PDF Research in Data Science Ellen Gasparovic Ebook Full Chapter

Research in Data Science Ellen
Gasparovic
Visit to download the full and correct content document:
https://textbookfull.com/product/research-in-data-science-ellen-gasparovic/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Analytics and Data Science: Advances in Research and

Pedagogy 1st Edition Amit V. Deokar
https://textbookfull.com/product/analytics-and-data-science-
advances-in-research-and-pedagogy-1st-edition-amit-v-deokar/
Data Science and Social Research II: Methods,

Technologies and Applications Paolo Mariani
https://textbookfull.com/product/data-science-and-social-
research-ii-methods-technologies-and-applications-paolo-mariani/
Frontiers In Data Science Matthias Dehmer
https://textbookfull.com/product/frontiers-in-data-science-
matthias-dehmer/
Ethics in Science Ethical Misconduct in Scientific

Research D'Angelo
https://textbookfull.com/product/ethics-in-science-ethical-
misconduct-in-scientific-research-dangelo/
Social Science Research Ethics in Africa Nico Nortjé
https://textbookfull.com/product/social-science-research-ethics-
in-africa-nico-nortje/
Real Time Stream Data Management Push Based Data in

Research Practice Wolfram Wingerath
https://textbookfull.com/product/real-time-stream-data-
management-push-based-data-in-research-practice-wolfram-
wingerath/
Frontiers in Data Science 1st Edition Matthias Dehmer
https://textbookfull.com/product/frontiers-in-data-science-1st-
edition-matthias-dehmer/
Uncertainty Modelling in Data Science Sébastien

Destercke (Ed.)
https://textbookfull.com/product/uncertainty-modelling-in-data-
science-sebastien-destercke-ed/
Advances in Panel Data Analysis in Applied Economic

Research Nicholas Tsounis
https://textbookfull.com/product/advances-in-panel-data-analysis-
in-applied-economic-research-nicholas-tsounis/
Association for Women in Mathematics Series
Ellen Gasparovic
Carlotta Domeniconi Editors
Research
in Data
Science
Volume 17
Series Editor
Kristin Lauter
Microsoft Research
Redmond,Washington, USA
Focusing on the groundbreaking work of women in mathematics past, present, and

future, Springer’s Association for Women in Mathematics Series presents the latest
research and proceedings of conferences worldwide organized by the Association
for Women in Mathematics (AWM). All works are peer-reviewed to meet the highest
standards of scientific literature, while presenting topics at the cutting edge of pure
and applied mathematics. Since its inception in 1971, The Association for Women in
Mathematics has been a non-profit organization designed to help encourage women
and girls to study and pursue active careers in mathematics and the mathematical
sciences and to promote equal opportunity and equal treatment of women and girls
in the mathematical sciences. Currently, the organization represents more than 3000
members and 200 institutions constituting a broad spectrum of the mathematical
community, in the United States and around the world.
More information about this series at http://www.springer.com/series/13764

Ellen Gasparovic • Carlotta Domeniconi
Editors
Research in Data Science
123
Editors
Ellen Gasparovic Carlotta Domeniconi
Department of Mathematics Department of Computer Science
Union College George Mason University
Schenectady, NY, USA Fairfax, VA, USA
ISSN 2364-5733 ISSN 2364-5741 (electronic)

ISBN 978-3-030-11565-4 ISBN 978-3-030-11566-1 (eBook)
https://doi.org/10.1007/978-3-030-11566-1
Library of Congress Control Number: 2019935518
Mathematics Subject Classification (2010): 62-07, 68P05, 68T05, 68P20, 62H30, 91C20, 68U10, 65D18,
62H35, 68W27, 68U05, 05E45, 55Q07
© The Author(s) and the Association for Women in Mathematics 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
The first Women in Data Science and Mathematics (WiSDM) Research Collabora-
tion Workshop was held on July 17–21, 2017, at the Institute for Computational and
Experimental Research in Mathematics (ICERM) in Providence, Rhode Island. In
addition to generous support from ICERM, the workshop was partially supported
by the Association for Women in Mathematics (AWM) ADVANCE grant funded by
the National Science Foundation. Additional support for some participant travel was
provided by the Center for Discrete Mathematics and Theoretical Computer Science
(DIMACS) in association with its Special Focus on Information Sharing and
Dynamic Data Analysis. The workshop was co-sponsored by Brown University’s
Data Science Initiative.
The 52 participants of the workshop included women from mathematics (theo-
retical and applied), statistics, computer science, and electrical and computer engi-
neering. They represented women at all career stages and in diverse career paths,
including faculty, postdoctoral fellows, graduate students, industrial scientists, and
government scientists. Based on their research interests and backgrounds, each
participant was assigned to one of six working groups headed by leading researchers
in the field of data science. In addition to intense research time, the week’s schedule
featured introductory talks from the group leaders, a panel discussion, an invited
lecture from Brown University biologist Sohini Ramachandran, and presentations
of the week’s work from each of the working groups.
This single-blind peer-reviewed volume is the proceedings of the first WiSDM
workshop and features accepted submissions from several of the working groups as
well as additional papers solicited from the wider community. Topics range from
the more theoretical to the more applied and computational.
The Cross-Disciplinary Field of Data Science
Data science is a cross-disciplinary field relying on methodologies and practices

from statistics, computer science, and mathematics. Data science seeks consistent
v
vi Preface
patterns in data for actionable prediction and decision-making. As such, it is an

applied field driven by problems in a variety of disciplines. Data science offers great
promise to those fields that traditionally lacked solid theory. Data generated within
such disciplines can now be utilized to produce powerful predictive models.
While data science has emerged as a prominent new research area, enrolling
record numbers and attracting talented researchers from many scientific fields, the
role of theoretical and applied mathematics has not yet been made highly visible.
Both the WiSDM workshop participants and the editors and authors of this volume
are committed to highlighting and establishing the fundamental role of mathematics
in the development of data science.
Project Descriptions
Led by Julie Mitchell, the goal of the first group’s project was to build and
optimize predictive models for molecular data using a range of machine learning
and informatics techniques applied to data generated from past molecular modeling
projects. Prior models had been built with around 100 experimental data points,
while other bimolecular models utilized over 50,000 data points. Hence, participants
focused on the applicability of various machine learning methods to data sets of
different sizes.
Under the guidance of Linda Ness, the second group considered the rep-
resentation of data as multi-scale features and measures. Recently, multi-scale
representation theorems from harmonic analysis and geometric measure theory have
been exploited to compute canonical multi-scale representations of data samples
for supervised and unsupervised learning, statistical fusion and construction of
confidence measures, and data visualization. The goal of this research collaboration
was threefold: to assess the applicability of multi-scale representation approaches
to various types of data, to introduce the approach to statistical researchers who
may be interested in statistical fusion and confidence measures, and to develop and
apply new multi-scale methods for representation of data as measures characterizing
mathematical properties of the data.
The members of the third group, led by Giseon Heo, looked at inferential models
founded in statistical and topological learning applied to pediatric obstructive sleep
apnea (OSA) data sets. OSA is a form of sleep-disordered breathing characterized
by recurrent episodes of partial or complete airway obstruction during sleep and
is prevalent in 1–5% of school-aged children. Chronic diseases such as OSA
are multifactorial disorders, necessitating different types of data to capture the
complex system. The team focused on analyses for time series signals from
polysomnography, survey questionnaires, and upper airway shapes. The participants
sought to develop a statistical and topological learning model that could accurately
predict OSA severity.
The goal of the fourth group’s project, under the direction of Deanna Needell,
was to apply stochastic signal processing for high-dimensional data. One mathe-
Preface vii
matical method that has gained a lot of recent attention in the ever-evolving field
of data science is the use of sparsity and stochastic designs. Sparsity captures the
idea that high-dimensional signals often contain a very small amount of intrinsic
information. Often, through randomized designs, signals can be captured using a
very small number of measurements. On the recovery side, stochastic methods can
accurately estimate signals from those measurements in the underdetermined set-
ting, as well as solve large-scale systems in the highly overdetermined setting. Par-
ticipants selected applications of interest, designed stochastic algorithms for those
frameworks, and ran experiments on synthetic data from those application areas.
The fifth working group, led by Carlotta Domeniconi, studied the hubness
phenomenon in high-dimensional spaces. Although data can easily contain tens of
thousands of features, data often have an intrinsic dimensionality that is embedded
within the full-dimensional space. Hubness causes certain data examples to appear
more often than others as neighbors of points, thus generating a skewed distribution
of nearest neighbor counts. The participants investigated the relationship between
the hubness phenomenon and the intrinsic dimensionality of data, with the ultimate
goal of recovering the subspaces data lie within. The findings could enable effective
subspace clustering of data, as well as outlier identification.
Finally, under the guidance of Emina Soljanin, the participants in the sixth
working group focused on codes for data storage with queues for data access. Large
volumes of data, which are being collected for the purpose of knowledge extraction,
have to be reliably, efficiently, and securely stored. Retrieval of large data files from
storage has to be fast (and often anonymous and private). This project was concerned
with big data storage and access, and its relevant mathematical disciplines include
algebraic coding and queueing theory. The members of the group looked at coding
and queueing problems in the era of big data, two interwoven and indispensable
aspects of big data storage and access.
Contributed Papers
In the first paper by Durgin et al., the authors develop a stochastic approach, based
on the sparse randomized Kaczmarz (SRK) algorithm, to perform support recovery
of corrupted and jointly sparse multiple measurement vectors (MMV). In the MMV
setting, one has access to multiple vectors, or signals, that are assumed to share a
support set. However, the measurement vectors are corrupted, meaning that each
measurement vector may have additional non-zeros that are not truly part of the
shared support the authors aim to recover. The authors also adapt the SRK algorithm
to the online setting where the measurements are streaming continuously.
Mani et al. address the challenges of learning with high-dimensional data,
focusing on the hubness phenomenon. The authors identify and discuss new geo-
metric relationships between hubness, data density, and data distance distribution.
The findings shed light on the role of hubness in the discovery of the intrinsic
dimensionality of data and thus in the design of effective methods to recover the
viii Preface
embedding subspaces of data. Potential research directions to leverage hubness for

