Compressive sensing of earth

observations 1st Edition Chen Chi-Hau

Visit to download the full and correct content document:
https://textbookfull.com/product/compressive-sensing-of-earth-observations-1st-editio
n-chen-chi-hau/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Compressive Sensing in Healthcare (Advances in

ubiquitous sensing applications for healthcare) 1st
Edition Mahdi Khosravy (Editor)

https://textbookfull.com/product/compressive-sensing-in-
healthcare-advances-in-ubiquitous-sensing-applications-for-
healthcare-1st-edition-mahdi-khosravy-editor/

Satellite Earth Observations and Their Impact on

Society and Policy 1st Edition Masami Onoda

https://textbookfull.com/product/satellite-earth-observations-
and-their-impact-on-society-and-policy-1st-edition-masami-onoda/

Geodynamics and Earth Tides Observations from Global to

Micro Scale Carla Braitenberg

https://textbookfull.com/product/geodynamics-and-earth-tides-
observations-from-global-to-micro-scale-carla-braitenberg/

Understanding Earth Observation The Electromagnetic

Foundation of Remote Sensing 1st Edition Domenico
Solimini (Auth.)

https://textbookfull.com/product/understanding-earth-observation-
the-electromagnetic-foundation-of-remote-sensing-1st-edition-
domenico-solimini-auth/
US Firms Business Competence in the Taiwanese IT
Industry 1st Edition Pi-Chi Chen

https://textbookfull.com/product/us-firms-business-competence-in-
the-taiwanese-it-industry-1st-edition-pi-chi-chen/

Remote Sensing Advances for Earth System Science The

ESA Changing Earth Science Network Projects 2011 2013
1st Edition Diego Fernández-Prieto

https://textbookfull.com/product/remote-sensing-advances-for-
earth-system-science-the-esa-changing-earth-science-network-
projects-2011-2013-1st-edition-diego-fernandez-prieto/

Key Technologies of Intelligentized Welding

Manufacturing Visual Sensing of Weld Pool Dynamic
Characters and Defect Prediction of GTAW Process
Zongyao Chen
https://textbookfull.com/product/key-technologies-of-
intelligentized-welding-manufacturing-visual-sensing-of-weld-
pool-dynamic-characters-and-defect-prediction-of-gtaw-process-
zongyao-chen/

Industrial Networks and Intelligent Systems 3rd

International Conference INISCOM 2017 Ho Chi Minh City
Vietnam September 4 2017 Proceedings 1st Edition
Yuanfang Chen
https://textbookfull.com/product/industrial-networks-and-
intelligent-systems-3rd-international-conference-iniscom-2017-ho-
chi-minh-city-vietnam-september-4-2017-proceedings-1st-edition-
yuanfang-chen/

The Beginnings of Chinese Civilization Chi Li

https://textbookfull.com/product/the-beginnings-of-chinese-
civilization-chi-li/
 COMPRESSIVE
SENSING OF EARTH
OBSERVATIONS
 Signal and Image Processing of
Earth Observations Series

Series Editor
C.H. Chen

Published Titles

Compressive Sensing of Earth Observations

C.H. Chen
Radar Imaging for Maritime Observation
Fabrizio Berizzi, Marco Martorella, and Elisa Giusti
Principles of Synthetic Aperture Radar Imaging: A System
Simulation Approach
Kun-Shan Chen
Remote Sensing Image Fusion
Luciano Alparone, Bruno Aiazzi, Stefano Baronti, and
Andrea Garzelli
 COMPRESSIVE
SENSING OF EARTH
OBSERVATIONS

Edited by
C.H. Chen
MATLAB R is a trademark of The MathWorks, Inc. and is used with permission. The MathWorks does not warrant the
accuracy of the text or exercises in this book. This book’s use or discussion of MATLAB R software or related products
does not constitute endorsement or sponsorship by The MathWorks of a particular pedagogical approach or particular
use of the MATLAB R software.

CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
c 2017 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Printed on acid-free paper

International Standard Book Number-13: 978-1-4987-7437-6 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been
made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity
of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright
holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we
may rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or
utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including
photocopying, microfilming, and recording, or in any information storage or retrieval system, without written
permission from the publishers.
For permission to photocopy or use material electronically from this work, please access www.copyright.com
(http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers,
MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of
users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been
arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.

Library of Congress Cataloging-in-Publication Data

Names: Chen, C. H. (Chi-hau), 1937- editor.

Title: Compressive sensing of earth observations / edited by C.H. Chen.
Description: Boca Raton, FL : Taylor & Francis, 2017. | Includes
bibliographical references and index.
Identifiers: LCCN 2016050152 | ISBN 9781498774376 (print : alk. paper)
Subjects: LCSH: Earth sciences--Remote sensing. | Compressed sensing
(Telecommunication)
Classification: LCC QE33.2.R4 C66 2017 | DDC 550.28/7--dc23
LC record available at https://lccn.loc.gov/2016050152

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
This book is dedicated to my wife, Wanda,
my sons, Ivan and Christopher,
and grandsons, Jeremy and Benjamin.
Contents

Preface ix

Editor xiii

Contributors xv

1 Compressed Sensing: From Theory to Praxis 1

Axel Flinth, Ali Hashemi, and Gitta Kutyniok

2 Compressive Sensing on the Sphere: Slepian Functions for

Applications in Geophysics 33
Zubair Khalid and Abubakr Muhammad

3 Compressive Sensing–Based High-Resolution Imaging and Tracking

of Targets and Human Vital Sign Detection behind Walls 55
Ahmad Hoorfar, Ozlem Kilic, and Aly E. Fathy

4 Recovery Guarantees for High-Resolution Radar Sensing with

Compressive Illumination 83
Nithin Sugavanam and Emre Ertin

5 Compressive Sensing for Inverse Synthetic Aperture Radar Imaging 105

Alessio Bacci, Elisa Giusti, Sonia Tomei, Davide Cataldo, Marco Martorella,
and Fabrizio Berizzi

6 A Novel Compressed Sensing–Based Algorithm for Space–Time

Signal Processing Using Airborne Radars 131
Jing Liu, Mahendra Mallick, Feng Lian, and Kaiyu Huang

7 Bayesian Sparse Estimation of Radar Targets in the Compressed

Sensing Framework 153
Stéphanie Bidon, Marie Lasserre, and François Le Chevalier

8 Virtual Experiments and Compressive Sensing for Subsurface

Microwave Tomography 177
Martina Bevacqua, Lorenzo Crocco, Loreto Di Donato, Tommaso Isernia,
and Roberta Palmeri

9 Seismic Source Monitoring with Compressive Sensing 199

Ismael Vera Rodriguez and Mauricio D. Sacchi

10 Seismic Data Regularization and Imaging Based on Compressive

Sensing and Sparse Optimization 227
Yanfei Wang and Jingjie Cao

vii
viii Contents

11 Land Use Classification with Sparse Models 257

Mohamed L. Mekhalfi, Farid Melgani, Yakoub Bazi, and Naif Alajlan

12 Compressive Sensing for Reconstruction, Classification, and

Detection of Hyperspectral Images 273
Bing Zhang, Wei Li, Lianru Gao, and Xu Sun

13 Structured Abundance Matrix Estimation for Land Cover

Hyperspectral Image Unmixing 299
Paris V. Giampouras, Konstantinos E. Themelis, Athanasios A. Rontogiannis,
and Konstantinos D. Koutroumbas

14 Parallel Coded Aperture Method for Hyperspectral Compressive

Sensing on GPU 313
Gabriel Martı́n, José Nascimento, and José Bioucas-Dias

15 Algorithms and Prototyping of a Compressive Hyperspectral Imager 329

Alessandro Barducci, Giulio Coluccia, Donatella Guzzi, Cinzia Lastri,
Enrico Magli, and Valentina Raimondi

Index 351
Preface

Compressive sensing (CS), a new area of signal processing, has been around for 15 years. Its
theory has now been mostly developed. With increased high-resolution sensor development
for remote sensing, we are dealing with an enormous amount of earth observation data. CS
is particularly important for meeting this challenge. CS exploits the sparsity of the remote
sensing data and also allows for reduced sampling. It can considerably simplify the data
acquisition design for remote sensing systems and speed up the process of extracting the
desired information from a large amount of data. It is believed that future remote sensing
systems will make extensive use of CS and will include CS as part of the system design.
Indeed, CS has enormous potential for many aspects of remote sensing.
Several recent efforts to implement or incorporate CS in remote sensing systems are
presented in this book. The book consists, somewhat loosely, of five parts. Part 1 provides
a basic review of the theory and its transition to practice in Chapter 1 and an overview
of the Slepian functions in Chapter 2. Part 2, containing Chapters 3 through 7, discusses
CS for radar and microwave remote sensing. This is the area that has benefited the most
from the use of CS. Part 3, with Chapters 8 through 10, covers CS for subsurface sensing.
Part 4, with Chapters 11 through 13, examines CS for multispectral and hyperspectral
images. Part 5, with Chapters 14 and 15, covers other issues of CS for remote sensing,
such as implementation involving hyperspectral remote sensing images. This book was
largely inspired by the well-written chapter “Compressive Remote Sensing,” by J. Ma et al.,
published as Chapter 5 of Signal and Image Processing for Remote Sensing, 2nd edition,
edited by C.H. Chen, CRC Press, 2012, and the well-edited book, Compressive Sensing for
Urban Radar , by M. Amin, CRC Press, 2014.
Digital signal processing has experienced significant growth since the development (or
discovery) of the fast Fourier transform in the early 1960s. Since then, the methodologies of
adaptive filtering, spectrum estimation, time–frequency analysis, wavelet transforms, neural
networks and machine learning, deconvolutions, and so on have dominated the rapid progress
in one-dimensional and two-dimensional (image) signal processing. The success of all of
these methodologies was largely built on sound mathematical development and advances.
CS was built on the mathematical theory of sparse representation. As with previous major
advances in signal processing, the many advances in CS and the meeting of continued
challenges was made possible by the close collaboration of mathematicians, engineers, and
scientists. Though this volume is more concerned with the application of CS for earth
observations, we are very fortunate to have leading mathematician Prof. Dr. Gitta Kutyniok
and her colleagues as the authors of Chapter 1, “Compressive Sensing: From Theory to
Praxis.” This chapter provides a clear and rigorous introduction to CS as a mathematical
topic in signal processing. A significant part of the chapter is devoted to some sample
applications such as medical imaging, data separation, analog-to-information converters,
and sparse channel estimation, among others. Chapter 2, “Compressive Sensing on the
Sphere: Slepian Functions for Applications in Geophysics,” was authored by Drs. Khalid
and Muhammad. As noted by the authors, the Slepian functions are particularly useful
for CS in geophysics, which deals with earth observations. The remaining chapters are
concerned with practical CS applications for earth observations.

ix
x Preface

Chapter 3, by Prof. Hoofar et al., titled “Compressive Sensing–Based High-Resolution

