficiently computed even in high dimensions. Our theoretical analysis establishes that any Fourier $s$-sparse function can be accurately reconstructed with high probability. This guarantee is complemented by several numerical tests demonstrating the high efficiency and versatile applicability for the exactly sparse case and also for the compressible case.

math.NA↗

arXiv2020

A Deterministic Algorithm for Constructing Multiple Rank-1 Lattices of Near-Optimal Size

In this paper we present the first known deterministic algorithm for the construction of multiple rank-1 lattices for the approximation of periodic functions of many variables. The algorithm works by converting a potentially large reconstructing single rank-1 lattice for some $ d $-dimensional frequency set $ I \subset [N]^d $ into a collection of much smaller rank-1 lattices which allow for accurate and efficient reconstruction of trigonometric polynomials with coefficients in $ I $ (and, therefore, for the approximation of multivariate periodic functions). The total number of sampling points in the resulting multiple rank-1 lattices is theoretically shown to be less than $ \mathcal{O}\left( |I| \log^{ 2 }(N |I|) \right) $ with constants independent of $d$, and by performing one-dimensional fast Fourier transforms on samples of trigonometric polynomials with Fourier support in $ I $ at these points, we obtain exact reconstruction of all Fourier coefficients in fewer than $ \mathcal{O}\left(d\,|I|\log^4 (N|I|)\right) $ total operations. Additionally, we present a second multiple rank-1 lattice construction algorithm which constructs lattices with even fewer sampling points at the cost of only being able to reconstruct exact trigonometric polynomials rather than having additional theoretical approximation. Both algorithms are tested numerically and surpass the theoretical bounds. Notably, we observe that the oversampling factors #samples$/|I|$ appear to grow only logarithmically in $ |I| $ for the first algorithm and appear near-optimally bounded by four in the second algorithm.

math.NA↗

arXiv2019

Sparse Harmonic Transforms II: Best $s$-Term Approximation Guarantees for Bounded Orthonormal Product Bases in Sublinear-Time

In this paper, we develop a sublinear-time compressive sensing algorithm for approximating functions of many variables which are compressible in a given Bounded Orthonormal Product Basis (BOPB). The resulting algorithm is shown to both have an associated best $s$-term recovery guarantee in the given BOPB, and also to work well numerically for solving sparse approximation problems involving functions contained in the span of fairly general sets of as many as $\sim10^{230}$ orthonormal basis functions. All code is made publicly available. As part of the proof of the main recovery guarantee new variants of the well known CoSaMP algorithm are proposed which can utilize any sufficiently accurate support identification procedure satisfying a {Support Identification Property (SIP)} in order to obtain strong sparse approximation guarantees. These new CoSaMP variants are then proven to have both runtime and recovery error behavior which are largely determined by the associated runtime and error behavior of the chosen support identification method. The main theoretical results of the paper are then shown by developing a sublinear-time support identification algorithm for general BOPB sets which is robust to arbitrary additive errors. Using this new support identification method to create a new CoSaMP variant then results in a new robust sublinear-time compressive sensing algorithm for BOPB-compressible functions of many variables.

math.NA↗

arXiv2019

Approximation of multivariate periodic functions based on sampling along multiple rank-1 lattices

In this work, we consider the approximate reconstruction of high-dimensional periodic functions based on sampling values. As sampling schemes, we utilize so-called reconstructing multiple rank-1 lattices, which combine several preferable properties such as easy constructability, the existence of high-dimensional fast Fourier transform algorithms, high reliability, and low oversampling factors. Especially, we show error estimates for functions from Sobolev Hilbert spaces of generalized mixed smoothness. For instance, when measuring the sampling error in the $L_2$-norm, we show sampling error estimates where the exponent of the main part reaches those of the optimal sampling rate except for an offset of $1/2+\varepsilon$, i.e., the exponent is almost a factor of two better up to the mentioned offset compared to single rank-1 lattice sampling. Various numerical tests in medium and high dimensions demonstrate the high performance and confirm the obtained theoretical results of multiple rank-1 lattice sampling.

math.NA↗

arXiv2018

NFFT meets Krylov methods: Fast matrix-vector products for the graph Laplacian of fully connected networks

The graph Laplacian is a standard tool in data science, machine learning, and image processing. The corresponding matrix inherits the complex structure of the underlying network and is in certain applications densely populated. This makes computations, in particular matrix-vector products, with the graph Laplacian a hard task. A typical application is the computation of a number of its eigenvalues and eigenvectors. Standard methods become infeasible as the number of nodes in the graph is too large. We propose the use of the fast summation based on the nonequispaced fast Fourier transform (NFFT) to perform the dense matrix-vector product with the graph Laplacian fast without ever forming the whole matrix. The enormous flexibility of the NFFT algorithm allows us to embed the accelerated multiplication into Lanczos-based eigenvalues routines or iterative linear system solvers and even consider other than the standard Gaussian kernels. We illustrate the feasibility of our approach on a number of test problems from image segmentation to semi-supervised learning based on graph-based PDEs. In particular, we compare our approach with the Nyström method. Moreover, we present and test an enhanced, hybrid version of the Nyström method, which internally uses the NFFT.

cs.LG↗

arXiv2017

High-dimensional sparse FFT based on sampling along multiple rank-1 lattices

The reconstruction of high-dimensional sparse signals is a challenging task in a wide range of applications. In order to deal with high-dimensional problems, efficient sparse fast Fourier transform algorithms are essential tools. The second and third authors have recently proposed a dimension-incremental approach, which only scales almost linear in the number of required sampling values and almost quadratic in the arithmetic complexity with respect to the spatial dimension $d$. Using reconstructing rank-1 lattices as sampling scheme, the method showed reliable reconstruction results in numerical tests but suffers from relatively large numbers of samples and arithmetic operations. Combining the preferable properties of reconstructing rank-1 lattices with small sample and arithmetic complexities, the first author developed the concept of multiple rank-1 lattices. In this paper, both concepts - dimension-incremental reconstruction and multiple rank-1 lattices - are coupled, which yields a distinctly improved high-dimensional sparse fast Fourier transform. Moreover, the resulting algorithm is analyzed in detail with respect to success probability, number of required samples, and arithmetic complexity. In comparison to single rank-1 lattices, the utilization of multiple rank-1 lattices results in a reduction in the complexities by an almost linear factor with respect to the sparsity. Various numerical tests confirm the theoretical results, the high performance, and the reliability of the proposed method.

math.NA↗

arXiv2016

Tight error bounds for rank-1 lattice sampling in spaces of hybrid mixed smoothness

We consider the approximate recovery of multivariate periodic functions from a discrete set of function values taken on a rank-$s$ integration lattice. The main result is the fact that any (non-)linear reconstruction algorithm taking function values on a rank-$s$ lattice of size $M$ has a dimension-independent lower bound of $2^{-(α+1)/2} M^{-α/2}$ when considering the optimal worst-case error with respect to function spaces of (hybrid) mixed smoothness $α>0$ on the $d$-torus. We complement this lower bound with upper bounds that coincide up to logarithmic terms. These upper bounds are obtained by a detailed analysis of a rank-1 lattice sampling strategy, where the rank-1 lattices are constructed by a component-by-component (CBC) method. This improves on earlier results obtained in [25] and [27]. The lattice (group) structure allows for an efficient approximation of the underlying function from its sampled values using a single one-dimensional fast Fourier transform. This is one reason why these algorithms keep attracting significant interest. We compare our results to recent (almost) optimal methods based upon samples on sparse grids.

math.NA↗