clustering and for subspace estimation are discussed.
Medina et al. introduce a heuristic framework for testing the multi-manifold
hypothesis for high-dimensional data. The proposed approach uses a multi-scale
variance-based computation of intrinsic dimension. Results on both synthetic and
real-world data are discussed.
In Heo et al., the authors work with complex and multidimensional obstructive
sleep apnea patient data. They use a combination of techniques from computational
geometry, machine learning and statistics (e.g., random forests, principal component
analysis, and cluster analysis), and computational topology to investigate the data
from many angles, with the forward-looking goal of potential usefulness for clinical
practitioners.
Munch and Stefanou show that the ∞ -cophenetic metric for phylogenetic trees
is actually an example of an interleaving distance. They achieve this by representing
phylogenetic trees as topological structures known as merge trees with labeled
leaves and by taking advantage of a generalized framework for interleavings on
categories with a flow.
In her paper, Ness proposes representing data sets via a dyadic set structure
coming from ordered binary feature functions coupled with a variant of a nerve
simplicial complex that is determined by the support of the dyadic measure and its
Betti numbers. She demonstrates the utility of her methods on a data quality data set
and makes a case for their usefulness in such venues as statistical fusion, inference,
and visualization.
Genctav et al. propose a mesh saliency measure which reflects local to global
integration of saliency. An effective feature construction followed by additive matrix
decomposition is performed to generate the saliency measure. Several proof-of-
concept illustrative results are shown.
Seeger et al. use support vector machines to create a classifier that can recognize
protein binding hotspots. Hotspots are the residues that provide at least 2 kcal/mol
to the binding energy, thus making them the key residues for design to alter affinity
or specificity or to target drug molecules to block protein binding. The work
combines one of the leading methods, KFC2, with features from Rosetta, a software
package for biomolecular design. A resulting classifier provides results with greater
accuracy, precision, recall, and specificity than KFC2 or Rosetta alone.
Aroutiounian et al. employ the skeletal shape model known as the Blum medial
axis to uncover features capturing the regional geometry of a given shape. They
apply their methods to both successfully classify a data set consisting of corpora
callosa of schizophrenic and healthy patients and to automatically decompose a
corpus callosum into regional substructures.
In the second paper by Durgin et al., the authors propose algorithms for the
compressive anomaly detection problem. They analyze existing approaches for
compressed sensing of jointly sparse signals to infer anomalous entries from
a sequence of observations. The authors first present probabilistic variations of
previous sparse anomaly models and utilize the new models to prove asymptotic
bounds on two different algorithms that infer the locations of the anomalies from
Preface ix
random and compressive measurements. Finally, the authors empirically validate

and compare their methods on a number of simulations.
In Grim et al., the authors analyzed simulated data of crowds exiting various con-
figurations of a given building. This entailed using a plethora of mathematical and
statistical techniques to investigate multiple aspects of crowd dynamics research,
including visualization, pattern learning, panic detection, and exit time estimation.
Adanova and Tari present a novel approach to perform continuous categorization
of ornaments based on data clustering. Traditional approaches use group theory to
identify symmetric patterns of ornaments. Simply relying on symmetrical groups,
though, fails to capture relations among ornaments when unusual base motifs are
used or color permutation is performed by the artist. To address this issue, the
authors avoid fitting an ornament into one of the predefined symmetry groups
and instead perform clustering based on content-suppressed images to detect
perceptually close ornaments.
Fairfax, VA, USA Carlotta Domeniconi

Schenectady, NY, USA Ellen Gasparovic
Acknowledgments
The first Women in Data Science and Mathematics workshop was a great success
thanks in large part to generous funding and support from ICERM, the National
Science Foundation (NSF-HRD 1500481) and the AWM (“Career Advancement for
Women Through Research-Focused Networks”), DIMACS, and Brown University’s
Data Science Initiative. The AWM and NSF have provided funds for a follow-
up research minisymposium during the 2019 SIAM Conference on Computational
Science and Engineering. The organizers of WiSDM, the editors, and the authors
in this volume are all tremendously grateful for the support for these unique
opportunities for collaboration and dissemination.
The editors would like to thank the chapter authors as well as the reviewers
who gave valuable feedback and suggestions to the authors. We would also like
to heartily thank the AWM and Springer for the opportunity to create this volume.
We look forward to many more WiSDM research collaboration workshops in the
future and to continuing to build the WiSDM research network.
xi
Contents
Sparse Randomized Kaczmarz for Support Recovery of Jointly

Sparse Corrupted Multiple Measurement Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Natalie Durgin, Rachel Grotheer, Chenxi Huang, Shuang Li, Anna Ma,
Deanna Needell, and Jing Qin
The Hubness Phenomenon in High-Dimensional Spaces . . . . . . . . . . . . . . . . . . . . 15
Priya Mani, Marilyn Vazquez, Jessica Ruth Metcalf-Burton,
Carlotta Domeniconi, Hillary Fairbanks, Gülce Bal, Elizabeth Beer,
and Sibel Tari
Heuristic Framework for Multiscale Testing of the Multi-Manifold
Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
F. Patricia Medina, Linda Ness, Melanie Weber,
and Karamatou Yacoubou Djima
Interdisciplinary Approaches to Automated Obstructive Sleep
Apnea Diagnosis Through High-Dimensional Multiple Scaled
Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Giseon Heo, Kathryn Leonard, Xu Wang, and Yi Zhou
The ∞ -Cophenetic Metric for Phylogenetic Trees
As an Interleaving Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Elizabeth Munch and Anastasios Stefanou
Inference of a Dyadic Measure and Its Simplicial Geometry
from Binary Feature Data and Application to Data Quality . . . . . . . . . . . . . . . . 129
Linda Ness
A Non-local Measure for Mesh Saliency via Feature Space Reduction . . . . 167
Asli Genctav, Murat Genctav, and Sibel Tari
xiii
xiv Contents
Feature Design for Protein Interface Hotspots Using

KFC2 and Rosetta. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Franziska Seeger, Anna Little, Yang Chen, Tina Woolf, Haiyan Cheng,
and Julie C. Mitchell
Geometry-Based Classification for Automated Schizophrenia
Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Robert Aroutiounian, Kathryn Leonard, Rosa Moreno, and Robben Teufel
Compressed Anomaly Detection with Multiple Mixed Observations . . . . . . 211
Analysis of Simulated Crowd Flow Exit Data: Visualization, Panic
Detection and Exit Time Convergence, Attribution, and Estimation . . . . . . 239
Anna Grim, Boris Iskra, Nianqiao Ju, Alona Kryshchenko,
F. Patricia Medina, Linda Ness, Melissa Ngamini, Megan Owen,
Randy Paffenroth, and Sui Tang
A Data Driven Modeling of Ornaments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Venera Adanova and Sibel Tari
Sparse Randomized Kaczmarz for
Support Recovery of Jointly Sparse
Corrupted Multiple Measurement
Vectors
Abstract While single measurement vector (SMV) models have been widely
studied in signal processing, there is a surging interest in addressing the multiple
measurement vectors (MMV) problem. In the MMV setting, more than one
measurement vector is available and the multiple signals to be recovered share
some commonalities such as a common support. Applications in which MMV is
a naturally occurring phenomenon include online streaming, medical imaging, and
video recovery. This work presents a stochastic iterative algorithm for the support
recovery of jointly sparse corrupted MMV. We present a variant of the sparse
randomized Kaczmarz algorithm for corrupted MMV and compare our proposed
method with an existing Kaczmarz type algorithm for MMV problems. We also
N. Durgin
Spiceworks, Austin, TX, USA
R. Grotheer ()
Goucher College, Baltimore, MD, USA
e-mail: rachel.grotheer@goucher.edu
C. Huang
Yale University, New Haven, CT, USA
S. Li
Colorado School of Mines, Golden, CO, USA
A. Ma
Claremont Graduate University, Claremont, CA, USA
e-mail: anna.ma@cgu.edu
D. Needell
University of California, Los Angeles, CA, USA
J. Qin
Montana State University, Bozeman, MT, USA
© The Author(s) and the Association for Women in Mathematics 2019 1

E. Gasparovic, C. Domeniconi (eds.), Research in Data Science,
Association for Women in Mathematics Series 17,
https://doi.org/10.1007/978-3-030-11566-1_1
2 N. Durgin et al.
showcase the usefulness of our approach in the online (streaming) setting and
provide empirical evidence that suggests the robustness of the proposed method
to the number of corruptions and the distribution from which the corruptions are
drawn.
1 Introduction
In recent years, there has been a drastic increase in the amount of available data.
This so-called “data deluge” has created a demand for fast, iterative algorithms that
can be used to process large-scale data. Stochastic iterative algorithms, such as the
randomized Kaczmarz or stochastic gradient descent algorithms, have become an
increasingly popular option for processing large-scale data [3, 10]. These methods
recover a signal X ∈ Rn given a vector of measurements Y ∈ Rm and a
measurement matrix Φ ∈ Rm×n such that:
Y = ΦX, (1)
without accessing the full measurement matrix in a single iteration. We refer to (1)
as a single measurement vector (SMV) model. In the multiple measurement vector
(MMV) setting, one may have thousands of measurement vectors Y (·,j ) pouring in
over time. Each measurement vector corresponds to a signal X(·,j ) , where signals
typically share a common property such as sparsity, smoothness, etc. For simplicity,
let Y = [Y (·,1) · · · Y (·,J ) ] ∈ Rm×J and X = [X(·,1) · · · X(·,J ) ] ∈ Rn×J . Since high-
dimensional data is typically sparse in nature, a commonality of particular interest
is joint sparsity, or when signals share the same support. The support of a vector v is
defined to be the set of indices for which v is nonzero, i.e., supp(v) = {i : vi = 0}.
Many algorithms have been developed for the MMV setting, especially in
applications such as line spectral estimation [13, 20] and modal analysis [14].
The authors in these works extend the previous SMV-based algorithms as well as
theoretical analysis in [4, 8, 19] to the MMV case. The theoretical bound in [14] also
indicates that MMV settings could make the problem of compressed signal recovery
much easier than in the SMV setting. In particular, the number of measurements
needed for perfect recovery in each signal decreases as the number of signals
increases, reducing the sample complexity per signal.
As a motivating example, consider diffuse optical tomography (DOT) where
the goal is to find small areas of high contrast corresponding to the location of
cancerous cells [2]. Since cancerous cells have a much larger absorption coefficient
than healthy cells, a two-dimensional medical image can be interpreted as a sparse
signal where each entry of the signal represents the absorption coefficient of a given
pixel and the nonzero entries correspond to tumor locations. In a hyperspectral DOT
setting, hundreds of different wavelengths are used to acquire a variety of images
of the same tissue, allowing practitioners to obtain a more accurate location of
tumors [11]. The hyperspectral imaging process results in a jointly sparse MMV,
Sparse Randomized Kaczmarz for Support Recovery of Jointly Sparse. . . 3
where each wavelength produces a different image (or signal), and the joint support
across all images indicates the locations of cancerous cells.
Signals may share a common support but it is improbable for them to be
perfectly accurate. Since sensing mechanisms are not impervious to error, signals
can contain corruptions. Other sources of corruption in signal processing include
spikes in power supply, defective hardware, and adversarial agents [12]. Going
back to the hyperspectral imaging example, “corruptions” in each signal may be
caused by noncancerous cells that absorb more light at a given wavelength than
their neighbors. For example, if a cell contains an anomalous amount of melanin,
then it absorbs more light at shorter wavelengths in the visible spectrum (i.e., violet
or blue light) compared to a typical noncancerous cell [5, 15]. This produces a large
nonzero absorption coefficient in the location of a healthy cell, i.e., a corruption.
Corrupt entries erroneously indicate the presence of cancerous cells in a location
with healthy cells.
Corruptions cause support recovery algorithms such as the MMV sparse random-
ized Kaczmarz (MMV-SRK) algorithm, which we describe in detail in Sect. 2, to fail
due to the algorithmic dependence on the row norms of the signal approximation
to estimate the support [1]. Thus, large corruptions in a signal with comparatively
small entries may erroneously be included in the support estimate given by these
algorithms. In the corrupt MMV setting, the availability of multiple measurement
vectors becomes vital to the estimate of the true support. Clearly, if only a single
measurement vector is available, there would be no way to distinguish a corrupt
nonzero entry without any additional assumptions on the signal or corruption.
Corrupt measurement signals have been studied in the context of the SMV model.
In [18] and [12], additive noise in the measurement scheme is assumed to be sparse.
Both works focus on the compressive sensing setting where m n.
The primary objective of this work is to design an algorithm for recovering the
support of jointly sparse, corrupt signals in the large-scale setting. We propose
a new online algorithm called sparse randomized Kaczmarz for corrupted MMV
(cMMV-SRK) for support recovery. Note that the proposed algorithm can recover
the signals with high accuracy based on our experiments, but we mainly focus on
support recovery in this work. Our experiments show that the proposed algorithm
outperforms the previously proposed Kaczmarz type algorithm in recovering the
joint support from MMV when the signals are corrupted.
1.1 Problem Formulation
Given a set of linear measurements Y (·,j ) ∈ Rm and a measurement matrix Φ ∈