Imaging and Tracking of Targets and Human Vital Sign Detection behind Walls,” outlines
their recent works in the application of CS in synthetic aperture radar (SAR), multiple-
input multiple-output (MIMO) radar imaging, and simultaneous tracking of multiple targets
behind multilayer walls or other layered media. The chapter also presents the use of CS in
detection of human vital signs in a cluttered environment and discusses the corresponding
modeling of rib-cage breathing. The chapter concludes with a discussion on hardware
design of low complexity receivers with reduced analog-to-digital conversion (ADC)
requirements suitable for SAR, MIMO, ultra-wideband, or stepped frequency. Chapter 4,
by Prof. Ertin and N. Suvaganam, discusses recovery guarantees for high-resolution
radar sensing with compressive illumination. The chapter presents a comprehensive radar
architecture that combines multitoned linear frequency modulated waveforms on transmit
with classical stretch processor and sub-Nyquist sampling on receive. The proposed
compressive illumination scheme has fewer random elements than previously proposed
compressive radar designs based on stochastic waveforms, resulting in reduced storage and
complexity for implementation. Bounds on the operator norm and mutual coherence are
presented such that high-resolution recovery is guaranteed for a sparse scene using sampling
rates that scale linearly with the scene sparsity. Chapter 5, by Dr. Bacci et al., is on CS
for inverse SAR imaging. The applicability of CS to the radar imaging problem has been
justified by observing that the radar signal is sparse when represented in the 2D Fourier
domain, since it can be represented by a superposition of a few prominent scatter responses
in the radar image plane with respect to the pixels in the image. Three different applications
of CS to the Inverse Synthetic Aperture Radar (ISAR) imaging problem are discussed,
namely image reconstruction from compressed data, resolution enhancement, and image
reconstruction from incomplete data. Chapter 6, by Prof. Liu et al., is on a novel CS–based
algorithm for space–time signal processing using airborne radars. The chapter presents
the general similar sensing matrix pursuit algorithm to reconstruct the sparse radar scene
directly using the test cell, which reduces the computational complexity significantly. The
algorithm can efficiently cope with the deterministic sensing matrix with high coherence;
it can estimate weak elements (targets) as well as strong elements (clutter) in the DOA-
Doppler plane accurately and distinguish the targets from clutter successfully.
Chapter 7, by Prof. Bidon et al., covers Bayesian sparse estimation of radar targets
in the CS framework. The chapter investigates the use of a wideband waveform jointly
with sparse recovery techniques to unambiguously reconstruct the target signal. Several
hierarchical Bayes models are presented where sparsity of the target scene is induced via
sparse-promoting prior. Advantages of the Bayes approach are clearly presented.
Chapter 8, by Prof. Di Donato et al., on virtual experiments and CS for subsurface
microwave tomography, presents a CS-inspired tomographic approach for processing ground-
penetrating radar (GPR) data. The approach exploits the recently introduced Virtual
Experiments framework, which is applied to surface and borehole GPR surveys for
quantitative characterization of buried targets. Chapter 9, by Drs. Rodriguez and Sacchi, on
seismic source monitoring with CS, illustrates practical aspects regarding the incorporation
of CS into a system for continuous monitoring of seismic activity. Solving the continuous
monitoring system as a sparse representation presents an advantage over the least-squares
solution in the form of enhanced resolution in event location, particularly when noise is
present in the observations. Sparsity makes the problem amenable to the incorporation of
CS. Chapter 10, by Drs. Wang and Cao, provides a detailed introduction to seismic data
regularization and imaging based on CS and sparse optimization. Numerical experiments
based on some reviewed models and methods prove the efficiency of CS theory and sparse
optimization.
Preface xi

Chapter 11, by Prof. Melgani et al., on land use classification with sparse models,
formulates the land use classification problem within a CS fusion framework. Residuals are
generated from the image reconstruction with dictionaries associated with the available set
of possible land uses and gathered to form a single-feature image pattern. The patterns
obtained from different types of spatial/spectral features are then used to provide the final
land use estimate. As demonstrated by experiments run on the basis of a public benchmark
database, the presented schemes can achieve significant classification accuracy gains over
recent reference methods. With the development of the CS paradigm, sparse representation-
based classification and detection have drawn much attention in hyperspectral imagery. With
this in mind, Chapter 12, by Prof. Zhang et al., on CS for reconstruction, classification, and
detection of hyperspectral images, introduces three reconstruction methods from random
projection, three representation-based classifiers and their extensions, and two one-class
representation-based target detectors and their improved versions. Chapter 13, by Prof.
Rontogiannis et al., proposes a novel semi-supervised algorithm for the problem of abundance
matrix estimation in land cover hyperspectral image unmixing. The abundance matrix being
sought is constrained to be simultaneously sparse and low rank.
Chapter 14, by Prof. Nascimento et al., covers the parallel coded aperture method for
hyperspectral CS on Graphics Processing Unit (GPU). Reduction of the large amount of
continuously acquired hyperspectral image data is much needed. Hyperspectral images are
often highly compressible owing to the very high spatial and spectral correlation. Therefore,
this imaging modality is a perfect candidate for applying CS technology. A simple process
can be implemented on the optics of the acquisition system onboard the satellite, while in the
CS model the reconstruction is performed at the ground station, where a variety of powerful
computational resources may be available. The chapter describes the computationally
efficient implementation of two reconstruction algorithms in graphics process units. The
hardware is able to exploit data parallelism through a single-instruction, multiple-data
computing architecture. Finally, Chapter 15, by Prof. Magli et al., on algorithms and
prototyping of a compressed hyperspectral imager, deals with the practical implementation
of a compressive hyperspectral imager, highlighting design and data quality issues. Details
of the instruments employed as well as image acquisition and recovery are discussed in
depth. From Chapters 14 and 15, we may conclude that implementation is a new challenge
in designing CS systems for hyperspectral images.
Based on the development thus far, we may conclude that radar remote sensing and
hyperspectral image sensing have been the major users of CS. Much continued effort is
expected to be driven by the need for system design.
The rapid advances in nearly all areas of science and engineering over the past half
century have been the result of research and publications. Although journal publications
and textbooks still play a fundamentally important role among publications, conference
publications and edited books have also become very important. Since the publication of
my first edited book 40 years ago, I have had a particular appreciation for the unique
role of edited books in advancing knowledge, which may often be overlooked by readers.
Contributors to edited books often do not get proper credit or recognition for making
the book possible. Preparing an edited book requires much effort from both editors and
contributors. Readers can only appreciate the rich content of an edited book through careful
reading of each chapter or certain chapters. I am sure that there will be many more books in
the foreseeable future on CS theory and applications, and I hope that readers will give the
chapters of this book a careful reading. As in past edited books, my thanks and appreciation
go to all contributors of the book for their important role in advancing technical knowledge.
C.H. Chen
University of Massachusetts Dartmouth
xii Preface

MATLAB R is a registered trademark of The MathWorks, Inc. For product information,

please contact:

The MathWorks, Inc.

3 Apple Hill Drive
Natick, MA 01760-2098
USA Tel: 508-647-7000
Fax: 508-647-7001
Email: info@mathworks.com
Web: www.mathworks.com
Editor

C.H. Chen was born on December 22, 1937. He received his PhD in electrical engineering
from Purdue University, West Lafayette, Indiana, in 1965, his MSEE from the University of
Tennessee, Knoxville, in 1962, and his BSEE from the National Taiwan University, Taipei,
in 1959.
He is currently the chancellor professor emeriti of electrical and computer engineering at
the University of Massachusetts Dartmouth, where he has been a faculty member since 1968.
His research areas encompass statistical pattern recognition and signal/image processing
with applications to remote sensing, geophysical, underwater acoustics, and nondestructive
testing problems, as well as computer vision for video surveillance, time-series analysis, and
neural networks.
Dr. Chen has edited and authored 34 books in his area of research. He is the editor
of Digital Waveform Processing and Recognition (CRC Press, 1982), Signal Processing
Handbook (Marcel Dekker, 1988), and Signal and Image Processing for Remote Sensing
(CRC Press, first edition 2006 and second edition 2012). He is the chief editor of Handbook
of Pattern Recognition and Computer Vision, volumes 1–5 (World Scientific Publishing,
1993, 1999, 2005, 2010, and 2016, respectively). He is the editor of Fuzzy Logic and Neural
Network Handbook (McGraw-Hill, 1966). In the field of remote sensing, he also edited
the books Information Processing for Remote Sensing and Frontiers of Remote Sensing
Information Processing (World Scientific Publishing, 1999 and 2003, respectively).
He served as associate editor of the IEEE Transactions on Acoustics, Speech, and Signal
Processing for 4 years, associate editor of the IEEE Transactions on Geoscience and Remote
Sensing for 15 years, and he has been a board member of Pattern Recognition since 2008.
Dr. Chen has been a Fellow of the Institute of Electrical and Electronic Engineers (IEEE)
since 1988, a Life Fellow of the IEEE since 2003, and Fellow of the International Association
of Pattern Recognition (IAPR) since 1996. He is also editor of a book series, Signal and
Image Processing of Earth Observations, from CRC Press.

xiii
Contributors

Naif Alajlan José Bioucas-Dias

College of Computer and Information Instituto de Telecomunicações
Sciences Instituto Superior Técnico
King Saud University Lisboa, Portugal
Riyadh, Saudi Arabia
Jingjie Cao
Alessio Bacci Hebei GEO University
CNIT Radar and Surveillance System Shijiazhung, Hebei, China
(RaSS) National Laboratory Davide Cataldo
and CNIT Radar and Surveillance System
Department of Information Engineering (RaSS) National Laboratory
University of Pisa and
Pisa, Italy Department of Information Engineering
Alessandro Barducci University of Pisa
SOFASI SRL Pisa, Italy
Viale Alessandro Guidoni Giulio Coluccia
Firenze, Italy Dipartimento di Elettronica e
Yakoub Bazi Telecomunicazioni
Politecnico di Torino
College of Computer and Information
Torino, Italy
Sciences
King Saud University Lorenzo Crocco
Riyadh, Saudi Arabia Institute for Electromagnetic Sensing of the
Environment
Fabrizio Berizzi
National Research Council of Italy
CNIT Radar and Surveillance System Naples, Italy
(RaSS) National Laboratory
Loreto Di Donato
and
Department of Electrical, Electronics and
Department of Information Engineering Computer Engineering (DIEEI)
University of Pisa University of Catania
Pisa, Italy Catania, Italy
Martina Bevacqua Emre Ertin
Department of Information Engineering, Department of Electrical and Computer
Infrastructures and Substainable Energy Engineering
University of “Mediterranea” of Reggio The Ohio State University
Calabria Columbus, Ohio
Reggio di Calabria, Italy
Aly E. Fathy
Stéphanie Bidon Antenna Research Laboratory
Department of Electronics Optronics and Department of Electrical Engineering and
Signal Computer Science
University of Toulouse–ISAE-Supaéro University of Tennessee
Toulouse, France Knoxville, Tennessee

xv
xvi Contributors

Axel Flinth Tommaso Isernia

Institut für Mathematik Department of Information Engineering,
Technische Universität Berlin Infrastructures and Substainable Energy
Berlin, Germany University of “Mediterranea” of Reggio
Calabria
Lianru Gao
Reggio di Calabria, Italy
Institute of Remote Sensing and Digital
Earth Zubair Khalid
Chinese Academy of Sciences Department of Electrical Engineering
Beijing, China School of Science and Engineering
Paris V. Giampouras Lahore University of Management Sciences
IAASARS Lahore, Pakistan
National Observatory of Athens
Penteli, Greece Ozlem Kilic
Department of Electrical and Computer
Elisa Giusti Science
CNIT Radar and Surveillance System The Catholic University of America
(RaSS) National Laboratory Washington, DC
and
Konstantinos D. Koutroumbas
Department of Information Engineering IAASARS
University of Pisa National Observatory of Athens
Pisa, Italy Penteli, Greece
Donatella Guzzi
Institute of Applied Physics Gitta Kutyniok
‘Nello Carrara’ – National Research Institut für Mathematik
Council of Italy (IFAC-CNR) Technische Universität Berlin
C.N.R.-IFAC Berlin, Germany
Sesto Fiorentino, Italy
Marie Lasserre
Ali Hashemi Department of Electronics Optronics and
Berlin International Graduate School in Signal
Model and Simulation based Research University of Toulouse–ISAE-Supaéro
(BIMoS) Toulouse, France
Institut für Mathematik
Technische Universitat Cinzia Lastri
Berlin, Germany Institute of Applied Physics
Ahmad Hoorfar ‘Nello Carrara’ – National Research
Antenna Research Laboratory Council of Italy (IFAC-CNR)
Department of Electrical and Computer C.N.R.-IFAC
Engineering Sesto Fiorentino, Italy
Villanova University
Villanova, Pennsylvania François Le Chevalier
Delft University of Technology
Kaiyu Huang Delft, The Netherlands
Department of Automation Science and
Technology Wei Li
School of Electronics and Information College of Information Science and
Engineering Technology
Xi’an Jiaotong University Beijing University of Chemical Technology
Xi’an, China Beijing, China
Contributors xvii