Rm×n such that:
Y (·,j ) = ΦX(·,j ) for j = 1, . . . J, (2)

4 N. Durgin et al.
with X(·,j ) ∈ Rn . We assume that the data is large-scale, meaning we cannot access
all of Φ at once (m and/or n is too large) and must only operate on one row of Φ
at a time. We allow the system to be overdetermined (m n) or underdetermined
(m n) and assume the X(·,j ) ’s are jointly sparse such that supp(X(·,j ) ) = S and
|S | = k. For an n-dimensional vector X(·,j ) , let X(·,j ) |s return X(·,j ) with zeros in
the n − s smallest (in magnitude) entries. We also assume that each column of X
contains one or more corruptions where a corruption is a nonzero entry occurring
outside the joint support. In other words, instead of supp(X(·,j ) ) ⊂ S , the joint
support set, the support of X(·,j ) is
supp(X(·,j ) ) = S ∪ Cj , Cj ⊂ {1, . . . , n},
where Cj is the “corrupt index set.” Note that the Cj are not necessarily the same
for every j . In this work, our goal is to recover the joint support S from the given
linear measurements Y .
The remainder of this manuscript is organized in the following way. Section 2
discusses the sparse randomized Kaczmarz method and the MMV-SRK algorithm.
Section 3 provides a discussion on how corruptions can negatively impact the
performance of MMV-SRK. Section 3 also presents our method, cMMV-SRK, a
variant of SRK which works in the corrupted signal setting. Numerical experiments
using this method are presented in Sect. 4 and we conclude with a summary of our
contributions and future directions in Sect. 5.
2 Related and Existing Work
2.1 Sparse Randomized Kaczmarz
In this work, we utilize the sparse randomized Kaczmarz (SRK) algorithm to

recover the support of each column of X. The original Kaczmarz algorithm was
first introduced in the early 1900s by Kaczmarz himself and was revitalized as
the algebraic-reconstruction technique in the early 1980s [6, 10]. The randomized
Kaczmarz algorithm (RK) discussed throughout the paper was first introduced by
Strohmer and Vershynin and enjoys an expected linear convergence rate to the
solution of a consistent linear system [17]. The SRK algorithm is another variant
designed specifically for overdetermined systems with sparse solutions. SRK has
been empirically shown to solve underdetermined systems with sparse solutions as
well [16].
Algorithm 1 outlines the SRK algorithm. Note that ties are broken lexicograph-
ically in all algorithms presented in this work. The estimated support size k̂ is a
parameter of the algorithm and is typically chosen to be larger than the true support
Algorithm 1 Sparse Randomized Kaczmarz

1: procedure SRK(Φ ∈ Rm×n , Y ∈ Rm , k̂, τ )
2: Initialize X0 = 0n×1
3: for t = 1, ..., τ do
Φ (i,·) 22
4: Choose row Φ (i,·) with probability
Φ 2F
5: Set support estimate S t = supp(X t−1 |k̂ )
6: Set row weights w ∈ Rn

1 : l ∈ St
wl = √1
t
: l ∈ S tc
c
S t is the complement set of S t
7: Set a = w · Φ (i,·) a ∈ Rn is the weighted row of Φ
Y i −aXt−1 T
8: Update X t = Xt−1 + a
a 22
9: end for
10: return X τ
11: end procedure
size k. In this variant, the algorithm runs for a specified number of iterations (up to
τ ). However, any stopping criteria one would use for an iterative algorithm, e.g.,
terminating after the residual meets a certain criteria, after the updates become
small, etc., can be used. Algorithm 1 also differs from the original presented
algorithm by Mansour and Yilmaz [16] in that at every iteration the support estimate
has size k̂ instead of starting with n and shrinking the size to k̂. We find that these
modifications do not significantly affect the behavior of SRK.
Algorithm 1 has been shown empirically to find the solution to overdetermined,
consistent (i.e., a solution exists) linear systems but there are no theoretical results
supporting this. One can make a few observations about the behavior of SRK
for support recovery. Concerning the support size estimate k̂, it is clear that if
k̂ < k then the probability that the true support is contained in the support of the
approximation is 0, i.e., P (S ⊂ supp(Xτ )) = 0. Additionally, if k̂ = n, then
P (S ⊂ supp(Xτ )) = 1. In regards to the choice of weighting, as t → ∞, √1t → 0
so that row elements inside the support estimate contribute to the approximation the
most. If one has a weighting function that decreases too rapidly, the true support
may not be captured in S t causing the algorithm to fail.
Although Algorithm 1 and the following algorithms require the Frobenius norm
of the matrix, Φ 2F , for row selection, practically speaking row selections can be
done uniformly at random to avoid using the full measurement matrix in a single
iteration. Indeed, it is advantageous to select the rows at random to avoid introducing
bias from rows with larger norms.
6 N. Durgin et al.
2.2 SRK for MMV
In this section we present an SRK-based approach to the MMV setting proposed by

Aggarwal and Majumdar [1]. Because we are assuming joint sparsity in the MMV
model, the estimated support of a signal reveals information about the support of all
signals. The authors of [1] present Algorithm 2 to leverage this idea. There are a few
key aspects to note about this version of the SRK algorithm. First, the algorithm runs
one iteration of SRK for every signal in the MMV model, then updates the support
estimate based on the row norms of the estimate Xt . Due to this, the algorithm does
not lend itself well to being extended for an online variant which only receives a
small number (possibly 1) of signals at a time. Second, the algorithm uses the same
selected row for each signal. It has been well observed that a random selection
scheme reduces the possibility of a poor choice of row ordering and it may be
advantageous to allow each signal to be projected onto a different randomly selected
row [7, 9].
Algorithm 2 Sparse Randomized Kaczmarz for MMV

1: procedure MMV-SRK(Φ ∈ Rm×n , Y ∈ Rm×J , k̂, τ )
2: Initialize X0 = 0n×J
3: for t = 1, . . . , τ do
Φ (i,·) 2
4: Choose row Φ (i,·) with probability 2
Φ 2F
5: Set support estimate S t : k̂ indices with largest row norm of Xt−1

1 : l ∈ St
wl = √1
t
: l ∈ S tc
c
S t is the complement set of S t
7: Set a = w · Φ (i,·) a is the weighted row of Φ
8: for j = 1, . . . , J do
Y (i,j ) −aXt−1
9: Update X t(·,j ) = Xt−1
(·,j ) +
(·,j )
aT
a 2
2
10: end for
11: Update Xt = [X t(·,1) | . . . |X t(·,J ) ]
12: end for
13: return X t
14: end procedure
3 Main Results
3.1 Corrupted MMV
To review, we are interested in constructing an algorithm that recovers the support

of jointly sparse corrupted high-dimensional MMV, that is, where we can only
access one row of the measurement matrix Φ at a time. To this end, we propose
Algorithm 3, which we refer to as cMMV-SRK. We first note that the base of this
algorithm is the SRK algorithm (Algorithm 1), which is an effective algorithm for
large-scale problems due to its low memory footprint, requiring only one row of the
measurement matrix to be used at a time. The proposed method, cMMV-SRK, is
adapted to the MMV setting using the intuition that the individual signals give us
information about the common support between all signals. We keep track of a bin
or tally vector b that estimates the true support of the signals. In particular, we use
the nonzeros in b to indicate the estimated joint support. This binning process allows
the algorithm to be robust in the face of corruptions in the signal, as the corruptions
will receive a low number of tallies compared to the entries in the true support
because the corruptions occur in random positions for every signal. Note that in
the corrupted MMV case, we expect Algorithm 2 to fail as the support estimate
step relies on the 2 -norm of the rows to be large if an index is in the support
and small otherwise. The corruptions may be so large that a single corruption in
a row could lead to misidentification of the corrupt entry being in the joint support.
Algorithm 3 can also account for signals being processed one at a time, as in an
online or “streaming” setting.
For each signal, let τ̃j be the number of SRK projections performed on X(·,j )
and let τ̃ = [τ̃1 · · · τ̃J ]. In the online setting, one can imagine that the amount of
time before the next signal is acquired may vary due to, for example, stalls in the
measurement process. The varying amount of time that the system has to process
each signal is one of the major challenges of support recovery in the online setting.
To improve the joint support estimate when τ̃j varies, we weight the binning based
τ˜
on τ = maxt τ˜t . In other words, we let bq = bq + τj , where bq is the qth entry of
b and τ̃j is the maximum number of inner iterations of SRK for the j th signal. This
reweighting scheme places a larger importance on support estimates which have had
more time (iterations) to improve. In the online setting where the τ̃j s are not known
a priori, τ can be set manually. We adopt the following notation for cMMV-SRK:
Sjt is the estimated support at the tth SRK iteration for X(·,j ) , and S t denotes the
joint support estimate.
If the number of inner iterations τ̃j is large enough, the support estimate should
be such that it contains the joint support (along with the corruption index). Because
we are tallying the support estimate after every τ iterations, it is clear that the entries
in the joint support will have an overwhelming number of tallies compared to all
other entries. The experimental results in the next section support these claims and
we leave the analytical study of these algorithms for future work.
8 N. Durgin et al.
4 Experiments
In this section, we compare Algorithms 2 and 3 under a variety of settings, specif-

ically comparing the robustness of each algorithm in the presence of corruptions
in the signal. To test this robustness, we vary the number of corrupt entries, the
distribution from which the corruptions are drawn, the number of columns in
X, and the number of projection computations τ̃j made for each signal. In what
follows, we will refer to τ as the number of SRK iterations. These experiments are
summarized in Table 1. In all experiments, the results are averaged over 40 trials
and the nonzero entries of X are drawn independently and identically distributed
(i.i.d.) from N (0, 1). On the horizontal axis we plot the number of “iterations”
where a single iteration is defined by a projection. In other words, the horizontal
axis represents every time Step 9 is performed in Algorithms 2 and 3.
Figure 1 compares Algorithms 2 and 3 with m = 1000, n = 100, and k = 10. The
support size estimate is k̂ = 1.5k. To create X, we uniformly at random select k of
n indexes to be the joint support S and set Y = ΦX. The corrupt entries are drawn
uniformly at random from {1, . . . N} \ S . To start off, each signal has one corrupt
entry. We choose corruptions i.i.d. from N (7, 1) to simulate corruptions being large
spikes in the signal (possibly caused by system malfunction). The maximum number
i.i.d
of SRK iterations for each signal is τ = 300. In Fig. 1a, we create Φ ∈ Rm×n ∼
N (0, 1) and J = 300. In Fig. 1b, the entries of Φ ∈ Rm×n are drawn from a
Algorithm 3 Sparse Randomized Kaczmarz for corrupted MMV

1: procedure CMMV-SRK( Φ ∈ Rm×n , Y ∈ Rm×J , k̂, τ̃ )
2: Initialize X0 = 0n×J , b = 0n×1 , τ = maxj τ˜j
3: for j = 1, . . . , J do
4: for t = 1, . . . , τ̃j do
Φ (i,·) 22
5: Choose row Φ (i,·) with probability
Φ 2F
t−1
6: Set support estimate Sjt = supp(X (·,j ) |k̂ )

1 : l ∈ St
wl = √1
t
: l ∈ St c
S t c is the complement set of S t

8: Set a = w · Φ (i,·) a is the weighted row of Φ
Y (i,j ) −aXt−1
9: Update X t(·,j ) = Xt−1
(·,j ) +
(·,j )
aT
a 2
2
10: end for
τ̃
11: If q ∈ Sjt then bq = bq + τj
12: Set initial support estimate for next signal S j = supp(b|k̂ )
13: end for
14: return Joint support estimate S J
15: end procedure
Table 1 Summary of experiment parameters

Figures Φ entries Distribution of corruptions Number of corruptions τ̃j J
Figure 1a Gaussian N (7, 1) 1 40 300
Figure 1b Uniform N (7, 1) 1 80 600
Figure 2a Gaussian N (0, 1) 1 40 300
Figure 2b Uniform N (0, 1) 1 80 600
Figure 3 Gaussian N (7, 1) Varies 50 300
Figure 4a Gaussian N (7, 1) Varies Varies 800
This table provides a summary of experiment parameters and figure references for each experiment
Fig. 1 Comparing cMMV-SRK and MMV-SRK for support recovery when there is a single
corrupt entry per signal whose magnitude is drawn from N (7, 1). (a) Φ ∼ N (0, 1). (b)
Φ ∼ U nif ([0, 1])
uniform distribution ranging from 0 to 1. We note that, in both cases, Algorithm 3

is able to recover the full support after a sufficient number of iterations, whereas
Algorithm 2 is only able to recover at most about 20% of the support, regardless of
the number of iterations. Since Algorithm 2 relies on row norms to estimate the joint
support, it is to be expected that the relatively large value of the corruption would
cause it to often be erroneously chosen to be part of the joint support estimate. As a
result, this experiment highlights the advantage of the binning in Algorithm 3 in the
presence of a single corruption with large magnitude.
In Fig. 2, we experiment further with the magnitude of the corruption. Here we
have that m = 1000, n = 100, and k = 10 but instead of the corrupt entries
being drawn from a normal distribution with mean 7 and standard deviation 1, it is
drawn from a standard normal distribution, as are the entries in the support. This
allows us to test the robustness of our method to the choice of distribution. Note that
Algorithm 2 is able to find an increasingly accurate approximation for the support in
this case, and will be able to recover the full support after a sufficiently large number
of iterations. Because the magnitudes of the corruptions are smaller, the algorithm
still has a chance of detecting the correct support using row norms to estimate S .
However, Algorithm 3 is able to obtain an accurate support estimate with a small
number of projections compared to Algorithm 2.
10 N. Durgin et al.
Fig. 2 Comparing cMMV-SRK and MMV-SRK for support recovery when there is a single
corrupt entry per signal whose magnitude is drawn from N (0, 1). (a) Φ ∼ N (0, 1). (b)
Φ ∼ U nif ([0, 1])
Fig. 3 Investigating the robustness of cMMV-SRK and MMV-SRK when a random number of
(multiple) corruptions are introduced. Here, a signal can have between 1 and 3 corruptions. The
corrupts magnitude of the corruptions are drawn from N (7, 1). (a) X matrix with 1–3 corruptions
per signal. (b) Performance of Algorithm 3 with multiple corruptions
In the next two experiments, we test the robustness of our proposed algorithm
against multiple corruptions. In Fig. 3, we allow for each signal to have multiple
corruptions. For each signal, i.e., column of X, we uniformly at random select an
integer from 1 to 3 to be the number of corruptions. The value of the corruptions
is drawn i.i.d. from N (7, 1) and an example of the resulting matrix can be seen in
Fig. 3a. The performance of the methods can be seen in Fig. 3b. We note that the
results of this experiment are very similar to those of the experiment in Fig. 1 since
the corruptions are drawn from the same distribution. Again, due to the use of row
norms, in the presence of multiple corruptions Algorithm 2 gives a less accurate
joint support estimate, recovering no more than about 15% of the support.
Fig. 4 Investigating the robustness of cMMV-SRK in a simulated online setting with a random
number of (multiple) corruptions. Here, a signal can have between 1 and 3 corruptions whose
magnitudes are drawn from N (7, 1) and we consider the over- and underdetermined linear system
settings. (a) Overdetermined linear system (m > n). (b) Underdetermined linear system (m < n)
Figure 4a shows the performance results for Algorithm 3 in a simulated online

setting. Instead of allowing the algorithm to loop for a fixed number of projections
for each signal, for approximately 90% of the signals τ̃j ∈ [5, 15] and for the
remaining signals τ̃j ∈ [95, 100]. The purpose of this is to simulate a variation in the
amount of time that a system has to work with a signal. The longer runs represent
stalls in the online setting. For each signal, we first draw a random Bernoulli variable
z with probability of success p = 0.1. If z = 1, then we choose an integer
in [95, 100] uniformly at random. If z = 0, then an integer in [5, 15] is chosen
uniformly at random. Algorithm 2 cannot be investigated under this setting as the
support estimate relies on processing all signals in every iteration. We note that with
respect to other parameters, the only difference between this experiment and that
in Fig. 3 is the size of J . We choose J to be large enough such that the maximal
number of projections made is 15,000 (as in Fig. 3).
The following experiment is motivated by compressed sensing and utilizes an
underdetermined linear systems as opposed to an overdetermined system. We repeat
the parameters as in Fig. 4a with the exception of the measurement matrix, which has
m = 100 rows and n = 500 columns, and the total number of signals J = 1500. The
results can be found in Fig. 4b. In the underdetermined case, the proposed algorithm
is still successful in recovering the support of the signal.
Finally, we test the robustness of our algorithm on the hyperspectral diffuse
optical tomography motivating problem discussed in the introduction. For this
experiment, we simulate absorption coefficient values for a two-dimensional circu-
lar sample of tissue of radius 25 cm, centered at the origin, with a circular tumor of
radius 5 cm centered at the point (−15,−10), see Fig. 5a. Each signal thus represents
a reconstruction of the absorption coefficient value at each point in a mesh of
size 541 over the sample area. The number of measurements corresponds to the
number of source–detector pairs in the imaging process. We use a random Gaussian
measurement matrix with m = 248 and n = 541, and have J = 200 total signals,
12 N. Durgin et al.
Fig. 5 Testing the robustness of cMMV-SRK in a real-world setting by using a simulated

hyperspectral diffuse optical tomography (hyDOT) model. Healthy cells are in Ω0 , while cancerous
cells are located in Ω1 . (a) Geometry of the real-world example. (b) X matrix with 1–3 corruptions
per signal/wavelength
Fig. 6 Investigating the

robustness of cMMV-SRK in
a simulated real-world setting
(hyperspectral diffuse optical
tomography) when random
multiple corruptions occur in
each signal. In this setting,
the measurement matrix is
underdetermined (m = 248,
n = 541)
each corresponding to a different wavelength at which the tissue was imaged. As

in previous experiments, 1–3 corruptions for each signal are drawn from a normal
distribution with mean equal to the average value of the absorption coefficient for the
cancerous cells at each wavelength, and standard deviation a quarter of the distance
between that value and the value of the absorption coefficient for the healthy cells.
The online setting is not used for this experiment. The results can be found in Fig. 6.
We see that the proposed algorithm is still successful in recovering the support of
the signal.
The experiments shown in this section highlight the usefulness of Algorithm 3
for support recovery of jointly sparse MMVs, especially in the presence of (even
quite large) corruptions. In each comparison between Algorithms 2 and 3, our
proposed method outperforms the previously proposed method for support recovery.
Additionally, the proposed method lends itself to the online setting.
5 Conclusion
In this work, we construct an algorithm for the support recovery of corrupted