Feng Lian Abubakr Muhammad

Department of Automation Science and Department of Electrical Engineering
Technology Center for Water Informatics & Technology
School of Electronics and Information (WIT)
Engineering SBA School of Science and Engineering,
Xi’an Jiaotong University LUMS
Xi’an, China Lahore, Pakistan

Jing Liu José Nascimento

Department of Automation Science and Instituto de Telecomunicações
Technology ISEL–Instituto Superior de Engenharia de
School of Electronics and Information Lisboa
Engineering Instituto Politécnico de Lisboa
Xi’an Jiaotong University Lisboa, Portugal
Xi’an, China
Roberta Palmeri
Enrico Magli Department of Information Engineering
Dipartimento di Elettronica e Infrastructures and Substainable Energy
Telecomunicazioni University “Mediterranea” of Reggio
Politecnico di Torino Calabria
Torino, Italy Reggio di Calabria, Italy

Marco Martorella Valentina Raimondi

CNIT Radar and Surveillance System
(RaSS) National Laboratory
and
Department of Information Engineering
University of Pisa
Pisa, Italy
Mahendra Mallick
Independent Consultant
Smith River, California
Gabriel Martı́n
Instituto de Telecomunicações
Lisboa, Portugal
Institute of Applied Physics
‘Nello Carrara’ – National Research
Council of Italy (IFAC-CNR)
C.N.R.-IFAC
Sesto Fiorentino, Italy
Athanasios A. Rontogiannis
IAASARS
National Observatory of Athens
Penteli, Greece
Mauricio D. Sacchi
Department of Physics
University of Alberta
Edmonton, AB, Canada

Mohamed L. Mekhalfi Nithin Sugavanam

Department of Information Engineering Department of Electrical and Computer
and Computer Science Engineering
University of Trento The Ohio State University
Trento, Italy Columbus, Ohio

Farid Melgani Xu Sun

Department of Information Engineering Institute of Remote Sensing and Digital
and Computer Science Earth
University of Trento Chinese Academy of Sciences
Trento, Italy Beijing, China
xviii Contributors

Konstantinos E. Themelis Yanfei Wang

IAASARS Key Laboratory of Petroleum Resources
National Observatory of Athens Research
Penteli, Greece Institute of Geology and Geophysics
Chinese Academy of Sciences
Sonia Tomei Beijing, China
CNIT Radar and Surveillance System
(RaSS) National Laboratory Bing Zhang
and Institute of Remote Sensing and Digital
Earth
Department of Information Engineering Chinese Academy of Sciences
University of Pisa Beijing, China
Pisa, Italy

Ismael Vera Rodriguez

Schlumberger Gould Research
Cambridge, UK
1
Compressed Sensing: From Theory to Praxis

Axel Flinth, Ali Hashemi, and Gitta Kutyniok

CONTENTS
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Recovery algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Convex programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Thresholding algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 Greedy algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Necessary and sufficient conditions of the measurement matrix
for sparse recovery . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 The RIP and the NSP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.2 Mutual coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.3 Random matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.3.1 Bounded orthonormal systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Some exemplary applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4.1 Compressed medical imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Data separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4.2.1 Orthonormal bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.2.2 Frames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.3 Analog-to-information converters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.4 Communication and networking domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.4.4.1 Error correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.4.2 Sparse channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.4.3 Spectrum sensing in CR networks . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1.1 Introduction
For a long time, the Shannon sampling theorem was the ruling paradigm in the signal
processing community when it came to choosing sampling rates. The theorem, proved
in 1949 by Claude E. Shannon, states that any function f : R → C having limited
bandwidth W (meaning that the support of the Fourier transform fˆ
is contained in the
n
interval [−W, W ]) can be exactly reconstructed from its values f ( 2W ) n∈Z . Put differently,
sampling a function at a rate at least two times higher than the bandwidth of the function
will provide enough information for perfect reconstruction of the signal of interest. The rate
two times higher than the bandwidth of a signal is often referred to as the Nyquist rate.

1
2 Compressive Sensing of Earth Observations

During the course of the last decade, a new sampling theory has emerged, known
as compressive sampling or compressed sensing, which was introduced about 10 years
ago as an effective and efficient way of sensing and acquiring data. This theory enables
faithful recovery of signals, images, and other types of data from highly sub-Nyquist-rate
samples. While Shannon sampling theory only utilizes the bandwidth information of the
signal, compressed sensing relies on the crucial observation that data we are interested in
acquiring typically are structured. To be concrete, they often possess a sparse or nearly
sparse representation in a certain basis or dictionary [5,11,25].
To state the problem mathematically, let x0 = ((x0 )i )ni=1 ∈ Rn be our signal of interest.
As prior information, we first assume that x0 is sparse by itself, that is, that its support
supp x0 = {i : (x0 )i 6= 0} contains few elements. Another way of stating this is to say that
the `0 -norm of the signal,

kxk0 := # supp x0 ,

is small. Furthermore, let A be an m × n matrix, which is typically called sensing matrix

or measurement matrix. Throughout this chapter, we will always assume that m < n. Then
the compressed sensing problem can be formulated as follows: Recover x0 from knowledge of

b = Ax0 .

The above equation is an underdetermined linear system of equations. The prior

information that the signal of interest is sparse makes it possible to still uniquely recover it
using compressed sensing techniques.
In this introductory chapter, we will first give an overview of different solution
strategies to the above problem. Then we will present the basic results about which
properties the measurement matrix has to have in order for those strategies to succeed.
In this part, we will also briefly discuss the case of the matrix being chosen at random.
In the final part, we will widen our horizon by discussing applications of the theory of
compressed sensing. In that part, it will be necessary to briefly discuss some generalizations
of compressed sensing.

1.2 Recovery algorithms

Given the prior knowledge that a linear system of equations Ax = b has a sparse solution
x0 , and that x0 is the sparsest among all solutions, it is evident that the `0 -minimization
problem,

min kxk0 subject to Ax = b, (1.1)

will have x0 as its solution. There is however no general strategy to solve Equation 1.1
efficiently. In fact, one can prove that solving Equation 1.1 for general A and b is N P -
hard [60]. Going into detail what this means goes beyond the scope of this introductory
chapter. To give the reader some intuition, let us just note that the N P -hardness of
Equation 1.1 implies that if we find an algorithm which can solve Equation 1.1 for any
matrix A and any vector b in polynomial time, we would have solved the famous millenium
problem P versus N P , which certainly is very much to ask for. This motivates the need for
formulating tractable relaxations of Equation 1.1.
Compressed Sensing 3

1.2.1 Convex programming

Since there often exist efficient solvers for convex optimization, it seems reasonable to search
for a convex surrogate of Equation 1.1. In some sense, the most natural way (see [20]) of
designing a convex program for identifying sparse vectors is to minimize the `1 -norm instead
of the `0 -norm, that is, to consider

min kxk1 subject to Ax = b. (1.2)

The algorithm, or rather strategy, of solving Equation 1.2 for recovering sparse solutions
of Ax = b is known as Basis Pursuit. Researchers have known for several decades that the
`1 -norm promotes sparsity. But it was as late as 2001 that the Basis Pursuit algorithm was
formulated [21]. The Basis Pursuit algorithm can be formulated as a linear program, which
in particular proves that it can be solved in polynomial time. It is however by no means
clear that the solution of the problem defined by Equation 1.2 is equal to the sparse signal
of interest x0 . Much of the theory in fact revolves around determining for which type of
matrices this is the case. We will discuss this in detail in Section 1.3.
The problem defined by Equation 1.2 assumes that the linear measurements b = Ax0
are exactly known. In applications, they are however often affected by noise; that is, we are
only given b = Ax0 + n for some unknown noise vector n. There exist several ways to handle
this situation. Assuming that some appropriate bound for the noise level knk2 is available,
say knk2 ≤ e, we can use regularized Basis Pursuit, given by

min kxk1 subject to kAx − bk2 ≤ e. (1.3)

A procedure of similar characteristic is the so-called Dantzig* selector [16]

min kxk1 subject to kA∗ (b − Ax)k∞ ≤ τ.

τ is thereby a tuning parameter. If one assumes that the√noise is randomly chosen according
to the normal distribution with variance σ2 , then τ = σ 2 log d is a good choice.
Another way to handle noise is to apply the so-called LASSO (least absolute shrinkage
and selection operator) [71]. The LASSO has its origin in statistics, and it again consists
of solving a minimization problem, this time given by

min kAx − bk2 subject to kxk1 ≤ τ,

where τ is a tuning parameter (which has to be chosen carefully).

1.2.2 Thresholding algorithms

For many practical applications such as image processing, linear programming methods face
implementation challenges, since the complexity of these algorithms grows superlinearly
with the input size. As an example, recall that the required average number of steps in the
simplex algorithm for finding a solution grows cubically with the size of the input [12]. In
order to decrease the computational complexity of `1 -norm minimization, iterative methods
have been proposed, which converge to a very good approximation of the exact solution of
the original linear program.
Let us consider the following unconstrained regularized least-squares (LS) problem:
1
min kb − Axk22 + λkxk1 .
x 2

* The name of the algorithm was chosen to honor George Dantzig, the founding father of linear

programming. In particular, it does not directly originate from the German name of the Polish town Gdańsk.
4 Compressive Sensing of Earth Observations

It can be shown that there exists a λ > 0 such that the solution of the above optimization
problem is equal to the one of Equation 1.2. This value of λ although cannot be calculated
a priori , when it is often empirically determined. Interested readers might want to have a
look at [75], where a more systematic approach is described.
Possibilities for efficient solvers are iterative methods such as steepest descend or
conjugate gradient approaches. These will converge to the minimizer of Equation 1.2 by
selecting an appropriate starting point as an initial value. Several algorithms have been
developed for this purpose [6,77]. One of the most prominent is called iterative hard
thresholding (IHT) [9,24,32] and will be explained in the sequel.
In the IHT algorithm, a sparse solution can be found by refining the estimate of sparse
signals by the following iterative procedure:

xi+1 = Hλ (xi − A∗ (Axi − b)), (1.4)

where A∗ is the (conjugate) transpose of the matrix A and H(·) is a the hard thresholding
operator given by √
0 : |xi | ≤ √λ,

Hλ (x)i =
xi : |xi | ≥ λ.
As can be seen from Equation 1.4, the input of the thresholding function only consists of
the multiplication of A and A∗ with xi and b. These operations have a significantly smaller
computational complexity than linear programming. This feature can be considered as the
main advantage of these types of algorithms as compared with the linear programming
approaches [65, p. 18].
The threshold in iterative algorithms can be chosen either constant in all iterations such
as in Equation 1.4, or adaptively, resulting in iterative shrinkage thresholding algorithms
(ISTA). Adaptively choosing the threshold can improve the performance compared with
IHT [24,32]. The iterative step then looks as follows:

xi+1 = σλδ (xi − δA∗ (Axi − b)),

where δ is a parameter for step size and σλ (·) is the soft shrinkage threshold function given by

σλ (xi ) = max(|xi | − λ, 0) sgn(xi ).