jointly sparse MMV. Our empirical results demonstrate that the proposed algorithm
outperforms the previously proposed method, cMMV-SRK, for support recovery
of jointly sparse MMV, specifically when corruptions are present in the signal. Fur-
thermore, empirical evidence indicates that our method is robust to the magnitude of
corruptions and the number of corruptions. This improvement is due to the fact that
the support estimate in Algorithm 2, as many other signal recovery approaches for
the jointly sparse MMV problem, depends on the row norms of the signals, which
in this case would be dominated by corruptions. In comparison, the estimate for
Algorithm 3 only depends on the number of times an index appears in the support
estimate of a signal. Finally, our method lends itself well to the online setting when
measurement vectors are streaming in continuously. We leave the analytical study
of our method for future work.
Acknowledgements The initial research for this effort was conducted at the Research Collabora-
tion Workshop for Women in Data Science and Mathematics, July 17–21, held at ICERM. Funding
for the workshop was provided by ICERM, AWM, and DIMACS (NSF Grant No. CCF-1144502).
SL was supported by NSF CAREER Grant No. CCF−1149225. DN was partially supported by
the Alfred P. Sloan Foundation, NSF CAREER #1348721, and NSF BIGDATA #1740325. JQ was
supported by NSF DMS-1818374.
References
1. H.K. Aggarwal, A. Majumdar, Extension of sparse randomized Kaczmarz algorithm for

multiple measurement vectors, in IEEE International Conference on Pattern Recognition, pp.
1014–1019 (2014)
2. S.R. Arridge, J.C. Schotland, Optical tomography: forward and inverse problems. Inverse
Probl. 25(12), 123010 (2009)
3. L. Bottou, Large-scale machine learning with stochastic gradient descent, in Proceedings of
COMPSTAT (Springer, Berlin, 2010), pp. 177–186
4. E.J. Candès, C. Fernandez-Granda, Towards a mathematical theory of super-resolution.
Commun. Pure Appl. Math. 67(6), 906–956 (2014)
5. L. Fodor, M. Elman, Y. Ullmann, Aesthetic Applications of Intense Pulsed Light (Springer,
London, 2011)
6. R. Gordon, R. Bender, G.T. Herman, Algebraic reconstruction techniques (ART) for three-
dimensional electron microscopy and X-ray photography. J. Theor. Biol. 29, 471–481 (1970)
7. C. Hamaker, D.C. Solmon, The angles between the null spaces of X-rays. J. Math. Anal. Appl.
62(1), 1–23 (1978)
8. R. Heckel, M. Soltanolkotabi, Generalized line spectral estimation via convex optimization.
IEEE Trans. Inf. Theory 64(6), 4001–4023 (2018)
9. G. Herman, L. Meyer, Algebraic reconstruction techniques can be made computationally
efficient. IEEE Trans. Med. Imaging 12(3), 600–609 (1993)
10. S. Kaczmarz, Angenäherte auflösung von systemen linearer gleichungen. Bull. Int. Acad. Pol.
Sci. Lett. Ser. A 35, 335–357 (1937)
14 N. Durgin et al.
11. F. Larusson, S. Fantini, E.L. Miller, Hyperspectral image reconstruction for diffuse optical
tomography. Biomed. Opt. Express 2(4), 946–965 (2011)
12. J.N. Laska, M.A. Davenport, R.G. Baraniuk, Exact signal recovery from sparsely corrupted
measurements through the pursuit of justice, in Asilomar Conference on Signals, Systems, and
Computers (IEEE, Piscataway, 2009), pp. 1556–1560
13. Y. Li, Y. Chi, Off-the-grid line spectrum denoising and estimation with multiple measurement
vectors. IEEE Trans. Signal Process. 64(5), 1257–1269 (2016)
14. S. Li, D. Yang, G. Tang, M.B. Wakin, Atomic norm minimization for modal analysis from
random and compressed samples. IEEE Trans. Signal Process. 66(7), 1817–1831 (2018)
15. G. Lu, B. Fei, Medical hyperspectral imaging: a review. J. Biomed. Opt. 19(1), 010901 (2014)
16. H. Mansour, O. Yilmaz, A fast randomized Kaczmarz algorithm for sparse solutions of
consistent linear systems (2013). arXiv preprint arXiv:1305.3803
17. T. Strohmer, R. Vershynin, Comments on the randomized Kaczmarz method. J. Fourier Anal.
Appl. 15(4), 437–440 (2009)
18. C. Studer, P. Kuppinger, G. Pope, H. Bolcskei, Recovery of sparsely corrupted signals. IEEE
Trans. Inf. Theory 58(5), 3115–3130 (2012)
19. G. Tang, B.N. Bhaskar, P. Shah, B. Recht, Compressed sensing off the grid. IEEE Trans. Inf.
Theory 59(11), 7465–7490 (2013)
20. Z. Yang, L. Xie, Exact joint sparse frequency recovery via optimization methods. IEEE Trans.
Signal Process. 64(19), 5145–5157 (2014)
The Hubness Phenomenon
in High-Dimensional Spaces
Priya Mani, Marilyn Vazquez, Jessica Ruth Metcalf-Burton,

Carlotta Domeniconi, Hillary Fairbanks, Gülce Bal, Elizabeth Beer,
and Sibel Tari
Abstract High-dimensional data analysis is often negatively affected by the curse

of dimensionality. In high-dimensional spaces, data becomes extremely sparse and
distances between points become indistinguishable. As a consequence, reliable
estimations of density, or meaningful distance-based similarity measures, cannot
be obtained. This issue is particularly prevalent in clustering, which is commonly
employed in exploratory data analysis. Another challenge for clustering high-
dimensional data is that data often exist in subspaces consisting of combinations
of dimensions, with different subspaces being relevant for different clusters. The
hubness phenomenon is a recently discovered aspect of high-dimensional spaces.
It is observed that the distribution of neighbor occurrences becomes skewed
in intrinsically high-dimensional data, with few points, the hubs, having high
occurrence counts. Hubness is observed to be more pronounced with increasing
dimensionality. Hubs are also known to exhibit useful clustering properties and
could be leveraged to mitigate the challenges in high-dimensional data analysis.
In this chapter, we identify new geometric relationships between hubness, data
density, and data distance distribution, as well as between hubness, subspaces, and
intrinsic dimensionality of data. In addition, we formulate various potential research
directions to leverage hubness for clustering and for subspace estimation.
P. Mani · M. Vazquez · C. Domeniconi ()

George Mason University, Fairfax, VA, USA
e-mail: pmani@masonlive.gmu.edu; mvazque3@masonlive.gmu.edu; cdomenic@gmu.edu
J. R. Metcalf-Burton
National Security Agency, Fort Meade, MD, USA
H. Fairbanks
University of Colorado Boulder, Boulder, CO, USA
e-mail: hillary.fairbanks@colorado.edu
G. Bal · S. Tari
Middle East Technical University, Cankaya, Turkey
E. Beer
Center for Computing Sciences, Institute for Defense Analyses, Alexandria, VA, USA
© The Author(s) and the Association for Women in Mathematics 2019 15

E. Gasparovic, C. Domeniconi (eds.), Research in Data Science,
Association for Women in Mathematics Series 17,
https://doi.org/10.1007/978-3-030-11566-1_2
16 P. Mani et al.
1 Introduction
One of the key disciplines contributing to data science is machine learning, which
seeks to discover meaningful patterns or structure within the data. Key machine
learning paradigms are supervised and unsupervised learning. The first makes use of
data labels; the second does not. Most data in real-life scenarios do not have labels,
either because labels are unknown or are too costly to obtain. As such, unsupervised
learning plays an important role in data-driven research. In particular, a fundamental
unsupervised learning problem, extremely common in exploratory data mining, is
clustering. The goal in clustering is to discover groups data based on a notion of
similarity.
An issue related to clustering is the so-called curse of dimensionality [2].
Data with thousands of dimensions abound in fields and applications as diverse
as bioinformatics, security and intrusion detection, and information and image
retrieval. Clustering algorithms can handle data with low dimensionality, but as
the dimensionality of the data increases, these algorithms tend to break down.
This is because in high-dimensional spaces data become extremely sparse and
their distances become indistinguishable, a phenomenon also known as distance
concentration. As a consequence reliable density estimation cannot be performed,
and this affects negatively any learning algorithms that compute distances (or
similarities) in the full-dimensional input space (also known as the embedding
space).
A common scenario with high-dimensional data is that several clusters may exist
in different subspaces comprised of different combinations of features. In many real-
world problems, points in a given region of the input space may cluster along a
given set of dimensions, while points located in another region may form a tight
group with respect to different dimensions. Each dimension could be relevant to at
least one of the clusters. Common global dimensionality reduction techniques are
unable to capture such local structure of the data. Thus, a proper feature selection
procedure should operate locally in the input space. Local feature selection allows
one to estimate to which degree features participate in the discovery of clusters. As
a result, many different subspace clustering methods have been proposed [10, 14,
17, 18].
Another aspect of high-dimensional spaces which has recently come to light
is the phenomenon of hubness. It is observed that the distribution of neighbor
occurrences becomes skewed in intrinsically high-dimensional data. This means that
there are few data points with a high neighbor occurrence count, which emerge as
hubs. Though hubs and related power laws have been observed in other contexts
such as the Internet and protein–protein interactions, the phenomenon referred to
here is the concept of skewness of degree distributions in the k-nearest neighbor
topologies of high-dimensional data. More details on this phenomenon are given
in Sect. 2. Recent studies [19] have shown that hubness is an inherent property of
intrinsically high-dimensional data and could potentially be leveraged to improve
Another random document with
no related content on Scribd:
Beethoven entitled the next movement ‘a devout song of praise,
offered by a convalescent to God, in the Lydian mode.’ It probably
owes its origin to the fact that Beethoven was taken seriously ill while
at work on this and the B-flat major quartet. It seems likely that
before this illness he had other plans for the quartet, and that the
Danza tedesca before mentioned was to find a place in it.
The movement is long in performance but relatively simple in

structure. The chorale melody, simply harmonized, is preceded by a
short, preludizing phrase; and its strophes are set apart from each
other by short interludes in the same manner. After the chorale has
been once given, there is an episode in D major (Neue Kraft fühlend)
of blissful, gently animated character. The chorale is then repeated,
the melody an octave higher than before, the interludes and the
accompaniment complicated by syncopations. Once again the D
major episode, highly elaborated. Following this, the chorale is
introduced once more; but the introductory phrase is greatly
lengthened and developed, and there are suggested entrances of
the theme in all the instruments; nor does the complete theme make
itself heard, but only the first phrase of it seems ultimately to soar
aloft, in yet a higher register than before. So that this last section
may be taken as a coda, or as an apotheosis.
The short march which follows calls for no comment. The final
allegro is introduced by recitative passages for the first violin, gaining
in passion, culminating in a dramatic run over the diminished
seventh chord which bears some resemblance to the opening of the
allegro of the first movement. There is a passing sigh before the last
movement begins, Allegro appassionato.
Compared with the quartet in A minor, that in B-flat major is simple. It

is more in the nature of a suite than in that of a sonata, though the
first movement presents beneath an apparently irregular outline the
basis of the classical sonata-form. At first glance the frequent
changes of not only key signature but time signature as well are
confusing. The key signatures are now two flats, six flats, two sharps
and one sharp; and at the beginning, the middle and the end of the
movement the time is now triple, now duple, now slow, now fast.
The slow measures are related to the introduction, which here as in

other works of Beethoven is recalled at times in the main body of the
movement. The allegro makes a false start, in which the main
outlines of the first theme are suggested. From the second start,
however, the movement follows a relatively normal course. The first
theme is compound. On the one hand, there are rapid groups of
sixteenths, which play an important part in the whole movement; on
the other, a rhythmical motive, rather than a theme, first announced
by the second violin, which is the motto of the piece. The second
theme is first presented in G-flat major by the second violin and
immediately taken up by the first. At the beginning of the
development section and again in the coda use is made of the
motive of the introduction.
The second movement, a Presto in B-flat minor in alla breve time,

with a Trio in 6/4 time, is short and in the manner of a folk-song or
dance. It has no inner relation with the first movement; but it may be
said to breathe something of its spirit into the following andante (D-
flat major, common time). The kernel of the melody of this movement
may be found in the first measure, given by viola and 'cello; and this
kernel was sown, so to speak, by the previous movement. The viola
develops it in the second measure and the phrase is immediately
after taken up by the first violin.
For the fourth movement there is a rapid German waltz—Alla danza

tedesca—in G major. The fifth is a simple cavatina. Karl Holz, one of
the members of the Schuppanzigh quartet, has reported that
Beethoven could not read over the score of this short movement
without tears in his eyes. As the sixth movement there is the fugue,
published as opus 133, with a new dedication to Archduke Rudolph,
which was, as we have said, written for this quartet, and one of the
themes of which seems related to the chromatic motives of the A
minor quartet, on the one hand, and of the C-sharp minor quartet, on
the other; or there is the brilliant rondo with which Beethoven
replaced it at the behest of the publishers, and which is the last of
Beethoven’s compositions.
The fourth of the last quartets, in C-sharp minor, is dedicated to Field

Marshal Baron von Sutterheim, who interested himself deeply in the
affairs of Beethoven’s family. It is in some respects the most elusive,
in others the most unusual of all. Its various movements are
designated by numbers; yet two of them are so short that they need
not be regarded as separate movements, but only as transitional or
introductory sections. These are the third and the sixth. Furthermore,
a definite pause is justifiable only between the fourth and fifth. Thus,
in spite of the numbers, the work is closely blended into a whole, of
which the separate parts are not only æsthetically united, but
thematically complementary.
The first movement is a slow fugue, on a chromatic motive that

makes us once again remember that Beethoven was working on this
and the two preceding quartets at the same time. The fugue unfolds
itself with greatest smoothness and seeming simplicity. The texture
of the music is extremely close until near the end, where wide skips
appear in the various parts, like the movement of a more vigorous
life soon to break free in subsequent sections from such strict
restraint of form. One will find a perfect skill in technical details, such
as the diminution of the theme which appears in the first violin at the
change of signature, and the augmentation in the 'cello part in the
stretto not far before the end.
The fugue ends on a C-sharp unison, following a chord of C-sharp

major in seven parts. Then, as if this single C-sharp bore within itself
a secret harmonic significance, i.e., as the leading note in the scale
of D major, the whole fabric slips up half a tone in the opening notes
of the following movement, allegro molto vivace, D major—in 6/8
time. One cannot but feel the relationship between the delicate
convolutions of this new theme and the fugue theme. The whole
second movement hardly moves away from the motives of the
opening measures. A sort of complement to them may be found in
the successions of fourths which begin to rise up in the twenty-fifth
measure; and much farther on a sequence of chords beginning in F-
sharp major suggests some variety. But on the whole the movement
plays upon one theme, which recurs at intervals as in a rondo, but
after episodes that offer only in the main an harmonic contrast.
The third movement, allegro moderato, in common time, is a

recitative, begun in F minor and leading to a half-cadence in the
dominant seventh harmonies of A major, in which key the following
movement opens. We have here an andante and seven variations,
variations so involved and recondite that, though they may be clearly
perceived in the score, they will strike the unfamiliar ear as aimless
and inexplicable music.
The theme itself is in the form of a dialogue between first and second
violins. It merges into the first variation without perceptible break in
the music. Here the theme is carried by the second violin, the first
filling the pauses with a descending figure. This clause of the theme
is then repeated by the viola, the 'cello taking the rôle of the first
violin. The second clause of the theme is similarly treated.
The remaining six variations are clearly set apart from each other by
changes in the time signature. There is a variation marked piu
mosso, really alla breve, which is a dialogue between first violin and
'cello, accompanied at first monotonously by the other two
instruments, later with more variety and animation. The next is an
andante moderato e lusinghiero, in which the theme is arranged as a
canon at the second, first between the two lower instruments, later
between the two higher. This leads to an adagio in 6/8 time, in which
the theme is broken up into passage work. The next and fifth
variation (allegretto, 2/4) is the most hidden of all. The notes of the
theme are separated and scattered here and there among the four
parts. But the sixth, an adagio in 9/4 time, is simpler. The seventh,
and last, is a sort of epilogue, a series of different statements of the
theme, at first hidden in triplet runs; then emerging after a long trill, in
its simplest form, in the key of C major; then in A major with an
elaborated accompaniment; in F major, simple again; and finally
brilliantly in A major.
The following Presto in E major, alla breve, is very long, but is none
the less symmetrical and regular in structure. It is in effect a scherzo
and trio. The scherzo is in the conventional two sections, both of
which are built upon the same subject. The second section is broken
by four measures (molto poco adagio!); and there is a false start of
the theme, following these, in G-sharp minor, suddenly broken by a
hold. This recalls the effect of the very opening of the movement, a
single measure, forte, by the 'cello, as if the instrument were starting
off boldly with the principal subject. But a full measure of silence
follows, giving the impression that the 'cello had been too precipitate.
The Trio section offers at first no change of key; but a new theme is
brought forward. Later the key changes to A major, and the rhythm is
broadened. A series of isolated pizzicato notes in the various
instruments prepares the return of the Scherzo (without repeats).
The Trio follows again; and there is a coda, growing more rapid, after
the Scherzo has been repeated for the second time.
A short adagio, beginning in G-sharp minor, forms the sixth

movement, modulating to the dominant seventh in C-sharp minor.
The last movement is in sonata form. There are clearly a first theme
and a second theme, arranged according to rule. But the coda is
very long; and, even more important, not only the first and second
themes, but secondary themes and motives are all vaguely or
definitely related to the themes of the earlier movements. The first
theme, for all its somewhat barbaric character, is akin to the theme of
the first allegro in D major. In the episodes which follow, the notes of
the first violin and of the 'cello, in contrary motion, give a distinct
impression of the opening fugue theme. The second theme itself—in
E major—brings back a breath of the Trio, and Dr. Riemann finds in
the accompaniment suggestions of the fourth variation. Only a
detailed analysis could reveal the elaborate and intricate polyphony
which is in every measure in the process of weaving.
After the C-sharp minor quartet, the last quartet—in F major, opus
135—appears outwardly simple. It shares with the first of the series
simplicity and regularity of form; and is, like the quartet in E-flat
major, calm and outspoken, rather than disturbed, gloomy, or
mysterious. It is the shortest of all the last quartets.
The first movement is in perfect sonata form. The first theme (viola)
has a gently questioning sound, which one may imagine mocked by
the first violin. The second theme, in C major, is light, almost in the
manner of Haydn. The movement builds itself logically out of the
opposition of these two motives, the one a little touched with
sadness and doubt, the other confidently gay. The Scherzo which
follows needs no analysis. Two themes, not very different in
character, are at the basis. The second is presented successively in
F, G, and A, climbing thus ever higher. The climax at which it arrives
is noteworthy. The first violin is almost acrobatic in the expression of
wild humor, over an accompaniment which for fifty measures
consists of the unvaried repetition of a single figure by the other
three instruments in unison. Following this fantastical scherzo there
is a short slow movement in D-flat major full of profound but not
tragic sentiment. The short theme, flowing and restrained, undergoes
four variations; the second in C-sharp minor, rather agitated in
character; the third in the tonic key, giving the melody to the 'cello;
and the fourth disguising the theme in short phrases (first violin). To
the last movement Beethoven gave the title, Der schwer gefasste
Entschluss. Two motives which occur in it are considered, the one as
a question: Muss es sein? the other as the answer: Es muss sein.
The former is heard only in the introduction, and in the measures
before the third section of the movement. The latter is the chief
theme. Whether or not these phrases are related to external
circumstances in Beethoven’s life, the proper interpretation of them
is essentially psychological. The question represents doubt and
distrust of self. The answer to such misgivings is one of deeds, not
words, of strong-willed determination and vigorous action. Of such
the final movement of the last quartet is expressive. Such seems the
decision which Beethoven put into terms of music.
FOOTNOTES:
[70] The famous Schuppanzigh quartet met every Friday morning at the house of
Prince Lichnowsky. Ignaz Schuppanzigh (b. 1776) was leader. Lichnowsky himself
frequently played the second violin. Franz Weiss (b. 1788), the youngest member,
hardly more than a boy, played the viola. Later he became the most famous of the
viola players in Vienna. The 'cellist was Nikolaus Kraft (born 1778).
[71] Förster (1748-1823) forms an important link between Haydn and Beethoven.
[72] 2d edition, Berlin, 1913, pp. 482, et seq.
[73] Beethoven’s Streichquartette.
[74] Only Schuppanzigh himself, and Weiss, the violist, remained of the original
four who first played Beethoven’s quartets opus 18 at the palace of Prince
Lichnowsky. The second violinist was now Karl Holz, and the 'cellist Joseph Linke.
CHAPTER XVII
THE STRING ENSEMBLE SINCE
BEETHOVEN
The general trend of development: Spohr, Cherubini, Schubert—
Mendelssohn, Schumann and Brahms, etc.—New developments:
César Franck, d’Indy, Chausson—The characteristics of the
Russian schools: Tschaikowsky, Borodine, Glazounoff and others
—Other national types: Grieg, Smetana, Dvořák—The three great
quartets since Schubert and what they represent; modern
quartets and the new quartet style: Debussy, Ravel, Schönberg—
Conclusion.
I
There is little history of the string quartet to record after the death of
Beethoven in 1827. It has undergone little or no change or
development in technique until nearly the present day. The last
quartets of Beethoven taxed the powers of the combined four
instruments to the uttermost. Such changes of form as are to be
noted in recent quartets are the adaptation of new ideas already and
first put to test in music for pianoforte, orchestra, or stage. The
growth of so-called modern systems of harmony affect the string
quartet, but did not originate in it. A tendency towards richer or fuller
scoring, towards continued use of pizzicato or other special effects,
and a few touches of new virtuosity here and there, reflect the
general interest of the century in the orchestra and its possibilities of
tone-coloring. But it is in the main true that after a study of the last
quartets of Beethoven few subsequent quartets present new
difficulties; and that, excepting only a few, the many with which we
shall have to do are the expressions of the genius of various
musicians, most of whom were more successful in other forms, or
whose qualities have been made elsewhere and otherwise more
familiar.
Less perhaps than any other form will the string quartet endure by
the sole virtue of being well written for the instruments. Take, for
example, the thirty-four quartets of Ludwig Spohr. Spohr was during
the first half of the nineteenth century the most respected musician in
Germany. He was renowned as a leader, and composer quite as
much as he was world-famous as a virtuoso. He was especially
skillful as a leader in quartet playing. He was among the first to bring
out the Beethoven quartets, opus 18, in Germany. He was under a
special engagement for three years to the rich amateur Tost in
Vienna to furnish chamber compositions. No composer ever
understood better the peculiar qualities of the string instruments;
none was ever more ambitious and at the same time more serious.
Yet excluding the violin concertos and an occasional performance of
his opera Jessonda, his music is already lost in the past. Together
with operas, masses, and symphonies, the quartets, quintets, and
quartet concertos, are rapidly being forgotten. The reason is that
Spohr was more conscientious than inspired. He stood in fear of the
commonplace. His melodies and harmonies are deliberately
chromatic, not spontaneous. Yet shy as he was of
commonplaceness in melody and harmony, he was insensitive to a
more serious commonplaceness.
When we consider what subtle systems of rhythm the semi-civilized