As an extended version of the ISTA method, the FISTA algorithm [6] has been proposed by
utilizing the solutions of two previous iterations in order to improve the convergence rate
of ISTA [6].

1.2.3 Greedy algorithms

The key idea of greedy algorithms is to first identify some subset in support of the original
signal and then refine it in a number of steps until a good estimate of the support is found.
In compressed sensing problems, the number of steps is typically equal to the sparsity level
s = kx0 k0 . These algorithms have the advantage of being very fast and easy to implement.
One of the simplest algorithms of this class is called matching pursuit (MP) [56]. To
explain this method, we first provide an equivalent representation of the compressed sensing
problem as follows:
n
X
b= ai xi , (1.5)
i=1
Compressed Sensing 5

where ai is the ith column of A. By assuming s-sparsity of the signal, the compressed sensing
problem in this formulation can be considered as finding the s “active” columns of A (i.e.,
the support of the vector x = (xi )ni=1 ) and the corresponding coefficients xi [65, p. 19].
MP approximates the source in s steps. In each step, the algorithm finds the column of
the sensing matrix A which is most correlated with the current residual r (initialized as b).
Once the column of A is revealed, the corresponding coefficient xi is revealed by minimizing
the error corresponding to the data fidelity, kb − xi ai k22 , through solving a LS problem [10].
In particular, the column ai∗ with the maximum absolute value of hr, ai i (most correlated
with residual) is selected as an active column in Equation 1.5 and xi = hr, ai∗ i, after
obtaining the inner products hb, ai i of b. Subsequently, the residual is updated according to

r ← r − xi ai∗ .

The main disadvantage of this approach is that it assumes that the columns of A are
orthogonal, which is not the case for most sensing matrices [65]. If that is the case however,
it is very fast and cheap, in particular when the matrix multiplication with A and A∗ can
be efficiently implemented (that being the main complexity bottleneck).
Orthogonal matching pursuit (OMP) [72] improves MP, making it work for more general
measurement matrices, via updating the found xi ’s in each step. This approach is explained
in detail as follows: At the first stage, the OMP algorithm takes a matrix A ∈ Rm,n , a
measurement vector b ∈ Rm , and an error threshold e > 0 as an input data. In the next
step, for starting the algorithm, the initial values for the solution x0 = 0, the residual value
r0 = b − Ax0 = b, and for the support S 0 = supp x0 = ∅ are assigned. If k denotes the
index of the iteration, the algorithm updates the support of recovered coefficients at each
iteration by choosing a new index i∗ such that [46]

| hai∗ , rk−1 i | ≤ | hai , rk−1 i | for all i.

Then, based on the updated support S k = S k−1 ∪ {i∗ }, OMP computes the residual
rk = y − Axk by updating xk in each step as follows [46]:

xk = argminx kAx − bk2 subject to supp x = S k .

Finally, the algorithm repeats through iterations until the stopping criterion krk k2 < e
is met, where k is the index of iteration.
Faster algorithms such as compressive sampling matched pursuit (CoSaMP) [61] improve
OMP by choosing more than k columns of the sensing matrix, recovering the signal
coefficients based on the projection onto the space of the selected columns, and rejecting
those which might not be in the true support.

1.3 Necessary and sufficient conditions of the measurement

matrix for sparse recovery
In this section, we will focus on the original problem of compressed sensing: Under which
assumptions on the measurement matrix A does a specific algorithm recover all s-sparse
vectors x0 from their respective measurements Ax0 ? Note that almost all of the results in
this section also hold for complex matrices and signals, but we shall stay in the real setting
for the sake of simplicity of the exposition.
For a subset S of {1, . . . , n}, let AS denote the submatrix formed by the columns ai
of A corresponding to indices i ∈ S. Now, let us begin by making a simple observation: If
6 Compressive Sensing of Earth Observations

there exists a submatrix of A consisting of 2s columns which is not injective, any algorithm
whose output only depends on A and b = Ax0 will fail to recover all s-sparse vectors. To
see this, let η be a nonzero vector in the kernel of a submatrix AS with S = {i1 , . . . i2s }.
Then the two s-sparse vectors x1 and x2 defined by
( (
η(i) : i ∈ {i1 , . . . is } , −η(i) : i ∈ {is+1 , . . . i2s } ,
x1 (i) = and x2 (i) =
0 : else, 0 : else,

will have the same measurements, b = Ax1 = Ax2 . Thus, they cannot be distinguished by
any recovery algorithm.
We have derived a first result. To state it formally, let us introduce the spark of a matrix
A [28]. The spark, denoted by spark(A), is defined as the cardinality of the smallest set
of columns in A which is linearly dependent. Using this naming, we have just proven the
following proposition:

Proposition 1.3.1 If A ∈ Rm,n is a matrix such that any s-sparse vector x0 can be
recovered from its linear measurements Ax0 , then necessarily spark(A) > 2s.

Since spark(A) ≤ m, we can in particular conclude that recovery of all s-sparse vectors
is only possible if the number of measurements, m, is at least equal to 2s. This will serve
as a lower benchmark in our further discussion.

1.3.1 The RIP and the NSP

Spark-based analyses will rarely yield any statements about the success of specific recovery
procedures. In particular, it is impossible to prove stability statements solely relying on the
spark, since it is possible to construct matrices with arbitrarily ill-conditioned submatrices
which are still invertible. Therefore, it is necessary to impose other conditions on A. The
arguably most celebrated and well-known condition is the so-called restricted isometry
property (RIP). It was introduced by Candès and Tao in the pioneer work of [14].

Definition 1.3.2 (restricted isometry property) Let A ∈ Rm,n . Then A satisfies the
(δ, k)-RIP , if all k-sparse vectors x satisfy

(1 − δ)kxk22 ≤ kAxk22 ≤ (1 + δ)kxk22 .

The smallest δ such that A satisfies the (δ, k)-RIP is called the kth restricted isometry
constant of A and is denoted by δk (A).

Note that despite the popularity of claiming that a matrix “has the RIP,” that statement
by itself has a vague meaning. Formally, one always has to specify the parameters k and δ
to make the statements precise. In reality, however, this is seldom done. The sparsity level
is most often clear from the context and δ is implicitly assumed to be sufficiently small. In
the following, we will also use this convention.
Next notice that δk < 1 implies spark(A) > k. Furthermore, by considering the definition
of submatrix, it is not hard to see that using standard theorems about the spectral properties
of Hermitian matrices, the inequality δk (A) ≥ δ is equivalent to

kA∗S AS − idS k ≤ δ (1.6)

for all S with |S| ≤ k. kAk thereby denotes the operator norm of the matrix A. This
equivalence will become very useful when proving the RIP, in particular, for random
matrices, since there exist a lot of powerful results about their spectral properties.
Compressed Sensing 7

There exist many proofs of the fact that the RIP-property implies that regularized Basis
Pursuit, that is, the problem defined by Equation 1.3, will recover any s-sparse signal in a
stable manner. Candès and Tao were in their original article able to construct a so-called
dual certificate * c ∈ Rn , only using the RIP of the matrix A. We will consider an alternative
route, following [33, Chs. 4, 6]. This route takes a detour to another key property, namely
the so-called null space property (NSP). The term “NSP” originates from the article [23],
but it was implicitly known before that.

Definition 1.3.3 Let A ∈ Rm,n . Then the matrix A has the s-N SP , if for every set S ⊆
{1, . . . , n} with |S| ≤ n and 0 6= η ∈ ker A,
X X
|ηi | < |ηi |.
i∈S i∈S
/

The s-NSP is in fact equivalent to the uniform success of Equation 1.2 at recovering
sparse vectors as the following result shows.

Lemma 1.3.4 Let A ∈ Rm,n . Then the following conditions are equivalent:

1. The program Equation 1.2 with b = Ax0 has x0 as its unique solution for every
s-sparse x0 .

2. A has the s-N SP .

Proof. (ii) ⇒ (i). Suppose that A has the s-NSP, and let x0 be an arbitrary s-sparse vector.
Suppose further that x0 is not the unique solution of Equation 1.2, but instead that there
exists x ∈ Rn with Ax = Ax0 and kxk1 ≤ kx0 k1 . Then, η = x − x0 ∈ ker A.
We will now prove that
X X
|ηi | ≥ |ηi |,
i∈supp x0 i∈supp
/ x0

which would be a contradiction to the s-NSP. Due to the triangle inequality,

kx0 + θηk1 = kθ(x0 + η) + (1 − θ)x0 k1 ≤ θkxk1 + (1 − θ)kx0 k1 ≤ kx0 k1

for every θ ∈ [0, 1]. Hence, for every such θ, we have

X X X
|x0 (i)| ≥ |x0 (i) + θη(i)| + θ|η(i)|. (1.7)
i∈supp x0 i∈supp x0 i∈supp
/ x0

If we now choose θ sufficiently close to zero, we can ensure that |x0 (i) + θη(i)| = |x0 (i)| +
θ sgn(x0 (i))η(i) for all i ∈ S. For such θ, inequality Equation 1.7 reads
X X X
|x0 (i)| ≥ |x0 (i)| + θ sgn(x0 (i))η(i) + θ|η(i)|
i∈supp x0 i∈supp x0 i∈supp
/ x0
X X
⇔ θ|η(i)| − θ sgn(x0 (i))η(i) ≤ 0.
i∈supp
/ x0 i∈supp x0

Dividing by θ > 0 and estimating sgn(x0 (i))η(i) ≤ |ηi | yields the claim.
* A dual certificate is, put simply, a vector in the range of A∗ with certain properties whose existence

implies that a vector x0 is the solution of a convex optimization problem. The properties of course depend
on which problem is considered.
8 Compressive Sensing of Earth Observations

(i) ⇒ (ii). Now, suppose that A does not have the s-NSP. We will construct a vector x0
which is not recovered by Equation 1.2. Due to the definition of the NSP, there exists an
η 6= 0 with Aη = 0 and an S with |S| ≤ s such that
X X
|ηi | ≥ |ηi |. (1.8)
i∈S i∈S
/

If we define
( (
η(i) : i ∈ S, −η(i) : i ∈
/ S,
x0 (i) = and x(i) =
0 : else 0 : else,

then x0 is s-sparse, Ax0 = Ax, and kxk1 ≤ kx0 k, due to Equation 1.8. Hence, x0 is not
uniquely recovered by Equation 1.2, and the proof is finished.
Let us next analyze the relation between the NSP and the RIP.

Theorem 1.3.5 ([33, Thm. 6.9]) If the 2s-th restricted isometry constant of A obeys
δ2s ≤ 13 , then A satisfies the s-N SP . In particular, Basis Pursuit will uniquely recover any
s-sparse vector from its measurements Ax0 .

There are two important facts about the NSP and the RIP to notice.

Remark 1.3.6

1. The NSP has a nice geometrical interpretation, as can be seen in Figure 1.1. For
a function f : Rn → R and a point x0 ∈ Rn , define the descent cone D(f, x0 ) as
the set of descent directions at x0 :

D(f, x0 ) = {v ∈ Rn : ∃ t > 0 such that f (x0 + tv) ≤ f (x0 )} .

A careful analysis of the proof of Lemma 1.3.4 shows that it in fact provides an
argument for the fact that
 
 X X 
D(k · k1 , x0 ) = v ∈ Rn : |v(i)| ≤ − sgn(x0 (i))v(i) .
 
i∈supp
/ x0 i∈supp x0

This formula makes it clear that the NSP is equivalent to the statement that
D(k·k1 , x0 )∩ker A = {0} for every s-sparse x0 . This geometrical characterization of
success of programs such as the one given by Equation 1.2 is directly used for many
results. Together with methods from high-dimensional geometry, it is possible
to provide very elegant arguments about statements on how many Gaussian
measurements (see Section 1.3.3) are needed to guarantee unique recovery with
high probability. The seminal work in this direction is [2].
2. The RIP is not equivalent to the success of Equation 1.2, see the discussion in [33,
p. 169].

As has been mentioned before, it was proven that regularized Basis Pursuit is stable. To
be precise, we speak of stability if the solution x∗ of the program Equation 1.3 with input
b = Ax0 is guaranteed to satisfy
σs (x0 )1
kx∗ − x0 k2 ≤ C1 √ + C2 e (1.9)
s
Compressed Sensing 9

x0 + D( f, x0)

x0 + ker A

x0

{ f (x) ≤ f (x0)}

FIGURE 1.1
An illustration of the situation ker A ∩ D(f, x0 ) = {0}. Note that there does not exist
any vector x ∈ x0 + ker A which satisfies f (x) < f (x0 ). Hence, x0 solves the program
min f (x) subject to Ax = Ax0 .

for every x0 ∈ Rn , where σs (x0 )1 denotes the s-term `1 -approximation error, that is,

σs (x0 )1 = min kx0 − x0 k1 .

x0 s-sparse

√
Indeed, a statement like Equation 1.9 holds under the assumption
√ δ2s ≤ 4/ 41. The strategy
resembles the one above: First, it is proven that δ2s ≤ 4/ 41 implies A to satisfy the so-
called robust null space property (Robust NSP), namely that

kxS k1 ≤ ρkxS c k1 + τkAxk2

for some 0 < ρ < 1 and τ > 0, with ρ and τ only dependent on δ2s . It is then argued that
the Robust NSP implies stability of Equation 1.3 . It should again be emphasized that one
of the first articles on Basis Pursuit [18] derives a similar assertion under the assumption
δ3s + 3δ4s < 2 without taking the detour to the Robust NSP.

1.3.2 Mutual coherence

Another quality measure for matrices of high relevance in the setting of compressed sensing is
the so-called mutual coherence. It was first introduced in [29] the special case of A = [Φ1 , Φ2 ],
where each matrix Φi is an orthogonal matrix, but has then been adopted accordingly for
general matrices. For the sake of simplicity of the exposition, we assume that all matrices
in this section have `2 -normalized columns. Then, the definition of mutual coherence reads
as follows:

Definition 1.3.7 Let ai denote the i-th column of the matrix A. Then the mutual coherence
of A is defined by

µ(A) = max | hai , aj i |.

i6=j

Small mutual coherence of a matrix is a strong property. For instance, a low mutual
coherence implies the RIP. The following proposition is a simplification of a theorem
from [72].
10 Compressive Sensing of Earth Observations

Proposition 1.3.8 For every matrix A and positive integer s, we have

δs (A) < (s − 1)µ(A).

In particular, spark(A) > (1 + µ(A)−1 ).

Sketch of proof. Considering the equivalence of spectral properties of submatrices of A

with the RIP (expressed in Equation 1.6), we have to analyze the eigenvalues of the matrices
M S = A∗S AS for S ⊆ {1, . . . , n} arbitrary with |S| ≤ s. The entries of M S are given by
S
Mij = hai , aj i. Hence, the diagonal entries MiiS are all equal to 1, due to the assumed
unit length of the columns. Besides, all other entries uniformly are bounded in magnitude
by µ(A). Now, one can apply the Gershgorin Disk Theorem [35], which states that all
eigenvalues λ of a matrix M lie in one of the circles Ci ⊆ C with center Mii = 1 and radius
Ri = j6=i |Mij |. This implies that the eigenvalues λ of M S all obey
P

X
S
|λ − 1| ≤ |Mij | ≤ (s − 1)µ,
j∈S\{i}

which is the statement we needed to prove.

The concept of mutual coherence has a major drawback: There is a fundamental lower
bound, which is referred to as Welch Bound [78]. In fact, the mutual coherence of a matrix
A ∈ Rm,n obeys
r
n−m
µ(A) ≥ . (1.10)
m(n − 1)

If, via Proposition 1.3.8, we intend to guarantee that δs (A) ≤ c, we require µ(A) ≤
c(s − 1)−1 . The Welch bound Equation 1.10 then enforces

(s − 1)2 (s − 1)2
r  
n−m c
≤ ⇐⇒ m 1 + ≥ 2 .
m(n − 1) s−1 n c (1 − n1 )

Thus, if we assume a regime where n & s2 (i.e., n ≥ Cs2 for some constant C), we
require an order of s2 measurements to—via the mutual coherence—ensure that δs (A) is
small, which is needed for ensuring success of Basis Pursuit for s-sparse signals. Ideally, we
would like to remove the square exponent, so that the number of required measurements
essentially scales as the degree of sparsity s. It is fortunately possible to directly prove that
certain types of random matrices in Rm,n satisfy the RIP with m in that range. We will
return to this issue in Section 1.3.3.
With the last paragraph in mind, one could argue that one should avoid using the notion
of mutual coherence for analyzing the performance of algorithms, and instead concentrate on
the RIP or NSP directly. There is however one situation when the RIP fails, and that is when
analyzing the performance of OMP. Indeed, there exist many matrices having the s-RIP for
s ∼ n for which there exist s-sparse vectors x0 such that the first step of OMP fails to pick a
correct index (see [27, Sec. 7]). In contrast, we may however prove statements guaranteeing
good performance of OMP involving the mutual coherence, such as the following one.

Theorem 1.3.9 (Simplified version of Theorem A and B from [72]) If µ(A)(2s − 1)

< 1, then OMP succeeds at recovering every s-sparse signal x0 from the input b = Ax0 in s
steps.
Another random document with
no related content on Scribd:
 CHAPTER XII
THE WORSHIP OF MITHRAS