races are masters of, we can but be astonished at the regularity of
our own systems. Only occasionally does a composer diverge from
the straight road of four-measure melody building. Yet is it not a little
subtlety even within this rigorous system that raises the great
composer above the commonplace? Certainly the ordinary in rhythm
most quickly wearies and disgusts the listener even if he is not
aware of it. Spohr’s rhythmical system was so little varied that
Wagner wrote of his opera Jessonda that it was 'alla Polacca’ almost
all the way through.
The thirty-five string quartets are fundamentally commonplace, for all

the chromaticism of their harmonies and melodies, and for all the
skillful treatment of the instruments. The double-quartets (four, in D
minor, E minor, E-flat major, and G minor) amount to compositions
for small string orchestra. There are, among the quartets, six so-
called ‘brilliant,’ which give to the first violin a solo rôle, and to the
other instruments merely accompaniment. It is hardly surprising that
the first violin is treated brilliantly in most of the quartets.
But the point is that Spohr’s quartets have not lived. In neatness of
form and in treatment of the instruments they do not fall below the
greatest. They are in these respects superior to those of Schumann
for example. The weakness of them is the weakness of the man’s
whole gift for composition; and they represent no change in the art of
writing string quartets.
Ludwig Spohr.
Another man whose quartets are theoretically as good as any is
Cherubini. Of the six, that in E-flat major, written in 1814, is still
occasionally heard.
On the other hand, Schubert, a man with less skill than either Spohr
or Cherubini, has written quartets which seem likely to prove
immortal. Fifteen are published in the complete Breitkopf and Härtel
edition of Schubert’s works. Of these the first eleven may be
considered preparatory to the last four. They show, however, what is
frequently ignored in considering the life and art of Schubert—an
unremitting effort on the part of the young composer to master the
principles of musical form.
The first of the great quartets, that in C minor—written in December,

1820—is but a fragment. Schubert completed but the first
movement. Why he neglected to add others remains unknown. But
the single movement is inspired throughout. The opening measures
give at once an example of the tremolo, of which Schubert made
great use in all his quartets. The general triplet rhythm is familiar in
all his later works. We have here the Schubert of the great songs, of
the B minor symphony, of the later pianoforte sonatas; warm,
intense, inspired.
Two quartets were written in 1824, that in A minor, published as opus

29, and that in D minor,[75] the best known of all his quartets. The A
minor is dedicated to Ignaz Schuppanzigh, with whom Schubert was
on friendly terms. The second movement of the quartet in D minor is
a series of variations on the song Der Tod und das Mädchen.
Finally there is the great quartet in G major, written in 1826, which

may be taken as representative throughout of the very best of
Schubert’s genius as it showed itself in the form. In it are to be found
all the qualities associated with Schubert especially. The opening
major triad, swelling to a powerful minor chord in eleven parts, and
the constant interchange of major and minor throughout the
movement; the tender second theme with its delicate folk-rhythm, its
unrestrained harmonies, its whispering softness in the variation after
the first statement; these could have been the work of Schubert
alone. Peculiar to Schubert’s treatment of the quartet are the
tremolo, and the general richness of scoring—the sixths for second
violin in the variation of the second theme, for example; the frequent
use of octaves and other double-stops, the eleven-voiced chord at
the beginning, and other such effects of fullness. There is little sign
of the polyphonic drawing which so distinguished the last quartets of
Beethoven. The quartet is made up of rich masses of sound that
glow warmly, and fade and brighten. The inner voices are used
measure after measure frankly to supply a richly vibrating harmony,
nothing more. And an occasional dialogue between two instruments
is all of polyphonic procedure one meets.
The beautiful andante in E minor begins with a melody for

violoncello, a true Schubertian melody, which is carried on for two
sections. Then a new spirit enters through hushed chords, and
breaks forth loudly in G minor. There follows a passage full of wild
passion. The agitated chords swell again and again to fortissimo. At
last they die away, only the monotonous F-sharp of the cello
suggests the throbbing of a despair not yet relieved. Over this the
first violin and the viola sing the opening melody. Later the hushed
tapping is given to other instruments and the cello takes up its
melody again. Once more the despair breaks wildly forth, and yet
again is hushed but not relieved. The sudden major in the ending
can not take from the movement its quality of unconsoled sadness.
The scherzo, in B minor, is built upon the constant imitation and play
of a single merry figure. The trio is in G major, one of those
seemingly naïve yet perfect movements such as Schubert alone
could write. There is only the swing of a waltz, only the melody that a
street gamin might carelessly whistle; but somewhere beneath it lies
genius. The interchange of phrases of the melody between the
different instruments, and the mellifluous counter-melodies, have
something the same sort of charm as the Scherzo of the symphony
in C major. The final movement is a rondo with a profusion of
themes. There are the familiar marks of Schubert: the triplet rhythm
(6/8), the shifting between major and minor; the full, harmonic style;
the naïve swing, the spontaneous and ever fresh melodies.
Schubert worked at the string quartet with special devotion.

Excepting the songs, his steady development toward perfect mastery
of his expression is nowhere better revealed than in the quartets.
Certainly the last two quartets are second only to the songs as proof
of his genius. There is that soft, whispering, quality in Schubert’s
music, for the expression of which the string quartet is a perfect
instrument. Much of Schubert is intimate, too, and happily suited to
the chamber. Less than any of the great composers did Schubert
make use of polyphonic skill. It is easy to say that he lacked it; but
what is hard to understand is how without it he could have
contributed to music some of its most precious possessions.
II
We may say that Schubert applied himself to the composition of
string quartets with a special devotion and ultimately with great
success; that certain qualities of his genius were suited to an
expression in this form. Mendelssohn applied himself to all branches
of music with equal facility and with evidently little preference. Most
of his chamber music for strings alone, however, belongs to the early
half of his successful career. This in the case of Mendelssohn does
not mean, as in the case of almost every other composer, that the
quartets may not be the expression of his fully-matured genius.
Mendelssohn never wrote anything better than the overture to
‘Midsummer Night’s Dream.’ This before he was twenty! But having
put his soul for once into a few quartets he passed on to other works.
There was a time when these quartets were considered a worthy

sequel to Beethoven’s. In the English translation of Lampadius’ ‘Life
of Mendelssohn’ occurs the sentence: ‘But in fact they [his works]
stand in need neither of approval nor defense: the most audacious
critic bows before the genius of their author; the power and weight of
public opinion would strike every calumniator dumb.’ And yet what
can now be said of Mendelssohn’s quartets save that they are
precise in form, elegant in detail?
There are six in all. The first, opus 12, is in E-flat major. The slow
introduction and the first allegro have all the well-known and now
often ridiculed marks of the ‘Songs Without Words’: short, regular
phrases; weak curves and feminine endings; commonplace
harmonies, monotonous repetitions, uninteresting accompaniment.
The second movement—a canzonetta—is interesting as
Mendelssohn could sometimes be in light pieces; but the andante
oozes honey again, and the final allegro is very long.
Is it unfair to dwell upon these wearisome deficiencies? Is there

anything substantially better in the last of the six, in the quartet in F
minor, opus 80? Here we have to do with one of the composer’s
agitated spells. There is a rough start and measures of tremolo for all
the instruments follow. This is the first theme, properly just eight
measures long and as thoroughly conventional as music well may
be. Then measures in recitative style, and again the first theme, and
its motives endlessly repeated. Suddenly the instruments in an
access of fury break into triplets; but this being calmed, the second
theme appears, as it should in A-flat major, a theme that positively
smirks.
But why attempt either analysis or description of works so patently

urbane? There is no meaning hidden in them; there is no richness of
sentiment; no harmonies out of new realms; no inspiration; nothing
really to study. Between the first two quartets mentioned and the last
in F minor there is a series of three (opus 44), one in D major, one in
E minor, and one in E-flat major. There is an ‘Andante, Scherzo,
Capriccio and Fugue’ for the four instruments, published as opus 81.
One turns to Schumann for a breath of more bracing air. Though