Few of us, perhaps, are inclined to recognize that, from its first
establishment down to the Mahommedan Invasion of the VIIth
century, the Roman Empire found itself constantly in the presence of
a bitter, determined, and often victorious enemy. Alexander had
conquered but had not destroyed the Persians; and, although the
magic of the hero’s personality held them faithful to him during his
too brief life, he was no sooner dead than they hastened to prove
that they had no intention of tamely giving up their nationality.
Peucestas, the Royal bodyguard who received the satrapy of Persia
itself on his master’s death, and was confirmed in it at the first
shuffling of the cards at Triparadisus, found it expedient to adopt the
Persian language and dress, with the result that his subjects
conceived for him an affection only equal to that which they
afterwards showed for Seleucus[768]. Later, when the rise of the
Parthian power under Arsaces brought about the defeat of Seleucus
II Callinicus, the opposition to European forms of government found
a centre further north[769], whence armies of lightly-equipped
horsemen were able to raid up to the Eastern shores of the
Mediterranean[770]. Thanks probably to the knowledge of this support
in reserve, when Western Asia found the military power of the Greek
kings becoming exhausted by internecine wars, she began to throw
off the alien civilization that she had in part acquired, and to return
more and more to Persian ways[771]. When the Romans in their turn
set to work to eat up the enfeebled Greek kingdoms, they quickly
found themselves in presence of a revived nationality as firmly held
and nearly as aggressive as their own, and henceforth Roman and
Parthian were seldom at peace. The long struggle with Mithridates,
who gave himself out as a descendant of Darius[772], taught the
Romans how strong was the reaction towards Persian nationality
even in Asia Minor, and the overthrow of Crassus by the Parthians
convinced his countrymen for a time of the folly of pushing their arms
too far eastwards.
With the establishment of the Empire, the antagonism between
Rome and Persia became still more strongly marked, and a struggle
commenced which lasted with little intermission until the foundation
of the Mahommedan Caliphate. In this struggle the advantage was
not always, as we should like to think, on the side of the Europeans.
While Augustus reigns, Horace boasts, there is no occasion to dread
the “dreadful Parthians[773]”; but Corbulo is perpetually fighting them,
and when Nero commits suicide, the legend immediately springs up
that the tyrant is not dead, but has only betaken himself beyond the
Euphrates to return with an army of Rome’s most dreaded enemies
to lay waste his rebellious country[774]. Towards the close of the first
Christian century, Trajan, fired, according to Gibbon, by the example
of Alexander, led an army into the East and achieved successes
which enabled him to add to his titles that of Parthicus[775]; but the
whole of his Oriental conquests were given back by the prudent
Hadrian on his succession to the throne. During the reign of Marcus
Aurelius, Avidius Cassius obtained some solid victories on the
frontier; but Macrinus is said to have bought off the Parthians with a
bribe of nearly two millions of money. The rise of the Sassanian
house and the retransfer of the leadership from the Parthians to their
kinsmen in Persia proper brought about the reform of the Persian
religion, and added another impulse to the increasing strength of
Persian national feeling. Alexander Severus may have gained some
successes in the field over Ardeshîr or Artaxerxes, the restorer of the
Persian monarchy[776]; but in the reign of the last named king’s son
and successor Sapor, the capture of the Emperor Valerian with his
whole army, and the subsequent ravaging of the Roman provinces in
Asia by the victors, showed the Republic how terrible was the might
of the restored kingdom[777]. Aurelian, the conqueror of Palmyra, did
much to restore the prestige of Roman arms in the East; and
although he was assassinated when on the march against Persia,
the Emperor Carus shortly after led a successful expedition into the
heart of the Persian kingdom[778]. In the reign of Diocletian, indeed,
the Persians lost five provinces to the Romans[779]; but under
Constantine the Great the Romans were again vanquished in the
field, and the Persians were only prevented by the heroic resistance
of the fortified town of Nisibis and an incursion into their Eastern
provinces of tribes from Central Asia from again overrunning the
Asiatic possessions of Rome[780].
Henceforward, the history of the long contest between the two great
empires—“the eyes,” as the Persian ambassador told Galerius, “of
the civilized world[781],” is the record of almost uninterrupted advance
on the part of Persia and of continual retreat on the side of Rome.
The patriotic enthusiasm of a Julian, and the military genius of a
Belisarius, aided by the dynastic revolutions common among
Oriental nations, might for a time arrest the progress of the
conquering Persians; but, bit by bit, the Asiatic provinces slipped out
of the grasp of the European masters of Constantinople. In 603 A.D.,
it looked as if Persia were at length in the position to deliver the final
blow in a war which had lasted for more than five centuries. By the
invasion of Chosroes and his successive captures of Antioch,
Jerusalem, and Egypt, it seemed as if the Persians had restored the
world-empire of Cambyses and Darius; but the Persians then
discovered, as Xerxes had done a millennium earlier, how
dangerous it is for Orientals, even when flushed by conquest, to
press Europeans too far. The Roman Emperor Heraclius, who never
before or afterwards gave much proof of military or political capacity,
from his besieged capital of Constantinople collected an army with
which he dashed into Persia in a manner worthy of Alexander
himself. After six brilliant campaigns he dictated to the Persians a
triumphant peace in the very heart of their empire[782]. A few years
later, and its shattered and disorganized remains fell an easy prey to
the Mahommedan invaders.
The effect of this long rivalry might have been expected to produce
in the Romans during its continuance a hearty dislike of the customs
and institutions of the nation opposed to them; but almost the exact
contrary was the result. It may be argued that Rome’s proved skill in
government was in no small measure due to her ready adoption of
all that seemed to her admirable in the nations that she overcame.
Or it may be that the influence which the memory of Alexander
exercised over all those who succeeded to his empire led them to
imitate him in his assumption of Persian manners. The fact remains
that, long before the division of the Roman Empire into East and
West, the Romans displayed a taste for Oriental luxury and
magnificence which seems entirely at variance with the simplicity
and austerity of the republican conquerors of Carthage. It is hardly
too much to say that while Alexander’s conscious aim was to make
Asia Greek, the Romans, on possessing themselves of his Asiatic
conquests, allowed themselves to become to a great extent
“Medized,” and showed an unexpected admiration for the habits and
culture of Alexander’s Persian subjects.
It may of course be said that this was in external matters only, and
that the “Persian furniture” which excited Horace’s wrath[783] might if it
stood alone be looked upon as merely a passing fashion; but the
Court ceremonial introduced by Diocletian argues a steady tendency
towards Persian customs and forms of government that must have
been in operation for centuries. The household of a Julian Caesar
was no differently arranged from that of a Roman noble of the
period, and his title of Prince of the Senate showed that he was only
looked upon as the first of his equals. But Diocletian was in all
respects but language a Persian emperor or Shah, and his style of
“Lord and God,” his diadem, his silken state dress, the elaborate
ritual of his court, and the long hierarchy of its officials, were all
designed to compel his subjects to recognize the fact[784]. As usual,
the official form of religion in the Roman Empire had for some time
given indications of the coming change in the form of government.
The sun had always been the principal natural object worshipped by
the Persians, and a high-priest of the Sun-God had sat upon the
Imperial throne of Rome in the form of the miserable Heliogabalus.
Only 13 years before Diocletian, Aurelian, son of another Sun-God’s
priestess and as virile and rugged as his predecessor was soft and
effeminate, had also made the Sun-God the object of his special
devotion and of an official worship. Hence Diocletian and his
colleague Galerius were assured in advance of the approval of a
large part of their subjects when they took the final plunge in 307
A.D.,and proclaimed Mithras, “the Unconquered Sun-God,” the
Protector of their Empire[785].
In spite of this, however, it is very difficult to say how Mithras
originally became known to the Romans. Plutarch says indeed that
his cult was first introduced by the Cilician pirates who were put
down by Pompey[786]. This is not likely to be literally true; for the
summary methods adopted by these sea-robbers towards their
Roman prisoners hardly gave much time for proselytism, while most
of the pirates whom Pompey spared at the close of his successful
operations he deported to Achaea, which was one of the few places
within the Empire where the Mithraic faith did not afterwards show
itself. What Plutarch’s story probably means is that the worship of
Mithras first came to Rome from Asia Minor, and there are many
facts which go to confirm this. M. Cumont, the historian of Mithraism,
has shown that, long before the Romans set foot in Asia, there were
many colonies of emigrants from Persia who with their magi or
priests had settled in Asia Minor, including in that phrase Galatia,
Phrygia, Lydia, and probably Cilicia[787]. When Rome began to
absorb these provinces, slaves, prisoners, and merchants from them
would naturally find their way to Rome, and in time would no doubt
draw together for the worship of their national deities in the way that
we have seen pursued by the worshippers of the Alexandrian Isis
and the Jewish exiles. The magi of Asia Minor were great supporters
of Mithridates, and the Mithridatic wars were no doubt responsible
for a large number of these immigrants.
Once introduced, however, the worship of Mithras spread like wild-
fire. The legions from the first took kindly to it, and this is the less
surprising when we find that many of them were recruited under the
earliest emperors in Anatolian states like Commagene, where the
cult was, if not indigenous, yet of very early growth[788]. Moreover the
wars of the Romans against the Persians kept them constantly in the
border provinces of the two empires, where the native populations
not infrequently changed masters. The enemy’s town that the legions
besieged one year might therefore give them a friendly reception the
next; and there was thus abundant opportunity for the acquaintance
of both sides with each others’ customs. When the Roman troops
marched back to Europe, as was constantly the case during the civil
wars which broke out on the downfall of the Julian house, they took
back with them the worship of the new god whom they had adopted,
and he thus became known through almost the whole of the Roman
Empire[789]. “From the shores of the Euxine to the north of Brittany
and to the fringe of the Sahara[790],” as M. Cumont says, its
monuments abound, and, he might have added, they have been met
with also in the Egyptian Delta, in Babylon, and on the northern
frontiers of India. In our own barbarous country we have found them
not only in London and York, but as far west as Gloucester and
Chester and as far north as Carlisle and Newcastle[791]. The Balkan
countries, like Italy, Germany, Southern France, and Spain, are full of
them; but there was one part of the Roman Empire into which they
did not penetrate freely. This was Greece, where the memories of
the Persian Wars long survived the independence of the country, and
where the descendants of those who fought at Salamis, Marathon,
and Thermopylae would have nothing to do with a god coming from
the invaders’ fatherland. It is only very lately that the remains of
Mithras-worship have been discovered at the Piraeus and at Patras,
in circumstances which show pretty clearly that it was there practised
only by foreigners[792].
Notwithstanding this popularity, it is not easy to say exactly what god
Mithras’ European worshippers considered him to be. If length of
ancestry went for anything in such matters, he might indeed claim a
greater antiquity than any deity of the later Roman Pantheon, with
the single exception of the Alexandrian gods. Mithras was certainly
worshipped in Vedic India, where his name of Mitra constantly
occurs in the sacred texts as the “shining one,” meaning apparently
the material sun[793]. He is there invoked in company with Varuna,
generally considered the god of the sky, and therefore according to
some, the prototype of the Greek Zeus and the Latin Jupiter[794]. His
appearance in a similar connection in the sacred books of the
Persians led the founders of the comparative study of religion to
think that he must have been one of the primitive gods of their
hypothetical Aryan race, and that his worship must go back to the
imaginary time when Persians and Hindus dwelt side by side in the
plains of Cashmere. But this theory is giving way before proof that
the original home of the Indo-European race was Europe, and has
been badly shaken by the discovery at Boghaz Keui of tablets
showing that the gods Mithra and Varuna were gods of the
Mitannians or Hittites[795] at some date earlier than 1500 B.C., and
therefore long before the appearance of the Persians in history. If the
worship of Mithras were not indigenous in Western Asia, it may
therefore well have come there independently of the Persians[796].
There is no doubt, however, that the roots of Mithras-worship went
very far down into the Persian religion. In the Yashts or hymns which
are the earliest evidence of primitive Iranian beliefs, Mithra—to use
the Avestic spelling of his name—frequently appears, not indeed as
the material sun, but as the “genius of the heavenly light” which
lightens the whole universe[797] and is the most beneficent among the
powers of Nature. Mithras is not here, however, the Supreme Being,
nor even the highest among the gods benevolent to man. This last
place is occupied in the Zend Avesta by Ahura Mazda, “the
omniscient lord,” who appears to be the Persian form of Varuna, the
god of the sky whom we have seen associated with Mitra in the
Vedas[798]. Nor is Mithras in the Zend Avesta one of the six
Amshaspands, the deified abstractions or personified attributes of
Ahura Mazda, who, in the later developments of the Persian religion,
occupy towards him much the same position that the “Roots” of
Simon Magus and the Aeons of the Pleroma among the Gnostics do
towards the Boundless Power or the Ineffable Bythos[799]. In the later
Avestic literature, he appears as the chief of the Izeds or Yazatas, a
race of genii created by Ahura Mazda, who are the protectors of his
universe and the helpers of mankind in their warfare against the
powers of darkness[800]. In the latest as in the earliest Persian view of
the personality of Mithras, therefore, it is plain that he occupies an
intermediate position between the Creator and man.
It is not, however, in the religion associated with the name of
Zoroaster that we must look for the origin of Mithraism. The date of
the sacred books of Mazdeism and the historical existence of
Zoroaster himself have recently been brought down to as late as the
VIIth century B.C.[801] and the appearance in Asia of the Persian tribes
as conquerors, whereas Mithras was, as we have seen, worshipped
in Asia Minor nearly a millennium earlier. Moreover, the strict dualism
which set Ahriman, the god of darkness and evil, in eternal and
perhaps equal opposition to Ormuzd, the god of light and goodness,
seems to have been unknown before the Sassanid reform in 226
A.D., by which time the worship of Mithras in Europe was at its
apogee[802]. M. Cumont is, therefore, doubtless right when he thinks
that Mithraism was derived not from Mazdeism, but from Magism or
the religion of the Magi, the tribe of Medes whose domination was
put an end to by Darius the son of Hystaspes, and whose name was
afterwards given to a priestly caste and has passed into our own
language as the root of the word “magic.”
That these Magi practised a religion different from that taught in the
Avestic literature is plain enough. The romantic story told by
Herodotus of the Magian who seized the throne of Persia during
Cambyses’ absence in Egypt on the pretence that he was the king’s
brother whom Cambyses had privily put to death[803], is fully
confirmed by Darius’ trilingual inscription on the Rock of Behistun,
first copied and deciphered by Sir Henry Rawlinson and lately
published in elaborate form by the British Museum[804]. Darius here
narrates how “a certain man, a Magian, Gaumata by name ... lied
unto the people (saying) ‘I am Smerdis, the son of Cyrus, the brother
of Cambyses.’ Then all the people revolted from Cambyses and
went over to him, even Persia and Media and the other provinces.”
Darius goes on to record that “thereupon Cambyses died by his own
hand[805],” that the seven Persian nobles overthrew the pretender
much in the way described by Herodotus, and that “I rebuilt the
temples of the gods, which that Gaumata, the Magian, had
destroyed. I restored that which had been taken away as it was in
the days of old[806].” This he tells us he did “by the grace of Ahura
Mazda,” and that by this grace he always acted. The memory of
these events was kept up by the festival of the Magophonia or
Massacre of the Magi which was yearly celebrated in Persia and
during which no Magus dared show himself in the streets[807]. Darius’
words show that there was a religious as well as a dynastic side to
the Magian revolt, though whether the false Smerdis restored the old
worship of the land, which he found in danger of being supplanted by
Zoroastrianism or the worship of Ahura Mazda, may still be doubtful.
In any event, the reformation or counter-reformation made by Darius
did not succeed in entirely uprooting the old Magian faith, for
Herodotus speaks of the Magi as still being in his time the priestly
caste among the Persians, and as acting as diviners and sacrificers
to the Achaemenian kings who ruled Persia up to Alexander’s
Conquest[808].
The Magian religion as it appears in Herodotus and other Greek
authors, however, seems to have shown none of the hostility to the
powers of darkness so apparent in the religious literature collected
by the Sassanian kings. “The whole circuit of the firmament” was,
according to Herodotus, their greatest god or Zeus; and he says that
they also “sacrifice to the sun and moon, to the earth, to fire and
water, and to the winds”; but that “they do not, like the Greeks,
believe the gods to have the same nature as men[809].” He also tells
us that later they borrowed from the Arabians and the Assyrians the
worship of a goddess whom he calls Mitra, and although he is
probably wrong as to the origin and sex of this deity, his evidence
shows that Semitic admixture counted for something in the Magian
worship. In other respects, the Magian seems to have been a
primitive faith given up to the worship of the powers of nature or
elements, which it did not personify in the anthropomorphic manner
of either the Semites or the Greeks, and to have paid little attention
to public ceremonies or ritual. It follows therefore that, like the
religions of many uncivilized people of the present day, it would draw
no very sharp distinction between good and evil gods, and would be
as ready to propitiate or make use of the evil, that is those hostile to
man, as the good or benevolent. Plutarch, who describes the religion
of the Magi more than three centuries after Herodotus, when the
name of Zoroaster the Persian prophet and the dualistic belief
favoured by his teaching had long been popularly known in the West,
says that the Magi of his time held Mithras to be the “Mediator” or
intermediary between “Oromazes” or Light on the one hand, and
“Areimanios” or Darkness and Ignorance on the other, and that they
used to make bloody sacrifices to the last-named in a place where
the sun never comes[810]. It is easy to see how such a cult, without
the control of public ceremonies and with its unabashed traffic with
the powers of evil, would be likely to degenerate into compulsion or
magic.
There was, however, another popular superstition or belief which,
about the time when Mithraism made its appearance in Europe, had
spread itself over Western Asia. This was the idea that the positions
and changes of the heavenly bodies exercise an influence over the
affairs of the world and the lot both of kingdoms and individual men.
It probably began in Babylonia, where the inhabitants had from
Sumerian times shown themselves great observers of the stars, and
had been accustomed to record the omens that they drew from their
motions for the guidance of the kings[811]. This kind of divination—or
astrology to call it by a familiar name—received a great impulse after
Alexander’s Conquest, in the first place from the break up of the
Euphratean priestly colleges before referred to, and the driving out of
the lesser priests therein to get their own living, and then from the
fact that the scientific enquiry and mathematical genius of the
Greeks had made the calculation of the positions of the heavenly
bodies at any given date and hour a fairly simple matter to be
determined without direct observation[812]. It was probably no mere
coincidence that the Chaldaei and the Mathematici, as the
astrologers called themselves, should have swarmed at Rome under
just those emperors in whose reigns Mithraism began to push itself
to the front[813].
While we may be sure that these factors, the religion of the Magi, the
practice of magic, and the astrological art, all counted in the
composition of the worship of Mithras, we yet know but very little of
its tenets. No work has come down to us from any devotee of
Mithras which will give us the same light on the way his worshippers
regarded him that the romance of Apuleius and the encomium of
Aelius Aristides have cast on the mental attitude of the devotees of
the Alexandrian cult. The extensive books of Eubulus and Pallas on
Mithras and the history of his worship, which Porphyry tells us were
extant from the reign of Hadrian down to his own time[814], are
entirely lost, and our only source of information, except a very few
scattered notices in the Fathers and in profane writers like the
Emperor Julian and Porphyry himself, are the sculptures and
inscriptions which have been found in his ruined chapels. These
texts and monuments the scholarly care of M. Cumont has gathered
into two large volumes, which will always remain the chief source
from which later enquirers must draw their materials[815]. From their
study he comes to the conclusion that, in the religion of Mithras,
there figured above him the Mazdean gods of good and evil
respectively called in the Zend Avesta Ahura Mazda and Angro
Mainyus, or in more familiar language, Ormuzd and Ahriman. Behind
and above these again, he would place a Supreme Being called
Zervan Akerene or Boundless Time, who seems to be without
attributes or qualities, and to have acted only as the progenitor of the
opposing couple. This is at first sight very probable, because the
Orphic doctrine, which, as we have seen, made Chronos or Time the
progenitor of all the gods, was widely spread in Asia Minor before
Alexander’s Conquest, and the Persian colonies formed there under
his successors must therefore have come in frequent contact with
this most accommodating of schools[816]. Traditions of a sect of
Zervanists in Western Asia, who taught that all things came from
Infinite Time, are also to be found[817]. But most of these are recorded
after Mithraism had become extinct; and M. Cumont’s proofs of the
existence of this dogma in the European religion of Mithras can be
reduced on final analysis to a quotation from a treatise by Theodore,
the Christian bishop of Mopsuestia who died in 428 A.D., directed, as
it would seem, against the “Magi” of his time, in which he admits that
their dogmas had never been written, and that the sectaries in
question, whom he calls Magusaeans, said “sometimes one thing
and deceived themselves, and sometimes another and deceived the
ignorant[818].” M. Cumont’s identification of the lion-headed statue
often found in Mithraic chapels with the Supreme God of the system
has been shown elsewhere to be open to serious question, and the
figure itself to be susceptible of another interpretation than that which
he would put upon it[819]. On the whole, therefore, while M. Cumont’s
mastery of his subject makes it very dangerous to differ from him, it
seems that his theory of a Boundless Time as the pinnacle of the
Mithraist pantheon cannot be considered as proved.
Whether Ormuzd and Ahriman played any important part in the
Roman worship of Mithras is also doubtful. With regard to the first-
named, both Greeks and Romans knew him well and identified him
unhesitatingly with Zeus and Jupiter[820]. Hence we should expect to
find him, if represented at all on the Mithraic sculptures, with the
well-known features, the thunderbolt, and the eagle, which long
before this time had become the conventional attributes of the
Roman as well as of the Homeric father of gods and men. We are
not entirely disappointed, for we find in a bas-relief formerly in a
chapel of Mithras at Sissek (the ancient Sissia in Pannonia) and now
in the Museum at Agram, the bull-slaying scene in which Mithras
figures and which will be presently described, surmounted by an
arch on which is ranged Jupiter seated on his throne, grasping the
thunderbolt, wielding the sceptre, and occupying the place of honour
in a group of gods among whom we may distinguish Mars and
Mercury[821]. In another bas-relief of the same scene, now at the
Rudolfinum in Klagenfurt, he is depicted in a similar position in an
assembly of the gods, which although much mutilated seems to
show Zeus or Jupiter in the centre with Hera or Juno by his side[822].
But the most conclusive of these monuments is the great bas-relief
found at Osterburken in the Odenwald, wherein the arch
surmounting the usual bull-slaying scene contains an assembly of
twelve gods with Zeus in the centre armed with thunderbolt and
sceptre, while around him are grouped Apollo, Ares, Heracles, Hera,
Athena, Aphrodite, Nike, Poseidon, Artemis, Hades, and perhaps
Persephone[823]. When by the side of these we put the many
inscriptions left by the legionaries to “the holy gods of the fatherland,
to Jupiter best and greatest, and to the Unconquered One”; to
“Jupiter best and greatest, and to the divine Sun, the Unconquered
Sun,” and other well-known names of Mithras, there can be no doubt
that his worshippers used to adore him together with the head of the
Roman Pantheon, and that they considered Mithras in some way the
subordinate of or inferior to Jupiter[824]. Yet there is nothing to show
that the Mithraists as such identified in any way this Jupiter Optimus
Maximus with the Persian Ahura Mazda, Oromasdes, or Ormuzd, or
that they ever knew him by any of these outlandish names.
The case is different with Ormuzd’s enemy Ahriman, who evidently
was known by his Persian name to the Roman worshippers of
Mithras. In the Vatican can be seen a triangular marble altar
dedicated by a clarissimus named Agrestius who was a high-priest
of Mithras, to “the god Arimanius[825],” and altars with similar
inscriptions have been found at Buda-Pesth[826]. At a Mithraic chapel
in York also, there was found a statue, now in the Museum of the
Philosophical Society in that city, which bears an inscription to the
same god Arimanius[827]. There is therefore fairly clear evidence that
the Mithraists recognized Ahriman under his Persian name, and that
they sacrificed to him, as Plutarch said the “magi” of his time did to
the god whom he calls Hades[828], and this agrees with Herodotus’
statement that the Persians used to do the same to “the god who is
said to be beneath the earth[829].” Although this gave occasion to the
Christian Fathers to accuse the Mithraists of worshipping the devil,
we are not thereby bound to conclude that they looked upon
Arimanius as an essentially evil being. It seems more probable that
they considered him, as the Greeks did their Hades or Pluto, as a
chthonian or subterranean power ruling over a place of darkness and
discomfort, where there were punishments indeed, but not as a deity
insusceptible of propitiation by sacrifice[830], or compulsion by other
means such as magic arts[831]. It has been shown elsewhere that his
image in a form which fairly represents his attributes in this capacity
appears with some frequency in the Mithraic chapels, where a
certain amount of mystery attached to its exhibition[832]. It seems to
follow from these considerations that the worshippers of Mithras
attributed to their special god no inferiority to Ahriman as M.
Cumont’s argument supposes, and that the only power whom they
acknowledged as higher than Mithras himself was the Roman
equivalent of Ormuzd, the Jupiter Optimus Maximus adored
throughout the Roman Empire of their time as the head of the
Pantheon[833].
The connection of Mithras with the sun is also by no means easy to
unravel. The Vedic Mitra was, as we have seen, originally the
material sun itself, and the many hundreds of votive inscriptions left
by the worshippers of Mithras to “the unconquered Sun Mithras[834],”
to the unconquered solar divinity (numen) Mithras[835], to the
unconquered Sun-God (deus) Mithra[836], and allusions in them to the
priests (sacerdotes), worshippers (cultores), and temples (templum)
of the same deity leave no doubt open that he was in Roman times a
sun-god[837]. Yet this does not necessarily mean that he was actually
the day-star visible to mankind, and the Greeks knew well enough
how to distinguish between Apollo the god of light who was once at
any rate a sun-god, and Helios the Sun itself[838]. On the Mithraic
sculptures, we frequently see the unmistakable figure of Mithras
riding in the chariot of the Sun-God driven by the divinity with long
hair and a rayed nimbus, whom we know to be this Helios or his
Roman equivalent, going through some ceremony of consecration
with him, receiving messages from him, and seated side by side with
him at a banquet which is evidently a ritual feast. M. Cumont
explains this by the theory that Mithras, while in Persia and in the
earliest Aryan traditions the genius of the celestial light only[839], no
sooner passed into Semitic countries and became affected by the
astrological theories of the Chaldaeans, than he was identified with
their sun-god Shamash[840], and this seems as reasonable a theory
as can be devised. Another way of accounting for what he calls the
“at first sight contradictory proposition” that Mithras at once was and
was not the sun[841], is to suppose that while the Mithraists wished
those who did not belong to their faith to believe that they