Schumann was first and foremost a composer for the pianoforte, and
though his quartets seem to be written in rather a pianoforte style,
yet there are flashes of inspiration in the music which must be
treasured, imperfect as the recording of them may be. There are
three quartets, composed in 1842 and dedicated to Mendelssohn. As
early as 1838 Schumann mentioned in letters to his sweetheart that
he had a string quartet in mind; but work in this direction was
seriously hindered by troubles with Wieck, which were growing daily
more acute. The second summer after his marriage, however, work
on the quartets was resumed; and the three were composed in the
short time of eight weeks, the last indeed apparently in five days (18-
22 July).
The first offers an harmonic innovation. The introduction is in A

minor, which is the principal key of the whole quartet; but the first
allegro is in F major. There is a Scherzo in A minor, with an
Intermezzo, not a Trio, in C major. In these first two movements the
habit of syncopation which gives much of his pianoforte music its
peculiar stamp is evident: in the first theme of the allegro; in the
measures which lead to the repetition of the first part; in the motive
of the Intermezzo, which is rhythmically similar to the first movement
and suggests some connection in Schumann’s mind. It is perhaps
the prevalence in all three quartets of the rhythmical devices which
we associate mostly with the pianoforte that raises a question of
propriety of style. The adagio is pure Schumann, in quality of melody
and accompaniment. Measures in the latter—noticeably the viola
figure which accompanies the first statement of the melody—look
upon the printed page like figures in a piano piece. Such figures are
not polyphonic. They are broken chords, the effect of which is
felicitous only on the pianoforte. The final presto suggests no little
the spirit of the first and last movements of the pianoforte quintet,
opus 44, which was composed in the following months. The whole
movement, except for a charming musette and a few following
measures of sustained chords just before the end, is built upon a
single figure.
The first movement of the next quartet (in F major) likewise suggests
the quintet. The style is smoothly imitative and compact; and the
theme beginning in the fifty-seventh measure casts a shadow before.
The Andante quasi Variazioni is most carefully wrought, and is rich in
sentiment. The Scherzo which follows—in C minor—is syncopated
throughout. The final allegro suggests the last movement of the B-
flat major symphony, the joyous Spring symphony written not long
before.
The last quartet (in A) may rank with the finest of his compositions.
Whether or not in theory the style is pianistic, the effect is rich and
sonorous. The syncopations are sometimes baffling, especially in the
last movement; but on the whole this quartet presents the essence of
Schumann’s genius in most ingratiating and appealing form. The
structure is free, reminding one in some ways of the D minor
symphony. But there is no rambling. The whole work is intense.
There is an economy of mood and of thematic material. One phrase
dominates the first movement; the Assai agitato is a series of terse
variations. There is a sustained Adagio in D major; and then a
vigorous finale in free rondo form, the chief theme of which is
undoubtedly related to the chief theme of the first movement.
It must be admitted that Schumann’s quartets are beautiful by

reason of their harmonies and melodies; that theirs is a fineness of
sentiment, not of style; that the luminous interweaving of separate
parts such as is found in the quartets of Haydn, Mozart, and
Beethoven, is not to be found in his. He follows rather Schubert, but
without Schubert’s instinct for instrumental color. So then one feels
that it happened that Schumann should seek expression thrice
through the medium of the string quartet; not that a certain quality of
inspiration within him demanded just that expression and none other.
His quartets represent neither a refinement nor an abstract of his
genius. They are of a piece with his pianoforte pieces and his songs;
as are likewise his symphonies. We admire and love all for the same
qualities.
Brahms, who for so many reasons we may think of as taking up

German music where Schumann left it, published only three string
quartets. That he had written many others which he had chosen to
discard before the two quartets, opus 51, were published in 1873, is
evident from the note to Dr. Billroth concerning a dedication.[76]
Several pianoforte quartets, and two sextets for two violins, two
violas and two violoncellos, opus 18 and opus 36, are closely related
to the string quartet. The sextets are especially noteworthy.
The first sextet, in B-flat major, has won more popular favor than
many other works by the same composer. The addition of two
instruments to the regular four brought with it the same sort of
problems which were mentioned in connection with Mozart’s
quintets: i.e., the avoidance of thickness in the scoring. The group of
six instruments is virtually a string orchestra; but the sextets of
Brahms are finely drawn, quite in the manner of a string quartet.
Especially in this first sextet have the various instruments a like
importance and independence.
The first theme of the first movement (cello) is wholly melodious. The
second theme, regularly brought forward in F major, is yet another
melody, and again is announced by the violoncello. A passage of
twenty-eight measures, over a pedal point on C, follows. This closes
the first section. The development is, as might be expected, full of
intricacies. The return of the first theme is brilliantly prepared,
beginning with announcing phrases in the low registers, swelling to a
powerful and complete statement in which the two violins join. The
second movement is a theme and variations in D minor. The theme
is shared alternately by first viola and first violin. The variations are
brilliant and daring, suggesting not a little the pianoforte variations on
a theme of Paganini’s. There is a Scherzo and Trio. The main motive
of the Scherzo serves as an accompaniment figure in the Trio; and
the Trio is noteworthy for being entirely fortissimo. The last
movement is a Rondo.
The second sextet, in G major, is outwardly less pleasing; and like

much of Brahms’ music is veiled from the casual or unfamiliar
listener.
The first movement (allegro non troppo) opens mysteriously with a

trill for first viola, which continues through the next thirty-two
measures. In the third the first violin announces, mezza voce, the
main theme of the movement; of which the chief characteristic is two
upward fifths (G—D—E-flat—B-flat). The second theme appears
after an unexpected modulation in D major, and is given to the first
violoncello. The striding fifths sound again in the closing measures of
the first section. The development begins with these fifths employed
as a canon, in contrary motion; and the same intervals play a
prominent part in the entire section. The recapitulation is regular. The
following Scherzo (Allegro non troppo, G minor) has a touch of
Slavic folk-music. There is a Trio section in G major. The slow
movement is, as in the earlier sextet, a theme and variations. The
last is in sonata form. The first theme may be divided into two wholly
contrasting sections, of which the second is melodiously arranged in
sixths. The second theme is given out regularly in D major by the
violoncello. There is a long coda, animato, which is practically a
repetition of much of the development section.
In these sextets and in the three quartets, written many years later,
we have the classical model faithfully reproduced. The separate
parts are handled with unfailing polyphonic skill; there is the special
refinement of expression which, hard to define, is unmistakable in a
work that is properly a string quartet.
Opus 51, No. 1, is in C minor. The first theme is given out at once by
the first violin; a theme characteristic of Brahms, of long phrases and
a certain swinging power. Within the broadly curving line there are
impatient breaks; and the effect of the whole is one of restlessness
and agitation. This is especially noticeable when, after a contrasting
section, the theme is repeated by viola and cello under an agitated
accompaniment, and leads to sharp accents. There is no little
resemblance between this theme and Brahms’ treatment of it, and
the theme of the first movement of the C minor symphony,
completed not long before. There is throughout this movement the
rhythm, like the sweep of angry waves, which tosses in the first
movement of the symphony; an agitation which the second theme
(B-flat major, first violin) cannot calm, which only momentarily—as
just after the second theme, here, and in the third section of the
movement—is subdued.

PDF Research in Data Science Ellen Gasparovic Ebook Full Chapter

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDF Research in Data Science Ellen Gasparovic Ebook Full Chapter

Uploaded by

Copyright:

Available Formats

Research in Data Science Ellen

Analytics and Data Science: Advances in Research and

Data Science and Social Research II: Methods,

Frontiers In Data Science Matthias Dehmer

Ethics in Science Ethical Misconduct in Scientific

Real Time Stream Data Management Push Based Data in

Frontiers in Data Science 1st Edition Matthias Dehmer

Uncertainty Modelling in Data Science Sébastien

Advances in Panel Data Analysis in Applied Economic

Focusing on the groundbreaking work of women in mathematics past, present, and

More information about this series at http://www.springer.com/series/13764

Research in Data Science

ISSN 2364-5733 ISSN 2364-5741 (electronic)

Library of Congress Control Number: 2019935518

© The Author(s) and the Association for Women in Mathematics 2019

The Cross-Disciplinary Field of Data Science

Data science is a cross-disciplinary field relying on methodologies and practices

patterns in data for actionable prediction and decision-making. As such, it is an

embedding subspaces of data. Potential research directions to leverage hubness for

random and compressive measurements. Finally, the authors empirically validate

Fairfax, VA, USA Carlotta Domeniconi

Sparse Randomized Kaczmarz for Support Recovery of Jointly

Feature Design for Protein Interface Hotspots Using

© The Author(s) and the Association for Women in Mathematics 2019 1

1.1 Problem Formulation

Given a set of linear measurements Y (·,j ) ∈ Rm and a measurement matrix Φ ∈

Y (·,j ) = ΦX(·,j ) for j = 1, . . . J, (2)

supp(X(·,j ) ) = S ∪ Cj , Cj ⊂ {1, . . . , n},

2 Related and Existing Work

2.1 Sparse Randomized Kaczmarz

In this work, we utilize the sparse randomized Kaczmarz (SRK) algorithm to

Algorithm 1 Sparse Randomized Kaczmarz

2.2 SRK for MMV

In this section we present an SRK-based approach to the MMV setting proposed by

Algorithm 2 Sparse Randomized Kaczmarz for MMV

3.1 Corrupted MMV

To review, we are interested in constructing an algorithm that recovers the support

In this section, we compare Algorithms 2 and 3 under a variety of settings, specif-

Algorithm 3 Sparse Randomized Kaczmarz for corrupted MMV

S t c is the complement set of S t

Table 1 Summary of experiment parameters

uniform distribution ranging from 0 to 1. We note that, in both cases, Algorithm 3

Figure 4a shows the performance results for Algorithm 3 in a simulated online

Fig. 5 Testing the robustness of cMMV-SRK in a real-world setting by using a simulated

Fig. 6 Investigating the

each corresponding to a different wavelength at which the tissue was imaged. As

In this work, we construct an algorithm for the support recovery of corrupted

1. H.K. Aggarwal, A. Majumdar, Extension of sparse randomized Kaczmarz algorithm for

Priya Mani, Marilyn Vazquez, Jessica Ruth Metcalf-Burton,

Abstract High-dimensional data analysis is often negatively affected by the curse

P. Mani · M. Vazquez · C. Domeniconi ()

© The Author(s) and the Association for Women in Mathematics 2019 15

The movement is long in performance but relatively simple in

Compared with the quartet in A minor, that in B-flat major is simple. It

The slow measures are related to the introduction, which here as in

The second movement, a Presto in B-flat minor in alla breve time,

For the fourth movement there is a rapid German waltz—Alla danza

The fourth of the last quartets, in C-sharp minor, is dedicated to Field

The first movement is a slow fugue, on a chromatic motive that

The fugue ends on a C-sharp unison, following a chord of C-sharp

The third movement, allegro moderato, in common time, is a

A short adagio, beginning in G-sharp minor, forms the sixth

[72] 2d edition, Berlin, 1913, pp. 482, et seq.