themselves worshipped the visible luminary, they yet instructed their
votaries in private that he was a deity superior to it and in fact the
power behind it. As we shall see, the two theories are by no means
irreconcilable, although absolute proof of neither can yet be offered.
One can speak with more certainty about the Legend or mythical
history of Mithras which M. Cumont has contrived with rare acumen
to reconstruct from the monuments found in his chapels. It is
comprised in eleven or twelve scenes or tableaux which we will take
in their order[842]. We first see the birth of the god, not from the head
of his father Zeus like Athena, or from his thigh like Dionysos, but
from a rock, which explains his epithet of “Petrogenes” or rock-born.
The god is represented in this scene as struggling from the rock in
which he is embedded below the waist, and always uplifts in one
hand a broad knife of which we shall afterwards see him make use,
and in the other a lighted torch[843]. He is here represented as a boy,
and wears the Phrygian cap or so-called cap of liberty which is his
distinctive attribute, while the torch is doubtless, as M. Cumont
surmises, symbolical of the light which he is bringing into the
world[844]. The rock is sometimes encircled by the folds of a large
serpent, probably here as elsewhere a symbol of the earth, and is in
the Mithraic chapel discovered at Housesteads in Northumberland
represented in the form of an egg, the upper part remaining on the
head of the nascent god like an egg-shell on that of a newly-hatched
chicken[845]. This is probably due to some confusion or identification
with the Orphic legend of the First-born or Phanes who sprang from
the cosmic egg; but the central idea of the rock-birth seems to be
that of the spark, hidden as it were in the stone and leaping forth
when struck. In one or two examples of the scene, the miraculous
birth is watched by a shepherd or shepherds, which leads M.
Cumont to draw a parallel between this and the Adoration of the
Shepherds at the Birth of Christ.
The next two scenes are more difficult to interpret with anything
approaching certainty. In one of them[846], Mithras is represented as
standing upright before a tree from which he cuts or tears a large
branch bearing leaves and fruit. He is here naked, save for the
distinctive cap; but immediately after, he is seen emerging from the
leafage fully clothed in Oriental dress. In the next scene—the relative
order of the scenes seems settled by the places they most often
occupy on different examples of the same sculptures[847]—Mithras in
the Phrygian cap, Persian trousers, and flowing mantle generally
worn by him, kneels on one knee drawing a bow, the arrows from
which strike a rock in the distance and draw from it a stream of water
which a kneeling man receives in his hands and lifts to his mouth[848].
Several variants of this scene exist, in one of which a suppliant is
kneeling before the archer-god and raising his hands towards him as
if in prayer; while in another, the rock may well be a cloud. M.
Cumont can only suggest with regard to these scenes, that the first
may be an allusion to the Fall of Man and his subsequently clothing
himself with leaves as described in the Book of Genesis, and that the
second scene may depict a prolonged drought upon earth, in which
man prays to Mithras and is delivered by the god’s miraculous
production of rain. He admits, however, that this is pure conjecture,
and that he knows no Indian, Persian, or Chaldaean legend or myth
to which the scenes in question can be certainly attached. It seems
therefore useless to discuss them further here.
Passing on, we come to a series of scenes, the meaning of which is
more easily intelligible. In all of these a bull plays a principal part. It is
abundantly clear that this bull is no terrestrial creature, but is the
Goshurun or Heavenly Bull of the Zend Avesta, from whose death
come forth not only man, but beasts, trees, and all the fruits of the
earth[849]. In the Mithraic sculptures, we see the Bull first sailing over
the waters in a cup-shaped boat[850] like the coracles still used on the
Euphrates, or escaping from a burning stable to which Mithras and a
companion have set fire[851]. Then he is depicted grazing peaceably
or raising his head now and then as if alarmed by some sudden
noise[852]. Next he is chased by Mithras, who seizes him by the
horns, mounts him[853], and after a furious gallop casts him over his
shoulders, generally holding him by the hind legs so that the horned
head dangles to the ground[854]. In this position, he is taken into the
cave which forms the chapel of Mithras.
Here, if the order in the most complete monuments be followed, we
break off to enter upon another set of scenes which illustrate the
relations between Mithras and the sun[855]. In what again seems to be
the first in order, we see Mithras upright with a person kneeling
before him who, from the rayed nimbus round his head, is evidently
the god Helios or Sol[856]. In one representation of this scene, Mithras
extends his left hand towards this nimbus as if to replace it on the
head of its wearer[857] from which it has been displaced in yet another
monument[858], while in the other, he displays an object not unlike a
Phrygian cap which may, however, be, as M. Cumont suggests,
something like a water-skin[859]. Generally, Mithras is represented as
holding this object over the bared head of the kneeling Sun-God, as
if to crown him with it[860]. Then we find Mithras with the ray-crowned
Sun-God upright beside him, while he grasps his hand in token, as it
would seem, of alliance or friendship[861]. If we accept the hint
afforded by the theory that the rock yielding water on being split by
the arrows of Mithras is really a cloud producing the fertilizing rain,
we may imagine that we have here the unconquered god removing
clouds which obscure the face of the great life-giving luminary and
restoring to him the crown of rays which enables him to shed his
kindly light upon the earth. The earth would thus be made fit for the
creation of man and other animals which, as we shall see, follows;
but in any event, the meaning of the scene which shows the alliance
is, as M. Cumont has pointed out, not doubtful[862]. In one monument,
where Mithras grasps the hand of the person we have identified with
the Sun-God before an altar, he at the same time draws his sword,
as if to perform the exchange of blood or blood-covenant usual in the
East on swearing alliance[863]. Possibly the crowning scene, as M.
Cumont also suggests[864], is to be connected with Tertullian’s
statement that in the initiation of the Mithraist to the degree of miles
or soldier, he was offered at the sword’s point a crown, which he cast
away from him saying that Mithras was his crown. If so, it would
afford some proof that the initiate here, as in the mysteries of Isis,
was made to impersonate the sun, which is on other grounds likely
enough.
We return to the scenes with the Bull, which here reach their climax.
This is the sacrifice of the Bull by Mithras, which forms the central
point of the whole legend. Its representation, generally in bas-relief,
was displayed in the most conspicuous position in the apse of the
Mithraic chapel, where it occupied the place of the modern altar-
piece, and such art as the Roman sculptors succeeded in displaying
was employed to make it as impressive and as striking as
possible[865]. It shows the god grasping with his left hand the nostrils
of the beast, and kneeling with his left knee in the middle of the Bull’s
back, while with his right hand he plunges the broad-bladed dagger
with which he was armed at his birth into the Bull’s shoulder[866]. A
dog leaps forward to lap the blood flowing from the wound, while at
the same time a scorpion seizes the Bull by the genitals. A serpent
also forms part of the group, but his position varies in the different
monuments, while that of the other animals does not. Sometimes, he
lifts his head towards the blood, as if to share it with the dog,
sometimes he is extended along the ground beneath the Bull’s belly
in apparent indifference to the tragedy enacted above him[867]. Before
the Bull stands generally a youth clothed like Mithras himself in
Phrygian cap, tunic, and mantle, as well as the anaxyrides or tight
trousers in which the Greeks depicted most Easterns, while another
youth similarly attired stands behind the dying victim. These two
human figures are alike in every particular save that one of them
bears a torch upright with the flame pointing upwards, while the other
holds a similar torch reversed so that the flame juts towards the
earth. We know from a Latin inscription that the torch-bearer with
uplifted torch was called Cautes, he with the reversed one
Cautopates, but of neither name has any satisfactory derivation or
etymology yet been discovered[868].
The meaning of the group as a whole can, however, be explained by
the documents of the later Persian religion. The Bundahish tells us
that Ahura Mazda created before all things the Bull Goshurun, who
was killed by Ahriman, the god of evil, and that from his side came
forth Gayômort, first of men, while from his tail there issued useful
seed-plants and trees, from his blood the vine, and from his seed the
different kinds of beasts[869]. Save that the bull-slayer is here not the
god of evil but the lord of light himself, the myth is evidently the same
in the Mithraic bas-reliefs, for in some of the earliest monuments the
Bull’s tail is actually shown sprouting into ears of wheat, while in
others the production of animals as a consequence of the Bull’s
death may be indicated, as well as the birth of the vine[870]. That the
dog plays the part of the guardian of the Bull’s soul is probable from
what we know of later Persian beliefs[871], while the scorpion as the
creature of Ahriman may be here represented as poisoning the seed
of future life at its source[872]. That Mithras is not supposed to kill the
Bull from enmity or other personal reasons, but in obedience to
orders from some higher power, is shown by the listening pose of his
head during the sacrifice. This is M. Cumont’s opinion[873], as also
that the serpent here takes no active part in the affair, but is merely a
symbolic representation of the earth[874]. The whole drama is clearly
shown as taking place in a cave or grotto, as appears from the arch
of rocks which surmounts, and, as it were, acts as a frame to, the
Tauroctony or bull-slaying scene in most Mithraic chapels. This cave,
according to Porphyry, represents the universe.
The Legend, however, does not end with the death of the Bull. In the
chapel at Heddernheim, the great slab on which the Tauroctony is
sculptured in bas-relief is pivoted so as to swing round and display
on its other face another scene which we find repeated in a slightly
different form on many monuments[875]. Mithras and the Sun-God are
here shown as partaking of a ritual feast or banquet in which grapes
seem to figure. At Heddernheim, the grapes are tendered to the two
gods over the body of the dead bull by the two torch-bearing figures
Cautes and Cautopates, while on an arch above them various
quadrupeds, dogs, a boar, a sheep, and a cow, are seen springing
into life. In other monuments, the same scene generally appears as
a banquet at which Mithras and Helios are seated side by side at a
table sometimes alone, but at others in company with different
persons who can hardly be any other than initiates or
worshippers[876]. That this represents some sort of sacrament where
a drink giving immortality was administered seems probable, and its
likeness to representations of the Last Supper is sufficient to explain
the complaint of Justin Martyr and other Fathers that the devil had
set on the Mithraists to imitate in this and other respects the Church
of Christ[877]. The final scene of all comes when we see Mithras
arresting the glorious chariot of the Sun-God drawn by four white
horses, and, mounting therein, being driven off by the ray-crowned
Helios himself to the abode of light above the firmament[878]. In this
also, it is easy to see a likeness between representations of the
Ascension of Mithras and that of Elijah or even of Christ[879].
However this may be, the Legend of Mithras, as thus portrayed,
shows with fair closeness the belief of his worshippers as to his
place in the scheme of the universe. Mithras was certainly not the
Supreme God, a rank in the system filled by Ahura Mazda, or his
Latin counterpart, Jupiter Best and Greatest[880]. But this being, like
the Platonic Zeus and the Gnostic Bythos, was considered too great
and too remote to concern himself with the doings of the visible
universe, in which Mithras acts as his vicegerent. Whether Mithras
was or was not considered as in some sort the double or antitype of
the Supreme Being cannot be said; but it is worth noticing that in the
Vedas, as among the Hittites, Varuna and Mitra form an inseparable
couple who are always invoked together, and that the same seems
to have been the case with Ahura Mazda and Mithra in the oldest
religious literature of the Persians[881]. It may therefore well be that
the learned doctors of the Mithraic theology regarded their Supreme
Being and Mithras as two aspects of the same god, an idea that, as
we have seen, was current at about the same period among the
Gnostics. It is, however, impossible to speak with certainty on such a
point in the absence of any writings by persons professing the
Mithraic faith, and it is highly improbable that the rugged soldiers
who formed the majority of the god’s worshippers ever troubled
themselves much about such questions. For them, no doubt, and for
all, perhaps, but a few carefully-chosen persons, Mithras was the
Demiurge or Divine Artizan of the universe[882], which he governs in
accordance with the laws of right and justice, protecting and
defending alike man and those animals and plants useful to him
which Mithras has himself created from his own spontaneous
goodness. Hence he was the only god to whom they admitted
allegiance, and although the existence of other heavenly beings was
not denied, it is probable that most of them were looked upon as
occupying at the best a position less important to us than that of
Mithras himself.
It is probable, moreover, that all the scenes in the Mithraic sculptures
in which we have seen the god taking part were considered as being
enacted before the creation of man and in some heaven or world
midway between the abode of Infinite Light and this earth. That the
grotto into which Mithras drags the primordial Bull is no earthly
cavern is plain from Porphyry’s remark that the Mithraic cave was an
image of the universe[883], as well as from the band of zodiacal
figures or the arch of rocks which sometimes encloses the bas-
reliefs, the sky being looked upon by the Babylonians as a rocky
vault. The sun and moon in their respective chariots also appear
above the principal scene; and a further hint as to its whereabouts
may be found in the fact that the flowing mantle of Mithras is
sometimes depicted as spangled with stars, thereby indicating that
the scenes in which he appears are supposed to take place in the
starry firmament. Hence is explained the epithet of μεσίτης or
Mediator, which Plutarch gives him[884], and which should be
interpreted not as intercessor but as he who occupies a position
midway between two places[885]. That the higher of these in this case
was the Garôtman or abode of Infinite Light of the Avestic literature,
there can, it would seem, be no question; but what was the lower?
Although the statement must be guarded with all the reserves
imposed upon us in all matters relating to the religion of Mithras by